iPro-MP: a BERT-based model to predict multiple prokaryotic promoters

Abstract

Background

Promoters, as essential cis-regulatory elements in prokaryotes, govern gene expression by mediating RNA polymerase binding through core motifs and long-range regulatory interactions, playing a pivotal role in cell metabolism and environmental adaptation. Hence, accurate identification of prokaryotic promoters is vital for understanding their biological functions. However, the existing tools for predicting prokaryotic promoters are mainly concentrated on individual model organisms, and their prediction accuracy needs to be further improved. To address these gaps, we develop iPro-MP, a transformer-based prokaryotic promoter prediction framework that we systematically evaluate across 23 phylogenetically diverse species, including both model and non-model organisms.

Results

iPro-MP utilizes a multi-head attention mechanism to capture textual information in DNA sequences and effectively learns the hidden patterns. Cross-species prediction demonstrates the necessity of constructing species-specific models. Through a series of experiments, iPro-MP shows outstanding performance, with the AUC exceeding 0.9 in 18 out of 23 species.

Conclusions

Our novel approach to predicting prokaryotic promoters, iPro-MP, provides the superiority to other existing tools, especially in predicting non-model organisms. Finally, for the convenience of other researchers, the source code and datasets of iPro-MP are freely available at https://github.com/Jackie-Suv/iPro-MP.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13059-025-03819-9.

Background

Gene transcription is a fundamental biological process that is tightly regulated by promoter regions [1–3]. Promoters initiate transcription by guiding RNA polymerase binding and play a central role in the regulation of gene expression [4–6]. In prokaryotes, promoters serve as key cis-regulatory elements that define transcription start sites (TSSs), thereby modulating gene expression levels [7–9]. Accurate identification of promoters is therefore essential for understanding gene regulatory mechanisms and cellular responses to environmental changes. However, despite their critical biological functions, promoter recognition in genomic sequences remains a challenging task, particularly in non-model organisms.

In recent years, the rapid development of next-generation sequencing (NGS) technologies has significantly advanced multi-omics research, such as genomics and transcriptomics [10]. Among these, differential RNA sequencing (dRNA-seq) [11] enables high-resolution, genome-wide mapping of TSSs by selectively sequencing primary transcripts with a 5′ triphosphate end [12]. This technique has been successfully applied to several prokaryotic species, including Helicobacter pylori (H. pylori) [13], Methylorubrum [14], Haloferax volcanii (H. volcanii) [15], and Streptomyces tsukubaensis (S. tsukubaensis) [16], leading to the construction of several prokaryotic promoter databases, such as RegulonDB [17], DBTBS [18], Pro54DB [19], and PPD [20].

However, with the exponential growth of sequencing data and the increasing cost and labor demands of experimental validation, traditional experimental methods are no longer sufficient to meet the large-scale needs of promoter identification. This has driven the development of efficient computational approaches for prokaryotic promoter prediction. Over the past decade, a variety of machine learning models have been proposed based on DNA sequence features [5, 21–25]. For instance, Lai et al. [26] proposed iProEP, a support vector machine (SVM)-based model [27, 28] that integrates pseudo k-tuple nucleotide composition with a position-correlation scoring function, achieving accuracies of 95.2% and 93.1% for Escherichia coli (E. coli) and Bacillus subtilis (B. subtilis), respectively. Zhang et al. [29] introduced MULTiPly, a two-layer prediction predictor capable of identifying E. coli promoters and their subtypes, with an accuracy of 86.9%. Subsequently, PromoterLCNN, a convolutional neural network (CNN)-based model, further improved the prediction accuracy to 88.6% [30]. In 2022, iPro-WAEL [31], a weighted average ensemble learning model, was developed to support promoter prediction in multiple prokaryotic species. Most recently, Du et al. [32] developed Prompt, a voting-based strategy to predict promoters of 16 prokaryotes. Although these computational models have yielded promising results, they are mainly limited to a few well-studied model organisms such as E. coli and B. subtilis, with limited applicability to a broader range of prokaryotic species. Hence, it is urgent to develop a powerful predictor capable of recognizing the promoters of multiple prokaryotic species.

To address this limitation, we developed iPro-MP, a DNABERT-based deep learning model tailored for promoter prediction across 23 prokaryotes. By leveraging a bidirectional self-attention mechanism, iPro-MP effectively capture both local sequence motifs and global contextual relationships in DNA sequences. Through a series of comparative experiments and evolutionary analyses, iPro-MP exhibited competitive performance compared to other existing tools for predicting prokaryotic promoters. To facilitate further research and application, we have made the source code and datasets publicly available at https://github.com/Jackie-Suv/iPro-MP.

