Deepvirusclassifier: a deep learning tool for classifying SARS-CoV-2 based on viral subtypes within the coronaviridae family

Open Access © The Author(s) 2024. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creativecommons.org/licenses/by/4.0/. RESEARCH Azevedoetal. BMC Bioinformatics (2024) 25:231 https://doi.org/10.1186/s12859-024-05754-1 BMC Bioinformatics Deepvirusclassifier: adeep learning tool forclassifying SARS-CoV-2 based onviral subtypes withinthecoronaviridae family Karolayne S. Azevedo1, Luísa C. de Souza1, Maria G. F. Coutinho1, Raquel de M. Barbosa1,3* and Marcelo A. C. Fernandes1,2,4* Abstract Purpose: In this study, we present DeepVirusClassifier, a tool capable of accurately classifying Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) viral sequences among other subtypes of the coronaviridae family. This classification is achieved through a deep neural network model that relies on convolutional neural networks (CNNs). Since viruses within the same family share similar genetic and structural characteristics, the classification process becomes more challenging, necessitating more robust models. With the rapid evolution of viral genomes and the increasing need for timely classification, we aimed to provide a robust and efficient tool that could increase the accuracy of viral identification and classification processes. Contribute to advancing research in viral genomics and assist in surveilling emerging viral strains. Methods: Based on a one-dimensional deep CNN, the proposed tool is capable of training and testing on the Coronaviridae family, including SARS-CoV-2. Our model’s performance was assessed using various metrics, including F1-score and AUROC. Additionally, artificial mutation tests were conducted to evaluate the model’s generalization ability across sequence variations. We also used the BLAST algorithm and conducted comprehensive processing time analyses for comparison. Results: DeepVirusClassifier demonstrated exceptional performance across several evaluation metrics in the training and testing phases. Indicating its robust learning capacity. Notably, during testing on more than 10,000 viral sequences, the model exhibited a more than 99% sensitivity for sequences with fewer than 2000 mutations. The tool achieves superior accuracy and significantly reduced processing times compared to the Basic Local Alignment Search Tool algorithm. Furthermore, the results appear more reliable than the work discussed in the text, indicating that the tool has great potential to revolutionize viral genomic research. Conclusion: DeepVirusClassifier is a powerful tool for accurately classifying viral sequences, specifically focusing on SARS-CoV-2 and other subtypes within the Coronaviridae family. The superiority of our model becomes evident through rigorous evaluation and comparison with existing methods. Introducing artificial mutations into the sequences demonstrates the tool’s ability to identify variations and significantly contributes to viral classification and genomic research. As viral surveillance *Correspondence: [email protected]; [email protected] 1 InovAI Lab, nPITI/IMD, Federal University of Rio Grande do Norte, Natal, RN 59078-970, Brazil 2 Bioinformatics Multidisciplinary Environment (BioME), Federal University of Rio Grande do Norte, Natal, RN 59078-970, Brazil 3 Department of Pharmacy and Pharmaceutical Technology, University of Seville, 41012 Seville, Spain 4 Department of Computer Engineering and Automation (DCA), Federal University of Rio Grande do Norte, Natal, RN 59078-970, Brazil Page 2 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 becomes increasingly critical, our model holds promise in aiding rapid and accurate identification of emerging viral strains. Keywords: SARS-CoV-2, Coronaviridae, Viral classification, Deep learning Introduction One particular virus has made of attention of the entire world, the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The virus belongs to the family Coronaviridae, which contains one of the largest viral genomes, ranging from 26,000 base pairs (bp) to 31,700 bp and is well known for infecting animals and humans [1].Viruses from the same family have similar genetic and structural characteristics, which makes the classification process more challenging. This is especially true considering that the selection or extraction of resources is essential to carry out such differentiation. Furthermore, viruses can undergo recombination events where genetic material from different viruses combines, blurring the lines between viral families. The SARS-CoV-2 causes the COVID-19 disease, which has caused the death of thousands of people worldwide due to its high virulence rate in conjunction with your rapid spread [2, 3]. The novel and timely classification systems are necessary for more insights into the evolution of underlying mechanisms of increased epidemicity and enhanced virulence compared to related lineages [4, 5]. Classifying and identifying viruses remains a crucial and relevant task, even with the end of the pandemic. It