Humanization (OASis)

简介

Humanization (OASis)是抗体序列人源化评分工具，不依赖于有限的人源种系（germline）序列，而是基于海量的天然人类抗体库（Observed Antibody Space, OAS）。该数据库包含来自数百个受试者的数亿条序列，这使得它能捕捉到更丰富、更多样化的抗体序列空间。将待评估的抗体序列切割成所有可能的、长度为9个氨基酸的短肽（9-mer），将这些短肽放到庞大的OAS数据库中进行搜索，找出每个短肽在真实人类抗体库中出现的频率，以及携带该短肽的个体数量。如果一个序列中的大多数短肽在人类抗体库中都很常见，那么它的OASis评分就高，意味着“看起来很人类”，免疫原性风险较低；反之，如果含有大量在人类中罕见的短肽，则评分低，提示可能需要进一步人源化改造。

参数说明

FASTA File

上传抗体FASTA序列文件，用于进行抗体人源化分析。

Output File

输出文件名称。

Numbering Scheme

可变区编号规则，支持 kabat、chothia、imgt、aho。

CDR Definition

CDR 区域定义规则，支持 kabat、chothia、imgt、north。

Iterations

Sapiens 模型运行迭代次数，用于发现更多人源化突变。

Humanize CDRs

是否允许对 CDR 区域进行人源化突变。

Max Sequences

单次处理的最大序列数量，默认 1000，可通过环境变量 MAX_SEQS 设置。

Fasta Only

仅输出人源化后的 FASTA 序列文件，可加快处理速度。

Scores Only

仅输出 Sapiens 模型位置-残基评分 CSV 文件。

Mean Score

输出每条序列对应的平均 Sapiens 分数 CSV 文件。

Output File

结果说明

Sapiens.xlsx

Sapiens.xlsx 为默认输出的评分结果文件，包含 Overview、VH、VL、OASis Curves 和抗体名称对应的详细结果 Sheet。

Overview

Overview Sheet 包含抗体人源化前后的总体评分结果及相关参数信息。

VH或者VL

VH/VL Sheet 包含抗体重链可变区的编号及对应氨基酸残基信息。

OASis Curves

OASis Curves Sheet 包含抗体序列在不同 OASis 百分位条件下的数据。

序列名称

序列名称 Sheet 包含该抗体的详细分析结果。

alignments.txt

序列比对文件，包含原始抗体序列与人源化后序列的比对结果。

humanized.fa

人源化后的抗体序列 FASTA 文件。

Mean Score Only 模式

输出mean-score-only.csv文件内容包括：

Scores Only 模式

输出 scores-only.csv文件内容包括：

参考文献

• David Prihoda, Jad Maamary, Andrew Waight, Veronica Juan, Laurence Fayadat-Dilman, Daniel Svozil & Danny A. Bitton (2022) BioPhi: A platform for antibody design, humanization, and humanness evaluation based on natural antibody repertoires and deep learning, mAbs, 14:1, DOI: https://doi.org/10.1080/19420862.2021.2020203

Humanization (OASis)

Introduction

Humanization (OASis) is an antibody sequence humanization scoring tool. Unlike methods that rely on a limited set of human germline sequences, it is based on the large-scale natural human antibody repertoire database, Observed Antibody Space (OAS). This database contains hundreds of millions of sequences from hundreds of subjects, allowing it to capture a broader and more diverse antibody sequence space.

The antibody sequence to be evaluated is divided into all possible peptides consisting of nine amino acids (9-mers). These peptides are then searched against the extensive OAS database to determine the frequency of each peptide in the natural human antibody repertoire, as well as the number of individuals carrying that peptide.

If most peptides in a sequence are commonly found in the human antibody repertoire, the sequence receives a high OASis score, indicating that it appears more “human-like” and may have a lower immunogenicity risk. Conversely, if a sequence contains a large number of peptides that are rarely observed in humans, it receives a lower score, suggesting that further humanization may be required.

Parameter

FASTA File

Upload an antibody FASTA sequence file for antibody humanization analysis.

Output File

The name of the output file.

Numbering Scheme

The numbering scheme used for the variable region. Supported options include kabat, chothia, imgt, and aho.

CDR Definition

The definition scheme used for CDR regions. Supported options include kabat, chothia, imgt, and north.

Iterations

The number of iterations performed by the Sapiens model to identify additional humanizing mutations.

Humanize CDRs

Whether humanizing mutations are allowed within CDR regions.

Max Sequences

The maximum number of sequences processed in a single run. The default value is 1000 and can be configured using the MAX_SEQS environment variable.

Fasta Only

Output only the humanized FASTA sequence file to accelerate processing.

Scores Only

Output only the CSV file containing Sapiens model position-residue scores.

Mean Score

Output a CSV file containing the average Sapiens score for each sequence.

Result

Sapiens.xlsx

Sapiens.xlsx is the default scoring result file. It contains the Overview, VH, VL, OASis Curves, and detailed result Sheets corresponding to the antibody names.

Overview

The Overview Sheet contains overall scoring results before and after antibody humanization, as well as related parameter information.

VH or VL

The VH/VL Sheets contain the numbering information and corresponding amino acid residues for the antibody heavy-chain or light-chain variable region.

OASis Curves

The OASis Curves Sheet contains sequence data under different OASis percentile conditions.

Sequence Name

The Sequence Name Sheet contains detailed analysis results for the corresponding antibody.

alignments.txt

A sequence alignment file containing alignment results between the original antibody sequences and the humanized sequences.

humanized.fa

A FASTA file containing the humanized antibody sequences.

Mean Score Only Mode

The output mean-score-only.csv file contains the following fields:

Scores Only Mode

The output scores-only.csv file contains the following fields:

References

• David Prihoda, Jad Maamary, Andrew Waight, Veronica Juan, Laurence Fayadat-Dilman, Daniel Svozil & Danny A. Bitton (2022). BioPhi: A platform for antibody design, humanization, and humanness evaluation based on natural antibody repertoires and deep learning. mAbs, 14(1), DOI: https://doi.org/10.1080/19420862.2021.2020203

Structure Prediction (OpenDDE)

简介

OpenDDE（Open Drug Discovery Engine）是 Aureka AI Research 推出的开源全原子生物分子基础模型，以多分子共折叠为核心，用于蛋白质、抗体、抗原、核酸和小分子等生物分子复合物的联合结构预测。模块将结构预测作为药物发现流程中的统一结构推理层，为分子设计、亲和力预测和结构优化等任务提供基础。

它解决了什么问题？

药物发现中的核心难题之一，是准确理解不同生物分子之间如何结合。

OpenDDE 试图解决三个关键问题：

1. 从序列推断复合物的全原子结构，而不只预测单个蛋白质。
2. 理解分子界面的几何关系，包括朝向、间距、互补性和空间冲突。
3. 建立可扩展的结构推理底座，让结构预测与分子设计共享同一套模型框架。

核心能力

1. 全原子结构预测

OpenDDE 可以联合建模蛋白质、核酸、配体、离子和修饰残基，并生成完整的原子坐标。

2. 抗体–抗原共折叠

这是 OpenDDE 当前最突出的能力。模型在多个抗体–抗原基准上表现优于报告中对比的开源模型，并在新发布、低同源的生物体系上保持较强泛化能力。

实际表现如何？

抗体–抗原基准

在按模型排序分数选择结果时，OpenDDE 的 DockQ 成功率为：

在 Oracle 选择下，即从多个候选中选择真实 DockQ 最优结果，成功率分别达到：

FoldBench 综合表现

OpenDDE 的提升主要集中在需要跨分子界面推理的任务，尤其是蛋白质–蛋白质和抗体–抗原体系。

参数说明

Single Mode

Single Mode 用于预测一个单体或复合物。Protein、DNA、RNA、Ligand 和 Ion 中至少需要提供一种分子输入；Modification、Covalent Bond 和 Template 只能补充已有分子信息，不能单独构成预测任务。

Protein、DNA 或 RNA FASTA 中的多条记录，以及配体和离子文件中的多个实体，都会合并到同一个待预测体系中。

Protein Sequence

蛋白质序列使用 FASTA 格式。每条记录表示同一体系中的一条独立蛋白链：

>protein1
EVQLVESGGGVVQPGRSLRLS
>protein2
NFMLTQPHSVSVL
>peptide
ELFPQWHLPIKIAAIIAS

允许的蛋白字符为 ACDEFGHIKLMNPQRSTVWYX。存在蛋白输入时，平台会自动查询并缓存蛋白 MSA。

DNA Sequence

DNA 序列使用 FASTA 格式，每条记录表示同一体系中的一条独立 DNA 链：

>dna1
ATGCGTACGT
>dna2
NNATGC

允许的 DNA 字符为 ATGCNX。

RNA Sequence

RNA 序列使用 FASTA 格式，每条记录表示同一体系中的一条独立 RNA 链：

>rna1
AUGCGUACGU

允许的 RNA 字符为 AUGCNX。当前平台不会为 RNA 查询 MSA。

FASTA 通用要求

• 每条序列前必须有以 > 开头的标题行；
• 不允许空记录，也不允许在第一个标题行前出现序列；
• 空行和以 ; 开头的注释行会被忽略；
• 序列不区分大小写，读取后统一转换为大写；
• 单个 FASTA 文件不能超过 10 MiB。

Ligand

Ligand 支持 SMILES、CCD Code 和三维配体结构文件。文本文件中每个非空、非注释行表示一个独立配体实体：

CC(=O)O
CCD,ATP
CCD_ATP_HY3

三维配体文件支持 .pdb、.sdf、.mol 和 .mol2。CCD,ATP,HY3 会被解析为一个 CCD_ATP_HY3 配体实体。配体蛋白或多肽必须通过 Protein Sequence 输入。

Ion

支持逗号或换行分隔的离子代码，可使用 :数量 指定同种离子的数量，例如 MG:2,ZN,CU:3。

Modification

每行使用 实体编号,CCD编号,1-based位置。实体按照 Protein → DNA → RNA → Ligand → Ion 的顺序编号，且修饰只能作用于 Protein、DNA 或 RNA。

Covalent Bond

每行使用 实体编号,位置编号,原子名称;实体编号,位置编号,原子名称。实体编号与 Modification 保持一致，小分子或离子通常使用位置 1。

Template

Template 仅用于蛋白质输入，支持 .a3m 或 .hhr 文件，不接受 PDB 或 CIF 文件。模板信息会应用到输入中的第一条蛋白链。

Format

用于输出蛋白质结构的格式，可选 cif 或 pdb，默认为 cif。

Seed

用于控制预测过程随机性的随机种子，最多使用前 5 个，默认为1,39,248,1970,20967。

Checkpoint

模型参数，general 或 abag。 abag为在抗原抗体复合物上微调后的版本，默认为general。

MSA

用于跳过MSA检索，默认为使用MSA。

Batch Mode

Batch Mode 用于在一个任务中批量预测多个蛋白单体或蛋白复合物。该

Protein Sequence

蛋白质序列文件，FASTA 格式，支持多条序列。每条记录表示一个需要预测的结构，并且每条记录名称必须唯一。如果某条记录包含多条链，则不同链之间应使用英文冒号 : 连接。示例内容：

>1
EVQLVESGGGVVQPGRSLRLS:NFMLTQPHSVSVL:ELFPQWHLPIKIAAIIAS
>2
YYCAKDRRDMGYFQHWGQGTLVTVSSQWYQQRPGSAPTTVIYEDNQRPSGTPPTFMIAVFLPIVVLIFKSILFL

这表示需要预测两个结构。第一条记录名称为 1，包含三条由英文冒号分隔的蛋白质链。第二条记录名称为 2，包含一条链。

Ligand

包含小分子信息的 TXT 格式文本文件。支持 SMILES 或 CCD Code。如果使用 SMILES 格式，每行应包含一个小分子；如果使用 CCD Code，每行可以包含一个或多个小分子，多个小分子之间用英文逗号分隔。每一行表示一个需要预测的结构，并且每一行必须以唯一名称开头。该名称可以与 RNA、DNA 或 Protein 参数中的名称相同，表示该行中的所有配体属于同一个结构。名称和所有配体之间使用英文冒号 : 分隔。示例内容：

1:CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP,HY3
lig:CC1CC(CN2NC[N+](=C3CCC(OC(F)(F)F)CC3)C2=O)CC(C)C1OC(C)(C)C(=O)[O-]

这表示两条记录。第一条名称为 1，包含三个配体，其中一个为 SMILES，两个为 CCD 代码。由于该名称存在于 RNA、DNA 或 Protein 示例记录中，因此表示它们属于同一个结构。第二条记录名称为 lig，包含一个 SMILES 格式的配体。由于该名称未出现在 RNA、DNA 或 Protein 示例记录中，因此表示一个新的结构。

注意：
1.不适用于配体蛋白或多肽的氨基酸序列格式输入。
2.在预测每个结构的序列数量相同的情况下，可为所有待预测结构设置相同的小分子信息，设置方式为输入一行小分子信息（可多个），且不设置结构名称，如CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP，表示为所有结构都加入小分子CC(=O)OC1C[NH+]2CCC1CC2与ATP。

Format

用于输出蛋白质结构的格式，可选 cif 或 pdb，默认为 cif。

Seed

用于控制预测过程随机性的随机种子，最多使用前 5 个，默认为1,39,248,1970,20967。

Checkpoint

模型参数，general 或 abag。 abag为在抗原抗体复合物上微调后的版本，默认为general。

MSA

用于跳过MSA检索，默认为使用MSA。

Enhanced Mode

Enhanced Mode 用于对一个单体或复合物进行高精度、多候选预测。该模式固定使用 500 个唯一随机 seed，不能提交自定义 Seed；其余输入均按一个预测体系组织。Enhanced Mode 会生成大量候选结构，运行时间和计算量明显高于 Single Mode，适合对候选覆盖度和预测精度要求较高的任务。

Protein、DNA、RNA、Ligand 和 Ion 中至少需要提供一种分子输入；Modification、Covalent Bond 和 Template 不能单独构成预测任务。

Protein Sequence

使用 FASTA 格式，每条记录表示同一体系中的一条独立蛋白链。允许字符为 ACDEFGHIKLMNPQRSTVWYX，平台会自动查询并缓存蛋白 MSA。

>protein1
EVQLVESGGGVVQPGRSLRLS
>protein2
NFMLTQPHSVSVL

DNA Sequence

使用 FASTA 格式，每条记录表示同一体系中的一条独立 DNA 链。允许字符为 ATGCNX。

>dna1
ATGCGTACGT

RNA Sequence

使用 FASTA 格式，每条记录表示同一体系中的一条独立 RNA 链。允许字符为 AUGCNX，当前不查询 RNA MSA。

>rna1
AUGCGUACGU

Protein、DNA 和 RNA 均须遵循以下 FASTA 要求：

• 每条序列前必须有以 > 开头的标题行；
• 不允许空记录，也不允许在第一个标题行前出现序列；
• 空行和以 ; 开头的注释行会被忽略；
• 序列不区分大小写，读取后统一转换为大写；
• 单个 FASTA 文件不能超过 10 MiB。

Ligand

支持 SMILES、CCD,ATP / CCD_ATP 形式的 CCD Code，以及 .pdb、.sdf、.mol、.mol2 三维配体文件。每个有效输入行表示一个独立配体实体，配体蛋白或多肽必须通过 Protein Sequence 输入。

Ion

支持逗号或换行分隔的离子代码，可使用 :数量 指定同种离子的数量，例如 MG:2,ZN,CU:3。

Modification

每行使用 实体编号,CCD编号,1-based位置。实体按照 Protein → DNA → RNA → Ligand → Ion 的顺序编号，且修饰只能作用于 Protein、DNA 或 RNA。

Covalent Bond

每行使用 实体编号,位置编号,原子名称;实体编号,位置编号,原子名称。实体编号与 Modification 保持一致，小分子或离子通常使用位置 1。

Template

仅在存在 Protein Sequence 时支持 .a3m 或 .hhr 模板文件；不接受 PDB 或 CIF 模板。模板信息应用于第一条蛋白链。

Format

用于输出蛋白质结构的格式，可选 cif 或 pdb，默认为 cif 。

Checkpoint

模型参数，general 或 abag。 abag为在抗原抗体复合物上微调后的版本，默认为general。

MSA

用于跳过MSA检索，默认为使用MSA。

结果说明

Single Mode

Single Mode 输出一个体系排名前 5 的结果：

• rank_1.pdb 至 rank_5.pdb：按置信度排序的结构；
• rank_1_confidence.json 至 rank_5_confidence.json：对应结构的置信度信息；
• pred_scores_opendde.csv：排名前 5 的候选分数；
• final_results.tar.gz：全部有效 seed/sample 的 CIF 结构和对应置信度 JSON；
• result_manifest.json：任务输入、模型、参数和结果文件清单。

pred_scores_opendde.csv 字段

pDockQ2 的质量区间沿用 DockQ 分级：

Batch Mode

Batch FASTA 中的每条记录都会独立排序，并在 results/<记录名称>/ 下输出该记录的 rank_1.pdb 至 rank_5.pdb 及对应置信度 JSON。

任务根目录还会输出：

• pred_scores_opendde.csv：汇总全部 Batch 记录的有效候选分数；
• final_results.tar.gz：按记录名称和 seed 组织的 CIF 结构及对应置信度 JSON；
• result_manifest.json：所有 Batch 记录及结果文件清单。

Enhanced Mode

Enhanced Mode 与 Single Mode 使用相同的排名输出结构，但候选来自500 个 seed：

• rank_1.pdb 至 rank_5.pdb 及对应置信度 JSON；
• pred_scores_opendde.csv：排名前 5 的候选分数；
• final_results.tar.gz：全部有效 seed/sample 的 CIF 结构和对应置信度 JSON；
• result_manifest.json：任务参数和结果文件清单。

参考文献

• OpenDDE Project, Aureka AI Research. 2026, Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine, arXiv, arXiv:2607.03787,DOI:10.48550/arXiv.2607.03787

Structure Prediction (OpenDDE)

Introduction

Structure Prediction (OpenDDE) is a biomolecular complex structure prediction tool based on OpenDDE (Open Drug Discovery Engine). OpenDDE is an open-source, all-atom biomolecular foundation model developed by Aureka AI Research. With multi-molecule co-folding as its core capability, it supports joint structure prediction for biomolecular complexes such as proteins, antibodies, antigens, nucleic acids, and small molecules. The module uses structure prediction as a unified structural reasoning layer in drug discovery workflows, providing a foundation for molecular design, affinity prediction, and structure optimization.

OpenDDE is mainly used to address how different biomolecules bind to each other in drug discovery, including inferring all-atom complex structures from sequences, understanding molecular interface geometry, and building a scalable structural reasoning foundation.

Parameters

Single Mode

Single Mode is used to predict one monomer or complex. At least one molecular input must be provided among Protein, DNA, RNA, Ligand, and Ion. Modification, Covalent Bond, and Template can only supplement existing molecular information and cannot independently define a prediction task.

Multiple records in Protein, DNA, or RNA FASTA files, as well as multiple entities in ligand and ion files, are merged into the same prediction system.

Protein Sequence

Protein sequence file in FASTA format. Each record represents an independent protein chain in the same system. Allowed protein characters are ACDEFGHIKLMNPQRSTVWYX. When protein input is provided, the platform automatically searches and caches protein MSA.

>protein1
EVQLVESGGGVVQPGRSLRLS
>protein2
NFMLTQPHSVSVL
>peptide
ELFPQWHLPIKIAAIIAS

DNA Sequence

DNA sequence file in FASTA format. Each record represents an independent DNA chain in the same system. Allowed DNA characters are ATGCNX.

>dna1
ATGCGTACGT
>dna2
NNATGC

RNA Sequence

RNA sequence file in FASTA format. Each record represents an independent RNA chain in the same system. Allowed RNA characters are AUGCNX. The platform currently does not search MSA for RNA.

>rna1
AUGCGUACGU

General FASTA Requirements

Protein, DNA, and RNA inputs must follow these FASTA requirements:

• Each sequence must have a header line starting with >;
• Empty records are not allowed, and sequence content is not allowed before the first header line;
• Empty lines and comment lines starting with ; are ignored;
• Sequences are case-insensitive and converted to uppercase after reading;
• A single FASTA file must not exceed 10 MiB.

Ligand

Ligand input supports SMILES, CCD Code, and 3D ligand structure files. Each non-empty, non-comment line in the text file represents an independent ligand entity.

CC(=O)O
CCD,ATP
CCD_ATP_HY3

3D ligand files support .pdb, .sdf, .mol, and .mol2. CCD,ATP,HY3 is parsed as one CCD_ATP_HY3 ligand entity. Protein ligands or peptide ligands must be provided through Protein Sequence.

Ion

Ion input supports ion codes separated by commas or line breaks. The number of ions of the same type can be specified using :count, such as MG:2,ZN,CU:3.

Modification

Modification information. Each line uses entity index,CCD code,1-based position. Entities are indexed in the order Protein → DNA → RNA → Ligand → Ion, and modifications can only be applied to Protein, DNA, or RNA.

Covalent Bond

Covalent bond information. Each line uses entity index,position index,atom name;entity index,position index,atom name. Entity indices are consistent with Modification. Small molecules or ions usually use position 1.

Template

Protein template file, only used for protein input. .a3m and .hhr files are supported, while PDB and CIF files are not accepted. Template information is applied to the first protein chain in the input.

Format

Output structure file format. The default value is cif. Available values are cif and pdb.

Seed

Random seed used to control prediction randomness. Up to the first 5 seeds are used. The default value is 1,39,248,1970,20967.

Checkpoint

Model checkpoint. The default value is general. Available values are general and abag. abag is the version fine-tuned on antigen-antibody complexes.

MSA

Whether to skip MSA search. MSA is used by default.

Batch Mode

Batch Mode is used to predict multiple protein monomers or protein complexes in one task.

Protein Sequence

Protein sequence file in FASTA format, supporting multiple sequences. Each record represents one structure to be predicted, and each record name must be unique. If a record contains multiple chains, different chains should be connected with an English colon :

>1
EVQLVESGGGVVQPGRSLRLS:NFMLTQPHSVSVL:ELFPQWHLPIKIAAIIAS
>2
YYCAKDRRDMGYFQHWGQGTLVTVSSQWYQQRPGSAPTTVIYEDNQRPSGTPPTFMIAVFLPIVVLIFKSILFL

Ligand

Small molecule information file in TXT format. SMILES and CCD Code are supported. If SMILES format is used, each line should contain one small molecule. If CCD Code is used, each line may contain one or more small molecules, separated by English commas. Each line represents one structure to be predicted and must start with a unique name. The name and all ligands are separated by an English colon :.

1:CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP,HY3
lig:CC1CC(CN2NC[N+](=C3CCC(OC(F)(F)F)CC3)C2=O)CC(C)C1OC(C)(C)C(=O)[O-]

Notes:

1. This input is not applicable to amino acid sequence input for protein ligands or peptide ligands.
2. When the number of sequences is the same for each structure, the same small molecule information can be applied to all structures by providing one line without a structure name, such as CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP, which adds small molecule CC(=O)OC1C[NH+]2CCC1CC2 and ATP to all structures.

Format

Output structure file format. The default value is cif. Available values are cif and pdb.

Seed

Random seed used to control prediction randomness. Up to the first 5 seeds are used. The default value is 1,39,248,1970,20967.

Checkpoint

Model checkpoint. The default value is general. Available values are general and abag. abag is the version fine-tuned on antigen-antibody complexes.

MSA

Whether to skip MSA search. MSA is used by default.

Enhanced Mode

Enhanced Mode is used for high-accuracy, multi-candidate prediction of one monomer or complex. This mode always uses 500 unique random seeds and does not accept custom Seed input. All other inputs are organized as one prediction system. Enhanced Mode generates a large number of candidate structures and requires more runtime and computation than Single Mode, making it suitable for tasks that require higher candidate coverage and prediction accuracy.

At least one molecular input must be provided among Protein, DNA, RNA, Ligand, and Ion. Modification, Covalent Bond, and Template cannot independently define a prediction task.

Protein Sequence

Protein sequence file in FASTA format. Each record represents an independent protein chain in the same system. Allowed characters are ACDEFGHIKLMNPQRSTVWYX, and the platform automatically searches and caches protein MSA.

>protein1
EVQLVESGGGVVQPGRSLRLS
>protein2
NFMLTQPHSVSVL

DNA Sequence

DNA sequence file in FASTA format. Each record represents an independent DNA chain in the same system. Allowed characters are ATGCNX.

>dna1
ATGCGTACGT

RNA Sequence

RNA sequence file in FASTA format. Each record represents an independent RNA chain in the same system. Allowed characters are AUGCNX. RNA MSA is not searched currently.

>rna1
AUGCGUACGU

General FASTA Requirements

Protein, DNA, and RNA inputs must follow these FASTA requirements:

• Each sequence must have a header line starting with >;
• Empty records are not allowed, and sequence content is not allowed before the first header line;
• Empty lines and comment lines starting with ; are ignored;
• Sequences are case-insensitive and converted to uppercase after reading;
• A single FASTA file must not exceed 10 MiB.