Results

iPro-MP exhibits excellent performance and robustness in multi-species promoter prediction

To evaluate the predictive capability and generalizability of iPro-MP, we conducted comprehensive experiments across 23 representative prokaryotic species. First, a comparative analysis of different k-mer sizes demonstrated that the 6-mer representation yielded the best performance (Additional file 1: Fig. S1 and Additional file 2: Table S1). This indicates that a larger k-mer provides the model with richer sequence semantics, thereby enhancing its ability to effectively capture promoter-specific features. Second, several cross-validation strategies were evaluated, including 5-fold, 10-fold, and repeated fivefold cross-validation (Additional file 2: Fig. S2 and Additional file 2: Table S2). The results showed minimal performance differences among these methods, whereas 5-fold cross-validation exhibited higher computational efficiency. Based on these findings, the 6-mer encoding scheme and 5-fold cross-validation were adopted as the final configuration for model training and evaluation.

As shown in Fig. 1a–d and Additional file 2: Table S3, iPro-MP achieved consistently high scores across four evaluation metrics using 5-fold cross-validation—Acc, AUC, area under the precision-recall curve (AUPRC), and MCC. Specifically, AUC values exceeded 0.9 in 17 species (73.9%), ranged from 0.8 to 0.9 in 5 species (21.7%), and fell below 0.8 in only one case (4.3%) (Fig. 1e). Similar trends were observed for AUPRC, indicating reliable model behavior even under data imbalance. While MCC values exhibited slightly greater variability, the majority still remained within a high-performance range, reflecting the model’s stability across species with different promoter sequence characteristics.

Fig. 1.

Comprehensive performance metrics evaluated using fivefold cross-validation: a Acc, b AUC, c AUPRC, d MCC, and e ROC curve. Numbers 1–23 correspond to the 23 prokaryotic species listed in Table 3

The robustness of iPro-MP was further validated on independent testing sets (Fig. 2 and Additional file 2: Table S4). Encouragingly, the model maintained high predictive performance, with 18 species (78.3%) achieving AUC > 0.9, four species (17.4%) between 0.8 and 0.9, and only one species (4.3%) below 0.8 (Fig. 2e). These results demonstrate that iPro-MP exhibits minimal performance degradation from training to independent testing, highlighting its strong generalization capability. Notably, the model performed well not only on common model organisms such as E. coli and B. subtilis, but also on phylogenetically distant or compositionally diverse species such as H. volcanii and S. tsukubaensis.

Fig. 2.

Comprehensive performance metrics evaluated using independent testing sets: a Acc, b AUC, c AUPRC, d MCC, e ROC curve, and f PR curve. Numbers 1–23 correspond to the 23 prokaryotic species listed in Table 3

It should be noted that despite the relative imbalance in the dataset, iPro-MP consistently achieved strong recall and F1 scores—averaging 0.824 and 0.816 under 5-fold cross-validation, and 0.834 and 0.832 on the independent test set, respectively. These results suggest that the model effectively identifies true promoter sequences while maintaining a favorable trade-off between precision and sensitivity, highlighting the model’s robustness in handling imbalanced datasets and its ability to generalize across diverse species.

The strong and consistent performance of iPro-MP across evolutionarily diverse species can be attributed to its underlying architecture, which is based on DNABERT. Unlike traditional machine learning approaches that rely on handcrafted features, DNABERT employs a multi-head self-attention mechanism capable of capturing both local and long-range dependencies within DNA sequences. This enables the model to learn complex regulatory signals and latent motif structures directly from the raw genomic sequence. Such contextual modeling allows iPro-MP to remain sensitive to global sequence patterns, enhancing its ability to recognize functionally relevant promoter elements—even in species that were not used during training. As a result, iPro-MP effectively generalizes promoter recognition beyond species boundaries, uncovering both conserved and species-specific regulatory elements with high fidelity.

iPro-MP reveals the species-specificity at the sequential level

To further evaluate the transferability of promoter recognition across species, we conducted a comprehensive cross-species prediction experiment. Specifically, for each of the 23 prokaryotic species, we trained a species-specific model using its own promoter data and tested its predictive performance on independent datasets from all other species. The resulting Acc, AUC, AUPRC, and MCC values are visualized in the form of heatmaps in Fig. 3a–d, respectively.

