Low-data interpretable deep learning prediction of antibody viscosity using a biophysically meaningful representation

Area of Science:

• Biophysics
• Computational Biology
• Machine Learning

Background:

• Deep learning models typically require large datasets, which are often unavailable for specific scientific problems.
• Predicting therapeutic antibody viscosity is crucial for drug development but has been hindered by data scarcity, limiting deep learning applications.

Purpose of the Study:

• To develop a generalizable deep learning model for predicting high-concentration therapeutic antibody viscosity using limited data.
• To introduce PfAbNet-viscosity, a novel 3D convolutional neural network architecture for viscosity prediction.

Main Methods:

• Utilized a biophysically meaningful representation: the electrostatic potential surface of the antibody variable region.
• Developed and trained a 3D convolutional neural network (PfAbNet-viscosity) using this representation as input.
• Evaluated model generalizability and accuracy with limited training data (as few as a dozen data points).

Main Results:

• PfAbNet-viscosity achieved high accuracy in predicting antibody viscosity even with minimal training data.
• Feature attribution analysis confirmed that the model learned key biophysical drivers of viscosity.
• Demonstrated the model's ability to generalize effectively under data-scarce conditions.

Conclusions:

• A novel deep learning approach (PfAbNet-viscosity) can accurately predict therapeutic antibody viscosity from limited data.
• The use of biophysically meaningful features is effective for developing generalizable models in data-limited scientific domains.
• The presented methodology shows potential applicability to other biological systems and prediction tasks.