Not all representations are equal: Comparing protein language models for antibody thermostability prediction

Predicting antibody thermostability is an important and challenging task in computational antibody design. Antibodies which are not thermostable may be incompatible with mass production and distribution. To this end, we assess how different protein language model (pLM) representations affect performance in the downstream task of predicting antibody thermostability. Our findings demonstrate that the choice of pLM has a large effect on predictor performance, even when data, model size, and hyperparameters are held stable. We also show that a performance boost may be obtained by combining pLM representations.