Fig. 3.

The heat map shows the values of a Acc, b AUC, c AUPRC, and d MCC in cross-species promoter validation. Numbers 1–23 correspond to the 23 prokaryotic species listed in Table 3

As expected, the highest performance for each model generally occurred when the training and testing species were the same (diagonal elements), confirming the reliability of the species-specific models. However, a number of off-diagonal elements also exhibited high prediction performance, suggesting that the promoter sequence features learned from certain species can generalize to others. Notably, models trained on Campylobacter jejuni strains (C. jejuni RM1221, C. jejuni subsp. jejuni 81,116, C. jejuni subsp. jejuni 81–176, and C. jejuni subsp. jejuni NCTC 11168) [33, 34] achieved high accuracy when predicting each other’s promoter sequences. This observation is consistent with their close phylogenetic relationships as shown in the evolutionary tree (Fig. 4a), and their shared promoter motif patterns (Fig. 4b) revealed by the Two Sample Logos [35]. As shown in the motif logos adjacent to the tree, these strains possess highly similar sequence elements in their − 10 and − 35 regions, indicating conserved cis-regulatory architecture.

Fig. 4.

The evolutionary relationships and the conserved sequences in the promoters. a Phylogenetic tree of 23 species combined with significant motifs. The nucleotide distribution and preference in the promoters of b 21 bacteria and c 2 archaea compared to non-promoters using Two Sample Logos tool

Similar cross-species transferability was observed between other phylogenetically related species, such as C. diphtheriae NCTC 13129 and C. glutamicum ATCC 13032, as well as between S. aureus subsp. aureus MW2 and S. epidermidis ATCC 12228, which also share highly similar promoter sequence signatures (Fig. 4b). These results indicate that phylogenetic proximity and sequence motif conservation play crucial roles in enabling effective promoter recognition across species boundaries.

In contrast, models trained on archaeal species such as H. volcanii DS2 and T. kodakarensis KOD1 performed poorly when applied to bacterial species, with accuracy values consistently below 0.7. This is likely due to fundamental differences in transcriptional regulatory mechanisms between archaea and bacteria [36–38]. As shown in Fig. 4c, archaeal promoters lack the clearly enriched − 10 and − 35 motifs typically found in bacterial promoters and instead exhibit distinct sequence preferences potentially related to the TATA-binding protein (TBP) and transcription factor B (TFB) binding sites [39]. In addition, we found that there was no difference in performance between GC-rich species and AT-rich species (Additional file 1: Fig. S3). These observations further emphasize the importance of constructing species-specific or clade-specific models for accurate promoter identification, particularly when dealing with evolutionarily distant organisms.

Generally, it can be clearly found that the results of most models in predicting other species promoters are not satisfactory. This result provides compelling evidence for the existence of strong species specificity among prokaryotic species. Given these circumstances, the development of species-specific prediction models becomes not just beneficial but essential. These models have the potential to enhance the accuracy of promoter prediction by leveraging species-specific data and knowledge.

iPro-MP learns discriminative representations that clearly separate promoters from non-promoters

To evaluate the internal feature representations learned by iPro-MP, we applied t-distributed stochastic neighbor embedding (t-SNE) [40] to project the high-dimensional embedding vectors of promoter and non-promoter sequences into a two-dimensional space. As shown in Fig. 5a, the visualized distributions reveal a clear separation between the two classes in most of the 23 species. Promoter sequences (yellow) consistently form tight and coherent clusters, whereas non-promoter sequences (purple) are more broadly scattered, indicating greater diversity in background genomic regions. This class-specific clustering suggests that iPro-MP effectively captures regulatory features that are shared among promoter sequences, such as conserved sequence motifs or positional patterns.

