A bacterial selenoprotein gene recognition method and terminal based on deep learning

Abstract

This invention relates to the field of biomedical technology, and more particularly to a method and terminal for identifying bacterial selenoprotein genes based on deep learning. This invention collects a large amount of gene sequence information from various known bacterial selenoprotein families, as well as stop-TGA codons from species that do not use selenoproteins. A deep learning model is then used to train and debug this model. By deeply mining the sequence and structural features downstream of the Sec-TGA codon, combined with parameter optimization and performance evaluation, a bacterial selenoprotein gene identification model capable of accurately identifying Sec-TGA codons in bacterial genomes is established. By inputting the TGA codon and downstream base sequence from the genome of the bacterial species to be tested into this model, the selenoprotein gene identification results can be obtained. This identification method significantly improves the speed and accuracy of selenoprotein gene identification in bacterial genomes. The analysis time for each genome can be completed within 10 minutes, greatly reducing labor and time costs, and lowering analysis costs and risks.

Description

Technical Field

[0001] This invention relates to the field of biomedical technology, and in particular to a method, terminal, and computer-readable storage medium for identifying bacterial selenoprotein genes based on deep learning. Background Technology

[0002] Selenium is an essential trace element for most organisms, possessing powerful antioxidant properties and playing a vital role in disease prevention, anti-aging, antiviral activity, anti-inflammation, and heavy metal detoxification. Selenium primarily inserts into selenocysteine ​​(Sec) in the form of amino acids, thereby participating in various metabolic pathways in the body. Most selenocysteine ​​proteins are involved in important redox processes in vivo, exhibiting functions such as scavenging peroxides, preventing cell damage, and participating in hormone synthesis.

[0003] In the past two decades, selenium has been recognized as an important new area of ​​research related to human health, longevity, and disease prevention, attracting widespread attention. Significant progress has been made in the study of the in vivo synthesis of sec and the molecular mechanisms of selenoprotein insertion. These mechanisms mainly include the TGA codon specifically encoding sec (Sec-TGA), a sec insertion sequence (SECIS) element located on the mRNA immediately following the sec-TGA, sec-specific tRNA, and several related proteases (Sec synthase SelA, sec-specific elongation factor SelB, selenophosphate synthase SelD, etc.). Since the most important function of TGA is as a protein synthesis stop codon (i.e., stop-TGA), the most critical issue in the selenoprotein synthesis mechanism is how to correctly distinguish between sec-TGA and stop-TGA.

[0004] SECIS elements are conserved RNA secondary structures unique to all selenoprotein genes and are essential cis-key factors in selenoprotein synthesis. The types and quantities of selenoproteins can vary considerably among different organisms. Effective identification of various selenoprotein-encoding genes in species (especially bacteria) will open an important window for a deeper understanding of the biological functions of selenium, exploring the molecular mechanisms of related diseases, and even the rational development of selenium-enriched biological products.

[0005] Traditionally, the identification of selenoproteins has relied on small-scale biological experiments, which is time-consuming and labor-intensive. Existing methods for selenoprotein gene sequencing fall into two categories: algorithms that rely on SECIS elements and those that do not. However, the known specificity of SECIS elements—key components in selenoprotein synthesis—is not very high, and the SECIS elements of a small number of known selenoprotein genes do not possess these characteristics. Therefore, the bSECISearch algorithm, developed based on SECIS elements, has a high false positive rate and produces some false negatives, affecting the final efficiency of bacterial selenoprotein gene identification. Furthermore, this technique is time-consuming, and although online analysis tools exist for some steps, they are not complete analytical procedures, resulting in only intermediate results with a high false positive rate. Algorithms that do not rely on SECIS elements are too time-consuming, computationally expensive, and do not offer convenient online analysis tools.

[0006] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method and terminal for identifying bacterial selenoprotein genes based on deep learning, which aims to solve the problem that the existing technology cannot quickly and effectively identify selenoproteins of various sequenced bacteria.

[0008] The technical solution of the present invention is as follows:

[0009] A deep learning-based method for identifying bacterial selenoprotein genes includes the following steps:

[0010] Obtain known Sec-TGA codons and corresponding selenoprotein-encoding genes as positive sample datasets;

[0011] Obtain the stop-TGA codons of species that do not use selenoproteins as a negative sample dataset;

[0012] The positive sample dataset and the negative sample dataset are segmented, vectorized, and then input into the deep learning model in matrix form for training and debugging to obtain the bacterial selenoprotein gene recognition model.

