Transfer learning and meta-learning for drug discovery

While deep learning accelerates drug discovery, its demand for large labeled datasets and computational power clashes with the resource-intensive reality of biomedical research, especially for novel targets lacking data. Transfer learning addresses this by adapting knowledge from related tasks, domains, or pre-trained models, minimizing reliance on the large scale of data. Meta-learning and multi-task learning extend these benefits: meta-learning enables rapid adaptation to new tasks with minimal examples, while multi-task learning shares insights across correlated tasks or objectives. Although conventional transfer learning risks overfitting or irrelevant knowledge transfer, recent advances tailor these methods to drug discovery's unique challenges, such as molecular structure representation and target specificity. This chapter analyzes how transfer learning, meta-learning, and multi-task learning tackle low-data challenges in drug discovery, covering core concepts, algorithmic innovations, and real-world impacts.