Chemical Structure Search

简介

ChemicaLite 是一个基于 RDKit 的 SQLite 数据库扩展，专为化学信息学应用设计。可在化合物库中搜索目标分子，支持子结构搜索和相似性搜索两种模式。

核心特性：

• 基于 SQLite 扩展架构，支持标准 SQL 查询
• 集成 RDKit 化学信息学工具包
• 支持子结构搜索和相似性搜索
• 使用 rdtree 索引实现高性能查询

适用场景：

• 化合物数据库管理：构建和管理大规模化合物库
• 虚拟筛选：基于子结构或相似性搜索候选化合物
• 化学空间分析：计算分子描述符和指纹

参数说明

Query

查询分子文件，支持多个文件，格式为 .sdf、.smi、.smiles

Private Library

私有化合物库文件路径，与 Public Library 二选一

Search Method

搜索方法，可选值：

• sim：相似性搜索，基于 Tanimoto 系数
• sub：子结构搜索，基于 SMARTS 匹配
• 默认为sim

Threshold

相似性阈值，范围 0.0-1.0，默认为 0.7，仅在相似性搜索时有效

Hits SDF

输出 SDF 文件路径，默认为 hits.sdf

Hits Info

命中信息 CSV 文件路径，默认为 hits.csv

结果说明

输出结果包括：

其中 SDF 文件包含以下分子属性：

其中 hits.csv 包含信息如下：

Chemical Structure Search

Introduction

ChemicaLite is a SQLite database extension built on RDKit, designed for cheminformatics applications. It enables searching for target molecules within compound libraries, supporting two modes: substructure search and similarity search.

Key features:

• SQLite extension architecture with standard SQL query support
• Integrated RDKit cheminformatics toolkit
• Substructure and similarity search support
• High-performance queries via rdtree indexing

Use cases:

• Compound database management: build and manage large-scale compound libraries
• Virtual screening: search candidate compounds by substructure or similarity
• Chemical space analysis: compute molecular descriptors and fingerprints

Parameters

Query

Query molecule file(s); multiple files supported. Accepted formats: .sdf, .smi, .smiles.

Private Library

Path to a private compound library file. Mutually exclusive with Public Library.

Search Method

Search algorithm. Options:

• sim — Similarity search based on Tanimoto coefficient
• sub — Substructure search based on SMARTS matching
• Default: sim

Threshold

Similarity threshold in the range 0.0–1.0. Default: 0.7. Applies to similarity search only.

Hits SDF

Output SDF file path. Default: hits.sdf.

Hits Info

Output CSV file path for hit information. Default: hits.csv.

Results

Results consist of two files:

SDF molecule properties:

hits.csv columns:

Peptide Design (PepCraft)

简介

PepCraft 是唯信开发的从头多肽生成模型，用于面向蛋白受体热点区域设计候选结合多肽。

用户提供受体序列、目标 hotspot、多肽长度和多肽类型后，PepCraft会生成多肽候选，并使用 Boltz-2 对受体-多肽复合物进行结构预测与打分，最终输出按综合评分排序的设计结果。

当前支持三种多肽类型：

• Linear：线性多肽。
• Disulfide：首尾半胱氨酸形成二硫键约束的多肽。
• Cyclic：head-to-tail 环肽。

相比于EvoBind等多肽设计方法，PepCraft在生成的质量和多样性方面具有显著优势，同时支持线性、环肽等各种多肽类型。

注：上图中PepSeek即为PepCraft

PepCraft 的核心流程为“候选生成 - 结构验证 - 指标评分 - 迭代优化”。候选多肽可由 PepMLM、随机生成、突变和交叉等方式产生；结构验证阶段使用 Boltz-2 预测复合物，并结合整体置信度、界面质量和 hotspot 接触情况进行综合排序。

为提升运行效率，流程会在每次任务开始时仅对受体序列搜索一次 MSA，后续所有候选多肽验证时复用该受体 MSA；多肽链始终使用 single-sequence mode，不单独搜索 MSA。

参数说明

Receptor Sequence

受体蛋白序列文件。支持标准 FASTA 单行序列、标准 FASTA 多行序列，以及无 header 的纯序列输入。

流程会自动进行格式检查与标准化，包括：

• 合并多行序列；
• 将序列统一为大写；
• 检查非法氨基酸字符；
• 确认输入仅包含一条受体序列。

示例：

>1SSC_1|Chain A|RIBONUCLEASE A
KETAAAKFERQHMDSSTSAASSSNYCNQMMKSRNLTKDRCKPVNTFVHESLADVQAVCSQKNVACKNGQTNCYQSYSTMSITDCRETGSSKYPNCAYKTTQANKHIIVACEG

Hotspot

受体上的目标结合热点残基，使用 1-indexed 编号。支持单个残基、多个残基和连续区间。

示例：

15,16
20-24,31

Peptide Length

目标多肽长度。

示例：

15

Peptide Type

多肽的化学类型或结构约束，用于限定生成多肽的理化性质，可选参数。

• linear：线性多肽，无环化约束
• disulfide
• cyclic：环化多肽，首尾或侧链形成环化结构

其中 disulfide 会约束多肽首尾为半胱氨酸，并在结构预测输入中加入首尾二硫键约束；cyclic 会在结构预测输入中设置环肽约束。

结果说明

PepCraft 输出打包结果 results.zip，其中包含按综合评分排序的候选多肽信息和对应结构文件。

主要输出文件包括：

top_designs.csv 输出以下信息：

结构文件仍按排名输出为 rank_1.cif、rank_2.cif 等；rank_1.cif 对应 top_designs.csv 第一行，rank_2.cif 对应第二行，以此类推。CSV 中不再包含结构文件路径或内部来源字段。

Peptide Design (PepCraft)

Introduction

PepCraft is a peptide design framework for generating candidate binding peptides targeting hotspot regions on protein receptors. Given a receptor sequence, target hotspot residues, peptide length, and peptide type, the workflow generates peptide candidates and evaluates them using Boltz-2 structure prediction and scoring. Final peptide designs are ranked according to a composite score.

Compared with peptide design methods such as EvoBind, PepCraft boasts prominent advantages in the quality and diversity of generated peptides and supports various peptide types including linear peptides and cyclic peptides.

Note: PepSeek in the figure above refers to PepCraft

Currently, three peptide types are supported:

• Linear: Linear peptides.
• Disulfide: Peptides constrained by a disulfide bond formed between N-terminal and C-terminal cysteines.
• Cyclic: Head-to-tail cyclic peptides.

The core PepCraft workflow consists of candidate generation → structure validation → metric scoring → iterative optimization. Candidate peptides can be generated using PepMLM, random generation, mutation, and crossover operations. During structure validation, Boltz-2 is used to predict receptor–peptide complex structures, which are subsequently ranked according to overall confidence, interface quality, and hotspot-contact metrics.

To improve computational efficiency, receptor MSA is searched only once at the beginning of each task and reused throughout all subsequent peptide evaluations. Peptide chains are always modeled in single-sequence mode without independent MSA searches.

Parameters

Receptor Sequence

Input receptor protein sequence file.

The following formats are supported:

• Standard FASTA with a single-line sequence
• Standard FASTA with a multi-line sequence
• Plain sequence without a FASTA header

The workflow automatically performs format validation and normalization, including:

• Merging multi-line sequences
• Converting sequences to uppercase
• Checking for invalid amino acid characters
• Ensuring that only one receptor sequence is provided

Example:

>1SSC_1|Chain A|RIBONUCLEASE A
KETAAAKFERQHMDSSTSAASSSNYCNQMMKSRNLTKDRCKPVNTFVHESLADVQAVCSQKNVACKNGQTNCYQSYSTMSITDCRETGSSKYPNCAYKTTQANKHIIVACEG

Hotspot

Target binding hotspot residues on the receptor, using 1-indexed residue numbering.

Supports individual residues, multiple residues, and residue ranges.

Examples:

15,16
20-24,31

Peptide Length

Target peptide length.

Example:

15

Peptide Type

Chemical type or structural constraint applied to generated peptides. Optional.

Available options:

• linear: Linear peptide without cyclization constraints
• disulfide: Disulfide-constrained peptide
• cyclic: Cyclic peptide

For disulfide, PepCraft enforces cysteine residues at both peptide termini and introduces a terminal disulfide bond constraint during structure prediction.

For cyclic, cyclic peptide constraints are applied during structure prediction.

Results

PepCraft produces a compressed result package named results.zip, containing ranked peptide candidates and their corresponding structure files.

Main output files include:

top_designs.csv contains the following information:

Structure files are output as:

• rank_1.cif
• rank_2.cif
• …

where rank_1.cif corresponds to the first row of top_designs.csv, rank_2.cif corresponds to the second row, and so on.

The CSV file does not contain structure file paths or internal provenance fields.

Molecular Docking (Gnina)

简介

基于深度学习的分子对接工具，采用卷积神经网络（CNN）评分函数对配体-受体结合构象进行打分和排序。Gnina在传统对接算法基础上引入深度学习评分，显著提升了对接精度和虚拟筛选效率，支持刚性对接、柔性残基对接和共价对接等多种模式。

核心技术

• CNN 评分函数：基于 PyTorch 的深度学习评分模型，训练于大规模蛋白质-配体复合物结构数据
• 多模式对接：支持标准刚性对接、柔性残基对接和共价对接
• 自动盒子构建：可根据参考配体自动计算搜索空间，简化参数设置
• 知识蒸馏优化：GNINA 1.3 引入蒸馏模型，在保持精度的同时大幅提升筛选速度

适用场景

• 虚拟筛选：从大型化合物库中快速发现潜在活性分子
• 先导化合物优化：分析配体与靶点的结合模式，指导结构改造
• 共价药物设计：支持共价键合分子的对接计算
• 分子动力学预处理：生成合理的初始结合构象用于后续模拟

参数说明

Receptor

受体结构文件，包含对接计算中保持刚性的受体部分。

Flex

柔性受体侧链文件，指定对接过程中允许柔性的受体侧链。

Ligand

配体结构文件，支持多种分子格式。

Flexres

柔性残基列表，以逗号分隔的 chain:resid 格式指定需要柔性的残基。

Flexdist Ligand

Flexdist 模式的参考配体，用于自动识别该配体附近的柔性残基。

Flexdist

柔性化距离阈值，自动将距离 flexdist_ligand 该范围内的残基设为柔性。

Flex Limit

柔性残基数量的硬上限，限制最多允许多少个残基柔性化。

Flex Max

最多保留的最近柔性残基数量，当柔性残基超过限制时只保留距离最近的。

Center X

搜索盒子中心的 X 坐标，用于定义对接搜索空间的位置。

Center Y

搜索盒子中心的 Y 坐标。

Center Z

搜索盒子中心的 Z 坐标。

Size X

搜索盒子在 X 方向的尺寸，设置时必须为正值。

Size Y

搜索盒子在 Y 方向的尺寸，设置时必须为正值。

Size Z

搜索盒子在 Z 方向的尺寸，设置时必须为正值。

Autobox Ligand

参考配体文件，用于自动计算搜索盒子的中心和尺寸，无需手动指定 center 和 size 参数。

Autobox Add

在自动计算的搜索盒子周围添加的额外填充距离，用于扩展搜索空间。

Scoring

用于选择打分函数（scoring function），即评估配体与受体结合好坏的数学模型。

• default（CNN 深度学习）
  gnina 默认使用的打分函数，基于卷积神经网络，在训练数据覆盖的体系上精度最高，适合对结果质量要求较高的场景。

• vina（经验式）
  AutoDock Vina 原版打分函数，最经典且广泛使用，速度快、兼容性好，是虚拟筛选中的常用基准。

• vinardo（经验式）
  Vina 的改进版本，在部分体系上精度优于原版 Vina，可作为 Vina 的替代选择。

• ad4_scoring（经验式）
  AutoDock 4 的打分函数，需配合 AD4 力场参数文件使用，适合已有 AD4 工作流的场景。

• dkoes_fast（知识式）
  dkoes 系列中速度最快的版本，精度相对较低，适合需要极高吞吐量的大规模粗筛。

• dkoes_scoring（知识式）
  dkoes 系列的标准版本，在速度与精度之间取得平衡，是该系列的推荐选择。

• dkoes_scoring_old（知识式）
  dkoes_scoring 的旧版实现，一般仅用于复现早期文献或历史计算结果。

CNN Scoring

CNN 评分模式，用于选择不同的深度学习评分策略。

• none
  CNN 完全不介入，由传统打分函数独立完成全部计算，精度较低，适合超大规模粗筛场景。

• rescore（默认）
  在传统方法完成构象搜索后，由 CNN 对所有姿势进行最终重打分和重排序，精度中高，是日常虚拟筛选的推荐模式。

• refinement
  在初始姿势生成后，用 CNN 分数引导进一步局部优化，精度较高，适合中等规模的精细筛选。

• metrorescore
  引入 Metropolis 采样以 CNN 分数驱动构象搜索，最终再执行 CNN 重打分，精度较高，适合构象空间复杂或结合口袋灵活的体系。

• metrorefine
  结合 Metropolis 采样与 CNN 引导的局部优化，精度很高，适合对少量重要化合物进行精细对接评估。

• all
  CNN 参与对接的全部阶段（搜索、优化、重打分），精度最高，计算代价也最大，适合对少量化合物进行最严格的精确评估。

Num Modes

输出的最大结合模式数量，即最终保留的候选构象数，默认为10

Covalent Receptor Atom

指定蛋白质中哪个原子与配体形成共价键

A:145:SG      # A链第145位半胱氨酸的硫原子
A:200:OG      # A链第200位丝氨酸的氧原子
B:63:NZ       # B链第63位赖氨酸的氨基氮原子

Covalent Lig Atom Pattern

SMARTS 模式，用于识别配体中参与共价键的原子。

C(=O)Cl           # 酰氯，与Cys/Ser/Lys反应
C=C               # 迈克尔受体（丙烯酰胺类），与Cys反应
[CH2]Br           # 卤代烷，烷基化反应
C(=O)[F,Cl,Br]    # 通用酰卤模式
[cH]1[cH][nH]c1   # 用于特定杂环弹头

Covalent Lig Atom Position

共价配体原子的初始放置坐标。

12.345,7.890,-3.210     # 从晶体结构中读取的弹头原子坐标
-5.100,22.300,8.750     # 从同源建模结构推测的坐标

Covalent Bond Order

共价键的键级，用于共价对接计算。

1       # 单键（最常见，如 Cys-S–C 烷基化产物）
2       # 双键（如与 Lys 形成的亚胺/席夫碱）
1.5     # 芳香键（较少用）

结果说明

输出结果包括：对接的压缩文件docked.sdf.gz、解压后的小分子文件docked.sdf和打分文件docked.csv。
打分文件docked.csv各指标说明：

参考文献

• McNutt A, Li Y, Meli R, Aggarwal R, Koes D R. GNINA 1.3: the next increment in molecular docking with deep learning[J]. J. Cheminformatics, 2025, 17:43.DOI: 10.1186/s13321-025-00973-x

Molecular Docking (Gnina)

Introduction

A deep learning-based molecular docking tool that employs convolutional neural network (CNN) scoring functions to score and rank ligand–receptor binding poses. Building upon traditional docking algorithms, Gnina introduces deep learning scoring, significantly improving docking accuracy and virtual screening efficiency. It supports rigid docking, flexible residue docking, and covalent docking, among other modes.

Core Technology

• CNN Scoring Function: A PyTorch-based deep learning scoring model trained on large-scale protein–ligand complex structural data.
• Multi-mode Docking: Supports standard rigid docking, flexible residue docking, and covalent docking.
• Automatic Box Construction: Automatically calculates the search space based on a reference ligand, simplifying parameter setup.
• Knowledge Distillation Optimization: GNINA 1.3 introduces distillation models that substantially boost screening speed while maintaining accuracy.

Use Cases

• Virtual Screening: Rapidly discover potentially active molecules from large compound libraries.
• Lead Compound Optimization: Analyze ligand–target binding modes to guide structural modifications.
• Covalent Drug Design: Support docking calculations for covalently binding molecules.
• Molecular Dynamics Preprocessing: Generate reasonable initial binding poses for subsequent simulations.

Parameters

Receptor

Receptor structure file containing the rigid portion of the receptor used in the docking calculation.

Flex

Flexible receptor sidechain file specifying receptor sidechains allowed to be flexible during docking.

Ligand

Ligand structure file supporting multiple molecular formats.

Flexres

Flexible residue list specifying residues to be made flexible in chain:resid format, comma-separated.

Flexdist Ligand

Reference ligand for flexdist mode, used to automatically identify flexible residues near this ligand.

Flexdist

Flexibilization distance threshold; residues within this distance from flexdist_ligand are automatically set as flexible.

Flex Limit

Hard limit on the number of flexible residues, restricting the maximum number of residues that can be made flexible.

Flex Max

Maximum number of nearest flexible residues to retain; when the number of flexible residues exceeds the limit, only the closest ones are kept.

Center X

X coordinate of the search box center, defining the position of the docking search space.

Center Y

Y coordinate of the search box center.

Center Z

Z coordinate of the search box center.

Size X

Search box dimension in the X direction; must be set to a positive value.

Size Y

Search box dimension in the Y direction; must be set to a positive value.

Size Z

Search box dimension in the Z direction; must be set to a positive value.

Autobox Ligand

Reference ligand file used to automatically calculate the search box center and size, eliminating the need to manually specify center and size parameters.

Autobox Add

Additional padding distance added around the automatically calculated search box to expand the search space.

Scoring

Scoring function selection, i.e., the mathematical model used to evaluate ligand–receptor binding quality.