[0013] The TGA codon and downstream base sequence are extracted from the genome of the bacterial species to be tested and input into the bacterial selenoprotein gene recognition model, and the selenoprotein gene recognition result of the bacterial species to be tested is output.

[0014] The aforementioned deep learning-based bacterial selenoprotein gene identification method, wherein the step of obtaining known Sec-TGA codons and corresponding selenoprotein-encoding genes as a positive sample dataset includes:

[0015] Obtain known selenoprotein families and representative sequences of selenoprotein families, and use the representative sequences of the selenoprotein families as seed sequences;

[0016] Obtain sequenced bacterial genome data, and use the homology search toolkit BLAST to perform bacterial genome similarity search based on the seed sequence to obtain similar fragments from the genomes of other species that are similar to the representative sequences of different selenoproteins;

[0017] Each of the similar segments is analyzed, and the Sec in the seed sequence is selected to correspond to the TGA codon in the similar segment. The analysis is then performed to determine whether each similar segment containing the TGA codon has an open reading frame sequence.

[0018] By retaining the TGA codons and open reading frame sequences containing the open reading frame sequences, and removing redundant sequences, a positive sample dataset of known Sec-TGA codons and corresponding selenoprotein-coding genes is obtained.

[0019] The bacterial selenoprotein gene identification method based on deep learning further includes, after obtaining a positive sample dataset of known Sec-TGA codons and corresponding selenoprotein encoding genes, optimizing the positive sample dataset.

[0020] The optimization process includes the following steps: extracting genomic sequences with downstream lengths of 100, 200, and 300 bases for each of the Sec-TGA codons, and removing redundant sequences to obtain the optimized positive sample dataset.

[0021] The aforementioned deep learning-based bacterial selenoprotein gene identification method, wherein the species that do not use selenoproteins include those that do not contain the selenoprotein synthesis pathway gene SelA, SelB, or SelD.

[0022] The bacterial selenoprotein gene identification method based on deep learning further includes, after obtaining the stop-TGA codons of species that do not use selenoproteins as a negative sample dataset, optimizing the negative sample dataset.

[0023] The optimization process includes the following steps: extracting genomic sequences downstream of each stop-TGA codon with lengths of 100, 200, and 300 bases, removing redundant sequences, and obtaining sequence data;

[0024] The sequence data was clustered using the MMseqs2 tool to obtain an optimized negative sample dataset.

[0025] The bacterial selenoprotein gene identification method based on deep learning, wherein after clustering the sequence data using the MMseqs2 tool, sequence data with similarity less than 50% are retained.

[0026] The aforementioned deep learning-based bacterial selenoprotein gene identification method, wherein the deep learning model is a BERT model based on the Transformer architecture;

[0027] The positive sample dataset includes 90% of the sequences of the selenoprotein family as the positive sample training set, 50% of the sequences of the selenoprotein family as the positive sample validation dataset, and the remainder as the positive sample test dataset.

[0028] The negative sample dataset randomly selects an equal number of sequences based on the specific number of sequences in the positive sample dataset, which are respectively the negative sample training dataset, the negative sample validation dataset, and the negative sample test dataset.

[0029] The aforementioned deep learning-based bacterial selenoprotein gene identification method, after extracting the TGA codon and downstream base sequence from the genome of the bacterial species to be tested and inputting them into the bacterial selenoprotein gene identification model, and before outputting the selenoprotein gene identification results of the bacterial species to be tested, further includes a homology search, comprising the following steps:

[0030] The open reading frame sequences of the Sec-TGA codons obtained after inputting the bacterial selenoprotein gene recognition model were subjected to homology search in a bacterial protein database, and the results were compared and analyzed.

[0031] The open reading frame sequence satisfies the condition that at least one known Sec-containing homologous protein exists, or that at least two of the first ten homologous proteins contain a Sec-TGA-cysteine ​​pair. The open reading frame sequence is then output as the selenoprotein gene recognition result of the bacterial species to be tested.

[0032] A terminal includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the deep learning-based bacterial selenoprotein gene recognition method.

[0033] A computer-readable storage medium storing a computer program thereon, the computer-readable storage medium storing a deep learning-based bacterial selenoprotein gene recognition program, the deep learning-based bacterial selenoprotein gene recognition program, when executed by a processor, implements the steps of the deep learning-based bacterial selenoprotein gene recognition method.

