Abstract

Graph Neural Networks (GNNs) and Transformer have emerged as dominant tools for AI-driven drug discovery. Many state-of-the-art methods first pre-train GNNs or the hybrid of GNNs and Transformer on a large molecular database and then fine-tune on downstream tasks. However, different from other domains such as computer vision (CV) or natural language processing (NLP), getting labels for molecular data of downstream tasks often requires resource-intensive wet-lab experiments. Besides, the pre-trained models are often of extremely high complexity with huge parameters. These often cause the fine-tuned model to over-fit the training data of downstream tasks and significantly deteriorate the performance. To alleviate these critical yet under-explored issues, we propose two straightforward yet effective strategies to attain better generalization performance: 1. MolAug, which enriches the molecular datasets of down-stream tasks with chemical homologies and enantiomers; 2. WordReg, which controls the complexity of the pre-trained models with a smoothness-inducing regularization built on dropout. Extensive experiments demonstrate that our proposed strategies achieve notable and consistent improvements over vanilla fine-tuning and yield multiple state-of-the-art results. Also, these strategies are model-agnostic and readily pluggable into fine-tuning of various pre-trained molecular graph models. We will release the code and the fine-tuned models.