Revealing the limits of covalent docking and advancing affinity prediction with covalent-aware multi-task learning

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Targeted covalent inhibitors (TCIs) are crucial in drug discovery, but computational prediction of their binding poses and affinity remains challenging.
• Existing computational tools often lack specificity for covalent interactions, hindering accurate drug-target affinity (DTA) prediction.

Purpose of the Study:

• To systematically evaluate covalent docking tools using a large, curated benchmark.
• To develop and validate a novel covalent-aware DTA prediction framework to enhance affinity ranking and hit prioritization for TCIs.

Main Methods:

• Curated a benchmark of 2172 covalent protein-ligand complexes from CovalentInDB 2.0 and evaluated four docking engines.
• Assessed the correlation between docking scores and experimental pIC50 values across 17 covalent targets.
• Developed CovMTL-DTA, a multi-task deep learning model integrating ligand graphs, protein embeddings, and cross-modal attention for covalent DTA prediction.

Main Results:

• Boltz-2 demonstrated the best pose-reproduction performance on the benchmark, though potential data leakage was noted.
• Score-affinity correlations from docking were generally weak and target-dependent (|r| < 0.2).
• CovMTL-DTA achieved a Pearson correlation of ~0.77 on an independent test set, outperforming existing methods.
• CovMTL-DTA successfully prioritized known EGFR covalent inhibitors in a virtual screening campaign.

Conclusions:

• Current covalent docking tools have limitations in pose prediction and affinity ranking.
• CovMTL-DTA offers a significant advancement in predicting drug-target affinity for TCIs.
• The developed framework improves hit prioritization and can accelerate covalent drug discovery efforts.