• none
  The CNN is completely uninvolved; all calculations are handled independently by the traditional scoring function. Accuracy is lower, making it suitable for ultra-large-scale coarse screening scenarios.
• rescore (default)
  After the traditional method completes conformational search, the CNN performs final rescoring and reranking of all poses. Accuracy is medium-to-high, making it the recommended mode for routine virtual screening.
• refinement
  After initial pose generation, CNN scores guide further local optimization. Accuracy is relatively high, suitable for medium-scale fine-grained screening.
• metrorescore
  Incorporates Metropolis sampling driven by CNN scores for conformational search, followed by a final CNN rescoring step. Accuracy is relatively high, suitable for systems with complex conformational spaces or flexible binding pockets.
• metrorefine
  Combines Metropolis sampling with CNN-guided local optimization. Accuracy is very high, suitable for detailed docking evaluation of a small number of important compounds.
• all
  The CNN participates in every stage of docking (search, optimization, and rescoring). Accuracy is the highest, but so is the computational cost, making it appropriate for the most rigorous and precise evaluation of a small set of compounds.

CNN Scoring

CNN scoring mode, used to select different deep learning scoring strategies.

• none
  The CNN is completely uninvolved; all calculations are handled independently by the traditional scoring function. Accuracy is lower, making it suitable for ultra-large-scale coarse screening scenarios.
• rescore (default)
  After the traditional method completes conformational search, the CNN performs final rescoring and reranking of all poses. Accuracy is medium-to-high, making it the recommended mode for routine virtual screening.
• refinement
  After initial pose generation, CNN scores guide further local optimization. Accuracy is relatively high, suitable for medium-scale fine-grained screening.
• metrorescore
  Incorporates Metropolis sampling driven by CNN scores for conformational search, followed by a final CNN rescoring step. Accuracy is relatively high, suitable for systems with complex conformational spaces or flexible binding pockets.
• metrorefine
  Combines Metropolis sampling with CNN-guided local optimization. Accuracy is very high, suitable for detailed docking evaluation of a small number of important compounds.
• all
  The CNN participates in every stage of docking (search, optimization, and rescoring). Accuracy is the highest, but so is the computational cost, making it appropriate for the most rigorous and precise evaluation of a small set of compounds.

Num Modes

Maximum number of binding modes to output, i.e., the final number of candidate poses retained. Default: 10.

Covalent Receptor Atom

Specifies which atom in the protein forms a covalent bond with the ligand.

A:145:SG      # Sulfur atom of Cysteine 145 on chain A
A:200:OG      # Oxygen atom of Serine 200 on chain A
B:63:NZ       # Amino nitrogen atom of Lysine 63 on chain B

Covalent Lig Atom Pattern

SMARTS pattern used to identify the atom in the ligand that participates in the covalent bond.

C(=O)Cl           # Acyl chloride; reacts with Cys/Ser/Lys
C=C               # Michael acceptor (acrylamide-like); reacts with Cys
[CH2]Br           # Haloalkane; alkylation reaction
C(=O)[F,Cl,Br]    # General acyl halide pattern
[cH]1[cH][nH]c1   # For specific heterocyclic warheads

Covalent Lig Atom Position

Initial placement coordinates of the covalent ligand atom.

12.345,7.890,-3.210     # Warhead atom coordinates read from a crystal structure
-5.100,22.300,8.750     # Coordinates inferred from a homology model

Covalent Bond Order

Bond order of the covalent bond, used in covalent docking calculations.

1       # Single bond (most common, e.g., Cys-S–C alkylation product)
2       # Double bond (e.g., imine/Schiff base formed with Lys)
1.5     # Aromatic bond (rarely used)

Results

The output includes a compressed docking file docked.sdf.gz, the extracted small molecule file docked.sdf, and a scoring file docked.csv.

Column descriptions for the scoring file docked.csv:

Reference

• McNutt A, Li Y, Meli R, Aggarwal R, Koes D R. GNINA 1.3: the next increment in molecular docking with deep learning[J]. J. Cheminformatics, 2025, 17:43.DOI: 10.1186/s13321-025-00973-x

Structure Minimization

简介

Structure Minimization 用于在 GB 隐式溶剂下对蛋白质/核酸/小分子/复合物结构进行能量最小化，在指定突变的情况下也支持对突变体进行能量最小化（蛋白突变和核酸突变都支持）。优化过程中可自动检测小分子配体并使用 GAFF 力场进行参数化。
Structure Minimization 提供两种最小化方法：

1. openmm（默认）：OpenMM 内置 LocalEnergyMinimizer（L-BFGS），在CPU和GPU计算平台上结果具有非确定性（结果不可重现）
2. capped-sd：自定义的确定性能量最速下降法（GPU 力求值 + NumPy 坐标更新），在CPU和GPU计算平台上结果均可重现

参数说明

Input File

输入的蛋白质/核酸/小分子/复合物 PDB 文件，必选项。如果存在残基编号间隙，可在 PDB 中提供 SEQRES 记录以便自动补全（晶体结构中一般都有SEQRES记录因此会自动补全）。

Mutations

突变指定，可选项。省略时进入 WT-only 模式（仅计算 WT 的结合自由能）。

mutations.txt文件内容示例：

#A100V （注释行，可省略）
A:100:VAL
A:100:VAL,A:105:LEU 

备注：如果Input File中没有链名，可以不指定链名，如100:VAL（表示第100个残基突变为VAL），但当有多条链都包含有指定的突变残基时会报错

Method

最小化方法，必选项，默认 openmm。

• openmm：OpenMM 内置 L-BFGS，速度快但 GPU 上非确定性
• capped-sd：自定义的确定性最速下降方法，结果可重现

Add Hydrogens

控制在结构准备过程中如何处理氢原子。

• --add-hydrogens：默认删除所有H，然后根据pH重建H原子
• --no-add-hydrogens：跳过 H 处理，使用原始输入结构中的H原子，适用于原始输入结构已经进行过H处理的PDB文件

Keep Hydrogens

控制是否保留输入结构中的原始氢原子，可选项。默认删除所有原始氢原子，随后根据设定的 pH 条件重新构建全部氢原子。

• --keep-hydrogens：保留输入结构中原始H原子，仅补缺失的H原子，适用于原始结构中已经包含了部分H原子，但仍然缺失H原子的PDB文件

pH

对Input File文件进行加氢时参考的pH状态，会根据pH值进行残基的质子化状态判定，默认 7.0

Ligand SMILES

小分子配体的 SMILES，可选项。用于确保小分子配体正确的键序和连接性，提供时会先去除配体 H 再进行键序匹配，完成后自动重新添加。（当输入结构没有提供键连关系和键序信息时对小分子配体很难做到准确加H，提供小分子配体的smiles可做到对小分子配体的准确加H）。

Ligand SMILES书写格式:
"OC[C@H]1O[C@@H](n2cnc3cc(Cl)c(Cl)cc32)[C@H](O)[C@@H]1O" （适用于Input File中只含有一种配体的情况）
"RFZ:OC[C@H]1O[C@@H](n2cnc3cc(Cl)c(Cl)cc32)[C@H](O)[C@@H]1O,W2R:O=C(Nc1cccc(Oc2ccc(Nc3ncnc4ccn(CCOCCO)c34)cc2Cl)c1)NC1CCCCC1" （适用于Input File中含有多种配体的情况，以逗号分隔）

Min Tolerance

能量最小化收敛精度 (kJ/mol/nm)，默认 1.0，值越小越精确。

Max Steps

能量最小化最大迭代步数，默认 5000。

Restraint Force

骨架位置限制力常数 (kJ/mol/nm^2)，默认 100.0，设为 0 表示不对骨架位置进行限制。

结果说明

WT-only 模式输出

突变模式输出

如何理解结果

1. openmm方法在CPU和GPU计算平台上结果均不可重现
2. capped-sd方法在CPU和GPU计算平台上结果均可重现

注意事项

1. 修饰残基：修饰残基嵌入蛋白链或者核酸链时会报错
2. 缺失残基：PDB 中存在残基编号间隙但缺少 SEQRES 记录时会报错
3. 突变支持：支持蛋白质残基突变（标准氨基酸）和核酸残基突变（DNA/RNA）
4. 小分子 SMILES：强烈建议提供 --ligand-smiles 以确保正确的键序和连接性

参考文献

• Friedrichs M. S., et al. Accelerating molecular dynamic simulations on graphics processing units[J]. J. Comput. Chem., 2009, 30: 864-872.DOI: 10.1002/jcc.21209
• Eastman P., Pande V. S. Efficient nonbonded interactions for molecular dynamics on a graphics processing unit[J]. J. Comput. Chem., 2010, 31: 1268-1272.DOI: 10.1002/jcc.21413
• Eastman P., Pande V. S. Constant constraint matrix approximation: A robust, parallelizable constraint method for molecular simulations[J]. J. Chem. Theor. Comput., 2010, 6: 434-437.DOI: 10.1021/ct900463w
• Eastman P., et al. OpenMM 8: Molecular Dynamics Simulation with Machine Learning Potentials[J]. J. Phys. Chem. B, 2023, 128: 109-116.DOI: 10.1021/acs.jpcb.3c06662

Structure Minimization

Introduction

Structure Minimization performs energy minimization on protein/nucleic acid/small-molecule/complex structures in GB implicit solvent . When mutations are specified, it also supports energy minimization of mutant structures (both protein and nucleic acid mutations are supported). During optimization, small-molecule ligands are automatically detected and parameterized using the GAFF force field.

Structure Minimization provides two minimization methods:

1. openmm (default): OpenMM’s built-in LocalEnergyMinimizer (L-BFGS). Results are non-deterministic on both CPU and GPU platforms (not reproducible).
2. capped-sd: A custom deterministic energy steepest descent method (GPU force evaluation + NumPy coordinate updates). Results are reproducible on both CPU and GPU platforms.

Parameters

Input File

Input protein/nucleic acid/small-molecule/complex PDB file. Required. If residue numbering gaps exist, a SEQRES record can be provided in the PDB for automatic completion (crystal structures typically contain SEQRES records, so completion is automatic).

Mutations

Mutation specification. Optional. When omitted, the tool enters WT-only mode.

Example mutations.txt file content:

#A100V (comment line, can be omitted)
A:100:VAL
A:100:VAL,A:105:LEU

Note: If the Input File does not contain chain names, the chain name can be omitted, e.g. 100:VAL (indicating residue 100 is mutated to VAL). However, an error will be raised when multiple chains contain the specified mutation residue.

Method

Minimization method. Required. Default: openmm.

• openmm: OpenMM’s built-in L-BFGS. Fast but non-deterministic on GPU.
• capped-sd: Custom deterministic steepest descent method. Results are reproducible.

Add Hydrogens

Controls how hydrogen atoms are handled during structure preparation.

• Add Hydrogens (default): Deletes all H atoms, then rebuilds them according to pH.
• No Add Hydrogens: Skips H processing and uses H atoms from the original input structure. Suitable for PDB files that have already been H-treated.

Keep Hydrogens

Controls whether original hydrogen atoms from the input structure are preserved. Optional. By default, all original H atoms are deleted and subsequently rebuilt according to the set pH condition.

• --keep-hydrogens: Preserves original H atoms from the input structure and only adds missing H atoms. Suitable for PDB files where the original structure already contains partial H atoms but still has missing H atoms.

pH

pH state referenced during hydrogen addition to the Input File. Residue protonation states are determined based on the pH value. Default: 7.0.

Ligand SMILES

SMILES string of the small-molecule ligand. Optional. Used to ensure correct bond order and connectivity of the small-molecule ligand. When provided, ligand H atoms are first removed for bond-order matching, then automatically re-added. (When the input structure does not provide bond connectivity and bond order information, accurate H addition for small-molecule ligands is difficult; providing the SMILES enables accurate H addition for the ligand.)

Ligand SMILES format:
"OC[C@H]1O[C@@H](n2cnc3cc(Cl)c(Cl)cc32)[C@H](O)[C@@H]1O" (for cases where the Input File contains only one ligand)
"RFZ:OC[C@H]1O[C@@H](n2cnc3cc(Cl)c(Cl)cc32)[C@H](O)[C@@H]1O,W2R:O=C(Nc1cccc(Oc2ccc(Nc3ncnc4ccn(CCOCCO)c34)cc2Cl)c1)NC1CCCCC1" (for cases where the Input File contains multiple ligands, comma-separated)

Min Tolerance

Energy minimization convergence tolerance (kJ/mol/nm). Default: 1.0. Smaller values are more precise.

Max Steps

Maximum number of energy minimization iterations. Default: 5000.

Restraint Force

Backbone position restraint force constant (kJ/mol/nm²). Default: 100.0. Set to 0 to disable backbone position restraints.

Results

WT-only Mode Output

Mutation Mode Output

Interpreting Results

1. The openmm method produces non-reproducible results on both CPU and GPU platforms.
2. The capped-sd method produces reproducible results on both CPU and GPU platforms.

Notes

1. Modified residues: An error will be raised if modified residues are embedded in the protein or nucleic acid chain.
2. Missing residues: An error will be raised if residue numbering gaps exist in the PDB but no SEQRES record is provided.
3. Mutation support: Supports protein residue mutations (standard amino acids) and nucleic acid residue mutations (DNA/RNA).
4. Small-molecule SMILES: It is strongly recommended to provide --ligand-smiles to ensure correct bond order and connectivity.

References

• Friedrichs M. S., et al. Accelerating molecular dynamic simulations on graphics processing units[J]. J. Comput. Chem., 2009, 30: 864-872. DOI: 10.1002/jcc.21209
• Eastman P., Pande V. S. Efficient nonbonded interactions for molecular dynamics on a graphics processing unit[J]. J. Comput. Chem., 2010, 31: 1268-1272. DOI: 10.1002/jcc.21413
• Eastman P., Pande V. S. Constant constraint matrix approximation: A robust, parallelizable constraint method for molecular simulations[J]. J. Chem. Theor. Comput., 2010, 6: 434-437. DOI: 10.1021/ct900463w
• Eastman P., et al. OpenMM 8: Molecular Dynamics Simulation with Machine Learning Potentials[J]. J. Phys. Chem. B, 2023, 128: 109-116. DOI: 10.1021/acs.jpcb.3c06662

Mutation Energy Calculation (ddG)

简介

Mutation Energy Calculation (ddG) 用于计算在蛋白质/核酸/小分子复合物结构中，由于突变而引起的结合自由能差（即突变能，ddG）。当不指定突变时可用于计算蛋白质/核酸/小分子复合物结构的结合自由能。支持蛋白突变和核酸（DNA/RNA）突变。

参数说明

Input File

输入的蛋白质/核酸/小分子/复合物 PDB 文件，必选项。如果存在残基编号间隙，可在 PDB 中提供 SEQRES 记录以便自动补全（晶体结构中一般都有SEQRES记录因此会自动补全）。

Receptor Chains

输入结构中受体链 ID（逗号分隔），默认为全部非配体链。

D
B,C

Receptor Residues

输入结构中受体残基号范围，默认为全部非配体链。

1-100,120 （如输入结构中没有包含链名，可不指定链名，但当有多条链都包含有指定的残基时会报错）
 A:1-100,B:200

Mutations

突变指定，可选项。省略时进入 WT-only 模式（仅计算 WT 的结合自由能）。

mutations.txt文件内容示例：

#A100V （注释行，可省略）
A:100:VAL
A:100:VAL,A:105:LEU 

备注：如果Input File中没有链名，可以不指定链名，如100:VAL（表示第100个残基突变为VAL），但当有多条链都包含有指定的突变残基时会报错

Ligand Chains

输入结构的受体链 ID（逗号分隔）。

注意：Ligand Chains、Ligand Residues和Ligand Name参数三选一

Ligand Residues

从输入结构中指定的小分子的名称。

501-520,530 （如Input File中没有包含链名，可不指定链名，但当有多条链都包含有指定的残基时会报错）
B:501-520

Ligand Name

从输入结构中指定的小分子的名称。

RFZ
LIG

Ligand SMILES

小分子配体的 SMILES，可选项。用于确保小分子配体正确的键序和连接性，提供时会先去除配体 H 再进行键序匹配，完成后自动重新添加。（当输入结构没有提供键连关系和键序信息时对小分子配体很难做到准确加H，提供小分子配体的smiles可做到对小分子配体的准确加H）。

Ligand SMILES书写格式:

"OC[C@H]1O[C@@H](n2cnc3cc(Cl)c(Cl)cc32)[C@H](O)[C@@H]1O" （适用于Input File中只含有一种配体的情况）
"RFZ:OC[C@H]1O[C@@H](n2cnc3cc(Cl)c(Cl)cc32)[C@H](O)[C@@H]1O,W2R:O=C(Nc1cccc(Oc2ccc(Nc3ncnc4ccn(CCOCCO)c34)cc2Cl)c1)NC1CCCCC1" （适用于Input File中含有多种配体的情况，以逗号分隔）

Add Hydrogens

控制在结构准备过程中如何处理氢原子。

• --add-hydrogens：默认删除所有H，然后根据pH重建H原子
• --no-add-hydrogens：跳过 H 处理，使用原始输入结构中的H原子，适用于原始输入结构已经进行过H处理的PDB文件

Keep Hydrogens

控制是否保留输入结构中的原始氢原子，可选项。默认删除所有原始氢原子，随后根据设定的 pH 条件重新构建全部氢原子。

• --keep-hydrogens：保留输入结构中原始H原子，仅补缺失的H原子，适用于原始结构中已经包含了部分H原子，但仍然缺失H原子的PDB文件

pH

对输入结构文件进行加氢时参考的pH状态，会根据pH值进行残基的质子化状态判定，默认 7.0

Energy Model

溶剂化模型，控制 PB/GB 静电相互作用的计算方法，必选项，默认 ALPB。

• GB：Generalized Born，广义 Born 模型，适用于一般的溶剂化能计算。
• ALPB：Analytical Linearized Poisson-Boltzmann，解析线性化泊松-玻尔兹曼模型，精度较高，适合需要更准确静电相互作用的场景。
• CHAGB：Charge-Dependent GB，电荷依赖型 GB 模型，考虑原子电荷变化对溶剂化的影响。
• CHAGBCAN：Charge-Dependent GB with canonical radii，使用标准原子半径的电荷依赖型 GB 模型。

Inradii

溶剂化半径，可选项

• inpqr：使用 PQR 文件中的 BONDI 半径
• bestgb：使用 GB 优化半径
• chagb：使用 CHAGB 专用半径（仅限 CHAGB/CHAGBCAN 模型）

