GBactPro: cross-species bacterial promoter prediction using element-resolved promoter definitions
Journal:
bioRxiv
Published Date:
Oct 23, 2026
Abstract
Accurate bacterial promoter prediction remains challenging across species, partly because conventional models often use fixed-length promoter windows that do not explicitly account for promoter architecture. Using promoter annotations derived from the Promoter Architecture Scanner (PAS), we found that training on sequences comprising the -35 element, spacer, and -10 element improved cross-species prediction compared with fixed-length promoter windows. Based on this element-resolved promoter definition, we developed GBactPro, a convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) model jointly trained on curated promoter data from 43 bacterial species. The jointly trained model achieved a mean AUROC of 0.9943 on held-out test sequences. Under matched-length benchmarking, GBactPro outperformed Promotech, iPro-MP, Prompt, and iPro-WAEL by 0.066-0.314 in mean AUROC. Genome-wide scanning of Escherichia coli achieved an AUROC of 0.979 and an AUPRC of 0.160 under stringent one-to-one promoter-site matching, increasing to 0.997 and 0.323, respectively, with a 100-bp positional tolerance. Unmatched predictions were enriched for H-NS binding, and independent H-NS-knockout data further supported the possibility that some predicted promoter-like regions are transcriptionally repressed. Together, these findings highlight the importance of promoter-sequence definition for cross-species prediction and demonstrate the potential of GBactPro for genome-wide promoter identification and biological interpretation.