[0034] Beneficial Effects: This invention provides a method and terminal for identifying bacterial selenoprotein genes based on deep learning. The method includes the following steps: obtaining known Sec-TGA codons and corresponding selenoprotein-encoding genes as a positive sample dataset; obtaining stop-TGA codons from species that do not use selenoproteins as a negative sample dataset; segmenting the positive and negative sample datasets, vectorizing them, and inputting them into a deep learning model in matrix form for training and debugging to obtain a bacterial selenoprotein gene identification model; extracting TGA codons and downstream base sequences from the genome of the bacterial species to be tested, inputting them into the bacterial selenoprotein gene identification model, and outputting the selenoprotein gene identification results of the bacterial species to be tested. This invention collects a large amount of gene sequence information from various known bacterial selenoprotein families, as well as stop-TGA codons from species that do not use selenoproteins. A deep learning model is then used to train and debug this model. By deeply mining the sequence and structural features downstream of the Sec-TGA codon, combined with parameter optimization and performance evaluation, a bacterial selenoprotein gene recognition model capable of accurately identifying Sec-TGA codons in bacterial genomes is established. Then, the TGA codons and downstream base sequences from the genome of the bacterial species to be tested are input into this model to obtain the selenoprotein gene recognition results. This recognition method significantly improves the speed and accuracy of selenoprotein gene recognition in bacterial genomes. The analysis time for each genome can be completed within 10 minutes, significantly faster than existing recognition methods, greatly reducing labor and time costs, and lowering analysis costs and risks. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a deep learning-based bacterial selenoprotein gene recognition method of the present invention.

[0036] Figure 2 This is a schematic diagram illustrating the principles of deep learning based on the Transformer architecture.

[0037] Figure 3 This is a schematic diagram of the complete process of a deep learning-based bacterial selenoprotein gene recognition method.

[0038] Figure 4 This is a schematic diagram of the homepage and result output page of the web version tool deep-Sep in Example 2. Detailed Implementation

[0039] This invention provides a method and terminal for bacterial selenoprotein gene recognition based on deep learning. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0041] Although bioinformatics has been increasingly applied across various fields of life sciences over the past decade, research in selenium-related bioinformatics has lagged significantly, particularly in the field of bacteria. Currently, there are very few analytical methods and tools available for identifying bacterial selenoprotein genes.

[0042] Existing methods for predicting selenoprotein genes fall into two categories:

[0043] 1) Algorithms relying on SECIS elements: The bSECISearch algorithm is currently the only widely used algorithm for finding selenoprotein-coding genes in bacterial genomes. It mainly utilizes some artificially defined conserved features of bacterial SECIS elements to predict selenoprotein genes. Although this algorithm has helped researchers discover some new selenoprotein families, the predictive results have a high false positive rate and long running time (e.g., analyzing a common bacterial genome takes at least half a day to a day) because these known features are relatively simple and lack specificity. Furthermore, some known selenoproteins cannot be identified by this algorithm.

[0044] 2) Algorithms independent of SECIS elements: These algorithms primarily utilize the characteristic that "almost all selenoproteins contain many homologous proteins containing cysteine ​​(Cys)." By collecting all Cys-containing sequences from protein databases (such as NCBI's nr database) and combining this with a homology search nucleic acid database, they search for possible Cys / TGA pairings and further analyze them to confirm possible selenoprotein genes in the species' genome. These methods are more time-consuming and require significant computational resources (such as computer clusters), and do not offer online analysis tools. Therefore, these existing techniques are far from meeting the massive data volume generated by the rapid development of current genome sequencing technology and cannot quickly and effectively identify the presence of selenoproteins (selenoproteome) in various sequenced bacteria.

[0045] Clearly, the bSECISearch algorithm, developed based on the SECIS element, has a high false positive rate and produces some false negative results, affecting the final efficiency of bacterial selenoprotein gene identification, and its running time is long. Algorithms that do not rely on the SECIS element are too time-consuming, consume too much computational resources, and cannot provide convenient online analysis tools.

[0046] Based on this, such as Figure 1 As shown, this invention provides a method for identifying bacterial selenoprotein genes based on deep learning, comprising the following steps:

[0047] Step S10: Obtain known Sec-TGA codons and corresponding selenoprotein-coding genes as positive sample datasets;

[0048] Step S20: Obtain the stop-TGA codons of species that do not use selenoproteins as a negative sample dataset;

[0049] Step S30: The positive sample dataset and the negative sample dataset are segmented, vectorized, and then input into the deep learning model in matrix form for training and debugging to obtain the bacterial selenoprotein gene recognition model.

