Reduction of monoclonal antibody viscosity using interpretable machine learning.

Area of Science:

• Biotechnology
• Protein Engineering
• Computational Biology

Background:

• Early identification of antibody candidates with favorable drug-like properties is crucial for efficient therapeutic development.
• Subcutaneous antibody formulations require low self-association to achieve high concentrations while minimizing viscosity, opalescence, and aggregation.

Purpose of the Study:

• To develop an interpretable machine learning model for predicting antibody (IgG1) viscosity based on variable (Fv) region sequences.
• To identify key Fv sequence features that correlate with antibody viscosity.

Main Methods:

• A machine learning model was trained on antibody viscosity data (>100 mg/mL mAb concentration) at pH 5.2.
• The model utilizes antibody sequences to predict viscosity, focusing on Fv region features: isoelectric point (pI), hydrophobic patch size, and negative charge patch count.
• Model performance was validated on training, test, and previously reported datasets.

Main Results:

• The model identified low Fv isoelectric points (pI < 6.3) as a primary predictor of high antibody viscosity across diverse antibody germlines and clinical-stage IgG1s.
• The developed model demonstrated high accuracy in identifying viscous antibodies.
• The interpretable nature of the model facilitated the design of mutations that experimentally reduced antibody viscosity.

Conclusions:

• An interpretable machine learning model can predict antibody viscosity from Fv sequences, aiding in early identification of drug-like antibody candidates.
• Low Fv pI is a significant factor contributing to high antibody viscosity.
• This approach can streamline antibody drug development by reducing experimental screening and improving candidate selection.