SE(3)-equivariant structure-based virtual screening model. Given a protein pocket + a compound library, it ranks compounds by predicted binding likelihood. Combines a receptor grid encoder (SE(3)-equivariant transformer) with a ligand GAT encoder, linked by trigonal attention.

Install

Requires Python 3.9, CUDA 11.7+ compatible GPU (or CPU fallback), and uv.

git clone https://github.com/j2ho/Motifscreen-pub.git
cd Motifscreen-pub
uv sync --extra preprocessing

The preprocessing extra pulls in RDKit + openbabel-wheel, both required for the prepare step. See INSTALL.md for CUDA-version alternatives and troubleshooting.

Quick inference

Two commands: prepare featurizes a protein + a compound library once; predict scores compounds against the prepared receptor.

# 1. Add hydrogens to your protein PDB (reduce is easiest)
reduce -BUILD receptor.pdb > receptor_h.pdb

# 2. Prepare: protein + compound library -> model input
uv run python motifscreen.py prepare \
    --protein receptor_h.pdb \
    --ligands compounds.sdf \
    --center=12.5,34.2,8.7 \
    --output prepared/

# 3. Predict: score compounds
uv run python motifscreen.py predict \
    --datapath prepared/ \
    --checkpoint models/epoch70.pkl \
    --base-config configs/training/endtoend.yaml \
    --output scores.csv

Output scores.csv:

target_id,compound_id,score,source_mol2
mytarget,ZINC000012345,0.912,all_ligands
mytarget,ZINC000098765,0.023,all_ligands

Score is in range [0, 1] where higher score is more likely to be a binder. Score should be evaluated as relative likeliness to be a binder within the same target. (e.g. target A could have active score 0.3 and decoy score 0.1 where target B could have active score of 0.7 and decoy score of 0.5)

Model checkpoint + endtoend.yaml config: checkpoint to be uploaded.

Benchmark reproduction

Download the pre-featurized benchmark tarball (~3.9 GB, ChEMBL-LR-107 + DUD-E-100 + LIT-PCBA-13) from Zenodo and run predict directly — no user-side prep needed:

# Fetch bench + verify
wget https://zenodo.org/records/21371299/files/motifscreen_benchmarks.tar.gz
echo "df2d57c7c2b4942af24216e6ed993d23c42ba3403cae9fdc42c41cae49d9f753  motifscreen_benchmarks.tar.gz" | sha256sum -c
tar xzf motifscreen_benchmarks.tar.gz

# Score DUD-E
uv run python motifscreen.py predict \
    --datapath motifscreen_benchmarks/dude_bench/prepared \
    --checkpoint {path_to_model_chkpt} \
    --base-config configs/training/endtoend.yaml \
    --gpus 0,1 \
    --output dude_scores.csv

Full reproduction scripts + benchmark-specific notes: see BENCHMARKS.md.

Model

• SE(3)-equivariant transformer for the receptor pocket grid (~800 grid points)
• GAT for ligand molecular graph
• Trigonal attention crossover between the two
• Outputs: motif labels on the grid, predicted key-atom positions, and a binding score for ranking

Architecture, losses, and training-time input featurization are described in the manuscript (TODO: preprint link at publish time).

ChEMBL data curation

The model was trained on a curated ChEMBL34-derived dataset along with PDBbind and BioLip data. If you want to use the ChEMBL dataset for training, but with the newer ChEMBL releases, the curation pipeline is a separate repository: j2ho/chembl-q. It filters ChEMBL activity data, selects artificial decoys with reasonable compound/target criteria, and provides a leakage-resistant train/test split by receptor sequence similarity against a supplied training-set FASTA (PDBbind + BioLip by default, easily swappable). Running chembl-q with the default settings reproduces the ChEMBL-LR benchmark this repo uses (may differ slightly from the original due to updates in ChEMBL).

Repository layout

Motifscreen-pub/
├── motifscreen.py                # CLI entry (prepare + predict subcommands)
├── src/
│   ├── cli/                      # prepare.py, predict.py
│   ├── data/                     # dataset, mol2 parsing, SASA, graph building
│   ├── model/                    # SE(3) + GAT + trigon attention
│   └── io/                       # protein/ligand file utilities
├── scripts/
│   ├── prepare_batch.sh          # multi-target parallel prepare
│   ├── gen_manifest.py           # auto-manifest from a directory of targets
│   ├── run_dude_bench.sh         # DUD-E-100 reproduction
│   ├── run_litpcba_bench.sh      # LIT-PCBA-14 reproduction
│   └── download_and_prepare_chembl_bench.sh
├── configs/
│   ├── training/endtoend.yaml    # architecture config (also needed by predict)
│   └── inference/                # benchmark configs
├── data/
│   ├── rosetta_params/           # bundled residue-type params (no Rosetta install needed)
│   ├── generic_potential/        # ligand atom typing
│   ├── dude_bench_manifest.tsv
│   ├── litpcba_bench_manifest.tsv
│   └── chembl_bench_manifest.tsv
├── INSTALL.md
├── INFERENCE.md
└── BENCHMARKS.md