is a widely applied task by many scientists worldwide. Virus classification is essential in several contexts, including areas related to genomics and viral surveillance. Furthermore, it supports the control, prevention, and treatment of future complications that these agents may cause in a population. This knowledge is valuable for the development of treatments, therapies, and vaccines for both known and emerging viruses [6]. This activity assigns a certain sequence to a specific group based on known genomic sequences which share common characteristics and traits [7]. The conventional methods for characteristics extraction of the virus are based on sequence alignment [8, 9]. Alignment-based techniques search for regions of similarity between biological sequences from a previously characterized reference sequence. These techniques can also be used for viral identification [7]. Alignment-based techniques are used in algorithms like Basic Local Alignment Search Tool (BLAST) [10], Megan alignment tool (MALT) [11], FASTQ preprocessor (FASTP) [12], ClustalW [13] and USEARCH [14]. However, these methods have some limitations: low accuracy and limited genomic sequence length used [8, 15]. The use of long genomic sequences implies a high computational cost due to the nature of the problem [16]. Works presented in [7, 8] draw attention to the evidence that alignment-based methods are not quite satisfactory when applied to genomes susceptible to large genetic variations, which is the case of the vast majority of the viruses. Furthermore, due to the high computational cost involved, alignment-based methods make it impossible to analyze a large number of complete genomes and in many cases, the structures need to be homologous [16]. In order to minimize these problems, free-alignment (FA) techniques emerged, which are based on features from linear algebra, information theory and statistical mechanics to calculate the similarity or distance between sequences [7, 8]. Page 3 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 According to [7, 17, 18], to provide the best results, the viral classification based on free-alignment algorithms uses the artificial intelligence approach based on machine learning (ML) techniques to perform the feature extraction of the genomic sequences. Moreover, alignment-free techniques encompass methods that explore new forms of representation of input data by patterns identified in genomic data, as suggested by the works [15, 19, 20].Recent studies indicate that ML algorithms and techniques have been widely used in research related to genomics, including viral classification, for offering a set of methods capable of identifying highly complex patterns in an automated, efficient way and with the minimal human intervention [21, 22]. Works in the literature show that machine learning based on Deep Learning (DL) techniques provides excellent results for genomic sequences applications, including classification problems [23, 24]. Mottaqi [22] and Lalmuanawma [25] show that among many ML algorithms, the Convolutional Neural Networks (CNN) have been frequently used for data analysis based on genomic sequence for their ability to extract intrinsic characteristics of the sequences and present promising results in their applications. However, most of these tools and techniques use genomic sequences of limited length or are aimed at other purposes such as protein prediction [26, 27]. Fabijańska proposes a deep viral genome classifier, named VGDC (Viral Genome Deep Classifier), able to identify viral subtypes from different families such as dengue, hepatitis B and C, HIV-1, and influenza A presented F1-score between 0.85 and 1 [28]. Tampuu etal. presented an architecture to recognize the presence of viruses by the raw metagenomic contigs of various human samples. The methodology proposed was named ViraMiner and made use of two CNNs. They reached a Receiver Operating Characteristic (AUROC) curve of 0.923 [29]. The work presented by Whata etal. used a CNN and a Bi-LSTM (bi-directional long short-term memory), which he called CNN-Bi-LSTM (convolutional neural network bidirectional long short-term memory). This model achieved a classification accuracy of 99.95% , AUC of 100.00% , specificity of 99.97% , and sensitivity of 99.97% as from 34 sequences from the SARS-CoV-2 virus and 295 samples from other viruses of the same family [30]. The study presented by Adetiba etal. used a CNN to perform a multiclass classification of genomic sequences of three viral subtypes, MERS-CoV (Middle East Respiratory Syndrome CoV), SARS-CoV (Severe Acute Respiratory Syndrome CoV), and SARSCoV -2 (Severe Acute Respiratory Syndrome Coronavirus 2). The authors used the GSP (Genomic Signal Processing) technique to transform the genomic sequences into RGB images and later applied them to a CNN, using only 300 samples for training. The model obtained an accuracy of 95% for MERS-CoV, 95% for SARS-CoV, and 95% for SARSCoV-2, titled by the authors DeepCOVID-19 [31]. Classification between SARS-CoV-2, MERS-CoV, SARS-CoV, hepatitis-A, dengue, and