Toward generalizable prediction of antibody thermostability using machine learning on sequence and structure features

Area of Science:

• Biotechnology and Pharmaceutical Sciences
• Computational Biology and Bioinformatics
• Protein Engineering and Design

Background:

• Monoclonal antibodies (mAbs) are increasingly important therapeutics, with over 100 FDA-approved.
• Multispecific biologics (msAbs) offer advantages by targeting multiple molecules.
• Single-chain variable fragments (scFvs), key components of msAbs, often suffer from poor thermostability, hindering therapeutic development.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting thermostable single-chain variable fragment (scFv) variants from sequence data.
• To compare the performance of pre-trained language models (PTLMs) and convolutional neural networks (CNNs) for thermostability prediction.
• To assess the generalizability of trained models on independent datasets and identify beneficial mutations.

Main Methods:

• Development of two machine learning approaches: a PTLM capturing sequence variation effects and a CNN trained on Rosetta energetic features.
• Training models using temperature-specific data (TS50 measurements) from multiple scFv libraries.
• Evaluation of model performance on out-of-distribution sequences and an independent monoclonal antibody dataset.

Main Results:

• A simple CNN model outperformed general PTLMs on out-of-distribution sequences (Spearman correlation of 0.4 vs. 0.15).
• An antibody-specific language model achieved superior performance compared to the CNN model (Spearman correlation of 0.52).
• Models trained on TS50 data successfully identified 18 residue positions and 5 mutations associated with thermostability in an independent mAb, demonstrating generalizability.

Conclusions:

• Computational models, particularly antibody-specific language models, can effectively predict and enhance the thermostability of scFvs.
• These predictive models offer a faster and more cost-effective alternative to experimental methods for antibody engineering.
• The developed approach shows broad applicability for improving antibody characteristics and optimizing large-scale production and delivery of biologics.