Improving substructural analysis with chemical weighted RMSE in symbolic regression

Substructural Analysis (SSA) is a ligand-based virtual screening method that ranks compounds by combining fragment weights, where each weight reflects the contribution of a molecular substructure to the separation of active and inactive compounds. However, conventional SSA weighting schemes are often static and may be less adaptive to heterogeneous and highly imbalanced bioactivity datasets. In this study, we proposed a hybrid SSA framework that used Genetic Algorithms (GA) to optimize fragment-level weights and Genetic Programming (GP) to evolve symbolic weighting equations guided by a Chemical Weighted Root Mean Square Error (CW-RMSE) fitness function. Enrichment factor at the top 1% of the ranked list (EF@1%) was used as the main early-recognition metric because it measured active-compound retrieval relative to random selection in the most practically relevant part of a screening list. The framework was evaluated using 15 ChEMBL-derived activity classes represented by molecular fingerprints. SSA-GP achieved higher median EF@1% than GA-based SSA in 14 of 15 classes, with statistically significant improvements in 13 of these classes using the Mann-Whitney U test (p < 0.05). Notable median EF@1% gains were observed for RNN (16.98 to 69.34), MMP1 (36.67 to 83.33), 5HT1A (3.13 to 46.25), and FXA (15.83 to 57.91), while AT1 was the exception in which GA performed better. These results indicated that SSA-GP, guided by the CW-RMSE fitness function, could improve early active-compound prioritization while retaining the interpretability advantage of symbolic SSA models. Further validation against non-SSA machine-learning baselines and expert chemical interpretation of the generated equations is recommended.