Ele Corr

启用 Debye-Huckel 静电屏蔽校正，默认关闭（不进行静电能校正）

Temp

温度 (K)，默认 298.15

Min Tolerance

能量最小化收敛精度 (kJ/mol/nm)，默认 1.0，值越小越精确。

Max Steps

能量最小化最大迭代步数，默认 5000。

Restraint Force

骨架位置限制力常数 (kJ/mol/nm^2)，默认 100.0，设为 0 表示不对骨架位置进行限制。

Output

结果输出文件，可选项，默认mutations.csv

结果说明

1. 突变能输出到mutations.csv文件中
   包含信息如下：

如果未指定突变，则进入WT-Only模式，csv文件中只有输入结构的结合自由能

mutations.csv:
mutation,WT_G_bind,MUT_G_bind,DDG
WT-only,-15.2300,N/A,N/A

2. 突变前后对应的PDB结构文件
   | 文件 | 说明 |
   |------|------|
   | WT_minimized.pdb | WT 能量最小化后的结构 |
   | MUT_<链名>_<残基号>_<突变残基名称>_minimized.pdb | MUT 能量最小化后的结构 |
   如果未指定突变，则进入WT-Only模式，则只输出WT_minimized.pdb结构

如何理解结果

1. DDG > 0：突变削弱结合（不利突变）
2. DDG < 0：突变增强结合（有利突变）

注意事项

1. 修饰残基：修饰残基嵌入蛋白链或者核酸链时会报错
2. 缺失残基：PDB 中存在残基编号间隙但缺少 SEQRES 记录时会报错
3. 突变支持：支持蛋白质残基突变（标准氨基酸）和核酸残基突变（DNA/RNA）
4. 配体指定：--ligand-chains、--ligand-residues、--ligand-name 三选一，至少提供一个
5. 小分子 SMILES：强烈建议提供 --ligand-smiles 以确保正确的键序和连接性

参考文献

• Friedrichs M. S., et al. Accelerating molecular dynamic simulations on graphics processing units[J]. J. Comput. Chem., 2009, 30: 864-872.DOI: 10.1002/jcc.21209
• Eastman P., Pande V. S. Efficient nonbonded interactions for molecular dynamics on a graphics processing unit[J]. J. Comput. Chem., 2010, 31: 1268-1272.DOI: 10.1002/jcc.21413
• Eastman P., Pande V. S. Constant constraint matrix approximation: A robust, parallelizable constraint method for molecular simulations[J]. J. Chem. Theor. Comput., 2010, 6: 434-437.DOI: 10.1021/ct900463w
• Eastman P., et al. OpenMM 8: Molecular Dynamics Simulation with Machine Learning Potentials[J]. J. Phys. Chem. B, 2023, 128: 109-116.DOI: 10.1021/acs.jpcb.3c06662

Mutation Energy Calculation (ddG)

Introduction

Mutation Energy Calculation (ddG) computes the change in binding free energy (i.e., mutation energy, ddG) caused by mutations in protein/nucleic acid/small-molecule complex structures. When no mutation is specified, it can be used to calculate the binding free energy of the protein/nucleic acid/small-molecule complex structure. Supports both protein mutations and nucleic acid (DNA/RNA) mutations.

Parameters

Input File

Input protein/nucleic acid/small-molecule/complex PDB file. Required. If residue numbering gaps exist, a SEQRES record can be provided in the PDB for automatic completion (crystal structures typically contain SEQRES records, so completion is automatic).

Mutations

Mutation specification. Optional. When omitted, the tool enters WT-only mode (only the WT binding free energy is calculated).

Example mutations.txt file content:

#A100V (comment line, can be omitted)
A:100:VAL
A:100:VAL,A:105:LEU

Note: If the Input File does not contain chain names, the chain name can be omitted, e.g. 100:VAL (indicating residue 100 is mutated to VAL). However, an error will be raised when multiple chains contain the specified mutation residue.

Receptor Residues

Receptor residue number range(s) from the input structure. Defaults to all non-ligand chains.

1-100,120 (if the Input File does not contain chain names, the chain name can be omitted; however, an error will be raised when multiple chains contain the specified residues)
A:1-100,B:200

Receptor Chains

Receptor chain ID(s) from the input structure (comma-separated). Defaults to all non-ligand chains.

D
B,C

Ligand Chains

Ligand chain ID(s) from the input structure (comma-separated). Defaults to all non-ligand chains.

D
B,C

Note: Exactly one of Ligand Chains, Ligand Residues, and Ligand Name must be provided.

Ligand Residues

Specify small-molecule residue name(s) from the input structure.

501-520,530 (if the Input File does not contain chain names, the chain name can be omitted; however, an error will be raised when multiple chains contain the specified residues)
B:501-520

Ligand Name

Specify small-molecule name(s) from the input structure.

RFZ
LIG

Ligand SMILES

SMILES string of the small-molecule ligand. Optional. Used to ensure correct bond order and connectivity of the small-molecule ligand. When provided, ligand H atoms are first removed for bond-order matching, then automatically re-added. (When the input structure does not provide bond connectivity and bond order information, accurate H addition for small-molecule ligands is difficult; providing the SMILES enables accurate H addition for the ligand.)
Ligand SMILES format:

"OC[C@H]1O[C@@H](n2cnc3cc(Cl)c(Cl)cc32)[C@H](O)[C@@H]1O" (for cases where the Input File contains only one ligand)
"RFZ:OC[C@H]1O[C@@H](n2cnc3cc(Cl)c(Cl)cc32)[C@H](O)[C@@H]1O,W2R:O=C(Nc1cccc(Oc2ccc(Nc3ncnc4ccn(CCOCCO)c34)cc2Cl)c1)NC1CCCCC1" (for cases where the Input File contains multiple ligands, comma-separated)

Add Hydrogens

Controls how hydrogen atoms are handled during structure preparation.

• --add-hydrogens (default): Deletes all H atoms, then rebuilds them according to pH.
• --no-add-hydrogens: Skips H processing and uses H atoms from the original input structure. Suitable for PDB files that have already been H-treated.

Keep Hydrogens

Controls whether original hydrogen atoms from the input structure are preserved. Optional. By default, all original H atoms are deleted and subsequently rebuilt according to the set pH condition.

• --keep-hydrogens: Preserves original H atoms from the input structure and only adds missing H atoms. Suitable for PDB files where the original structure already contains partial H atoms but still has missing H atoms.

pH

pH state referenced during hydrogen addition to the input structure file. Residue protonation states are determined based on the pH value. Default: 7.0.

Energy Model

Solvation model. Controls the calculation method for PB/GB electrostatic interactions. Required. Default: ALPB.

• GB: Generalized Born model. Suitable for general solvation energy calculations.
• ALPB: Analytical Linearized Poisson-Boltzmann model. Higher accuracy, suitable for scenarios requiring more accurate electrostatic interactions.
• CHAGB: Charge-Dependent GB model. Considers the effect of atomic charge changes on solvation.
• CHAGBCAN: Charge-Dependent GB with canonical radii. Uses standard atomic radii.

Inradii

Solvation radii. Optional.

• inpqr: Uses BONDI radii from the PQR file.
• bestgb: Uses GB-optimized radii.
• chagb: Uses CHAGB-specific radii (for CHAGB/CHAGBCAN models only).

Ele Corr

Enable Debye-Huckel electrostatic shielding correction. Disabled by default (no electrostatic energy correction).

Temp

Temperature (K). Default: 298.15.

Min Tolerance

Energy minimization convergence tolerance (kJ/mol/nm). Default: 1.0. Smaller values are more precise.

Max Steps

Maximum number of energy minimization iterations. Default: 5000.

Restraint Force

Backbone position restraint force constant (kJ/mol/nm²). Default: 100.0. Set to 0 to disable backbone position restraints.

Results

Result output file. Optional. Default: mutations.csv.

1. Mutation energies are output to the mutations.csv file, containing the following columns:

If no mutation is specified, the tool enters WT-only mode, and the CSV file contains only the binding free energy of the input structure:

mutations.csv:
mutation,WT_G_bind,MUT_G_bind,DDG
WT-only,-15.2300,N/A,N/A

2. PDB structure files corresponding to pre- and post-mutation states:

If no mutation is specified, the tool enters WT-only mode and only outputs WT_minimized.pdb.

Interpreting Results

1. DDG > 0: The mutation weakens binding (unfavorable mutation).
2. DDG < 0: The mutation strengthens binding (favorable mutation).

Notes

1. Modified residues: An error will be raised if modified residues are embedded in the protein or nucleic acid chain.
2. Missing residues: An error will be raised if residue numbering gaps exist in the PDB but no SEQRES record is provided.
3. Mutation support: Supports protein residue mutations (standard amino acids) and nucleic acid residue mutations (DNA/RNA).
4. Ligand specification: Exactly one of --ligand-chains, --ligand-residues, and --ligand-name must be provided.
5. Small-molecule SMILES: It is strongly recommended to provide --ligand-smiles to ensure correct bond order and connectivity.

References

• Friedrichs M. S., et al. Accelerating molecular dynamic simulations on graphics processing units[J]. J. Comput. Chem., 2009, 30: 864-872. DOI: 10.1002/jcc.21209
• Eastman P., Pande V. S. Efficient nonbonded interactions for molecular dynamics on a graphics processing unit[J]. J. Comput. Chem., 2010, 31: 1268-1272. DOI: 10.1002/jcc.21413
• Eastman P., Pande V. S. Constant constraint matrix approximation: A robust, parallelizable constraint method for molecular simulations[J]. J. Chem. Theor. Comput., 2010, 6: 434-437. DOI: 10.1021/ct900463w
• Eastman P., et al. OpenMM 8: Molecular Dynamics Simulation with Machine Learning Potentials[J]. J. Phys. Chem. B, 2023, 128: 109-116. DOI: 10.1021/acs.jpcb.3c06662

Peptide Property Prediction (PeptiVerse)

简介

基于PeptiVerse深度学习模型的多肽ADMET性质预测工具，支持溶血性、溶解性、细胞穿透性、毒性、膜通透性、半衰期等多种性质的批量预测。输入支持标准氨基酸序列和 SMILES 化学结构两种格式，适用于线性肽、环肽及修饰肽的虚拟筛选与性质评估。

适用场景

• 多肽药物早期筛选：快速评估候选多肽的成药性关键性质
• 安全性评价：预测溶血性、毒性等安全性指标
• 递送潜力评估：评估细胞穿透性和膜通透能力
• 靶点结合分析：预测多肽与目标蛋白的结合亲和力

参数说明

Peptide Sequence

Input Peptides

输入的FASTA 格式多肽序列文件：

>id1
ZCVBDSWERTA
>id2
WERTAZCV

Property

预测属性名称，必填，支持多选。

• permeability_penetrance：细胞穿透性，预测多肽进入或穿透细胞膜的能力
• hemolysis：溶血性，预测多肽破坏红细胞膜并引发溶血的风险
• nf：抗污性 / 非特异性吸附，预测多肽发生非特异性蛋白吸附的倾向
• solubility：溶解性，预测多肽在水相环境中的溶解能力
• halflife：半衰期，预测多肽在体内的稳定性和半衰期表现

Uncertainty

是否计算预测不确定性。启用后输出结果包含不确定性估计值，有助于评估预测可靠性。

• true：计算不确定性
• false：不计算不确定性

Output

预测结果的输出文件路径，默认输出为 results.csv。

Peptide Smiles

Input Peptides

输入的多肽文件，支持 SMILES 格式：

N[C@@H](CC(C)C)C(=O)N[C@@H](CC(=O)N)C(=O)N[C@@H](CCC(=O)N)C(=O)N[C@@H](CC(=CN2)C1=C2C=CC=C1)C(=O)N[C@@H](CC(=O)O)C(=O)N[C@@H](CO)C(=O)N[C@@H](CC(=CN2)C1=C2C=CC=C1)C(=O)N[C@@H](Cc1ccccc1)C(=O)N[C@@H](CC(=O)N)C(=O)N[C@@H](CCC(=O)O)C(=O)N[C@@H](CCSC)C(=O)N[C@@H]([C@H](CC)C)C(=O)N[C@@H](CCC(=O)N)C(=O)N[C@@H](CCCCN)C(=O)N[C@@H]([C@H](O)C)C(=O)NCC(=O)N[C@@H](CO)C(=O)O 

Property

预测属性名称，必填，支持多选。

• hemolysis：溶血性，预测多肽破坏红细胞膜并引发溶血的风险
• nf：抗污性 / 非特异性吸附，预测多肽发生非特异性蛋白吸附的倾向
• solubility：溶解性，预测多肽在水相环境中的溶解能力
• permeability_penetrance：细胞穿透性，预测多肽进入或穿透细胞膜的能力
• toxicity：毒性，预测多肽潜在毒性风险
• permeability_pampa：PAMPA 通透性，预测多肽在人工膜通透性实验中的表现
• permeability_caco2：Caco-2 通透性，预测多肽在 Caco-2 细胞模型中的通透能力
• halflife：半衰期，预测多肽在体内的稳定性和半衰期表现

Uncertainty

是否计算预测不确定性。启用后输出结果包含不确定性估计值，有助于评估预测可靠性。

• true：计算不确定性
• false：不计算不确定性

Output

预测结果的输出文件路径，默认输出为 results.csv。

结果说明

输出结果包括 results.csv 预测结果表格，包含每条多肽的各项预测性质及对应的不确定性。

results.csv 包含信息如下：

• 不确定性类型说明

注意：不确定性指标仅在 Uncertainty 选择 true 时输出。

参考文献

• Zhang Y, Tang S, Chen T, Mahood E, Vincoff S, Chatterjee P. PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction[bioRxiv]. 2025.DOI: 10.64898/2025.12.31.697180

Peptide Property Prediction (PeptiVerse)

Introduction

A deep learning-based multi-property prediction tool for peptides, supporting batch prediction of properties including hemolysis, solubility, cell penetration, toxicity, membrane permeability, half-life, and binding affinity. Input supports both standard amino acid sequences and SMILES chemical structure formats, making it suitable for virtual screening and property evaluation of linear peptides, cyclic peptides, and modified peptides.

Use Cases

• Early-stage peptide drug screening: Rapid assessment of key druggability properties for candidate peptides
• Safety evaluation: Prediction of safety-related indicators such as hemolysis and toxicity
• Delivery potential assessment: Evaluation of cell penetration and membrane permeability
• Target binding analysis: Prediction of binding affinity between peptides and target proteins

Parameters

Peptide Sequence

Input Peptides

Input peptide sequence file in FASTA format:

>id1
ZCVBDSWERTA
>id2
WERTAZCV

Property

The properties to predict. Required; multiple selections supported.

• permeability_penetrance: Cell penetration — predicts the ability of a peptide to enter or traverse the cell membrane
• hemolysis: Hemolysis — predicts the risk of a peptide disrupting red blood cell membranes and causing hemolysis
• nf: Non-fouling / non-specific adsorption — predicts the tendency of a peptide to undergo non-specific protein adsorption
• solubility: Solubility — predicts the ability of a peptide to dissolve in an aqueous environment
• halflife: Half-life — predicts the in vivo stability and half-life performance of a peptide

Peptide SMILES

Input Peptides

Input peptide file in SMILES format:

N[C@@H](CC(C)C)C(=O)N[C@@H](CC(=O)N)C(=O)...

Property

The properties to predict. Required; multiple selections supported.

• hemolysis: Hemolysis — predicts the risk of a peptide disrupting red blood cell membranes and causing hemolysis
• nf: Non-fouling / non-specific adsorption — predicts the tendency of a peptide to undergo non-specific protein adsorption
• solubility: Solubility — predicts the ability of a peptide to dissolve in an aqueous environment
• permeability_penetrance: Cell penetration — predicts the ability of a peptide to enter or traverse the cell membrane
• toxicity: Toxicity — predicts the potential toxicity risk of a peptide
• permeability_pampa: PAMPA permeability — predicts peptide performance in artificial membrane permeability assays
• permeability_caco2: Caco-2 permeability — predicts peptide permeability in the Caco-2 cell model
• halflife: Half-life — predicts the in vivo stability and half-life performance of a peptide
• binding_affinity: Binding affinity — predicts the binding strength between a peptide and a target protein. If no Target Sequence is provided, this property will be automatically skipped.

Uncertainty

Whether to calculate prediction uncertainty. When enabled, the output includes uncertainty estimates to help assess prediction reliability.

• true: Calculate uncertainty
• false: Do not calculate uncertainty

Output

Output file path for prediction results. Defaults to results.csv.

Results

The output is a results.csv prediction table containing the predicted properties and corresponding uncertainty estimates for each peptide.

Uncertainty Type Explanation

Note: Uncertainty columns are only included in the output when Uncertainty is set to true.

Reference

• Zhang Y, Tang S, Chen T, Mahood E, Vincoff S, Chatterjee P. PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction [bioRxiv]. 2025. DOI: 10.64898/2025.12.31.697180

Extract Fv and Analyze Contacts

简介

从输入 PDB 文件中自动提取抗体 Fv 区域及邻近分子片段，生成包含 Fv 与伙伴链的截断 PDB 和 Fv 序列文件，并进行界面（interface）和氢键（hydrogen bond）相互作用计算。

核心技术

• 自识别 VH/VL：基于保守的 Cys-Trp motif 和 FR1 起始序列特征自动识别可变区，无需外部数据库
• 编号方案：内置 IMGT、Kabat、Chothia、Martin、CCG 五种 CDR 位置定义
• Fv 截断：按各方案对应的 Fv 末端位置截断，去除 CH1/CL 恒定区（如 IMGT≈128，Kabat≈113）
• 伙伴链保留：通过 NeighborSearch 识别与 Fv 原子距离在截止范围内的所有链（抗原、小分子、离子等），一并写入输出 PDB

适用场景

• 抗体结构预处理：自动提取 Fv 区域用于后续人源化、亲和力成熟等流程
• 分子相互作用分析：计算抗体与抗原、配体之间的界面接触和氢键网络
• 结构数据准备：生成标准化的 Fv 结构文件和序列文件用于下游分析

参数说明

Input Structure

输入的抗体 PDB 结构文件，需包含完整的抗体结构及可能结合的抗原、配体或其他分子。输入时请限制抗体及其相互作用的对象是一对一的，例如一个轻重连构成的抗体对应抗原，而非多个抗体对应一个抗原

Numbering Scheme

Fv 编号方案，用于确定 CDR 位置和 Fv 截断点。

• IMGT：国际免疫基因组学标准，Fv 末端约 128 位
• Kabat：基于序列变异性定义，Fv 末端约 113 位
• Chothia：基于结构环区定义
• Martin：基于 Kabat 的修订版本
• CCG：癌症基因组学联盟方案

Contact Cutoff Distance

Fv 与邻近分子的接触截止距离，用于识别需要保留的伙伴链。单位 Å，默认 10.0 Å。

结果说明

输出结果包括：

Extract Fv and Analyze Contacts

Introduction

Automatically extracts the antibody Fv region and neighboring molecular fragments from an input PDB file, generates a truncated PDB containing Fv with partner chains and an Fv sequence file, and calculates interface and hydrogen bond interactions.