Fig. 5.

t-SNE visualization of a both promoters and non-promoters, and b only promoters in a two-dimensional feature space based on the token embeddings from the final hidden layer. c Distribution of attention weights for promoter sequences. Each row represents a species, and each column corresponds to the position of the first nucleotide of a 6-mer token relative to the TSS (position 0)

Furthermore, the separation observed across phylogenetically distant species demonstrates that the learned representations are not merely species-specific, but reflect generalizable biological properties of prokaryotic promoters. In some species, particularly those with well-defined − 10 and − 35 motifs, the promoter and non-promoter groups are almost completely separable. In contrast, species with weaker motif signals or atypical GC content exhibit more overlap between the classes, consistent with the known challenges in promoter annotation in these genomes. In addition, promoters from all species were embedded in the same space (Fig. 5b). Promoters from different species generally form distinct clusters, with several phylogenetically related species, e.g., C. jejuni strains, showing closely positioned or partially overlapping clusters, while others, such as C. glutamicum and S. elongatus, occupy well-separated regions. This result highlights the species-specific nature of promoter sequence patterns, yet also suggests that iPro-MP captures shared sequence features.

To further enhance the interpretability of our model, the attention weights learned by iPro-MP were extracted and visualized (Fig. 5c). Notably, in most species, the attention is strongly concentrated around position − 10, which aligns well with the known biological architecture of bacterial promoters. However, in the case of archaea (species 11 and 21), the attention weights are primarily concentrated around position –26, which corresponds to the location of the TATA-box-like core promoter element that is evolutionarily conserved among archaeal species [39]. In a few species, additional attention is observed near the TSS, reflecting subtle species-specific variations in promoter architecture. These patterns not only highlight the evolutionary conservation of core regulatory motifs but also underscore the model’s ability to capture subtle differences across species. In summary, the high-attention 6-mers concentrated upstream of the TSS validate that iPro-MP leverages functional promoter motifs, thereby offering both predictive power and biological insight. This feature-level visualization underscores the model’s interpretability and highlights its ability to generalize across diverse prokaryotic lineages. These results provide strong evidence that iPro-MP learns biologically meaningful and transferable representations, which contribute to its robust performance in promoter classification across diverse prokaryotic taxa.

iPro-MP outperforms classical and deep learning baselines across species

To comprehensively assess the predictive capacity of iPro-MP, we benchmarked it against four baseline classifiers—random forest (RF) [41], eXtreme Gradient Boosting (XGBoost) [42], logistic regression (LR) [43], and long short-term memory networks (LSTM) [44]—using promoter classification tasks in 23 diverse prokaryotic species. All models were trained using 5-fold cross-validation and optimized hyperparameters (Additional file 2: Table S5), and performance was evaluated using Acc, AUC, AUPRC, and MCC (Additional file 2: Table S6).

Across all four metrics, iPro-MP consistently outperformed the competing methods (Fig. 6a–d). Specifically, iPro-MP achieved the highest average values across species for all metrics (Table 1): Acc (mean = 0.890), AUC (mean = 0.935), AUPRC (mean = 0.903), and MCC (mean = 0.752). In contrast, the second-best performing model (RF) achieved Acc = 0.839, AUC = 0.882, AUPRC = 0.847, and MCC = 0.688, indicating a substantial performance gap. Next is XGBoost, whose performance is close to that of RF. However, methods like LR and LSTM lagged behind, especially in MCC and AUPRC, suggesting that they struggled to maintain class balance and precision under imbalanced data distributions.

Fig. 6.

Performance comparison of different classification algorithms in identifying promoters through fivefold cross-validation in terms of a Acc, b AUC, c AUPRC and d MCC. Numbers 1–23 correspond to the 23 prokaryotic species listed in Table 3

Table 1.

Comparison of the average prediction performance of iPro-MP and other methods using the independent testing sets

Methods	Acc	AUC	AUPRC	MCC
iPro-MP	0.890	0.935	0.903	0.752
RF	0.839	0.882	0.847	0.688
XGBoost	0.833	0.876	0.842	0.675
LR	0.803	0.843	0.790	0.612
LSTM	0.806	0.850	0.792	0.621

The performance gap is especially pronounced in challenging species. For example, in H. volcanii DS2—an archaeon with distinct transcriptional regulatory elements—iPro-MP reached an MCC of 0.876, while RF and LR only achieved 0.820 and 0.747, respectively. Similarly, in T. kodakarensis KOD1, iPro-MP achieved an MCC of 0.853, significantly outperforming RF (0.824) and XGBoost (0.793). Even in well-characterized species like E. coli, iPro-MP outperformed all alternatives in AUC (0.878 vs. 0.689–0.823) and AUPRC (0.807 vs. 0.596–0.737).

