GBactPro: cross-species bacterial promoter prediction using element-resolved promoter definitions

Accurate bacterial promoter prediction supports genome annotation, but many models use fixed-length promoter windows anchored to transcription start sites (TSSs) rather than promoter-element boundaries. We developed the General Bacterial Promoter Predictor (GBactPro), a convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) framework using promoter sequences resolved by the Promoter Architecture Scanner. Using Escherichia coli-trained models evaluated across 43 bacterial species, we compared eight input designs and identified the contiguous -35/spacer/-10 core as optimal (mean AUROC = 0.9859). Species-specific models showed broad transfer in a 43 x 43 evaluation. Joint training on all 43 species achieved a mean AUROC of 0.9943 across their held-out test sets and outperformed four established predictors under matched-length benchmarks. In genome-wide E. coli scanning, GBactPro achieved AUROC and AUPRC values of 0.979 and 0.160 under stringent full-containment matching (0-bp tolerance), increasing to 0.997 and 0.323 under a 100-bp tolerance. False-positive predictions overlapped H-NS binding regions more frequently than true positives (24.4% versus 18.8%), and predictions also overlapped promoters derepressed in an H-NS-knockout dataset. Prediction precision varied across four TSS classes. These findings demonstrate that element-resolved promoter definitions improve cross-species prediction and support genome-wide identification of candidate bacterial promoters.