FGFR2 Inhibitor SAR Mining from Patent WO2020231990

What this research found

Medicinal-chemistry patents often publish their potency tables as pictures, which blocks automated structure-activity analysis. Starting from patent WO2020231990, which claims heterocyclic inhibitors of fibroblast growth factor receptor 2 (FGFR2), 45 of 52 example compounds were recovered and verified against PubChem, then scored geometrically against a 1.77 angstrom co-crystal of the FGFR2 kinase domain in place of the unreadable activity bins. A random forest trained on those scores reproduced them with an R-squared of 0.36 to 0.40 across three cross-validation schemes and identified the 4-amino-pyrrolopyrimidine hinge-binder core as its leading structural driver.

• Extraction recovered 45 of 52 patent examples as PubChem-verified structures, an 86.5% yield with no duplicate identifiers. Pairwise Tanimoto distances spanning 0.328 to 0.954 indicate a chemically diverse series rather than one tight analogue family, though 33 of the 45 sit in a single pyrrolopyrimidine cluster.
• The patent's master activity table is a raster image in both its PDF and HTML forms, so no per-compound potency bin could be read. A geometric proxy target was substituted, combining alignment similarity to the co-crystal ligand, heavy-atom pocket contacts of 10 to 44 per molecule, and ligand-to-pocket centroid distances of 1.09 to 5.40 angstroms.
• The random forest generalises consistently but only moderately, with out-of-bag R-squared 0.386, 5-fold 0.364 (RMSE 1.586) and leave-one-out 0.400 (RMSE 1.541), against a training R-squared of 0.699.
• Impurity-based, permutation, and TreeSHAP importances all converge on the same seven drivers: Bertz topological complexity, polar surface area, exact molecular weight, aromatic-ring count, total atom count, Crippen log P, and a single fingerprint bit.
• That bit maps to the 4-amino-pyrrolo[2,3-d]pyrimidine core and fires in 22 of 45 compounds. Related informative bits pick out the pyrimidine nitrogen-carbon-nitrogen hinge triad (34 of 45), the 4-amino hydrogen-bond donor (32 of 45) and the methacrylamide covalent warhead (9 of 45), recovering the established kinase hinge-binder pharmacophore rather than proposing a new one.

How it was done

The patent full text was retrieved from Patentscope and Google Patents, example blocks were parsed on their chemical names, and each was matched to a PubChem structure by exact lookup with a fuzzy fallback above a 0.75 similarity score. Structures were standardised in RDKit and described by 23 two-dimensional descriptors plus a 2,048-bit Morgan fingerprint, with three-dimensional conformers built by distance geometry and minimised under the MMFF94s force field. In parallel a 1.77 angstrom FGFR2 kinase-domain co-crystal was retrieved and its 23-residue ATP pocket annotated with the canonical landmarks: gatekeeper V564, the VEYA hinge motif, the DFG motif, and catalytic lysine K517. With no readable potency column, every ligand was scored on a standardised composite of alignment similarity, pocket contacts, and centroid distance, and a 1,000-tree random forest was fitted across all 2,071 features with out-of-bag, 5-fold, and leave-one-out error reported side by side.

Data sources

• Patent WO2020231990, Heterocyclic Inhibitors of FGFR2 and Methods of Use Thereof (Mirati Therapeutics, 2020) — 52 example blocks parsed, 45 curated
• PubChem — name-to-structure resolution and verification of extracted compounds
• RCSB Protein Data Bank — FGFR2 kinase-domain co-crystal 1OOU at 1.77 angstroms, with 3B2T, 1OOO and 1OOQ retained as alternates
• UniProt P21802 (FGFR2_HUMAN) — tyrosine-kinase domain, residues 458 to 768
• Subbiah et al., Proceedings of the National Academy of Sciences 121:e2317756121 (2024) — lirafugratinib discovery
• Turner et al., Journal of Medicinal Chemistry 65:1481 (2022) — fragment-to-lead selective FGFR2 inhibitors