Deep cross-modal representation fusion learning for enhanced drug target affinity prediction

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Drug target affinity (DTA) prediction is crucial for efficient drug discovery.
• Experimental DTA estimation is costly and time-consuming.
• Protein and drug structural flexibility poses challenges for traditional methods.

Purpose of the Study:

• To develop an accurate and efficient computational model for DTA prediction.
• To leverage multi-modal deep learning for enhanced DTA estimation.
• To reduce the need for extensive experimental validation in drug discovery.

Main Methods:

• A three-stage deep cross-modal representation fusion framework (PCBERT-GAT-DFFNN-DTA).
• Utilized ProtBERT for protein sequence embeddings and ChemBERT/GAT for drug sequence/structure embeddings.
• Employed deep feed-forward neural networks for final affinity prediction.

Main Results:

• The PCBERT-GAT-DFFNN-DTA model outperformed baseline methods in S1-S3 settings.
• Achieved strong performance on KIBA and competitive results on Davis/Metz datasets in the S4 blind setting.
• Demonstrated significant improvements in R² scores compared to LLMDTA across all datasets.

Conclusions:

• The proposed framework effectively captures complex drug-protein interactions for reliable DTA prediction.
• PCBERT-GAT-DFFNN-DTA offers a promising computational approach to accelerate drug discovery.
• The multi-modal fusion strategy enhances prediction accuracy for drug target binding affinity.