Fig. 1: The workflow of penetratin optimization.

a First, we pretrained our model with UniProt data. Secondly, we constructed Amps Training Pairs and Penetratin Training Pairs datasets (cf. Methods) by a degenerative workflow, and feed these to finetune the pretrained model. The deepAMP-POM was used to optimize penetratin model, and the b deepAMP-predict were used to rank these candidates. First round18 sequences were selected from ranked candidates for c in vitro and in vivo experimental validation.