DeepRank-Ab Inference Pipeline

DeepRank-Ab is a scoring function for ranking antibody-antigen docking models based on geometric deep learning.

📄 Publication
https://www.biorxiv.org/content/10.64898/2025.12.03.691974v1

This repository provides the full inference pipeline for the model described in the paper.

---

🚀 Features

• PDB Processing
  • Split ensemble models
  • Extract chain sequences
  • Merge chains for downstream analysis
• FASTA Conversion
  • Generate FASTA files for CDR annotation and ESM embeddings
• ESM Embeddings
  • Compute embeddings using esm2_t33_650M_UR50D
• Graph Construction
  • Build atom-level graphs with precomputed node and edge features
• Prediction
  • Inference with pretrained EGNN models and output predicted DockQ

---

📦 Installation

1. Clone the repository

git clone https://github.com/haddocking/DeepRank-Ab
cd DeepRank-Ab

2. Create and activate the environment

mamba env create -f environment-gpu.yml
mamba activate deeprank-ab

3. Install ANARCI

ANARCI is required for CDR annotation.

Installation instructions:
https://github.com/oxpig/ANARCI

Note: hmmscan is already included in the environment.
If you encounter issues, follow the workaround here:
oxpig/ANARCI#102

---

🔧 Usage

The inference pipeline is executed through:

DeepRank-Ab/scripts/inference.py

Run the pipeline

usage: inference.py [-h] [--heavy_chain_id HEAVY_CHAIN_ID] [--light_chain_id LIGHT_CHAIN_ID]
                    [--antigen_chain_id ANTIGEN_CHAIN_ID]
                    pdb_file

pdb_file is required; heavy_chain_id, light_chain_id, and antigen_chain_id are optional. If not provided, they will be detected automatically.

Example

python3 scripts/inference.py example/test.pdb 

This will:

• Create a workspace
• Generate ESM embeddings
• Annotate CDRs
• Build atom-level graphs
• Cluster nodes
• Predict DockQ scores
• Save output files (.csv and .hdf5)

---

🧬 Input Requirements

• PDB file
  Antibody--antigen structure. Can be a single model or an ensemble.

• Heavy chain ID (optional)
  Example: H

• Light chain ID (optional)
  Example: L

• Antigen chain ID (optional)
  Example: A

---

⚙️ Large-Scale Inference

We provide a helper script for running DeepRank-Ab on large batches of complexes. Adapt it to your dataset as needed.

Example:

python3 scripts/run_batch_inference.sh

---

📫 Support

For issues or questions, please open a GitHub issue.