These results demonstrate that transformer-based models like iPro-MP are better equipped to capture the complex sequence patterns and regulatory logic in prokaryotic promoters. Unlike conventional methods that rely on fixed features or local dependencies, iPro-MP benefits from the global context modeling capability of DNABERT, enabling more accurate promoter recognition across a wide evolutionary range.

iPro-MP outperforms existing tools in multi-species promoter prediction

To validate the effectiveness and robustness of iPro-MP, we compared its performance with three state-of-the-art tools for prokaryotic promoter prediction: Prompt [32], PromoterLCNN [30], and iPro-WAEL [31]. These methods were selected based on the availability of webservers or open-source implementations, allowing for reproducible benchmarking on the same 23 independent testing sets used in this study.

As shown in Fig. 7 and Additional file 2: Table S7, iPro-MP consistently outperformed the competing methods across most species. Specifically, iPro-MP achieved the highest Acc (mean = 0.890), AUC (mean = 0.935), AUPRC (mean = 0.903), MCC (mean = 0.752), and F1 score (mean = 0.832) among all compared methods (Table 2). Moreover, when examined at the species level, iPro-MP achieves the best performance in 21 out of 23 species, further confirming its robustness and cross-species generalizability. Importantly, although iPro-MP utilizes a transformer-based architecture, it maintains competitive runtime efficiency (average 29.09 s) and memory usage (around 3.6 GB), which outperforms all other tools. To explore alternative validation methods in the benchmarking process, we repeated the training procedures of the other three tools using fivefold cross-validation, while strictly adhering to the official training and evaluation protocols provided by each tool (Additional file 1: Fig. S4 and Additional file 2: Table S8). iPro-MP exhibited the most robust and consistently top-performing results in all scenarios. These results collectively highlight the practical advantages of iPro-MP in terms of both predictive accuracy and computational efficiency.

Fig. 7.

Performance comparison of iPro-MP with existing tools based on the independent testing sets in terms of a Acc and b MCC. Numbers 1–23 correspond to the 23 prokaryotic species listed in Table 3

Table 2.

Comparison of the average prediction performance of iPro-MP and existing tools using the independent testing sets

Tools	Acc	AUC	AUPRC	MCC	F1	Runtime (s)
iPro-MP	0.890	0.935	0.903	0.752	0.832	29.09
Prompt	0.788	0.835	0.798	0.611	0.737	34.48
PromoterLCNN	0.755	0.793	0.767	0.565	0.706	35.32
iPro-WAEL	0.798	0.838	0.811	0.606	0.747	34.40

While there were a few exceptions where other methods performed slightly better—e.g., PromoterLCNN achieved higher accuracy in E. coli (0.832 vs. 0.819) and iPro-WAEL outperformed iPro-MP in P. riograndensis SBR5 (0.802 vs. 0.793)—iPro-MP still demonstrated significantly superior performance in MCC, indicating better overall classification reliability and robustness, particularly under class imbalance. These results highlight iPro-MP’s strong generalization capability and its advantage in learning universal promoter features across phylogenetically diverse prokaryotes, offering a substantial improvement over previous methods.

Discussion

In this study, we developed iPro-MP, the first DNABERT-based framework for promoter prediction across multiple prokaryotic species. This model addresses key limitations of existing methods, including narrow species coverage and suboptimal predictive performance, by leveraging a transformer architecture with multi-head self-attention to learn both local and long-range dependencies in DNA sequences.

Through rigorous benchmarking on 23 prokaryotic species using 5-fold cross-validation and independent testing sets, iPro-MP consistently outperformed existing tools in terms of Acc and MCC. Importantly, our cross-species prediction analysis highlighted both the potential and limitations of model generalization, underscoring the value of species-specific modeling in certain phylogenetic contexts. Moreover, t-SNE visualization and motif conservation analysis demonstrated that iPro-MP captures biologically meaningful regulatory features, even in non-model organisms.

The satisfactory performance of iPro-MP can be attributed to several key design choices. First, the use of DNABERT enables the model to capture complex contextual patterns in DNA sequences. This is particularly effective for modeling promoter regions, which often contain dispersed regulatory motifs and variable structural elements. Second, the training dataset spans 23 prokaryotic species with diverse genomic characteristics, improving the model’s ability to generalize across phylogenetic clades. Third, the integration of both CDS and intergenic sequences as negative samples forces the model to learn more refined boundaries between promoters and non-promoters, beyond simply identifying open reading frames. Finally, the lightweight classification head with GELU activation, dropout regularization, and layer normalization contributes to model stability and generalization.

