Accurate Identification of Covalently Ligandable Cysteines Using CCSite.
TL;DR
CCSite is presented, a protein language model-based framework for discovering covalently ligandable cysteines from protein sequences that uniquely integrates low-rank adaptation of ESM Cambrian with a cysteine-centered encoder-decoder module to capture local microenvironment features and long-range contextual information.
Abstract
Targeted covalent inhibition is an important strategy in modern drug discovery, with cysteine being the most common residue targeted for covalent ligands. Accurate identification of covalently ligandable cysteines is therefore essential, especially for traditionally "undruggable" targets. However, structure-based methods depend on available and reliable protein structures, while sequence-based methods remain scarce and require further improvement. Here, we present CCSite, a protein language model-based framework for discovering covalently ligandable cysteines from protein sequences. It uniquely integrates low-rank adaptation of ESM Cambrian (LoRA-ESMC) with a cysteine-centered encoder-decoder module to capture local microenvironment features and long-range contextual information. Benchmarking and independent evaluation showed that CCSite achieved competitive performance without requiring 3D structures. Moreover, a real-world application revealed that CCSite was capable of prospectively identifying experimentally validated covalent cysteines, and a large-scale screen of human pathogenic X-to-Cys mutations further identified over 2,000 neo-cysteines as covalently ligandable candidates for future covalent drug development.