AffinityLM: Binding-Site Informed Multitask Language Model for Drug-Target Affinity Prediction

Drug-target affinity (DTA) prediction is crucial for drug discovery and repurposing. Current DTA prediction models are trained on datasets biased towards abundance of unique molecules relative to target proteins. Such models excel at molecule screening but struggle with inverse drug screening and generalization to novel targets. This data bias limitation hinders drug repurposing efforts and the development of truly general solutions capable of understanding interactions between any given drug and target. To address these challenges, we present AffinityLM, a multitask transformer model that creates comprehensive joint-feature representations of drug-target compounds. Our approach leverages both molecule-biased binding affinity and protein-biased binding site data, thus balancing the protein-molecule data distribution. This approach enables learning from a significantly more diverse dataset with more drug-target pairs, with the potential to improve accuracy and generalization. Despite being trained on a relatively smaller subset of data, comparative evaluations against state-of-the-art models on standard benchmarks like DAVIS and KIBA show that AffinityLM matches or surpasses existing models for binding affinity prediction. Our results demonstrate that models forced to learn rich feature spaces for drug-target interactions offer superior performance and versatility in DTA prediction tasks, paving the way for more effective and broadly applicable drug discovery strategies.