I tried to reproduce the well-known PDBbind data-leakage effect with a small 3D GNN on a laptop - and couldn't. Two independent diagnostics show why: leakage only inflates models strong enough to memorize.
Jan 9, 2025
A three-tier QSAR validation framework: internal statistics, cross-validation, and external checks, with the thresholds used for papers and regulatory submissions.
Nov 18, 2024
A practical guide to three advanced 3D fingerprinting methods (PLEC, SPLIF, and E3FP) and how to choose between them when featurizing docking poses for ML-based drug discovery models.
Jul 12, 2024
Practical introduction to DataWarrior as a free, chemistry-aware workbench for data visualization, filtering, and focused library generation, based on Bruno Villoutreix’s tutorial series.
Feb 7, 2024
Step-by-step guide to constructing an open-source drug discovery pipeline with AI and chemistry tools, from data to visualization.
Oct 1, 2023
A random forest on the B3DB blood-brain barrier set, scored on a held-out split, then opened with impurity importance and SHAP. TopoPSA is in the descriptor block only if you select that block on purpose.
May 24, 2023
A 3D-QSAR model maps aligned molecular fields onto potency. The statistics say whether it predicts. The contours say where a chemist might change the molecule. Both depend on the alignment.
Apr 2, 2023
Supervised learning needs a label on each example. Unsupervised learning looks for structure when that label is absent. The difference is in the table, before any algorithm is chosen.
Oct 29, 2022
ChEMBL, BindingDB, PubChem, PDBbind, and BRENDA answer different questions. The useful choice is the table that holds the measurement you mean to model.
Oct 3, 2022
How a ChEMBL IC50 table for one target becomes a labelled set: one row per molecule, a mean potency, and a class that leaves the uncertain middle out.
Sep 7, 2022