AbMelt: Learning antibody thermostability from molecular dynamics.

Area of Science:

• Biophysics
• Computational Biology
• Protein Engineering

Background:

• Antibody thermostability is critical for drug development but difficult to predict from sequence or structure alone.
• Current prediction methods often lack direct entropic information, limiting accuracy.

Purpose of the Study:

• To develop a novel computational method, AbMelt, for predicting antibody thermostability.
• To identify key molecular descriptors related to antibody flexibility that correlate with thermal stability endpoints.

Main Methods:

• Utilized molecular dynamics simulations at three temperatures to model the flexibility of homologous antibody structures.
• Learned relevant descriptors from simulations to predict temperatures of aggregation (T_agg), melt onset (T_m,on), and melt (T_m).
• Employed machine learning regression for robust prediction on a held-out test set.

Main Results:

• The deviation in radius of gyration for complementarity determining regions at 400 K was the strongest descriptor for T_agg (r_p = -0.68).
• Deviation of internal molecular contacts at 350 K strongly correlated with both T_m,on (r_p = -0.74) and T_m (r_p = -0.69).
• AbMelt achieved robust performance (R² values of 0.57-0.60) and outperformed traditional models, even when trained on <5% of the data.

Conclusions:

• AbMelt successfully models antibody flexibility using molecular dynamics to predict key thermostability parameters.
• The method provides accurate and robust predictions, offering a valuable tool for antibody engineering and development.
• AbMelt is available for predicting thermostability measurements of antibody variable fragments.