Pilot gap analysis: COCONUT 2.0 versus ChEMBL 37 (November 2026)

Full application of a gap-targeting rule for natural-product-inspired library design. Natural-product reference: COCONUT 2.0 (September 2026 csv_lite; 738,827 rows, 434,058 unique structures). Synthetic reference: ChEMBL 37 (2,762,587 structures after removing 97,913 NP matches and 37,319 parse failures). Under the strictest definition of absence, 65.8%, 50.0% and 46.5% of recurring NP scaffolds were absent at the exact, generic and ring-system levels; at ≤20 synthetic occurrences the fractions were 90.3%, 78.7% and 74.2%. Gap scaffolds had a higher median Fsp³ than covered scaffolds (0.467 vs 0.400; Mann–Whitney p ≈ 3 × 10⁻²⁵⁶), supporting the prioritisation on three-dimensionality, with a small effect size.