Recently, the research team led by Chen Luonan from the Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences (UCAS) / Shanghai Jiao Tong University, in collaboration with MOMED BIOTECH, published an online research paper in the journal PNAS titled "Mining Lysine Post-Translational Modification Site by Integrating Protein Language Model Representations with Structural Context." This study proposes a deep learning framework that integrates protein language models with atomic-level structural features, achieving accurate prediction of various lysine post-translational modification (PTM) sites. The functional impact of these sites was further validated through molecular dynamics simulations, offering new insights for PTM-related mechanistic research and drug target development. The co-first authors are Dr. Luo Mengqi from the Hangzhou Institute for Advanced Study, UCAS, and Dr. Zhu Xiaohong from MOMED BIOTECH. The co-corresponding authors are Dr. Chen Luonan from the Hangzhou Institute for Advanced Study, UCAS / Shanghai Jiao Tong University, Dr. Warshel from the University of Southern California / MOMED BIOTECH, and Dr. Bai Chen.

Lysine PTMs are critical molecular events that regulate protein function, signal transduction, and disease progression. However, their experimental identification still faces challenges such as high costs and low throughput. Most existing computational methods rely on manually designed sequence features, struggle to effectively integrate three-dimensional structural information, and are often limited to a single PTM type, resulting in poor generalization. Constructing an intelligent prediction model capable of integrating multi-source information and uniformly mining multiple PTM types has become a bottleneck that needs to be broken in this field.

This study proposes a dual-module deep learning framework. On one hand, it utilizes the protein language model ESM-2 to extract semantic features from sequences. On the other hand, based on structural data provided by AlphaFold, it constructs atomic-level contact maps and captures spatial structural information through a Graph Convolutional Network (GCN). After fusing these two types of features, PTM site classification is achieved via a Multilayer Perceptron. This framework demonstrates stable and excellent predictive performance across six common types of lysine PTMs (including acetylation, succinylation, crotonylation, etc.), with F1-scores reaching up to 80.9% and AUC up to 88.3%. It maintains strong generalization ability under different data partitioning strategies, with a small fluctuation range in results upon repeated evaluation (±1.0–1.5%).

This study not only proposes a unified prediction framework scalable to multiple PTM types but also integrates AI predictions with biological function validation through dynamic simulations, overcoming the limitations of traditional static prediction methods. This "prediction-validation" closed-loop process provides important tools and a theoretical basis for future PTM functional exploration, disease mechanism analysis, and the design of targeted drugs based on key modification sites.

Interested readers can access the original research paper at:https://www.pnas.org/doi/10.1073/pnas.2529141123

The code and data associated with this article have been made publicly available on the GitHub platform(https://github.com/qi29/lysine-PTM-site-Mining).