Enhancing molecular property prediction with auxiliary learning and task-specific adaptation

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Pretrained Graph Neural Networks (GNNs) are vital for molecular property prediction.
• Standard fine-tuning of GNNs often results in poor generalization.
• Adapting pretrained models effectively is crucial for advancing molecular tasks.

Purpose of the Study:

• To enhance the generalization of pretrained GNNs for molecular property prediction.
• To investigate auxiliary learning strategies for adapting GNNs.
• To address the challenge of negative transfer in model adaptation.

Main Methods:

• Jointly training pretrained GNNs with multiple auxiliary tasks.
• Developing strategies to measure auxiliary task relevance.
• Implementing adaptive gradient combination and bi-level optimization for task weighting.
• Proposing a novel gradient surgery technique, Rotation of Conflicting Gradients ($$).

Main Results:

• Proposed methods improved performance by up to 7.7% compared to traditional fine-tuning.
• Demonstrated the efficacy of auxiliary learning in enhancing GNN generalizability.
• Showcased the effectiveness of novel gradient surgery techniques.

Conclusions:

• Auxiliary task integration is an effective strategy for improving pretrained GNN generalizability in molecular property prediction.
• The proposed framework offers a significant advancement over the standard pretraining-fine-tuning approach.
• This work provides valuable tools for drug discovery, particularly for data-limited scenarios.