Ligand

Ligand input supports SMILES, CCD Code in CCD,ATP / CCD_ATP format, and 3D ligand files in .pdb, .sdf, .mol, and .mol2 formats. Each valid input line represents an independent ligand entity. Protein ligands or peptide ligands must be provided through Protein Sequence.

Ion

Ion input supports ion codes separated by commas or line breaks. The number of ions of the same type can be specified using :count, such as MG:2,ZN,CU:3.

Modification

Modification information. Each line uses entity index,CCD code,1-based position. Entities are indexed in the order Protein → DNA → RNA → Ligand → Ion, and modifications can only be applied to Protein, DNA, or RNA.

Covalent Bond

Covalent bond information. Each line uses entity index,position index,atom name;entity index,position index,atom name. Entity indices are consistent with Modification. Small molecules or ions usually use position 1.

Template

Protein template file. .a3m and .hhr template files are supported only when Protein Sequence is provided. PDB and CIF templates are not accepted. Template information is applied to the first protein chain.

Format

Output structure file format. The default value is cif. Available values are cif and pdb.

Checkpoint

Model checkpoint. The default value is general. Available values are general and abag. abag is the version fine-tuned on antigen-antibody complexes.

MSA

Whether to skip MSA search. MSA is used by default.

Results

Single Mode

Single Mode outputs the top 5 ranked results for one system:

Batch Mode

Each record in the Batch FASTA is ranked independently. The corresponding rank_1.pdb to rank_5.pdb files and confidence JSON files are output under results/<record_name>/.

The task root directory also outputs:

Enhanced Mode

Enhanced Mode uses the same ranked output structure as Single Mode, but candidates are generated from 500 seeds:

The pred_scores_opendde.csv file includes the following information:

pDockQ2 quality ranges follow DockQ grading:

References

• OpenDDE Project, Aureka AI Research. 2026, Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine, arXiv, arXiv:2607.03787, DOI: 10.48550/arXiv.2607.03787

Pareto Set

简介

在多目标优化问题中，计算Pareto集，并进行排序。
Pareto 集：所有不被任何其他解支配的解的集合。
Pareto 支配（Pareto Dominance）：对于多目标优化问题，解 A 支配 解 B 当且仅当：

• A 在所有目标上都不比 B 差
• A 在至少一个目标上严格比 B 好

参数说明

CSV

进行Pareto分析的数据文件，CSV格式，包含多列（多目标）。

Columns

指定数据文件中，用于分析的多列（多目标）名称及排序方向，使用文本文件，每行定义一个列名，以及排序方向（1表示升序，数值越小越优；0表示降序，数值越大越优），示例如下：

ddG(ThermoMPNN),1
dP(ESMIF),0

表示参与多目标优化的共2列，其中ddG(ThermoMPNN)列使用升序（数值越小越优），dP(ESMIF)列使用降序（数值越大越优）。

Output

输出排序结果文件，默认为pareto_ranked.csv。

结果说明

输出排序结果文件pareto_ranked.csv，新增列名Pareto_Rank表示pareto集分层排序结果，先按Pareto_Rank升序排列，再按各目标列及其排序方向排序。

Pareto Set

Introduction

In multi-objective optimization problems, compute the Pareto set and perform ranking.
Pareto Set: The set of all solutions that are not dominated by any other solution.
Pareto Dominance: For a multi-objective optimization problem, solution A dominates solution B if and only if:

• A is no worse than B in all objectives
• A is strictly better than B in at least one objective

Parameters

CSV

Data file for Pareto analysis in CSV format, containing multiple columns (multiple objectives).

Columns

Specifies the column names and sorting directions for the multiple columns (multiple objectives) to be analyzed. Use a text file, with each line defining a column name and its sorting direction (1 for ascending, smaller values are better; 0 for descending, larger values are better). Example:

ddG(ThermoMPNN),1
dP(ESMIF),0

This indicates a total of 2 columns are involved in multi-objective optimization, where the ddG(ThermoMPNN) column uses ascending order (smaller values are better), and the dP(ESMIF) column uses descending order (larger values are better).

Output

Output file for the ranking results, default is pareto_ranked.csv.

Results

Outputs the ranking result file pareto_ranked.csv, with an added column Pareto_Rank representing the Pareto set hierarchical ranking results. The rows are first sorted by Pareto_Rank in ascending order, then by each objective column according to their specified sorting directions.

Structure Prediction (OpenDDE)

简介

OpenDDE（Open Drug Discovery Engine）是 Aureka AI Research 推出的开源全原子生物分子基础模型，以多分子共折叠为核心，用于蛋白质、抗体、抗原、核酸和小分子等生物分子复合物的联合结构预测。模块将结构预测作为药物发现流程中的统一结构推理层，为分子设计、亲和力预测和结构优化等任务提供基础。

它解决了什么问题？

药物发现中的核心难题之一，是准确理解不同生物分子之间如何结合。

OpenDDE 试图解决三个关键问题：

1. 从序列推断复合物的全原子结构，而不只预测单个蛋白质。
2. 理解分子界面的几何关系，包括朝向、间距、互补性和空间冲突。
3. 建立可扩展的结构推理底座，让结构预测与分子设计共享同一套模型框架。

核心能力

1. 全原子结构预测

OpenDDE 可以联合建模蛋白质、核酸、配体、离子和修饰残基，并生成完整的原子坐标。

2. 抗体–抗原共折叠

这是 OpenDDE 当前最突出的能力。模型在多个抗体–抗原基准上表现优于报告中对比的开源模型，并在新发布、低同源的生物体系上保持较强泛化能力。

实际表现如何？

抗体–抗原基准

在按模型排序分数选择结果时，OpenDDE 的 DockQ 成功率为：

在 Oracle 选择下，即从多个候选中选择真实 DockQ 最优结果，成功率分别达到：

FoldBench 综合表现

OpenDDE 的提升主要集中在需要跨分子界面推理的任务，尤其是蛋白质–蛋白质和抗体–抗原体系。

参数说明

Single Mode

Single Mode 用于预测一个单体或复合物。Protein、DNA、RNA、Ligand 和 Ion 中至少需要提供一种分子输入；Modification、Covalent Bond 和 Template 只能补充已有分子信息，不能单独构成预测任务。

Protein、DNA 或 RNA FASTA 中的多条记录，以及配体和离子文件中的多个实体，都会合并到同一个待预测体系中。

Protein Sequence

蛋白质序列使用 FASTA 格式。每条记录表示同一体系中的一条独立蛋白链：

>protein1
EVQLVESGGGVVQPGRSLRLS
>protein2
NFMLTQPHSVSVL
>peptide
ELFPQWHLPIKIAAIIAS

允许的蛋白字符为 ACDEFGHIKLMNPQRSTVWYX。存在蛋白输入时，平台会自动查询并缓存蛋白 MSA。

DNA Sequence

DNA 序列使用 FASTA 格式，每条记录表示同一体系中的一条独立 DNA 链：

>dna1
ATGCGTACGT
>dna2
NNATGC

允许的 DNA 字符为 ATGCNX。

RNA Sequence

RNA 序列使用 FASTA 格式，每条记录表示同一体系中的一条独立 RNA 链：

>rna1
AUGCGUACGU

允许的 RNA 字符为 AUGCNX。当前平台不会为 RNA 查询 MSA。

FASTA 通用要求

• 每条序列前必须有以 > 开头的标题行；
• 不允许空记录，也不允许在第一个标题行前出现序列；
• 空行和以 ; 开头的注释行会被忽略；
• 序列不区分大小写，读取后统一转换为大写；
• 单个 FASTA 文件不能超过 10 MiB。

Ligand

Ligand 支持 SMILES、CCD Code 和三维配体结构文件。文本文件中每个非空、非注释行表示一个独立配体实体：

CC(=O)O
CCD,ATP
CCD_ATP_HY3

三维配体文件支持 .pdb、.sdf、.mol 和 .mol2。CCD,ATP,HY3 会被解析为一个 CCD_ATP_HY3 配体实体。配体蛋白或多肽必须通过 Protein Sequence 输入。

Ion

支持逗号或换行分隔的离子代码，可使用 :数量 指定同种离子的数量，例如 MG:2,ZN,CU:3。

Modification

每行使用 实体编号,CCD编号,1-based位置。实体按照 Protein → DNA → RNA → Ligand → Ion 的顺序编号，且修饰只能作用于 Protein、DNA 或 RNA。

Covalent Bond

每行使用 实体编号,位置编号,原子名称;实体编号,位置编号,原子名称。实体编号与 Modification 保持一致，小分子或离子通常使用位置 1。

Template

Template 仅用于蛋白质输入，支持 .a3m 或 .hhr 文件，不接受 PDB 或 CIF 文件。模板信息会应用到输入中的第一条蛋白链。

Format

用于输出蛋白质结构的格式，可选 cif 或 pdb，默认为 cif。

Seed

用于控制预测过程随机性的随机种子，最多使用前 5 个，默认为1,39,248,1970,20967。

Checkpoint

模型参数，general 或 abag。 abag为在抗原抗体复合物上微调后的版本，默认为general。

MSA

用于跳过MSA检索，默认为使用MSA。

Batch Mode

Batch Mode 用于在一个任务中批量预测多个蛋白单体或蛋白复合物。该

动态计算量

≤ 5 条序列：500 计算量 / 条
第 6–100 条序列：300 计算量 / 条
超过 100 条的部分：100 计算量 / 条

Protein Sequence

蛋白质序列文件，FASTA 格式，支持多条序列。每条记录表示一个需要预测的结构，并且每条记录名称必须唯一。如果某条记录包含多条链，则不同链之间应使用英文冒号 : 连接。示例内容：

>1
EVQLVESGGGVVQPGRSLRLS:NFMLTQPHSVSVL:ELFPQWHLPIKIAAIIAS
>2
YYCAKDRRDMGYFQHWGQGTLVTVSSQWYQQRPGSAPTTVIYEDNQRPSGTPPTFMIAVFLPIVVLIFKSILFL

这表示需要预测两个结构。第一条记录名称为 1，包含三条由英文冒号分隔的蛋白质链。第二条记录名称为 2，包含一条链。

Ligand

包含小分子信息的 TXT 格式文本文件。支持 SMILES 或 CCD Code。如果使用 SMILES 格式，每行应包含一个小分子；如果使用 CCD Code，每行可以包含一个或多个小分子，多个小分子之间用英文逗号分隔。每一行表示一个需要预测的结构，并且每一行必须以唯一名称开头。该名称可以与 RNA、DNA 或 Protein 参数中的名称相同，表示该行中的所有配体属于同一个结构。名称和所有配体之间使用英文冒号 : 分隔。示例内容：

1:CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP,HY3
lig:CC1CC(CN2NC[N+](=C3CCC(OC(F)(F)F)CC3)C2=O)CC(C)C1OC(C)(C)C(=O)[O-]

这表示两条记录。第一条名称为 1，包含三个配体，其中一个为 SMILES，两个为 CCD 代码。由于该名称存在于 RNA、DNA 或 Protein 示例记录中，因此表示它们属于同一个结构。第二条记录名称为 lig，包含一个 SMILES 格式的配体。由于该名称未出现在 RNA、DNA 或 Protein 示例记录中，因此表示一个新的结构。

注意：
1.不适用于配体蛋白或多肽的氨基酸序列格式输入。
2.在预测每个结构的序列数量相同的情况下，可为所有待预测结构设置相同的小分子信息，设置方式为输入一行小分子信息（可多个），且不设置结构名称，如CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP，表示为所有结构都加入小分子CC(=O)OC1C[NH+]2CCC1CC2与ATP。

Format

用于输出蛋白质结构的格式，可选 cif 或 pdb，默认为 cif。

Seed

用于控制预测过程随机性的随机种子，最多使用前 5 个，默认为1,39,248,1970,20967。

Checkpoint

模型参数，general 或 abag。 abag为在抗原抗体复合物上微调后的版本，默认为general。

MSA

用于跳过MSA检索，默认为使用MSA。

Enhanced Mode

Enhanced Mode 用于对一个单体或复合物进行高精度、多候选预测。该模式固定使用 500 个唯一随机 seed，不能提交自定义 Seed；其余输入均按一个预测体系组织。Enhanced Mode 会生成大量候选结构，运行时间和计算量明显高于 Single Mode，适合对候选覆盖度和预测精度要求较高的任务。

Protein、DNA、RNA、Ligand 和 Ion 中至少需要提供一种分子输入；Modification、Covalent Bond 和 Template 不能单独构成预测任务。

Protein Sequence

使用 FASTA 格式，每条记录表示同一体系中的一条独立蛋白链。允许字符为 ACDEFGHIKLMNPQRSTVWYX，平台会自动查询并缓存蛋白 MSA。

>protein1
EVQLVESGGGVVQPGRSLRLS
>protein2
NFMLTQPHSVSVL

DNA Sequence

使用 FASTA 格式，每条记录表示同一体系中的一条独立 DNA 链。允许字符为 ATGCNX。

>dna1
ATGCGTACGT

RNA Sequence

使用 FASTA 格式，每条记录表示同一体系中的一条独立 RNA 链。允许字符为 AUGCNX，当前不查询 RNA MSA。

>rna1
AUGCGUACGU

Protein、DNA 和 RNA 均须遵循以下 FASTA 要求：

• 每条序列前必须有以 > 开头的标题行；
• 不允许空记录，也不允许在第一个标题行前出现序列；
• 空行和以 ; 开头的注释行会被忽略；
• 序列不区分大小写，读取后统一转换为大写；
• 单个 FASTA 文件不能超过 10 MiB。

Ligand

支持 SMILES、CCD,ATP / CCD_ATP 形式的 CCD Code，以及 .pdb、.sdf、.mol、.mol2 三维配体文件。每个有效输入行表示一个独立配体实体，配体蛋白或多肽必须通过 Protein Sequence 输入。

Ion

支持逗号或换行分隔的离子代码，可使用 :数量 指定同种离子的数量，例如 MG:2,ZN,CU:3。

Modification

每行使用 实体编号,CCD编号,1-based位置。实体按照 Protein → DNA → RNA → Ligand → Ion 的顺序编号，且修饰只能作用于 Protein、DNA 或 RNA。

Covalent Bond

每行使用 实体编号,位置编号,原子名称;实体编号,位置编号,原子名称。实体编号与 Modification 保持一致，小分子或离子通常使用位置 1。

Template

仅在存在 Protein Sequence 时支持 .a3m 或 .hhr 模板文件；不接受 PDB 或 CIF 模板。模板信息应用于第一条蛋白链。

Format

用于输出蛋白质结构的格式，可选 cif 或 pdb，默认为 cif 。

Checkpoint

模型参数，general 或 abag。 abag为在抗原抗体复合物上微调后的版本，默认为general。

MSA

用于跳过MSA检索，默认为使用MSA。

结果说明

Single Mode

Single Mode 输出一个体系排名前 5 的结果：

• rank_1.pdb 至 rank_5.pdb：按置信度排序的结构；
• rank_1_confidence.json 至 rank_5_confidence.json：对应结构的置信度信息；
• pred_scores_opendde.csv：排名前 5 的候选分数；
• final_results.tar.gz：全部有效 seed/sample 的 CIF 结构和对应置信度 JSON；
• result_manifest.json：任务输入、模型、参数和结果文件清单。

pred_scores_opendde.csv 字段

pDockQ2 的质量区间沿用 DockQ 分级：

Batch Mode

Batch FASTA 中的每条记录都会独立排序，并在 results/<记录名称>/ 下输出该记录的 rank_1.pdb 至 rank_5.pdb 及对应置信度 JSON。

任务根目录还会输出：

• pred_scores_opendde.csv：汇总全部 Batch 记录的有效候选分数；
• final_results.tar.gz：按记录名称和 seed 组织的 CIF 结构及对应置信度 JSON；
• result_manifest.json：所有 Batch 记录及结果文件清单。

Enhanced Mode

Enhanced Mode 与 Single Mode 使用相同的排名输出结构，但候选来自500 个 seed：

• rank_1.pdb 至 rank_5.pdb 及对应置信度 JSON；
• pred_scores_opendde.csv：排名前 5 的候选分数；
• final_results.tar.gz：全部有效 seed/sample 的 CIF 结构和对应置信度 JSON；
• result_manifest.json：任务参数和结果文件清单。

参考文献

• OpenDDE Project, Aureka AI Research. 2026, Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine, arXiv, arXiv:2607.03787,DOI:10.48550/arXiv.2607.03787

Structure Prediction (OpenDDE)

Introduction

Structure Prediction (OpenDDE) is a biomolecular complex structure prediction tool based on OpenDDE (Open Drug Discovery Engine). OpenDDE is an open-source, all-atom biomolecular foundation model developed by Aureka AI Research. With multi-molecule co-folding as its core capability, it supports joint structure prediction for biomolecular complexes such as proteins, antibodies, antigens, nucleic acids, and small molecules. The module uses structure prediction as a unified structural reasoning layer in drug discovery workflows, providing a foundation for molecular design, affinity prediction, and structure optimization.

OpenDDE is mainly used to address how different biomolecules bind to each other in drug discovery, including inferring all-atom complex structures from sequences, understanding molecular interface geometry, and building a scalable structural reasoning foundation.

Parameters

Single Mode

Single Mode is used to predict one monomer or complex. At least one molecular input must be provided among Protein, DNA, RNA, Ligand, and Ion. Modification, Covalent Bond, and Template can only supplement existing molecular information and cannot independently define a prediction task.

Multiple records in Protein, DNA, or RNA FASTA files, as well as multiple entities in ligand and ion files, are merged into the same prediction system.

Protein Sequence

Protein sequence file in FASTA format. Each record represents an independent protein chain in the same system. Allowed protein characters are ACDEFGHIKLMNPQRSTVWYX. When protein input is provided, the platform automatically searches and caches protein MSA.

>protein1
EVQLVESGGGVVQPGRSLRLS
>protein2
NFMLTQPHSVSVL
>peptide
ELFPQWHLPIKIAAIIAS

DNA Sequence

DNA sequence file in FASTA format. Each record represents an independent DNA chain in the same system. Allowed DNA characters are ATGCNX.

>dna1
ATGCGTACGT
>dna2
NNATGC

RNA Sequence

RNA sequence file in FASTA format. Each record represents an independent RNA chain in the same system. Allowed RNA characters are AUGCNX. The platform currently does not search MSA for RNA.

>rna1
AUGCGUACGU

General FASTA Requirements

Protein, DNA, and RNA inputs must follow these FASTA requirements:

• Each sequence must have a header line starting with >;
• Empty records are not allowed, and sequence content is not allowed before the first header line;
• Empty lines and comment lines starting with ; are ignored;
• Sequences are case-insensitive and converted to uppercase after reading;
• A single FASTA file must not exceed 10 MiB.

Ligand

Ligand input supports SMILES, CCD Code, and 3D ligand structure files. Each non-empty, non-comment line in the text file represents an independent ligand entity.

CC(=O)O
CCD,ATP
CCD_ATP_HY3

3D ligand files support .pdb, .sdf, .mol, and .mol2. CCD,ATP,HY3 is parsed as one CCD_ATP_HY3 ligand entity. Protein ligands or peptide ligands must be provided through Protein Sequence.

Ion

Ion input supports ion codes separated by commas or line breaks. The number of ions of the same type can be specified using :count, such as MG:2,ZN,CU:3.

Modification

Modification information. Each line uses entity index,CCD code,1-based position. Entities are indexed in the order Protein → DNA → RNA → Ligand → Ion, and modifications can only be applied to Protein, DNA, or RNA.

Covalent Bond

Covalent bond information. Each line uses entity index,position index,atom name;entity index,position index,atom name. Entity indices are consistent with Modification. Small molecules or ions usually use position 1.

Template

Protein template file, only used for protein input. .a3m and .hhr files are supported, while PDB and CIF files are not accepted. Template information is applied to the first protein chain in the input.

Format

Output structure file format. The default value is cif. Available values are cif and pdb.

Seed

Random seed used to control prediction randomness. Up to the first 5 seeds are used. The default value is 1,39,248,1970,20967.

Checkpoint

Model checkpoint. The default value is general. Available values are general and abag. abag is the version fine-tuned on antigen-antibody complexes.

MSA

Whether to skip MSA search. MSA is used by default.

Batch Mode

Batch Mode is used to predict multiple protein monomers or protein complexes in one task.

Dynamic Tokens

≤ 5 sequences：500 compute units per sequence
6–100 sequences：300 compute units per sequence
Beyond 100 sequences：100 compute units per sequence

Protein Sequence

Protein sequence file in FASTA format, supporting multiple sequences. Each record represents one structure to be predicted, and each record name must be unique. If a record contains multiple chains, different chains should be connected with an English colon :

>1
EVQLVESGGGVVQPGRSLRLS:NFMLTQPHSVSVL:ELFPQWHLPIKIAAIIAS
>2
YYCAKDRRDMGYFQHWGQGTLVTVSSQWYQQRPGSAPTTVIYEDNQRPSGTPPTFMIAVFLPIVVLIFKSILFL

Ligand

Small molecule information file in TXT format. SMILES and CCD Code are supported. If SMILES format is used, each line should contain one small molecule. If CCD Code is used, each line may contain one or more small molecules, separated by English commas. Each line represents one structure to be predicted and must start with a unique name. The name and all ligands are separated by an English colon :.

1:CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP,HY3
lig:CC1CC(CN2NC[N+](=C3CCC(OC(F)(F)F)CC3)C2=O)CC(C)C1OC(C)(C)C(=O)[O-]

Notes:

1. This input is not applicable to amino acid sequence input for protein ligands or peptide ligands.
2. When the number of sequences is the same for each structure, the same small molecule information can be applied to all structures by providing one line without a structure name, such as CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP, which adds small molecule CC(=O)OC1C[NH+]2CCC1CC2 and ATP to all structures.

Format

Output structure file format. The default value is cif. Available values are cif and pdb.

Seed

Random seed used to control prediction randomness. Up to the first 5 seeds are used. The default value is 1,39,248,1970,20967.

Checkpoint

Model checkpoint. The default value is general. Available values are general and abag. abag is the version fine-tuned on antigen-antibody complexes.

MSA

Whether to skip MSA search. MSA is used by default.

Enhanced Mode

Enhanced Mode is used for high-accuracy, multi-candidate prediction of one monomer or complex. This mode always uses 500 unique random seeds and does not accept custom Seed input. All other inputs are organized as one prediction system. Enhanced Mode generates a large number of candidate structures and requires more runtime and computation than Single Mode, making it suitable for tasks that require higher candidate coverage and prediction accuracy.

At least one molecular input must be provided among Protein, DNA, RNA, Ligand, and Ion. Modification, Covalent Bond, and Template cannot independently define a prediction task.

Protein Sequence

Protein sequence file in FASTA format. Each record represents an independent protein chain in the same system. Allowed characters are ACDEFGHIKLMNPQRSTVWYX, and the platform automatically searches and caches protein MSA.

>protein1
EVQLVESGGGVVQPGRSLRLS
>protein2
NFMLTQPHSVSVL

DNA Sequence

DNA sequence file in FASTA format. Each record represents an independent DNA chain in the same system. Allowed characters are ATGCNX.

>dna1
ATGCGTACGT

RNA Sequence

RNA sequence file in FASTA format. Each record represents an independent RNA chain in the same system. Allowed characters are AUGCNX. RNA MSA is not searched currently.

>rna1
AUGCGUACGU

General FASTA Requirements

Protein, DNA, and RNA inputs must follow these FASTA requirements:

• Each sequence must have a header line starting with >;
• Empty records are not allowed, and sequence content is not allowed before the first header line;
• Empty lines and comment lines starting with ; are ignored;
• Sequences are case-insensitive and converted to uppercase after reading;
• A single FASTA file must not exceed 10 MiB.

Ligand

Ligand input supports SMILES, CCD Code in CCD,ATP / CCD_ATP format, and 3D ligand files in .pdb, .sdf, .mol, and .mol2 formats. Each valid input line represents an independent ligand entity. Protein ligands or peptide ligands must be provided through Protein Sequence.

Ion

Ion input supports ion codes separated by commas or line breaks. The number of ions of the same type can be specified using :count, such as MG:2,ZN,CU:3.

Modification

Modification information. Each line uses entity index,CCD code,1-based position. Entities are indexed in the order Protein → DNA → RNA → Ligand → Ion, and modifications can only be applied to Protein, DNA, or RNA.

Covalent Bond

Covalent bond information. Each line uses entity index,position index,atom name;entity index,position index,atom name. Entity indices are consistent with Modification. Small molecules or ions usually use position 1.

Template

Protein template file. .a3m and .hhr template files are supported only when Protein Sequence is provided. PDB and CIF templates are not accepted. Template information is applied to the first protein chain.

Format

Output structure file format. The default value is cif. Available values are cif and pdb.

Checkpoint

Model checkpoint. The default value is general. Available values are general and abag. abag is the version fine-tuned on antigen-antibody complexes.

MSA

Whether to skip MSA search. MSA is used by default.

