GenPTM: A Generalizable Framework for Protein Post-Translational Modification Information Extraction from the Scientific Literature

Background: Protein post-translational modification (PTM) plays a pivotal role in cellular activities and biological processes. Although several databases curate PTM information, most cover only a limited number of PTMs. Although the scientific literature continues to accumulate a vast amount of PTM-related knowledge, these databases are not updated regularly. This growing information gap highlights the need for automated information extraction (IE) systems that can identify modified proteins and their specific amino acid sites directly from the literature. While numerous PTMs have been reported in scientific articles, most existing tools are designed only for a few specific PTMs, and developing separate systems for every PTM is not feasible. Methods: To address this challenge, we developed GenPTM, a generalized and adaptable IE tool that identifies modified proteins and sites from PubMed abstracts using a unified text representation strategy. GenPTM replaces PTMspecific modification and chemical group mentions with generic placeholders, allowing the model to focus on shared textual patterns that express modification events. A BiomedBERT-based classifier is fine-tuned to determine whether a candidate protein or site is truly modified, and a post-processing module assembles the final protein, site, or protein–site pair predictions. Results: Trained on five major PTM types (e.g., Ubiquitination, Phosphorylation) and evaluated on eight additional PTMs, including PTMs that are not frequently mentioned (e.g., Citrullination, AMPylation), GenPTM achieves F1- scores ranging from 92% to 96% across all PTMs for three different evaluation categories. Conclusions: These results exhibit strong generalization capability of GenPTM by providing a viable solution for PTM-agnostic IE and automated PTM knowledge discovery in proteomics.