TGF-M: Topology-augmented geometric features enhance molecular property prediction

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Design
• Machine Learning for Molecular Properties

Background:

• Accurate molecular property prediction is vital for AI-driven drug design (AIDD).
• Existing models struggle to balance predictive accuracy with computational complexity.
• Molecular topology and geometry offer crucial spatial information but increase model complexity.

Purpose of the Study:

• To introduce TGF-M (Topology-augmented Geometric Features for Molecular Property Prediction), a novel model for molecular property prediction.
• To optimize feature extraction for enhanced information capture and improved accuracy.
• To reduce model complexity and computational cost.

Main Methods:

• Developed TGF-M, a model integrating topological and geometric molecular features.
• Optimized feature extraction to balance information richness and computational efficiency.
• Evaluated performance on the re-segmented PCQM4Mv2 dataset for HOMO-LUMO gap prediction.

Main Results:

• TGF-M achieved a low mean absolute error (MAE) of 0.0647 in HOMO-LUMO gap prediction.
• The model utilizes only 6.4M parameters, significantly less than state-of-the-art models.
• Demonstrated comparable performance to existing methods with less than one-tenth the parameters.

Conclusions:

• TGF-M effectively leverages molecular topology and geometry for accurate property prediction.
• The model offers a favorable balance between accuracy and computational efficiency.
• TGF-M shows significant potential for advancing AI-driven drug design.