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

Le résumé fourni par la source

This record is the v1.0.0 source snapshot of PLABench, the code behind the paper of the same name. PLABench is a leakage-controlled benchmark for protein-ligand binding affinity prediction: it runs nine models, sequence-based and structure-based, over the same targets under the same metrics, and varies the input structure so that pose quality can be distinguished from model quality. Development happens at github.com/BioinfoMachineLearning/PLABench. The zip is the v1.0.0 tag, byte for byte. What is in the zip PLABench-v1.0.0.zip (8.4 MB, 366 files) holds the benchmark package plabench/, the runner run_benchmark.py, the Hydra configs in configs/, the data preparation, training, scoring, and leakage scripts in scripts/, the conda environment files in environments/, the figure and table code in analysis/, and the scored results in results/. forks/patches/ holds the changes to the nine model forks as git patches, 33 KB for all of them. The forks are git submodules in the repository, which a zip cannot carry. Checking out the pinned upstream commit listed in forks/README.md and running git am on the matching patch reproduces each fork's tree exactly. flowr_root is unchanged from its pinned commit and has no patch. No model weights are in this record, and nothing is vendored from the nine upstream projects beyond those patches. Running it git clone --recurse-submodules https://github.com/BioinfoMachineLearning/PLABench.git cd PLABench bash scripts/download_checkpoints.sh bash scripts/download_third_party.sh download_checkpoints.sh pulls the weights PLABench trained from the data record, 10.5281/zenodo.22716174, which also carries the benchmark inputs, the AlphaFold 3 and Boltz-2 structures, and every prediction and metric in the paper. download_third_party.sh fetches the other models' weights from their own releases and checks each one against the SHA256 in checkpoints/THIRD_PARTY.tsv. README.md covers what each corpus needs and which ones cannot be redistributed. License MIT, in LICENSE. The patches in forks/patches/ modify code that stays under each upstream project's own license, listed with the forks in forks/README.md. Data and weights are not covered by this record; the data record gives their terms per path.

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