Results

Single Mode

Single Mode outputs the top 5 ranked results for one system:

Batch Mode

Each record in the Batch FASTA is ranked independently. The corresponding rank_1.pdb to rank_5.pdb files and confidence JSON files are output under results/<record_name>/.

The task root directory also outputs:

Enhanced Mode

Enhanced Mode uses the same ranked output structure as Single Mode, but candidates are generated from 500 seeds:

The pred_scores_opendde.csv file includes the following information:

pDockQ2 quality ranges follow DockQ grading:

References

• OpenDDE Project, Aureka AI Research. 2026, Folding, Reasoning, and Scaling with Open-source Drug Discovery Engine, arXiv, arXiv:2607.03787, DOI: 10.48550/arXiv.2607.03787

Protein Design (ABACUS-T)

简介

基于结构的序列重设计或逆向折叠。基于ABACUS-T模型实现，该模型是一种多模态逆向折叠模型，以氨基酸序列空间去噪扩散执行逆折叠，整合原子侧链 / 配体相互作用、预训练蛋白语言模型、多骨架构象及 MSA 进化信息，实现更精确逆折叠，且无需固定预设残基即可生成功能活性序列。利用该模型重新设计的蛋白质展现出显著改良效果。例如，一种阿洛糖结合蛋白的亲和力提升了17倍，同时仍然保留其必要的构象变化。总体而言，ABACUS-T为在生物技术应用中高效重新设计具备特定功能的蛋白质提供了极具前景的解决方案。

在ABACUS-T每个去噪步骤中，不仅残基类型会被解码，侧链构象也会同步解码。此外，每个去噪步骤都会通过自条件机制，将前一步的输出氨基酸序列(其中包含预先嵌入的ESM蛋白语言模型信息以及侧链原子结构)作为输入的一部分。为促进生成不仅可折叠且具有功能活性的蛋白序列，ABACUS-T在以单一主链为输入条件的基础上，还可进一步使用三类可选输入，包括与该主链结合的配体分子的原子结构、该蛋白的多种主链构象状态、目标蛋白的多序列比对(MSA)。

与其他模型比较结果：

ABACUS-T在pLDDT和scRMSD指标上与LigandMPNN和CarbonDesign相当，并且在残基组成指标上与LM-design类似，但在其余对比中ABACUS-T均以显著优势优于其他方法。尤其是，ABACUS-T对口袋残基的高回收率与其较小的scRMSD相结合，可能在生成具有功能活性的序列方面相较先前方法取得关键性进步(上图B)。随后，为了评估ABACUS-T在侧链构象建模方面的准确性，使用ABACUS-T在天然主链结构上对天然序列进行侧链重构，并将结果与数种已有方法进行比较。上图C展示了各方法在正确预测χ1或χ2扭转角的侧链比例分布。ABACUS-T相对于其他方法实现了显著改善的χ1和χ2预测。

可接受多个主链结构作为输入的ABACUS-TMSD的结果与以下方案进行了比较：使用ABACUS-T的单状态设计；使用ABACUS-TMSD分别对每个构象单独进行单状态设计；使用ProteinMPNN的单状态设计；使用ProteinMPNN-MSD的多状态设计。比较基于所有残基的天然侧链类型恢复率，以及那些在不同构象间表现出较大结构变化的残基(称为动态残基)的恢复率(上图D)。虽然所有方法在动态残基上的恢复率均低于整体残基，但ABACUS-TMSD和ProteinMPNN-MSD的多状态设计均优于其对应基础模型的单状态设计。此外，ABACUS-TMSD相较其他模型表现出了显著提升，对于所有残基和动态残基，绝对恢复率提升均超过0.05。

参数说明

Single State

在单一蛋白构象状态下进行序列设计

Structure

蛋白的结构文件，PDB或CIF格式（或对应的压缩格式，如：cif.gz）。

Number of Designs

设计的序列数量，目前最大支持1000条。

Chain

指定需要设计的链名，支持多链，如A,B。注意：输入多链时，也仅按单链进行设计，并合并结果，因为当前模型不支持跨链设计。。

Type

设计残基模式：固定（Fix，指定Position参数中的残基在设计时保持不变）或者设计（Design，指定Position参数中的残基进行设计，而其他未指定残基在设计时保持不变）。

Position

设置链上的氨基酸序号，对设置的氨基酸根据Type选项进行固定或设计。格式为链名+残基编号（从1开始的位置顺序编号），支持范围设置，多段区域时用逗号分隔，如：A85-100,A114,B1-100表示：
Chain参数指定的A链上，第85至100,114号残基，B链上，第1至100号残基，进行Type指定形式的固定或设计。
链名也可省略，如：85-100,114，表示Chain参数指定的所有链上的相同位置残基，进行固定或设计。
注意：Chain与Position参数必须至少设置一个。

Sampling Temp

采样氨基酸概率分布时使用的“温度系数”，用来控制随机性/多样性。建议的取值为0.1、0.15、0.2、0.25、0.3。较高的值会导致更多的多样性。默认：0.1，数值范围为0 ~ 1.5。

Use MSA

是否采用MSA信息，引入进化约束，并结合模型残基分布与进化分布加权，修正生成残基的概率偏好，从而提升序列生成准确性与可靠性。
注意： 开启MSA，会显著增加计算时间，MSA查询和处理时间约60分钟左右。请耐心等待！

Output

输出设计的序列文件，FASTA格式，默认为design_sstate.fasta。

Multiple States

在同一蛋白链的多个构象状态下，进行序列设计。

Multiple States Structures

同一蛋白链的多个构象结构的打包压缩文件，支持的压缩格式：.zip、.gz、.tar、.tar.gz、.tgz、.tar.bz2、.tbz2、.tar.xz、.txz
注意：这些结构中应包含同一蛋白链，目前支持的最大结构数量为100

Number of Designs

设计的序列数量，目前最大支持1000条。

Reference Chain

指定其中一个结构的一条链作为参考链进行设计，格式为文件名(不含后缀名),链名，如1gub,A表示指定结构1gub中的A链作为参考链进行设计。

Type

同Single State模式。

Position

同Single State模式。

Sampling Temp

同Single State模式。

Use MSA

同Single State模式。

Output

输出设计的序列文件，FASTA格式，默认为design_mstate.fasta。

结果说明

Single State 模式

输出设计的序列文件design_sstate.fasta，文件内容如下：

>original
AAPLAATASKFLGCAYGAQQAPGFAQYWNKLTPENGGKWGSVEAVRDQMDWSTLDAAYRFAQANQMPFQMHVMVWGNQQPEWIKTLRPAEQRREIEQWFAAVAQRYPDIALLEVVNEPLNDPPSKADTGGGNYLQALGGNGDSGWEWVLQSFRLARRHFPHTKLMINDYSITSSAQATQKYLQIVRLLQRENLVDAIGVQEHAFETTPEVAVSVHRDNLDALAATGLPIYITEFDLDGPTDAQQLADYKRVFPVFWEHPAVHGITLWGFRPGLWRDKEAAYLIRADGTERPALTWLRDYVAAHPG.
>rank1; overall_identity=0.675; violation=0
QKPMAAGAAKFLGCAYSSSQAPGFENYWNAVTPENGGKWGSIEPTRDQMNWTELDNAYNFAQQHNFPFRHHVLVWGQQQPQWIHSLPPAEQRAAIEEWFKAVAARYPNIDQLEVVNEPLHEPPNSNGYGGGNYIEALGGNGETGWEWVVEAFRLARKYFPNTKLMINDYNITNSSANTDRYLEIVNQLQKENLVDVIGVQCHSFSTQPGVPVSVHRANLDRLASTGLPIMITEFDIDGPTDQIQLNEYKRVFPVYWNHPAVVGVTMWGYRPGLWRHAQNAFLINADGSERPALVWLRDYVAAHPR
>rank2; overall_identity=0.695; violation=1
QKPMAAGAAKWLGCAYGSGQAPGFETYFNAVTPENGGKWGSIEPTRNQMNWTELDAAYNFAQAHNMPFRHHVLVWGQQQPQWISSLPPAEQRKAIEEWFAAVAARYPNIDHLEVVNEPLNDPPNSNGYGGGNYIEALGGNGASGWEWVVEAFRLARKYFPNTKLMINDYNIENSSANTDRYLEIVNLLQKENLVDVIGTQCHSFCTQPGVPMSVHRANLDRLASTGLPIMITEFDIDGPTDQIQLNEYKRVFPVYWNHPAVVGVTLWGYRPGLWRHAENAYLVNADGTERPALVWLRDYVAAHPR

说明：

将 violation 作为评估设计序列结构质量的核心指标，具体包括以下几类违反几何约束或结构不合理的类型：

输出复合物序列的批量结构预测模式文件abacusT_complex_batch.fasta

输出设计后的结构打包文件，例如：result_single.tar.gz

multiple 模式

输出结果文件design_mstate.fasta，文件内容如下：

>original
AAEYAVVLKTLSNPFWVDMKKGIEDEAKTLGVSVDIFASPSEGDFQSQLQLFEDLSNKNYKGIAFAPLSSVNLVMPVARAWKKGIYLVNLDEKIDMDNLKKAGGNVEAFVTTDNVAVGAKGASFIIDKLGAEGGEVAIIEGKAGNASGEARRNGATEAFKKASQIKLVASQPADWDRIKALDVATNVLQRNPNIKAIYCANDTMAMGVAQAVANAGKTGKVLVVGTDGIPEARKMVEAGQMTATVAQNPADIGATGLKLMVDAEKSGKVIPLDKAPEFKLVDSILVTQ
>rank1; overall_identity=0.898; violations=5,0,3,0; avg_violation=2.00
AKEYAAVLKTLSNPFWVDMKKGIEAEAKKLGVSVDIFAAPSEGDLAAQLNLFEDLSNKNYKGIMLAPLSSTNLVNPVAEAWKKGIYLVNLDEKIDMDNLKKAGGNVEAFVTTDNVAVGAKGASFIIDKLGAEGGEVAIIEGKAGNASGEARRNGATEAFKKASQIKLVASQPADWDRIKALDVATNVLQRNPNIKAIYCANDTMAMGVAQAVANAGKTGKVLVVGTDGIPEARKMVEAGQMTATVAQNPADIGATGLKLMVDAEKSGKVIPLDKAPEFKLVDSILVTQ
>rank2; overall_identity=0.905; violations=5,1,3,0; avg_violation=2.25
AKEYAAVLKTLSNPFWVDMKKGIEAEAKKLGVSVDIFAAPSEGDLQSQLNLFEDLSNKNYKGIMLAPLSSTNLVNPVAKAWKKGIYLVNLDEKIDMDNLKKAGGNVEAFVTTDNVAVGAKGASFIIDKLGAEGGEVAIIEGKAGNASGEARRNGATEAFKKASQIKLVASQPADWDRIKALDVATNVLQRNPNIKAIYCANDTMAMGVAQAVANAGKTGKVLVVGTDGIPEARKMVEAGQMTATVAQNPADIGATGLKLMVDAEKSGKVIPLDKAPEFKLVDSILVTQ

说明：

注意：每个构象的violation评估与single模式保持一致

输出设计后的结构打包文件，例如：result_multiple.tar.gz

参考文献

• Liu, Y., Wu, R., Wang, X. et al. Enhancing functional proteins through multimodal inverse folding with ABACUS-T. Nat Commun 16, 10177 (2025).DOI:10.1038/s41467-025-65175-3

Protein Design (ABACUS-T)

Introduction

Structure-based sequence redesign or inverse folding. Implemented based on the ABACUS‑T model, which is a multimodal inverse folding model that performs inverse folding via denoising diffusion in amino acid sequence space. It integrates atomic side‑chain/ligand interactions, a pre‑trained protein language model, multiple backbone conformations, and MSA evolutionary information to achieve more accurate inverse folding, and can generate functionally active sequences without requiring fixed pre‑defined residues. Proteins redesigned with this model show significant improvements. For example, the affinity of an allose‑binding protein was enhanced by 17‑fold while still retaining its essential conformational changes. Overall, ABACUS‑T offers a highly promising solution for efficiently redesigning proteins with specific functions in biotechnological applications.

In each denoising step of ABACUS‑T, not only the residue type is decoded, but the side‑chain conformation is also decoded simultaneously. In addition, each denoising step incorporates, via a self‑conditioning mechanism, the amino acid sequence from the previous step (which contains pre‑embedded ESM protein language model information and side‑chain atomic structures) as part of the input. To facilitate the generation of protein sequences that are not only foldable but also functionally active, ABACUS‑T, on top of taking a single backbone as input condition, can further use three optional inputs: the atomic structure of ligand molecules bound to that backbone, multiple backbone conformational states of the protein, and multiple sequence alignments (MSA) of the target protein.

Comparison with other models:

ABACUS‑T is comparable to LigandMPNN and CarbonDesign in terms of pLDDT and scRMSD, and similar to LM‑design in residue composition, but outperforms all other methods with significant margins in all other comparisons. Notably, the high recovery of pocket residues combined with low scRMSD may represent a critical advance over previous methods for generating functionally active sequences (Figure B above). Subsequently, to evaluate the accuracy of ABACUS‑T in side‑chain conformation modelling, we used ABACUS‑T to reconstruct side chains of native sequences on native backbone structures and compared the results with several existing methods. Figure C above shows the fraction of side chains for which each method correctly predicts the χ1 or χ2 torsion angles. ABACUS‑T achieves significantly improved χ1 and χ2 predictions over other methods.

The results of ABACUS‑T^MSD, which can accept multiple backbone structures as input, were compared with the following schemes: single‑state design using ABACUS‑T; single‑state design performed separately on each conformation using ABACUS‑T^MSD; single‑state design using ProteinMPNN; and multi‑state design using ProteinMPNN‑MSD. The comparison was based on the recovery rate of native side‑chain types for all residues, as well as for residues that exhibit large structural changes between conformations (called dynamic residues) (Figure D above). Although all methods show lower recovery rates on dynamic residues than on overall residues, the multi‑state designs of ABACUS‑T^MSD and ProteinMPNN‑MSD both outperform the single‑state designs of their respective base models. Furthermore, ABACUS‑T^MSD shows a significant improvement over other models, with absolute recovery gains exceeding 0.05 for both all residues and dynamic residues.

Parameters

Single State

Sequence design under a single protein conformational state.

Structure

The protein structure file, in PDB or CIF format (or their compressed equivalents, e.g., cif.gz).

Number of Designs

Number of designed sequences; currently supports up to 1000.

Chain

Specify the chain name(s) to be designed. Supports multiple chains, e.g., A,B. Note: When multiple chains are input, the design is still performed as single‑chain designs and results are merged, because the current model does not support cross‑chain design.

Type

Residue design mode: Fix (specified residues in the Position parameter remain unchanged during design) or Design (specified residues in the Position parameter are designed, while other unspecified residues remain unchanged during design).

Position

Set the amino acid sequence numbers on the chain. The residues set here are either fixed or designed according to the Type option. The format is chain name + residue number (sequential position starting from 1). Supports range specification; multiple regions separated by commas, e.g., A85‑100,A114,B1‑100 means:

• For the chain specified by the Chain parameter: residues 85‑100 and 114 on chain A, and residues 1‑100 on chain B, are fixed or designed in the form specified by Type.

The chain name can be omitted, e.g., 85‑100,114, which means the same positions on all chains specified by the Chain parameter are fixed or designed.

Note: At least one of Chain and Position must be set.

Sampling Temp

The “temperature coefficient” used when sampling amino acid probability distributions, controlling randomness/diversity. Recommended values are 0.1, 0.15, 0.2, 0.25, 0.3. Higher values lead to more diversity. Default: 0.1. Value range: 0 ~ 1.5.

Use MSA

Whether to use MSA information, introducing evolutionary constraints, and combining model residue distributions with evolutionary distributions to weight and correct the probability preference of generated residues, thereby improving the accuracy and reliability of sequence generation.

Note: Enabling MSA will significantly increase computation time. MSA retrieval and processing takes about 60 minutes. Please be patient!

Output

Output the designed sequences in FASTA format. Default: design_sstate.fasta.

Multiple States

Sequence design under multiple conformational states of the same protein chain.

Multiple States Structures

A compressed archive containing multiple conformational structures of the same protein chain. Supported compression formats: .zip, .gz, .tar, .tar.gz, .tgz, .tar.bz2, .tbz2, .tar.xz, .txz.

Note: These structures should contain the same protein chain. Currently supports a maximum of 100 structures.

Number of Designs

Number of designed sequences; currently supports up to 1000.

Reference Chain

Specify one chain from one of the structures as the reference chain for design. Format: filename (without extension), chain name, e.g., 1gub,A means using chain A from structure 1gub as the reference chain for design.

Type

Same as in Single State mode.

Position

Same as in Single State mode.

Sampling Temp

Same as in Single State mode.

Use MSA

Same as in Single State mode.

Output

Output the designed sequences in FASTA format. Default: design_mstate.fasta.

Results

Single State Mode

The designed sequence file design_sstate.fasta is generated. Example content:

>original
AAPLAATASKFLGCAYGAQQAPGFAQYWNKLTPENGGKWGSVEAVRDQMDWSTLDAAYRFAQANQMPFQMHVMVWGNQQPEWIKTLRPAEQRREIEQWFAAVAQRYPDIALLEVVNEPLNDPPSKADTGGGNYLQALGGNGDSGWEWVLQSFRLARRHFPHTKLMINDYSITSSAQATQKYLQIVRLLQRENLVDAIGVQEHAFETTPEVAVSVHRDNLDALAATGLPIYITEFDLDGPTDAQQLADYKRVFPVFWEHPAVHGITLWGFRPGLWRDKEAAYLIRADGTERPALTWLRDYVAAHPG.
>rank1; overall_identity=0.675; violation=0
QKPMAAGAAKFLGCAYSSSQAPGFENYWNAVTPENGGKWGSIEPTRDQMNWTELDNAYNFAQQHNFPFRHHVLVWGQQQPQWIHSLPPAEQRAAIEEWFKAVAARYPNIDQLEVVNEPLHEPPNSNGYGGGNYIEALGGNGETGWEWVVEAFRLARKYFPNTKLMINDYNITNSSANTDRYLEIVNQLQKENLVDVIGVQCHSFSTQPGVPVSVHRANLDRLASTGLPIMITEFDIDGPTDQIQLNEYKRVFPVYWNHPAVVGVTMWGYRPGLWRHAQNAFLINADGSERPALVWLRDYVAAHPR
>rank2; overall_identity=0.695; violation=1
QKPMAAGAAKWLGCAYGSGQAPGFETYFNAVTPENGGKWGSIEPTRNQMNWTELDAAYNFAQAHNMPFRHHVLVWGQQQPQWISSLPPAEQRKAIEEWFAAVAARYPNIDHLEVVNEPLNDPPNSNGYGGGNYIEALGGNGASGWEWVVEAFRLARKYFPNTKLMINDYNIENSSANTDRYLEIVNLLQKENLVDVIGTQCHSFCTQPGVPMSVHRANLDRLASTGLPIMITEFDIDGPTDQIQLNEYKRVFPVYWNHPAVVGVTLWGYRPGLWRHAENAYLVNADGTERPALVWLRDYVAAHPR

The design_sstate.fasta file includes the following information:

violation is the core metric for evaluating designed sequence structure quality. It includes the following violation types:

The batch structure prediction mode file for complex sequences, abacusT_complex_batch.fasta, is also generated.

A packaged structure file after design is generated, for example result_single.tar.gz.

Multiple Mode

The result file design_mstate.fasta is generated. Example content:

>original
AAEYAVVLKTLSNPFWVDMKKGIEDEAKTLGVSVDIFASPSEGDFQSQLQLFEDLSNKNYKGIAFAPLSSVNLVMPVARAWKKGIYLVNLDEKIDMDNLKKAGGNVEAFVTTDNVAVGAKGASFIIDKLGAEGGEVAIIEGKAGNASGEARRNGATEAFKKASQIKLVASQPADWDRIKALDVATNVLQRNPNIKAIYCANDTMAMGVAQAVANAGKTGKVLVVGTDGIPEARKMVEAGQMTATVAQNPADIGATGLKLMVDAEKSGKVIPLDKAPEFKLVDSILVTQ
>rank1; overall_identity=0.898; violations=5,0,3,0; avg_violation=2.00
AKEYAAVLKTLSNPFWVDMKKGIEAEAKKLGVSVDIFAAPSEGDLAAQLNLFEDLSNKNYKGIMLAPLSSTNLVNPVAEAWKKGIYLVNLDEKIDMDNLKKAGGNVEAFVTTDNVAVGAKGASFIIDKLGAEGGEVAIIEGKAGNASGEARRNGATEAFKKASQIKLVASQPADWDRIKALDVATNVLQRNPNIKAIYCANDTMAMGVAQAVANAGKTGKVLVVGTDGIPEARKMVEAGQMTATVAQNPADIGATGLKLMVDAEKSGKVIPLDKAPEFKLVDSILVTQ
>rank2; overall_identity=0.905; violations=5,1,3,0; avg_violation=2.25
AKEYAAVLKTLSNPFWVDMKKGIEAEAKKLGVSVDIFAAPSEGDLQSQLNLFEDLSNKNYKGIMLAPLSSTNLVNPVAKAWKKGIYLVNLDEKIDMDNLKKAGGNVEAFVTTDNVAVGAKGASFIIDKLGAEGGEVAIIEGKAGNASGEARRNGATEAFKKASQIKLVASQPADWDRIKALDVATNVLQRNPNIKAIYCANDTMAMGVAQAVANAGKTGKVLVVGTDGIPEARKMVEAGQMTATVAQNPADIGATGLKLMVDAEKSGKVIPLDKAPEFKLVDSILVTQ

The design_mstate.fasta file includes the following information:

The violation evaluation for each conformational state is consistent with Single State mode.

A packaged structure file after design is generated, for example result_multiple.tar.gz.

References

• Liu Y., Wu R., Wang X., et al. 2025, Enhancing functional proteins through multimodal inverse folding with ABACUS-T, Nature Communications, 16:10177, DOI: 10.1038/s41467-025-65175-3

Protein Stability (ESM3 & SaProt)

简介

Protein Stability (ESM3 & SaProt) 用于预测蛋白热力学稳定性，集成经微调的 ESM3ΔG 和 SaProtΔG 模型。模块可基于 PDB/CIF 结构文件，对指定蛋白链生成残基级稳定性预测，并输出链级或复合物级平均 ΔG；同时支持单点突变扫描，预测突变引起的稳定性变化（ΔΔG），用于辅助筛选更稳定的候选蛋白或突变位点。

参数说明

Absolute Stability (Single Chain)

计算结构中每条蛋白单链的绝对稳定性 dG。

Structure

需计算单链绝对稳定性的蛋白结构文件，支持 PDB/CIF 格式；同时支持批量结构输入，可将批量结构文件压缩打包。支持格式包括 .zip、.tar、.tar.gz、.tgz、.tar.bz2、.tbz2、.tar.xz、.txz。

Absolute Stability (Complex)

计算结构中复合物（指定链）的绝对稳定性 dG，同时计算相应单链的 dG。

Complex

蛋白复合物结构文件，支持 PDB/CIF 格式；同时支持批量结构输入，可将批量结构文件压缩打包。支持格式包括 .zip、.tar、.tar.gz、.tgz、.tar.bz2、.tbz2、.tar.xz、.txz。

Target Chain

指定复合物中作为受体的链，支持多链，使用逗号分隔，如 A,B。

Binder Chain

指定复合物中作为配体的链，支持多链，使用逗号分隔，如 A,B。未指定时，默认选取结构中除 Target Chain 外的所有链作为配体。

Mutation Energy of Stability

计算结构中指定链饱和单点突变对应的稳定性变化 ddG。

Structure

蛋白结构文件，支持 PDB/CIF 格式；同时支持批量结构输入，可将批量结构文件压缩打包。支持格式包括 .zip、.tar、.tar.gz、.tgz、.tar.bz2、.tbz2、.tar.xz、.txz。

Chain

指定需进行单点饱和突变分析的蛋白链，支持多链，使用逗号分隔，如 A,B。未指定时使用全部链。

结果说明

Absolute Stability 预测输出 absolute_stability.csv，包含信息如下：

Absolute Stability 预测还会输出逐残基明细表 absolute_stability_details.csv，包含信息如下：

Mutation Energy of Stability 预测输出 mutation_stability.csv，包含信息如下：

参考文献

• Cho Y., Tsuboyama K., Litberg T. J., et al. 2026, Accurate protein stability prediction for small domains using mega-scale experiments, bioRxiv, Preprint, DOI: 10.64898/2026.05.19.726285

Protein Stability (ESM3 & SaProt)

Introduction

Protein Stability (ESM3 & SaProt) predicts protein thermodynamic stability using fine-tuned ESM3ΔG and SaProtΔG models. Based on PDB/CIF structure files, the module generates residue-level stability predictions for specified protein chains and outputs chain-level or complex-level average ΔG. It also supports single-point mutation scanning to predict stability changes caused by mutations (ΔΔG), helping screen more stable candidate proteins or mutation sites.

