Predicting Protein–Ligand Interactions with Deep Learning Co-Folding and Docking Architectures Through the Lens of Conformational Landscapes and Local Frustration

Abstract

This Zenodo record archives the delivered-pose docking evaluation associated with the study “Predicting Protein–Ligand Interactions with Deep Learning Co-Folding and Docking Architectures Through the Lens of Conformational Landscapes and Local Frustration.” It provides evaluation code, delivered DiffDock and EquiBind ligand poses, static experimental references, per-pose results, and provenance records. It supports reanalysis of the delivered poses, not reproduction of the original docking runs. The record contains: misato_benchmark_code.zip: a snapshot of the analysis repository at commit 55b8b45, including evaluation scripts, environment specifications, inventories, cohort accounting, and processed CSV results. misato_delivered_poses.zip: 101,920 supplied SDF files, including non-primary ranked outputs. The delivered primary-pose cohorts comprise 8,419 DiffDock and 8,565 EquiBind poses. misato_reference_cifs.zip: 8,601 asymmetric-unit crystal structures obtained from the RCSB Protein Data Bank for static-reference evaluation. FILE_GUIDE.md and archive_manifest.json: file descriptions, provenance, archive sizes, and SHA-256 checksums. The original MISATO dataset, including its QM and MD HDF5 files, trajectories, and restart files, is not rehosted here. The supplied predictions used static crystal receptors and QM-derived ligand conformers. Original docking input-preparation scripts, run settings, and complete failure logs are not included.