A deep learning framework for predicting molecular property based on multi-type features fusion

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Molecular property prediction is crucial for drug discovery.
• Existing methods like sequence-based and graph-based representations have limitations in capturing comprehensive molecular information.
• Integrating diverse molecular features can enhance prediction accuracy.

Purpose of the Study:

• To propose a novel deep learning fusion model (DLF-MFF) for molecular property prediction.
• To leverage the strengths of multiple molecular representations simultaneously.
• To improve the accuracy and efficiency of predicting molecular properties and identifying potential drug candidates.

Main Methods:

• Extracting four types of molecular features: fingerprints, 2D graphs, 3D graphs, and images.
• Utilizing four deep learning frameworks to learn features from each representation individually.
• Integrating these features into a unified representation for property prediction.
• Comparing DLF-MFF against seven state-of-the-art methods on six benchmark datasets.

Main Results:

• DLF-MFF achieved state-of-the-art performance across six benchmark datasets for molecular property prediction.
• The model successfully identified potential anti-SARS-CoV-2 inhibitors from a library of 2500 drugs.
• DLF-MFF demonstrated superior performance in predicting 3CL protease inhibition and binding affinity.

Conclusions:

• The proposed DLF-MFF model effectively integrates multi-type molecular features for enhanced property prediction.
• This approach offers a promising direction for accurate molecular property prediction and drug repurposing, particularly for infectious diseases like COVID-19.