influenza was proposed by Gunasekaran etal. Therefore, the authors use the CNN, CNN-LSTM, and CNN-Bidirectional LSTM architectures with k-mers to verify which architectures present better performance. According to the tests performed, it was observed that CNN and CNN-Bidirectional LSTM with k-mers offered the highest accuracy metrics, reaching 93.16% and 93.13% , respectively [32]. A neural network called miRNA proposed by Lopez-Rincon etal. was applied at viral classification. The Page 4 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 architecture has a few layers and was also used to classify viruses from the Coronaviridae family. This model showed an accuracy of 98% , specificity of 0.9939, and sensitivity of 1.00 [24]. Several viral genomic sequences of different sizes were analyzed by [33], which used the area under the receiver operating characteristic (AUROC) as their performance metric. The research obtained AUROC values of 0.95, 0.93, 0.97, and 0.98, for the genomic sizes 300, 500, 1000, and 3000 bp, respectively. The architecture used was called DeepVirFinder and consists of a CNN of multiple layers [33]. Given this context, the present work aims to present a technique capable of classifying the Coronaviridae family’s viruses and recognizing the SARS-Cov-2 virus. That approach uses the CNN that receives complete genomic sequences of cDNA as input, codified by the one-hot-encoding technique. The proposed method has high metrics and has been tested with over 10,000 complete SARS-CoV-2 sequences. Thus, this work makes the following specific contributions: • Develop an alignment-free method to classify SARS-CoV-2 sequences between viruses from the same family, well known in the literature. • Develop a deep learning algorithm that can efficiently classify the complete cDNA sequences of the virus. • Comparison of the performance of the proposed model with the BLAST algorithm, recognized as the gold standard among alignment-free techniques, in terms of the number of samples found or correctly classified and the processing time taken by both tools to present their results. • Utilization of a DL technique to analyze large datasets, enabling the efficient classification of numerous viral sequences in a short amount of time. • Reduced computational cost when classifying many sequences compared to traditional established alignment-free methods. • Use of partially mutated cDNA sequences to test the generalization and efficiency of the model in covering future mutations that may occur in the virus. Results Training andvalidation As mentioned in “Database and data balancing” section, the dataset used for training the network comprises 501 samples referring to the Non-SARS group and receiving label 0 and 501 samples from SARS, in which they obtained label 1. In this way, we obtained a training set balanced and homogeneous consisting of 1002 samples. Cross-validation was used to train and validate the classification model (see “CNN architecture and parameters” section). The performance metrics for the k-fold ( k=5 ) cross-validation corresponded to the average between all the values obtained in each fold. The classification results of validation (after training) were presented through the confusion matrix (see Fig.1), the AUROC (see Fig.2), and measured by the sensitivity, specificity, precision, accuracy, and F1-score metrics (see Table1). As a result, the model results in maximum performance values for the training and validation sets, as shown in Table1. Figure1 presents the results of the mean classification of the samples referring to the validation set (SARS-CoV-2 and Not SARS-CoV-2) and shows that for all subsets, all Page 5 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 sequences were correctly grouped according to their respective class. The ROC curve for this problem is shown in Fig.2 and presents sensitivity and specificity values equal to 100% , according to Table1. Figures3 and 4 illustrate the training and validation learning curve for accuracy and loss, respectively. Each iteration point represents the mean and standard deviations of the fivefold cross-validation. The accuracy learning curve of training and validation (see Fig.3) corroborates with the results presented in Table1, and these curves show that the model does not suffer from overfitting (high variance) or underfitting (high bias). Furthermore, the reduced difference (almost zero) between the training and validation curves consolidates the absence of overfitting. The training was concluded after 10 epochs with 72 iterations, as shown in Figs.3 and 4. It is observed that the error was stabilized after the 30th iteration (see Fig.4). SARS‑Cov‑2 prediction tests Similar to the methodology used in [16], two tests were performed to evaluate the SARS-Cov-2 prediction of the proposed deep learning model after training. The tests were composed of samples not used in the training stage, that is, samples that remained Table 1 Performance metrics results for the classification of SARS-Cov-2 from the architecture