Molecular graph-based invariant representation learning with environmental inference and subgraph generation for...

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Molecular representation learning (MRL) is key for predicting molecular properties and drug discovery.
• Current MRL methods struggle with out-of-distribution (OOD) data due to distribution shifts.
• Real-world molecular data often involves dynamic and uncertain environments, necessitating robust generalization.

Purpose of the Study:

• To develop a novel framework, EISG (Integrating Environmental Inference and Subgraph Generation), to improve MRL generalization on OOD data.
• To capture the invariance of molecular graphs across different environmental conditions.
• To enhance the robustness of molecular models in dynamic chemical environments.

Main Methods:

• Proposed EISG framework integrating environmental inference and subgraph generation.
• Developed an unsupervised environmental classification model to identify latent distribution variables.
• Designed a subgraph extractor utilizing information bottleneck theory to extract invariant representations.

Main Results:

• EISG demonstrated strong generalization capabilities across various OOD settings.
• The framework successfully identified invariant graph representations in different environments.
• Experimental results validated the model's improved performance on OOD data.

Conclusions:

• The proposed EISG framework effectively addresses the challenge of OOD generalization in MRL.
• Capturing environmental invariance is crucial for building robust molecular models.
• EISG offers a promising approach for reliable molecular property prediction in diverse real-world scenarios.