Capturing induced-fit effects: A geometry-aware and interpretable framework for robust drug-target affinity...

Area of Science:

• Computational Chemistry and Molecular Modeling
• Pharmacology and Drug Discovery
• Artificial Intelligence in Bioinformatics

Background:

• Accurate drug-target affinity (DTA) prediction is crucial for efficient virtual screening in drug discovery.
• Conventional DTA models often use static molecular representations, neglecting dynamic induced-fit effects crucial for binding.
• Deep learning models in DTA prediction can suffer from a "black-box" nature, limiting interpretability.

Purpose of the Study:

• To develop a geometry-aware and interpretable framework (DCR-DTA) for accurate DTA prediction.
• To explicitly model bidirectional induced-fit interactions by focusing on stable structural anchors.
• To enhance pretrained representations by mitigating feature anisotropy and capturing interaction manifold geometry.

Main Methods:

• Proposed a novel framework, DCR-DTA, integrating Dynamic Contextual Regularization.
• Employed a geometry-aware approach prioritizing stable structural anchors over raw 3D displacements.
• Validated the model on Davis and KIBA benchmarks, evaluating performance using MSE, r_m², and Concordance Index (CI).

Main Results:

• DCR-DTA consistently outperformed state-of-the-art baselines, especially in challenging cold-start scenarios.
• Achieved superior external validation metrics: r_m² score of 0.787 on KIBA and a Concordance Index (CI) of 0.902.
• Visualization analyses confirmed that the model learns biologically intuitive and discriminative interaction patterns.

Conclusions:

• Explicitly modeling interaction dynamics and representation geometry is essential for robust and explainable DTA prediction.
• DCR-DTA offers a significant advancement in understanding binding mechanisms and improving drug discovery pipelines.
• The proposed framework provides interpretable insights, addressing the limitations of traditional "black-box" deep learning models.