proposed in this work for the validation set Metrics Performance Sensitivity 100% Specificity 100% Precision 100% Accuracy 100% F1-score 1 10 Predicted Class 0 1 True Class 99 101 Fig. 1 Confusion matrix of the proposed approach for the classification problem of distinguishing between SARS-CoV-2 and Non-SARS-CoV-2 samples. Non-SARS-CoV-2 samples are represented by label 0, and SARS-CoV-2 samples are represented by label 1. The model is capable of correctly classifying all samples according to their respective classes Page 6 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 from the initial dataset belonging to the SARS-CoV-2 virus (see “Pre-processing and data mapping” section). The tests, called Prediction test 1 and Prediction test 2, are described below. Prediction test 1 Of the remaining 16,891 SARS-CoV-2 samples from the initial dataset, 12,000 were randomly chosen to compose this experiment. These samples obtained label 1 indicating 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 False Positive Rate 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 True Positive Rate Class 1 Class 2 Fig. 2 AUROC curve for classification of SARS-CoV-2 and Non SARS-CoV-2 Fig. 3 The learning curve of training and validation accuracy of the training set using fivefold cross-validation Page 7 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 that they were SARS-CoV-2. The objective of this experiment was to test the model for identifying SARS-CoV-2. Prediction test 2 For this experiment, 10,000 samples of SARS-CoV-2 were used (of the remaining 16,891 SARS-CoV-2 samples from the initial dataset), in which they were divided into two groups, each with 5000 samples. In one of these groups, we applied the artificial mutation method discussed in “Artificial mutation technique” sectionto investigate the architecture’s sensitivity and robustness to possible mutations in the SARS-CoV-2 virus. In this way, a group was created with 5000 samples of the SARS-CoV-2 virus, which suffered artificial mutations, and another group, also with 5000 samples, which did not undergo any mutation. The artificial mutation strategy used Vmax =31,029 and γ=5% , i.e., Nmut =1551 nucleotides have changed per sequence. Prediction test results The results of Prediction tests 1 and 2 are shown in Table2. For prediction test 1, 11,996 were correctly classified to their respective group (SARS-CoV-2), and only 4 samples were not classified correctly, reaching 99.99% , 100% , 99, 94% , and 99, 96% for the sensitivity, precision, F1-score, and accuracy, respectively. As described above, prediction test 2 verified the ability of the trained model to classify SARS-CoV-2 samples even after changing their genomic structure through the artificial mutation technique in half of the dataset samples. Even applying modifications to the sequences, the model is quite sensitive to possible mutations that the sequences may suffer, reaching a sensitivity value of Fig. 4 The learning curve of training and validation loss of the training set using fivefold cross-validation Table 2 Results associated with prediction tests 1 and 2 Pt Sensitivity (%) Precision (%) F1‑score Accuracy (%) Pt-1 99.99 100 0.94 99.96 Pt-2 99.77 100 0.88 99.96 Page 8 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 99.77% . This result strongly attests to the model’s ability to generalize, given that, even with the samples changing, the network can identify who is SARS-CoV-2 through low false negative results (accuracy about 99.96% ). The results obtained through the experiments carried out and detailed in “Pre-processing and data mapping” section, are promising, consistent with the performance obtained in the network training phase. Furthermore, the sensitivity and precision values derived from the set of experiments remain high regardless of the class labels, which is very important, considering that high rates of false negatives directly corroborate the increase in infected people. The biological implications of these results are significant, as they showcase the robustness and high accuracy of the model in detecting SARS-CoV-2 even in the presence of artificial mutations. This underscores the model’s potential for practical applications in viral detection and classification, with implications for disease diagnosis and management. The high sensitivity of the tool is crucial in virus detection, as it minimizes the risks of false positives, ensuring reliable virus identification. High precision reduces unnecessary alerts or classification errors, which can have biological and public health consequences as viruses undergo mutations over time. A model that remains sensitive to these changes is invaluable for real-world applications, especially in the detection of new viral strains. The results obtained with this tool demonstrate the model’s resilience, high precision, and potential for practical applications in viral detection and classification, supporting diagnosis, disease management, and the detection of new viral variants. Finally, the proposed model’s characteristics