Beyond promoter prediction, iPro-MP has potential applications in fields such as synthetic biology and functional genomics. Recent advances in deep learning have enabled the generation of synthetic promoters with high expression strength [45–47], yet current generative models are limited by the lack of training data for non-model organisms. By enabling large-scale and accurate identification of native promoters across diverse prokaryotic genomes, iPro-MP can help expand promoter datasets for underexplored species. This, in turn, may support the development of more generalizable generative frameworks for synthetic promoter design, particularly in industrially or medically relevant prokaryotes.

Although iPro-MP achieves strong performance overall, prediction accuracy remains suboptimal for a few species, particularly those with weak promoter motifs or atypical regulatory architectures, e.g., P. putida and S. meliloti. This highlights the need for further investigation into species-specific promoter patterns and possibly incorporating epigenetic or structural features. In addition, recent advances in bioinformatics have demonstrated the strong problem-solving potential of universal models across a variety of biological prediction tasks, from protein function annotation to cross-species regulatory element detection. Inspired by this trend, we aim to extend our work by developing a universal promoter prediction framework capable of generalizing across a broader range of prokaryotic species. Such a model would reduce the dependency on organism-specific fine-tuning and enable scalable application to newly sequenced or poorly annotated species.

Conclusion

Overall, iPro-MP represents a powerful and generalizable tool for prokaryotic promoter annotation. By enhancing our ability to decode regulatory regions across diverse microbial genomes, this framework holds promise for advancing research in microbial gene regulation, synthetic biology, microbial ecology, and the study of pathogenic mechanisms. Future efforts may further improve its utility through integration with multi-omics data and domain-specific fine-tuning strategies.

Methods

The main workflow of iPro-MP is illustrated in Fig. 8, including three modules:

1. Data collection and processing module: experimentally validated promoter sequences were curated from the Prokaryotic Promoter Database (PPD). Redundant sequences were removed using CD-HIT with a similarity threshold of 0.8. Negative samples were generated from long coding sequences (CDS) and convergent intergenic regions using a sliding window strategy. Finally, the benchmark dataset was divided into a training set and an independent testing set.

2. Classification module: all sequences in the training set were tokenized into overlapping 6-mers and subsequently used to fine-tune the DNABERT model. The final prediction was generated through two fully connected layers.

3. Evaluation module: 5-fold cross-validation and independent testing were applied to evaluate performance, as well as feature analysis, cross-species prediction, phylogenetic analysis, and comparative test.

Fig. 8.

The flowchart of iPro-MP. a Promoter data collection and processing module, where all data are de-redundant and divided into training sets and testing sets. b Classification module, which captures contextual information hidden in the DNA sequence based on the DNABERT to make prediction. c Evaluation module, which assesses the performance of iPro-MP through a series of comparisons

Data collection and preprocessing

In our previous work, we constructed the prokaryotic promoter database (PPD) [20], which includes experimentally validated promoter data from 63 prokaryotic species. The PPD provides a robust foundation for this study. To ensure sufficient training data and minimize biases associated with small sample sizes, positive samples were selected according to the following rules: (1) Species with more than 1,000 experimentally validated promoters were included from the PPD. Notably, although B. subtilis has fewer than 1000 promoters, it is widely used in promoter prediction research and is therefore included in the benchmark dataset; (2) For each species, promoter sequences were filtered using the CD-HIT tool [48] with a sequence identity threshold of 0.8 to remove redundancy. As a result, a total of 107,286 promoter sequences (81 bp in length, [−60:20] where 0 represents TSS) were collected from 23 prokaryotic organisms using the uniform preprocessing and labeling strategy.

To generate a reliable negative sample dataset, we downloaded the whole genome sequences and corresponding annotation files of the 23 prokaryotic organisms from the GenBank database. Negative samples were derived from two genomic regions unlikely to contain promoters: (1) coding sequence (CDS) regions longer than 2000 bp, and (2) convergent intergenic regions longer than 81 bp. From these regions, we generated 81 bp subsequences using a sliding window method with a step size of 1 [49]. Subsequently, redundant sequences were removed using the CD-HIT program with a 0.8 threshold (Additional file 2: Table S9). Finally, the same number of sequences as positive samples was randomly selected from each of the two types of negative samples. This resulted in a final dataset with a 1:2 ratio of positive to negative samples (Table 3). After all the above steps, the benchmark dataset was randomly divided into training and independent testing sets in a ratio of 4:1 (Fig. 8a).

