LightDTA: lightweight drug-target affinity prediction via random-walk network embedding and knowledge distillation.

Accurately predicting drug targeting affinity is crucial in the field of drug discovery. With the rapid development of artificial intelligence, many deep learning methods have been proposed for drug target affinity prediction tasks. However, most existing methods rely heavily on a detailed description of biochemical attributes of inputs; besides, the model architecture is getting increasingly complex just to achieve a slight performance gain. Together, these poses great challenges for real-world employments and applications. This study proposes a new lightweight framework, LightDTA, which combines knowledge distillation and random walk algorithms to predict drug target affinity. It adopts a lightweight network-based protein representation and eliminates the tedious process of collecting detailed biochemical properties. A knowledge distillation framework is further introduced to enrich molecular-level knowledge and enhance predictive capability while not affecting the model efficiency. Comprehensive experiments show that LightDTA achieves state-of-the-art performance in both classification and regression tasks, with only 61% of the memory requirements of the suboptimal baseline model. It also achieves a 7 × speedup in inference time. Therefore, the proposed method offers a highly efficient and accurate model for real-world prediction of drug-target affinities. The code for LightDTA is publicly available at: https://github.com/Huang-zilin/LightDTA-final .