Machine Learning Models for Predicting Monoclonal Antibody Biophysical Properties from Molecular Dynamics Simulations...

Area of Science:

• Biotechnology
• Computational Biology
• Pharmaceutical Sciences

Background:

• Monoclonal antibodies (mAbs) are crucial therapeutics, but their development is lengthy and costly.
• Assessing mAb developability early is vital for therapeutic success.
• Key factors include biophysical properties like aggregation, solubility, and viscosity.

Purpose of the Study:

• To develop and validate machine learning (ML) and deep learning (DL) models for predicting mAb biophysical properties.
• To assess the predictive performance of these models against existing methods.
• To introduce a novel DL model, DeepSP, for predicting spatial aggregation and charge properties.

Main Methods:

• Utilized a dataset of 12 biophysical properties for 137 antibodies.
• Employed full-length antibody molecular dynamics simulations.
• Applied machine learning techniques and a new deep learning model (DeepSP) for property prediction.

Main Results:

• Developed ML models that outperform previous methods in predicting most biophysical properties.
• The DeepSP model achieved comparable predictive accuracy to molecular dynamics simulations.
• DeepSP significantly reduced computational time for property prediction.

Conclusions:

• The developed ML and DL models offer efficient and accurate prediction of mAb developability.
• DeepSP provides a faster alternative to traditional simulations for assessing antibody properties.
• Freely available code and a web application (AbDev) facilitate broader adoption and accelerate drug development.