Multimodal framework for early developability assessment to accelerate protein and antibody development

Protein and antibody therapeutics, distinguished by exceptional activity, specificity, and precise biological functions, are crucial in modern pharmaceuticals. Their clinical success depends on developability, encompassing physicochemical properties to ensure stability and safety and are governed by protein characteristics and formulation composition. However, formulation development of biopharmaceutical drugs remains hindered by inefficient and expensive trial-and-error experiments. In this study, we constructed four developability-related datasets, each containing both protein and excipient information. The datasets cover conformational stability with 986 entries, colloidal stability with 919 entries, viscosity with 900 entries, and solubility with 749 entries, where each data entry represents a single measurement of a specific developability property under certain protein and formulation conditions such as pH, ionic strength, excipient, etc. The multimodal deep learning framework (Formulation_Protein) was designed to capture the complex interplay among three-dimensional structural data, sequence, amino acid composition descriptors, and formulation compositions to enable accurate prediction of four protein formulation developability parameters. This architecture leverages transfer learning in conjunction with conventional machine learning algorithms to enable comprehensive feature representation and prediction. Formulation_Protein demonstrated good performance, obtaining average accuracies of 0.925 for conformational stability, 0.858 for colloidal stability, 0.917 for viscosity, and 0.742 for solubility on the test sets. Further experimental validation was performed on 9 proteins across 65 formulations. In conclusion, this study present Formulation_Protein, a multimodal deep learning framework, for comprehensive developability assessment in early stage to accelerate protein and antibody development.