Core Technologies

• VH/VL Auto-identification: Automatically identifies variable regions based on conserved Cys-Trp motifs and FR1 starting sequence features without external databases
• Numbering Schemes: Built-in definitions for five CDR positioning schemes: IMGT, Kabat, Chothia, Martin, and CCG
• Fv Truncation: Truncates at Fv terminus positions according to each scheme to remove CH1/CL constant regions (e.g., IMGT≈128, Kabat≈113)
• Partner Chain Retention: Uses NeighborSearch to identify all chains within the cutoff distance of Fv atoms (antigens, small molecules, ions, etc.) and writes them into the output PDB

Use Cases

• Antibody structure preprocessing: Extract Fv regions for downstream humanization, affinity maturation, and other workflows
• Molecular interaction analysis: Calculate interface contacts and hydrogen bond networks between antibodies and antigens/ligands
• Structural data preparation: Generate standardized Fv structure and sequence files for downstream analysis

Parameters

Input Structure

Input antibody PDB structure file, which should contain the complete antibody structure and any bound antigens, ligands, or other molecules.

Numbering Scheme

Fv numbering scheme used to determine CDR positions and Fv truncation points.

• IMGT: International ImMunoGeneTics standard, Fv terminus around position 128
• Kabat: Defines CDRs based on sequence variability, Fv terminus around position 113
• Chothia: Defines CDRs based on structural loop regions
• Martin: Revised version based on Kabat
• CCG: Cancer Genome Consortium scheme

Contact Cutoff Distance

Contact cutoff distance between Fv and neighboring molecules for identifying partner chains to retain. Unit: Å, default 10.0 Å.

Results

The output includes the following files:

ImmuneReport

简介

对 Immunogenicity Prediction (AlphaMHC v3.0 beta)和 Immunogenicity Prediction (WeADApt v4.1.0)、Immunogenicity Prediction (WeADApt v4.2)、Immunogenicity Prediction (WeADApt v4.3) 四个免疫原性评估模块的结果进行汇总，生成分子和表位级别的整合报告。该模块为流程编排组件，需配合上游免疫原性预测模块使用。

参数说明

Input Directory

基础输入目录，仅作为省略的输入文件参数的默认路径前缀。

FASTA

FASTA 格式的氨基酸序列文件。

AlphaMHC v3 Molecule Score

AlphaMHC v3.0 分子评分 CSV 文件。

AlphaMHC v3 Epitope Score

AlphaMHC v3.0 表位评分 CSV 文件。

WeAdapt v4.1 Molecule Score

WeAdapt v4.1 分子评分 CSV 文件。

WeAdapt v4.1 Epitope Score

WeAdapt v4.1 表位评分 CSV 文件。

WeAdapt v4.2 Molecule Score

WeAdapt v4.2 分子评分 CSV 文件。

WeAdapt v4.2 Epitope Score

WeAdapt v4.2 表位评分 CSV 文件。

WeAdapt v4.3 Molecule Score

WeAdapt v4.3 分子评分 CSV 文件。

WeAdapt v4.3 Epitope Score

WeAdapt v4.3 表位评分 CSV 文件。

Molecule Summary

分子汇总 CSV 输出路径。

Epitope Summary

表位汇总 CSV 输出路径。

Errors

记录级错误 CSV 输出路径。

结果说明

输出结果包括：

molecule_summary.csv文件包含信息如下：

epitope_summary.csv文件包含信息如下：

ImmuneReport

Introduction

Aggregates results from four immunogenicity assessment modules ( Immunogenicity Prediction (AlphaMHC v3.0 beta) and Immunogenicity Prediction (WeADApt v4.1.0)、Immunogenicity Prediction (WeADApt v4.2)、Immunogenicity Prediction (WeADApt v4.3)) to generate integrated molecule-level and epitope-level reports. This module is a workflow orchestration component and must be used in conjunction with upstream immunogenicity prediction modules.

Parameters

Input Directory

Base input directory used only as the default path prefix for omitted input file arguments.

FASTA

Amino acid sequence file in FASTA format.

AlphaMHC v3 Molecule Score

AlphaMHC v3.0 molecule score CSV file.

AlphaMHC v3 Epitope Score

AlphaMHC v3.0 epitope score CSV file.

WeAdapt v4.1 Molecule Score

WeAdapt v4.1 molecule score CSV file.

WeAdapt v4.1 Epitope Score

WeAdapt v4.1 epitope score CSV file.

WeAdapt v4.2 Molecule Score

WeAdapt v4.2 molecule score CSV file.

WeAdapt v4.2 Epitope Score

WeAdapt v4.2 epitope score CSV file.

WeAdapt v4.3 Molecule Score

WeAdapt v4.3 molecule score CSV file.

WeAdapt v4.3 Epitope Score

WeAdapt v4.3 epitope score CSV file.

Molecule Summary

Molecule summary CSV output path.

Epitope Summary

Epitope summary CSV output path.

Errors

Record-level error CSV output path.

Results

The output includes the following files:

The molecule_summary.csv file contains the following columns:

The epitope_summary.csv file contains the following columns:

Antibody Numbering v3

简介

基于 ANARCI 和 mafft 的抗体序列编号工具，支持 FV 和 FC 批量编号。

Antibody Numbering 是一个抗体序列编号工具，用于将抗体氨基酸序列映射到标准化编号体系。编号后的序列具有统一的位置参照，使得不同抗体之间的同源比对、CDR 精确定位、突变分析等工作成为可能。

抗体序列的氨基酸残基数因克隆不同而差异较大，直接比较两条原始序列很难确定哪些位置是同源的。编号方案通过为每个残基赋予标准化编号来解决这个问题，使研究人员能够准确识别 CDR 和 FR 的边界。

适用场景：

• 抗体工程：精确定位 CDR 区域，指导人源化、亲和力成熟等改造工作。
• 序列分析：对不同来源的抗体进行同源比较，识别保守位点和突变热点。
• 数据库标准化：将序列统一到 IMGT 等标准编号体系，便于入库和检索。
• 质量评估：通过编号结果识别缺失、插入、非典型残基等异常。

FV 编号使用 ANARCI 引擎，自动识别输入序列中的可变区结构域，支持单条序列中包含多个结构域的情况。编号结果包含每个残基的标准化编号、CDR/FR 区域标注及链类型判定。支持 IMGT、Kabat、Chothia、Martin、AHo、CCG 等方案。

FC 编号使用 mafft 多序列比对引擎，将输入序列与已知恒定区模板进行比对，通过匹配率判定同种型和亚型。适用于同型鉴定、Fc 工程改造等下游分析。支持 EU 和 Kabat 方案。

每个编号方案生成独立的 JSON 和 CSV 结果文件。FV 还会生成未覆盖片段 FASTA，FC 还会生成模板匹配率 CSV。summary.jsonl 包含各方案的处理统计，failed.fasta 收集编号失败的序列。

参数说明

FV Numbering

该模式针对抗体的Fv区序列（包括重链 VH 和轻链 VL），通过指定编号规则（如 Kabat、Chothia、或 IMGT等）对氨基酸残基进行标准化编号。

FASTA File

上传需要进行抗体编号的氨基酸序列文件。支持批量提交多条序列，文件内容应使用 FASTA 格式。

Numbering Scheme

可变区编号规则，支持IMGT、Kabat、Chothia、Martin、AHo、CCG可多选。

FC Numbering

通常用于抗体的EU、Kabat标准化编号。

FASTA File

上传需要进行抗体恒定区编号的氨基酸序列文件。支持批量提交多条序列，文件内容应使用 FASTA 格式。

Numbering Scheme

恒定区编号规则：eu，kabat。默认为eu。

结果说明

输出结果包含以下文件：

FV Numbering模式输出的output_{scheme}.csv文件包含信息如下：

FC Numbering模式输出的output_{scheme}.csv文件包含信息如下：

FC Numbering模式输出的output_{scheme}_match_rate.csv文件包含信息如下：

Antibody Numbering v3

Introduction

An antibody sequence numbering tool based on ANARCI and mafft, supporting batch numbering for FV and FC regions.

Antibody Numbering is a tool that maps antibody amino acid sequences to standardized numbering schemes. Numbered sequences share a unified positional reference, enabling homologous alignment across different antibodies, precise CDR localization, and mutation analysis.

The number of amino acid residues in antibody sequences varies widely across clones, making it difficult to identify homologous positions by comparing raw sequences directly. Numbering schemes resolve this by assigning each residue a standardized identifier, allowing researchers to accurately delineate CDR and FR boundaries.

Use cases:

• Antibody engineering: Precisely locate CDR regions to guide humanization, affinity maturation, and other modifications.
• Sequence analysis: Perform homologous comparisons across antibodies from different sources, identifying conserved sites and mutation hotspots.
• Database standardization: Unify sequences under standard numbering schemes such as IMGT for streamlined archiving and retrieval.
• Quality assessment: Detect anomalies such as deletions, insertions, and atypical residues from numbering results.

The FV numbering module uses the ANARCI engine to automatically identify variable-region domains in input sequences, supporting cases where a single sequence contains multiple domains. Results include standardized residue numbering, CDR/FR region annotations, and chain-type classification. Supported schemes include IMGT, Kabat, Chothia, North, Martin, AHo, and CCG.

The FC numbering module uses the mafft multiple-sequence-alignment engine to align input sequences against known constant-region templates, determining isotype and subtype by match rate. Applicable for isotype identification and Fc engineering downstream analyses. Supported schemes include EU and Kabat.

Each numbering scheme generates independent JSON and CSV result files. FV numbering also produces an unassigned-segment FASTA file, and FC numbering produces a template match-rate CSV. summary.jsonl contains per-scheme processing statistics, and failed.fasta collects sequences that failed numbering.

Parameters

FV Numbering

This mode targets Fv-region sequences of antibodies (including heavy chain VH and light chain VL), applying a standardized numbering scheme (e.g., Kabat, Chothia, IMGT) to amino acid residues.

FASTA File

Upload the amino acid sequence file for antibody numbering. Batch submission of multiple sequences is supported; file content must be in FASTA format.

Numbering Scheme

Variable-region numbering rules. Supports IMGT, Kabat, Chothia, Martin, AHo, and CCG. Multiple selection is allowed.

FC Numbering

Commonly used for standardized EU and Kabat numbering of antibody constant regions.

FASTA File

Upload the amino acid sequence file for antibody constant-region numbering. Batch submission of multiple sequences is supported; file content must be in FASTA format.

Numbering Scheme

Constant-region numbering rules: eu, kabat. Default is eu.

Results

Output results include the following files:

The output_{scheme}.csv files produced by both FV Numbering modes contain the following columns:

The output_{scheme}_match_rate.csv files produced by the FC Numbering mode contain the following columns:

FC Numbering mode output output_{scheme}.csv contains the following fields:

Filter Antibody Sequences

简介

基于 ANARCI 的抗体序列快速分类工具，将输入 FASTA 文件中的序列自动划分为可编号、不可编号和异常序列三类，并分别输出到独立的 FASTA 文件中。

核心技术

• ANARCI 编号判定：调用 ANARCI 引擎尝试对每个序列进行编号，根据编号成功与否判定序列类别
• 三分类输出：自动将序列归类为可编号（numberable）、不可编号（unnumberable）和异常（invalid）
• 并行处理：支持多核并行加速，提升大批量序列处理效率
• 多编号方案：内置 IMGT、Kabat、Chothia、Martin 等主流编号方案

适用场景

• 批量序列质控：快速筛选出可被标准编号体系识别的抗体序列
• 数据清洗：从混合序列集中分离异常或低质量序列
• 下游分析预处理：为后续的抗体编号、人源化、CDR 分析等流程准备合格的输入数据

参数说明

Input FASTA

输入的抗体氨基酸序列文件，需为标准 FASTA 格式，支持单条或多条序列。
注意：仅包含完整或可识别 Fv（可变区）结构域的序列才能被 ANARCI 正确编号。

Numbering Scheme

ANARCI 编号方案，用于判定序列是否可被识别为抗体可变区并进行编号。

• IMGT：国际免疫基因组学标准，最常用
• Kabat：基于序列变异性定义
• Chothia：基于结构环区定义
• Martin：基于 Kabat 的修订版本

默认使用 IMGT。

Numberable Output

可编号序列（即成功识别为 Fv 区域的序列）的输出文件路径。
这些序列包含可解析的抗体可变区结构域，可被 ANARCI 成功编号，并适用于下游分析（如 CDR 定位、人源化等）。
默认输出文件为 numberable.fasta。

Unnumberable Output

不可编号序列的输出文件路径。
这些序列不包含可识别的 Fv 区域，或与标准抗体可变区差异过大，因此无法被 ANARCI 识别和编号。
默认输出文件为 unnumberable.fasta。

Invalid Output

异常序列的输出文件路径。
这些序列存在格式错误（如 FASTA 不规范）、包含非标准氨基酸，或其他导致无法解析的问题。
默认输出文件为 invalid.fasta。

输出结果包括以下文件：

Filter Antibody Sequences

Introduction

A rapid antibody sequence classification tool based on ANARCI that automatically partitions sequences from an input FASTA file into three categories: numberable, unnumberable, and invalid, exporting each to separate FASTA files.

Core Technologies

• ANARCI Numbering Assessment: Calls the ANARCI engine to attempt numbering on each sequence, classifying based on success or failure
• Three-Class Output: Automatically categorizes sequences as numberable, unnumberable, or invalid
• Parallel Processing: Supports multi-core parallel acceleration for improved throughput on large batch jobs
• Multiple Numbering Schemes: Built-in support for mainstream schemes including IMGT, Kabat, Chothia, and Martin

Use Cases

• Batch sequence quality control: Quickly filter antibody sequences recognizable by standard numbering systems
• Data cleaning: Separate abnormal or low-quality sequences from mixed sequence sets
• Downstream analysis preprocessing: Prepare qualified input data for subsequent antibody numbering, humanization, and CDR analysis workflows

Parameters

Input FASTA

Input antibody amino acid sequence file in standard FASTA format, supporting single or multiple sequences.

Note: Only sequences containing complete or recognizable Fv (variable region) domains can be correctly numbered by ANARCI.

Numbering Scheme

ANARCI numbering scheme used to determine whether a sequence can be recognized as an antibody variable region and subsequently numbered.

• IMGT: International ImMunoGeneTics information system standard, most commonly used
• Kabat: Based on sequence variability definitions
• Chothia: Based on structural loop definitions
• Martin: Revised version based on Kabat

Default: IMGT.

Numberable Output

Output file path for numberable sequences (i.e., sequences successfully identified as Fv regions).

These sequences contain parseable antibody variable region domains that can be successfully numbered by ANARCI, and are suitable for downstream analyses (e.g., CDR localization, humanization, etc.).

Default output file: numberable.fasta.

Unnumberable Output

Output file path for unnumberable sequences.

These sequences do not contain recognizable Fv regions or deviate too far from standard antibody variable regions, and therefore cannot be recognized or numbered by ANARCI.

Default output file: unnumberable.fasta.

Invalid Output

Output file path for invalid sequences.

These sequences have formatting errors (e.g., non-standard FASTA), contain non-standard amino acids, or have other issues that prevent parsing.

Default output file: invalid.fasta.

Results

The output includes the following files:

Split Antibody Chain

简介

Split Antibody Chain 是一个用于拆分抗体链的工具，能够将混合的抗体序列分离为重链、轻链和非抗体序列。

核心思想
本项目采用 基于抗体编号方案的链分类 策略：

• 重链识别 根据指定的编号方案识别并输出抗体重链序列
• 轻链识别 根据指定的编号方案识别并输出抗体轻链序列
• 非抗体过滤 识别并分离不符合抗体特征的序列

该流程以"基于 IMGT/Kabat/Chothia 编号方案的抗体链分类"为核心，实现抗体序列的自动化拆分和分类功能。

参数说明

Input File

输入文件路径，FASTA 格式，为必选参数。
注意：仅包含完整或可识别 Fv（可变区）结构域的序列才能被 ANARCI 识别为抗体重链和轻链。

Numbering Scheme

抗体编号方案，可选值包括 imgt、kabat 或 chothia。该方案用于链分类的标准依据。

Heavy Chain

输出包含抗体重链序列的文件名称

Light Chain

输出包含抗体轻链序列的文件名称

Non-Antibody

输出包含非抗体序列的文件名称

结果说明

输出结果包括以下 FASTA 格式文件：

所有输出文件均为 FASTA 格式，每条记录包含序列标识符和氨基酸序列。

Split Antibody Chain

Introduction

Split Antibody Chain is a tool for splitting mixed antibody sequences into heavy chains, light chains, and non-antibody sequences.

Core concept

This tool adopts a numbering-scheme-based chain classification strategy:

• Heavy chain recognition: Identifies and outputs antibody heavy chain sequences according to the specified numbering scheme.
• Light chain recognition: Identifies and outputs antibody light chain sequences according to the specified numbering scheme.
• Non-antibody filtering: Identifies and separates sequences that do not meet antibody characteristics.

