Simple Graph Sampling-Based Meta-Learning for Molecular Property Prediction
Wang Chushi, Jia Shiyu, Ren Xinyue · 2024
We focuses on the challenges and advances in molecular property prediction within the field of drug discovery, particularly when addressing small sample sizes. In seeking to enhance predictive accuracy while reducing costs and time expenditure, novel computational approaches are explored. Notably, the integration of meta-learning and contrastive learning has demonstrated potential, as evidenced in methods. Despite these advancements, persistent issues such as slow model training and sensitivity to data noise undermine practical applicability. Our research introduces a new optimization strategy that incorporates random noise during subgraph sampling in the GS-Meta model to tackle these challenges. The approach significantly increases the robustness and accuracy of molecular property predictions. Our findings, corroborated by robust performance metrics, indicate the promising potential of random noise optimization in molecular property prediction models. Future research should continue to explore these techniques' generalizability across larger datasets and the impact of different noise injection strategies.