Sequence-Only Prediction of Antibody Fab Thermostability Using Protein Language Model Embeddings

Antibody developability is a critical consideration in therapeutic manufacturing, with conformational stability--measured by melting temperature (Tm)--being a key determinant of manufacturability. Traditional experimental approaches to assess Tm are costly and low-throughput, while statistical models offer limited predictive power and generalizability. In this study, we present a scalable, sequence-only machine learning (ML) approach to predict the thermostability of antigen-binding fragments (Fabs) using embeddings from IgBERT, a protein language model fine-tuned on antibody sequences. We trained a Random Forest regressor on 133 proprietary antibody sequences with measured Fab Tm values, achieving a Pearson correlation coefficient (PCC) of 0.77 on a held-out internal test set. When applied to a Jain set of clinical-stage therapeutics, the model retained predictive power (PCC = 0.28), outperforming baseline statistical and sequence-similarity methods. Embedding proximity in t-SNE space correlated with prediction accuracy, suggesting that model generalizability is influenced by distributional similarity rather than sequence identity. Our results demonstrate that PLM-derived embeddings encode biophysical features relevant to Fab thermostability and offer a fast, interpretable, and scalable route to de-risk antibody lead candidates.