Fig. 2From: Pre-trained protein language model sheds new light on the prediction of Arabidopsis protein–protein interactionsAUPR values of combinations between nine pLMs from ESM and three machine learning algorithms. Of the different ESM models, ESM-1v was fine-tuned for predicting variant effects and contained five models with different random seeds. ESM-1b differs from ESM-1 mainly in higher learning rate, dropout after word embedding, learned positional embeddings, final layer norm before the output, and tied input/output word embedding. A The results from the independent dataset C2 where only one protein in each pair appeared in the training dataset (i.e., C1), while B corresponds to the results from the independent dataset C3 where no protein in each pair appeared in the training dataset (i.e., C1)Back to article page