Quantifying Multi-Objective Optimization in Generative Chemistry for Molecule Design
- 1Novartis Biomedical Research
- 2Novartis Campus
Abstract
Multi objective optimization (MOO) is central to molecular design, yet its impact on generative chemistry outcomes remains poorly quantified. Consequently, it is unclear how different MOO strategies influence chemical space exploration and property optimization in generative methods. Here, we introduce a benchmarking framework for evaluating MOO strategies in generative chemistry. Six MOO approaches are highlighted including arithmetic and geometric means, non dominated sorting, Chebyshev scalarization, generational distance, and binary thresholding within a lead optimization setting. Across 60 scaffolds derived from six potency datasets, we explore a many-objective task for the simultaneous optimization of various compound properties, including potency, solubility and lipophilicity. We evaluate methods using metrics aligned to recurring medicinal chemistry use cases or decision scenarios and assess differences across datasets and scaffolds. We find that the MOO choice substantially changes both optimization outcomes and explored chemical space across datasets, scaffold and intended use cases. Overall, our results provide practical, use case driven guidance for selecting and comparing MOO strategies in generative chemistry workflows.
Related
- From Computational Chemistry to Generative Models: A Survey of AI-Driven Small-Molecule Drug Discovery
- Quantifying the Coherence Wall: FeMoCo Quantum Chemistry from 48 to 108 Qubits on IBM Heron r2
- Robust Multi-Objective Reinforcement Learning Improves Joint Predicted-Potency and Drug-Likeness Yield in EGFR Molecule Generation
- Single-molecule cold chemistry: Isotopic branching ratios of the H2O+ + HD reaction at cryogenic temperatures
- Built in Minutes, Checked by Script: Instructor-Made Interactive Chemistry Study Guides with Generative AI
BibTeX
@misc{torrenperaire2026quantifying,
title = {{Quantifying Multi-Objective Optimization in Generative Chemistry for Molecule Design}},
author = {Paula Torren-Peraire and Nadine Schneider and Raquel Rodríguez-Pérez and Robin A. Fairhurst and Jessica Lanini},
howpublished = {ChemRxiv},
year = {2026},
month = aug,
doi = {10.26434/chemrxiv.15008103/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008103/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Quantifying Multi-Objective Optimization in Generative Chemistry for Molecule Design
AU - Paula Torren-Peraire
AU - Nadine Schneider
AU - Raquel Rodríguez-Pérez
AU - Robin A. Fairhurst
AU - Jessica Lanini
PY - 2026
DA - 2026/08/31
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008103/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008103/v1
AB - Multi objective optimization (MOO) is central to molecular design, yet its impact on generative chemistry outcomes remains poorly quantified. Consequently, it is unclear how different MOO strategies influence chemical space exploration and property optimization in generative methods. Here, we introduce a benchmarking framework for evaluating MOO strategies in generative chemistry. Six MOO approaches are highlighted including arithmetic and geometric means, non dominated sorting, Chebyshev scalarization, generational distance, and binary thresholding within a lead optimization setting. Across 60 scaffolds derived from six potency datasets, we explore a many-objective task for the simultaneous optimization of various compound properties, including potency, solubility and lipophilicity. We evaluate methods using metrics aligned to recurring medicinal chemistry use cases or decision scenarios and assess differences across datasets and scaffolds. We find that the MOO choice substantially changes both optimization outcomes and explored chemical space across datasets, scaffold and intended use cases. Overall, our results provide practical, use case driven guidance for selecting and comparing MOO strategies in generative chemistry workflows.
ER -