Consensus modeling of sex hormone-binding globulin binding using machine learning and deep learning approaches

Sex hormone-binding globulin (SHBG) regulates the bioavailability of circulating androgens and estrogens through high-affinity binding. Chemicals that bind SHBG may alter the equilibrium between bound and free hormone fractions, making SHBG binding activity a relevant endpoint for endocrine disruption assessment. However, experimental assessment of SHBG binding is resource intensive with low throughput, limiting its application to large chemical inventories. To enable large-scale screening, we developed machine learning, deep learning, and consensus models using a dataset of 125 chemicals tested in a human SHBG competitive binding assay. Random forest, support vector machine, and multilayer perceptron models were trained using molecular descriptors, while a graph isomorphism network with edge features was trained using molecular graph representations. A consensus model was then generated by averaging the predicted probabilities from the four individual models. Model performance was evaluated using 100 iterations of 5-fold cross-validation and external validation with an independent dataset curated from ChEMBL. All models showed good predictive performance, with accuracy ranging from 0.811 to 0.855 in internal validation and from 0.793–0.890 in external validation. The consensus model had the best overall performance, yielding accuracy of 0.855 and 0.884 in internal and external validations, respectively. Confidence analysis further showed that prediction accuracy increased with prediction confidence, indicating the practical utility of confidence estimates for decision-making. These results indicate that consensus modeling can reliably predict SHBG-binding activity and provide a rapid, cost-effective approach for screening large chemical inventories. The proposed models can support the prioritization of chemicals for experimental testing, facilitate the identification of potential endocrine-disrupting substances, and complement New Approach Methodologies (NAMs) used in chemical safety assessment.