and results will be compared and discussed with works found in the literature below. Methods The viral classification tool proposed in this work utilized genomic data from the cDNA of nine viral subtypes belonging to the Coronaviridae family, including SARS-CoV-2. The dataset underwent preprocessing, including balancing, transforming, and mapping viral sequences (see “Pre-processing and data mapping” section) to construct a homogeneous and balanced dataset. Subsequently, the CNN trained and processed the data, capable of extracting intrinsic features from the sequences, providing us with the classification result as either SARS-CoV-2 or non-SARS-CoV-2. Figure5 below displays the flowchart of activities. Database anddata balancing The National Genomics Data Center (NGDC) provides open and free access to a set of database resources that have the resources of the New Coronavirus 2019 Data Resource - 2019nCoVR. The 2019nCoV maintains daily updates and brings together a comprehensive collection of genomic sequences and clinical information, not only about SARS-CoV-2 but also regarding other viruses that belong to the coronaviridae family worldwide and from other traditional repositories, such as the National Center for Biotechnology Information - NCBI [34]. The 2019nCoV was the chosen repository to download the dataset. Sequences belonging to the coronaviridae family were selected, whose size ranges from 25,000 to 35,000 bp, covering the size of all viruses in the family without losing any crucial genetic information. The selected host was the Homo Sapiens. The download of the dataset used in this research was carried out in August 2020, when the variants of concern were not yet available. Page 9 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 The database used is formed by 17,893 genomic sequences of nine types of viruses of the coronaviridae family, coming from 62 different countries. Figure6 shows all countries with genomic samples on the database. It is observed that the United States has the Fig. 5 Overview of the proposed technique Fig. 6 Countries that contain genomic samples of the coronaviridae family in the database Page 16 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 It is important to note that the designations si,m and ski , m refer to the same element, where ki identifies the exact position of the nucleotide that will undergo alteration in the sequence si,m . Discussion Blast comparison The strategy proposed in this work was compared with the BLAST algorithm. The comparison obtained results associated with the correctness rate in the classification of sequences through various values of artificial position mutation rate (see “Artificial mutation technique” section) and the average processing time to classify these sequences. In the comparison, 34 sequences belonging to the Coronaviridae family were used (17 SARS-CoV-2 and 17 Not SARS-CoV-2) that did not participate in the deep learning training. The BLAST software version 2.13.0 made available by the NCBI [34] was downloaded and installed locally. The BLAST software used a database of 6,180,834 Betacoronavirus sequences (updated Sep 8, 2022) found in [34]. The database was also downloaded for local use. Using the BLAST software locally, accessing a local database allows a fairer comparison in terms of processing time with the deep learning strategy proposed in this work. The same computer used to run BLAST with its database was also used to train and run the CNN strategy. The computer has the following configurations: Intel(R) core(TM) i7-10700 CPU 2.9 GHz, 128 GBytes of RAM, 512 GBytes NVMe HD and an NVIDIA GeForce RTX 3060 GPU with 12 GBytes of RAM. Figure10 presents the relationship between the artificial position mutation rate (see “Artificial mutation technique” section) applied in the 34 test sequences and the correctness rate (in percentage terms) of both the BLAST and the proposed CNN. It is possible to observe that up to γ≈2% ( Nmut ≈620 nucleotides), the correctness rate for BLAST and CNN-based strategy is the same, that is, 100% . However, for values of γ>2% , the correctness rate of BLAST drops rapidly to 50% , in which γ≈19% ( Nmut ≈5895 nucleotides). On the other hand, the proposal based on CNN has a correctness rate of 100% up to γ≈13% ( Nmut ≈4033 nucleotides) and decays more slowly than BLAST, with γ>13% . For γ≈19% , a proposal based on CNN has a correctness rate of around 95.88% and BLAST around 50% . For values of γ between ≈32% ( Nmut ≈9,929 nucleotides) and ≈45% ( Nmut ≈13,963 nucleotides), the correctness rate of BLAST rapidly decays to zero while the proposal with CNN decays more slowly to 50% . Table6 presents the values of correctness rate, artificial position mutation rate, γ , and the number of nucleotides that mutated, Nmut , for each point in the graphs shown in Fig.10. Table7 presents the average processing time obtained for BLAST and CNN at each point presented in the graphs in Fig.10. The data presented for CNN are the time required to perform the inference of the 34 test sequences, given