The workflow centers on “antibody chain classification based on IMGT/Kabat/Chothia numbering schemes,” achieving automated splitting and classification of antibody sequences.

Parameters

Input File

Input file path in FASTA format. Required.

Note: Only sequences containing complete or recognizable Fv (variable region) domains can be recognized by ANARCI as antibody heavy or light chains.

Numbering Scheme

Antibody numbering scheme. Supported values: imgt, kabat, or chothia. This scheme serves as the standard basis for chain classification.

Heavy Chain

Output filename for sequences identified as antibody heavy chains.

Light Chain

Output filename for sequences identified as antibody light chains.

Non-Antibody

Output filename for sequences identified as non-antibody sequences.

Results

Output files include the following FASTA-format files:

All output files are in FASTA format; each record contains a sequence identifier and the amino acid sequence.

Small Molecule PK Predictor

简介

基于 MolPK 模型的药代动力学（PK）参数批量预测工具。利用预训练深度学习模型，从小分子结构（SMILES）及实验条件（物种、给药途径、剂量）预测 PK 参数，支持多种输入格式和灵活的批处理场景。

核心技术

• MolPK 深度学习模型：基于分子表征学习预测药代动力学参数
• 多格式输入：支持 SMILES 文本、CSV 表格、SDF 结构文件
• 灵活批处理：支持单条预测和批量预测，自动识别 CSV 列名
• 双格式输出：同时输出 CSV 结果表和带 PK 属性的 SDF 结构文件

适用场景

• 药物筛选早期：快速评估候选分子的药代动力学性质
• 批量预测：对大型化合物库进行高通量 PK 参数预测
• 实验设计辅助：为体内实验提供剂量和物种选择的参考依据

参数说明

Input

输入的待预测文件，支持 .smi（SMILES 文本）、.csv（表格）或 .sdf（结构文件）格式。为必填参数。

Species

实验物种，可选值为 rat（大鼠）、mou（小鼠）、dog（犬）、hum（人）。用于指定 PK 预测对应的物种背景。

Route

给药途径，可选值为 iv（静脉注射）、po（口服）。不同给药途径对 PK 曲线有显著影响。

Dose

给药剂量，单位为 mg/kg。用于指定预测时对应的剂量条件。

Output

预测结果的 CSV 输出路径。默认输出为 pred_pk_value.csv。

Output SDF

带 PK 预测属性的 SDF 结构文件输出路径。默认输出为 pred_with_pk_value.sdf。

SMILES Column

当输入为 CSV 文件时，指定包含 SMILES 字符串的列名。

Species Column

当输入为 CSV 文件时，指定包含物种信息的列名。

Route Column

当输入为 CSV 文件时，指定包含给药途径信息的列名。

Dose Column

当输入为 CSV 文件时，指定包含剂量信息的列名。

结果说明

输出结果包括：

输出的预测结果文件文件pred_pk_value.csv：

Small Molecule PK Predictor

Introduction

A batch pharmacokinetic (PK) parameter prediction tool . Utilizes a pretrained deep learning model to predict PK parameters from molecular structures (SMILES) and experimental conditions (species, administration route, dose), supporting multiple input formats and flexible batch processing scenarios.

Core Technologies

• MolPK Deep Learning Model: Predicts pharmacokinetic parameters based on molecular representation learning
• Multi-format Input: Supports SMILES text, CSV tables, and SDF structure files
• Flexible Batch Processing: Supports both single and batch prediction with automatic CSV column detection
• Dual-format Output: Simultaneously outputs CSV result tables and SDF structure files with PK attributes

Use Cases

• Early drug screening: Rapidly evaluate pharmacokinetic properties of candidate molecules
• Batch prediction: High-throughput PK parameter prediction for large compound libraries
• Experimental design support: Provides reference for dose and species selection in in vivo studies

Parameters

Input

Input file for prediction, supporting .smi (SMILES text), .csv (table), or .sdf (structure file) formats. This is a required parameter.

Species

Experimental species, with options rat, mou (mouse), dog, or hum (human). Specifies the species background for PK prediction.

Route

Administration route, with options iv (intravenous) or po (oral). Different routes significantly affect PK profiles.

Dose

Administration dose in mg/kg. Specifies the dose condition for prediction.

Output

Output CSV path for prediction results. Default: pred_pk_value.csv.

Output SDF

Output SDF structure file path with PK prediction attributes. Default: pred_with_pk_value.sdf.

SMILES Column

When input is a CSV file, specifies the column name containing SMILES strings.

Species Column

When input is a CSV file, specifies the column name containing species information.

Route Column

When input is a CSV file, specifies the column name containing administration route information.

Dose Column

When input is a CSV file, specifies the column name containing dose information.

Results

The output includes the following files:

The predicted results are output to pred_pk_value.csv:

Protein Evolution

简介

蛋白进化分析，快速找到能够协同作用的多重突变组合。基于MULTI-evolve框架实现，面向蛋白工程中的候选突变发现与组合优化，提供单点突变与多点突变两种工作模式：前者利用蛋白语言模型进行零样本评估，快速发现潜在有利的单点突变；后者基于实验测得的突变数据训练监督模型，并在候选突变池上进一步搜索高阶组合突变。该流程将蛋白语言模型、表观互作（epistasis）建模和后续实验构建衔接为一套端到端方案；其中单点突变部分实际整合了 5 个 ESM-1v 模型(esm1v_t33_650M_UR90S_1-5)、1 个 ESM-2 3B 模型(esm2_t36_3B_UR50D)，以及结构感知的 ESM-IF1，多点突变部分则以全连接神经网络为核心预测器来学习序列与性质之间的映射。

使用流程：
1，计算步骤，先使用单点突变模式，获取优势单点突变（一般选择排名靠前的15-20个）
2，湿实验步骤，对第一步选择的单点突变，及其所有两点突变的组合（100~200个组合），进行湿实验验证，获取突变对应的湿实验数据，请使用性质数据的比值（Fold-Change，FC值），即： 突变后的性质/野生型的性质。
3，计算步骤，使用多点突变模式，输入第二步的湿实验结果，进行模型训练，并预测多点突变组合对应的FC值，给出推荐的优势多点突变组合。

参数说明

Single Point Mutation

利用多个蛋白语言模型对蛋白单点突变的潜在效应进行突变概率预测，帮助研究者高效筛选更有希望进入后续实验验证的候选单点突变。模块并行提供 4 种筛选策略：ESM、ESM-IF、ESM-z 和 ESM-IF-z。突变位置从1开始按残基顺序编号。

• 在 ESM 筛选中，每个 ESM 序列子模型都会在野生型序列背景下，分别计算目标位点上突变氨基酸与野生型氨基酸的条件概率，取对数后作差，得到该子模型对该单点突变的原始分数；随后再对所有序列子模型的分数取平均，作为最终的 ESM 综合得分。
• 在 ESM-IF 筛选中，模型会结合输入的蛋白结构信息，对每个结构分别计算目标位点上突变残基与野生型残基的结构条件打分，并以两者差值作为该结构下的原始分数；当输入多个结构文件或多个构象时，再对各结构得到的分数取平均，作为最终的 ESM-IF 综合得分。
• ESM-z 和 ESM-IF-z 则是在对应原始得分的基础上，进一步进行 z-score 标准化处理，使不同突变位置之间的分数更便于横向比较与排序。
  注意： z-score指的是一种标准化方法，模块提供了两种标准化方法，由Normalization控制。

Structure

输入蛋白结构文件，支持 PDB 或 CIF 格式，用于结构模型评分。支持输入同一结构的批量构象（需压缩文件格式，支持：.zip，.tar， .tar.gz， .tgz，.tar.bz2， .tbz2，.tar.xz， .txz），模块会分别计算每个构象中的突变评分，再取不同构象的平均值，以降低单一构象带来的偏差。

Chain

指定链名，进行单点突变推荐，多链时用逗号分隔，如A,B。如果不指定该参数，则对结构中的每条链都会进行单点突变推荐。

TopN

设置每种集成方法对每条链推荐的候选单点突变数量，默认20。

Positions Excluded

需排除的突变位点的位置，使用链名+残基位置编号(从1开始按顺序)，如：A100表示A链中位置顺序编号100的残基进行排除。多位置时使用逗号分隔，支持范围符号，例如：A10-20,A25,B30-36,B40表示：排除A链编号10至20、25的残基，B链编号30至36、40的残基`。

Normalization

z-score 标准化的分组方式，可选 aa_substitution_type 和 aa_mutation，默认为：aa_substitution_type。
两种方法说明如下：

aa_substitution_type ：按具体替换类型分组标准化。例如所有突变位置中， A→L的突变单独作为一组（如：A10L，A35L，A128L），所有G→V的突变为另一组；该方式更关注“从哪种氨基酸变成哪种氨基酸”。
aa_mutation ： 按突变后的目标氨基酸分组标准化。例如 A10P、G25P、L80P 都会归到 P 这一组；该方式更关注“最终变成了什么氨基酸”。

Single Point Mutation File

指定输出结果csv文件的名称。默认：SP_Mutation.csv

Multiple Point Mutations

基于实验数据训练预测模型，对候选突变进行自动筛选与组合，生成可用于实验验证的优势多点突变方案。该模式的典型使用场景是：先通过单点或双点突变实验获得一定规模的功能数据，再训练模型预测更高阶组合突变（通常为 >=3 位点）的潜在表现。

Structure

输入蛋白结构文件，支持 PDB 或 CIF 格式。

Training Data

输入.csv格式文件，CSV必须包含以下列：
mutation ：指定结构中的突变信息，使用原始残基+链名+残基位置编号(从1开始按顺序)+突变后的残基，如：KA100N表示A链中位置顺序编号100的残基K，突变为N。多点突变时用分号分隔，如：GA48R;DB106A
property ： 突变对应的性质变化倍数，即性质数据的比值（Fold-Change，FC值），即： 突变后的性质/野生型的性质。
注意：
1.突变样本数量需要大于20条
2.模块会对输入内容进行检查；若存在数据错误，请查看 stderr.txt。

Single Mutations

用于进行多点组合突变的单点突变文件，同样使用原始残基+链名+残基位置编号(从1开始按顺序)+突变后的残基，输入格式如下：

TA192V
TB192K
AC167R
NA72A

注意：
1.如果不指定该参数，默认会将训练数据中的所有单点突变，进行组合，然后预测推荐。
2.模块会对输入内容进行检查；若存在数据错误，请查看 stderr.txt。

Top variants

指定为每类组合突变推荐的TopN数量，默认为：3，即：三点组合突变推荐3个，四点组合突变推荐3个，五点组合突变推荐3个，…，最多推荐十点组合突变。

Mutiple Point Mutation File

指定输出结果csv文件的名称。默认：MP_Mutation.csv

结果说明

单点突变模式下，结果输出SP_Mutation.csv，内容如下：

说明：

• 模块基于ESM、ESM-IF、ESM-z 和 ESM-IF-z 4 种推荐方法对饱和单点突变进行筛选，每种推荐方法均按照对应的打分规则对候选突变进行排序，并依次选取前TopN个且位点不重复的突变作为推荐结果；被推荐的突变在对应列中记为1，未被推荐则记0

在多点突变模式下，结果输出MP_Mutation.csv，结果内容如下：

说明：

同时，输出 MP_Mutation.tar.gz，其中包含最终合并结果 CSV。压缩包内包含以下文件：

• MP_Mutation.csv
• MP_Mutation_all.csv

其中，MP_Mutation_all.csv 为全部筛选变体的完整结果文件。

参考文献

• Vincent Q. Tran et al. ,Rapid directed evolution guided by protein language models and epistatic interactions.DOI:10.1126/science.aea1820

Protein Evolution

Introduction

Protein evolution analysis for rapidly identifying synergistic multi-site mutation combinations, based on the MULTI-evolve framework. This module is designed for candidate mutation discovery and combinatorial optimization in protein engineering. It provides two working modes: single-point mutation and multi-point mutation.

The single-point mutation mode uses protein language models for zero-shot evaluation to rapidly identify potentially beneficial single mutations. The multi-point mutation mode trains a supervised model using experimentally measured mutation data and further searches for higher-order combinatorial mutations within the candidate mutation pool.

This workflow integrates protein language models, epistasis modeling, and experimental validation into an end-to-end pipeline. The single-point mutation module integrates five ESM-1v models (esm1v_t33_650M_UR90S_1-5), one ESM-2 3B model (esm2_t36_3B_UR50D), and structure-aware ESM-IF1. The multi-point mutation module uses a fully connected neural network as the core predictor to learn the mapping between sequence and functional properties.

Workflow:

1. Computational step: Use single-point mutation mode to obtain advantageous single mutations (typically selecting the top 15–20 candidates).
2. Experimental step: Perform wet-lab validation on selected single mutations and their pairwise combinations (100–200 variants). Measure experimental properties and compute Fold-Change (FC), defined as: mutant property / wild-type property.
3. Computational step: Use multi-point mutation mode with experimental data from step 2 to train the model and predict FC values for higher-order mutation combinations, then identify optimal multi-point variants.

Parameters

Single Point Mutation

This module uses multiple protein language models to predict the potential effects of single-point mutations, enabling efficient screening of promising candidates for experimental validation. Four screening strategies are provided: ESM, ESM-IF, ESM-z, and ESM-IF-z. Residue indexing starts from 1.

• In the ESM strategy, each ESM sub-model computes the conditional probability difference between the mutant amino acid and the wild-type amino acid at the target position under the wild-type sequence background. The log-probability difference is used as the raw score for each sub-model, and the final ESM score is obtained by averaging across all sub-models.
• In the ESM-IF strategy, structure information is incorporated. For each structure, a structural conditional score is computed for the mutation and wild type at the target position. The difference is used as the raw score. If multiple structures or conformations are provided, the final ESM-IF score is the average across all structures.
• ESM-z and ESM-IF-z apply z-score normalization to the corresponding raw scores, enabling better comparison and ranking across mutation sites.

Note: Z-score refers to a standardization method. Two normalization strategies are supported and controlled by the Normalization parameter.

Structure

Input protein structure file in PDB or CIF format for structure-based scoring. Multiple conformations of the same structure are supported (compressed formats: .zip, .tar, .tar.gz, .tgz, .tar.bz2, .tbz2, .tar.xz, .txz). Scores are averaged across conformations to reduce bias from a single structure.

Chain

Specify chain IDs for single-point mutation recommendation. Multiple chains are separated by commas (e.g., A,B). If not specified, all chains in the structure will be analyzed.

TopN

Number of candidate single-point mutations recommended per chain for each integrated method. Default: 20.

Positions Excluded

Residues to exclude from mutation analysis. Format: Chain + residue index (starting from 1), e.g., A100. Multiple positions can be separated by commas and ranges are supported, e.g., A10-20,A25,B30-36,B40.

Normalization

Defines the grouping strategy for z-score normalization. Options: aa_substitution_type (default) and aa_mutation.

• aa_substitution_type: Groups mutations by substitution type (e.g., A→L, G→V). Focuses on “which amino acid is replaced by which”.
• aa_mutation: Groups mutations by the resulting amino acid. Focuses on “what amino acid it becomes”.

Single Point Mutation File

Output CSV file name for single-point mutation results. Default: SP_Mutation.csv.

Multi-point Mutation

This module trains predictive models using experimental data to automatically screen and combine mutations, generating high-order mutation designs for experimental validation. A typical workflow involves generating experimental data from single- or double-point mutations, then training a model to predict higher-order combinations (≥3 sites).

Structure

Input protein structure file in PDB or CIF format.

Training Data

Input .csv file containing the following required columns:

• mutation: Mutation information in the format WildTypeResidue + Chain + Position + MutantResidue, e.g., KA100N. For multi-point mutations, use semicolons, e.g., GA48R;DB106A.
• property: Experimental fold-change (FC), defined as: mutant property / wild-type property.

Note:

1. The number of mutation samples must be greater than 20.
2. The module will check the input content; if there are data errors, please refer to stderr.txt.

Single Mutations

Single-point mutation file used for combinatorial generation. Format:

TA192V
TB192K
AC167R
NA72A

Notes:

1. If not specified, all single mutations in the training set will be used for combination.
2. Input is validated; check stderr.txt for errors.

Top variants

Number of top-ranked variants returned per mutation order. Default: 3 (e.g., top 3 for triple, quadruple, etc., up to decuple mutations).

Multiple Point Mutation File

Output CSV file name. Default: MP_Mutationcsv.

Results

Single-point Mutation Output (SP_Mutation.csv)

Field Description:

• Chain ID: Chain where mutation is recommended
• Mutations: Mutation in format WildTypeResidue + Position + MutantResidue (index starts from 1), e.g., F26L
• ESM / ESM-IF / ESM-z / ESM-IF-z: Whether selected by each method (1 = yes, 0 = no)
• Count: Number of methods that selected the mutation

Each method ranks candidates independently and selects top-N non-redundant mutations.

