MMF-MCP: A Deep Transfer Learning Model Based on Multimodal Information Fusion for Molecular Feature Extraction and...

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Accurate prediction of molecular carcinogenicity is vital for cancer prevention and drug development.
• Current deep learning models face limitations in data quality, feature richness, accuracy, robustness, and interpretability.

Purpose of the Study:

• To develop an improved deep transfer learning model for molecular carcinogenicity prediction.
• To enhance accuracy, robustness, and interpretability using multimodal information fusion.

Main Methods:

• Proposed MMF-MCP model utilizing deep transfer learning and multimodal information fusion.
• Extracted molecular graph features (GATs) and fingerprint features (CNNs).
• Processed molecular images using SE-ResNet18; applied transfer learning by pretraining on mutagenicity data.

Main Results:

• MMF-MCP achieved high performance on benchmark carcinogenicity datasets (ACC: 0.8452, AUC: 0.8513, SE: 0.8571, SP: 0.8333).
• Significantly outperformed existing state-of-the-art molecular carcinogenicity prediction methods.
• Visualization results demonstrated strong interpretability, aiding in understanding critical molecular features.

Conclusions:

• MMF-MCP offers a robust and interpretable approach for molecular carcinogenicity prediction.
• The model's performance and generalization ability are enhanced through transfer learning and multimodal fusion.
• Provides valuable insights for cancer research, prevention, and pharmaceutical development.