Advanced Machine Learning Models Accurately Predict Asparagine Deamidation in Biologics, Boosting Stability

The spontaneous deamidation of asparagine residues is a critical challenge in developing stable and effective protein therapeutics. This post-translational modification can alter a protein's structure, function, and immunogenicity, leading to reduced efficacy and shelf-life. Existing predictive models often suffer from limited generalizability and sequence similarity biases, making it difficult to reliably identify deamidation hotspots early in the drug development pipeline. A more robust in-silico prediction method is needed to mitigate costly late-stage interventions.

Researchers developed advanced machine learning models, integrating protein language models (PLMs) like ESM2 with graph neural networks (GNNs), to predict asparagine deamidation. The models were trained on a comprehensive dataset comprising 591 asparagine sites from over 105 distinct protein molecules. To counter data leakage and ensure robust performance estimates for novel deamidation sites, a specialized peptide grouping strategy was implemented. This approach aimed to provide more accurate and unbiased predictions compared to traditional methods.

The study demonstrated that, when controlling for sequence similarity bias, PLMs achieved prediction accuracy comparable to traditional feature-based models that rely on amino acid composition and predicted secondary structure/solvent accessibility, while offering significant computational efficiencies. The integration of a GNN-based pipeline further enhanced predictive power.

This GNN-based approach increased prediction accuracy by approximately ~8% compared to language model-only tools and delivered a substantial 15%-25% improvement over conventional motif-based prediction methods. This methodological framework enables more reliable and rapid in-silico identification of deamidation liabilities, crucial for protein therapeutic design.