Multi-point Mutation Output (MP_Mutation.csv)

Field Description:

• Variant ID: Variant index (consistent with all file)
• Chain ID: Chains with mutations; multiple chains separated by semicolon
• Mutations Number: Number of mutations per chain
• Mutations: Mutation list; within-chain separated by /, between chains by ;
• Sequence: Full amino acid sequence; multiple chains separated by :
• Average: Mean predicted score; higher indicates better predicted performance

The output package MP_Mutation.tar.gz contains:

• MP_Mutation.csv
• MP_Mutation_all.csv (complete results)

References

• Vincent Q. Tran et al. ,Rapid directed evolution guided by protein language models and epistatic interactions.DOI:10.1126/science.aea1820

Enzyme pH Optimum Prediction (EpHod)

简介

EpHod 是一个基于机器学习的酶最适 pH（pHopt）预测工具，旨在从氨基酸序列直接预测酶的最适工作 pH 值。

核心思想是通过蛋白质语言模型 ESM1v 提取酶序列特征，结合残差注意力机制（RLAT）和支持向量回归（SVR）进行集成预测。模型直接从序列数据中学习与 pHopt 相关的结构和生物物理特征，包括残基与催化中心的距离、溶剂分子可及性等。

参数说明

Enzyme Fasta File

输入的酶序列 FASTA 文件路径，必选项

FASTA 文件每条序列以 > 开头，格式示例：

>Q2YPV0 | Brucella abortus | 4.2.1.11 | 8.5 | 0.366
MTAIIDIVGREILDSRGNPTVEVDVVLEDGSFGRAAVPSGASTGAHEAVELRDGGSRYLGKGVEKAVEVVNGKIFDAIAGMDAESQLLIDQTLIDLDGSANKGNLGANAILGVSLAVAKAAAQASGLPLYRYVGGTNAHVLPVPMMNIINGGAHADNPIDFQEFMILPVGATSIREAVRYGSEVFHTLKKRLKDAGHNTNVGDEGGFAPNLKNAQAALDFIMESIEKAGFKPGEDIALGLDCAATEFFKDGNYVYEGERKTRDPKAQAKYLAKLASDYPIVTIEDGMAEDDWEGWKYLTDLIGNKCQLVGDDLFVTNSARLRDGIRLGVANSILVKVNQIGSLSETLDAVETAHKAGYTAVMSHRSGETEDSTIADLAVATNCGQIKTGSLARSDRTAYNQLIRIEEELGKQARYAGRSALKLL

Predicted Results

输出预测结果文件名，默认为 prediction.csv

结果说明

预测结果为 CSV 文件，包含以下列：

如何理解结果

1. Ensemble 为推荐使用的预测值，综合了注意力机制和回归模型的优势
2. 预测值为数值型 pH 值，代表酶的最适工作 pH
3. 建议结合预测置信度（序列头部的 pLDDT 值）综合判断结果可靠性
4. 序列长度超过 1022 残基时会被截断处理

参考文献

• Gado J E, Knotts M, Shaw A Y, et al. Machine learning prediction of enzyme optimum pH[J]. Nature Machine Intelligence, 2025, 7(5): 716-729. DOI: 10.1038/s42256-025-01026-6

Enzyme pH Optimum Prediction (EpHod)

Introduction

EpHod is a machine learning tool for predicting enzyme optimum pH (pHopt) directly from amino acid sequences.

The core approach uses the protein language model ESM1v to extract enzyme sequence features, combined with Residual Light Attention (RLAT) and Support Vector Regression (SVR) for ensemble prediction. The model learns structural and biophysical features directly from sequence data that relate to pHopt, including residue proximity to catalytic centers and solvent accessibility.

Parameters

Enzyme Fasta File

Path to input enzyme sequence FASTA file, required
Each sequence in the FASTA file starts with >, example format:

>Q2YPV0 | Brucella abortus | 4.2.1.11 | 8.5 | 0.366
MTAIIDIVGREILDSRGNPTVEVDVVLEDGSFGRAAVPSGASTGAHEAVELRDGGSRYLGKGVEKAVEVVNGKIFDAIAGMDAESQLLIDQTLIDLDGSANKGNLGANAILGVSLAVAKAAAQASGLPLYRYVGGTNAHVLPVPMMNIINGGAHADNPIDFQEFMILPVGATSIREAVRYGSEVFHTLKKRLKDAGHNTNVGDEGGFAPNLKNAQAALDFIMESIEKAGFKPGEDIALGLDCAATEFFKDGNYVYEGERKTRDPKAQAKYLAKLASDYPIVTIEDGMAEDDWEGWKYLTDLIGNKCQLVGDDLFVTNSARLRDGIRLGVANSILVKVNQIGSLSETLDAVETAHKAGYTAVMSHRSGETEDSTIADLAVATNCGQIKTGSLARSDRTAYNQLIRIEEELGKQARYAGRSALKLL

Predicted Results

Output prediction result filename, default prediction.csv

Results

Prediction result is a CSV file with the following columns:

How to Interpret Results

1. Ensemble is the recommended prediction value, combining attention mechanism and regression model advantages
2. The predicted value is a numeric pH value representing the enzyme’s optimum working pH
3. It is recommended to consider prediction confidence (pLDDT value in sequence header) when evaluating results
4. Sequences longer than 1022 residues will be truncated

Reference

• Gado J E, Knotts M, Shaw A Y, et al. Machine learning prediction of enzyme optimum pH[J]. Nature Machine Intelligence, 2025, 7(5): 716-729. DOI: 10.1038/s42256-025-01026-6

Enzyme-Substrate Prediction (ESP)

简介

ESP (Enzyme-Substrate Prediction) 是一个用于预测酶-底物反应活性的机器学习工具，旨在为实验筛选提供优先级排序。
它要解决的问题是：在候选组合数量较大时，如何优先挑出更可能发生反应的酶-底物对，从而降低实验试错成本。
ESP 的核心思想是联合利用两类信息：

1. 代谢物表征：通过 GNN（图神经网络）提取底物的分子特征。
2. 酶表征：通过 ESM1b（蛋白语言模型）提取酶的序列特征。
   最终将两类表征拼接后，使用 XGBoost 模型预测反应活性分数。

参数说明

Enzyme–Substrate Pair

输入的底物-酶对列表文件，支持 .csv、.xlsx、.xls 格式，必选项
文件应包含两列：substrate 和 enzyme

substrate,enzyme
C00069,MARLPFYLLVISTLLLVVTADSFLARPPSSSFLHALSNKRASTPASLPSCSLDFLLQTRGGTAANAATTALPTSALVERKGGAAVALEGGKTLWEKSKVWVFIGLWYFFNVAFNIYNKKVLNALPLPWTVSIAQLGLGALYTMFLWLVRARKMPTIAAPEMKTLSILGVLHAVSHITAITSLGAGAVSFTHIVKSAEPFFSAVFAGLFFGQFFSLPVYAALIPVVSGVAYASLKELTFTWLSFWCAMASNVVCAARGVVVKGMMGGKPTQSKDLTSSNMYSVLTILAALVLLPFGALVEGPGLHAAWKAAAAHPSLTNGGTELAKYLVYSGLTFFLYNEVAFAALESLHPISHAVANTIKRVVIIVVSVLVFRNPMSTQSIIGSSTAVIGVLLYSLAKHYCK

Substrate Column

底物列的列名，默认为 substrate

Enzyme Column

酶列的列名，默认为 enzyme

Predicted Results

输出结果文件名，默认为 predictions.csv

结果说明

输出文件

输出的结果文件，CSV 格式，包含以下列：

如何理解结果

1. Prediction：预测分数，值越高表示该酶-底物对越可能发生反应
2. 该分数本质是概率型分数，不应直接等同于"反应一定发生"
3. 建议将输出用于候选排序，按分数从高到低组织实验顺序（Top-K 优先）
4. metabolite_similarity_score 反映底物与训练集的相似程度，可作为预测可信度的参考

注意事项

1. 底物输入格式：支持 KEGG Compound ID（如 C00069）、SMILES 或 InChI 格式
2. 输入质量影响显著：底物 ID 合法性、SMILES 有效性会直接影响预测稳定性
3. 结果使用原则：建议在同一任务上下文内做相对比较，不建议跨任务直接比较绝对分数阈值

参考文献

• Kroll A, Ranjan S, Engqvist M K M, Lercher M J. A general model to predict small molecule substrates of enzymes based on machine and deep learning[J]. Nature Communications, 2023, 14(1): 2787. DOI:10.1038/s41467-023-38347-2

Enzyme-Substrate Prediction (ESP)

Introduction

ESP (Enzyme-Substrate Prediction) is a machine learning tool for predicting enzyme-substrate reaction activity, designed to provide priority ranking for experimental screening.
It addresses the problem: when the number of candidate combinations is large, how to prioritize enzyme-substrate pairs that are more likely to react, thereby reducing experimental trial-and-error costs.
The core idea of ESP is to jointly utilize two types of information:

1. Metabolite representation: Extract substrate molecular features through GNN (Graph Neural Network)
2. Enzyme representation: Extract enzyme sequence features through ESM1b (protein language model)
   Finally, the two types of representations are concatenated and used with XGBoost model to predict reaction activity scores.

Parameters

Input File (CSV/Excel)

Input file containing substrate-enzyme pairs, supports .csv, .xlsx, .xls format, required

The file should contain two columns: substrate and enzyme

substrate,enzyme
C00069,MARLPFYLLVISTLLLVVTADSFLARPPSSSFLHALSNKRASTPASLPSCSLDFLLQTRGGTAANAATTALPTSALVERKGGAAVALEGGKTLWEKSKVWVFIGLWYFFNVAFNIYNKKVLNALPLPWTVSIAQLGLGALYTMFLWLVRARKMPTIAAPEMKTLSILGVLHAVSHITAITSLGAGAVSFTHIVKSAEPFFSAVFAGLFFGQFFSLPVYAALIPVVSGVAYASLKELTFTWLSFWCAMASNVVCAARGVVVKGMMGGKPTQSKDLTSSNMYSVLTILAALVLLPFGALVEGPGLHAAWKAAAAHPSLTNGGTELAKYLVYSGLTFFLYNEVAFAALESLHPISHAVANTIKRVVIIVVSVLVFRNPMSTQSIIGSSTAVIGVLLYSLAKHYCK
C00002,MKGRRRRRREYCKFALLLVLYTLVLLLVPSVLDGGRDGDKGAEHCPGLQRSLGVWSLEAAAAGEREQGAEARAAEEGGANQSPRFPSNLSGAVGEAVSREKQHIYVHATWRTGSSFLGELFNQHPDVFYLYEPMWHLWQALYPGDAESLQGALRDMLRSLFRCDFSVLRLYAPPGDPAARAPDTANLTTAALFRWRTNKVICSPPLCPGAPRARAEVGLVEDTACERSCPPVAIRALEAECRKYPVVVIKDVRLLDLGVLVPLLRDPGLNLKVVQLFRDPRAVHNSRLKSRQGLLRESIQVLRTRQRGDRFHRVLLAHGVGARPGGQSRALPAAPRADFFLTGALEVICEAWLRDLLFARGAPAWLRRRYLRLRYEDLVRQPRAQLRRLLRFSGLRALAALDAFALNMTRGAAYGADRPFHLSARDAREAVHAWRERLSREQVRQVEAACAPAMRLLAYPRSGEEGDAEQPREGETPLEMDADGAT

Substrate Column

Column name for substrate, default substrate

Enzyme Column

Column name for enzyme, default enzyme

Output File

Output result filename, default predictions.csv

Results

Output File

Output result file in CSV format, containing the following columns:

How to Interpret Results

1. Prediction: Prediction score, higher values indicate the enzyme-substrate pair is more likely to react
2. This score is essentially a probability-like score and should not be directly equated with “reaction will definitely occur”
3. It is recommended to use the output for candidate ranking, organizing experimental order from high to low scores (Top-K priority)
4. metabolite_similarity_score reflects the similarity between substrate and training set, which can be used as a reference for prediction reliability

Notes

1. Substrate input format: Supports KEGG Compound ID (e.g., C00069), SMILES, or InChI format
2. Input quality significantly affects results: Substrate ID validity and SMILES validity directly affect prediction stability
3. Result usage principle: It is recommended to make relative comparisons within the same task context, and avoid directly comparing absolute score thresholds across tasks

Reference

• Kroll A, Ranjan S, Engqvist M K M, Lercher M J. A general model to predict small molecule substrates of enzymes based on machine and deep learning[J]. Nature Communications, 2023, 14(1): 2787. DOI: 10.1038/s41467-023-38347-2

Catalytic Optimum Predictor (CatOpt)

简介

Catalytic Optimum Predictor (CatOpt） 是一个基于深度学习的酶催化剂特性预测工具，用于从蛋白质序列预测酶的最适 pH 、最适温度和热变性温度。

CatOpt 的核心思想是利用蛋白质语言模型 ESM2 提取酶序列的高维特征表征，结合多头自注意力机制的多尺度卷积神经网络，实现高精度的酶催化特性预测。

参数说明

参数说明

Input Dataset

输入数据集路径，CSV格式
输入文件应包含 sequence 列，每行为蛋白质的氨基酸序列。

sequence
MARLPFYLLVISTLLLVVTADSFLARPPSSSFLHALSNKRASTPASLPSCSLDFLLQTRGGTAANAATTALPTSALVERKGGAAVALEGGKTLWEKSKVWVFIGLWYFFNVAFNIYNKKVLNALPLPWTVSIAQLGLGALYTMFLWLVRARKMPTIAAPEMKTLSILGVLHAVSHITAITSLGAGAVSFTHIVKSAEPFFSAVFAGLFFGQFFSLPVYAALIPVVSGVAYASLKELTFTWLSFWCAMASNVVCAARGVVVKGMMGGKPTQSKDLTSSNMYSVLTILAALVLLPFGALVEGPGLHAAWKAAAAHPSLTNGGTELAKYLVYSGLTFFLYNEVAFAALESLHPISHAVANTIKRVVIIVVSVLVFRNPMSTQSIIGSSTAVIGVLLYSLAKHYCK
MKGRRRRRREYCKFALLLVLYTLVLLLVPSVLDGGRDGDKGAEHCPGLQRSLGVWSLEAAAAGEREQGAEARAAEEGGANQSPRFPSNLSGAVGEAVSREKQHIYVHATWRTGSSFLGELFNQHPDVFYLYEPMWHLWQALYPGDAESLQGALRDMLRSLFRCDFSVLRLYAPPGDPAARAPDTANLTTAALFRWRTNKVICSPPLCPGAPRARAEVGLVEDTACERSCPPVAIRALEAECRKYPVVVIKDVRLLDLGVLVPLLRDPGLNLKVVQLFRDPRAVHNSRLKSRQGLLRESIQVLRTRQRGDRFHRVLLAHGVGARPGGQSRALPAAPRADFFLTGALEVICEAWLRDLLFARGAPAWLRRRYLRLRYEDLVRQPRAQLRRLLRFSGLRALAALDAFALNMTRGAAYGADRPFHLSARDAREAVHAWRERLSREQVRQVEAACAPAMRLLAYPRSGEEGDAEQPREGETPLEMDADGAT

Task

预测任务类型：pHopt（最适 pH）、topt（最适温度）、tm（热变性温度）

Output Path

输出结果文件路径,默认为prediction_results.csv

结果说明

输出文件

CSV 格式，包含以下列：

预测值范围

注意事项

1. 输入格式：仅支持标准氨基酸字母序列，不含特殊字符或未知氨基酸（X 除外）

参考文献

• Qiu S, Wang N K, Lu Y, et al. Deep Learning-Based Prediction of Enzyme Optimal pH and Design of Point Mutations to Improve Acid Resistance[J]. ACS Synthetic Biology, 2025, 14(12): 4897-4906.DOI: 10.1021/acssynbio.5c00679

Catalytic Optimum Predictor (CatOpt)

Introduction

CatOpt is a deep learning-based tool for predicting enzyme catalytic properties, including optimal pH, optimal temperature, and melting temperature from protein sequences.

The core idea of CatOpt is to leverage the ESM2 protein language model to extract high-dimensional sequence features, combined with a multi-scale convolutional neural network with multi-head self-attention mechanism, achieving high-precision enzyme catalytic property prediction.

Parameters

Input Dataset

Path to the input dataset in CSV format.
The input file must contain a sequence column, with each row representing a protein amino acid sequence.

sequence
MARLPFYLLVISTLLLVVTADSFLARPPSSSFLHALSNKRASTPASLPSCSLDFLLQTRGGTAANAATTALPTSALVERKGGAAVALEGGKTLWEKSKVWVFIGLWYFFNVAFNIYNKKVLNALPLPWTVSIAQLGLGALYTMFLWLVRARKMPTIAAPEMKTLSILGVLHAVSHITAITSLGAGAVSFTHIVKSAEPFFSAVFAGLFFGQFFSLPVYAALIPVVSGVAYASLKELTFTWLSFWCAMASNVVCAARGVVVKGMMGGKPTQSKDLTSSNMYSVLTILAALVLLPFGALVEGPGLHAAWKAAAAHPSLTNGGTELAKYLVYSGLTFFLYNEVAFAALESLHPISHAVANTIKRVVIIVVSVLVFRNPMSTQSIIGSSTAVIGVLLYSLAKHYCK
MKGRRRRRREYCKFALLLVLYTLVLLLVPSVLDGGRDGDKGAEHCPGLQRSLGVWSLEAAAAGEREQGAEARAAEEGGANQSPRFPSNLSGAVGEAVSREKQHIYVHATWRTGSSFLGELFNQHPDVFYLYEPMWHLWQALYPGDAESLQGALRDMLRSLFRCDFSVLRLYAPPGDPAARAPDTANLTTAALFRWRTNKVICSPPLCPGAPRARAEVGLVEDTACERSCPPVAIRALEAECRKYPVVVIKDVRLLDLGVLVPLLRDPGLNLKVVQLFRDPRAVHNSRLKSRQGLLRESIQVLRTRQRGDRFHRVLLAHGVGARPGGQSRALPAAPRADFFLTGALEVICEAWLRDLLFARGAPAWLRRRYLRLRYEDLVRQPRAQLRRLLRFSGLRALAALDAFALNMTRGAAYGADRPFHLSARDAREAVHAWRERLSREQVRQVEAACAPAMRLLAYPRSGEEGDAEQPREGETPLEMDADGAT

Task

Prediction task type: pHopt (optimal pH), topt (optimal temperature), tm (melting temperature)

Output Path

Path to the output results file. Default: prediction_results.csv

Results

Output File

CSV format with the following columns:

Prediction Range

Notes

1. Input Format: Only standard amino acid letter sequences are supported; no special characters or unknown amino acids (except X)

Reference

• Qiu S, Wang N K, Lu Y, et al. Deep Learning-Based Prediction of Enzyme Optimal pH and Design of Point Mutations to Improve Acid Resistance[J]. ACS Synthetic Biology, 2025, 14(12): 4897-4906.DOI: 10.1021/acssynbio.5c00679

Protein Contacts Profile

简介

对结构预测模型（如：Boltz2/Protenix/AF3等）预测的一组蛋白单体或复合物结构进行全面分析，包括：二级结构、溶剂可及性、疏水性、残基接触、结构置信度等等，对分析结果进行统一整理和对比展示。

