EXPLAINABLE ARTIFICIAL INTELLIGENCE (XAI) FOR PREDICTING PROTEIN AGGREGATION AND VISCOSITY IN HIGH-CONCENTRATION LYOPHILIZED BIOLOGICS DURING LONG TERM STORAGE

*Dr. Akhilesh Kumar

High-concentration biologics are increasingly developed for subcutaneous administration because they enable smaller injection volumes and improve patient convenience. However, increasing protein concentration can substantially increase viscosity, protein-protein interactions, self-association, and aggregation. Lyophilization can improve the storage stability of protein therapeutics, but freezing, dehydration, residual moisture, glass-transition behaviour, reconstitution, and long-term storage may introduce additional degradation risks. Conventional stability development is largely dependent on experimental screening and accelerated studies, which can be time-consuming and may not adequately capture complex nonlinear interactions. This study proposes an Explainable Artificial Intelligence (XAI) framework for predicting aggregation and viscosity in high-concentration lyophilized biologics during long-term storage. The framework integrates protein molecular descriptors, formulation composition, lyophilization parameters, storage conditions, and analytical stability measurements. Random Forest and gradient-boosting models are proposed for prediction, while SHapley Additive exPlanations (SHAP) are used to identify the variables responsible for individual predictions. Protein concentration, pH, ionic strength, stabilizer-to-protein ratio, residual moisture, glass-transition temperature, storage temperature, storage duration, and molecular interaction characteristics are considered key predictive variables. The framework separately models aggregation and viscosity while identifying shared molecular and formulation drivers. Existing evidence indicates that concentrated protein viscosity is strongly affected by intermolecular interactions, while lyophilized stability depends on formulation composition, freezing history, residual moisture, and storage conditions. The proposed XAI approach can convert complex multidimensional stability data into interpretable predictions and experimentally testable formulation hypotheses. It provides a transparent computational strategy for accelerating formulation development and improving long-term stability-risk assessment.