[0050] Step S40: Extract the TGA codon and downstream base sequence from the genome of the bacterial species to be tested, and input them into the bacterial selenoprotein gene recognition model to output the selenoprotein gene recognition result of the bacterial species to be tested.

[0051] In this embodiment, a large amount of gene sequence information from various known bacterial selenoprotein families and the stop-TGA codon from species that do not use selenoproteins are collected. A deep learning model is then used to train and debug this model. By deeply mining the sequence and structural features downstream of the Sec-TGA codon, combined with parameter optimization and performance evaluation, a bacterial selenoprotein gene recognition model capable of accurately identifying Sec-TGA codons in bacterial genomes is established. Then, the TGA codon and downstream base sequence from the genome of the bacterial species to be tested are input into this model to obtain the selenoprotein gene recognition results. This recognition method significantly improves the speed and accuracy of selenoprotein gene recognition in bacterial genomes. The analysis time for each genome can be completed within 10 minutes, significantly faster than existing recognition methods, greatly reducing labor and time costs, and lowering analysis costs and risks.

[0052] Specifically, the deep learning-based bacterial selenoprotein gene identification method is superior to existing identification methods in terms of interactivity, amount of data to be analyzed, response speed, selenoprotein identification efficiency, and code maintenance. It can also improve user experience and facilitate software maintenance and management for developers.

[0053] In some embodiments, step S10, the step of obtaining known Sec-TGA codons and corresponding selenoprotein-coding genes as a positive sample dataset, includes:

[0054] Step S11: Obtain known selenoprotein families and representative sequences of selenoprotein families, and use the representative sequences of the selenoprotein families as seed sequences;

[0055] Step S12: Obtain the sequenced bacterial genome data, and use the homology search toolkit BLAST to perform bacterial genome similarity search based on the seed sequence to obtain similar fragments of other species genomes that are similar to the representative sequences of different selenoproteins;

[0056] Step S13: Analyze each of the similar segments, select the Sec in the seed sequence corresponding to the TGA codon in the similar segment, and analyze whether each similar segment containing the TGA codon has an open reading frame sequence;

[0057] Step S14: Retain the TGA codons and open reading frame sequences containing the open reading frame sequences, and remove redundant sequences to obtain a positive sample dataset of known Sec-TGA codons and corresponding selenoprotein-coding genes.

[0058] Specifically, known selenoprotein families (>80 families) and their representative sequences were collected from databases such as NCBI (National Center for Biotechnology Information) and published literature. These representative sequences were used as seed sequences. Simultaneously, sequenced bacterial genome data (over 270,000 species genomes) and the BLAST (Basic Local Alignment Search Tool) homology search toolkit were downloaded from NCBI to perform bacterial genome similarity searches (using tblastn from the BLAST package) to obtain more genome fragments from other species with similarity to different selenoprotein representative sequences (similar fragments). Each similar fragment was then analyzed, selecting Sec (selenocysteine) / TGA pairings and analyzing whether each genome fragment containing that TGA had a reasonable open reading frame. Within the frame (ORF, i.e., a sequence with the potential to encode a protein), retain the TGA codons with reasonable ORFs (i.e., the TGA codons encoding selenocysteine, Sec-TGA) and their ORF sequences, further remove redundant sequences, and obtain a positive sample dataset of known Sec-TGA codons and corresponding selenoprotein-encoding genes.

[0059] In some embodiments, after obtaining the positive sample dataset of known Sec-TGA codons and corresponding selenoprotein-coding genes in step S14, the method further includes: optimizing the positive sample dataset; the optimization process includes the steps of: extracting genomic sequences downstream of each Sec-TGA codon with lengths of 100, 200, and 300 bases respectively, and removing redundant sequences to obtain the optimized positive sample dataset. By optimizing the positive sample dataset, the accuracy of the trained bacterial selenoprotein gene recognition model in identifying the target bacterial species can be improved.

[0060] In some embodiments, the selenoprotein-free species include those that do not contain the selenoprotein synthesis pathway gene SelA (selenocysteine ​​synthase), the selenoprotein synthesis pathway gene SelB (selenocysteine-specific elongation factor), or the selenoprotein synthesis pathway gene SelD (selenophosphate synthase).

