CASP17_ligand

A consensus pipeline for ligand-pose prediction. Several co-folding methods run independently on the same target; their poses are checked for physical validity, clustered by shape overlap, and the five largest clusters become the five submitted models. This repository drives AlphaFold3, Boltz-2 and Protenix; we also used SeedFold, which has no public release.

This is the code MULTICOM used in CASP17: 12 RNA-ligand targets, 11 protein-ligand targets, one 13-chain RNA-protein complex, and fragment screens over 1209 and 647 compounds.

Why four methods

Across the 20 pose targets, the model we ranked first came from Protenix 9 times, AlphaFold3 4, SeedFold 4, and Boltz-2 3. None of them is reliably best, and confidence scores do not tell you which one to trust: on T2414 Boltz-2 reported an interface ipTM of 0.953 and still put the ligand 9–10 Å from the pocket the other three methods agreed on.

So we do not pick a method. We generate 100 models per method per target, throw away the ones that are physically impossible, and let the remaining poses vote by clustering. A pose that three methods land on independently is a better bet than a pose one method is confident about.

Requirements

Each method keeps its own conda environment; this package only dispatches to them and post-processes what comes back.

conda create -n casp17_ligand python=3.10
conda activate casp17_ligand
conda install -c conda-forge pymol-open-source   # pip has no working build
pip install -e .

You also need, separately:

• Boltz-2 and Protenix in their own conda environments (boltz, protenix). casp17_ligand/utils/env_utils.py finds them via conda info --base.
• AlphaFold3 as a docker image, plus its databases and weights. Weights require a request to DeepMind — see google-deepmind/alphafold3.
• A PoseBench environment if you want the relax step. casp17_ligand/utils/relax_pose.py and casp17_ligand/utils/minimal_relax.py run there under OpenMM, not in this one.

Copy env.example to .env and fill in where your databases, weights and CASP group code live:

cp env.example .env
set -a; source .env; set +a

CASP_GROUP_ID has no default — submissions filed under the wrong group code fail silently, so the scripts refuse to guess.

Boltz-2 downloads its own weights on first run; bash scripts/setup_weights.sh links them into weights/boltz/. Protenix expects its checkpoints under weights/protenix/ (override with protenix_root_dir in the model config).

Running one target end to end

A target needs two files from the prediction center and one CSV row. The two files are what the per-method input generators read; the CSV row is what the ensemble stage reads.

TGT=T2409
SERIES=CASP17_T

mkdir -p data/casp17_data/sequences/$SERIES data/casp17_data/smiles/$SERIES

# Sequence. Check view=sequence against view=all by hand before trusting it --
# mirrors and server feeds have been wrong.
curl -s "https://predictioncenter.org/casp17/target.cgi?target=$TGT&view=sequence" \
  > data/casp17_data/sequences/$SERIES/$TGT.txt

# Ligand SMILES (tab-separated: ID, Name, SMILES, Task)
curl -s "https://predictioncenter.org/download_area/CASP17/extra_experiments/ligands/$TGT.smiles.txt" \
  > data/casp17_data/smiles/$SERIES/$TGT.tsv

Then add one row to data/test_cases/casp17_T/ensemble_inputs.csv:

target,entity_type,protein_sequence,ligand_name,ligand_smiles,experimental_affinity,notes
T2409,protein,MNTSGLGWMSATEMAAQ...,U04,CCOC(=O)NC1=CC=CC=C1,,472 aa amidase

RNA targets use the same schema with rna_sequence in place of protein_sequence, rna as the entity type, and CASP17_R as the series.

# 1. Per-method inputs (YAML for Boltz-2, JSON for AF3 and Protenix)
python casp17_ligand/data/boltz2_input_preparation.py   --config-name boltz2_input_preparation_casp17_T
python casp17_ligand/data/af3_input_preparation.py      --config-name af3_input_preparation_casp17_T
python casp17_ligand/data/protenix_input_preparation.py --config-name protenix_input_preparation_casp17_T

# 2. Inference: two rounds x 50 models per method
bash casp17/scripts/run_local_pipeline_T.sh $TGT        # Boltz-2 + Protenix, two GPUs
bash casp17/scripts/run_af3_casp17_T.sh $TGT            # AlphaFold3, MSA once then r1 r2

# 3. Filter, cluster, rank
bash scripts/run_ensemble_r1r2_pipeline.sh $TGT 16 --yes

# 4. Write the five CASP LG submission files
python casp17/scripts/generate_submission.py \
  --cluster-csv outputs/ensemble_r1r2/cluster_experiment_detail_latest_$TGT.csv \
  --output-dir casp17/submissions/casp17_T \
  --smiles-dir data/casp17_data/smiles/CASP17_T

Step 2 assumes the AF3 image is named alphafold3-3.0.3_casp17; set AF3_IMAGE if yours differs. RNA targets run through run_local_pipeline_R.sh and run_protenix_rna_dualckpt.sh instead.

Multi-chain complexes (the M series) skip step 1: a 13-chain assembly does not fit the one-row schema, so its inputs are written by hand. Copy data/test_cases/casp17_M/boltz2_inputs/M2415_input.yaml and the Protenix JSON next to it as templates, check them with python casp17/scripts/verify_target_inputs.py, then run run_boltz2_r1r2_casp17_M.sh and run_protenix_dualckpt_casp17_M.sh. Boltz-2 needs the narrow-batch route there: 13 chains OOM at the default batch width, so that driver collects 4 models per round over 25 rounds.

Step 3 does the work. Each predicted structure is split into receptor PDB and ligand SDF, run through PoseBusters, and relaxed only if it fails — a relax that does not fix the failure is discarded rather than shipped. Surviving poses are clustered with Butina on SuCOS shape overlap, and cluster representatives are ordered by cluster size.

One step is not optional and not automated: look at the poses in PyMOL before submitting.

python casp17_ligand/analysis/generate_alignment_zips.py --targets $TGT

This writes a .pml that loads all five models aligned on the receptor. We caught a wrong-pocket MODEL 1 this way more than once.

Layout

casp17_ligand/       the package: input prep, inference dispatch, ensemble, analysis
casp17/scripts/      inference drivers, submission writer, CASP format validator
scripts/             ensemble pipeline entry point, weight setup
scripts/tests/       unit tests (pytest)
configs/             Hydra configs for the R and T series
data/test_cases/     per-target input manifests

Upstream

We ran these from their own checkouts rather than vendoring them:

AlphaFold3	google-deepmind/alphafold3	4ca8a65
Boltz-2	jwohlwend/boltz	cb04aec
Protenix	bytedance/Protenix	21c0b8b
PoseBench	BioinfoMachineLearning/PoseBench	c5d728d
RTMScore	sc8668/RTMScore	3280de1

casp17/scripts/LG_validation.py is the CASP organisers' format checker, patched to accept receptor lines and to take its paths from the environment.