Fig. 1

The schematic diagram of ESMAraPPI. Arabidopsis PPIs from the IntAct database with MIscore ≥ 0.45 were collected as positive samples. We also compiled 10 times negative samples to construct an original dataset. Then, we divided the original dataset into three datasets (i.e., C1, C2 and C3). C1 was the training dataset, while C2 and C3 were two independent test datasets. The representations of protein pair were extracted from ESM-1b, and Hadamard product was applied before inputting to a 4-layer MLP. The final output was an interaction score between 0 and 1 (a prediction score ≥ 0.5 corresponded to a positive interaction)