Escaping the Drug-Bias Trap: Using Debiasing Design to Improve Interpretability and Generalization of Drug-Target...

Area of Science:

• Computational chemistry
• Bioinformatics
• Artificial intelligence in drug discovery

Background:

• Virtual screening is crucial for AI-assisted drug discovery due to high experimental costs.
• Existing deep learning models for drug-target interactions (DTIs) suffer from a drug-bias trap, overestimating accuracy and lacking interpretability.
• This bias leads to underutilization of protein information, questioning model generalizability.

Purpose of the Study:

• To introduce UdanDTI, an innovative deep-learning architecture for predicting drug-protein interactions.
• To address the limitations of existing DTI models, specifically the drug-bias trap and lack of interpretability.
• To enhance the biological interpretability and generalizability of DTI prediction models.

Main Methods:

• Developed UdanDTI, a deep-learning architecture featuring an unbalanced dual-branch system.
• Incorporated an attentive aggregation module to improve the utilization of protein information.
• Evaluated UdanDTI on various public datasets and compared its performance against state-of-the-art models.

Main Results:

• UdanDTI demonstrated superior performance across in-domain, cross-domain, and structural interpretability settings.
• The model achieved exceptional accuracy in predicting drug responses for specific Epidermal Growth Factor Receptor (EGFR) mutations in non-small cell lung cancer.
• Results were consistent with experimental findings and showed UdanDTI complements molecular docking software like DiffDock.

Conclusions:

• UdanDTI effectively overcomes the drug-bias trap in DTI prediction models.
• The architecture enhances biological interpretability and generalizability, offering reliable predictions.
• UdanDTI shows significant potential for accelerating drug discovery, particularly for targeted therapies like those for EGFR mutations.