Multimodal deep learning with a joint uncertainty quantification scheme for drug-target interaction prediction

Area of Science:

• Computational chemistry
• Bioinformatics
• Artificial intelligence in drug discovery

Background:

• AI-driven drug-target interaction (DTI) prediction is crucial but limited by model and data uncertainties.
• Uncertainty affects the reliability and accuracy of DTI prediction models.

Purpose of the Study:

• Introduce EUQTri-DTI, an evidence-guided, uncertainty quantification-based multimodal deep learning framework for DTI prediction.
• Address limitations in DTI prediction accuracy and reliability caused by uncertainties.

Main Methods:

• Integrate three modality-specific networks: 1D protein sequences, 2D molecular images, and 3D drug structures.
• Employ a bidirectional cross-attention mechanism for inter-modal information exchange.
• Implement a joint uncertainty quantification scheme using evidential uncertainty and prediction entropy.

Main Results:

• EUQTri-DTI demonstrates stable and competitive performance across three benchmark datasets (DrugBank, KIBA, Davis).
• Achieved high ROC-AUC (up to 93.45%) and PR-AUC (up to 85.32%) values, outperforming baseline methods.
• Uncertainty analysis confirms higher uncertainty for misclassified samples, enabling reliability-aware decision-making.

Conclusions:

• EUQTri-DTI offers improved DTI prediction accuracy and provides sample-level confidence assessment.
• The framework shows potential for uncertainty-aware virtual screening in drug discovery.
• Combines predictive performance with reliability assessment for more robust AI in drug development.