Ranking allosteric CDK2 inhibitors: Free energy calculations do not outperform ligand size, whereas a machine-learning docking score does

Complete inputs, scripts, raw program outputs and analyses for a computational study of allosteric inhibitors of cyclin-dependent kinase 2 (CDK2) from the 2-(arylalkylamino)benzoic (anthranilic) acid series, which bind a pocket adjacent to the C-helix rather than the ATP site. The study benchmarks conformer analysis, molecular docking and end-point free energy calculations (MM/GBSA) against experimental potencies taken from a single assay, using ligand heavy-atom count as the null model at every stage, and then tests the selected protocol on 30 further compounds of the same series under an analysis plan that was written, cryptographically hashed and time-stamped before those simulations were started. This deposit accompanies the manuscript "Ranking allosteric CDK2 inhibitors: Free energy calculations do not outperform ligand size, whereas a machine-learning docking score does" (submitted; article DOI to be added on acceptance). It is intended to let the series be reused as a benchmark for structure-based affinity prediction on allosteric kinase ligands. Contents - cdk2_allosteric_data.zip — the full working tree: curated potency data and PDB metadata (00_envanter/); CREST conformer ensembles with GFN2-xTB, r²SCAN-3c and UMA energies (01_crest_test/); superposed receptors, docking box, GNINA poses and redocking/cross-docking results (03_docking/); the 37-compound series with its ChEMBL structure corrections and docking-based ranking (04_afinite/); all Amber topologies, MD inputs, production outputs, trajectory analyses and MM/GBSA calculations for 108 simulations (05_md/); the time-stamped analysis plan, its three dated amendments, checksums and the validation result (06_dogrulama/); every workflow script (bin/); and the manuscript sources with the scripts that generate its figures, tables and Supporting Information (makale_bjp/). - yorunge_kuru/ — solvent-stripped production trajectories (protein and ligand, 100 ps frame spacing) with matching dry topologies, for all 108 simulations: six crystallised ligands in three replicas of 100–500 ns, and 30 validation compounds in three replicas of 12 ns. - README_DATA.md — layout, naming conventions and provenance notes. - MANIFEST.sha256 — SHA-256 checksum of every file in the archive. Notes for reuse. Solvated trajectories (about 58 GB of NetCDF) are not included; the solvent-stripped trajectories provided here reproduce every reported analysis. Each calculation directory carries a provenance file (KAYNAK_BILGISI.txt) recording the date, machine, program versions, full command lines and input checksums. Residue numbers in the Amber topologies are CDK2 (UniProt P24941) numbers plus one, because PDB entry 7RWF begins at residue 0, so topology LYS 34 is Lys33. System names 88O, 8IQ, 8IL, 7TW, 80E and 7TH are PDB chemical component identifiers of the crystallised ligands; V01–V30 are the validation compounds, mapped to ChEMBL identifiers in 06_dogrulama/dogrulama_seti.csv; the suffixes _t2 and _t3 denote the second and third independent replicas. Most directory names and script comments are in Turkish; the README gives a glossary of the recurring terms. Software. Amber 26 and AmberTools 26 (pmemd.cuda, tleap, antechamber, cpptraj, MMPBSA.py), CREST 3.0.2, xtb 6.7.1, ORCA 6.1.0, fairchem UMA (uma-s-1p1), GNINA 1.3.3, Open Babel 3.2.1, RDKit, and Python 3.12 with NumPy, SciPy and Matplotlib. Source data. Crystal structures from the Protein Data Bank (7RWE, 7RWF, 7RXO, 7S4T, 7S7A, 7S84, 7S85, 8FOW, 8FP0); potency data from ChEMBL document CHEMBL5380784, assay CHEMBL5381439.