Parameters

Absolute Stability (Single Chain)

Calculate the absolute stability dG of each protein single chain in the structure.

Structure

Protein structure file for calculating single-chain absolute stability, supporting PDB/CIF formats. Batch structure input is also supported by compressing multiple structure files into an archive. Supported formats include .zip, .tar, .tar.gz, .tgz, .tar.bz2, .tbz2, .tar.xz, and .txz.

Absolute Stability (Complex)

Calculate the absolute stability dG of the complex specified by chains in the structure, and also calculate the corresponding single-chain dG values.

Complex

Protein complex structure file, supporting PDB/CIF formats. Batch structure input is also supported by compressing multiple structure files into an archive. Supported formats include .zip, .tar, .tar.gz, .tgz, .tar.bz2, .tbz2, .tar.xz, and .txz.

Target Chain

Specify the chains used as the receptor in the complex. Multiple chains are supported and separated by commas, such as A,B.

Binder Chain

Specify the chains used as the ligand in the complex. Multiple chains are supported and separated by commas, such as A,B. If not specified, all chains except Target Chain in the structure are used as the ligand by default.

Mutation Energy of Stability

Calculate the stability change ddG for saturated single-point mutations on specified chains in the structure.

Structure

Protein structure file, supporting PDB/CIF formats. Batch structure input is also supported by compressing multiple structure files into an archive. Supported formats include .zip, .tar, .tar.gz, .tgz, .tar.bz2, .tbz2, .tar.xz, and .txz.

Chain

Specify the protein chains for saturated single-point mutation analysis. Multiple chains are supported and separated by commas, such as A,B. If not specified, all chains are used.

Results

Absolute Stability prediction outputs absolute_stability.csv, which includes the following information:

Absolute Stability prediction also outputs the residue-level detail table absolute_stability_details.csv, which includes the following information:

Mutation Energy of Stability prediction outputs mutation_stability.csv, which includes the following information:

References

• Cho Y., Tsuboyama K., Litberg T. J., et al. 2026, Accurate protein stability prediction for small domains using mega-scale experiments, bioRxiv, Preprint, DOI: 10.64898/2026.05.19.726285

MD (GMX2024) v2

简介

基于 GROMACS 的分子动力学模拟自动化脚本，覆盖能量最小化、NVT/NPT 温压平衡、成品模拟全流程，并集成 PLUMED Metadynamics 增强采样与距离/角度/二面角限制性模拟支持。

适用场景：

1. 非膜体系或膜体系的多阶段平衡与生产模拟（mini → nvt → npt → md）
2. 基于已有模拟结果的续跑（延长模拟或更换 MDP 参数续跑）
3. 距离/角度/二面角限制性模拟
4. 配合 PLUMED 进行构象空间增强采样

参数说明

System Topology

体系的拓扑输入文件，.top格式。全新模拟必选。

System Structure

体系的结构输入文件，.gro格式。全新模拟必选。

System ITP

体系的参数输入文件包，支持 .zip、.tgz 和 .tar.gz格式。全新模拟必选。

Index File

体系的索引输入文件，.ndx格式。全新模拟必选。

MDP File

用于分子动力学模拟的 MDP 参数输入文件，支持单个 .mdp文件或压缩包（.zip/.tgz/.tar.gz）。全新模拟和续跑方式二（新 MDP 续跑）必填。

MDP Sequence

MDP 输入文件的执行顺序，逗号分隔。文件名可不带 .mdp 后缀（自动补全）。不指定时按阶段前缀（mini/min/em → nvt → npt → md）自动排序，大小写不敏感，支持 _、-、. 分隔符。

CPT File

Checkpoint 文件（.cpt），仅全新模拟且首个 MDP 文件不是能量最小化时生效，为第一个动力学阶段提供初始速度和初始结构。续跑模式无效。

Path File

历史模拟路径文件，用于续跑。指定后进入续跑模式。如指定 Time 参数则进入续跑方式一（在历史模拟的参数基础上继续延长指定时长）；如指定 MDP File参数则进入续跑方式二（在历史模拟的结果基础上更换 MDP 参数进行续跑）。

Time

续跑延长时间（单位 ns），仅续跑方式一（延长模拟）使用。脚本从上次日志读取积分步长 dt 并计算新增步数。

Restraint File

限制性模拟时输入的限制参数文件，定义距离/角度/二面角限制。原子可用全局号、残基号,原子名 或 链名,残基号,原子名 指定。

Restraint MDP

启用限制的 MDP 阶段，逗号分隔。文件名可不带 .mdp 后缀（自动补全）。当分子动力学模拟运行到指定的 MDP 阶段时，会读入Restraint File内容进行限制性模拟。minimization阶段不支持进行限制，如命中会告警并跳过。

Metad Dat File

PLUMED 配置文件（.dat），用于 Metadynamics 增强采样。

Metad Ndx File

Metadynamics 增强采样的索引文件（.ndx）。

结果说明

输出结果包括：

能量分析覆盖所有非优化阶段（nvt/npt/md），各段能量数据按时间顺序合并为连续时间轴。

MD (GMX2024) v2

Introduction

A GROMACS-based molecular dynamics automation script covering the full workflow of energy minimization, NVT/NPT equilibration, and production simulation, with integrated PLUMED Metadynamics enhanced sampling and distance/angle/dihedral restraint support.

Applicable scenarios:

1. Multi-stage equilibration and production simulation for non-membrane or membrane systems (mini → nvt → npt → md)
2. Restart from existing simulation results (extend simulation or restart with new MDP parameters)
3. Distance/angle/dihedral restraint simulations
4. Enhanced conformational sampling with PLUMED

Parameters

System Topology

System topology input file in .top format. Required for fresh simulations.

System Structure

System structure input file in .gro format. Required for fresh simulations.

System ITP

System parameter input archive, supporting .zip, .tgz, and .tar.gz formats. Required for fresh simulations.

Index File

System index input file in .ndx format. Required for fresh simulations.

MDP File

MDP parameter input file for molecular dynamics simulation, accepting a single .mdp file or an archive (.zip/.tgz/.tar.gz). Required for fresh simulations and restart mode 2 (new MDP restart).

MDP Sequence

Execution order of MDP input files, comma-separated. File names may omit the .mdp extension (auto-completed). When not specified, files are auto-sorted by stage prefix (mini/min/em → nvt → npt → md), case-insensitive, supporting _, -, . separators.

CPT File

Checkpoint file (.cpt), effective only in fresh simulations where the first MDP file is not energy minimization, providing initial velocities and initial structure for the first dynamics stage. Invalid in restart mode.

Path File

History simulation path file for restart. Enters restart mode when specified. If the Time parameter is specified, enters restart mode 1 (extends the simulation for a specified duration based on the historical simulation’s parameters); if the MDP File parameter is specified, enters restart mode 2 (restarts with new MDP parameters based on the historical simulation’s results).

Time

Extended simulation time (in ns), used only in restart mode 1 (extend simulation). The script reads the integration step dt from the previous log and computes the new number of steps.

Restraint File

Restraint parameter input file for restraint simulations, defining distance/angle/dihedral restraints. Atoms may be specified by global number, resnum,atomname, or chain,resnum,atomname.

Restraint MDP

MDP stages that enable restraints, comma-separated. File names may omit the .mdp extension (auto-completed). When the molecular dynamics simulation reaches a specified MDP stage, the Restraint File content is read to perform restraint simulation. Minimization stages do not support restraints; if hit, a warning is issued and the stage is skipped.

Metad Dat File

PLUMED configuration file (.dat) for Metadynamics enhanced sampling.

Metad Ndx File

Index file (.ndx) for Metadynamics enhanced sampling.

Results

The output includes:

The energy analysis covers all non-minimization stages (nvt/npt/md); the energy data of each segment is merged into a continuous timeline in chronological order.

GMX MDP Generation

简介

基于模板自动生成 GROMACS MDP 文件的工具，根据系统组分和模拟类型选择对应模板，并通过参数化配置生成能量最小化（MINI）、NVT/NPT 平衡、分子动力学（MD）、拉伸分子动力学（SMD）和退火 MD （Anneal_MD）所需的 MDP 文件。

适用场景：

1. 标准分子动力学模拟流程（mini → nvt → npt → md）
2. 膜蛋白或含配体体系的模拟参数配置
3. 拉伸分子动力学（SMD）和退火模拟的 MDP 生成
4. 使用自定义 MDP 文件或压缩包快速部署

参数说明

Components

系统组分，可多选，可选项：membrane、receptor、ligand。solvent 始终隐含，无需指定。

Types

生成的 MDP 文件类型，可多选，可选项：mini、nvt、npt、md、smd、anneal_md。smd 和 anneal_md 不可同时选择。脚本根据优先级（smd > anneal_md > md > npt > nvt > mini）确定 target 类型（优先级最高的为target类型），target 文件接收所有参数修改。

User MDP

用户提供自定义 MDP 文件或压缩包，指定后所有其他参数设置均忽略。支持 .mdp、.tar.gz、.tgz、.zip 格式。

Duration

模拟时长，单位 ns。默认值：md/anneal_md 为 100 ns，smd 为 10 ns。仅 target 文件生效。

Time Step

时间步长，单位 ps，默认 0.002。仅 target 文件生效。

NSTXOUT Compressed

压缩坐标写入频率，默认 50000（步）。仅 target 文件生效。

Coulomb Type

静电相互作用方法，可选项：Cut-off、Ewald、PME，默认 PME。仅 target 文件生效。

VDW Type

范德华相互作用方法，可选项：Cut-off、PME，默认 Cut-off。仅 target 文件生效。

Temperature Coupling

温度耦合方法，可选项：Berendsen、V-rescale，默认 V-rescale。仅 target 文件生效。

Reference Temperature

参考温度，单位 K，支持单个值（广播到所有 tc-grps）或多个值（数量须等于 tc-grps 数）。默认膜体系 310 K，非膜体系 300 K。对所有含温度耦合的文件生效。

Temperature Groups

温度耦合组名，默认根据 --components 自动生成。自定义时数量须与实际组数一致。

Pressure Coupling

压力耦合方法，可选项：Berendsen、Parrinello-Rahman、C-rescale，默认 C-rescale。仅 target 文件生效。

Pressure Coupling Type

压力耦合类型，可选项：isotropic、semiisotropic、anisotropic、surface-tension。默认膜体系 semiisotropic，非膜体系 isotropic。仅 target 文件生效。

Reference Pressure

参考压力，单位 bar，支持单个值（广播）或多个值。数量须与 pcoupltype 对应（isotropic=1, semiisotropic=2, anisotropic=6, surface-tension=2）。仅 target 文件生效。

Annealing

退火类型，每个 tc-grps 一个值，可选项：single、periodic、none。none 表示禁用 nvt 中的退火参数。anneal_md 类型必填，四个退火参数须同时指定。

Annealing Npoints

每个 tc-grps 的退火点数，数量须等于 tc-grps 数。

Annealing Time

退火时间点，总数须等于所有 annealing-npoints 之和。

Annealing Temp

退火温度点，总数须等于所有 annealing-npoints 之和。

Pull Group Name

SMD 拉伸组定义，必填。分号分隔不同 coord，冒号分隔同一 coord 内的组。组名支持名称（如Protein）、链名（如A）、残基范围（如A35-100）。

Pull Coord Geometry

每个 coord 的几何类型，分号分隔，数量须等于 coord 数。可选项及所需组数：

• distance：2 组
• direction：2 组
• direction-periodic：2 组
• angle：4 组
• dihedral：6 组

Pull Coord Type

每个 coord 的拉伸类型，分号分隔，默认 umbrella。支持单个值广播。可选项：umbrella、constraint、constant-force。

Pull Coord Rate

每个 coord 的拉伸速率，分号分隔，默认 0.005。constant-force 类型不消耗值。

Pull Coord K

每个 coord 的力常数，分号分隔，默认 1000。

Pull Coord Dim

每个 coord 的维度/方向向量，分号分隔，默认 Y Y Y。仅对 distance、direction、direction-periodic 有效。

Custom

自定义 MDP 字段覆盖，格式为 key=val;key=val，在所有其他参数之后执行，优先级最高。

Output

输出文件名，默认 mdp.tar.gz。

结果说明

输出结果包括：

GMX MDP Generation

Introduction

GROMACS MDP Generator is a template-based tool for automatically generating GROMACS MDP files. The module selects appropriate templates according to system components and simulation types, then generates parameterized MDP files for energy minimization (MINI), NVT/NPT equilibration, molecular dynamics (MD), steered molecular dynamics (SMD), and annealing molecular dynamics (Anneal MD).

Applicable scenarios:

1. Standard molecular dynamics workflow: mini → nvt → npt → md
2. Simulation parameter setup for membrane protein or ligand-containing systems
3. MDP generation for steered molecular dynamics (SMD) and annealing simulations
4. Rapid deployment using custom MDP files or compressed archives

Parameters

Components

System components. Multiple options can be selected. Available values: membrane, receptor, ligand. solvent is always implied and does not need to be specified.

Types

Types of MDP files to generate. Multiple options can be selected. Available values: mini, nvt, npt, md, smd, anneal_md. smd and anneal_md cannot be selected at the same time. The module determines the target type by priority: smd > anneal_md > md > npt > nvt > mini. The highest-priority type is used as the target file and receives all parameter modifications.

User MDP

Custom MDP file or compressed archive provided by the user. When specified, all other parameter settings are ignored. Supported formats: .mdp, .tar.gz, .tgz, .zip.

Duration

Simulation duration, in ns. Default value: 100 ns for md/anneal_md, and 10 ns for smd. Only applies to the target file.

Time Step

Time step, in ps. Default value: 0.002. Only applies to the target file.

NSTXOUT Compressed

Compressed coordinate output frequency, in steps. Default value: 50000. Only applies to the target file.

Coulomb Type

Electrostatic interaction method. Default value: PME. Available values: Cut-off, Ewald, PME. Only applies to the target file.

VDW Type

Van der Waals interaction method. Default value: Cut-off. Available values: Cut-off, PME. Only applies to the target file.

Temperature Coupling

Temperature coupling method. Default value: V-rescale. Available values: Berendsen, V-rescale. Only applies to the target file.

Reference Temperature

Reference temperature, in K. A single value can be broadcast to all temperature coupling groups, or multiple values can be provided with the same count as the temperature coupling groups. The default is 310 K for membrane systems and 300 K for non-membrane systems. Applies to all files containing temperature coupling.

Temperature Groups

Temperature coupling group names. By default, groups are generated automatically according to Components. When customized, the number of groups must match the actual group count.

Pressure Coupling

Pressure coupling method. Default value: C-rescale. Available values: Berendsen, Parrinello-Rahman, C-rescale. Only applies to the target file.

Pressure Coupling Type

Pressure coupling type. The default is semiisotropic for membrane systems and isotropic for non-membrane systems. Available values: isotropic, semiisotropic, anisotropic, surface-tension. Only applies to the target file.

Reference Pressure

Reference pressure, in bar. A single value can be broadcast, or multiple values can be provided. The number of values must match the pressure coupling type: isotropic=1, semiisotropic=2, anisotropic=6, surface-tension=2. Only applies to the target file.

Annealing

Annealing type, with one value for each temperature coupling group. Available values: single, periodic, none. none disables annealing parameters in NVT. When anneal_md is selected, this parameter is required, and Annealing, Annealing Npoints, Annealing Time, and Annealing Temp must be specified together.

Annealing Npoints

Number of annealing points for each temperature coupling group. The number of values must equal the number of temperature coupling groups.

Annealing Time

Annealing time points. The total number of values must equal the sum of all Annealing Npoints.

Annealing Temp

Annealing temperature points. The total number of values must equal the sum of all Annealing Npoints.

Pull Group Name

SMD pull group definition. Required for SMD mode. Different coords are separated by semicolons, and groups within the same coord are separated by colons. Group names can be standard group names such as Protein, chain IDs such as A, or residue ranges such as A35-100.

Pull Coord Geometry

Geometry type for each coord, separated by semicolons. The number of values must equal the number of coords. Available values and required group counts are listed below:

Pull Coord Type

Pull type for each coord, separated by semicolons. Default value: umbrella. A single value can be broadcast. Available values: umbrella, constraint, constant-force.

Pull Coord Rate

Pull rate for each coord, separated by semicolons. Default value: 0.005. constant-force type does not consume this value.

Pull Coord K

Force constant for each coord, separated by semicolons. Default value: 1000.

Pull Coord Dim

Dimension or direction vector for each coord, separated by semicolons. Default value: Y Y Y. Only valid for distance, direction, and direction-periodic.

Custom

Custom MDP field overrides, in the format key=val;key=val. This parameter is applied after all other parameters and has the highest priority.

Output

Output file name. Default value: mdp.tar.gz.

Results

The output files include:

GMX Solvation

简介

统一的 GROMACS 体系溶剂化与离子化工具。根据输入自动判定工作模式，完成结构拼接、拓扑合并、建盒子、加水、加离子与索引生成，输出可直接用于分子动力学模拟的体系文件。

适用场景：

• 非膜体系：对受体（蛋白/核酸）单独或与配体一起溶剂化，按需加离子。
• 膜体系：处理已预先溶剂化的膜，可嵌入受体与配体，按需加离子并做电荷中和。

参数说明

Membrane Topology

膜拓扑输入文件，top 格式，膜体系必需。与 Membrane Structure、Membrane ITP 三者必须同时提供。

Membrane Structure

膜结构输入文件，gro 格式，膜体系必需，应为已预先溶剂化的膜结构。与 Membrane Topology、Membrane ITP 三者必须同时提供。

Membrane ITP

膜参数输入文件包，tar.gz 格式，膜体系必需。与 Membrane Topology、Membrane Structure 三者必须同时提供。

Receptor Topology

受体拓扑输入文件，top 格式。蛋白/核酸体系必须。提供时须与 Receptor Structure、Receptor Itp 三者必须同时提供。

Receptor Structure

受体结构输入文件，gro 格式。蛋白/核酸体系必须。提供时须与 Receptor Topology、Receptor Itp 三者必须同时提供。

Receptor ITP

受体参数输入文件包，tar.gz 格式。蛋白/核酸体系必须。提供时须与 Receptor Topology、Receptor Structure 三者必须同时提供。

Ligand Topology

配体拓扑输入文件，top 格式。提供时须与 Ligand Structure、Ligand Itp 三者必须同时提供。

Ligand Structure

配体结构输入文件，gro 格式。提供时须与 Ligand Topology、Ligand Itp 三者必须同时提供。

Ligand ITP

配体参数输入文件包，tar.gz 格式。提供时须与 Ligand Topology、Ligand Structure 三者必须同时提供。

All Component Parameters

包含体系中所有组分参数化结果的压缩输入文件，可包含蛋白、核酸、膜、配体等组分的参数文件，用于后续 GMX 体系构建与模拟。

Ions

添加的离子类型，不同离子类型用 : ，如 NA:CL。支持 NA、K、CL、CA、MG、ZN。默认不添加离子，体系会自动添加若干离子进行电中性处理。

Ions Number

各离子的添加数量，不同离子的数量用 : 分隔，如50:60。与 Ions参数顺序需一一对应，与 Ions Concentration 二选一。

Ions Concentration

各离子的添加浓度，单位 mol/L，用 : 分隔，如0.15:0.2。与 Ions参数顺序需一一对应，与 Ions Number 二选一。仅提供 --ions 而Ions Number 和 Ions Concentration均未给时默认 0.15 mol/L。

Box Type

盒子类型，默认 cubic。仅非膜模式有效；给体系添加溶剂盒子时作为 gmx editconf -bt 的盒子类型。膜模式忽略。

• cubic：立方体盒
• triclinic：一般三斜盒
• dodecahedron：近似球形、体积更小。通常用于蛋白或小分子体系，因为它能在保证同样最小距离的前提下，减少约 30% 的水分子数，节约计算量。
• octahedron：八面体盒

Distance

盒子边距，单位 nm，默认 1.0。仅非膜模式有效。给体系添加溶剂盒子时溶质到盒子边界的最短距离。膜模式忽略（膜已预溶剂化，自带溶剂盒子）。

Add Water

是否添加溶剂（--add-water / --no-add-water），默认开启（即添加溶剂）。--no-add-water 时不加新溶剂（输入中已有的水保留）。仅非膜模式有效。膜模式忽略（膜已预溶剂化，自带溶剂盒子）。
备注：是否真正加水还取决于拓扑是否 #include 了水模型。

结果说明

输出结果文件包括：

GMX Solvation

Introduction

A unified GROMACS system solvation and ionization tool. It automatically detects the working mode from the inputs and performs structure assembly, topology merging, box building, solvation, ion addition, and index generation, producing system files ready for molecular dynamics simulation.

Applicable scenarios:

• Non-membrane systems: solvate a receptor (protein/nucleic acid) alone or together with a ligand, with ions added as needed.
• Membrane systems: process a pre-solvated membrane, with optional embedded receptor and ligand, adding ions as needed and charge-neutralizing.

Parameters

Membrane Topology

Membrane topology input file, top format, required for membrane systems. Must be provided together with Membrane Structure and Membrane ITP (all three required together).

Membrane Structure

Membrane structure input file, gro format, required for membrane systems; should be a pre-solvated membrane structure. Must be provided together with Membrane Topology and Membrane ITP (all three required together).

Membrane ITP

Membrane parameter input file package, tar.gz format, required for membrane systems. Must be provided together with Membrane Topology and Membrane Structure (all three required together).

Receptor Topology

Receptor topology input file, top format, required for protein/nucleic acid systems. Must be provided together with Receptor Structure and Receptor ITP (all three required together).

Receptor Structure

Receptor structure input file, gro format, required for protein/nucleic acid systems. Must be provided together with Receptor Topology and Receptor ITP (all three required together).

Receptor ITP

Receptor parameter input file package, tar.gz format, required for protein/nucleic acid systems. Must be provided together with Receptor Topology and Receptor Structure (all three required together).

Ligand Topology

Ligand topology input file, top format. Must be provided together with Ligand Structure and Ligand ITP (all three required together).

Ligand Structure

Ligand structure input file, gro format. Must be provided together with Ligand Topology and Ligand ITP (all three required together).

Ligand ITP

Ligand parameter input file package, tar.gz format. Must be provided together with Ligand Topology and Ligand Structure (all three required together).

All Component Parameters

Compressed input file containing the parameterization results for all components in the system, including proteins, nucleic acids, membranes, and ligands, for subsequent GMX system setup and simulation.

Ions

Ion types to add, with different ion types separated by :, e.g. NA:CL. Supports NA, K, CL, CA, MG, ZN. By default no ions are added; the system automatically adds some ions for charge neutralization.

Ions Number

Number of each ion to add, with the counts of different ions separated by :, e.g. 50:60. Must correspond one-to-one with the order of the Ions parameter; mutually exclusive with Ions Concentration.

Ions Concentration

Concentration of each ion to add, in mol/L, separated by :, e.g. 0.15:0.2. Must correspond one-to-one with the order of the Ions parameter; mutually exclusive with Ions Number. When only --ions is given and neither Ions Number nor Ions Concentration is provided, defaults to 0.15 mol/L.

Box Type

Box type for gmx editconf -bt. Defaults to cubic. Used only in non-membrane mode when creating the solvent box; ignored in membrane mode.

• cubic: cubic box
• triclinic: general triclinic box
• dodecahedron: nearly spherical and smaller in volume. Commonly used for proteins or small molecules because it can reduce the number of water molecules by about 30% while maintaining the same minimum distance, saving computational cost.
• octahedron: octahedral box

Distance

Box margin, in nm, default 1.0. Effective only in non-membrane mode. The shortest distance from the solute to the box boundary when adding a solvent box to the system. Ignored in membrane mode (the membrane is pre-solvated and brings its own solvent box).

Add Water

Whether to add solvent (--add-water / --no-add-water), on by default (i.e., solvent is added). With --no-add-water, no new solvent is added (pre-existing water in the inputs is kept). Effective only in non-membrane mode. Ignored in membrane mode (the membrane is pre-solvated and brings its own solvent box).
Note: whether water is actually added also depends on whether the topology #includes a water model.

