MotifScreen virtual screening benchmarks: ChEMBL-LR-107 + DUD-E + LIT-PCBA (v1.0, prepared)

Pre-processed benchmark data for reproducing MotifScreen virtual screening results across one new benchmark and two conventional benchmarks. DUD-E and LIT-PCBA are shipped predict-ready; protein grid, receptor properties, canonicalized MMFF94-charged all-in-one ligand mol2, and BRICS keyatoms. Users can go directly to predict. ChEMBL-LR is shipped in a lightly-prepared "raw" form (protein grid + per-compound batch mol2s) and requires one prepare-time step to compute BRICS keyatoms before predict; see the "ChEMBL" section below. ChEMBL-LR chembl_bench/chembl_bench_raw/ : ChEMBL-LR-107 (new benchmark, our contribution) 107 targets from ChEMBL v34. Targets are selected to have no leakage to popular protein-ligand training datasets; PDBbind v.2020 and BioLip2. Designed to test whether model gains generalize across data-bias regimes. Details on how the targets were tercile-balanced by AVE bias (36 top / 35 mid / 36 bottom). AVE bias (Wallach & Heifets 2018, JCIM 58:916, doi:10.1021/acs.jcim.7b00403) are in additional descriptions. The tercile split lets users see how model quality depends on train-test similarity. Ships raw per-compound batch mol2s + Rosetta-protonated PDBs + baked training-time grid/prop npz. Requires one prep-time step before predict: run motifscreen prepare per target to compute BRICS keyatoms (<target>.keyatom.def.npz). See scripts/download_and_prepare_chembl_bench.sh in the code repository. Tercile labels: data/chembl_lr_tercile_labels.tsv (in the code repository). DUD-E dude_bench/ : DUD-E (100 targets) Prepared DUD-E targets from Mysinger et al. (2012, J Med Chem 55:6582, doi:10.1021/jm300687e). Original data source: http://dud.docking.org. Two of the canonical 102 DUD-E targets (fgfr1, kif11) are excluded due to incomplete raw data in our mirror. Manifest: dude_bench/manifest.tsv with pocket centers (COM of crystal_ligand.mol2 per target). LIT-PCBA litpcba_bench/ : LIT-PCBA (13 targets) Prepared LIT-PCBA targets from Tran-Nguyen, Jacquemard, Rognan (2020, JCIM 60:4263, doi:10.1021/acs.jcim.0c00155). Original data source: https://drugdesign.unistra.fr/LIT-PCBA/. Two of the canonical 15 LIT-PCBA targets are excluded in pre-baked inputs (FEN1, MAPK1), but users can run these two targets themselves by downloading raw data from the original source and running motifscreen prepare. Manifest: litpcba_bench/manifest.tsv. Per-target file layout DUD-E and LIT-PCBA (predict-ready): <benchmark>/prepared/<target>/ ├── receptor.pdb Reference receptor conformation ├── <target>.grid.npz Baked receptor pocket grid (after protonation) ├── <target>.prop.npz Baked receptor per-atom properties (after protonation) ├── all_ligands.mol2 Actives + decoys concatenated, MMFF94-charged └── <target>.keyatom.def.npz BRICS keyatoms per compound (dict by compound ID) <benchmark>/labels/ └── <target>.actives.txt One active-compound ID per line ChEMBL-LR (raw, needs one prepare step): chembl_bench/chembl_bench_raw/<target>/ ├── receptor.pdb Reference receptor conformation ├── <target>.grid.npz Baked receptor pocket grid ├── <target>.prop.npz Baked receptor per-atom properties ├── batch_mol2s/<CID>_b.mol2 Per-compound mol2s (individual files) └── active_smiles_clu.csv Active labels + SMILES clustering info Usage Verify the download: sha256sum motifscreen_benchmarks.tar.gz # expected: df2d57c7c2b4942af24216e6ed993d23c42ba3403cae9fdc42c41cae49d9f753 Extract: tar xzf motifscreen_benchmarks.tar.gz DUD-E / LIT-PCBA for direct prediction: # Change dude_bench -> litpcba_bench as needed. # Use --gpus 0 for single GPU, --gpus 0,1 for two. uv run python motifscreen.py predict \ --datapath motifscreen_aff_benchmarks/dude_bench/prepared \ --checkpoint <path>/epoch70.pkl \ --base-config <path>/endtoend.yaml \ --gpus 0,1 \ --output results/dude_scores.csv ChEMBL-LR for one prep step then prediction: bash scripts/download_and_prepare_chembl_bench.sh \ --input-dir motifscreen_aff_benchmarks/chembl_bench/chembl_bench_raw \ --output motifscreen_aff_benchmarks/chembl_bench/prepared # ...then predict on the resulting prepared/ dir as above. See BENCHMARKS.md in the MotifScreen code repository for the full workflow including per-target AUROC / EF@1% metric computation. Data provenance ChEMBL-LR-107: assembled by the MotifScreen authors from ChEMBL v34 (Zdrazil et al. 2024, doi:10.1093/nar/gkad1004). Distributed under CC-BY 4.0. DUD-E: Mysinger, Carchia, Irwin, Shoichet (2012) J Med Chem 55:6582. doi:10.1021/jm300687e. Freely available for academic and commercial use. LIT-PCBA: Tran-Nguyen, Jacquemard, Rognan (2020) J Chem Inf Model 60:4263. doi:10.1021/acs.jcim.0c00155. Verification sha256: df2d57c7c2b4942af24216e6ed993d23c42ba3403cae9fdc42c41cae49d9f753 License CC-BY 4.0 for the ChEMBL-LR contribution and for the derived preparations of DUD-E and LIT-PCBA data. Users must cite the original data source publications listed under Data provenance.