核心思路：以参考链为分析目标，将二级结构、溶剂可及性、疏水性、残基接触、模型置信度，以及可选的同源序列和保守性信息等等汇总到同一份报告中。

模块工作流 ：

• Gemmi：用于读取 PDB/mmCIF 结构，并从查询模型中规范化链、残基、原子和实体类型。
• DSSP：通过提取每个残基的二级结构和溶剂可及性。
• 二硫键计算模块：根据模型坐标计算以残基为中心的接触注释，并推断二硫键。
• Kyte-Doolittle 疏水性评分：逐一计算残基疏水性。
• 类 pLDDT 的置信度提取：当输入模型将置信度编码在原子的 B-factor字段中时，提取残基置信度数值。
• BLAST+：在配置好的本地序列数据库中搜索同源蛋白序列。
• Clustal Omega：对查询序列和命中序列进行比对，进行保守性分析。
• 相互作用类型：对残基对相互作用进行跨结构统计，生成CSV文件，比较不同结构之间相互作用出现的频率与类型。
• 统计汇总：基于上述分析结果，汇总输出多个CSV文件。
• HTML 渲染：将上述分析结果渲染为交互式 HTML 报告和按链导出的 PDF。

相互作用类型判断的阈值：

参数说明

Structures

输入蛋白结构，允许单个以及批量输出。单结构输入，支持 .pdb、.cif、.mmcif 结构格式批量输入需要以压缩包的形式，支持：.zip、.tar、.tar.gz、.tgz、.tar.bz2、.tbz2、.tar.xz、.txz，批量输入最大支持100个结构。

Sequence database

指定通过BLAST+搜索筛选的序列数据库。可选：SwissProt（UniProt知识库中的Swiss-Prot数据库）和 PDBAA（从实验确定的三维结构数据库PDB衍生的序列）。

E-value

指定BLAST+搜索识别序列匹配的保留阈值。E值表示给定序列比对的统计显著性。E值越低（越接近零），匹配的显著性越高，可选阈值：
1e-4、1e-5、1e-6、1e-7、1e-8、1e-9、1e-10、1e-11、1e-12。默认：1e-6

Homolog Display Limit

设置在序列搜索时，保留的最大序列数，可选数量：1~25。默认：5

Columns

指定一行序列展示的残基数量。默认：120

VDW

分析相互作用时，是否输出范德华相互作用，默认不输出。

结果说明

输出protein_contacts_profile.html，内容展示如下:
注意: 若输入超过80个结构，HTML文档有可能过大而导致浏览器无法正常显示。

说明：
A. 摘要
在图的最上方，是当前活动面板的摘要区。

• Reference Chain 表示当前正在查看的参考链；
• Residue Span 表示这条链在图中覆盖的残基范围；
• Model Count 表示本次一起参与比较的模型数量；
• Output Status 表示当前链面板的结果状态；
• Weak Contact Cutoff和Strong Contact Cutoff
• Hydropathy Window、E-value Threshold、Weak Contact Cutoff、Strong Contact Cutoff、Homolog Display Limit 和 Database 则对应这张图生成时使用的关键参数，说明如下：

B. 二级结构与置信度
在结构部分，每个模型都会用图形标出参考链上的二级结构单元。

• Alpha Helix、3₁₀ Helix 和 Pi Helix 都用卷曲波浪线表示，颜色不同，分别对应 α 螺旋、3₁₀ 螺旋和 π 螺旋。
• Strand 用箭头表示，箭头方向反映链段方向。
• Alpha Turn 和 Beta Turn 则表示两类更短的紧转角区域，通常出现在连接不同二级结构的局部片段中。

如果输入结构中带有逐残基置信度信息，结构轨道还会按 pLDDT 着色。

• 深蓝色表示 pLDDT >= 90，通常可以把这部分看作局部较可信的区域；
• 青蓝色表示 70 <= pLDDT < 90，整体骨架往往已经比较可靠；
• 黄色表示 50 <= pLDDT < 70，说明这一段需要更谨慎地解释；
• 橙红色表示 pLDDT < 50，这类区域往往更灵活，也更容易出现不稳定或低置信度的构象。

把“结构形态”和“置信度颜色”结合起来看：如果某个螺旋或链段在不同模型里都出现，但颜色差异明显，就说明该局部形态可能存在不确定性。

C.如果当前参考链被识别为抗体样链，还会额外显示一条 Antibody Numbering 轨道，并在界面中提供 Kabat、IMGT、Chothia 三种编号方案切换。

• 这条轨道主要用来标出抗体可变区中的 CDR1、CDR2、CDR3 位置，帮助快速判断互补决定区落在序列的哪一段；
• 切换编号方案时，变化的是 CDR 的边界定义和标签归属，而不是原始氨基酸序列本身；
• 若参考链被识别为重链，则这些 CDR 标记对应重链可变区；若被识别为轻链，则对应轻链可变区；

如果序列并不像抗体可变区，或者当前环境无法完成抗体编号，这条轨道可能不会显示。

D. 序列与同源信息

在结构轨道下方，报告会显示参考链的序列轴。

• Query Sequence 表示目标序列本身；
• Exact Match 表示该位置与查询序列完全一致；`
• Similar Substitution 表示氨基酸虽然不同，但仍属于较保守的替换。

阅读时，它能帮助你区分“模型之间结构有差异，但序列背景本身很保守”和“这一段本来就在同源序列里变化较大”这两种情况。

E. Accessibility与Hydropathy

在序列信息下面，报告会显示两条理化性质轨道。
Accessibility 用来表示残基在结构表面的暴露程度。

• 深蓝色对应埋藏程度较高的区域;
• 灰蓝色表示中间状态;
• 浅蓝色表示更容易暴露于溶剂环境;
• 金色则提示该位置表面暴露很明显;

对于结构分析，这条轨道适合用来判断一个位点更像是核心残基、表面残基，还是潜在的界面区域。

Hydropathy 则描述序列在局部窗口中的疏水或亲水倾向。

• 橙色表示偏疏水，常见于埋藏区、界面内部或与脂质环境相容的区域；
• 灰色表示性质居中；蓝色表示偏亲水，更常出现在暴露于水相环境的位置;

这条轨道结合Accessibility轨道一起观察，能更好的判断某一段结构是否符合直觉：例如，一个明显暴露的区域如果同时又很疏水，就值得进一步留意它是否参与界面作用或是否处在特殊构象环境中。

F. 接触与符号

接触轨道用来描述参考链残基与其他链、配体或小分子的相互作用。
蛋白-蛋白接触使用字母显示，字母表示接触对方所在的链。颜色用于区分接触强度：

• 较深的接触标记对应强接触，当前阈值为最短非氢原子距离小于 3.2 A；
• 较浅的标记对应弱接触，距离位于 3.2 A 到 3.7 A 之间。
• 若同一个位点同时存在多种接触，图中还会用额外外框加以标识，提醒这一位置可能处在更复杂的界面环境里；

除了链字母，轨道里还会出现一些专门的符号。

• S 用来标记二硫键；绿色通常表示链内二硫键，青色表示链间二硫键；
• # 表示接触对方与当前位置具有相同的残基编号和残基类型，常见于对称相关的接触关系；
• 对于非蛋白组分，模块也会用符号做简写显示，例如核酸、离子、糖类、卟啉样配体或其他小分子；参考下方的符号对照表

建议先看“有没有接触”，再看“接触对象是谁”，最后再结合结构轨道判断这些接触是否稳定、是否集中出现在同一个局部区域。

符号对照表：

G. 底部图例
分成三列：

• 左侧 Structure 主要解释螺旋、折叠链和转角这些结构符号以及抗体编号；
• 中间 Tracks 主要解释序列比对、可及性、疏水性和置信度颜色；
• 右侧 Contacts 主要解释强弱接触、二硫键和各类配体符号。

接触残基对的详细信息文件contact_details.csv，示例如下：

说明：

接触残基的详细信息文件contact_residue_details.csv，示例如下：

说明：

接触残基汇总文件contact_consensus.csv，示例如下：

说明：

结构聚类信息tm_clusters.csv，示例如下：

说明：

用于聚类的相似性分数(TM_score)矩阵tm_score_matrix.csv，示例如下：

复合物中所有可能相互作用的列表cross_structure_interaction.csv，示例如下：

说明：

输出protein_contacts_profile_results.tar.gz，会包含HTML、PDF、CSV文档。

参考文献

• Robert, X., Guillon, C. and Gouet, P. (2025) FoldScript: a web server for the efficient analysis of AI-generated 3D protein models, Nucleic Acids Res., 53(W1):W277-W282, DOI: https://doi.org/10.1093/nar/gkaf326
• https://foldscript.ibcp.fr

Protein Contacts Profile

Introduction

Performs comprehensive analysis on a set of protein monomer or complex structures predicted by structure prediction models (e.g., Boltz2 / Protenix / AF3), including secondary structure, solvent accessibility, hydrophobicity, residue contacts, structural confidence, and more. Analysis results are organized and presented in a unified comparative view.

Core concept: Using the reference chain as the analysis target, the report aggregates secondary structure, solvent accessibility, hydrophobicity, residue contacts, model confidence, and optionally homologous sequences and conservation information into a single document.

Module workflow:

• Gemmi: Reads PDB/mmCIF structures and normalizes chains, residues, atoms, and entity types from the query models.
• DSSP: Extracts per-residue secondary structure and solvent accessibility.
• Disulfide bond calculator: Computes residue-centric contact annotations from model coordinates and infers disulfide bonds.
• Kyte-Doolittle hydrophobicity scoring: Computes per-residue hydrophobicity.
• pLDDT-like confidence extraction: When the input model encodes confidence in atomic B-factor fields, per-residue confidence values are extracted.
• BLAST+: Searches homologous protein sequences against the configured local sequence database.
• Clustal Omega: Aligns query and hit sequences for conservation analysis.
• Interaction types: Performs cross-structure statistics on residue-pair interactions, generating a CSV file comparing the frequency and types of interactions across different structures.
• Statistical summary: Aggregates the above analysis results into multiple CSV files.
• HTML rendering: Renders the analysis results into an interactive HTML report and per-chain PDF exports.

Interaction type thresholds:

Parameters

Structures

Input protein structures, supporting both single and batch submission. Single-structure input supports .pdb, .cif, and .mmcif formats. Batch input must be provided as a compressed archive, supporting .zip, .tar, .tar.gz, .tgz, .tar.bz2, .tbz2, .tar.xz, .txz. Maximum 100 structures per batch.

Sequence database

Specifies the sequence database for BLAST+ search filtering. Options: SwissProt (from the UniProt Knowledgebase) and PDBAA (derived from experimentally determined 3D structures in PDB).

E-value

Specifies the retention threshold for BLAST+ sequence match significance. The E-value indicates the statistical significance of a given sequence alignment. A lower E-value (closer to zero) indicates higher significance. Available thresholds: 1e-4, 1e-5, 1e-6, 1e-7, 1e-8, 1e-9, 1e-10, 1e-11, 1e-12. Default: 1e-6

Homolog Display Limit

Sets the maximum number of sequences retained during sequence search. Range: 1–25. Default: 5

Columns

Specifies the number of residues displayed per line. Default: 120

VDW

Whether to output van der Waals interactions during interaction analysis. Default: false

Results

Outputs protein_contacts_profile.html, displayed as follows:

Note: If more than 80 structures are input, the HTML document may become too large for browsers to display properly.

A. Summary

At the top of the figure is the summary area for the current active panel.

• Reference Chain: the reference chain currently being viewed;
• Residue Span: the residue range covered by this chain in the figure;
• Model Count: the number of models participating in this comparison;
• Output Status: the result status of the current chain panel;
• Weak Contact Cutoff and Strong Contact Cutoff
• Hydropathy Window, E-value Threshold, Weak Contact Cutoff, Strong Contact Cutoff, Homolog Display Limit, and Database correspond to key parameters used during figure generation, described below:

B. Secondary Structure and Confidence

In the structure section, each model marks the secondary structure elements on the reference chain.

• Alpha Helix, 3₁₀ Helix, and Pi Helix are all represented by coiled waves in different colors, corresponding to α-helix, 3₁₀-helix, and π-helix, respectively.
• Strand is represented by arrows, with arrow direction reflecting strand orientation.
• Alpha Turn and Beta Turn denote two classes of shorter tight-turn regions, typically appearing in local segments connecting different secondary structures.

If the input structures contain per-residue confidence information, the structure tracks are colored by pLDDT.

• Dark blue: pLDDT >= 90, typically regarded as locally highly reliable regions;
• Cyan: 70 <= pLDDT < 90, generally indicating a fairly reliable backbone;
• Yellow: 50 <= pLDDT < 70, suggesting this segment should be interpreted with caution;
• Orange-red: pLDDT < 50, regions that are often more flexible and prone to unstable or low-confidence conformations.

Combining “structural morphology” with “confidence color”: if a helix or strand appears across multiple models but with noticeably different colors, it indicates potential uncertainty in the local conformation.

C. Antibody Numbering

If the current reference chain is recognized as an antibody-like chain, an additional Antibody Numbering track is displayed, with toggle options for Kabat, IMGT, and Chothia numbering schemes in the interface.

• This track primarily marks the positions of CDR1, CDR2, and CDR3 in the antibody variable region, helping to quickly determine which sequence segment the complementarity-determining regions fall into;
• When switching numbering schemes, what changes are the CDR boundary definitions and label assignments, not the underlying amino acid sequence itself;
• If the reference chain is recognized as a heavy chain, the CDR labels correspond to the heavy chain variable region; if recognized as a light chain, they correspond to the light chain variable region;

If the sequence does not resemble an antibody variable region, or if antibody numbering cannot be completed in the current environment, this track may not be displayed.

D. Sequence and Homology Information

Below the structure tracks, the report displays the reference chain sequence axis.

• Query Sequence: the target sequence itself;
• Exact Match: indicates complete identity with the query sequence at this position;
• Similar Substitution: indicates that the amino acid differs but belongs to a relatively conserved substitution.

When reading, this helps distinguish between “structural differences across models but a highly conserved sequence background” and “this segment is inherently variable among homologous sequences.”

E. Accessibility and Hydropathy

Below the sequence information, the report displays two physicochemical property tracks.

Accessibility indicates the degree of surface exposure of a residue in the structure.

• Dark blue corresponds to more buried regions;
• Gray-blue indicates an intermediate state;
• Light blue indicates greater exposure to solvent;
• Gold indicates clearly high surface exposure;

For structural analysis, this track is useful for judging whether a site is more likely a core residue, a surface residue, or a potential interface region.

Hydropathy describes the hydrophobic or hydrophilic tendency of the sequence in a local window.

• Orange indicates hydrophobic bias, commonly found in buried regions, internal interfaces, or lipid-compatible environments;
• Gray indicates intermediate properties; blue indicates hydrophilic bias, more frequently appearing in positions exposed to aqueous environments;

Observing this track together with the Accessibility track can better help determine whether a structural segment matches intuition: for example, a clearly exposed region that is also highly hydrophobic warrants further attention to whether it participates in interface interactions or resides in a special conformational environment.

F. Contacts and Symbols

Contact tracks describe interactions between reference chain residues and other chains, ligands, or small molecules.

Protein-protein contacts are displayed as letters indicating the chain of the contact partner. Colors distinguish contact strength:

• Darker contact markers correspond to strong contacts, currently defined as shortest non-hydrogen atom distance < 3.2 Å;
• Lighter markers correspond to weak contacts, with distances between 3.2 Å and 3.7 Å.
• If multiple contact types exist at the same site, additional borders are used to highlight that this position may be in a more complex interface environment.

In addition to chain letters, the track contains specialized symbols.

• S marks disulfide bonds; green typically indicates intra-chain disulfide bonds, cyan indicates inter-chain disulfide bonds;
• # indicates that the contact partner shares the same residue number and residue type as the current position, commonly seen in symmetry-related contacts;
• For non-protein components, the module uses shorthand symbols, e.g., nucleic acids, ions, sugars, porphyrin-like ligands, or other small molecules; refer to the Symbol Legend below.

Recommended reading order: first check “whether there is a contact”, then “who the contact partner is”, and finally combine with structural tracks to judge whether these contacts are stable and whether they are concentrated in the same local region.

Symbol Legend:

G. Bottom Legend

Divided into three columns:

• Left Structure: mainly explains structural symbols for helices, strands, and turns, as well as antibody numbering;
• Middle Tracks: mainly explains sequence alignment, accessibility, hydrophobicity, and confidence colors;
• Right Contacts: mainly explains strong/weak contacts, disulfide bonds, and various ligand symbols.

Contact Details

File: contact_details.csv

Example:

Field descriptions:

Contact Residue Details

File: contact_residue_details.csv

Example:

Field descriptions:

Contact Consensus

File: contact_consensus.csv

Example:

Field descriptions:

Structural Clustering Information

File: tm_clusters.csv

Example:

Field descriptions:

Similarity score (TM-score) matrix used for clustering: tm_score_matrix.csv

Example:

Cross-Structure Interaction List

File: cross_structure_interaction.csv

Example:

Field descriptions:

The output protein_contacts_profile_results.tar.gz contains HTML, PDF, and CSV documents.

