Parameter Efficient Deep Learning Models for Multi-Target Binding Affinity and hERG Cardiotoxicity Prediction

Accurately predicting binding affinity and toxicity is critical to drug discovery, offering the potential to reduce development costs and enhance safety profiles. Traditional computational methods, such as molecular docking and quantitative structure-activity relationship (QSAR) models, rely heavily on handcrafted features and domain expertise, limiting their scalability and generalizability to novel compounds. Recent advances in chemical language models (CLMs), which leverage Simplified Molecular Input Line Entry System (SMILES) representations, offer a scalable alternative by capturing complex molecular properties through learned representations. In this study, we propose a dual-paradigm framework for binding affinity prediction of 3 protein targets and cardiotoxicity for human ether-á-go-go related gene (hERG). The first component utilizes a specialized graph neural network (GNN) that operates directly on molecular graphs, treating atoms as nodes and bonds as edges. The second component employs a CLM fine-tuned via Low-Rank Adaptation (LoRA), enabling efficient adaptation of large pre-trained models to specific toxicological tasks with minimal parameter overhead. Our hybrid framework demonstrates superior performance, achieving an average AUROC of 0.92 across three protein targets, while the LoRA-adapted CLM attains an AUROC of 0.93 for cardiotoxicity prediction, offering a 98% reduction in trainable parameters, outperforming the performance of existing models.