Multi-channel learning for integrating structural hierarchies into context-dependent molecular representation.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Accurate molecular property prediction is crucial for drug discovery but hindered by data scarcity and complex relationships.
• Existing self-supervised learning (SSL) methods for molecules often neglect important chemical knowledge, limiting their effectiveness.

Purpose of the Study:

• To develop a novel multi-channel pre-training framework for molecular SSL that incorporates chemical knowledge.
• To improve the robustness and generalizability of molecular machine learning models, particularly for predicting structure-activity relationships and identifying activity cliffs.

Main Methods:

• A multi-channel pre-training framework leveraging molecular structural hierarchy.
• Distinct pre-training tasks across channels to embed molecular information.
• Task-specific aggregation of channel information during fine-tuning.

Main Results:

• The proposed framework demonstrates competitive performance across various molecular property prediction benchmarks.
• Significant advantages were observed in challenging scenarios, such as predicting activity cliffs, which are critical in drug discovery.

Conclusions:

• The multi-channel SSL approach effectively learns and integrates diverse chemical knowledge for enhanced molecular property prediction.
• This framework offers a promising solution for overcoming limitations in current molecular machine learning models, particularly in complex drug discovery tasks.