[0061] In some embodiments, step S20, which involves obtaining stop-TGA codons from species that do not use selenoproteins as a negative sample dataset, includes: randomly selecting representative species from different bacterial branches from species that do not use selenoproteins (i.e., species that do not contain the selenoprotein synthesis pathway gene SelA, SelB, or SelD), extracting gene sequences from all annotated genes of these species that use TGA (i.e., TGA codons as termination signals, stop-TGA) as stop codons, and using stop-TGA codons as a negative sample dataset.

[0062] In some implementations, after obtaining stop-TGA codons from species that do not use selenoproteins as a negative sample dataset, the method further includes: optimizing the negative sample dataset; the optimization process includes the steps of: extracting genomic sequences downstream of each stop-TGA codon with lengths of 100, 200, and 300 bases respectively, removing redundant sequences to obtain sequence data; and using the MMseqs2 tool to perform clustering processing on the sequence data to obtain the optimized negative sample dataset. By optimizing the negative sample dataset, the accuracy of the trained bacterial selenoprotein gene recognition model in identifying the target bacterial species can be improved.

[0063] In some implementations, after clustering the sequence data using the MMseqs2 tool, sequence data with similarity less than 50% are retained in order to obtain the downstream sequence features of the stop-TGA codons of different gene families as comprehensively as possible.

[0064] In some implementations, the segmentation process includes: segmenting both the positive sample dataset and the negative sample dataset using a sliding window of 3 bases per unit (i.e., 3-mer), so that after quantization, the data can be input into the deep learning model in matrix form for training and debugging.

[0065] In some implementations, the deep learning model is a BERT (Bidirectional Encoder Representations from Transformers) model based on the Transformer architecture. The deep learning principle based on the Transformer architecture is as follows: Figure 2 As shown, by deeply mining the sequence / structural features downstream of the Sec-TGA codon, and combining parameter optimization and performance evaluation, an optimal deep learning model that can accurately identify Sec-TGA codons in the bacterial genome is established, namely the bacterial selenoprotein gene recognition model.

[0066] In some implementations, the positive sample dataset includes 90% of the sequences of the selenoprotein family as a positive sample training set, 50% of the sequences of the selenoprotein family as a positive sample validation dataset, and the remainder as a positive sample test dataset.

[0067] Specifically, the deep learning model randomly selected 90% of the sequences from 40 selenoprotein families in the positive sample dataset as the positive sample training set, and then randomly selected 50% of the sequences from 30 selenoprotein families as the positive sample validation dataset. Finally, all the remaining families and sequences were used as the positive sample test dataset.

[0068] In some implementations, the negative sample dataset randomly selects an equal number of sequences based on the specific number of sequences in the positive sample dataset, which are respectively the negative sample training dataset, the negative sample validation dataset, and the negative sample test dataset.

[0069] Specifically, in step S30, the training and debugging process includes pre-training and fine-tuning; the training and debugging steps include:

[0070] 1) Establish various Sec-TGA models and conduct performance evaluation: For TGA downstream base sequences under different length conditions, firstly, the deep learning model is pre-trained on the training set corresponding to that length without classification labels for a total of 3 rounds of training, aiming to allow the model to learn extensively and mine the potential features of Sec-TGA codon downstream sequences.

[0071] After pre-training, the model enters the fine-tuning stage, using a training set with classification labels (including positive and negative sample training sets) to train the model effectively to distinguish between Sec-TGA and stop-TGA codons. The model is trained for a total of 10 rounds. After each round of training, the model is validated and its performance is evaluated on the validation dataset (including positive and negative sample validation datasets). This process continuously optimizes the model parameters, and finally, the deep learning model that performs best on the validation dataset is retained for each length condition. That is, an optimal model is obtained for each of the following conditions: the TGA downstream sequence lengths of 100, 200, and 300 bases.

[0072] 2) Establishment of the optimal Sec-TGA codon recognition model (bacterial selenoprotein gene recognition model): The optimal model under different downstream sequence lengths in step 1) was selected, and its performance was tested using independent test datasets (including positive and negative sample test datasets). Multiple evaluation metrics were used to assess model performance; these metrics included the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), accuracy, precision, recall, and F1 score. Through testing, it was confirmed that the sequence with a downstream sequence length of 200 digits showed the best overall performance and could therefore be considered the optimal Sec-TGA codon recognition model.

[0073] Furthermore, for any bacterial species genome, after extracting all possible TGA codons and the downstream 200 base sequences of the genome, the initial predicted set of Sec-TGA codons and corresponding selenoprotein ORF sequences of the species can be obtained by using the optimal recognition model obtained in step 2) above.

