Multi-Objective Inverse QSAR Based on Mixed-Integer Linear Programming and Machine Learning

Read full text Summarise with VIRAT
Authors: Bowen Song, Hiroto Saigo
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This work extends the mol-infer framework for inverse QSAR/QSPR to better support multi-criteria molecular design through improved molecular representation and MILP-based inverse inference. We explicitly incorporate several chemically meaningful structural quantities, including ROTB, HBA/HBD, TPSA, and aromaticity-related proxies, as derived descriptors and optimization objectives. To enable their multi-objective control in Stage 4, we introduce an additional constraint block (C3) that translates the rules of these environment-dependent properties into linear inequalities. We also extend the baseline 2L+CC descriptor set by introducing atomic environment configurations (AECs), which encode atom-centered local chemical contexts. These descriptor extensions are designed not only to improve prediction performance in Stage 3, but also to provide a convenient foundation for MILP-based multi-objective control in Stage 4. Experiments show that, compared to the baseline 2L+CC results reported in [20], the proposed descriptors achieve higher ML scores in the selected data sets. Moreover, in our experiments, the multi-objective MILP formulation finds feasible solutions within a 10-minute time limit and enables explicit control over several chemically meaningful structural properties. This controlled generation also leads to molecules with improved drug-likeness and synthetic accessibility.

Fields of Study & Subjects

Computational Drug Discovery Methods Machine Learning in Materials Science Advanced Chemical Physics Studies

Publication details

Journal / source
ChemRxiv
Publication date
31 March 2026
Language
EN
Document type
Preprint