Symbolic Fragment Weighting for Substructural Analysis: Genetic Programming vs Genetic Algorithm on ChEMBL

Substructural Analysis is a scalable and chemically interpretable approach to ligand-based virtual screening, but ranking quality depends strongly on the fragmentweighting model used. This study benchmarks Genetic Programming for evolving symbolic fragment-weighting equations against a Genetic Algorithm baseline that directly optimises integer fragment weights under a fixed Substructural Analysis scoring structure.

Using a ChEMBL-derived benchmark of 15 activity classes represented by MACCS 166-bit fingerprints, we conducted 10 runs per class under a stratified 90/10 train-test protocol. Early-recognition performance was evaluated using enrichment factor at the top 1% (EF@1%) on the test partition. Between-method differences were tested using a two-sided Mann-Whitney U test (α = 0.05), with Benjamini-Hochberg adjustment across activity classes.

The Genetic Algorithm baseline achieved higher median EF@1% than the Genetic Programming approach in most activity classes, with statistically significant advantages in 11 of 15 classes before adjustment and 10 of 15 classes after Benjamini-Hochberg adjustment. The Genetic Programming approach remained competitive for COX, 5HT1A, 5HT3, and ACHE; for COX, it produced a higher median EF@1% than the Genetic Algorithm baseline (19.05 vs 16.67), although the difference was not statistically significant (p = 0.333).

These findings suggest that symbolic fragment-weighting is feasible and inspectable, but that enrichment performance is sensitive to how the optimisation objective is defined for early-recognition screening.

Genetic Programming can evolve symbolic Substructural Analysis weighting equations; however, the observed performance gap suggests that future GP fitness designs should be more directly aligned with early-recognition objectives.