[0074] In some embodiments, after extracting the TGA codon and downstream base sequence from the genome of the bacterial species to be tested and inputting them into the bacterial selenoprotein gene recognition model, before outputting the selenoprotein gene recognition result of the bacterial species to be tested, a homology search is also included, comprising the steps of: performing a homology search on the open reading frame sequence of the Sec-TGA codon obtained after inputting it into the bacterial selenoprotein gene recognition model in a bacterial protein database, and comparing and analyzing the results; the open reading frame sequence satisfies the condition that there is at least one known Sec-containing homologous protein, or that at least two of the top ten homologous proteins contain a Sec-TGA-cysteine ​​pairing, and the open reading frame sequence is output as the selenoprotein gene recognition result of the bacterial species to be tested.

[0075] Specifically, this process utilizes homology searching to further eliminate false positives in the initial prediction set. Using the protein homology search tool DIAMOND, homology sequences are searched in bacterial protein databases for each ORF containing the Sec-TGA codon. The results are compared and analyzed to check the conservation of the sequences flanking each possible Sec-TGA. The ORF sequence is retained if the following criteria are met: (1) at least one known homologous protein containing Sec (selenocysteine) exists, or (2) at least two of the top ten homologous proteins in the results contain a pairing of Sec-TGA and Cys (cysteine). By utilizing homology searching, the accuracy of selenoprotein gene identification results for the tested bacterial species is further improved, effectively eliminating false positives.

[0076] In some implementations, the deep learning-based bacterial selenoprotein gene identification method can identify all Sec-TGA codons and corresponding selenoprotein gene sets of the test substance, and can classify them into known selenoprotein families and new selenoprotein families based on whether they have been reported, thereby realizing the identification of species selenoprotein genes and the discovery of new selenoprotein families.

[0077] Specifically, the complete process of the deep learning-based bacterial selenoprotein gene recognition method is as follows: Figure 3 As shown, all TGA codons were extracted from the bacterial genome, and then the downstream base sequences of TGA were collected and input into the optimal Sec-TGA codon recognition model. The results were used for microbial genome homology search, and then cross-species conservation analysis was performed on the sequences surrounding TGA codons to finally obtain the selenoprotein gene set. The optimal Sec-TGA codon recognition model was first pre-trained, including the following steps: using the downstream base sequences of Sec-TGA as the positive sample set and the downstream base sequences of stop-TGA as the negative sample set, performing sequence segmentation and vectorization (3-mers) on the positive and negative sample sets, and then training and debugging the BERT deep learning model based on the Transformer architecture to establish various Sec-TGA models and conduct performance evaluation to obtain the optimal Sec-TGA codon recognition model.

[0078] In addition, the present invention also provides a terminal, the terminal including a memory, a processor and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the deep learning-based bacterial selenoprotein gene recognition method.

[0079] In some embodiments, the memory may be an internal storage unit of the terminal, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the terminal, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Further, the memory may include both internal and external storage units. The memory is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory stores a deep learning-based bacterial selenoprotein gene recognition program, which can be executed by a processor to realize the deep learning-based bacterial selenoprotein gene recognition method of the present invention.

[0080] In some embodiments, the processor may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory or process data, such as executing the deep learning-based bacterial selenoprotein gene recognition method.

[0081] In addition, the present invention also provides a computer-readable storage medium storing a computer program thereon, the computer-readable storage medium storing a deep learning-based bacterial selenoprotein gene recognition program, the deep learning-based bacterial selenoprotein gene recognition program being executed by a processor to implement the steps of the deep learning-based bacterial selenoprotein gene recognition method.

[0082] In some implementations, to make this method more convenient for researchers, an online prediction tool for bacterial selenoprotein genes has been established based on a deep learning-based bacterial selenoprotein gene identification method. Users only need to provide the bacterial genome sequence of interest, and they can obtain all relevant information about the selenoproteins that may be contained therein (including complete gene sequences and protein sequences) within an acceptable time (online analysis of a 1 million base sequence takes about 1-2 minutes), thus realizing the identification of bacterial selenoprotein genes.

[0083] In some implementations, the deep learning-based bacterial selenoprotein gene recognition method can also be used in the field of other element-binding protein recognition, such as iron, zinc, and copper. This provides a good research idea and technical template for the application of deep learning in the field of trace element utilization and function, and helps to promote the development of related industries.

[0084] In some implementations, deep learning models such as convolutional neural networks, generative adversarial networks, and recurrent neural networks can be used to train bacterial selenoprotein gene recognition models and apply them to bacterial selenoprotein gene recognition.

