Dataset- and Endpoint-Guided Co-Evolutionary Feature Engineering of Molecular Descriptor Spaces: Benchmarking on 14 Datasets and DILI Prediction
Abstract Feature engineering remains a central challenge in Quantitative structure–activity relationship (QSAR) modeling, as the quality and relevance of molecular descriptors (MDs) directly determine predictive performance. Traditional approaches rely on the overproduction and subsequent selection of descriptors, a paradigm that often leads to redundancy and suboptimal representations. This limitation motivates the development of approaches that directly optimize descriptor representations rather than relying on post hoc selection. In this study, we introduce a cooperative coevolutionary extension of the Algorithm to explore and optimize descriptor configuration spaces (AExOp-DCS), termed C-AExOp-DCS, which explores multiple MD configuration spaces (DCSs) while evaluating descriptors in a cross-DCS cooperative context. The method combines supervised and unsupervised relevance criteria through the Choquet integral and introduces context-dependent descriptor evaluation by recalculating mean decrease impurity (MDI) in the presence of collaborators from other DCS populations. The proposed approach was validated on 14 benchmark data sets covering pharmacological and ecotoxicological end points, where it generally achieved improved or comparable predictive performance relative to the original noncooperative AExOp-DCS, while remaining competitive with previously reported state-of-the-art QSAR models. Statistical analysis indicated a higher probability of improved generalization performance. Furthermore, C-AExOp-DCS-derived models matched or outperformed advanced machine-learning and deep-learning approaches for DILI prediction across heterogeneous external validation sets. SHAP and LIME analyses revealed a stable recurrent explanatory signature across the top-performing random forest models, comprising a five-member dominant and a nine-member extended QuBiLS-MAS descriptor core, with highly consistent global SHAP-importance profiles (Pearson r = 0.978–0.992; Spearman ρ = 0.921–0.934). Structural interpretation revealed complementary molecular representations consistent with physicochemical and structural determinants previously associated with DILI. Overall, these results support cooperative optimization of MD configuration spaces as an effective strategy for identifying informative and complementary molecular representations with strong predictive utility and interpretability in complex QSAR and predictive toxicology applications.
Authors
- César R. García‐Jacas (ORCID: https://orcid.org/0000-0002-3962-7658)
- Yovani Marrero‐Ponce (ORCID: https://orcid.org/0000-0003-2721-1142)
- Luis A. García‐González (ORCID: https://orcid.org/0000-0003-4021-3345)
- Gabriel Lizama‐Uc (ORCID: https://orcid.org/0000-0002-0092-3636)
- Carlos A. Brizuela (ORCID: https://orcid.org/0000-0003-4621-0380)
- Sergio Andrés Águila
Institutions
- Instituto Tecnológico de Mérida (MX)
- Tecnológico Nacional de México (MX)
- Secretaría de Ciencia, Humanidades, Tecnología e Innovación (MX)
- Centro de Investigación Científica y de Educación Superior de Ensenada (MX)
- Universidad San Francisco de Quito (EC)
- Universidad Panamericana (MX)
- Universidad Nacional Autónoma de México (MX)
Publication Details
- Journal
- ACS Omega
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1021/acsomega.6c10514
- Primary Topic
- Computational Drug Discovery Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Consejo Nacional de Ciencia y Tecnología
- Dirección General de Asuntos del Personal Académico, Universidad Nacional Autónoma de México