CD40 CRD1 Antibody Shortlisting Pipeline

Computational pipeline to identify and rank antibody candidates that bind the CRD1 domain (residues 1-42) of CD40, using structure predictions from AlphaFold2, AlphaFold3, and Boltz followed by physics-based refinement and independent validation.

Pipeline Overview

01  Prescreen CRD1 contacts          (AF2/AF3/Boltz CIFs -> contact metrics)
     |
02  Filter candidates                (CRD1 fraction >= 0.5, require >= 2 sources)
     |
03  Convert best CIF -> PDB          (standardise chains: A/H/L)
     |
04  Prepare HADDOCK3 runs            (CDR-epitope AIRs + H-L anchoring + emref)
     |
    run_haddock_refine_batch.sh      (execute HADDOCK3 refinement)
     |
05  Collect HADDOCK3 scores          (parse capri_ss.tsv energy terms)
     |
06  Rosetta interface analysis       (shape complementarity, H-bonds, CDR contacts)
     |
07  Select best models               (merge HADDOCK + Rosetta, best per candidate)
     |
08  Final ranking                    (CRD category assignment, rank by HADDOCK score)

Result: 38 candidates -> 21 pass screening -> top 5 shortlisted for experimental follow-up.

Prerequisites

• Python 3.10+
• BioPython -- CIF/PDB parsing (pip install biopython)
• HADDOCK3 -- physics-based refinement (haddock3 docs)
• PyRosetta -- interface analysis (conda install pyrosetta in a dedicated env)
• Standard scientific Python: numpy, csv, pathlib

Setup

1. Clone the repository

git clone https://github.com/Ganesh-5347/CD40.git
cd CD40

2. Download structure predictions and HADDOCK3

bash haddock-project/scripts/download_predictions.sh

This downloads and sets up:

• haddock-project/inputs/ -- 1,330 CIF files (82 MB compressed, 333 MB extracted)
  • af3/ -- AlphaFold3: 38 candidates x 5 seeds = 190 CIFs
  • boltz/ -- Boltz: 38 candidates x 5 samples = 190 CIFs
  • af2/ -- AlphaFold2-multimer: 38 candidates x 5 architectures x 5 seeds = 950 CIFs
• haddock3/ -- HADDOCK3 source (cloned from GitHub)

3. Install dependencies

# HADDOCK3
pip install -e haddock3/

# Python packages
pip install biopython numpy

# PyRosetta (for step 06, requires separate conda env)
conda create -n rosetta python=3.10
conda activate rosetta
conda install pyrosetta -c https://conda.rosetta.com

Running the Pipeline

All scripts are run from the repository root. Each script has --help for full argument documentation.

# Stage 1: Prescreen all CIF models for CRD1 contacts
python haddock-project/scripts/01_prescreen_contacts.py

# Stage 2: Filter to candidates with CRD1 signal in >= 2 sources
python haddock-project/scripts/02_filter_candidates.py

# Stage 3: Convert best CIF per candidate/source to PDB
python haddock-project/scripts/03_build_complex_pdb.py

# Stage 4: Prepare HADDOCK3 refinement (split chains, generate AIRs + configs)
python haddock-project/scripts/04_prepare_haddock_runs.py

# Execute HADDOCK3 refinement (requires haddock3 on PATH)
bash haddock-project/scripts/run_haddock_refine_batch.sh

# Stage 5: Collect HADDOCK3 energy scores
python haddock-project/scripts/05_collect_haddock_scores.py

# Stage 6: Rosetta interface analysis (requires PyRosetta)
python haddock-project/scripts/06_rosetta_interface_analysis.py

# Stage 7: Merge HADDOCK + Rosetta scores, select best model per candidate
python haddock-project/scripts/07_select_best_models.py

# Stage 8: Assign CRD categories and produce final ranking
python haddock-project/scripts/08_final_ranking.py

Key Results

Metric	Value
Candidates screened	38
Models scanned (CIF)	1330 (950 AF2 + 190 AF3 + 190 Boltz)
Passed CRD1 filter	21
HADDOCK3 refinements	48 (one per candidate/source)
Rosetta validations	48
Shortlisted (top 5)	Ab-36.1, Ab-9.1, Ab-26.1, Ab-1.1, Ab-23.1

Top candidate: Ab-36.1 -- HADDOCK score -378.2, shape complementarity 0.80, 25 H-bonds, 6/6 CDR loops engaged, 97% CRD1-specific.

Output Files

All outputs are written under haddock-project/outputs/ (excluded from git):

Path	Description
outputs/prescreen/prescreen_contacts.csv	Per-model contact metrics (1330 rows)
outputs/prescreen/prescreen_summary.csv	Per-candidate summary with keep/reject
outputs/prescreen/prescreen_keep.txt	Passing candidate IDs (21)
outputs/complex_pdb/{af,af2,boltz}/*.pdb	Best CIF converted to PDB (48 files)
outputs/haddock_refine/*/run/	HADDOCK3 refinement outputs
outputs/reports/refine_scores.csv	HADDOCK3 energy scores
outputs/reports/rosetta_interface.csv	Rosetta interface metrics
outputs/reports/combined_scores.csv	Merged best model per candidate (21 rows)
outputs/reports/final_ranking.csv	Final ranked list with CRD categories