Machine Learning for Colloidal Stability and Aggregation Risk in Biopharmaceutical Formulations: Evidence, Limits, and Practical Use
Year: 2026
Type: article
Study design: Systematic Review
Abstract: Colloidal instability remains a dominant cause of product failure, manufacturing attrition, and safety risk across biopharmaceutical modalities including monoclonal antibodies (mAbs), bispecific antibodies, antibody-drug conjugates (ADCs), and mRNA-lipid nanoparticle systems. Protein aggregation is a critical quality attribute (CQA) under ICH Q6B, linked to immunogenicity, reduced potency, and adverse patient outcomes. Despite the transformative impact of machine learning (ML) on protein structure prediction and molecular design, its application to formulation-dependent colloidal stability prediction remains fragmented, poorly benchmarked, and largely disconnected from regulatory frameworks. This review systematically examines ML approaches for predicting aggregation propensity, viscosity, solubility, liquid-liquid phase separation, and shelf-life across biopharmaceutical modalities. We critically assess experimental data sources, feature engineering strategies, and ML architectures spanning classical models, deep learning, graph neural networks, and protein language models, alongside the emerging role of explainable AI (XAI). No standardised, cross-modality ML benchmarking framework for colloidal stability currently exists -- a gap that constrains generalisation, reproducibility, and regulatory acceptance. Principal unresolved challenges include dataset scarcity, label noise, external validation deficits, and proprietary data silos. A decade roadmap for integrating physics-informed ML, autonomous formulation laboratories, and foundation models into next-generation biologics development is proposed.
Source: Entropic and Disordered Matter
DOI: 10.3390/edm1010003
Author: Carlos Victor Montefusco-Pereira
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: diamond
OpenAlex ID: W7203641979