PLABench: data for leakage-controlled benchmarking reveals generalization limits of deep learning for protein–ligand binding affinity prediction

Le résumé fourni par la source

PLABench is a leakage-controlled benchmark for protein-ligand binding affinity prediction. This deposit contains the inputs, predicted structures, trained weights, and scored predictions from the paper “Leakage-controlled benchmarking reveals generalization limits of deep learning for protein-ligand binding affinity prediction”. The code and the instructions for using any of this are at github.com/BioinfoMachineLearning/PLABench. Contents 01_benchmark_inputs.tar.xz (12 MB). Split partitions, the filtered and full ChEMBL35 sets with their removal ledger, the CASP16 labels, SMILES and stage-1 inputs, the standardized CASF and CSAR-HiQ affinity tables, every Davis and KIBA fold, and SOURCES.tsv. 02_af3_structures.tar.xz (270 MB). The predicted structures the benchmark scored: AlphaFold 3 for ChEMBL35 and CASP16, Boltz-2 for CASP16, CASF and CSAR-HiQ, and the template-guided CASP16 rung. 03_checkpoints.tar.gz (1.1 GB). The weights PLABench trained, plus the third-party weight inventory. 04_predictions_and_metrics.tar.gz (12 MB). Every per-model prediction, the leakage tables and the scored metrics. Using it Every path inside is relative to the repository root, so unpack from there: git clone --recurse-submodules https://github.com/BioinfoMachineLearning/PLABench.git cd PLABench for f in /path/to/0*.tar.*; do tar xf "$f"; done Run sha256sum -c SHA256SUMS.txt from the directory holding the tarballs to check them. The four archives are 1.4 GB to download and 5.8 GB unpacked. Davis and KIBA Complete: train, validation and test, for the warm-start arm and for both cold-start arms. Everything but the test folds is in the pickle format the MixingDTA authors released, because the target sequence repeats on every row and pickle stores it once, which is 28 MB against 621 MB for the same folds as CSV. Running python scripts/cv/export_davis_kiba_folds.py from the repository root writes CSVs with the same four columns as the test files. What is not included, and why PLABench artifacts are deposited. Corpora other people built are linked instead, and data/SOURCES.tsv in archive 01 gives the origin and the license of every path. These are the cases that change what you get. PDBbind v2020 forbids redistribution without written permission. Archive 01 carries the 4,465 / 497 refined partition as compound_id, split, and nothing else. Rejoin it against huggingface.co/datasets/photonmz/pdbbindpp-2020. CSAR-HiQ was never released under terms that allow redistribution, and its portal is gone, so the 36 and 51 structure sets are not here. Their standardized PDB code, SMILES, and pKd tables are, which is enough to rescore once you have the complexes. Binding MOAD’s download page still serves the CSAR-NRC HiQ set and its update; Binding MOAD is sunset and static now, so fetching the 87 entries from rcsb.org by PDB code is the safer route. CASF-2013 and CASF-2016 come from the CASF authors under their own terms, so the core sets are not here either. Request them at pdbbind.org.cn/casf.php. The Boltz-2 poses for both, and the standardized affinity tables, are in archives 02 and 01. The experimental CASP16 stage-2 complexes and the stage-1 submissions of the other predictor groups are the Prediction Center’s to publish, not ours, and are at predictioncenter.org/casp16. Nothing in the paper needs them: the CASP16 metrics are scored from the labels in archive 01, and configs/manifests/ in the repository records which targets each stage-2 run covered, so scripts/collect_results.py reports the same 93 of 93 without them. The eight data/*_prepared directories hold no bytes of their own. MFE wants one directory per complex with protein.pdb and ligand.mol2 inside, so they are 2,214 symlinks into CASF, CSAR-HiQ, and the Boltz-2 poses. Rebuild them with python scripts/data_prep/link_mfe_inputs.py once the corpora are in place. The four Boltz-2 ones need only archive 02 and work straight after unpacking. Archive 03 holds only the weights PLABench trained. Boltz-2, FLOWR.ROOT, FlowDock, BA-Pred, Graph_RG/Haiping, ESM3, and the MixingDTA authors’ Davis and KIBA warm-start models stay with their own releases. checkpoints/THIRD_PARTY.tsv gives the URL, SHA256, size, and license of each file, and bash scripts/download_third_party.sh fetches and verifies them. Licenses Terms differ by archive, which is why the record-level license is Other (Open). The ChEMBL35 sets inherit ChEMBL’s CC-BY-SA 3.0 and share-alike propagates. The Davis and KIBA folds and the target class map are CC-BY 4.0. The AlphaFold 3 structures are covered by the AlphaFold 3 Output Terms of Use, non-commercially; NOTICE_ALPHAFOLD3.txt in archive 02 carries the modification notice those terms require. No third-party checkpoint is deposited here, and each keeps the terms of its own release, recorded per file in checkpoints/THIRD_PARTY.tsv. The code is MIT.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.