[0085] The following examples further illustrate the present invention in detail. It should also be understood that the following examples are only for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-essential improvements and adjustments made by those skilled in the art based on the above description of the present invention are within the scope of protection of the present invention.

[0086] Example 1

[0087] This embodiment provides a local version of a deep learning-based bacterial selenoprotein gene identification method for terminal devices to identify selenoprotein genes in the bacterial species to be analyzed, including the following steps:

[0088] 1) Download the bacterial genome sequence to be analyzed from the NCBI website (e.g., Clostridium cadaverisstrain IFB3C5, genome size 3,619,347 bp), the sequence file format is FASTA;

[0089] 2) Extract all TGA codons and their downstream 200 base pairs that may contain a valid ORF from the original genome sequence file;

[0090] 3) Based on the optimal deep learning model of Sec-TGA codons obtained through training, and after model screening, the initial prediction set of Sec-TGA codons and corresponding selenoprotein ORF sequences of this species is obtained (output in text file format).

[0091] 4) Based on the above results, the Diamond homology search software was used to conduct homology sequence search and alignment analysis of the protein database of representative bacterial species. If either of the following two rules is met, namely (1) there is at least one known homologous protein containing Sec (selenocysteine); (2) among the top ten homologous proteins in the search results, at least two homologous proteins have a pairing of Sec-TGA and Cys (cysteine), then the predicted selenoprotein gene sequence can be preserved;

[0092] 5) Obtain the final Sec-TGA codons and corresponding selenoprotein gene sequence set for this species, and output the results in text format (the Clostridium cadaveris strain IFB3C5 genome contains 10 known selenoprotein genes and 2 new selenoprotein genes, the number of known selenoproteins being consistent with the previous analysis results of searching the genome of this species using all known selenoprotein families).

[0093] The entire process described above takes approximately 9 minutes.

[0094] Example 2

[0095] This embodiment provides a network-based deep learning-based bacterial selenoprotein gene identification method for terminal devices, which identifies selenoprotein genes in the bacterial species to be analyzed, including the following steps:

[0096] 1) Go to the homepage of the web version tool deep-Sep, and enter the bacterial genome sequence to be analyzed (e.g., Syntrophobacter fumaroxidans MPOB) in the input sequence box; Considering the timely and real-time feedback of prediction results, the input sequence length of the current web version tool is temporarily limited to within 1 million base pairs (if you want to analyze a complete bacterial genome, you can pre-divide it into several parts and input them separately).

[0097] 2) Click “Submit” and the background program will start running (the specific analysis process is the same as in Example 1);

[0098] 3) The results page lists all selenoprotein gene sequences and corresponding protein sequences contained in the input genome sequence, and also provides the specific location of each selenoprotein gene in the genome sequence (in this example, the first 1 million base sequences of the species Desulfosarcina ovata subsp. ovata strain oXyS1 were input, and a total of 4 known selenoprotein genes were found, which is consistent with the previous analysis results of searching this part of the genome sequence using all known selenoprotein families).

[0099] The above process takes approximately 1 minute and 30 seconds to complete. The homepage and results output page of the web-based tool deep-Sep are shown below. Figure 4 As shown.

[0100] In summary, this invention provides a deep learning-based bacterial selenoprotein gene identification method and terminal. The deep learning-based bacterial selenoprotein gene identification method includes the following steps: obtaining known Sec-TGA codons and corresponding selenoprotein-encoding genes as a positive sample dataset; obtaining stop-TGA codons from species that do not use selenoproteins as a negative sample dataset; segmenting the positive and negative sample datasets, vectorizing them, and inputting them into a deep learning model in matrix form for training and debugging to obtain a bacterial selenoprotein gene identification model; extracting TGA codons and downstream base sequences from the genome of the bacterial species to be tested, inputting them into the bacterial selenoprotein gene identification model, and outputting the selenoprotein gene identification result of the bacterial species to be tested. This invention collects a large amount of gene sequence information from various known bacterial selenoprotein families, as well as stop-TGA codons from species that do not use selenoproteins. A deep learning model is then used to train and debug this model. By deeply mining the sequence and structural features downstream of the Sec-TGA codon, combined with parameter optimization and performance evaluation, a bacterial selenoprotein gene recognition model capable of accurately identifying Sec-TGA codons in bacterial genomes is established. Then, the TGA codons and downstream base sequences from the genome of the bacterial species to be tested are input into this model to obtain the selenoprotein gene recognition results. This recognition method significantly improves the speed and accuracy of selenoprotein gene recognition in bacterial genomes. The analysis time for each genome can be completed within 10 minutes, significantly faster than existing recognition methods, greatly reducing labor and time costs, and lowering analysis costs and risks.

