Comparison of methods for generating sequence embeddings.

The paratope-weighted embedding is computed as a weighted average of amino acid embeddings, with weights determined by their predicted probabilities of belonging to a paratope, while the averaged embedding is a uniform mean across all amino acid embeddings. Performance is assessed using the R2 score of a linear model predicting binding affinity or expression using the sequence embedding as input, across different protein language models (PLMs) and datasets. A paired Wilcoxon test across the six PLMs and three datasets (18 points) shows that paratope weighting does not provide a statistically significant improvement for affinity prediction (P = 0.393). (XLSX)