Targeted covalent inhibitors are a potent but underutilized class of therapeutics. Existing computational covalent screening methods are often limited by the need for prior knowledge of the ligandable target site and by constrained throughput, impeding their utility in early drug discovery. CovSite is presented as a blind covalent screening framework designed to identify candidate reactive residues across the entire protein surface using only a protein structure and an electrophile SMILES string. The framework aims to improve generalizability, accuracy, and throughput for covalent screening workflows.
CovSite applies a sequential pipeline of four orthogonal physicochemical filters to prioritize potential reactive sites and reduce search space. The filters are:
Nucleophile identification: detection of candidate nucleophilic residues on the protein surface.
Solvent accessibility: assessment of residue exposure to solvent, an important determinant of ligand accessibility.
Environment-dependent deprotonation prediction: evaluation of local environment effects that influence the protonation state and nucleophilicity of candidate residues.
Semi-quantum-mechanical reactivity ranking: a semi-quantum approach to rank residue–electrophile reactivity and prioritize likely covalent interactions.
Combining these orthogonal criteria allows CovSite to filter broadly across the protein surface while retaining chemically plausible reactive sites for downstream consideration.
While many covalent discovery efforts focus on cysteine, CovSite extends nucleophilic coverage to include serine, threonine, lysine, histidine, and tyrosine. This broader residue set increases the range of potential ligandable sites and supports discovery for targets lacking a suitably positioned cysteine. The framework therefore enables blind searches for reactive residues across multiple chemically relevant nucleophile classes rather than being restricted to a single residue type.
CovSite was validated using a dataset of 2,062 diverse covalent protein–ligand complexes spanning six nucleophilic residue types. On a held-out benchmark comprising 207 cysteine-targeted complexes, CovSite achieved a 98.5% blind target site hit rate. In this validation it reduced the candidate search space by 97.8%, substantially narrowing the number of sites requiring further evaluation.
The reported target-site hit detection using CovSite exceeds the 53–62% accuracy reported for popular covalent screening tools when those tools were applied under non-blind conditions on the same benchmark set. These comparative figures indicate improved blind-site detection performance and more aggressive search-space reduction relative to existing approaches.
A key design goal for CovSite is high throughput. The authors report that screening a 200-residue protein completes in two to three minutes on standard hardware. This level of efficiency addresses a common bottleneck in covalent screening: the time and computational cost of evaluating many potential residues and ligand combinations. The speed of CovSite supports its use in early-stage discovery and in iterative workflows where many protein–electrophile pairings must be evaluated rapidly.
CovSite is presented not only as a blind, target-agnostic detector of reactive residues, but also as an enabler for ligand-specific workflows. Because it requires only protein structure and electrophile SMILES, CovSite can be integrated into iterative, machine-learning–driven covalent inhibitor generation pipelines that were previously impractical with older tools. The authors position CovSite as a platform technology that can support computationally guided covalent drug discovery for novel and understudied targets by facilitating rapid, blind detection of viable covalent attachment points.
This work is reported as a preprint on bioRxiv and thus has not been peer reviewed. The source notes funding from the National Cancer Institute and the Prostate Cancer Foundation. Specific implementation details beyond the high-level pipeline, as well as per-residue or per-target examples from the validation set, are not reported in the provided source summary. Readers should interpret performance metrics within the context of a preprint report pending peer review.
CovSite delivers a blind, high-throughput framework for covalent screening that combines four orthogonal physicochemical filters, expands nucleophilic coverage beyond cysteine, and demonstrates strong benchmark performance with substantial search-space reduction. Its reported speed and blind-site accuracy suggest utility as a platform for early-stage covalent inhibitor discovery and for integration into machine-learning–driven design workflows. The findings are from a preprint and have not yet undergone peer review.