FTMMR: Fusion Transformer for Integrating Multiple Molecular Representations

Area of Science:

• Computational chemistry
• Drug discovery
• Materials science

Background:

• Molecular property prediction is crucial for bio-chemical applications.
• Existing methods struggle with fusing diverse molecular representations (1D, 2D, 3D).
• Limited data and heterogeneous data characteristics pose challenges.

Purpose of the Study:

• To develop an advanced framework for integrating multiple molecular representations.
• To enhance the accuracy and efficiency of molecular property prediction.
• To address the challenges of data heterogeneity and limited chemical data.

Main Methods:

• Introduced the Fusion Transformer for Multiple Molecular Representations (FTMMR) framework.
• Employed distinct representation-specific networks for each molecular representation.
• Utilized a fusion transformer architecture to integrate information.
• Implemented self-supervised learning and a combinatorial contrastive loss function.

Main Results:

• FTMMR successfully generated fused molecular representations.
• The framework effectively aligned heterogeneous molecular representations.
• Evaluated on seven benchmark datasets, FTMMR demonstrated superior performance.
• Outperformed existing fusion and self-supervised learning methods.

Conclusions:

• FTMMR offers a powerful approach for molecular property prediction.
• The fusion strategy and self-supervised learning effectively handle diverse molecular data.
• This framework advances the field of computational chemistry and drug discovery.