Results

The output result files include:

GMX Parameterization

简介

GROMACS 膜、受体（蛋白和核酸）、配体（小分子）、溶剂参数化工具，为膜/蛋白/核酸/小分子生成分子动力学模拟所需的拓扑文件、结构文件与参数文件（含位置限制文件）。

适用场景：

• 单组件参数化：只提供膜/受体/配体单一组件输入，输出单一组件参数化后的拓扑文件、结构文件和参数文件（含位置限制文件）。
• 多组件参数化：提供膜+受体+配体+溶剂等多组件输入，输出多组件参数后的拓扑文件、结构文件和参数文件（含位置限制文件）。多组件的力场/水模型统一。

参数说明

Membrane

膜结构输入文件，pdb 格式。提供即触发 membrane 的参数化。

Protein

蛋白结构输入文件，pdb 格式。提供即触发 receptor 的参数化。

Nucleic Acid

核酸结构输入文件，pdb 格式。提供即触发 receptor 的参数化，含核酸时力场必须为 Amber 力场。

Ligand

配体结构输入文件，单文件（pdb/sdf/mol2）或压缩包（zip/tar.gz/tgz/tar）。提供即触发 ligand 参数化，含配体时力场必须为 Amber 力场。

Ligand SMILES

配体结构的SMILES 模板，当配体加氢不准确时，可提供其SMILES用于准确加氢。格式为 Ligand1文件名或者Ligand1文件名前缀:SMILES;Ligand2文件名或者Ligand2文件名前缀:SMILES。当只有一个Ligand时也可以直接输入SMILES。

Ligand Charges

配体所带的电荷。格式为 Ligand1文件名或者Ligand1文件名前缀:Charge;Ligand2文件名或者Ligand2文件名前缀:Charge。当只有一个Ligand时也可以直接输入Charge值。

All Component Structures

包含体系中所有组分结构的压缩输入文件，可包含蛋白、核酸、膜、配体等组分，用于统一进行 GMX 参数化。

Ligand Multiplicity

配体的自旋多重度。格式为 Ligand1文件名或者Ligand1文件名前缀:Multiplicity;Ligand2文件名或者Ligand2文件名前缀:Multiplicity。当只有一个Ligand时也可以直接输入Multiplicity。

Charge Type

配体的电荷计算类型，bcc或者 resp。默认bcc。

Add Hydrogens

配体参数化过程中的加氢开关（--add-hydrogens / --no-add-hydrogens），默认开启加氢。

Solvent

外部水/离子输入文件，pdb 格式。模块会根据情况将外部水/离子的参数化结果合并到membrane、receptor 或者 ligand的拓扑文件、结构文件和参数文件（含位置限制文件）中。

Force Field

参数化过程中使用的力场，默认 amber14sb_parmbsc1。注意组合约束：含配体时必须选择 Amber 力场；含膜时只能选择 Amber/CHARMM 力场；含核酸时必须选择 Amber 力场。不兼容组合会报错。

• Amber（支持核酸、蛋白、膜、配体）：amber94、amber96、amber99、amber99sb、amber99sb-ildn、amberGS、amber03、amber14sb_parmbsc1
• CHARMM（仅蛋白、膜）：charmm27、charmm36-jul2020、charmm36-jul2022
• OPLS-AA（仅蛋白）：oplsaa
• GROMOS（仅蛋白）：gromos43a1、gromos43a2、gromos45a3、gromos53a5、gromos53a6、gromos54a7

Water Model

参数化过程中使用的水模型，默认 tip3p。膜不支持 none。

tip3p：三点水模型（默认）
spc / spce：简单点电荷水模型及扩展版
tip4p：四点水模型
tip5p：五点水模型
none：不使用水模型

结果说明

输出结果文件包括：

GMX Parameterization

Introduction

A GROMACS parameterization tool for membrane, receptor (protein and nucleic acid), ligand (small molecule), and solvent. It generates the topology, structure, and parameter (including position-restraint files) files required for molecular dynamics simulation of membrane/protein/nucleic acid/small molecules.

Applicable scenarios:

• Single-component parameterization: provide a single component input (membrane/receptor/ligand); output the parameterized topology, structure, and parameter (including position-restraint files) files for that single component.
• Multi-component parameterization: provide multiple component inputs such as membrane+receptor+ligand+solvent; output the parameterized topology, structure, and parameter (including position-restraint files) files for the multiple components (with a unified force field / water model).

Parameters

Membrane

Membrane structure input file, pdb format. Providing it triggers membrane parameterization.

Protein

Protein structure input file, pdb format. Providing it triggers receptor parameterization.

Nucleic Acid

Nucleic acid structure input file, pdb format. Providing it triggers receptor parameterization; the force field must be Amber when nucleic acid is present.

Ligand

Ligand structure input file, a single file (pdb/sdf/mol2) or an archive (zip/tar.gz/tgz/tar). Providing it triggers ligand parameterization; the force field must be amber when a ligand is present.

Ligand SMILES

SMILES template for the ligand structure. When hydrogen addition for the ligand is inaccurate, its SMILES can be supplied for accurate hydrogen addition. Format: Ligand1 filename or Ligand1 filename prefix:SMILES;Ligand2 filename or Ligand2 filename prefix:SMILES. When there is only one ligand, the SMILES can be entered directly.

Ligand Charges

Charge carried by the ligand. Format: Ligand1 filename or Ligand1 filename prefix:Charge;Ligand2 filename or Ligand2 filename prefix:Charge. When there is only one ligand, the Charge value can be entered directly.

All Component Structures

Compressed input file containing the structures of all components in the system, such as proteins, nucleic acids, membranes, and ligands, for unified GMX parameterization.

Ligand Multiplicity

Spin multiplicity of the ligand. Format: Ligand1 filename or Ligand1 filename prefix:Multiplicity;Ligand2 filename or Ligand2 filename prefix:Multiplicity. When there is only one ligand, the Multiplicity can be entered directly.

Charge Type

Charge calculation type for the ligand, bcc or resp. Default bcc.

Add Hydrogens

Hydrogen-addition switch during ligand parameterization (--add-hydrogens / --no-add-hydrogens), hydrogen addition on by default.

Solvent

External water/ion input file, pdb format. The module merges the external water/ion parameterization results into the topology, structure, and parameter (including position-restraint files) files of membrane, receptor, or ligand as appropriate.

Force Field

Force field used during parameterization, default amber14sb_parmbsc1. Note combination constraints: when a ligand is present, an Amber force field must be selected; when a membrane is present, only Amber/CHARMM force fields can be selected; when nucleic acid is present, an Amber force field must be selected. Incompatible combinations raise an error.

• Amber (supports nucleic acid, protein, membrane, ligand): amber94, amber96, amber99, amber99sb, amber99sb-ildn, amberGS, amber03, amber14sb_parmbsc1
• CHARMM (protein, membrane only): charmm27, charmm36-jul2020, charmm36-jul2022
• OPLS-AA (protein only): oplsaa
• GROMOS (protein only): gromos43a1, gromos43a2, gromos45a3, gromos53a5, gromos53a6, gromos54a7

Water Model

Water model used during parameterization, default tip3p. Membranes do not support none.

tip3p: three-site water model (default)
spc / spce: Simple Point Charge water model and extended variant
tip4p: four-site water models
tip5p: five-site water models
none: no water model

Results

The output result files include:

MD PDB Prepare

简介

基于 PDBFixer 的 PDB 结构预处理工具，用于将复合 PDB 文件自动拆分为膜、蛋白质、核酸、配体等组分（水分子与未匹配离子归入溶剂组分），并对蛋白质和核酸组分执行缺失残基补全和缺失重原子添加。

适用场景：

• 膜蛋白体系：从预测结构或实验结构中分离膜脂质、蛋白质、核酸及小分子配体
• 核酸-蛋白复合物：同时处理核酸与蛋白质，自动归属金属离子，并将水分子与未匹配离子归入溶剂组分
• 通用 PDB 预处理：对任意 PDB 结构进行组分拆分与结构修复

参数说明

Input File

输入 PDB 文件，必选参数。

结果说明

输出结果包括：

MD PDB Prepare

Introduction

A PDB structure preprocessing tool based on PDBFixer, designed to automatically split composite PDB files into components — membrane, protein, nucleic acid, and ligands (with water molecules and unmatched ions routed to the solvent component) — and perform missing residue completion and heavy atom addition for protein and nucleic acid components.

Applicable scenarios:

• Membrane protein systems: separating membrane lipids, proteins, nucleic acids, and small-molecule ligands from predicted or experimental structures
• Nucleic acid-protein complexes: processing both nucleic acids and proteins simultaneously, with automatic assignment of metal ions and routing of water molecules and unmatched ions into the solvent component
• General PDB preprocessing: component splitting and structural repair for arbitrary PDB structures

Parameters

Input File

Input PDB file, required parameter.

Results

The output includes:

Protein Design (Caliby)

简介

基于结构的蛋白质序列重设计（逆折叠）。基于 Caliby（Ensemble-conditioned protein sequence design）方法：以三维骨架为条件，用 Potts 模型在氨基酸序列空间采样，生成与目标结构相容、能量更优的设计序列。与仅依赖单一静态骨架的经典逆向折叠方法不同，Caliby 并不追求复现天然序列，而是通过对目标结构周围的多个变体进行建模（先用 Protpardelle-1c 对输入骨架做部分扩散，生成一组合成构象，再在系综约束下设计序列），增强结构一致性、自洽性与设计多样性，在benchmark中显著优于ProteinMPNN 和 ChromaDesign 等主流方法。特别是在 de novo 骨架和先前被 SolubleMPNN 判为“不可设计”的 binder 上，Caliby 展现出强大的适应性与拓展性，扩大了蛋白设计的可能空间。

核心思路：在固定（或局部可动）主链几何下，学习位点偏好场 (h) 与位点成对耦合 (J)，构成 Potts 能量 (U)；采样时在该能量景观上搜索低能序列。评价与筛选以 Potts Energy（(U)） 为主指标（数值越低通常越好），并辅以相对野生型的序列一致性。

实验结果表明，集成条件的Caliby在天然骨架和全新骨架上的表现均优于最先进的模型ProteinMPNN和ChromaDesign。在85个长度在100到1000之间的天然骨架上，集成条件的Caliby实现了更好的自洽性。即使是那些先前设计性较差、ProteinMPNN和ChromaDesign无法成功设计的具有挑战性的天然骨架，集成条件的Caliby也能成功设计。

典型应用场景：

• 抗体 / 纳米抗体 CDR 或结合界面局部重设计。
• 固定催化位点、结合口袋关键残基（Fix），对其余骨架做重设计（Design）。
• 多链复合物界面共设计。

参数说明

Structure

蛋白结构文件，格式支持.pdb、.cif

Chain

指定进行 Design/Fix 的链 ID，逗号分隔，如 A 或 A,B。

Type

设计残基模式：固定（Fix，指定Position参数中的残基在设计时保持不变）或者设计（Design，指定Position参数中的残基进行设计，而其他未指定残基在设计时保持不变）。

Position

设置链上的残基序号，对设置的残基根据Type选项进行固定或设计。格式为链名+残基编号（从1开始的位置顺序编号），支持范围设置，多段区域时用逗号分隔，如：A85-100,A114,B1-100表示：
Chain参数指定的A链上，第85至100,114号残基，B链上，第1至100号残基，进行Type指定形式的固定或设计。
链名也可省略，如：85-100,114，表示Chain参数指定的所有链上的相同位置残基，进行固定或设计。
注意：Chain与Position参数必须至少设置一个。

Omit AAs

设计时全局排除的残基类型，逗号分隔单字母代码，如 C,M。用于避免二硫键半胱氨酸或甲硫氨酸氧化等问题残基。

Restrict AA Type

逐个位点残基类型限制文本文件，每行设置链名+残基编号（从1开始的位置顺序编号）及其类型限制，如：

A6-10:QR
A18:HS
20-30:QR

表示A链上的6-10号残基，设计时，限制残基类型为Q和R，不能变为其他类型残基；A链18号残基，限制类型为H与S。
支持省略链名，表示应用到Chain参数指定的所有链。

Sampling Temp

采样氨基酸概率分布时使用的“温度系数”，用来控制随机性/多样性。建议的取值为0.1、0.15、0.2、0.25、0.3。较高的值会导致更多的多样性。默认：0.1，数值范围为0 ~ 1.5。

Number of Designs

设计的序列数量，目前最大支持1000条。

Seed

随机种子，影响系综与采样可复现性。

结果说明

汇总文件 caliby_mutants.csv
按复合物全序列去重后的设计列表，按 Potts 能量排序，便于挑选独特序列：

Name,Duplicate_Count,Potts_Energy
A_Y34I_G56N_P58T_E74T_B_V29I_S31N_T57E_H91A,4,-1982.1284
A_Y34M_G56N_P58A_E74T_B_V29I_S31N_T57E_H91A,6,-1981.9233

序列文件 caliby_designed.fasta

含野生型与去重后的设计序列。标题示例：

>A_WT_B_WT
...
>A_Y34I_G56N_..._B_V29I_...; overall_identity=0.965; nonfix_identity=0.758; potts_energy=-1982.1284; duplicate_count=4
...

batch格式的复合物序列 caliby_complex_batch.fasta，完整复合物序列，多链以 : 连接，便于下游批量重折叠。

设计结构的打包压缩文件 caliby_designed.tar.gz

参考文献

• Shuai, R.W., Lu, T., Bhatti, S., Kouba, P. & Huang, P.-S. Ensemble-conditioned protein sequence design with Caliby. bioRxiv (2025). DOI:10.1101/2025.09.30.679633

Protein Design (Caliby)

Introduction

Structure-based protein sequence redesign (inverse folding) based on Caliby (Ensemble-conditioned Protein Sequence Design). Conditioned on a three-dimensional protein backbone, Caliby uses a Potts model to sample the amino acid sequence space and generate designed sequences that are compatible with the target structure and have favorable energy scores.

Unlike conventional inverse folding methods that rely on a single static backbone, Caliby does not aim to reproduce the native sequence. Instead, it models multiple structural variants around the target structure. Protpardelle-1c is first used to partially diffuse the input backbone and generate an ensemble of synthetic conformations, after which sequence design is performed under ensemble constraints. This strategy improves structural consistency, self-consistency, and sequence diversity. In benchmark evaluations, Caliby significantly outperformed mainstream methods such as ProteinMPNN and ChromaDesign. In particular, Caliby demonstrated strong adaptability and design capability on de novo backbones and binders previously considered difficult to design by SolubleMPNN, thereby expanding the accessible protein design space.

Core Concept: Under a fixed or locally flexible backbone geometry, Caliby learns site-specific preference fields (h) and pairwise couplings between sites (J) to construct a Potts energy function (U). Sequence sampling is then performed over this energy landscape to identify low-energy sequences. Design evaluation and filtering primarily use Potts Energy (U), where lower values generally indicate better sequence–structure compatibility, together with sequence identity relative to the wild-type (WT) sequence.

Experimental results show that ensemble-conditioned Caliby outperforms state-of-the-art models such as ProteinMPNN and ChromaDesign on both native and de novo backbones. Across 85 native backbones ranging from 100 to 1000 residues in length, ensemble-conditioned Caliby achieved improved self-consistency. It was also able to successfully design challenging native backbones for which ProteinMPNN and ChromaDesign showed poor designability.

Typical Applications:

• Local redesign of antibody / nanobody CDRs or binding interfaces.
• Fixing catalytic sites or key binding-pocket residues (Fix) while redesigning the remaining backbone positions (Design).
• Interface co-design for multi-chain protein complexes.

Parameters

Structure

Protein structure file. Supported formats: .pdb and .cif.

Chain

Specify the chain IDs to be used for Design/Fix operations. Multiple chain IDs should be separated by commas, e.g., A or A,B.

Type

Residue design mode:

• Fix: Residues specified by the Position parameter remain unchanged during design.
• Design: Residues specified by the Position parameter are redesigned, while all unspecified residues remain unchanged.

Position

Specify residue positions on the selected chains. The specified residues are fixed or redesigned according to the Type parameter.

The format is chain ID + residue number, where residue numbers correspond to sequential positions starting from 1. Residue ranges are supported, and multiple regions should be separated by commas.

For example:

A85-100,A114,B1-100

This indicates that residues 85–100 and 114 of chain A, and residues 1–100 of chain B specified by the Chain parameter, will be processed according to the selected Type.

The chain ID can also be omitted. For example:

85-100,114

This applies the specified positions to all chains defined by the Chain parameter.

Note: At least one of the Chain or Position parameters must be specified.

Omit AAs

Specify amino acid types to be globally excluded during sequence design. Use single-letter amino acid codes separated by commas, e.g., C,M.

This parameter can be used to prevent the introduction of undesired residues, such as cysteine or methionine, when avoiding potential issues related to disulfide bonds or methionine oxidation.

Restrict AA Type

A text file used to restrict the allowed amino acid types at specific residue positions.

Each line specifies a chain ID + residue number (sequentially numbered from 1) and the allowed amino acid types. For example:

A6-10:QR
A18:HS
20-30:QR

This indicates that residues 6–10 of chain A can only be designed as Q or R, while residue 18 of chain A can only be designed as H or S.

The chain ID can be omitted, in which case the restriction is applied to the corresponding positions of all chains specified by the Chain parameter.

Sampling Temp

The temperature coefficient used when sampling from the amino acid probability distribution. It controls the randomness and diversity of generated sequences.

Recommended values are 0.1, 0.15, 0.2, 0.25, and 0.3. Higher values generally result in greater sequence diversity.

Default: 0.1
Range: 0 ~ 1.5

Number of Designs

Number of sequences to design. Currently, up to 1000 sequences are supported.

Seed

Random seed that affects ensemble generation and sequence sampling reproducibility.

Results

Summary file caliby_mutants.csv

This file contains the deduplicated design results based on full complex sequences. Designs are ranked by Potts Energy to facilitate the selection of unique sequences:

Name,Duplicate_Count,Potts_Energy
A_Y34I_G56N_P58T_E74T_B_V29I_S31N_T57E_H91A,4,-1982.1284
A_Y34M_G56N_P58A_E74T_B_V29I_S31N_T57E_H91A,6,-1981.9233

Sequence file caliby_designed.fasta

Contains the wild-type sequence and deduplicated designed sequences. Example headers:

>A_WT_B_WT
...
>A_Y34I_G56N_..._B_V29I_...; overall_identity=0.965; nonfix_identity=0.758; potts_energy=-1982.1284; duplicate_count=4
...

Batch-formatted complex sequence file caliby_complex_batch.fasta, containing complete complex sequences. For multi-chain complexes, individual chain sequences are separated by : for use in downstream batch refolding or structure prediction.

Compressed archive of designed structures: caliby_designed.tar.gz

References

• Shuai, R.W., Lu, T., Bhatti, S., Kouba, P. & Huang, P.-S. Ensemble-conditioned protein sequence design with Caliby. bioRxiv (2025). DOI:10.1101/2025.09.30.679633

Nanobody Design (mBER)

简介

基于靶点结构设计纳米抗体。模块基于 mBER 实现，mBER 建立在 ColabDesign 框架之上，基于靶点结构与纳米抗体的 FR 模板序列，通过反向传播（backpropagation）优化纳米抗体 CDR 区序列。

设计过程如下图所示：

其核心创新在于两大条件约束：

工作流程分为三步：

• 模板生成：准备靶点结构（可截断），使用 NanoBodyBuilder2 创建 VHH 结构模板。
• 轨迹优化：基于 AlphaFold-Multimer 的损失函数，迭代优化连续序列表示，结合 ESM-2 的序列先验。
• 评估：使用 AlphaFold-Multimer 置信度指标（ipTM、pLDDT）重新折叠并评分，筛选最优候选。

实验验证规模与性能

mBER 经历了迄今为止最大规模的从头抗体设计验证实验：

• 设计规模：针对 436 个多样化人源受体，设计了超过 100 万条 VHH 序列。
• 实验筛选：对 145 个靶点测试了 1,153,241 个设计，测量了超过 1 亿次抗体-靶点相互作用。
• 成功率：在 45% 的测试靶点（65/145）中检测到特异性结合。
• 高精度成功率：经过严格计算过滤后，针对特定表位的命中率最高可达 38%–40%。

这一数据使其成为当前开源抗体设计工具中实验成功率最高的方案之一。

参数说明

Structure

靶点结构文件，PDB 或 CIF 格式。

Chains

指定从 Structure 中提取相应的链作为靶点，可多选，例如 A,B。如不设置该参数，表示提取结构中的所有链作为靶点。

Hotspot

设置靶点结构中的结合位点，格式使用链名和残基位置，例如 A45,A67,A80-82 表示 A 链中编号 45、67、80 至 82 的残基作为结合位点。

Template

纳米抗体模板序列。若不设置该参数，则使用默认模板序列。模板序列使用文本格式，非 FASTA，支持输入多条，每行放置一条；设计过程中会循环使用每条模板序列。

示例：

EVQLVESGGGLVQPGGSLRLSCAAS7AMAWFRQAPGQEREFVAGI6TLYADSVRGRFTNSRDNSKNTLYLQMNSLRAEDTAVYYCAA17WGQGTLVTVSS

其中：

• 字母表示序列中明确的残基，设计中保持不变。
• 数值表示该位置要设计的长度，例如 17 表示该位置将设计 17 个残基，通常用于 CDR 区域设计。

默认纳米抗体模板：

• 7EOW — Caplacizumab（卡普赛珠单抗）：靶点为 vWF A1 域；作用为阻断 vWF 与血小板结合，抑制血栓形成。
• 7XL0 — Vobarilizumab（ALX-0061，沃巴利珠单抗）：靶点为 IL-6R（+ 白蛋白结合）；作用为抑制 IL-6 信号并延长半衰期。
• 8COH — TPP-3444（Gefurulimab / ALXN1720 组成部分）：靶点为补体 C5；作用为抑制补体激活。
• 8Z8V — ALB8（Ozoralizumab / ATN-103 组件）：靶点为人血清白蛋白（HSA）；作用为延长药物半衰期。
• Gontivimab（ALX-0171，格替韦单抗）：靶点为 RSV F 蛋白；作用为阻断病毒融合，抑制感染。
• Isecarosmab（M-6495 / ALX-1141，艾司卡索单抗）：靶点为 ADAMTS-5；作用为抑制软骨降解，具有抗炎作用。
• Sonelokimab：靶点为 IL-17A / IL-17F；作用为双重抑制炎症因子，增强抗炎效果。

Number of Designs

设计完成后最终输出的序列数量。

结果说明

输出成功过滤的设计结果 mBER_designed.csv，包含信息如下：

输出成功过滤的设计结果对应的序列文件 mBER_designed.fasta。

输出 Batch 格式的复合物序列文件 mBER_designed_complex_batch.fasta。

输出打包压缩文件 mBER_final_results.tar.gz。

参考文献

• Swanson E., Nichols M., Ravichandran S., et al. 2025, mBER: Controllable de novo antibody design with million-scale experimental screening, bioRxiv, Preprint, DOI: 10.1101/2025.09.26.678877

Nanobody Design (mBER)

Introduction

This module designs nanobodies based on a target structure. It is implemented using mBER, which is built on the ColabDesign framework. Given a target structure and a nanobody FR template sequence, mBER optimizes nanobody CDR sequences through backpropagation.

The design process is shown below:

Its core innovation is based on two conditional constraints:

The workflow has three steps:

• Template generation: Prepare the target structure, optionally truncated, and use NanoBodyBuilder2 to create a VHH structural template.
• Trajectory optimization: Iteratively optimize continuous sequence representations using AlphaFold-Multimer loss functions together with ESM-2 sequence priors.
• Evaluation: Refold and score candidates using AlphaFold-Multimer confidence metrics, including ipTM and pLDDT, and select the best candidates.

Experimental validation scale and performance

mBER has undergone the largest de novo antibody design validation experiment to date:

• Design scale: More than 1 million VHH sequences were designed against 436 diverse human receptors.
• Experimental screening: 1,153,241 designs were tested against 145 targets, with over 100 million antibody-target interactions measured.
• Success rate: Specific binding was detected for 45% of tested targets, corresponding to 65 out of 145 targets.
• High-precision success rate: After strict computational filtering, hit rates for specific epitopes reached up to 38%–40%.

These results make mBER one of the open-source antibody design tools with the highest experimental success rates.

Parameters

Structure

Target structure file in PDB or CIF format.

Chains

Chains extracted from Structure as the target. Multiple chains can be selected, for example A,B. If not specified, all chains in the structure are used as the target.

Hotspot

Binding sites on the target structure. The format uses chain name and residue position, for example A45,A67,A80-82, meaning residues 45, 67, and 80 to 82 on chain A are used as binding sites.

Template

Nanobody template sequence. If not specified, the default template sequences are used. Template sequences use plain text format rather than FASTA. Multiple sequences are supported, with one sequence per line; each template sequence is used cyclically during design.

Example:

EVQLVESGGGLVQPGGSLRLSCAAS7AMAWFRQAPGQEREFVAGI6TLYADSVRGRFTNSRDNSKNTLYLQMNSLRAEDTAVYYCAA17WGQGTLVTVSS

Where:

• Letters indicate fixed residues in the sequence that remain unchanged during design.
• Numbers indicate the length to design at that position. For example, 17 means 17 residues will be designed at that position, usually for CDR region design.

Default nanobody templates:

• 7EOW — Caplacizumab: Target is the vWF A1 domain; function is blocking vWF-platelet binding and inhibiting thrombosis.
• 7XL0 — Vobarilizumab (ALX-0061): Target is IL-6R (+ albumin binding); function is inhibiting IL-6 signaling and extending half-life.
• 8COH — TPP-3444 (component of Gefurulimab / ALXN1720): Target is complement C5; function is inhibiting complement activation.
• 8Z8V — ALB8 (component of Ozoralizumab / ATN-103): Target is human serum albumin (HSA); function is extending drug half-life.
• Gontivimab (ALX-0171): Target is RSV F protein; function is blocking viral fusion and inhibiting infection.
• Isecarosmab (M-6495 / ALX-1141): Target is ADAMTS-5; function is inhibiting cartilage degradation with anti-inflammatory activity.
• Sonelokimab: Targets IL-17A / IL-17F; function is dual inhibition of inflammatory cytokines and enhanced anti-inflammatory effect.

