SignalP 6.0 predicts all five types of signal peptides using protein language models

Abstract

Signal peptides (SPs) are short amino acid sequences that control protein secretion and translocation in all living organisms. SPs can be predicted from sequence data, but existing algorithms are unable to detect all known types of SPs. We introduce SignalP 6.0, a machine learning model that detects all five SP types and is applicable to metagenomic data.

Fig. 1. Modeling SP structure using protein LMs.

a, Region structures of the five SP types. Twin arginine (RR)-translocated SPs feature a twin-arginine motif, while SPs cleaved by SPase II feature a C-terminal lipobox. Sec/SPIII SPs have no substructure. b, Protein LM training procedure. BERT learns protein features by predicting masked amino acids in sequences from UniRef100. c, t-Distributed stochastic neighbor embedding (t-SNE) projection of protein representations before prediction training. Different SP types form distinct clusters, separated from sequences without SPs. d, SignalP 6.0 architecture. An amino acid sequence is passed through the LM, and the resulting representation serves as input for the CRF, which predicts region probabilities at each position and the SP type. CS, cleavage site.

Fig. 2. SignalP 6.0 shows strong performance on all types and organism groups.

a, SP detection performance (ARC, Archaea; EUK, Eukarya; NEG, Gram-negative bacteria; POS, Gram-positive bacteria). SignalP 6.0 substantially improves performance on underrepresented types. b, CS prediction performance. SignalP 6.0 has improved precision for all categories. c, Dependence of performance on identity to sequences in the training data. At sequence identities lower than 60%, SignalP 6.0 outperforms SignalP 5.0.