Table 3.

The benchmark dataset about 23 prokaryotic promoters

Species	GenBank accession number	CD-HIT (positive)		Training			Testing
		Before	After	Positive	CDS	Intergenic	Positive	CDS	Intergenic
1. Acinetobacter baumannii ATCC 17978	CP000521.1	1540	1111	889	889	889	222	222	222
2. Bradyrhizobium japonicum USDA 110	NC_004463.1	15,933	14,779	11,823	11,823	11,823	2956	2956	2956
3. Burkholderia cenocepacia J2315	AM747720.1 AM747721.1 AM747722.1	10,831	9663	7730	7730	7730	1933	1933	1933
4. Campylobacter jejuni RM1221	NC_003912.7	2166	1995	1596	1596	1596	399	399	399
5. Campylobacter jejuni subsp. jejuni 81,116	NC_009839.1	1942	1790	1432	1432	1432	358	358	358
6. Campylobacter jejuni subsp. jejuni 81–176	NC_008787.1	2129	1944	1555	1555	1555	389	389	389
7. Campylobacter jejuni subsp. jejuni NCTC 11168	NC_002163.1	1905	1745	1396	1396	1396	349	349	349
8. Corynebacterium diphtheriae NCTC 13129	NC_002935.2	1656	1518	1214	1214	1214	304	304	304
9. Corynebacterium glutamicum ATCC 13032	BX927147.1	3581	3031	2425	2425	2425	606	606	606
10. Escherichia coli str K-12substr. MG1655	NC_000913.3	8616	6230	4984	4984	4984	1246	1246	1246
11. Haloferax volcanii DS2	NC_013967.1	4749	4394	3515	3515	3515	879	879	879
12. Helicobacter pylori strain 26,695	NC_000915.1	2233	2062	1650	1650	1650	412	412	412
13. Nostoc sp. PCC7120	NC_003272.1	13,705	12,560	10,048	10,048	10,048	2512	2512	2512
14. Paenibacillus riograndensis SBR5	NZ_LN831776.1	2351	2276	1821	1821	1821	455	455	455
15. Pseudomonas putida KT2440	NC_002947.3	7938	6881	5505	5505	5505	1376	1376	1376
16. Shigella flexneri 5a str. M90T	NZ_CP037923.1	14,051	9760	7808	7808	7808	1952	1952	1952
17. Sinorhizobium meliloti 1021	NC_003047.1	17,003	14,482	11,586	11,586	11,586	2896	2896	2896
18. Staphylococcus aureus subsp. aureus MW2	NC_003923.1	2821	1969	1575	1575	1575	394	394	394
19. Staphylococcus epidermidis ATCC 12228	NC_004461.1	2207	1647	1318	1318	1318	329	329	329
20. Synechococcus elongatus PCC 7942	CP000100.1	1473	1410	1128	1128	1,128	282	282	282
21. Thermococcus kodakarensis KOD1	NC_006624.1	2720	2413	1930	1930	1,930	483	483	483
22. Xanthomonas campestris pv. campestris B100	AM920689.1	3067	2963	2370	2370	2,370	593	593	593
23. Bacillus subtilis subsp. subtilis str. 168	NC_000964.3	691	663	530	530	530	133	133	133

DNABERT model

Bidirectional Encoder Representations from Transformers (BERT) is a pre-trained language representation model that retains the encoder part of Transformers [50], including the multi-head attention mechanism and the position-wise fully connected feed-forward neural network, while also expanding the function of bidirectional learning. It employs the Masked Language Model (MLM) as a fundamental component of its pre-training strategy, enabling the model to develop a deep bidirectional understanding of language. The relevant formulas are as follows:

$$\begin{array}{c}MultiHead\left(Q,,,K,,,V\right)=Concat\left({\mathit{head}}_1,,,{\mathit{head}}_2,,,\cdots ,,,{\mathit{head}}_h\right)W^O\end{array}$$ 1

where

$$\begin{array}{c}{\mathit{head}}_i=Attention(Q{W}_i^Q,K{W}_i^K,V{W}_i^V)\end{array}$$ 2

$$\begin{array}{c}Attention\left(Q,,,K,,,V\right)=softmax\left(\frac{Q{K}^T}{\sqrt{d_k}}\right)V\end{array}$$ 3

In the above formulas, $Q$, $K$, and $V$ represent query vector, key vector, and value vector, respectively. ${\mathit{head}}_i$ refers to the attention layer, and $d_k$ is the dimension of the vector.