Reference

• Robert, X., Guillon, C. and Gouet, P. (2025) FoldScript: a web server for the efficient analysis of AI-generated 3D protein models, Nucleic Acids Res., 53(W1):W277-W282, DOI: https://doi.org/10.1093/nar/gkaf326
• https://foldscript.ibcp.fr

MD Dipole

简介

模块能够计算出动力学轨迹体系的总偶极矩以及其波动情况。通过这些数据，可以计算出例如低介电常数介质的介电常数。对于具有净电荷的分子，其净电荷会在分子质心处进行扣除。

参数说明

Path File

MD模拟后得到的路径文件，可以在GMX MD Run (GMX2024)模块或者AlphaAutoMD (GMX2024)模块中获取。

System Group

选择需要计算的结构组别：Backbone，Protein，DNA，RNA，Complex。
可以根据PDB中小分子的名称填写组别名称。
注意：其中Complex指的是蛋白-小分子复合物体系。

Custom Resid

自定义需要计算的残基编号，连续参数可用“-”表示，不连续残基用逗号隔开，例如：1-10，15。
注意：
1.使用该参数时必须指定完整分子的残基范围，不允许截断结构或遗漏残基。
2.残基编号参考system.gro文件

Custom Atom

自定义需要计算的原子编号，连续参数可用“-”表示，不连续残基用逗号隔开，例如：1-10，15。
注意：
1.使用该参数时必须指定完整分子的残基范围，不允许截断结构或遗漏残基。
2.原子编号参考system.gro文件

Skip Time (ns)

每一帧的间隔时间（单位ns）

结果说明

输出结果包括：

其中aver.csv包括信息如下：

其中Mtot.csv包括信息如下：

MD Dipole

Introduction

This module calculates the total dipole moment of a molecular dynamics trajectory system and its fluctuations. Based on these data, properties such as the dielectric constant of low-dielectric media can be derived. For molecules with a net charge, the net charge is subtracted at the molecular center of mass before the dipole moment calculation.

Parameters

Path File

The trajectory file obtained after MD simulations. It can be generated by the GMX MD Run (GMX2024) module or the AlphaAutoMD (GMX2024) module.

System Group

Select the structural group to be included in the calculation: Backbone, Protein, DNA, RNA, or Complex.
Custom group names can also be specified based on the names of small molecules defined in the PDB file.

Note: Complex refers to a protein–small-molecule complex system.

Custom Resid

Specify custom residue indices for calculation. Continuous ranges can be denoted using “-”, and non-contiguous residues should be separated by commas, e.g., 1-10,15.

Note:

1. When using this parameter, the complete residue range of the molecule must be specified. Truncated structures or missing residues are not allowed.
2. Residue numbering should follow the system.gro file.

Custom Atom

Specify custom atom indices for calculation. Continuous ranges can be denoted using “-”, and non-contiguous atoms should be separated by commas, e.g., 1-10,15.

Note:

1. When using this parameter, the complete atom range of the molecule must be specified. Truncated structures or missing atoms are not allowed.
2. Atom numbering should follow the system.gro file.

Skip Time (ns)

Time interval between successive frames used in the calculation (unit: nanoseconds).

Results

The output results include the following files:

aver.csv File Contents

Mtot.csv File Contents

MMGBSA

简介

MMGBSA计算受体与配体之间的结合自由能，并且提供能量分解数据、结合常数（Ka）、抑制剂常数（Ki）。熵的计算采用的是张增辉教授的相互作用熵的方法，该方法直接从分子动力学模拟计算结合自由能的熵组分（相互作用熵或-TΔS），但是需要选取结构稳定部分进行计算。
本模块提供四种计算方法，其中 Trajectory 是针对进行分子动力学模拟轨迹后计算其受配体之间的结合自由能；One Structure 是针对一帧pdb结构或者对接得到的PDB结构计算受配体之间的结合自由能，MMGBSA of One Structure 计算流程中可以直接输入PDB进行计算。
Index Name 是输入受配体组别名称；Custom Name 则是输入受配体的在PDB中的残基编号。

参数说明

Trajectory方法

Path File

MD模拟后得到的路径文件，可以在MD (GMX2024)模块或者AlphaAutoMD模块中获取。

Receptor Name

受体名称，可以为Protein、DNA、RNA。

Ligand Name

配体名称，可以为Protein、DNA、RNA。如果为小分子，填写其在PDB中的名称。如果体系中除了蛋白以外为配体（包括小分子）可用Other表示。

Reference Structure (GRO)

参考结构。默认system.gro。可以在GMX MD Run (GMX2024)模块的输出结果文件中找到。当周期性边界条件处理不当时可以使用该参数。

Start Time (ps)

起始帧时间，单位ps。最好选取RMSD稳定时区进行计算，以消除结构不稳定时导致的整体熵偏大。

End Time (ps)

结束帧时间，单位ps。最好选取RMSD稳定时区进行计算，以消除结构不稳定时导致的整体熵偏大。

Skip Time (ps)

间隔时间，单位ps。

Index File

索引文件，ndx格式。当体系中存在膜结构的时候可以通过Membrane Solvation模块获取index.ndx文件。膜模拟中的受体为receptor，配体为ligand，膜为membrane。

Custom Receptor

定义受体组别进行结合能计算。组别中填写的为蛋白氨基酸或者核酸碱基的序号。例如1-213或者1-211,212-213。蛋白氨基酸或者核酸碱基序号从1开始重新编号，与初始pdb氨基酸编号无关。

Custom Ligand

定义配体组别进行结合能计算。组别中填写的为蛋白氨基酸或者核酸碱基的序号。例如1-213或者1-211,212-213。蛋白氨基酸或者核酸碱基序号从1开始重新编号，与初始pdb氨基酸编号无关。

One Structure方法

System Topology

拓扑文件，由MD Solvation模块或者Membrane Solvation模块得到。

System GRO

结构文件，.gro格式，由MD Solvation模块或者Membrane Solvation模块得到。

System ITP

体系参数压缩文件，tar.gz格式。由MD Solvation模块或者Membrane Solvation模块得到。

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.
Duan L, Liu X, Zhang JZ. Interaction Entropy: A New Paradigm for Highly Efficient and Reliable Computation of Protein-Ligand Binding Free Energy. J Am Chem Soc. 2016 May 4;138(17):5722-8.
https://github.com/Electrostatics/apbs/blob/main/docs/index.rst

MMGBSA

Introduction

MMGBSA calculates the binding free energy between a receptor and a ligand, providing energy decomposition data, binding constants (Ka), and inhibitor constants (Ki). The entropy calculation uses the method of interaction entropy proposed by Professor Zhang Zenghui, which directly calculates the entropy component of the binding free energy (interaction entropy or -TΔS) from molecular dynamics simulations, but requires selecting stable structural regions for calculation. This module provides four calculation methods. Among them, Trajectory calculates the binding free energy between the receptor and ligand after molecular dynamics simulation trajectories; One Structure calculates the binding free energy between the receptor and ligand for a single frame PDB structure or a docked PDB structure; MMGBSA of One Structure allows direct input of a PDB for calculation. Index Name is the input receptor-ligand group name, while Custom Name is the input residue numbers of the receptor-ligand in the PDB.

Parameter Description

Trajectory Method

Path File

Path file obtained after MD simulation, available in the MD (GMX2024) module or AlphaAutoMD module.

Receptor Name

Name of the receptor, can be Protein, DNA, or RNA.

Ligand Name

Name of the ligand, can be Protein, DNA, or RNA. If it is a small molecule, enter its name in the PDB. Use ‘Other’ if the system contains a ligand (including small molecules) other than proteins.

Reference Structure (GRO)

Reference structure. Default: system.gro.
This file can be found in the output results of the GMX MD Run (GMX2024) module.
Use this parameter when periodic boundary conditions are not handled properly.

Start Time (ps)

Start frame time in ps. It is best to select a stable RMSD region for calculation to eliminate large overall entropy caused by structural instability.

End Time (ps)

End frame time in ps. It is best to select a stable RMSD region for calculation to eliminate large overall entropy caused by structural instability.

Skip Time (ps)

Time interval in ps.

Index File

Index file in ndx format. When there is a membrane structure in the system, you can obtain the index.ndx file through the Membrane Solvation module. In membrane simulations, the receptor is ‘receptor’, the ligand is ‘ligand’, and the membrane is ‘membrane’.

Custom Receptor

Define receptor groups for binding energy calculation. Enter the sequence numbers of protein amino acids or nucleic acid bases in the group. For example, 1-213 or 1-211,212-213. Protein amino acid or nucleic acid base numbers start from 1 and are independent of the initial pdb amino acid numbering.

Custom Ligand

Define ligand groups for binding energy calculation. Enter the sequence numbers of protein amino acids or nucleic acid bases in the group. For example, 1-213 or 1-211,212-213. Protein amino acid or nucleic acid base numbers start from 1 and are independent of the initial pdb amino acid numbering.

One Structure Method

System Topology

Topology file obtained from the MD Solvation module or Membrane Solvation module.

System GRO

Structure file in .gro format obtained from the MD Solvation module or Membrane Solvation module.

System ITP

System parameter compression file in tar.gz format. Obtained from the MD Solvation module or Membrane Solvation module.

Result Description

The output includes:

References

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.
Duan L, Liu X, Zhang JZ. Interaction Entropy: A New Paradigm for Highly Efficient and Reliable Computation of Protein-Ligand Binding Free Energy. J Am Chem Soc. 2016 May 4;138(17):5722-8.
https://github.com/Electrostatics/apbs/blob/main/docs/index.rst

Alanine Scan (MMGBSA)

简介

Alanine Scan (MMGBSA)是计算丙氨酸突变后的结合自由能，并且提供能量分解数据、结合常数（Ka）、抑制剂常数（Ki）。熵的计算采用的是张增辉教授的相互作用熵的方法，该方法直接从分子动力学模拟计算结合自由能的熵组分（相互作用熵或-TΔS），但是需要选取结构稳定部分进行计算。
本模块提供四种计算方法，其中 Trajectory 是针对进行分子动力学模拟轨迹后计算其受配体之间的结合自由能；One Structure 是针对一帧pdb结构或者对接得到的PDB结构计算受配体之间的结合自由能，MMGBSA of One Structure 计算流程中可以直接输入PDB进行计算。
Index Name 是输入受配体组别名称；Custom Name 则是输入受配体的在PDB中的残基编号。

参数说明

Trajectory方法

Path File

MD模拟后得到的路径文件，可以在MD (GMX2024)模块或者AlphaAutoMD模块中获取。

Receptor Name

受体名称，可以为Protein、DNA、RNA。

Ligand Name

配体名称，可以为Protein、DNA、RNA。如果为小分子，填写其在PDB中的名称。如果体系中除了蛋白以外为配体（包括小分子）可用Other表示。

Reference Structure (GRO)

参考结构。默认system.gro。可以在GMX MD Run (GMX2024)模块的输出结果文件中找到。当周期性边界条件处理不当时可以使用该参数。

Mutation Residue

突变扫描为丙氨酸（ALA）的氨基酸位置。格式为‘32-34,36’。蛋白氨基酸或者核酸碱基序号从1开始重新编号，与初始pdb氨基酸编号无关。

Force File

丙氨酸扫描时使用的力场。

Start Time (ps)

起始帧时间，单位ps。最好选取RMSD稳定时区进行计算，以消除结构不稳定时导致的整体熵偏大。

End Time (ps)

结束帧时间，单位ps。最好选取RMSD稳定时区进行计算，以消除结构不稳定时导致的整体熵偏大。

Skip Time (ps)

间隔时间，单位ps。

Index File

索引文件，ndx格式。当体系中存在膜结构的时候可以通过Membrane Solvation模块获取index.ndx文件。膜模拟中的受体为receptor，配体为ligand，膜为membrane。只能是AMBER力场下构建的膜体系才能进行计算。

Custom Receptor

定义受体组别进行结合能计算。组别中填写的为蛋白氨基酸或者核酸碱基的序号。例如1-213或者1-211,212-213。蛋白氨基酸或者核酸碱基序号从1开始重新编号，与初始pdb氨基酸编号无关。

Custom Ligand

定义配体组别进行结合能计算。组别中填写的为蛋白氨基酸或者核酸碱基的序号。例如1-213或者1-211,212-213。蛋白氨基酸或者核酸碱基序号从1开始重新编号，与初始pdb氨基酸编号无关。

One Structure方法

System Topology

拓扑文件，由MD Solvation模块或者Membrane Solvation模块得到。

System GRO

结构文件，.gro格式，由MD Solvation模块或者Membrane Solvation模块得到。

System ITP

体系参数压缩文件，tar.gz格式。由MD Solvation模块或者Membrane Solvation模块得到。

结果说明

输出结果包括：

ALA_Scan_Results.csv，包含信息如下：

Ka 越大表示结合力强，Ki 越小表示抑制效果强。

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.
Duan L, Liu X, Zhang JZ. Interaction Entropy: A New Paradigm for Highly Efficient and Reliable Computation of Protein-Ligand Binding Free Energy. J Am Chem Soc. 2016 May 4;138(17):5722-8.
https://github.com/Electrostatics/apbs/blob/main/docs/index.rst

Alanine Scan (MMGBSA)

Introduction

Alanine Scan (MMGBSA) calculates the binding free energy after alanine mutations and provides energy decomposition data, binding constants (Ka), and inhibitor constants (Ki). The entropy calculation uses the method of interaction entropy proposed by Professor Zhang Zenghui, which directly calculates the entropy component of the binding free energy (interaction entropy or -TΔS) from molecular dynamics simulations, but requires selecting stable structural regions for calculation. This module provides four calculation methods. Among them, Trajectory calculates the binding free energy between the receptor and ligand after molecular dynamics simulation trajectories; One Structure calculates the binding free energy between the receptor and ligand for a single frame PDB structure or a docked PDB structure; MMGBSA of One Structure allows direct input of a PDB for calculation. Index Name is the input receptor-ligand group name, while Custom Name is the input residue numbers of the receptor-ligand in the PDB.

Parameters

Trajectory Method

Path File

Path file obtained after MD simulation, available in the MD (GMX2024) module or AlphaAutoMD module.

Receptor Name

Name of the receptor, can be Protein, DNA, or RNA.

Ligand Name

Name of the ligand, can be Protein, DNA, or RNA. If it is a small molecule, enter its name in the PDB. Use ‘Other’ if the system contains a ligand (including small molecules) other than proteins.

Reference Structure (GRO)

Reference structure. Default: system.gro.
This file can be found in the output results of the GMX MD Run (GMX2024) module.
Use this parameter when periodic boundary conditions are not handled properly.

Mutation Residue

The mutation scans for the amino acid location of alanine (ALA). Must followed the format is ‘32-34,36’. The protein amino acid or nucleic acid number is re-numbered from 1, independent of the initial pdb amino acid number.

Force File

Force field used for alanine scanning.

Start Time (ps)

Start frame time in ps. It is best to select a stable RMSD region for calculation to eliminate large overall entropy caused by structural instability.

End Time (ps)

End frame time in ps. It is best to select a stable RMSD region for calculation to eliminate large overall entropy caused by structural instability.

Skip Time (ps)

Time interval in ps.

Index File

Index file in ndx format. When there is a membrane structure in the system, you can obtain the index.ndx file through the Membrane Solvation module. In membrane simulations, the receptor is ‘receptor’, the ligand is ‘ligand’, and the membrane is ‘membrane’. Only membrane systems built under the AMBER force field can be calculated.

Custom Receptor

Define receptor groups for binding energy calculation. Enter the sequence numbers of protein amino acids or nucleic acid bases in the group. For example, 1-213 or 1-211,212-213. Protein amino acid or nucleic acid base numbers start from 1 and are independent of the initial pdb amino acid numbering.

Custom Ligand

Define ligand groups for binding energy calculation. Enter the sequence numbers of protein amino acids or nucleic acid bases in the group. For example, 1-213 or 1-211,212-213. Protein amino acid or nucleic acid base numbers start from 1 and are independent of the initial pdb amino acid numbering.

One Structure Method

System Topology

Topology file obtained from the MD Solvation module or Membrane Solvation module.

System GRO

Structure file in .gro format obtained from the MD Solvation module or Membrane Solvation module.

System ITP

System parameter compression file in tar.gz format. Obtained from the MD Solvation module or Membrane Solvation module.

Results

The output includes:

ALA_Scan_Results.csv Contents

Larger Ka indicates stronger binding affinity, while smaller Ki indicates stronger inhibitory effect.

References

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.
Duan L, Liu X, Zhang JZ. Interaction Entropy: A New Paradigm for Highly Efficient and Reliable Computation of Protein-Ligand Binding Free Energy. J Am Chem Soc. 2016 May 4;138(17):5722-8.
https://github.com/Electrostatics/apbs/blob/main/docs/index.rst

Cleavage Site Prediction (PeptideCutter)

预测蛋白质序列中潜在的蛋白酶或化学试剂切割位点。模块基于PeptideCutter工具对应的文献资料复现。PeptideCutter 是瑞士生物信息学研究所（SIB）Expasy 平台提供的专业生物信息学工具。
支持的蛋白酶或化学试剂的切割规则如下：

*注：脯氨酸内肽酶仅能切割序列不超过30个氨基酸的底物。一种特殊的β螺旋结构域调控蛋白质水解：参见 Fulop 等，1998 年。

Trypsin Exceptions (Blocking Rules)

参数说明

Input File

上传蛋白的序列文件，只能提交单链序列，FASTA格式

Enzymes

选择切割的切割酶或化学物质，输入all表示选择全部，同时支持多个输入，输入方式如：Tryps;Ch_hi（输出对应的缩写，使用;分隔）。仅限上方切割规则表中酶和化学物质。

结果说明

输出All_in_One.csv，内容为输入序列中的切割点表，内容如下：

说明：

输出All_in_One.html，内容将包含所有链的切割信息，展出如下：
All_in_One.html

输出clvg_site_pred_results.tar.gz，包含所有序列各自的csv以及HTML报告结果。

参考文献

• Gasteiger, E. et al. (2005). Protein Identification and Analysis Tools on the ExPASy Server. In: Walker, J.M. (eds) The Proteomics Protocols Handbook. Springer Protocols Handbooks. Humana Press. https://doi.org/10.1385/1-59259-890-0:571DOI:10.1385/1-59259-890-0:571

Cleavage Site Prediction (PeptideCutter)

Predict potential protease or chemical reagent cleavage sites in protein sequences.
This module is reproduced based on the literature associated with the PeptideCutter tool.
PeptideCutter is a professional bioinformatics tool provided by the ExPASy platform of the Swiss Institute of Bioinformatics (SIB).

The supported protease and chemical reagent cleavage rules are listed below:

*Note: Proline endopeptidase can only cleave substrates with sequences shorter than 30 amino acids.
A special β-propeller domain regulates protein hydrolysis. See Fulop et al., 1998.

Trypsin Exceptions (Blocking Rules)