Kairoszhe/RSI-Lab-docking

RSI-Lab CDK2 exploratory docking

Public research archive. This repository is publicly accessible and can be downloaded without authentication. It includes original research prompts, model responses, execution receipts, and historical local machine paths. A focused public-release review found no credential or sensitive-personal-information blockers. Original scientific files were not modified. See dataset_manifest.json, SHA256SUMS, and each archive's EXPORT_MEMBER_HASHES.json for integrity verification.

Three distinct subsets

Subset	Candidate inputs	Observed seed jobs	Scope
agentic_pilot	80 across four disjoint cohorts	92 across six campaigns; 35 distinct docked compounds	Real Owner/Inspector, Lab, Meta, two paired validation cohorts, then development
legacy_screen_120	120	240	Earlier as-supplied chemical states; 118 compounds have two valid poses
reference_control	1 native 1DI8 DTQ ligand	2	Descriptive redocking control; not part of compound ranking

These subsets have different preparation histories. Do not merge them as one uniform scientific protocol. Validation parent/candidate arms share their prospectively paired cohort: repeated compound/seed observations across arms are not independent new compounds. Unobserved candidate scores are missing, not zero. All 334 job observations and invalid-pose outcomes are retained.

All docking uses standalone AutoDock Vina 1.2.7, seeds 17/43, native MODEL 1, with lower affinity in kcal/mol being better. The qualified compound score is the median of both seeds only when both completed, preserved chemical identity, and passed the configured PoseBusters checks. No pose is replaced after checks. There is no calibrated official score threshold and no formal >=100 selection claim. The agentic study ends with five compounds, not a formal screening list.

The agentic Meta gate accepted its candidate through the cost path, with 31.2924% fewer validation Lab tokens and no pose-valid coverage regression. The quality path did not pass. This tiny development gate is not evidence of general efficacy, biological potency, or statistical superiority.

Chemical and source limitations

Agentic inputs use MolScrub 0.3.2 rules at the exploratory pH interval 7.5-8.0, two tautomer-reaction rounds, and KEEPMAX/KEEPMIN filtering. Only one unique retained state with preserved heavy-atom/isotope identity and stereochemistry is admitted. Of 350 source candidates, 268 were admitted, 52 were ambiguous, and 30 failed state validation. A unique retained result is not measured dominance, a population estimate, or exhaustive microstate coverage. Upstream invalid reaction products may be skipped; acid/base sanitization can retain input. The interval is not a verified common assay pH or a sensitivity study.

The unchanged pilot receptor has missing loops, removed waters and residues, and default residue protonation; ligand preparation does not make the receptor pH-calibrated. Reference-state equivalence and frame-preserving redocking RMSD are descriptive controls, not a validation of ligand ranking.

Structures and assay provenance come from the first 1000 lexical-assay records in a frozen ChEMBL_37 CDK2 query (CHEMBL301), retrieved 2026-10-20, not an unbiased target-wide population. Historical identities and all returned prepared states were excluded; the split is deterministic and scaffold-disjoint. The separately archived raw source includes measured activity fields solely for source reconstruction; those values were not exposed as Lab inputs or used to choose/rank the cohorts. Do not present docking scores as IC50 labels.

Contents and reproduction

The five root tables are directly loadable with Hugging Face Datasets or pandas. *_inputs_poses.tar.gz packages contain candidate inputs, scientific protocols, native poses, identity/PoseBusters results and source references. The agentic package also contains the exact raw ChEMBL snapshot and retrieval metadata under agentic_pilot/upstream/. See SCHEMA.md for archive paths and table joins.

audit_*.tar.gz packages preserve the original run bytes, including duplicate frozen inputs and full model receipts. The prior-attempt archive contains seven original model attempts needed for overhead reconciliation; six have verified usage and one has unknown usage. Unknown cost is not zero. These archives are audit evidence, not directly relocatable resumable studies: create a new study directory to rerun. Do not overwrite or silently reuse original receipts.

The GitHub code revision is recorded in dataset_manifest.json.code_ref when publication is finalized. A null value means that cross-reference is pending; source hashes remain available. Download a specific HF revision and verify SHA256SUMS before using the data. License terms and citations are separated by component in LICENSES_AND_SOURCES.md; no new project-code license is granted.