DNABERT leverages BERT's bidirectional attention mechanism to learn both local and global contextual relationships within DNA sequences [51]. By pre-training on large-scale DNA datasets, DNABERT is capable of capturing complex patterns and dependencies in the genome, thereby providing robust feature representations for downstream tasks such as gene prediction [52, 53], motif discovery [54, 55], and regulatory element identification [56, 57]. DNABERT employs a sophisticated tokenization strategy that transforms DNA sequences into k-mer tokens, thereby enabling the application of transformer-based architectures to genomic data (Fig. 8b). A k-mer represents a contiguous subsequence of length k nucleotides extracted from the DNA sequence, analogous to word tokens in natural language processing (NLP). In this study, we fine-tuned the DNABERT model using the benchmark dataset of 23 species and applied the fine-tuned model for cross-species validation.

Fine-tuning of DNABERT model

In this study, a fine-tuning approach was applied to the pre-trained DNABERT model for the classification of DNA sequences with high accuracy. To encode input sequences for DNABERT, a k-mer tokenization strategy with k values of 3, 4, 5, and 6 was employed, enabling the model to capture sequence patterns at multiple resolutions.

To ensure reproducibility, all random number generators were initialized with a fixed seed, and PyTorch’s CuDNN backend was configured to accelerate the training and inference process. The recognition model integrated the pre-trained model with additional fully connected layers, layer normalization, dropout, and GELU activation functions to enhance model capacity and prevent overfitting. More specifically, during fine-tuning, we used the Adam optimizer with a learning rate of 1e − 5, a dropout rate of 0.3, and trained the model for up to 100 epochs. The batch size was set to 64. An early stopping strategy was applied with a patience of five epochs and a minimum delta of 1e − 4 in validation AUC.

Performance evaluation

To compare the quality of computational predictors, a set of metrics (Fig. 8c), including, sensitivity (Sn), specificity (Sp), overall accuracy (Acc), and Matthews correlation coefficient (MCC), was used to evaluate prediction performance [58–63].

$$\begin{array}{c}\left\{\begin{array}{c}Sn=\frac{\mathit{TP}}{TP+FN}0\le Sn\le 1\\ Sp=\frac{\mathit{TN}}{TN+FP}0\le Sp\le 1\\ Acc=\frac{TP+TN}{TP+TN+FP+FN}0\le Acc\le 1\\ MCC=\frac{TP\times TN-FP\times FN}{\sqrt{\left(T,P,+,F,N\right)\times \left(T,N,+,F,N\right)\times \left(T,P,+,F,P\right)\times \left(T,N,+,F,P\right)}}0\le MCC\le 1\end{array}\right)\end{array}$$ 4

where TP, FN, TN, and FP represent the number of true positives, false positives, true negatives, and false negatives, respectively.

Moreover, an alternative intuitive approach for model comparison is through the analysis of receiver operating characteristic (ROC) curves, which visually demonstrate relative performance across different decision thresholds. The area under the ROC curve (AUC) serves as a critical performance metric for evaluating binary classification models [64, 65], quantifying their inherent ability to discriminate between positive and negative classes. The higher the AUC, the better the performance. Besides, to save computing time, we used 5-fold cross-validation to train the model and the independent dataset to evaluate our model and other tools.

Authors’ contributions

Hao Lin and Hao Lyu conceived and designed the experiments. Wei Su, Yuhe Yang, Yafei Zhao, Shishi Yuan, Xueqin Xie, Yuduo Hao, Hongqi Zhang and Dongxin Ye performed the analysis and wrote the paper. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (62172078, 62402089) and China Postdoctoral Science Foundation (2023TQ0047, GZC20230380).

Data availability

The promoter sequences for 23 species were obtained from the Prokaryotic Promoter Database (PPD) (https://lin-group.cn/database/ppd/) [20]. The genome files and annotation files for 23 species were obtained from GenBank under the accession number in Table 3. The benchmark datasets and source codes for the iPro-MP framework are freely available at the GitHub repository (https://github.com/Jackie-Suv/iPro-MP) [66] under MIT license, which includes all necessary materials for local deployment, including species-specific training data, and detailed instructions for both prediction and model retraining. Meanwhile, we also provided the pretrained models at Zenodo (10.5281/zenodo.15180139) [67].

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.