Harnessing Fc/FcRn Affinity Data from Patents with Different Machine Learning Methods

Area of Science:

• Biotechnology
• Pharmacology
• Computational Biology

Background:

• Monoclonal antibodies have long half-lives due to Fc-FcRn binding.
• Fc engineering enhances this pharmacokinetic property, leading to new drug approvals.
• Existing methods for Fc variant discovery include structure-guided design and mutagenesis.

Purpose of the Study:

• To explore machine learning for predicting Fc variants with enhanced neonatal receptor (FcRn) binding affinity.
• To develop and validate predictive models for Fc-FcRn interactions.
• To identify novel Fc variants for improved therapeutic antibody half-lives.

Main Methods:

• Compiled 1323 Fc variants from patent literature affecting FcRn affinity.
• Trained machine learning algorithms, including support vector regressor (SVR), to predict FcRn binding.
• Validated models using 10-fold cross-validation, in silico mutagenesis, and experimental surface plasmon resonance (SPR).

Main Results:

• A support vector regressor (SVR) model achieved the best performance.
• The optimal SVR model used six features and 1251 training examples.
• The best model demonstrated a low mean absolute error (MAE) of <0.17 for log(K_D) predictions.

Conclusions:

• Machine learning, specifically SVR, can accurately predict Fc variants with altered FcRn binding affinity.
• This computational approach offers a novel strategy for discovering antibody variants with enhanced pharmacokinetic properties.
• The findings suggest a powerful tool for accelerating the development of next-generation therapeutic antibodies.