that the training is performed only once. However, the time for training the CNN was approximately 341 (8) s ki,m=          A if s ki= T T if ski=A C if ski=G G if ski=C N if s ki =T . Page 17 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 s (around 6 min). It is possible to observe that CNN has a constant processing time while BLAST has a variable processing time that depends on the value of γ . For sequences with many mutations, γ>25.78 ( Nmut >8000 ), BLAST has a faster response (shorter processing time) than for sequences with few mutations γ<3.22 ( Nmut <1000 ). Sequences with many mutations allow BLAST to reduce the search space due to the high dissimilarity between the query sequence and the sequences stored in the base. On the other hand, when the value of g decreases, the BLAST processing time increases to obtain a better similarity value between the query sequence and the sequences stored in the base. 0236 13 19 23 26 29 32 35 39 42 45 0 10 20 30 40 50 60 70 80 90 100 Fig. 10 Comparison of the correctness rate between BLAST and CNN (proposed in this work) for a test set of 34 sequences according to the increase of the artificial position mutation rate, γ Table 6 Values of correctness rate, artificial position mutation rate, γ , and the number of nucleotides that mutated, Nmut , for each point in the graphs shown in Fig. 10 γ ( % ) Nmut BLAST CNN Correctness rate ( % ) Correctness rate ( % ) 0.32 100 100.00 100.00 1.61 500 100.00 100.00 3.22 1000 91.18 100.00 6.45 2000 79.41 100.00 12.89 4000 61.76 100.00 19.34 6000 50.00 95.88 22.56 7000 50.00 93.53 25.78 8000 50.00 87.65 29.01 9000 50.00 76.47 32.23 10,000 50.00 67.65 35.45 11,000 41.18 61.76 38.67 12,000 29.41 58.82 41.90 13,000 11.76 54.71 45.12 14,000 0.00 52.35 Page 18 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 The gain in CNN processing time over BLAST is significant, being around 2600 times faster for γ=45.12% ( Nmut =14,000 ) and 130,000 times faster for γ=0.32% ( Nmut =100 ). It is essential to point out that BLAST needs a database of sequences already stored to find or classify the viral genome, and with this, it needs to carry out a search procedure which can take a long time. CNN stores the information needed to classify the viral genome in its models after the training process. After training, the CNN performs only a simple inference process, not needing to perform a search and a database. The proposed CNN model can be an excellent alternative and ally in the rapid virus classification process, given its high sensitivity in detecting changes in the virus structure (represented by random mutations in its nucleotides), corroborating SARS-Cov-2 surveillance. In addition, this model enables the analysis of more significant amounts of complete genomic samples, at a lower computational cost, compared to techniques that use alignment and even BLAST. State oftheart comparison The Tables8 and 9 summarize a set of approaches from the main works found in the literature, and addressed in this article, that perform viral classification using CNNs and viral sequences as input data with the aim of maintain a fairer comparison with the proposed technique. Characteristics such as the number of layers and size of genomic sequences will be presented in Table8. When applying longer sequences, the works presented in [28, 29, 33] had a considerable reduction in the performance of their models. This point implied the use of more extensive networks as in [28] and the reduction of sequence sizes as in works [29, 33]. Regarding [24], despite making use of complete genomic sequences and presenting a smaller number of layers, the author makes use of a small dataset for the training and Table 7 Time processing, artificial position mutation rate, γ , and the number of nucleotides that mutated, Nmut , for each point in the graphs shown in Fig. 10 γ ( % ) Nmut BLAST CNN Time processing (s) Time processing (s) 0.32 100 94,261.48 ( ≈26.2 h) 0.33 1.61 500 94,261.48 ( ≈26.2 h) 0.35 3.22 1000 93,202.74 ( ≈25.9 h) 0.39 6.45 2000 92,172.83 ( ≈25.6 h) 0.45 12.89 4000 91,176.66 ( ≈25.3 h) 0.68 19.34 6000 64,122.58 ( ≈17.8 h) 0.68 22.56 7000 24,587.36 ( ≈6.8 h) 0.68 25.78 8000 68,17.63 ( ≈1.9 h) 0.68 29.01 9000 44,35.43 ( ≈1.2 h) 0.68 32.23 10,000 2155.14 ( ≈0.6 h) 0.68 35.45 11,000 1940.66 ( ≈0.5 h) 0.68 38.67 12,000 1831.33 ( ≈0.5 h) 0.69 41.90 13,000 1832.6 ( ≈0.5 h) 0.69 45.12 14,000 1801.26 ( ≈0.5 h) 0.68 Page 19 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 validation of his model, which may lead to generalization problems and consequently on the performance of your network by presenting new samples. Table9 compares the performance results of the proposed architecture with the available results of the models in Table8. Although it presents an architecture with many layers, the variation in the performance values of the VGDC architecture was observed as the size of the genomic sequences used in the network increased. Although it uses two convolutional branches, the ViraMiner tool achieved 92.3% and 32% of the sensitivity and precision