Number of Designs

Final number of sequences to output after design.

Results

The filtered design results are output as mBER_designed.csv, containing the following information:

The sequence file corresponding to the filtered design results is output as mBER_designed.fasta.

The Batch-format complex sequence file is output as mBER_designed_complex_batch.fasta.

The compressed archive is output as mBER_final_results.tar.gz.

References

• Swanson E., Nichols M., Ravichandran S., et al. 2025, mBER: Controllable de novo antibody design with million-scale experimental screening, bioRxiv, Preprint, DOI: 10.1101/2025.09.26.678877

Biologics Long-term Stability

简介

抗体注射液在储存过程中，分子可能逐渐发生聚集、降解等变化，影响纯度与安全性。稳定性研究会在不同温度下存放样品，定期取样检测，观察这些变化随时间的发展趋势，为货架期（有效期）评估提供依据。

本模块的作用是：在已有短期、多温度稳定性实测数据的基础上，用动力学模型拟合降解规律，并外推至更长存放时间（默认 36 个月），给出各组分未来含量的预测值及 95% 预测区间，用于加速稳定性数据上的长期趋势分析。

LT/HT 双分支动力学源于 Bunc et al.（2022）ATM；整体预测流程（多温度联合拟合、Monte Carlo 预测区间等）参考 Dillon et al.（2024）AKM 方法。

从湿实验到预测：整体流程

背景知识

SEC、单体与 HMW（示例一相关）

典型转变过程： 刚配制时单体约占 98%、HMW 约占 2%；随存放时间延长，部分单体分子碰撞、折叠异常或化学修饰，逐渐聚集成 HMW，表现为 单体峰面积占比下降、HMW 峰面积占比上升。温度越高，这一变化通常越快。

IEC 与电荷异构体（示例二相关）

模型选择：双组分与三组分

模型由输入数据CSV文件中的列组成自动确定，无需手动切换：

列名非默认值时，通过 Temperature Column、Time Column、Component A/B/C Column 等参数映射。

输入数据

数据格式示例

temp,time,A,B
5,0.0,98.8,1.2
25,3.0,89.7,10.3
40,3.0,79.3,20.7

数值一般保留一位小数；每行一条记录（某一温度、某一时间点）。质量要求： 建议 ≥ 2～3 个不同存放温度（具体温度点按稳定性方案确定，常见如 5 / 25 / 40°C，亦可为其他组合）；每温 ≥ 4 个时间点（含 t=0）；高温长时间下单体 A 应有可观测下降；t=0 时组分之和 ≤ 101%。

参数说明

Input File

用于预测的数据文件，必传参数。CSV 格式，含存放温度、取样时间（月）及各组分含量列。

Temperature Column

温度列列名。默认 temp。

Time Column

取样时间列列名，单位为月。默认 time。

Component A / B / C Column

各组分含量列列名，单位为 %。默认 A、B、C。Component C Column 为可选；存在对应列时自动启用三组分模型。

Pathway

动力学通路模式，可选值：

• auto（默认）：数据较丰富且 LT/HT 拟合明显优于 simple 时选用 LT/HT，否则选用 simple
• simple：强制单路径模型，全温度共用一套速率参数
• lt_ht：强制 LT/HT 双分支模型（Dillon 文献方法）

Fit Mode

拟合模式，可选值：

• joint（默认）：A+B (+C)≈100时，采用联合建模。
• separate：不满足A+B (+C)≈100时，A，B（，C）单独建模。

Predict Temps

预测温度列表，逗号分隔，单位 °C。默认取 Input File 中的全部温度。

Predict Step

预测时间步长，单位月。默认 0.5。

Predict Max

预测时间上限，单位月。默认 36。

MC

蒙特卡洛 Bootstrap 重采样次数。默认 5000；次数越大，预测区间估计越稳定，计算耗时越长。

输出结果

示例一：SEC 单体 + HMW（双组分）

1. 湿实验做了什么

某抗体注射液开展 SEC 稳定性考察：

观察到的现象： 刚入组时（t=0）单体约 98%、HMW 约 2%；随存放时间延长，单体下降、HMW 上升；本例中 40°C 降得最快、5°C 最慢——符合「温度越高、聚集越快」的直观认识。

2. 初始实验数据（节选）

将上述检测结果整理为 CSV（不含 C 列 → 模块自动选用双组分 A→B 模型）。表中 A = 单体%，B = HMW%；每行对应「某一温度、某一时间点」的一次 SEC 检测结果。本例含三个温度共 12 条记录，此处节选其中 6 条：

temp,time,A,B
5,0.0,98.8,1.2
5,6.0,93.8,6.2
25,0.0,98.5,1.5
25,6.0,81.6,18.4
40,0.0,98.8,1.2
40,3.0,79.3,20.7

3. 拟合结果解读

拟合结果： a2b，simple；RMSE 0.13%；A₀=98.7%，B₀=1.3%；k(25°C)=0.032/月，Ea=10.4 kcal/mol。

4. 长期预测结果（36 个月）

报告表述示例： 基于 5 / 25 / 40°C SEC 稳定性数据外推，预测 25°C 存放 36 个月时单体为 31.7%（95% PI：30.6%–33.0%），HMW 为 68.3%（95% PI：67.1%–69.4%）。

示例二：IEC 主峰 + 酸性 + 碱性（三组分）

1. 湿实验做了什么

某抗体开展 IEC（离子交换色谱）稳定性考察，追踪电荷异构体变化：

观察到的现象： 刚入组时（t=0）主峰约 37%、酸性约 7%、碱性约 56%；随存放时间延长，主峰下降，酸性与碱性组分上升；本例中 40°C 变化最快、5°C 最慢——符合「温度越高、电荷异构体转化越快」的趋势。降解产物中 碱性组分增幅大于酸性组分，提示 A→C 通路相对更活跃。

2. 初始实验数据（节选）

将上述检测结果整理为 CSV（含 C 列 → 模块自动选用三组分 A→B+C 模型）。表中 A = 主峰%，B = 酸性%，C = 碱性%；每行对应「某一温度、某一时间点」的一次 IEC 检测结果。本例含三个温度共 16 条记录，此处节选其中 6 条：

temp,time,A,B,C
5,0.0,37.1,7.4,55.5
5,24.0,30.6,8.6,60.8
25,0.0,37.3,7.1,55.6
25,6.0,17.0,15.4,67.6
40,0.0,37.2,7.3,55.5
40,2.0,5.4,25.5,69.1

3. 拟合结果解读

拟合结果： a2bc，simple；RMSE 0.12%；A₀=37.2%，B₀=7.3%，C₀=55.6%；k_B(25°C)=0.052/月，k_C(25°C)=0.078/月（A→C 更快）。

4. 长期预测结果（36 个月）

报告表述示例： 基于 5 / 25 / 40°C IEC 稳定性数据外推，预测 25°C 存放 36 个月时主峰为 0.3%（95% PI：0.1%–0.6%），酸性组分为 22.1%（95% PI：21.4%–22.8%），碱性组分为 77.6%（95% PI：76.9%–78.3%）。

补充说明

• 拟合质量： 查阅 report.pdf 中实测散点是否落在预测曲线及 PI 阴影内。
• 预测区间较宽： 可补充高温长时间数据，或通过 Predict Max 缩短外推范围。
• MC： 正式分析建议默认 5000；快速预览可适当减小。

参考文献

• Bunc M., et al. 2022, Aggregation Time Machine, Journal of Medicinal Chemistry, 65, 2623-2632, DOI: 10.1021/acs.jmedchem.1c02010
• Dillon M., et al. 2024, Predicting the Long-Term Stability of Biologics with Short-Term Data, Molecular Pharmaceutics, 21, 4673-4687, DOI: 10.1021/acs.molpharmaceut.4c00609

Biologics Long-term Stability

Introduction

During storage of antibody injections, molecules may gradually aggregate, degrade, or undergo other changes that affect purity and safety. Stability studies store samples at different temperatures and test them at scheduled time points to observe how these changes develop over time, providing evidence for shelf-life assessment.

This module uses existing short-term, multi-temperature stability measurements to fit degradation kinetics and extrapolate to longer storage durations, with a default maximum of 36 months. It reports predicted future component contents and 95% prediction intervals, supporting long-term trend analysis from accelerated stability data.

LT/HT dual-branch kinetics is based on the ATM method from Bunc et al. (2022). The overall prediction workflow, including joint multi-temperature fitting and Monte Carlo prediction intervals, follows the AKM approach described by Dillon et al. (2024).

From Wet Experiment to Prediction: Overall Workflow

Background

SEC, Monomer, and HMW (Related to Example 1)

Typical transition process: At initial preparation, monomer is about 98% and HMW is about 2%. As storage time increases, some monomer molecules collide, misfold, or undergo chemical modification and gradually aggregate into HMW. This appears as a decrease in monomer peak area percentage and an increase in HMW peak area percentage. Higher temperature usually accelerates this change.

IEC and Charge Variants (Related to Example 2)

Model Selection: Two Components and Three Components

The model is automatically determined by the columns present in the input CSV file; no manual switching is required:

If column names are not the defaults, they can be mapped using Temperature Column, Time Column, and Component A/B/C Column.

Input Data

Data Format Example

temp,time,A,B
5,0.0,98.8,1.2
25,3.0,89.7,10.3
40,3.0,79.3,20.7

Values are generally kept to one decimal place. Each row is one record for a specific temperature and time point. Quality requirements: It is recommended to include at least 2-3 different storage temperatures (specific temperature points follow the stability protocol, commonly 5 / 25 / 40°C, but other combinations are also possible); at least 4 time points per temperature, including t=0; under high-temperature long-duration conditions, monomer A should show an observable decrease; at t=0, the sum of components should be ≤ 101%.

Parameters

Input File

Data file used for prediction. Required. CSV format, containing storage temperature, sampling time in months, and component content columns.

Temperature Column

Temperature column name. Default temp.

Time Column

Sampling time column name, in months. Default time.

Component A / B / C Column

Column names for component contents, in %. Defaults are A, B, and C. Component C Column is optional; when the corresponding column exists, the three-component model is enabled automatically.

Pathway

Kinetic pathway mode. Available values:

• auto (default): uses LT/HT when the dataset is sufficiently rich and LT/HT fits clearly better than simple; otherwise uses simple
• simple: forces a single-pathway model, with all temperatures sharing one set of rate parameters
• lt_ht: forces the LT/HT dual-branch model, following the Dillon method

Fit Mode

Fitting mode. Available values:

• joint (default): uses joint modeling when A+B (+C)≈100.
• separate: models A, B, and optionally C separately when A+B (+C)≈100 is not satisfied.

Predict Temps

List of prediction temperatures, comma-separated, in °C. By default, all temperatures in Input File are used.

Predict Step

Prediction time step, in months. Default 0.5.

Predict Max

Maximum prediction time, in months. Default 36.

MC

Number of Monte Carlo bootstrap resampling iterations. Default 5000; larger values provide more stable prediction interval estimates but require longer computation time.

Results

Example 1: SEC Monomer + HMW (Two Components)

1. What Was Done in the Wet Experiment

An antibody injection was evaluated in an SEC stability study:

Observed phenomenon: At t=0, monomer is about 98% and HMW is about 2%. As storage time increases, monomer decreases and HMW increases. In this example, 40°C decreases fastest and 5°C slowest, consistent with the intuitive expectation that higher temperature accelerates aggregation.

2. Initial Experimental Data (Excerpt)

The measured results are organized as a CSV file (without a C column → the module automatically selects the two-component A→B model). In the table, A = monomer %, B = HMW %. Each row corresponds to one SEC measurement at a specific temperature and time point. The full example contains 12 records across three temperatures; 6 records are shown here:

temp,time,A,B
5,0.0,98.8,1.2
5,6.0,93.8,6.2
25,0.0,98.5,1.5
25,6.0,81.6,18.4
40,0.0,98.8,1.2
40,3.0,79.3,20.7

3. Interpretation of Fitting Results

Fitting result: a2b, simple; RMSE 0.13%; A₀=98.7%, B₀=1.3%; k(25°C)=0.032/month, Ea=10.4 kcal/mol.

4. Long-term Prediction Results (36 Months)

Example report statement: Based on extrapolation from 5 / 25 / 40°C SEC stability data, the predicted monomer content after 36 months at 25°C is 31.7% (95% PI: 30.6%–33.0%), and HMW is 68.3% (95% PI: 67.1%–69.4%).

Example 2: IEC Main Peak + Acidic + Basic (Three Components)

1. What Was Done in the Wet Experiment

An antibody was evaluated in an IEC (Ion Exchange Chromatography) stability study to track charge variant changes:

Observed phenomenon: At t=0, the main peak is about 37%, acidic is about 7%, and basic is about 56%. As storage time increases, the main peak decreases while acidic and basic components increase. In this example, 40°C changes fastest and 5°C slowest, consistent with the trend that higher temperature accelerates charge variant conversion. Among degradation products, the basic component increases more than the acidic component, suggesting that the A→C pathway is relatively more active.

2. Initial Experimental Data (Excerpt)

The measured results are organized as a CSV file (with a C column → the module automatically selects the three-component A→B+C model). In the table, A = main peak %, B = acidic %, C = basic %. Each row corresponds to one IEC measurement at a specific temperature and time point. The full example contains 16 records across three temperatures; 6 records are shown here:

temp,time,A,B,C
5,0.0,37.1,7.4,55.5
5,24.0,30.6,8.6,60.8
25,0.0,37.3,7.1,55.6
25,6.0,17.0,15.4,67.6
40,0.0,37.2,7.3,55.5
40,2.0,5.4,25.5,69.1

3. Interpretation of Fitting Results

Fitting result: a2bc, simple; RMSE 0.12%; A₀=37.2%, B₀=7.3%, C₀=55.6%; k_B(25°C)=0.052/month, k_C(25°C)=0.078/month (A→C is faster).

4. Long-term Prediction Results (36 Months)

Example report statement: Based on extrapolation from 5 / 25 / 40°C IEC stability data, the predicted main peak after 36 months at 25°C is 0.3% (95% PI: 0.1%–0.6%), acidic component is 22.1% (95% PI: 21.4%–22.8%), and basic component is 77.6% (95% PI: 76.9%–78.3%).

Supplementary Notes

• Fitting quality: Check whether measured scatter points in report.pdf fall within the prediction curves and PI shading.
• Wide prediction intervals: Add high-temperature long-duration data, or shorten the extrapolation range using Predict Max.
• MC: The default 5000 is recommended for formal analysis; it can be reduced for quick preview.

References

• Bunc M., et al. 2022, Aggregation Time Machine, Journal of Medicinal Chemistry, 65, 2623-2632, DOI: 10.1021/acs.jmedchem.1c02010
• Dillon M., et al. 2024, Predicting the Long-Term Stability of Biologics with Short-Term Data, Molecular Pharmaceutics, 21, 4673-4687, DOI: 10.1021/acs.molpharmaceut.4c00609

Protein Sturcture Search (Foldseek)

模块用于在PDB100数据库中检索与输入蛋白复合物相似的结构。模块在接收复合物结构文件后，使用Foldseek-Multimer完成复合物结构比对，并将命中结构的复合物评分、链配对关系、链级比对指标和残基位置映射整理为 CSV 文件，便于进一步筛选和分析结构相似的蛋白复合物。

参数说明

Complex Structure

输入待检索的蛋白复合物结构，支持 PDB/CIF，结构中至少需要包含两条不同的蛋白链。

Alignment Type

指定链级结构比对方式。可选：TM-align 和 3Di+AA 。默认：3Di+AA

• 3Di/AA：使用3Di结构字母表进行快速预筛选，并使用基于“3Di结构字母表 + BLOSUM62”的 Smith-Waterman-Gotoh 算法进行比对（局部比对）。
• TM-align：使用3Di结构字母表进行快速预筛选，并使用 TM-align 进行比对（全局比对）。

结果说明

输出 alignment.csv，每行表示一个复合物命中结果：

输出 alignment_detail.csv，每行表示一组链配对中的相同或相似残基映射：

参考文献

• Kim W, Mirdita M, Levy Karin E, Gilchrist CLM, Schweke H, Söding J, Levy ED, Steinegger M. Rapid and sensitive protein complex alignment with Foldseek-Multimer.DOI:0.1038/s41592-025-02593-7

Protein Structure Search (Foldseek)

This module is used to search the PDB100 database for structures similar to an input protein complex. After receiving a complex structure file, the module uses Foldseek-Multimer to perform protein complex structure alignment and organizes the complex-level scores, chain-pairing relationships, chain-level alignment metrics, and residue position mappings of the matched structures into CSV files, facilitating further screening and analysis of structurally similar protein complexes.

Parameters

Complex Structure

The protein complex structure to be searched. PDB and CIF formats are supported. The structure must contain at least two distinct protein chains.

Alignment Type

Specifies the chain-level structural alignment method. Available options: TM-align and 3Di+AA. Default: 3Di+AA

• 3Di+AA: Uses the 3Di structural alphabet for rapid prefiltering and performs local alignment using the Smith-Waterman-Gotoh algorithm based on the “3Di structural alphabet + BLOSUM62”.
• TM-align: Uses the 3Di structural alphabet for rapid prefiltering and performs global alignment using TM-align.

Results

The module outputs alignment.csv, where each row represents one protein complex hit:

The module outputs alignment_detail.csv, where each row represents the mapping of identical or similar residues in one chain pairing:

References

• Kim W, Mirdita M, Levy Karin E, Gilchrist CLM, Schweke H, Söding J, Levy ED, Steinegger M. Rapid and sensitive protein complex alignment with Foldseek-Multimer.DOI:0.1038/s41592-025-02593-7

GMX MD [MuXi]

简介

提交GROMACS对应文件，从而进行分子动力学模拟，得到平衡模拟后得到的轨迹文件。

参数说明

GRO File

提交模拟体系的gro文件。该文件可以从MD Solvation或者Membrane Solvation模块获取。

Topology File

提交模拟体系的top文件。该文件可以从MD Solvation或者Membrane Solvation模块获取。

ITP File

提交模拟体系的itp文件。该文件可以从MD Solvation或者Membrane Solvation模块获取。

Minimize MDP File

提交进行最小化的参数化文件，文件格式为mdp。可以根据GMX MDP Generation模块的Minimization方法或者**GMX MDP Generation (Auto)**生成。

NPT MDP File

提交进行等压等温的参数化文件，文件格式为mdp。可以根据GMX MDP Generation模块的NPT方法或者**GMX MDP Generation (Auto)**生成。

MD MDP File

提交进行平衡模拟的参数化文件，文件格式为mdp。可以根据GMX MDP Generation模块的MD方法或者**GMX MDP Generation (Auto)**生成。

结果说明

输出结果包括：

参考文献

GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. M. J. Abraham, T. Murtola, R. Schulz, S. Páll, J. C. Smith, B. Hess, E. Lindahl, SoftwareX, 1, (2015), 19-25.

GMX MD [MuXi]

Introduction

Submit corresponding files to GROMACS to perform molecular dynamics simulations and obtain trajectory files after equilibrium simulations.

Parameter Description

GRO File

Submit the gro file of the simulated system. This file can be obtained from the MD Solvation or Membrane Solvation modules.

Topology File

Submit the top file of the simulated system. This file can be obtained from the MD Solvation or Membrane Solvation modules.

ITP File

Submit the itp file of the simulated system. This file can be obtained from the MD Solvation or Membrane Solvation modules.

Minimize MDP File

Submit the script file for minimization, in mdp format. This file can be generated using the GMX MDP Generation module with the Minimization method or GMX MDP Generation (Auto).

NPT MDP File

Submit the script file for NPT (isothermal-isobaric) simulation, in mdp format. This file can be generated using the GMX MDP Generation module with the NPT method or GMX MDP Generation (Auto).

MD MDP File

Submit the script file for equilibrium simulation, in mdp format. This file can be generated using the GMX MDP Generation module with the MD method or GMX MDP Generation (Auto).

Result Description

The output results include:

Reference

GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. M. J. Abraham, T. Murtola, R. Schulz, S. Páll, J. C. Smith, B. Hess, E. Lindahl, SoftwareX, 1, (2015), 19-25.

Binder Screening (pH-dependent)

简介

Binder Screening (pH-dependent) 是一个用于筛选 pH 敏感型 binder 序列的结构分析工具。该工具基于 pH_score、ddG_elec、氢键能量以及 HIS-带电残基配对分析，对输入的 target-binder 复合物结构进行多维度评估。

核心思想
该模块主要通过以下指标进行筛选：

• pH Score：基于组氨酸氢键模式评估界面 pH 敏感性。
• ddG Elec：使用 OpenMM 计算不同 pH 条件下的静电结合能变化。
• Hydrogen Bond Energy：分析 binder 内部 HIS-HIS、HIS-LYS、HIS-ARG 氢键网络的稳定性。
• HIS-Charged Residue Pairs：统计 HIS 与目标区域阳离子或阴离子残基的空间邻近关系。

该工具支持 PDB、CIF 和 mmCIF 格式输入，适用于界面 HIS 筛选、埋藏氢键网络筛选以及反向设计筛选等场景。

参数说明

Input Files

输入结构文件，必选参数。支持单个文件、目录或压缩包输入，可接受 .pdb、.cif、.mmcif 格式，以及 .zip、.tar.gz、.tgz 压缩包。

Binder Chains

Binder 链 ID，例如 A,B。与 Binder Residues 二选一。

Binder Residues

Binder 残基范围，例如 A:1-50,B:30 或 1-50,60-80。与 Binder Chains 二选一；未指定链名且多个链含相同残基编号时会报错。

Target Chains

Target 链 ID，格式与 Binder Chains 相同。未指定 Target Chains 和 Target Residues 时，默认使用除 binder 外的所有残基作为 target。

Target Residues

Target 残基范围，格式与 Binder Residues 相同。未指定 Target Chains 和 Target Residues 时，默认使用除 binder 外的所有残基作为 target。

Strategy

筛选策略，默认值为 interface。

• interface：界面 HIS 筛选策略，使用 pH_score、ddG_elec 过滤和 HIS-阳离子残基配对计数。
• network：埋藏网络筛选策略，检测 binder 内部 HIS 氢键网络能量。
• reverse：反向界面 HIS 筛选策略，使用 ddG_elec 和 HIS-阴离子残基配对计数，不计算 pH_score。
• both：同时运行 interface 和 network 两种筛选策略。

Minimize Method

结构能量最小化方法，默认值为 openmm，适用于所有筛选策略。

• openmm：使用 OpenMM 内置 LocalEnergyMinimizer，速度较快，但结果不完全可复现。
• capped-sd：自定义 capped steepest descent 优化，包含三阶段流程：冲突消除、氢原子与侧链弛豫、全原子优化，结果可复现。

Interface Cutoff

界面残基距离阈值，默认值为 6.0 Å。仅用于 interface 和 reverse 策略。

ddG Elec Threshold

ddG_elec 阈值，默认值为 0.0 kcal/mol。仅用于 interface 和 reverse 策略。
在 interface 策略中，ddG_elec >= threshold 表示通过筛选，即低 pH 下静电排斥增强。
在 reverse 策略中，ddG_elec <= threshold 表示通过筛选，即低 pH 下静电吸引增强。

Skip ddG Elec

是否跳过 ddG_elec 计算。仅用于 interface 策略。reverse 策略不支持跳过 ddG_elec，因为该指标是反向筛选的核心指标。

Min HIS Count

HIS-带电残基配对的最小数量，默认值为 1。仅用于 interface 和 reverse 策略。
在 interface 策略中，统计 HIS 附近的 target 阳离子残基 K、R、H。
在 reverse 策略中，统计 HIS 附近的 target 阴离子残基 D、E。

Buried SASA Ratio

埋藏残基 SASA 比例阈值，默认值为 0.2。当残基 SASA 比例，即实际 SASA / 标准氨基酸 SASA，低于该阈值时，该残基被认为是埋藏残基。仅用于 network 策略。

Network HBond Distance

网络氢键距离阈值，默认值为 3.5 Å。仅用于 network 策略。当 HIS 受体原子 ND1/NE2 与配对供体原子之间的距离低于该阈值时，记录为氢键。供体原子包括 ARG NH1/NH2/NE、LYS NZ 或 HIS ND1/NE2。

HBond Energy Threshold

氢键能量阈值，默认值为 -0.5 kcal/mol。仅用于 network 策略。当 binder 内部以 HIS 为受体、HIS/ARG/LYS 为供体的最强氢键能量低于该阈值时，结构通过筛选。

结果说明

输出结果包括：

其中 {strategy}_top5_{basename}_pH{N}.pdb 信息如下：

1. top5 规则：按各策略排序规则取每种策略排名前 5 的 passed PDB 结构。interface 按 ddG_elec 降序，设置 Skip ddG Elec 时按 pH_score 降序；reverse 按 ddG_elec 升序；network 按 best_hbond_energy 升序。
2. 命名格式：{strategy}_top5_{basename}_pH{N}.pdb，其中 {strategy} 为 interface、network 或 reverse，{basename} 为结构文件名去掉 .pdb 后的名称。interface 和 reverse 每个结构包含 _pH5 与 _pH7 两套，最多 5×2=10 个；network 每个结构仅输出 _pH7，最多 5 个；both 即 interface + network，最多 15 个。
3. 通过筛选的结构不足 5 个时，按实际数量输出。

其中 interface_screening_summary.csv 包含信息如下：

其中 network_screening_summary.csv 包含信息如下：

其中 reverse_screening_summary.csv 包含信息如下：

参考文献

• Ahn G, et al. 2025, Computational design of pH-sensitive binders, bioRxiv, DOI: https://doi.org/10.1101/2025.09.29.678932

Binder Screening (pH-dependent)

Introduction

Binder Screening (pH-dependent) is a structural screening tool for pH-sensitive binder sequences. It evaluates input target-binder complex structures using pH_score, ddG_elec, hydrogen bond energy, and HIS-charged residue pair analysis.

