Year: 2026
Type: preprint
Abstract: Predicting compounds that are more potent than those represented in the training data remains a difficult and practically important problem for quantitative structure-activity relationship (QSAR) modeling. This challenge is particularly relevant during lead optimization, where the goal is to design or identify analogs that are more potent than compounds already measured. To address this problem, we developed a new Pairwise-Difference QSAR (PD-QSAR) approach that learns changes in biological activity from differences in molecular descriptors between compound pairs. We evaluated three strategies for aggregating the resulting PD-QSAR predictions: mean aggregation, k-nearest-neighbor aggregation (k = 3), and 10% quantile aggregation, and compared their performance with Direct QSAR and Chemprop using a benchmark derived from publicly available ChEMBL 35 data. The primary benchmark included 103 human single-protein target datasets, comprising 33,707 activity records. For each target, the 10% most potent activity records were withheld for external validation. Among the approaches evaluated, 10% quantile aggregation provided the lowest median external validation RMSE (0.518) and outperformed mean and kNN aggregation on external validation RMSE for 100 of 103 targets, Direct QSAR for 102 targets, and Chemprop for 94 targets. These results support PD-QSAR with quantile aggregation as a useful strategy for prioritizing analogs expected to extend beyond the potency range represented by previously measured compounds.
Source: ChemRxiv
Authors: Sankalp Jain, Alexey V. Zakharov
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7222400298