Benchmarking Large Language Models for Assay-Aware Drug-Target Binding Affinity Prediction

Abstract

Machine learning models for molecular discovery often train on bioactivity data from heterogeneous assays, yet assay context is rarely explicitly represented. We investigate whether natural-language assay descriptions provide predictive information beyond molecular and protein representations. We develop ACL-DTI (Assay-aware Contrastive Learning-based Drug Target Interaction) and MAMMAL-AsAP, an assay-context extension of MAMMAL-DTI. Across five triplet-disjoint random splits, assay-aware models achieve competitive or improved performance over assay-agnostic baselines. Among the evaluated models, MAMMAL-AsAP achieves the highest (R^2) of 0.7198 +/- 0.0085. Masking experiments show that ACL-DTI is robust to incomplete assay descriptions, exhibiting the smallest reductions in R^2 among the evaluated models when masking the target name, assay technique, or both: 0.0228 +/- 0.0064, 0.0406 +/- 0.0113, and 0.0620 +/- 0.0190, respectively. These results indicate that assay descriptions contain predictive information not fully captured by molecular structures and protein sequences, and reveal architecture-dependent differences in how this information is exploited.