Core Concept

The module performs screening using the following metrics:

• pH Score: Evaluates interface pH sensitivity based on histidine hydrogen bond patterns.
• ddG Elec: Uses OpenMM to calculate electrostatic binding energy changes across pH conditions.
• Hydrogen Bond Energy: Analyzes the stability of HIS-HIS, HIS-LYS, and HIS-ARG hydrogen bond networks within the binder.
• HIS-Charged Residue Pairs: Counts spatially proximal residue pairs between HIS and charged target residues.

The tool supports PDB, CIF, and mmCIF input formats and is suitable for interface HIS screening, buried hydrogen bond network screening, and reverse design screening.

Parameters

Input Files

Input structure files. Required. Supports a single file, directory, or archive. Accepted formats include .pdb, .cif, .mmcif, as well as .zip, .tar.gz, and .tgz archives.

Binder Chains

Binder chain IDs, for example A,B. Mutually exclusive with Binder Residues.

Binder Residues

Binder residue range, for example A:1-50,B:30 or 1-50,60-80. Mutually exclusive with Binder Chains; omitting chain IDs can fail when multiple chains contain the specified residue numbers.

Target Chains

Target chain IDs, using the same format as Binder Chains. If neither Target Chains nor Target Residues is specified, all residues excluding the binder are used as the target.

Target Residues

Target residue range, using the same format as Binder Residues. If neither Target Chains nor Target Residues is specified, all residues excluding the binder are used as the target.

Strategy

Screening strategy, default interface.

• interface: Interface HIS screening using pH_score, ddG_elec filtering, and HIS-cationic residue pair counting.
• network: Buried network screening that detects HIS network hydrogen bond energy within the binder.
• reverse: Reverse interface HIS screening using ddG_elec and HIS-anionic residue pair counting, without pH_score calculation.
• both: Runs both interface and network screening.

Minimize Method

Structure minimization method, default openmm. Available for all strategies.

• openmm: Uses OpenMM built-in LocalEnergyMinimizer. It is fast but not fully reproducible.
• capped-sd: Uses custom capped steepest descent optimization with three phases: clash resolution, hydrogen and sidechain relaxation, and full atom optimization. This method is reproducible.

Interface Cutoff

Interface residue distance threshold, default 6.0 Å. Only used for interface and reverse strategies.

ddG Elec Threshold

ddG_elec threshold, default 0.0 kcal/mol. Only used for interface and reverse strategies.
For the interface strategy, ddG_elec >= threshold passes, indicating enhanced electrostatic repulsion at low pH.
For the reverse strategy, ddG_elec <= threshold passes, indicating enhanced electrostatic attraction at low pH.

Skip ddG Elec

Whether to skip ddG_elec calculation. Only used for the interface strategy. The reverse strategy does not support skipping ddG_elec because it is the core screening metric.

Min HIS Count

Minimum number of HIS-charged residue pairs, default 1. Only used for interface and reverse strategies.
For the interface strategy, HIS residues near target cationic residues K, R, and H are counted.
For the reverse strategy, HIS residues near target anionic residues D and E are counted.

Buried SASA Ratio

Buried residue SASA ratio threshold, default 0.2. A residue is considered buried when its SASA ratio, calculated as actual SASA / standard amino acid SASA, is below this threshold. Only used for the network strategy.

Network HBond Distance

Network hydrogen bond distance cutoff, default 3.5 Å. Only used for the network strategy. A hydrogen bond is recorded when the distance between the HIS acceptor atom ND1/NE2 and the partner donor atom is below this cutoff. Donor atoms include ARG NH1/NH2/NE, LYS NZ, or HIS ND1/NE2.

HBond Energy Threshold

Hydrogen bond energy threshold, default -0.5 kcal/mol. Only used for the network strategy. A structure passes when the strongest hydrogen bond energy, with HIS as acceptor and HIS/ARG/LYS as donor within the binder, is below this threshold.

Results

The output includes the following files:

Details for {strategy}_top5_{basename}_pH{N}.pdb are as follows:

1. Top 5 rule: the top 5 passed PDB structures are selected for each strategy according to the corresponding sorting rule. interface is sorted by ddG_elec in descending order, or by pH_score in descending order when Skip ddG Elec is enabled; reverse is sorted by ddG_elec in ascending order; network is sorted by best_hbond_energy in ascending order.
2. Naming format: {strategy}_top5_{basename}_pH{N}.pdb, where {strategy} is interface, network, or reverse, and {basename} is the input structure file name without .pdb. For interface and reverse, each structure includes both _pH5 and _pH7 versions, up to 5×2=10 files. For network, each structure only includes _pH7, up to 5 files. For both, meaning interface + network, up to 15 files are generated.
3. If fewer than 5 structures pass screening, the actual number of passed structures is exported.

The interface_screening_summary.csv file includes the following columns:

The network_screening_summary.csv file includes the following columns:

The reverse_screening_summary.csv file includes the following columns:

References

• Ahn G, et al. 2025, Computational design of pH-sensitive binders, bioRxiv, DOI: https://doi.org/10.1101/2025.09.29.678932

Binder Generation (pH-dependent)

简介

Binder Generation (pH-dependent) 是一个基于 ProteinMPNN 的 pH 敏感结合物序列生成工具。该工具通过在 binder 界面位置引入组氨酸（HIS）偏向，或在 binder 核心区域放置 HIS-HIS、HIS-ARG、HIS-LYS 网络突变，生成具有 pH 响应性的结合物序列。支持 PDB、CIF 和 mmCIF 格式输入。

核心思想
该模块主要支持以下设计策略：

• Interface HIS Strategy：在 target 正电残基 K/R/H 附近的 binder 位置施加 HIS 偏向，使低 pH 下静电排斥增强。
• Buried Network Strategy：在 binder 核心区域引入 HIS-阳离子残基网络，通过核心氢键实现 pH 敏感性。
• Reverse Design：在 target 负电残基 D/E 附近引入 HIS，使低 pH 下静电吸引增强。

该流程以结构输入、策略位点识别、ProteinMPNN 序列生成、HIS 数量过滤和 top-n 序列汇总为核心，输出候选 pH 敏感 binder 序列。

参数说明

Input Structure

输入结构文件，必选参数。支持单个文件、目录或压缩包输入，可接受 .pdb、.cif、.mmcif 格式，以及 .zip、.tar.gz、.tgz 压缩包。

Binder Chains

Binder 链 ID，例如 A,B。与 Binder Residues 二选一。

Binder Residues

Binder 残基范围，例如 A:1-50,B:30 或 1-50,60-80。与 Binder Chains 二选一；未指定链名且多个链含相同残基编号时会报错。

Target Chains

Target 链 ID，格式与 Binder Chains 相同。未指定 Target Chains 和 Target Residues 时，默认使用除 binder 外的所有残基作为 target。

Target Residues

Target 残基范围，格式与 Binder Residues 相同。未指定 Target Chains 和 Target Residues 时，默认使用除 binder 外的所有残基作为 target。

Strategy

设计策略，默认值为 interface。

• interface：界面 HIS 策略，在 target 正电残基附近的 binder 位置施加 HIS 偏向。
• network：埋藏网络策略，在 binder 核心引入 HIS-阳离子残基网络，周围邻域无偏向重设计。
• reverse：反向设计策略，与 interface 类似，但靶向 target 负电残基 D/E。
• both：同时运行 interface 和 network 策略。

Network Type

network 策略采用的网络类型，默认值为 all。仅用于 network 策略。

• his-his：仅使用 HIS-HIS 网络。
• his-arg：仅使用 HIS-ARG 网络，包含两个方向。
• his-lys：仅使用 HIS-LYS 网络，包含两个方向。
• all：尝试所有网络类型。

Bias Levels

HIS 偏向强度列表，使用逗号分隔，默认值为 2.0,3.0,5.0。仅用于 interface 和 reverse 策略。

Charged Cutoff

Target 带电残基附近 binder 残基的距离阈值，默认值为 6.0 Å。仅用于 interface 和 reverse 策略。

Interface Cutoff

界面残基距离阈值，默认值为 6.0 Å。仅用于 interface 和 reverse 策略。

Buried SASA Ratio

埋藏残基 SASA 比值阈值，默认值为 0.2。当残基 SASA 比值小于该阈值时，该残基视为埋藏残基。仅用于 network 策略。

Network Distance

埋藏网络配对的最大 CA-CA 距离，默认值为 6.0 Å。仅用于 network 策略。

Local Redesign Radius

局部重设计邻域半径，默认值为 6.0 Å。表示埋藏网络位置周围参与重设计的残基范围。仅用于 network 策略。

Omit AAs

屏蔽的氨基酸列表，例如 CS 表示屏蔽 Cys 和 Ser。生成序列中不会包含指定屏蔽的氨基酸。默认不屏蔽，所有策略均可使用。

Min HIS

HIS 数量下限，默认值为 1。

Max HIS

HIS 数量上限，默认使用渐进式动态阈值，直到覆盖 top-n 序列。

Num Sequences

生成序列数量。对于 interface 和 reverse 策略，表示每个设计在每个 bias 水平下生成的序列数，默认值为 100。对于 network 策略，表示每个设计中每对埋藏网络、每种网络类型生成的序列数，默认值为 100。

Top N

跨设计合并去重后最终输出的序列数，默认值为 100。所有策略均可使用。

结果说明

输出结果包括：

其中 generation_summary_all.csv 和 generation_summary.csv 包含信息如下：

当设置 Min HIS 或 Max HIS 时，除输出 generation_summary.csv、all_top_sequences.fasta 和 all_top_binder_sequences.fasta 外，还会额外输出一批默认的补充结果：

补充结果是程序根据经验默认生成的结果，可作为正式计算结果的补充。

参考文献

• Dauparas J, Anishchenko I, Bennett N, et al. 2022, Robust deep learning based protein sequence design using ProteinMPNN, Science, 378:6615, DOI: https://doi.org/10.1126/science.add2187
• Ahn G, et al. 2025, Computational design of pH-sensitive binders, bioRxiv, DOI: https://doi.org/10.1101/2025.09.29.678932

Binder Generation (pH-dependent)

Introduction

Binder Generation (pH-dependent) is a ProteinMPNN-based sequence generation tool for designing pH-sensitive binders. It generates pH-responsive binder sequences by introducing histidine (HIS) bias at binder interface positions or by placing HIS-HIS, HIS-ARG, and HIS-LYS network mutations in the binder core. It supports PDB, CIF, and mmCIF input formats.

Core Concept

The module supports the following design strategies:

• Interface HIS Strategy: Applies HIS bias to binder positions near positively charged target residues K/R/H, enhancing electrostatic repulsion at low pH.
• Buried Network Strategy: Introduces HIS-cationic residue networks in the binder core to create pH sensitivity through core hydrogen bonds.
• Reverse Design: Introduces HIS near negatively charged target residues D/E, enhancing electrostatic attraction at low pH.

The workflow centers on structure input, strategy-specific site identification, ProteinMPNN sequence generation, HIS count filtering, and top-n sequence summarization.

Parameters

Input Structure

Input structure file. Required. Supports a single file, directory, or archive. Accepted formats include .pdb, .cif, .mmcif, as well as .zip, .tar.gz, and .tgz archives.

Binder Chains

Binder chain IDs, for example A,B. Mutually exclusive with Binder Residues.

Binder Residues

Binder residue range, for example A:1-50,B:30 or 1-50,60-80. Mutually exclusive with Binder Chains; omitting chain IDs can fail when multiple chains contain the specified residue numbers.

Target Chains

Target chain IDs, using the same format as Binder Chains. If neither Target Chains nor Target Residues is specified, all residues excluding the binder are used as the target.

Target Residues

Target residue range, using the same format as Binder Residues. If neither Target Chains nor Target Residues is specified, all residues excluding the binder are used as the target.

Strategy

Design strategy, default interface.

• interface: Interface HIS strategy that applies HIS bias to binder positions near positively charged target residues.
• network: Buried network strategy that introduces HIS-cationic residue networks in the binder core and redesigns the surrounding neighborhood without bias.
• reverse: Reverse design strategy similar to interface, but targeting negatively charged target residues D/E.
• both: Runs both interface and network strategies.

Network Type

Network type used by the network strategy, default all. Only used for the network strategy.

• his-his: HIS-HIS networks only.
• his-arg: HIS-ARG networks only, in both directions.
• his-lys: HIS-LYS networks only, in both directions.
• all: Attempts all network types.

Bias Levels

Comma-separated list of HIS bias strengths, default 2.0,3.0,5.0. Only used for interface and reverse strategies.

Charged Cutoff

Distance cutoff for binder residues near charged target residues, default 6.0 Å. Only used for interface and reverse strategies.

Interface Cutoff

Interface residue distance cutoff, default 6.0 Å. Only used for interface and reverse strategies.

Buried SASA Ratio

Buried residue SASA ratio threshold, default 0.2. A residue is considered buried when its SASA ratio is below this threshold. Only used for the network strategy.

Network Distance

Maximum CA-CA distance for buried network pairs, default 6.0 Å. Only used for the network strategy.

Local Redesign Radius

Local redesign neighborhood radius, default 6.0 Å. It defines the range of residues around buried network positions that participate in redesign. Only used for the network strategy.

Omit AAs

Amino acids to omit. For example, CS omits Cys and Ser. Generated sequences will not contain the specified omitted amino acids. By default, no amino acids are omitted. Available for all strategies.

Min HIS

Minimum number of HIS residues, default 1.

Max HIS

Maximum number of HIS residues. By default, a progressive dynamic threshold is used until the top-n sequences are covered.

Num Sequences

Number of sequences to generate. For interface and reverse strategies, this is the number of sequences generated for each design at each bias level, default 100. For the network strategy, this is the number of sequences generated for each buried network pair and each network type in each design, default 100.

Top N

Final number of sequences to output after merging and deduplication across designs, default 100. Available for all strategies.

Results

The output includes the following files:

The generation_summary_all.csv and generation_summary.csv files include the following columns:

When Min HIS or Max HIS is set, in addition to generation_summary.csv, all_top_sequences.fasta, and all_top_binder_sequences.fasta, the following supplementary outputs are also generated by default:

The supplementary results are generated by the program using default empirical rules and can be used as additional references for the main calculation results.

References

• Dauparas J, Anishchenko I, Bennett N, et al. 2022, Robust deep learning based protein sequence design using ProteinMPNN, Science, 378:6615, DOI: https://doi.org/10.1126/science.add2187
• Ahn G, et al. 2025, Computational design of pH-sensitive binders, bioRxiv, DOI: https://doi.org/10.1101/2025.09.29.678932

Humanization Report v2.5

简介

Humanization Report v2.5是抗体人源化设计报告生成模块，用于生成最终的抗体人源化设计报告以及相应的专利实施例段落。

更新日志：

• 同时支持人源化和高通量人源化流程

参数说明

方法1：Humanization Report

Graft Policy

Grafting模块生成的Graft Policy文件。

Mutate Policy

Back Mutation Grouping模块生成的Policy文件。

Antibody Type

抗体类型，Antibody 标准双链抗体，Nanobody 纳米抗体。

Germline Score File

Grafting模块生成的score文件，JSON格式

Mutation Score File

Mutation模块生成的score文件，CSV格式

方法2：Humanization HTS Report

Graft Policy

Grafting模块生成的Graft Policy文件。

Mutate Policy

Back Mutation Grouping模块生成的Policy文件。

Antibody Type

抗体类型，Antibody 标准双链抗体，Nanobody 纳米抗体。

Germline Score File

Grafting模块生成的score文件，JSON格式

Mutation Score File

Mutation模块生成的score文件，CSV格式

Antibody RMSD File

抗体结构RMSD文件，由Antibody RMSD模块生成，CSV格式

Antibody RMSD Top

从RMSD排序中取前N个RMSD值小的抗体

Folding Stability File

Absolute Folding Stability模块预测生成的蛋白稳定性文件，CSV格式

结果说明

输出结果包括：

其中batch_registration_template.xlsx包含如下信息：

其中hotspot_summary.xlsx包含如下信息：

Humanization Report v2.5

Introduction

The Humanization Report v2.5 is a module for generating reports on antibody humanization design, including the final antibody humanization design report and corresponding paragraphs for patent implementation examples.

Update Log：

• support both humanization and high-throughput humanization.

Parameters

Method 1: Humanization Report

Graft Policy

The Graft Policy file generated by the Grafting module.

Mutate Policy

The Policy file generated by the Back Mutation Grouping module.

Antibody Type

Antibody type, Antibody or Nanobody

Germline Score File

Graft germline score file in JSON format generated by the Grafting module

Mutation Score File

Mutation score file in csv format generated by the Mutation module

Method 2: Humanization HTS Report

Graft Policy

The Graft Policy file generated by the Grafting module.

Mutate Policy

The Policy file generated by the Back Mutation Grouping module.

Antibody Type

Antibody type, Antibody or Nanobody

Germline Score File

Graft germline score file in JSON format generated by the Grafting module

Mutation Score File

Mutation score file in csv format generated by the Mutation module

Antibody RMSD File

Antibody structure RMSD file generated by Antibody RMSD module

Antibody RMSD Top

Select the top N antibodies with the smallest RMSD values from the RMSD ranking

Folding Stability File

Protein folding stability file generated by Absolute Folding Stability module in CSV format

Results

The output results include:

The batch_registration_template.xlsx file contains the following information:

The hotspot_summary.xlsx file contains the following information:

Mutation Score

简介

Mutation Score是抗体人源化设计中核心模块，是一个基于结构的自动化评分模块。该模块基于抗体的结构信息以及CDR嫁接后的序列信息，对移植抗体（graft）后的FR区的每个氨基酸在替换前后的变化程度进行量化评分。评分越高，说明CDR嫁接过程中氨基酸的替换对CDR区的构象可能影响较大，需要进行回复突变。模块输出每个氨基酸的打分值，用于抗体人源化设计流程中后续的分组和人源化抗体序列的生成。

参数说明

Sequence File

抗体Fv区序列文件，FASTA格式。

Model File

抗体结构文件，PDB格式。

Grafted Sequence

抗体CDR区Graft后的序列文件，FASTA格式。

Germline Hits

抗体FR区序列比对结果文件，FASTA格式

Interface Score

界面相互作用计算结果，包含原子/残基级别的接触信息

Hbond Score

氢键计算结果，包含供体-受体对、距离和角度信息

Output Score

指定输出打分文件的名称，CSV格式。

Antibody Type

抗体类型：

• Antibody：常规抗体
• Nanobody：纳米抗体

Numbering Type

抗体编号规则：kabat，imgt，chothia，martin，aho，ccg

结果说明

输出结果文件为score.csv，包含信息如下：

Mutation Score

Introduction

Mutation Score is a core module in antibody humanization design, serving as a structure-based automated scoring module. This module quantitatively scores the degree of change for each amino acid in the FR region after grafting CDRs based on the antibody’s structure and the sequence information post-CDR grafting. A higher score indicates that the replacement of amino acids during CDR grafting may have a significant impact on the CDR region’s conformation, suggesting the need for revertant mutations. The module outputs a score for each amino acid, which is used in subsequent grouping and generation of humanized antibody sequences in the antibody humanization design process.

Parameters

Sequence File

Sequence file of the antibody Fv region in FASTA format.

Model File

Antibody structure file in PDB format.

Grafted Sequence

Sequence file of the antibody CDR region after grafting in FASTA format.

Output Score

Specify the name of the output scoring file in CSV format.

Interface Score

Interface interaction calculation results, including atom/residue-level contact information

Hbond Score

Interface interaction calculation results, including atom/residue-level contact information

Antibody Type

Type of antibody:

• Antibody: Conventional antibody
• Nanobody: Nanobody

Numbering Type

Antibody numbering type: kabat，imgt，chothia,martin,aho,ccg

Results

The output result file is named score.csv and includes the following information:

Grafting v2.5

简介

Grafting模块是移植抗体的CDR到特定的框架区模板上，通常用于人源化设计。版本：v2.5

参数说明

Antibody Sequence File

抗体序列文件，FASTA格式

Numbering Type

抗体编号规则：kabat，imgt，chothia，martin，aho，ccg

Output File

指定输出抗体graft后的序列文件名称，FASTA格式

Output Policy

指定输出graft策略文件，JSON格式

Germline Score

指定输出抗体FR区序列比对同源性打分文件

Species

待移植模板的物种：human,cat,dog,默认为human

Germline

指定轻链或重链使用特定germline模板，也可都指定，写法如下：

seq_name1:germline_name1,seq_name2:germline_name2

其中链名来自于流程第一步输入的fasta文件。
例1：以下语句为链"Infliximab.H"指定了模板"IGHV3-7*01"：

Infliximab.H:IGHV3-7*01

例2：以下语句为两条链分别指定了模板：

Infliximab.H:IGHV3-7*01,Infliximab.L:IGKV1-39*01

Template V Sequence

指定抗体可变区 V 基因 的参考模板序列，FASTA格式。

Template J Sequence

指定抗体可变区 J 基因 的参考模板序列，FASTA格式。

Germline Hits

指定输出FR区序列比对结果文件，FASTA格式

Number of Hits

指定输出命中序列的数目

结果说明

输出结果包括：

Grafting v2.5

Introduction

The Grafting module is used to graft the CDRs of an antibody onto a specific framework region template, typically used for humanization design. Version: v2.5

Parameters

Antibody Sequence File

Antibody sequence file in FASTA format.

Numbering Type

Antibody numbering type: kabat，imgt，chothia,martin,aho,ccg

Output File

Specify the output file name for the grafted antibody sequence in FASTA format.

Output Policy

Specify the output grafting strategy file in JSON format.

Germline Score

Species of the template to be transplanted: human, cat, dog; default: human.

Species

Species of the sequence: human, cat, dog; default is human.

Germline

Specify the specific germline template to be used for the light chain or heavy chain, or both, in the format:

seq_name1:germline_name1,seq_name2:germline_name2

Where the chain names come from the FASTA file input in the first step of the process.
Example 1: The following statement specifies the template “IGHV3-7*01” for the chain “Infliximab.H”:

Infliximab.H:IGHV3-7*01

Example 2: The following statement specifies templates for two chains separately:

Infliximab.H:IGHV3-7*01,Infliximab.L:IGKV1-39*01

Template V Sequence

Specify the reference template sequence of the antibody V gene in FASTA format.

Template J Sequence

Specify the reference template sequence of the antibody J gene in FASTA format.

Germline Hits

Specify the output file for the FR region sequence alignment results in FASTA format.

Number of Hits

Specify the number of sequences to output.

Results

The output includes:

Back Mutation Grouping v2.6

简介

该模块是抗体人源化设计流程中分组模块，根据Mutation Score模块输出的回复突变评分表对回复突变进行分组。

更新内容：

• 模块同时支持人源化和高通量人源化流程。

参数说明

方法1：Mutate

Grafted Chain

抗体CDR区嫁接后序列文件，FASTA格式，由Grafting模块生成

Raw Chain

抗体序列文件，FASTA格式

Mutation Score

人源化突变评分文件，CSV格式，由Mutation Score模块生成

Output File

指定输出的突变序列文件名称，FASTA格式

Cutoff

打分分组的截断值，逗号分割，例如：2,5,10表示将氨基酸突变评分大于10的为一组，5~10的氨基酸为一组，小于2的氨基酸分为一组。

Output Policy

指定输出的回复突变的文件

Type

普通抗体Antibody或者纳米抗体Nanobody

方法2：HTS Mutate

Grafted Chain

抗体CDR区嫁接后序列文件，FASTA格式，由Grafting模块生成

Raw Chain

抗体序列文件，FASTA格式

Mutation Score

人源化突变评分文件，CSV格式，由Mutation Score模块生成

Output File

指定输出的突变序列文件名称，FASTA格式

Cutoff

打分分组的截断值，逗号分割，例如：2,5,10表示将氨基酸突变评分大于10的为一组，5~10的氨基酸为一组，小于2的氨基酸分为一组。

Output Policy

指定输出的回复突变的文件

Type

普通抗体Antibody或者纳米抗体Nanobody

Combination Min Cutoff

突变组合的截断值，Mutation Score模块中输出的氨基酸回复突变打分大于截断值的氨基酸参与生成突变组合

Combination Max Cutoff

高于截断值的突变自动进行回复突变

Combination Site Cutoff

每条链回复突变打分在Combination Min Cutoff与Combination Max Cutoff之间的选择打分前n个位置进行组合突变

结果说明

根据不同截断值得到突变分组结果文件mutate_policy.json。

高通量方法HTS Mutate中根据组合突变截断值得到的突变分组结果文件combination_mutate_policy.json，高通量人源化设计流程。

Back Mutation Grouping v2.6

Introduction

Back mutation grouping in the antibody humanization, which groups the back mutations based on the mutation scoring table generated by the Mutation Score module.

