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

Targeted covalent inhibitors (TCIs) have become an important modality in modern drug discovery, but computational tools for covalent pose prediction and quantitative affinity ranking remain underdeveloped. We constructed a large, structure and activity-resolved benchmark to systematically evaluate covalent docking and to develop a covalent-aware drug-target affinity (DTA) prediction framework. Starting from CovalentInDB 2.0 and related structural resources, we curated 2172 high quality covalent protein-ligand complexes spanning diverse protein classes and nine electrophilic warhead types, and used them to benchmark four docking engines (AutoDock4, CovDock in the Schrödinger Suite, GNINA and Boltz-2) in a self-docking setting. Boltz-2 shows the strongest pose-reproduction performance on our structure-resolved benchmark. However, because co-folding engines are trained on broad PDB corpora and our benchmark is also derived from PDB-resolved complexes, potential train-test overlap is likely; thus, Boltz-2 results are reported as a reference upper bound rather than a leakage-free estimate of prospective generalization. Across 17 covalent targets with quantitative IC₅₀ data, we further assessed the relationship between docking scores and experimental pIC₅₀ values and found that score-affinity correlations are generally weak and highly target dependent, with |r| < 0.2 for most target-software pairs and even pronounced negative correlations for several systems. We propose CovMTL-DTA to overcome these limitations, a covalent-aware multi-task DTA model that integrates ligand molecular graphs augmented with SMARTS-based warhead descriptors, pretrained protein sequence embeddings, cross-modal ligand-protein attention, and a task-relation module for inter-target transfer. Trained on curated covalent ligand-target pairs, the model outperforms classical machine-learning regressors and state-of-the-art deep DTA baselines, achieving a Pearson correlation of ∼0.77 with reduced RMSE and MAE on an independent test set. In an EGFR-focused virtual screening of ∼14 000 Michael-acceptor-containing compounds, the model prioritizes three clinically relevant EGFR covalent inhibitors within the top 1% of the ranked library and identifies structurally novel, favorable physicochemical properties hits. Our benchmark and model highlight both the strengths and limitations of current covalent docking and demonstrate how covalent-specific representations and multi-task learning can substantially improve affinity prediction and hit prioritization in covalent drug discovery.