Classification Pretraining Enhances Performance of Target-Based Affinity Prediction for Reproductive Toxicity Assessment

Reproductive toxicity is challenging to assess because
of its diverse
phenotypes and complex mechanisms; target-prediction models enable
early identification of toxic potential. However, affinity prediction
models suffer from limited training data and poor generalization to
novel chemical scaffolds, which hinders their application in prospective
toxicity assessment. Here, we propose a classification pretraining-regression
fine-tuning framework that leverages large-scale binary activity data
to learn generalizable compound-protein interaction patterns. Systematic
evaluations across six dual-encoder models and three data-split strategies
demonstrate consistent performance gains. Under the most stringent
cluster-split scenario, our framework achieved an average R2 improvement of 0.324, demonstrating substantially
enhanced generalization to novel chemotypes. Using the best-performing
model (R2 = 0.797; MSE = 0.370), we predicted
compound affinities across 81 reproductive targets, generating target
affinity spectra that distinguished toxicity effects, identified both
broad-spectrum and target-specific structural alerts, and revealed
key targets for six sex-specific reproductive diseases. By predicting
binding affinity, our framework not only enhances generalization but
also enables interpretable, mechanism-guided reproductive toxicity
assessment.