Update Log：

• support both humanization and high-throughput humanization.

Parameters

Method 1: Mutate

Grafted Chain

Sequence file of the antibody CDR region after grafting, in FASTA format, generated by the Grafting module.

Raw Chain

Sequence file of the antibody, in FASTA format.

Mutation Score

Humanization mutation score file, in CSV format, generated by the Mutation Score module.

Output File

Specify the name of the output mutation sequence file, in FASTA format.

Cutoff

Cutoff values for scoring grouping, separated by commas. For example, 2,5,10 indicates grouping amino acid mutations with scores greater than 10 in one group, amino acids with scores between 5 and 10 in another group, and amino acids with scores less than 2 in a separate group.

Output Policy

Specify the file for the output of back mutations.

Type

Antibody or Nanobody

Method 2: HTS Mutate

Grafted Chain

Sequence file of the antibody CDR region after grafting, in FASTA format, generated by the Grafting module.

Raw Chain

Sequence file of the antibody, in FASTA format.

Mutation Score

Humanization mutation score file, in CSV format, generated by the Mutation Score module.

Output File

Specify the name of the output mutation sequence file, in FASTA format.

Cutoff

Cutoff values for scoring grouping, separated by commas. For example, 2,5,10 indicates grouping amino acid mutations with scores greater than 10 in one group, amino acids with scores between 5 and 10 in another group, and amino acids with scores less than 2 in a separate group.

Output Policy

Specify the file for the output of back mutations.

Type

Antibody or Nanobody

Combination Min Cutoff

Cutoff value for mutation combinations. Amino acids with scores (generated from Mutation Score module) greater than the cutoff value are involved in the mutation combinations.

Combination Max Cutoff

Mutations above the cutoff value automatically undergo reversion mutations.

Combination Site Cutoff

For each chain, select the top n positions with back mutation scores between the Combination Min Cutoff and the Combination Max Cutoff for combination mutations.

Results

The mutation grouping results file mutate_policy.json is generated based on different cutoff values.
In HTS Mutate, the mutation grouping results file combination_mutate_policy.json is generated based on combination cutoff values.

Structure Prediction (ESMFold2)

简介

ESMFold2生物分子结构预测模型，定义了速度和准确性的新前沿，基于底层蛋白语言模型 ESMC 的预训练表示构建。是首个在单序列输入（不依赖多序列比对 MSA）条件下，性能超越 AlphaFold3 的蛋白质结构预测模型，并在两周内预测了 11亿个蛋白质的原子级结构。
ESMFold2 并非孤立存在，其能力完全来源于 ESMC（EvolutionaryScale Model of Chemistry）这一蛋白质生物学"世界模型"：

相比 ESM2，ESMC 在模型规模和长距离结构理解能力上进一步提升，为 ESMFold2 提供了更强的序列表征基础。

ESMFold2 的三大核心技术创新：

1. 稳定递归更新机制（Stable Recurrent Updates）
   解决的问题：传统折叠网络在多次循环迭代时会出现特征爆炸或梯度不稳定。
   技术方案：通过 Zero-Order-Hold 离散化的负对角矩阵约束，确保折叠层可以任意次数循环而不出现数值不稳定。
   意义：实现了推理时的"算力换精度"灵活权衡。
2. 截断扩散采样（Truncated Diffusion Sampling）
   解决的问题：标准扩散模型需要从纯噪声开始，经过约200步去噪，前期高噪声阶段的计算对最终精度贡献极低。
   技术方案：跳过早期高噪声阶段的无效计算，将扩散步数从 200步减少到68步。
   效果：速度提升 3倍，且精度无损失。
3. 滑动窗口原子注意力（Sliding Window Atom Attention）
   解决的问题：原子级扩散模型中，传统成对偏置注意力的计算复杂度为 O(N²)（N为原子数），长序列计算不可行。
   技术方案：用 FlashAttention 原生支持的滑动窗口注意力替代传统全局对偏置注意力。
   效果：计算复杂度从 O(N²) 降低到 O(N)，使得长蛋白质和大复合物的原子级建模成为可能。

FoldBench 基准测试对比：

仅凭单条序列，ESMFold2 就在最难的抗体-抗原任务上超越了依赖多序列比对的 AlphaFold3。绝大多数宏基因组蛋白质没有已知同源序列，无法构建 MSA，单序列预测能力才是打开未知蛋白质宇宙的钥匙。

参数说明

Single Mode

Protein Sequence

蛋白的序列文件，FASTA格式，支持多条序列。
注意：多蛋白复合物结构预测，其氨基酸序列输入格式如下：

>protein1
AASJ...
>protein2
AASJ...
>peptide
ASDF...

DNA Sequence

DNA核酸的序列文件，FASTA格式，支持多条序列。

RNA Sequence

RNA核酸分子的序列文件，FASTA格式，支持多条序列。

Ligand

包含小分子信息的 TXT 格式文本文件。支持 SMILES 或 CCD Code（Chemical Component Dictionary 编号）格式。如果使用 SMILES 格式，每行应包含一个小分子；如果使用 CCD Code，每行可以包含一个或多个小分子，多个小分子用英文逗号分隔，并以 CCD 作为前缀。示例如下：

CC(=O)OC1C[NH+]2CCC1CC2
CCD,ATP,HY3,P1L
CCD,MG

注意：该参数不适用于以氨基酸序列格式输入的配体蛋白或多肽。

Modification

包含翻译后修饰（PTM）信息的 TXT 格式文本文件。每一行包含一条 PTM 信息，由三部分组成：

• 发生 PTM 的序列顺序编号
• PTM 类型对应的 CCD 编号
• 发生 PTM 的残基位置编号

三部分之间使用英文逗号分隔。例如，1,HY3,1 表示第一条序列的第一个残基发生 HY3 类型的 PTM。HY3 是 CCD 编号，表示 3-hydroxyproline，即脯氨酸羟基化修饰。

注意：

• 序列顺序编号按照 Protein、DNA 和 RNA 参数中序列的顺序和数量依次编号，从 1 开始。例如，如果有 2 条蛋白质序列、1 条 DNA 序列和 1 条 RNA 序列，则编号为：第一条蛋白质序列编号为 1，第二条蛋白质序列编号为 2，DNA 序列编号为 3，RNA 序列编号为 4。
• CCD 介绍可参考 https://www.wwpdb.org/data/ccd。CCD 编号查询网站为 https://www.ebi.ac.uk/pdbe-srv/pdbechem/。

1,HY3,1
1,P1L,5
2,HY3,3

Covalent Bond

包含共价键信息的 TXT 格式文本文件。每一行包含一条共价键信息。每条共价键信息包含两个原子信息项，每个原子信息项由三部分组成：

• 序列或小分子的顺序编号。该编号遵循 Modification 中定义的序列编号规则，小分子编号追加在序列编号之后。
• 原子所在残基的位置编号。如果该残基是小分子，则编号为 1。
• 原子的标准名称，按照 CCD 中的定义填写。

三部分之间使用英文逗号分隔。例如，3,1,CA 表示第三个实体（序列或小分子）中第一个残基（或小分子）的 CA 原子。

一条共价键由两个原子信息项组成，两个原子信息项之间使用英文分号分隔。例如，1,1,CA;2,1,CA 表示由两个原子构成的共价键：第一个原子是 1,1,CA，第二个原子是 2,1,CA。

包含多条共价键信息的文件示例如下：

1,1,CA;2,1,CA
1,1,CA;3,1,CHA

MSA

包含蛋白质多序列比对（Multiple Sequence Alignment, MSA）信息的文件。该参数为默认:False，仅适用于 Protein Sequence 中的蛋白质序列，不适用于 DNA、RNA 或小分子配体。

MSA 可以为模型提供同源序列中的进化信息，例如保守位点、共变关系和插入缺失模式，从而帮助模型更好地判断残基之间的空间关系。对于蛋白复合物、远缘同源蛋白或结构较复杂的目标，输入 MSA 可能有助于提升预测准确性。如果不提供 MSA，模型会使用单序列模式进行预测。

Fast Mode

控制结构预测时是否使用更快的 EsmFold2-Fast 模型推理。默认值为 False，表示默认使用标准 EsmFold2 模型进行结构预测。Fast模型适合快速的批量结构预测，但是预测的精度会相对标准模型有所下降。
注意： 不支持MSA查询

Seed

用于控制预测过程随机性的随机种子。

Output Score

结构评分结果文件名，默认值为 pred_scores_esmfold2.csv。

Format

输出结构的格式，支持PDB或CIF格式，默认为CIF格式。

Batch Mode

动态计算量

≤ 5 条序列：500 计算量 / 条
第 6–100 条序列：300 计算量 / 条
超过 100 条的部分：100 计算量 / 条

Protein Sequence

蛋白质序列文件，FASTA 格式，支持多条序列。每条记录表示一个需要预测的结构，并且每条记录名称必须唯一。如果某条记录包含多条链，则不同链之间应使用英文冒号 : 连接。示例内容：

>1
EVQLVESGGGVVQPGRSLRLS:NFMLTQPHSVSVL:ELFPQWHLPIKIAAIIAS
>2
YYCAKDRRDMGYFQHWGQGTLVTVSSQWYQQRPGSAPTTVIYEDNQRPSGTPPTFMIAVFLPIVVLIFKSILFL

这表示需要预测两个结构。第一条记录名称为 1，包含三条由英文冒号分隔的蛋白质链。第二条记录名称为 2，包含一条链。

DNA Sequence

DNA 序列文件，FASTA 格式，支持多条序列。每条序列记录表示一个需要预测的结构，且记录名称必须唯一。该名称可以与 Protein 参数中的记录名称相同，表示该 DNA 序列属于同一个结构。如果某条记录包含多条链，则不同链之间应使用英文冒号 : 连接。示例内容：

>dna
GACCTCT:CCTAGCT
>1
CCTAGCT

这表示两条记录。第一条名称为 dna，包含两条由英文冒号分隔的 DNA 链。由于该名称未出现在 Protein 示例记录中，因此表示一个新的结构。第二条记录名称为 1，包含一条 DNA 链。由于该名称存在于 Protein 示例记录中，因此表示它们属于同一个结构，即该结构同时包含 Protein 和 DNA 序列。

RNA Sequence

RNA 序列文件，FASTA 格式，支持多条序列。每条序列记录表示一个需要预测的结构，且记录名称必须唯一。该名称可以与 DNA 或 Protein 参数中的记录名称相同，表示该 RNA 序列属于同一个结构。如果某条记录包含多条链，则不同链之间应使用英文冒号 : 连接。示例内容：

>1
AGCU
>rna
AGGCU:UGAUC

这表示两条记录。第一条名称为 1，是一条单链。由于该名称存在于 DNA 和 Protein 示例记录中，因此表示它们属于同一个结构，即该结构包含 Protein、DNA 和这条 RNA 序列。第二条记录名称为 rna，包含两条由英文冒号分隔的 RNA 链。由于该名称未出现在 DNA 或 Protein 示例记录中，因此表示一个新的结构。

Ligand

包含小分子信息的 TXT 格式文本文件。支持 SMILES 或 CCD Code。如果使用 SMILES 格式，每行应包含一个小分子；如果使用 CCD Code，每行可以包含一个或多个小分子，多个小分子之间用英文逗号分隔。每一行表示一个需要预测的结构，并且每一行必须以唯一名称开头。该名称可以与 RNA、DNA 或 Protein 参数中的名称相同，表示该行中的所有配体属于同一个结构。名称和所有配体之间使用英文冒号 : 分隔。示例内容：

1:CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP,HY3
lig:CC1CC(CN2NC[N+](=C3CCC(OC(F)(F)F)CC3)C2=O)CC(C)C1OC(C)(C)C(=O)[O-]

这表示两条记录。第一条名称为 1，包含三个配体，其中一个为 SMILES，两个为 CCD 代码。由于该名称存在于 RNA、DNA 或 Protein 示例记录中，因此表示它们属于同一个结构。第二条记录名称为 lig，包含一个 SMILES 格式的配体。由于该名称未出现在 RNA、DNA 或 Protein 示例记录中，因此表示一个新的结构。

注意：
1.不适用于配体蛋白或多肽的氨基酸序列格式输入。
2.在预测每个结构的序列数量相同的情况下，可为所有待预测结构设置相同的小分子信息，设置方式为输入一行小分子信息（可多个），且不设置结构名称，如CC(=O)OC1C[NH+]2CCC1CC2:CCD,ATP，表示为所有结构都加入小分子CC(=O)OC1C[NH+]2CCC1CC2与ATP。

Modification

包含翻译后修饰（PTM）信息的 TXT 格式文本文件。每个 PTM 条目的格式与 Single mode 中一致，可参考 Single mode 中的定义。在 Batch mode 中，每一行定义一个结构的所有 PTM 信息，必须以唯一名称开头。该名称必须存在于前面定义的 Protein、DNA 或 RNA 记录中，不同字段之间使用英文冒号 : 分隔。示例内容：

1:1,HY3,1:1,P1L,5:2,HY3,3
2:1,HY3,1:2,HY3,3

这表示名称为 1 的结构（Protein、DNA 或 RNA）有三个 PTM，而名称为 2 的结构有两个 PTM。
注意： 当每个待预测结构的序列数量相同时，可以输入一行不指定结构名称的 PTM 信息，将相同 PTM 信息应用到所有结构。例如：1,HY3,1:1,P1L,5:2,HY3,3 表示这些 PTM 信息会应用到所有结构。

Covalent Bond

包含共价键信息的 TXT 格式文本文件。每条共价键条目的格式与 Single mode 中一致，可参考 Single mode 中的定义。在 Batch mode 中，每一行定义一个结构的所有共价键信息，必须以唯一名称开头。该名称必须存在于前面定义的 Protein、DNA 或 RNA 记录中，不同字段之间使用英文冒号 : 分隔。示例内容：

1:1,1,CA;2,1,CA:1,1,CA;3,1,CHA
2:1,1,CA;3,1,CHA

这表示名称为 1 的结构（Protein、DNA 或 RNA）有两条共价键，而名称为 2 的结构有一条共价键。
注意：当每个待预测结构的序列数量相同时，可以输入一行不指定结构名称的共价键信息，将相同共价键信息应用到所有结构。例如：1,1,CA;3,1,CHA 表示该共价键信息会应用到所有结构。

MSA

该参数为默认:False，仅适用于 Protein Sequence 中的蛋白质序列，不适用于 DNA、RNA 或小分子配体。

MSA 可以为模型提供同源序列中的进化信息，例如保守位点、共变关系和插入缺失模式，从而帮助模型更好地判断残基之间的空间关系。对于蛋白复合物、远缘同源蛋白或结构较复杂的目标，输入 MSA 可能有助于提升预测准确性。

Fast Mode

控制结构预测时是否使用更快的 EsmFold2-Fast 模型推理。默认值为 False，表示默认使用标准 EsmFold2 模型进行结构预测。Fast模型适合快速的批量结构预测，但是预测的精度会相对标准模型有所下降。
注意： 不支持MSA查询

Seed

用于控制预测过程随机性的随机种子。

Output Score

结构评分结果文件名，默认值为 pred_scores_esmfold2.csv。

Format

输出结构的格式，支持PDB或CIF格式，默认为CIF格式。

结果说明

• Single Mode
  输出文件包括排名前 5 的复合物结构（rank_1-5.cif）、pred_scores_esmfold2.csv
pred_scores_esmfold2.csv 文件包含以下信息：

pDockQ2阈值（继承自 DockQ）：

• Batch Mode

输出final_results.tar.gz、pred_scores_esmfold2.csv
final_results.tar.gz文件为Batch模式下生成一个所有预测结果的打包文件，包含预测结构PDB文件、打分CSV文件。

参考文献

• Candido S., Hayes T., Derry A., et al. 2026, Language Modeling Materializes a World Model of Protein Biology, bioRxiv, Preprint, DOI: 10.64898/2026.06.03.729735

Structure Prediction (ESMFold2)

Introduction

ESMFold2 is a biomolecular structure prediction model that defines a new frontier in speed and accuracy. It is built on pretrained representations from the underlying protein language model ESMC. It is the first protein structure prediction model that outperforms AlphaFold3 under single-sequence input conditions, without relying on multiple sequence alignment (MSA), and predicted atomic-level structures for 1.1 billion proteins within two weeks.

ESMFold2 does not exist in isolation; its capability comes entirely from ESMC (EvolutionaryScale Model of Chemistry), a protein biology “world model”:

Compared with ESM2, ESMC further improves model scale and long-range structural understanding, providing a stronger sequence representation foundation for ESMFold2.

Three Core Technical Innovations of ESMFold2:

1. Stable Recurrent Updates

Problem addressed: Traditional folding networks may suffer from feature explosion or gradient instability during multiple recurrent iterations.

Technical approach: Negative diagonal matrix constraints discretized by Zero-Order-Hold ensure that folding layers can be recurrently updated for arbitrary numbers of cycles without numerical instability.

Significance: Enables flexible inference-time tradeoff between compute and accuracy.

2. Truncated Diffusion Sampling

Problem addressed: Standard diffusion models start from pure noise and require about 200 denoising steps, while computation in the early high-noise stage contributes very little to final accuracy.

Technical approach: Skips ineffective computation in the early high-noise stage and reduces diffusion steps from 200 to 68.

Effect: 3-fold speed improvement with no accuracy loss.

3. Sliding Window Atom Attention

Problem addressed: In atom-level diffusion models, the computational complexity of traditional pair-bias attention is O(N^2) (N is the number of atoms), making long-sequence computation infeasible.

Technical approach: Replaces traditional global pair-bias attention with FlashAttention-native sliding window attention.

Effect: Reduces computational complexity from O(N^2) to O(N), making atom-level modeling of long proteins and large complexes possible.

FoldBench Benchmark Comparison:

Using only a single sequence, ESMFold2 surpasses MSA-dependent AlphaFold3 on the most difficult antibody-antigen task. Most metagenomic proteins have no known homologous sequences and cannot form MSAs; single-sequence prediction capability is the key to opening the unknown protein universe.

Parameters

Single Mode

Protein Sequence

Protein sequence file in FASTA format, supporting multiple sequences.

Note: For multi-protein complex structure prediction, the amino acid sequence input format is as follows:

>protein1
AASJ...
>protein2
AASJ...
>peptide
ASDF...

DNA Sequence

DNA nucleic acid sequence file in FASTA format, supporting multiple sequences.

RNA Sequence

RNA nucleic acid molecule sequence file in FASTA format, supporting multiple sequences.

Ligand

TXT text file containing small molecule information. Supports SMILES or CCD Code (Chemical Component Dictionary code) format. If using SMILES, each line should contain one small molecule. If using CCD Code, each line may contain one or more small molecules, separated by English commas, and prefixed with CCD. Examples:

CC(=O)OC1C[NH+]2CCC1CC2
CCD,ATP,HY3,P1L
CCD,MG

Note: This parameter is not applicable to ligand proteins or peptides input in amino acid sequence format.

Modification

TXT text file containing post-translational modification (PTM) information. Each line contains one PTM entry composed of three parts:

• Sequence order index where PTM occurs
• CCD code corresponding to the PTM type
• Residue position index where PTM occurs

The three parts are separated by English commas. For example, 1,HY3,1 means that the first residue of the first sequence has an HY3-type PTM. HY3 is a CCD code representing 3-hydroxyproline, namely hydroxyproline modification.

Note:

• Sequence order indices are assigned according to the order and number of sequences in Protein, DNA, and RNA parameters, starting from 1. For example, if there are 2 protein sequences, 1 DNA sequence, and 1 RNA sequence, then the first protein sequence is indexed as 1, the second protein sequence as 2, the DNA sequence as 3, and the RNA sequence as 4.
• CCD information can be found at https://www.wwpdb.org/data/ccd. CCD code lookup is available at https://www.ebi.ac.uk/pdbe-srv/pdbechem/.

1,HY3,1
1,P1L,5
2,HY3,3

Covalent Bond

TXT text file containing covalent bond information. Each line contains one covalent bond entry. Each covalent bond entry contains two atom information items, and each atom information item consists of three parts:

• Order index of the sequence or small molecule. This index follows the sequence index rules defined in Modification; small molecule indices are appended after sequence indices.
• Position index of the residue containing the atom. If the residue is a small molecule, the index is 1.
• Standard atom name, filled according to the CCD definition.

The three parts are separated by English commas. For example, 3,1,CA indicates the CA atom of the first residue (or small molecule) in the third entity (sequence or small molecule).

A covalent bond consists of two atom information items, separated by an English semicolon. For example, 1,1,CA;2,1,CA represents a covalent bond composed of two atoms: the first atom is 1,1,CA, and the second atom is 2,1,CA.

An example file containing multiple covalent bonds is as follows:

1,1,CA;2,1,CA
1,1,CA;3,1,CHA

MSA

File containing protein Multiple Sequence Alignment (MSA) information. This parameter defaults to False, applies only to protein sequences in Protein Sequence, and does not apply to DNA, RNA, or small molecule ligands.

MSA can provide evolutionary information from homologous sequences, such as conserved sites, covariation relationships, and insertion/deletion patterns, helping the model better infer spatial relationships between residues. For protein complexes, remote homologous proteins, or structurally complex targets, providing MSA may help improve prediction accuracy. If MSA is not provided, the model uses single-sequence mode for prediction.

Fast Mode

Controls whether to use the faster EsmFold2-Fast model during structure prediction. The default value is False, meaning the standard EsmFold2 model is used by default. The Fast model is suitable for rapid batch structure prediction, but its accuracy may be lower than the standard model.

Note: MSA query is not supported.

Seed

Random seed used to control randomness during prediction.

Output Score

Structure score result file name, default pred_scores_esmfold2.csv.

Format

Output format for the structure, supporting PDB or CIF, default is CIF.

Batch Mode

Dynamic Tokens

≤ 5 sequences：500 compute units per sequence
6–100 sequences：300 compute units per sequence
Beyond 100 sequences：100 compute units per sequence

Protein Sequence

Protein sequence file in FASTA format, supporting multiple sequences. Each record represents one structure to predict, and each record name must be unique. If a record contains multiple chains, chains should be connected with an English colon :. Example:

>1
EVQLVESGGGVVQPGRSLRLS:NFMLTQPHSVSVL:ELFPQWHLPIKIAAIIAS
>2
YYCAKDRRDMGYFQHWGQGTLVTVSSQWYQQRPGSAPTTVIYEDNQRPSGTPPTFMIAVFLPIVVLIFKSILFL

This indicates two structures to predict. The first record is named 1 and contains three protein chains separated by English colons. The second record is named 2 and contains one chain.

DNA Sequence

DNA sequence file in FASTA format, supporting multiple sequences. Each sequence record represents one structure to predict, and the record name must be unique. The name can be the same as a record name in the Protein parameter, indicating that this DNA sequence belongs to the same structure. If a record contains multiple chains, chains should be connected with an English colon :. Example:

>dna
GACCTCT:CCTAGCT
>1
CCTAGCT

This indicates two records. The first is named dna and contains two DNA chains separated by an English colon. Since this name does not appear in the Protein example records, it represents a new structure. The second record is named 1 and contains one DNA chain. Since this name exists in the Protein example record, they belong to the same structure, meaning this structure contains both Protein and DNA sequences.

RNA Sequence

RNA sequence file in FASTA format, supporting multiple sequences. Each sequence record represents one structure to predict, and the record name must be unique. The name can be the same as a name in DNA or Protein parameters, indicating that this RNA sequence belongs to the same structure. If a record contains multiple chains, chains should be connected with an English colon :. Example:

>1
AGCU
>rna
AGGCU:UGAUC

This indicates two records. The first is named 1 and is a single chain. Since this name exists in the DNA and Protein example records, they belong to the same structure, meaning this structure contains Protein, DNA, and this RNA sequence. The second record is named rna and contains two RNA chains separated by an English colon. Since this name does not appear in DNA or Protein example records, it represents a new structure.

Ligand

TXT text file containing small molecule information. Supports SMILES or CCD Code. If using SMILES, each line should contain one small molecule. If using CCD Code, each line may contain one or more small molecules, separated by English commas. Each line represents one structure to predict and must begin with a unique name. This name can be the same as a name in RNA, DNA, or Protein parameters, indicating that all ligands in this line belong to the same structure. The name and all ligands are separated by an English colon :. Example: