DB-IRES: a deep learning model based on ensemble learning for predicting internal ribosome entry sites

Abstract

Background

Internal ribosome entry sites (IRES) are cap-independent translation initiation elements present in specific viral and cellular mRNAs. They facilitate direct ribosome recruitment for protein synthesis, bypassing the need for the canonical 5’ cap structure. Due to their pivotal roles in viral pathogenesis and cellular translational regulation, precise identification of IRES is crucial for advancing mechanistic studies and exploring potential therapeutic interventions. Nonetheless, manual identification is both labor-intensive and costly, while current computational methods exhibit limitations in accuracy and robustness.

Results

DB-IRES is a novel deep learning model designed for this purpose, incorporating densely connected 1D convolutional neural network blocks, bi-directional gated recurrent units, and a self-attention mechanism within an ensemble learning framework. Trained with a five-fold cross-validation strategy, the final integrated model exhibits superior and more robust predictive performance compared to existing methods on an independent test set. The model consistently achieves enhanced discriminatory capabilities across multiple evaluation metrics, thereby validating the effectiveness of its hybrid architecture and ensemble design.

Conclusions

DB-IRES serves as a reliable and precise computational tool for predicting IRES elements. Its improved performance enables more in-depth functional investigations of IRES biology and supports wider applications in RNA research and the development of associated therapeutics.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12859-026-06551-8.

Background

The classical model of gene expression translation initiation is predicated on the presence of the cap structure at the 5′ end of mRNA [1–3]. However, when cap-dependent translation initiation mechanisms are inhibited, translation can proceed in a cap-independent manner through the utilization of internal ribosome entry sites (IRES). The original IRES was identified by Pelletier and Sonenberg in Poliovirus [4]. Subsequent research revealed that IRES elements are also present in various viruses, including encephalomyocarditis virus, foot-and-mouth disease virus, and human immunodeficiency virus [5–7]. IRES elements have also been identified in cellular systems, where they occur in eukaryotic mRNAs [8, 9], such as XIAP [10], TP53 [11], and APAF1 [12]. IRES serves as a crucial component in viral genomes and represents a potential therapeutic target for the treatment of viral infections and associated diseases [13–15]. Consequently, the identification of IRES elements is of paramount importance for elucidating the mechanisms underlying IRES function and for the development of treatments for related diseases.

Bicistronic reporter vectors serve as the primary method for assessing internal ribosome entry sites (IRES) [16]. This technique involves the construction of two mRNA vectors: one that utilizes the 5′ cap structure and another that relies solely on IRES [17, 18]. Following their construction, these vectors are introduced into cells for activity assays, which are subsequently compared. However, this method is not cost-effective, demands high experimental precision, and often yields false positive result [19] due to the presence of concealed promoters and splice sites. Consequently, additional control experiments are typically necessary to mitigate these issues [20]. Therefore, a computational method has been developed to efficiently and accurately evaluate the significance of IRES.

A range of machine learning and deep learning tools is currently available for predicting internal ribosome entry sites (IRES). IRESFinder [21] employs k-mer features [22–24] to differentiate IRES using logistic regression models; however, it is limited to predicting human IRES. IRSS [25] utilizes RNALFold [26] to derive secondary structures, which are subsequently aligned with known IRES structures to assess IRES scores, but this tool is restricted to viral genome sequences and is no longer operational. IRESPred [27] applies multiple support vector machines (SVM) [28, 29] to predict the IRES of both viruses and cells, relying on 35 features that necessitate complex calculations and require information from the predicted secondary structure. IRESpy [30] employs the XGBoost [31, 32] model, which is trained using a combination of K-mer, QMFE [33, 34], and triple features [35, 36]. Our findings indicate that its performance in terms of generalization ability across various test sets is suboptimal.

DeepIRES [37] is the pioneering deep learning model for mRNA IRES prediction. It utilizes RNA sequences of 174 nucleotides as input, which undergo processing by hybrid neural networks to generate prediction scores. However, DeepIRES exhibits limitations in accurately predicting negative data classes.

Beyond the specific domain of IRES prediction, the broader landscape of biomedical machine learning is rapidly evolving towards more interpretable and robust deep learning frameworks for omics classification. Recent studies have successfully integrated advanced architectures with interpretability; for instance, IMPACT provides interpretable microbial phenotype analysis by leveraging characteristic traits [38]. In the realm of deep learning architectures, frameworks like scClassify2 have demonstrated the power of message passing mechanisms for precise cell state identification [39]. Furthermore, the importance of comprehensive representation learning in spatially resolved omics has been highlighted in recent multi-view characterizations [40]. These developments underscore the necessity for models that not only achieve high predictive accuracy but also offer insights into underlying biological mechanisms. Aligning with these broader trends, DB-IRES incorporates a hybrid deep learning architecture with a self-attention mechanism to enhance both prediction robustness and feature interpretability for RNA sequence analysis.

Therefore, we propose an enhanced DB-IRES model, an advanced framework aimed at improving prediction reliability. This model utilizes a sophisticated architecture to capture complex hierarchical features and long-range dependencies within the data, effectively identifying the most salient information for prediction. Its primary strength resides in aggregating the outputs of multiple independently trained learners, which together yield a more stable and accurate final prediction compared to traditional single-model approaches.

Methods

Data preparation

We utilized the data from Zhao, Chen, and colleagues in DeepIRES, which was derived from a high-throughput bicistronic assay [41]. They subsequently filtered this dataset using the CD-HIT-EST program [42–44], which has been a useful method for filtering data [45], resulting in a total of 18,527 sequences after screening. Within this dataset, samples exhibiting IRES activity above 600 were classified as positive. The final dataset comprised 1,917 positive samples and 16,610 negative samples, yielding an approximate IRES to non-IRES ratio of 1:9. Following this ratio, the entire dataset was partitioned into training and test sets while maintaining the original positive and negative class weights. The training set was further divided into a balanced and an imbalanced test set, resulting in 1,728 positive samples and 14,964 negative samples in the training data, 189 positive samples and 1,664 negative samples in the imbalanced test set, and 189 positive samples and 189 negative samples in the balanced test set.

Model architecture

We have developed a deep learning model that integrates DenseNet [46–50] (for text convolution specially), bidirectional GRU [51–53], and a self-attention module [54–56]. The model accepts an L × 4 matrix, which is generated through One-Hot encoding [57–60] of mRNA sequences. This matrix is then processed by our special Dense Block, each of which includes a transition layer to reduce the number of channels, with the exception of the final Dense Block. After feature extraction within the Dense Block, the resulting features are fed into a bidirectional GRU to capture long-distance dependencies. Subsequently, we apply the self-attention module for final processing. Upon completion of this stage, a fully connected layer is employed to produce probability scores. The detailed model structure is illustrated in Fig. 1.

Fig. 1.

Embed sequences as One-Hot encoding (A), followed by the application of a DenseNet for text convolution on the One-Hot features (B). The output from the DenseNet is then processed using a BiGRU and Self-Attention mechanism, culminating in the final prediction probability generated through a fully connected layer and a Sigmoid function (C), predicting the test set data across each fold of a fivefold cross-validation and averaging the prediction probabilities to derive the final classification probability for positive or negative outcomes (D)

DenseBlock of text convolution

In order to make full use of the features of different abstract levels in the sequence, we design DenseBlock (Text Convolutional Dense Block, TCDB) based on text convolution. The inspiration of this architecture comes from DenseNet, and its core idea is dense connection, that is, the input of each layer comes from the feature graph stitching of all previous layers [46]. Because of the natural advantage of one-dimensional convolution processing sequence, we choose to combine the idea of text convolution and Dense Block to design TCDB, and obtain good generalization ability in subsequent experiments.

Considering the effectiveness of one-dimensional convolution in processing sequential data, we integrate text convolution with the dense connection mechanism to construct TCDB, which improves feature reuse and enhances gradient propagation. This design allows the model to capture both local and global patterns in biological sequences, leading to improved generalization performance.

Bidirectional GRU

After processing the original One-hot features with TCDB, we use bidirectional GRU for feature capture of high-level, long-distance sequence dependencies. GRU is a special temporal neural network. Similar to LSTM [61], GRU can also handle gradient vanishing or explosion problems in traditional RNNs. GRU requires fewer parameters than LSTM and shorter training time, so we chose GRU to capture the long-distance dependencies implicit in the original sequence. Each GRU contains an update gate and a reset gate. The function of GRU can be expressed as follows:

1

2

3

4

For notational simplicity, bias terms are omitted in Eqs. (1–3). In the actual implementation, bias terms are included in all GRU computations.

GRU unit updates hidden state by gating mechanism: firstly, update gate and reset gate are obtained by sigmoid activation of input and previous hidden state respectively; then reset gate controls the influence of previous state on candidate hidden state , which is calculated by tanh activation; finally, current hidden state is generated by weighted summation of and by update gate .

Self-attention module

After extracting high-level sequential features through the TCDB and bidirectional GRU modules, we further introduce a self-attention mechanism to capture the most informative positions within the sequence for IRES prediction.

Let denote the output of the BiGRU layer, where L is the sequence length and d is the hidden feature dimension. Each represents the hidden state at position i.

First, each hidden state is transformed into an intermediate representation:

5

where and are trainable parameters, and is the dimension of the attention space.

Next, the importance of each position is evaluated using a trainable context vector :

6

where denotes the normalized attention weight of the i-th position.

Finally, the sequence representation is obtained as the weighted sum of all hidden states:

7

Cross-validation bagging

Bagging [62] is a highly effective ensemble learning [63, 64] technique. The approach we employ is akin to the concept of Bagging, yet we utilize cross-validation to construct the base model. Initially, we perform traditional fivefold cross-validation, a critical step in deep learning. Upon completion of this phase, we save the five models trained for subsequent integration. For predicting the test set, we implement a soft voting strategy: the test set is input into each of the five models, yielding five prediction probabilities for each sample. We then compute the average of these probabilities to derive the final prediction probability. By comparing the average performance metrics of the individual fold models (ACC, MCC) with those of the integrated models, we observed a significant enhancement in performance attributable to the integrated strategy. This underscores the efficacy of the method employed.

Performance evaluation metrics

To assess the performance of the model, we selected several common metrics: accuracy (ACC), sensitivity (SN), specificity (SP), Matthews correlation coefficient (MCC) [65], area under the ROC curve (AUC), and F1 score. When evaluated in a test set, the accuracy metric appears to be unevenly distributed.

8

9

10

11

12

Training details

The input RNA sequences were encoded as integer indices with a fixed length of 174 and transformed into one-hot embeddings of dimension 4 using a non-trainable embedding layer.

The feature extraction module consists of densely connected convolutional blocks. Specifically, four dense blocks were employed, each containing three convolutional layers with a growth rate of 48 and a kernel size of 3. A dropout rate of 0.27 was applied within the convolutional blocks to reduce overfitting, and transition layers were inserted between adjacent dense blocks.

The extracted features were then fed into a bidirectional GRU layer with 18 hidden units in each direction, resulting in a 36-dimensional output representation for each sequence position. A self-attention mechanism with an attention dimension of 12 was subsequently applied to aggregate sequence information.

The attention output was passed through a fully connected layer with 256 units, followed by a sigmoid activation function to produce the final prediction.

The model was trained using the Adam optimizer with a learning rate of 0.0011 and binary cross-entropy as the loss function. A batch size of 32 was used during training. To address class imbalance, class weights were computed from the training data and applied during optimization.

A five-fold cross-validation strategy was adopted for performance evaluation. In each fold, the model was trained for up to 100 epochs with early stopping based on the validation AUC, using a patience of 20 epochs. The model with the best validation AUC was saved and used for evaluation. Model performance was assessed using accuracy, sensitivity, specificity, Matthews correlation coefficient (MCC), AUC, AUPR, and F1-score as summarized in Table 1 .

Table 1.

Model hyperparameters and training settings

Sequence length	174
Number of dense blocks	4
Layers per dense block	3
Growth rate	48
Kernel size	3
Dropout rate	0.27
BiGRU units	18 (per direction)
Attention dimension	12
Fully connected layer size	256
Optimizer	Adam
Learning rate	0.0011
Loss function	Binary cross-entropy
Batch size	32
Epochs	100
Early stopping	Patience = 20 (monitor: validation AUC)

Class weight

To address the class imbalance problem (approximately 1:9), we employed a cost-sensitive learning strategy by assigning higher weights to minority-class samples during training. This approach avoids altering the original data distribution and is widely used in deep learning-based classification tasks.

Results and discussion

Implementation details of baselines

To ensure a fair and rigorous comparison with state-of-the-art methods, we standardized the implementation details across all competing baselines. We evaluated IRESFinder, IRESpy, DeepIRES, and several conventional machine learning models under consistent experimental conditions. The specific settings are as follows:

1. Data Consistency: All methods were evaluated on exactly the same training and test splits as DB-IRES. No changes were made to the original data distribution.

2. Sequence Preprocessing: We standardized the input sequence length to 174 nt for all methods. For sequence-based tools like IRESFinder and IRESpy, we applied no truncation or padding, ensuring that all models analyzed identical sequence information.

3. Feature Representation: For conventional machine learning baselines (SVM, Random Forest, KNN, and Logistic Regression), we employed the One-Hot encoding scheme, which is identical to the input representation used by DB-IRES. This eliminates potential performance biases arising from differences in feature engineering.

4. Thresholding: To ensure an "apples-to-apples" comparison of evaluation metrics (such as MCC and Accuracy), we applied a uniform classification threshold of 0.5 for all methods. For tools providing prediction scores, we normalized their outputs to the [0, 1] range and used 0.5 as the decision boundary.

5. Model deployment:

   1. Machine learning models: Trained directly on the full training set using scikit-learn with default hyperparameters.
   2. DeepIRES: We utilized the pre-trained model released by the authors to represent its best-reported performance on this dataset.
   3. IRESFinder: Downloaded and executed locally to ensure batch processing capability.
   4. IRESpy: Evaluated via the official web server using the standardized inputs.

These standardized settings ensure that the reported performance differences reflect the inherent advantages of the model architectures rather than variations in experimental setup.

Comparison with conventional machine learning models

To assess the performance of the proposed DB-IRES model, a comparative analysis was performed against four conventional machine learning models: random forest (RF) [66, 67], SVM, K-Nearest Neighbors (KNN) [68], and AdaBoost [69, 70], using both balanced and imbalanced datasets. The results, illustrated in a radar chart (Fig. 2) and ROC curves (Fig. 3), as well as summarized in Tables 2 and 3, indicate the effectiveness and robustness of DB-IRES.

Fig. 2.

Perform between DB-IRES and machine learning models on test sets

Fig. 3.

AUC curve of DB-IRES and machine learning models on test sets

Table 2.

Comparison between DB-IRES and machine learning models on the balanced test set. Bold values indicate the best performance for each evaluation metric across all models

Model	ACC	SN	SP	F1	MCC
RF	0.5000	0.0000	1.0000	0.0000	0.0000
SVM	0.5000	0.0000	1.0000	0.0000	0.0000
KNN	0.5079	0.0212	0.9947	0.0412	0.0695
ADBoost	0.5132	0.0265	1.0000	0.0515	0.1158
DB-IRES	0.7963	0.7619	0.8307	0.789	0.594

Bold values indicate the best performance for eachmetric.

Table 3.

Comparison between DB-IRES and machine learning models on the imbalanced test set. Bold values indicate the best performance for each evaluation metric across all models

Model	ACC	SN	SP	F1	MCC
RF	0.8980	0.0000	1.0000	0.0000	0.0000
SVM	0.8980	0.0000	1.0000	0.0000	0.0000
KNN	0.8985	0.0212	0.9982	0.0408	0.0955
ADBoost	0.8969	0.0265	0.9958	0.0498	0.0839
DB-IRES	0.8014	0.7679	0.8059	0.439	0.3958

Bold values indicate the best performance for eachmetric.

The radar charts presented in the two test sets (Fig. 2) indicate that DB-IRES exhibits generally better performance compared to other machine learning models, as evidenced by the larger area of its radar chart. While all machine learning models demonstrate improved SP and higher ACC on imbalanced test sets, this advantage arises from their inability to effectively differentiate between positive and negative samples, as reflected in their extremely low SN and MCC.

The superiority of DB-IRES is further validated through ROC analysis in Fig. 3. On the balanced dataset (left panel), the ROC curve of DB-IRES is closer to the top-left corner, outperforming other models. On the more challenging imbalanced dataset (right panel), DB-IRES maintains competitive performance, achieving the highest AUC, although the performance gap with some competitors diminishes. This consistent performance in both data scenarios suggests that the model has good discriminatory ability and generalization capability. In contrast, conventional machine learning models show relatively lower performance on the balanced dataset, as indicated by their ROC curves being closer to the random baseline.

Effect of cross-validation bagging

Figure 4 illustrates the performance of individual fold models (dotted line) in a five-fold cross-validation and the final integrated voting model (red line) across seven evaluation metrics (ACC, SN, SP, F1 score, MCC, AUC, AUPR).

Fig. 4.

Comparision between Bagging and model of 5 folds cross-validation

To improve clarity, the detailed numerical results of each fold, as well as their mean performance, are additionally reported in Tables 4, 5.

Table 4.

Comparision between Bagging and model of 5 folds cross-validation on balanced test. Bold values indicate the best performance for each evaluation metric across all models

Model	ACC	SN	SP	F1	MCC	AUC	AUPR
Fold1	0.7831	0.7354	0.8307	0.7722	0.5687	0.8544	0.8720
Fold2	0.7778	0.7090	0.8466	0.7614	0.5609	0.8457	0.8704
Fold3	0.7751	0.7566	0.7937	0.7709	0.5506	0.8621	0.8856
Fold4	0.7487	0.7778	0.7196	0.7558	0.4982	0.8407	0.8348
Fold5	0.7646	0.7143	0.8148	0.7521	0.5318	0.8472	0.8677
Mean(fivefold)	0.7698 ± 0.0122	0.7386 ± 0.0258	0.8011 ± 0.0443	0.7625 ± 0.0080	0.5420 ± 0.0252	0.8500 ± 0.0075	0.8661 ± 0.0168
DB-IRES(Bagging)	0.7963	0.7619	0.8307	0.789	0.594	0.8674	0.8864

Bold values indicate the best performance for eachmetric.

Table 5.

Comparision between Bagging and model of 5 folds cross-validation on imbalanced test. Bold values indicate the best performance for each evaluation metric across all models

Model	ACC	SN	SP	F1	MCC	AUC	AUPR
Fold1	0.8171	0.7354	0.8263	0.4506	0.4034	0.8484	0.4967
Fold2	0.8036	0.7090	0.8143	0.4241	0.3713	0.8349	0.4909
Fold3	0.7723	0.7566	0.7740	0.4040	0.3576	0.8444	0.4854
Fold4	0.7372	0.7778	0.7326	0.3764	0.3313	0.8387	0.4121
Fold5	0.7863	0.7143	0.7945	0.4054	0.3522	0.8363	0.4473
Mean(fivefold)	0.7833 ± 0.0276	0.7386 ± 0.0258	0.7883 ± 0.0331	0.4121 ± 0.0245	0.3632 ± 0.0239	0.8405 ± 0.0051	0.4665 ± 0.0322
DB-IRES(Bagging)	0.8014	0.7679	0.8059	0.439	0.3958	0.8557	0.5036

Bold values indicate the best performance for eachmetric.

The integrated voting model demonstrates improved performance compared to the single-fold models across most metrics, exhibiting greater stability. Notable improvements can be observed in SN, SP, and MCC, suggesting that the voting approach helps mitigate the variability and bias inherent in individual models. The improvements in ACC and AUPR further indicate an overall enhancement in model performance after applying the Bagging strategy.

These findings suggest that the proposed cross-validation Bagging technique effectively leverages the diversity among fold models. By integrating predictions from multiple models while preserving their individual strengths, it reduces the impact of variations in training data distributions and enhances the model’s generalization ability. Moreover, this approach consistently improves performance on both balanced and imbalanced test datasets, indicating its robustness under different data distributions.

Effect of different modules

DB-IRES consists of multiple modules, including a DenseBlock for text convolution, a bidirectional GRU, a self-attention module, and a fully connected layer. To investigate the effects of these different modules, we constructed several baseline deep learning models for performance comparison. TC_GRU_ATT denotes a model that integrates text convolution, a bidirectional GRU, and self-attention. TCDB_GRU refers to a model that employs DenseBlock in conjunction with text convolution and a bidirectional GRU. TCDB signifies a model that utilizes only DenseBlocks for text convolution.

To fully demonstrate the effectiveness and robustness of each module in DB-IRES, we conducted two module ablation experiments. Initially, to eliminate the impact of Bagging ensemble during cross-validation, we performed a fivefold cross-validation comparison on the training set between the model under evaluation in the ablation experiment and the model without cross-validation ensemble in DB-IRES. We preserved the cross-validated models for merging with the models on the test set following the same procedure as DB-IRES. This methodology not only mitigates potential confounding factors introduced by ensemble learning but also optimizes computational resource utilization.

The comprehensive analysis of the box plot (Fig. 5) reveals that, within this highly imbalanced training set, the application of a Dense Block for text convolution enhances the model’s capacity to differentiate between positive and negative samples compared to traditional text convolution methods. Furthermore, the results from the test set (Tables 6, 7) corroborate these findings, demonstrating that the Dense Block approach facilitates improved generalization capabilities for the model.

Fig. 5.

Comparison between DB-IRES and deep learning models of 5 cross validation on train set

Table 6.

Comparison on balanced test set between DB-IRES and Deep Learning models. Bold values indicate the best performance for each evaluation metric across all models

Model	ACC	SN	SP	F1	MCC	AUC	AUPR
TC_GRU_ATT	0.7249	0.7037	0.746	0.7189	0.4501	0.7992	0.7871
TCDB_GRU	0.7778	0.7871	0.7725	0.7789	0.5556	0.8648	0.8809
TCDB	0.7725	0.7725	0.8254	0.7598	0.5481	0.8523	0.8724
DB-IRES	0.7963	0.7619	0.8307	0.789	0.594	0.8674	0.8864

Bold values indicate the best performance for eachmetric.

Table 7.

Comparison on imbalanced test set between DB-IRES and Deep Learning models. Bold values indicate the best performance for each evaluation metric across all models

Model	ACC	SN	SP	F1	MCC	AUC	AUPR
TC_GRU_ATT	0.7264	0.7037	0.727	0.3441	0.2819	0.7891	0.3013
TCDB_GRU	0.776	0.7831	0.7752	0.4163	0.3756	0.8575	0.5036
TCDB	0.8025	0.7196	0.8119	0.4263	0.3754	0.8487	0.4852
DB-IRES	0.8014	0.7679	0.8059	0.439	0.3958	0.8557	0.5036

Bold values indicate the best performance for eachmetric.

Through the evaluation of TCDB_GRU and DB-IRES using fivefold cross-validation, we observed that the incorporation of the self-attention module resulted in a slight decrease in SN, while SP exhibited an increase. Consequently, both ACC and MCC showed improvement, and the performance of the F1 score also enhanced. This suggests that the self-attention module contributes to distinguishing between positive and negative classes. The results from the two test sets further corroborate this conclusion. On the balanced test set, the MCC for DB-IRES reached 0.594, surpassing the MCC of TCDB_GRU, which was 0.5556 and did not utilize the self-attention module. A similar trend was observed on the imbalanced test set, where DB-IRES achieved an MCC of 0.3958, compared to 0.3756 for TCDB_GRU without the self-attention module. These findings suggest that the addition of the self-attention module improves DB-IRES’s ability to perform the classification task.

Specifically, TCDB shows competitive performance, as evidenced by its results compared with traditional text convolution methods under both balanced and imbalanced test scenarios.

A comparison of TCDB_GRU and TCDB shows that the incorporation of GRU improves SN. Although the improvements in MCC and F1 are limited, the combination of DenseBlock, bidirectional GRU, and self-attention in DB-IRES leads to competitive overall performance, as reflected in both box plots and test set results (Tables 6 and 7).

To further validate the robustness of our model and quantify the contribution of each module, we performed paired t-tests based on the fivefold cross-validation results. The detailed performance metrics for each fold are provided in Supplementary Table S1.

The statistical analysis provides strong evidence that the proposed Text Convolutional Dense Block (TCDB) is the primary driver of the model’s superior performance. Specifically, DB-IRES significantly outperformed the baseline model lacking TCDB (TC_GRU_ATT) on both the balanced test set (Mean MCC: 0.5420 ± 0.0281 vs. 0.4143 ± 0.0239, p = 0.00025) and the imbalanced test set (Mean MCC: 0.3632 ± 0.0267 vs. 0.2616 ± 0.0163, p = 0.00119). These highly significant results (p < 0.01) confirm that the dense connection mechanism in TCDB is crucial for effective feature extraction and reuse in IRES prediction.

Regarding the comparisons with TCDB and TCDB_GRU, the differences in MCC did not reach statistical significance (p > 0.05). For instance, on the balanced test set, the p-values were 0.20 (vs. TCDB) and 0.66 (vs. TCDB_GRU). This suggests that while TCDB establishes a high-performance baseline, the marginal improvements contributed by the BiGRU and self-attention modules are relatively subtle in the fivefold setting. However, DB-IRES consistently achieved higher mean scores across all metrics compared to these baselines, indicating that the combination of these modules contributes to a more comprehensive feature representation. Furthermore, as demonstrated in “Comparison with conventional machine learning models” section, the ensemble learning strategy (Bagging) effectively stabilizes these improvements, ensuring that DB-IRES achieves the most robust and reliable performance overall.

Comparison with different coding ways

In the course of the experiment, we explored various encoding methods to investigate additional possibilities for the model. Ultimately, we employed three encoding techniques: Onehot, which utilizes only OneHot encoding as depicted in Fig. 6; NCP, which relies solely on NCP encoding [48, 71–73], which serves as a commonly employed feature extraction technique in computational biology studies of RNA and DNA sequences and is fundamentally based on the chemical characteristics of the constituent bases [74, 75]; and Onehot_NCP, which concatenates the feature matrices derived from both Onehot and NCP encodings to reconstruct the original features. Each of these encoding methods was processed using an identical model structure, with the final results presented in the bar chart in Fig. 6.

Fig. 6.

Comparison of different codings

The comprehensive evaluation of balanced and unbalanced datasets, as illustrated in the Fig. 6, demonstrates that one-hot encoding outperforms all other methods across seven metrics: ACC, SN, SP, F1, MCC, AUC, and AUPR. Its performance remains stable and exceeds that of individual NCP encoding and the combined Onehot and NCP encoding (Onehot_NCP). This finding suggests that the module in DB-IRES effectively extracts relevant features from sequences. It successfully captures and represents feature information from concise sequence data without incurring performance loss due to the simplicity of the feature input. Rather, it exhibits notable advantages in model generalization and discriminability. In contrast, more complex feature combinations, such as Onehot_NCP, or alternative encoding schemes, such as NCP, did not yield performance enhancements in this task.

Comparison with other deep learning models

To further validate the superior discriminatory capability of DB-IRES for IRES detection, we compared its performance with that of IRESFinder, IRESpy, and DeepIRES on two separate test sets, which is shown in Fig. 7.

Fig. 7.

Comparison between DB-IRES and other models

In two test sets, the DB-IRES achieved a more optimal balance between SN and SP, demonstrating its capability to effectively identify IRES elements while maintaining a low misjudgment rate. Additionally, the MCC and F1 values were significantly higher than those of IRESFinder and IRESpy, indicating stronger predictive stability and enhanced discriminative ability. Although the performance of DeepIRES on ACC is comparable to that of DB-IRES, its MCC and F1 values remain lower, suggesting that the stability and generalization of its classification boundary are relatively inadequate. In summary, the DB-IRES exhibited steady and balanced performance across both balanced and imbalanced test sets, maintaining higher SN and SP even amid variations in data distribution. This advantage under diverse data conditions further substantiates that DB-IRES effectively captures the internal characteristics of sequences, demonstrating excellent generalization and stability, thereby validating its effectiveness in the task of accurate IRES forecasting, as well as its robustness and reliability.

To further validate the robustness of DB-IRES in handling the class imbalance, we performed additional visual evaluations. Figure 8 displays the Precision-Recall (PR) curves and Calibration curves for both test sets. As shown in Fig. 8A, DB-IRES achieves a high Average Precision (AP) of 0.8864 on the balanced test set and a competitive AP of 0.5036 on the imbalanced test set. The PR curves illustrate that the model maintains strong discriminative power for the minority class even under severe imbalance, supporting the high MCC and F1-scores reported above. Furthermore, the calibration curves in Fig. 8B confirm that the predicted probabilities are well-calibrated with the true outcomes, ensuring the reliability of the model’s confidence estimates for downstream biological validation.

Fig. 8.

Detailed performance evaluation of DB-IRES on balanced and imbalanced test sets. (A) Precision-recall (PR) curves showing the trade-off between precision and recall. The average precision (AP) scores are 0.8864 (balanced) and 0.5036 (imbalanced) (B) Calibration curves showing the relationship between predicted probabilities and actual fractions of positives. The ‘Voting’ curve represents the DB-IRES model, demonstrating good alignment with the perfectly calibrated line

Motif analysis

To investigate the nucleotide distribution and sequence preference in the vicinity of IRES regions, we performed a comprehensive motif analysis using MEME software [76]. A motif is a short, conserved nucleotide or amino acid sequence pattern that is biologically significant, often corresponding to functional elements such as protein-binding sites or regulatory domains [77]. The analysis successfully identified several significant sequence motifs, one of which is visually represented in the sequence logo diagram below in Fig. 8A–D.

The motif pattern reveals a striking predominance of purine bases (adenine [A] and guanine [G]) within the core IRES regions. This purine-rich characteristic is biologically significant for multiple reasons. First, purine-rich sequences have a strong tendency to form stable secondary structures, such as stem-loops and complex RNA folds, which are essential for maintaining the structural integrity of IRES elements. Second, these motifs likely serve as recognition sites for trans-acting factors, including specific ribosomal subunits and RNA-binding proteins that facilitate the cap-independent translation initiation process. The high guanine content, in particular, may promote the formation of G-quadruplex structures, which have been implicated in translational regulation.

This finding provides molecular-level insights into the sequence characteristics of IRES elements, suggesting that purine-rich motifs may contribute to biological activity by facilitating structural stability and protein interaction capabilities. These observations align with known IRES mechanisms, though further experimental validation is required to fully establish the causal role of these specific motifs as illustrated in Fig. 9.

Fig. 9.

Motifs by MEME(A–D) Sequence logos of four distinct motifs identified from the dataset. Motif (A) has a length of 10 nucleotides, whereas motifs (B–D) each have a length of 4 nucleotides. In each logo, the height of each nucleotide indicates its relative frequency at that position, and the total height of the stack represents the information content. The horizontal axis denotes the position within the motif

UMAP visualization analysis

UMAP (uniform manifold approximation and projection) is a nonlinear dimensionality reduction technique that effectively preserves both local and global data structures when visualizing high-dimensional datasets [78]. We employed UMAP to evaluate the feature processing capability of our model by projecting the sample features into a two-dimensional space for visual analysis. As clearly demonstrated in the Fig. 10, the left panel displays the UMAP visualization of original one-hot encoded features, where negative samples are represented in light blue and positive samples in dark blue. The distribution pattern reveals a highly scattered and random arrangement of data points, with substantial overlap between the two classes. This intermixed clustering makes it extremely challenging to distinguish positive from negative samples based solely on the raw one-hot encoding, indicating the limited discriminative power of the initial feature representation. In striking contrast, the right panel presents the UMAP projection of features extracted through DB-IRES’s sophisticated architecture, which integrates DenseBlock, Bidirectional GRU (BiGRU), and Self-Attention mechanisms. The transformed feature space demonstrates remarkable clustering quality, with negative samples (light blue) and positive samples (dark blue) forming two distinct, well-separated clusters. The clear boundary between these clusters indicates that our model has effectively learned discriminative representations that capture the essential characteristics distinguishing IRES elements from non-IRES sequences. This visual evidence supports the conclusion that DB-IRES possesses superior feature extraction capability, successfully transforming the originally indistinguishable raw features into a well-structured latent space where positive and negative samples form distinct clusters. While UMAP is primarily a qualitative visualization tool, this observed improvement in feature separation is consistent with the high performance indicated by our quantitative metrics (e.g., AUC and MCC), suggesting that the model effectively captures the necessary discriminative features.

Fig. 10.

Visualization of original model input and sigmoid output

Limitation

Although the proposed method adopts a cost-sensitive learning strategy by using class weights to address the class imbalance problem, this approach still has certain limitations. Specifically, class weighting does not generate additional synthetic samples and therefore cannot fully increase the diversity of minority-class data. In addition, it assigns fixed penalties based on class frequency, which may not explicitly emphasize hard-to-classify samples during training, as in focal loss-based methods. Therefore, while class weighting provides a simple and effective solution for the imbalance setting in this study (approximately 1:9 ratio), its performance may be less optimal under more severe imbalance conditions or datasets with more complex distributions. "Additionally, the current DB-IRES model is trained and optimized on fixed-length RNA sequences of 174 nucleotides to maintain consistency with the DeepIRES benchmark dataset. This constraint limits the model’s direct application to sequences of varying lengths, such as full-length viral UTRs or cellular 5’UTRs. For practical applications involving sequences longer than 174 nt, we recommend employing a sliding window strategy. The sequence should be scanned with a window of 174 nt and a specific stride (e.g., 50 nt), predicting the IRES probability for each window. The region with the highest average score or maximum score can then be identified as a potential IRES candidate. Conversely, for sequences shorter than 174 nt, padding with ‘N’ (ambiguous nucleotides) or truncation to the nearest functional region can be applied before inputting them into the model. We hope to incorporate variable-length processing capabilities in future versions to enhance the model’s generalizability."

Conclusions

In this study, we developed DB-IRES, a deep learning model designed for the accurate identification of internal ribosome entry sites (IRES) in messenger RNA (mRNA) sequences. The model integrates TCDB, BiGRU, and a self-attention mechanism to jointly capture local sequence patterns, long-range dependencies, and key discriminative positions. Moreover, a five-fold cross-validation bagging strategy was employed to ensemble multiple base learners, effectively enhancing prediction stability and robustness.

Extensive experiments demonstrated the superior performance of DB-IRES compared with both conventional machine learning algorithms and state-of-the-art deep learning methods. Notably, DB-IRES achieved higher ACC, SN, MCC, and AUC across both balanced and imbalanced datasets, indicating its strong generalization capability and resilience to class imbalance. The UMAP visualization further confirmed that DB-IRES effectively transforms indistinguishable raw sequence representations into a well-separated latent feature space, while motif analysis revealed purine-rich sequence patterns biologically consistent with known IRES mechanisms.

Overall, DB-IRES provides a reliable and high-performance computational framework for IRES prediction. The biological motif analysis and feature visualization presented in this study offer valuable biological insights into the sequence characteristics of cap-independent translation mechanisms. In future work, we aim to extend the model to cross-species IRES datasets and apply formal feature attribution methods (such as attention-based visualization) to further elucidate the decision-making logic of the model.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

IRES

Internal ribosome entry site

TCDB

Text convolutional dense block

AUC

Area under the curve

AUPR

Area under the precision–recall curve

ACC

Accuracy

SN

Sensitivity

SP

Specificity

MCC

Matthews correlation coefficient

Author contributions

Song wrote main manuscript text and prepared most figures. Jia gave much guide while writing. Sun helped writing to make manuscript being better to read.

Funding

The authors declare that financial support was received for the research and/or publication of this article. This work was partially supported by the Scientific Research Plan of the Department of Education of Jiangxi Province, China (GJJ2400909, GJJ2402711). These funders had no role in the study design, data collection and analysis, decision to publish or preparation of manuscript.

Data availability

The source code and processed datasets generated in this study are publicly available. (https://github.com/lingxuanjue/StackIRESFinding) The datasets analyzed in this study were derived from previously published work. Specifically, the data were obtained from Weingarten-Gabbay et al. (Science, https://doi.org/10.1126/science.aad4939). The original datasets are available through the publication and its supplementary materials. All data sources have been clearly indicated to ensure transparency and reproducibility.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.