values, even using relatively short sequences. The DeepVirFinder architecture provided only the AUROC values obtained in its model, reaching the maximum value of 96.68% for samples with 3000 bp. Despite having obtained the sensitivity value of 100% and accuracy of 98% . The work presented by [24] obtained the AUROC value of 92% . The results obtained in the proposed model are superior for all architectures and performance metrics presented in Table9, indicating the high performance and robustness of the model. The DeepVirusClassifier showcases a robust learning capacity, as demonstrated by its ability to achieve exceptional performance when tested on a large dataset comprising more than 10,000 viral sequences. It maintains a sensitivity of over 99% for sequences with fewer than 2000 mutations. Conclusion Classification and prediction of viral sequences using deep neural networks (DNN) have shown great promise in recent years. This work proposes a tool, called DeepVirusClassifier, which uses a DNN-type CNN capable of classifying SARS CoV 2 through a binary classification based on complete genomic cDNA sequences among eight viral subtypes belonging Table 8 Comparison from the proposed architecture with related works References Codification Layers Sequence length Fabijańska e Grabowski [28] ASCII 30 3257–24,751 bp Ren et al. [33] One-hot encoded 6 150–3000 bp Tampuu et al. [29] One-hot encoded 2 CNNs with 7 layers each 300 bp Lopez-Rincon et al. [24] Assigned values from 0 to 1 to the channels 10 31,029 Proposed architecture One-hot encoded 26 31,029 Table 9 Performance metrics comparison from the proposed architecture with related works Ref. Accuracy Precision Sensibility Specificity F1‑score AUROC [28] 0.99–1 0.83–1 0.84–1 0.99–1 0.83–1 – [33] – – – – – 0.8635 0.9210 0.9496 0.9668 [29] 0.90 0.90 0.32 – – 0.923 [24] 0.985 0.98 1 0.9939 0.9797 0.92 This work 1 1 1 1 1 1 Page 20 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 to the same family. For this experiment, the cross-validation technique with k=5 folder was used, which reached maximum values in all evaluation metrics for the 960 samples used in training. More than 10,000 sequences were used to test the performance of the DNN after training. An artificial mutation technique was also used to test the generalizability of the model with sensitivity greater than 99% for less than 2000 mutations in the sequence. A test set consisting of 34 samples from the two classes experienced different position mutation rates and was processed by the model proposed in this work in conjunction with the BLAST algorithm to verify its performance in terms of accuracy rate according to the two classes. Taking into account results of accuracy and processing time, the proposed tool appears to be superior. To establish the superiority and practical applicability of our model, we carried out a comparative analysis with existing viral classification works in the literature, our results surpassed them. The proposed model was superior, indicating that the tool proposed in this work can be applied to classify viruses from the Coronaviridae family and viruses from different species. While the text primarily concentrates on classifying sequences from SARS-CoV-2 and the Coronaviridae family, the model architecture is versatile and has the potential to be adapted for classifying sequences from other viral families or applied to various sequence classification tasks. Our research signifies a substantial advancement in the field of viral sequence classification, opening the door to more precise and efficient tools in virology and bioinformatics and establishing itself as a reference for future research. DeepVirusClassifier significantly contributes as a foundation for early disease detection and diagnosis, genomic surveillance, and drug development, and even aids in identifying specific viral strains. Acknowledgements The authors wish to acknowledge the financial support of the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) and Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) for their financial support. Author contributions All the authors have contributed in various degrees to ensure the quality of this work (e.g., KSA, LCdS, MGFC, RdMB and MACF conceived the idea and experiments; KSA, LCdS, MGFC, RdMB and MACF designed and performed the experiments; KSA, LCdS, MGFC, RdMB and MACF analyzed the data; KSA, LCdS, MGFC, RdMB and MACF wrote the paper. MACF coordinated the project). All authors have read and agreed to the published version of the manuscript. Funding Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) - Finance Code 001 Availability of data and materials The datasets generated and/or analysed during the current study are available in the Mendeley Data repository, https:// data. mende ley. com/ datas ets/ zmhsn 2gz7w/1. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests No competing interest is declared. Received: 23 August 2023 Accepted: 19 March 2024 References 1. Wang H, et al. The genetic sequence, origin, and diagnosis of SARS-CoV-2. Eur J Clin Microbiol Infect Dis. 2020;39:1–7. Page 21 of 21 Azevedoetal. BMC Bioinformatics (2024) 25:231 2. Maghdid HS, Ghafoor KZ, Sadiq AS, Curran K, Rabie K. A novel AI-enabled framework to diagnose coronavirus COVID 19 using smartphone embedded sensors: design study; 2020. arXiv: 2003. 07434. 3. Chowdhury MEH, et al. Can AI help in screening viral and COVID-19 pneumonia? IEEE Access. 2020;8:132665–76. 4. Toyoshima Y, Nemoto K, Matsumoto S, Nakamura Y, Kiyotani K. SARS-CoV-2 genomic variations associated with mortality rate of COVID-19. J Hum Genet. 2020;65:1075–82. 5. Remita MA, et al. A machine learning approach for viral genome classification. BMC Bioinform. 2017;18:1–11. 6. Cobbin JC, Charon J, Harvey E, Holmes EC, Mahar JE. Current challenges to virus discovery by meta-transcriptomics. Curr Opin Virol. 2021;51:48–55. 7. Lebatteux D, Remita AM, Diallo AB. Toward an alignment-free method for feature extraction and accurate classification of viral sequences. J Comput Biol. 2019;26:519–35. 8. Zielezinski A, Vinga S, Almeida J, Karlowski WM. Alignment-free sequence comparison: benefits, applications, and tools. Genome Biol. 2017;18:1–17. 9. Nooij S, Schmitz D, Vennema H, Kroneman A, Koopmans MP. Overview of virus metagenomic classification methods and their biological applications. Front Microbiol. 2018;9:749. 10. Altschul SF, et al. Gapped blast and psi-blast: a new generation of protein database search programs. Nucleic Acids Res. 1997;25:3389–402. 11. Vågene ÅJ, et al. Salmonella enterica genomes from victims of a major sixteenth-century epidemic in Mexico. Nat Ecol Evol. 2018;2:520–8. 12. Lipman DJ, Pearson WR. Rapid and sensitive protein similarity searches. Science. 1985;227:1435–41. 13. Thompson JD, Higgins DG, Gibson TJ. CLUSTAL W: improving the sensitivity of progressive multiple sequence alignment through sequence weighting, position-specific gap penalties and weight matrix choice. Nucleic Acids Res. 1994;22:4673–80. 14. Edgar RC. Search and clustering orders of magnitude faster than blast. Bioinformatics. 2010;26:2460–1. 15. de Souza LC, Azevedo KS, de Souza JG, Barbosa RdM, Fernandes MA. New proposal of viral genome representation applied in the classification of SARS-CoV-2 with deep learning. BMC Bioinform. 2023;24:1–19. 16. Randhawa GS, et al. Machine learning using intrinsic genomic signatures for rapid classification of novel pathogens: COVID-19 case study. PLoS ONE. 2020;15: e0232391. 17. Randhawa GS, Hill KA, Kari L. ML-DSP: machine learning with digital signal processing for ultrafast, accurate, and scalable genome classification at all taxonomic levels. BMC Genom. 2019;20:267. 18. Ren J, Ahlgren NA, Lu YY, Fuhrman JA, Sun F. VirFinder: a novel k-mer based tool for identifying viral sequences from assembled metagenomic data. Microbiome. 2017;5:1–20. 19. Coutinho MG, Câmara GB, Barbosa RdM, Fernandes MA. SARS-CoV-2 virus classification based on stacked sparse autoencoder. Comput Struct Biotechnol J. 2023;21:284–98. 20. Hu L, Yang Y, Tang Z, He Y, Luo X. FCAN-MOPSO: an improved fuzzy-based graph clustering algorithm for complex networks with multi-objective particle swarm optimization. IEEE Trans Fuzzy Syst; 2023. 21. Davenport T, Kalakota R. The potential for artificial intelligence in healthcare. Future Healthc J. 2019;6:94. 22. Mottaqi MS, Mohammadipanah F, Sajedi H. Contribution of machine learning approaches in response to SARSCoV-2 infection. Inform Med Unlocked. 2021;100526. 23. Park Y, Kellis M. Deep learning for regulatory genomics. Nat Biotechnol. 2015;33:825–6. 24. Lopez-Rincon A, et al. Accurate identification of SARS-CoV-2 from viral genome sequences using deep learning. bioRxiv; 2020. 25. Lalmuanawma S, Hussain J, Chhakchhuak L. Applications of machine learning and artificial intelligence for COVID19 (SARS-CoV-2) pandemic: a review. Chaos Solitons Fractals. 2020;110059. 26. Eraslan G, Avsec Ž, Gagneur J, Theis FJ. Deep learning: new computational modelling techniques for genomics. Nat Rev Genet. 2019;20:389–403. 27. Zou J, et al. A primer on deep learning in genomics. Nat Genet. 2019;51:12–8. 28. Fabijańska A, Grabowski S. Viral genome deep classifier. IEEE Access. 2019;7:81297–307. 29. Tampuu A, Bzhalava Z, Dillner J, Vicente R. Viraminer: deep learning on raw DNA sequences for identifying viral genomes in human samples. PLoS ONE. 2019;14: e0222271. 30. Whata A, Chimedza C. Deep learning for SARS CoV-2 genome sequences. IEEE Access. 2021;9:59597–611. 31. Adetiba E, et al. DeepCOVID-19: a model for identification of COVID-19 virus sequences with genomic signal processing and deep learning. Cogent Eng. 2022;9:2017580. 32. Gunasekaran H, et al. Analysis of DNA sequence classification using CNN and hybrid models. Comput Math Methods Med. 2021;2021. 33. Ren J, et al. Identifying viruses from metagenomic data using deep learning. Quant Biol. 2020;8:1–14. 34. NCBI. GenBank overview; 2020. https:// www. ncbi. nlm. nih. gov/ genba nk/. 35. Kingma DP, Ba J. Adam: a method for stochastic optimization; 2014. arXiv: 1412. 6980. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.