[0101] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for identifying bacterial selenoprotein genes based on deep learning, characterized in that, Including the following steps: Obtain known Sec-TGA codons and corresponding selenoprotein-encoding genes as positive sample datasets; Obtain the stop-TGA codons of species that do not use selenoproteins as a negative sample dataset; The positive sample dataset and the negative sample dataset are segmented, vectorized, and then input into the deep learning model in matrix form for training and debugging to obtain the bacterial selenoprotein gene recognition model. The TGA codon and downstream base sequence are extracted from the genome of the bacterial species to be tested and input into the bacterial selenoprotein gene recognition model, and the selenoprotein gene recognition results of the bacterial species to be tested are output. The step of obtaining known Sec-TGA codons and corresponding selenoprotein-encoding genes as positive sample datasets includes: Obtain known selenoprotein families and representative sequences of selenoprotein families, and use the representative sequences of the selenoprotein families as seed sequences; Obtain sequenced bacterial genome data, and use the homology search toolkit BLAST to perform bacterial genome similarity search based on the seed sequence to obtain similar fragments from the genomes of other species that are similar to the representative sequences of different selenoproteins; Each of the similar segments is analyzed, and the Sec in the seed sequence is selected to correspond to the TGA codon in the similar segment. The analysis is then performed to determine whether each similar segment containing the TGA codon has an open reading frame sequence. Retain the TGA codons and open reading frame sequences containing the open reading frame sequences, and after removing redundant sequences, obtain a positive sample dataset of known Sec-TGA codons and corresponding selenoprotein-encoding genes; The deep learning model is the BERT model based on the Transformer architecture; The positive sample dataset includes 90% of the sequences of the selenoprotein family as the positive sample training set, 50% of the sequences of the selenoprotein family as the positive sample validation dataset, and the remainder as the positive sample test dataset. The negative sample dataset randomly selects an equal number of sequences based on the specific number of sequences in the positive sample dataset, which are respectively the negative sample training set, the negative sample validation dataset, and the negative sample test dataset; After extracting the TGA codon and downstream base sequence from the genome of the bacterial species to be tested and inputting them into the bacterial selenoprotein gene recognition model, before outputting the selenoprotein gene recognition results of the bacterial species to be tested, homology search is also included, including the following steps: The open reading frame sequences of the Sec-TGA codons obtained after inputting the bacterial selenoprotein gene recognition model were subjected to homology search in the bacterial protein database, and the results were compared and analyzed. The open reading frame sequence satisfies the condition that at least one known Sec-containing homologous protein exists, or that at least two of the first ten homologous proteins contain a Sec-TGA-cysteine ​​pair. The open reading frame sequence is then output as the selenoprotein gene recognition result of the bacterial species to be tested.

2. The method for bacterial selenoprotein gene recognition based on deep learning according to claim 1, characterized in that, After obtaining a positive sample dataset containing known Sec-TGA codons and corresponding selenoprotein-encoding genes, the method further includes: optimizing the positive sample dataset. The optimization process includes the following steps: extracting genomic sequences with downstream lengths of 100, 200, and 300 bases for each of the Sec-TGA codons, and removing redundant sequences to obtain the optimized positive sample dataset.

3. The method for bacterial selenoprotein gene recognition based on deep learning according to claim 1, characterized in that, The species that do not use selenoproteins include those that do not contain the selenoprotein synthesis pathway gene SelA, SelB, or SelD.

4. The method for bacterial selenoprotein gene recognition based on deep learning according to claim 1, characterized in that, After obtaining the stop-TGA codons of species that do not use selenoproteins as a negative sample dataset, the method further includes: optimizing the negative sample dataset; The optimization process includes the following steps: extracting genomic sequences downstream of each stop-TGA codon with lengths of 100, 200, and 300 bases, removing redundant sequences, and obtaining sequence data; The sequence data was clustered using the MMseqs2 tool to obtain an optimized negative sample dataset.

5. The method for bacterial selenoprotein gene recognition based on deep learning according to claim 4, characterized in that, After clustering the sequence data using the MMseqs2 tool, sequence data with a similarity of less than 50% are retained.

6. A terminal, characterized in that, The terminal includes a memory, a processor, and a program stored in the memory and executable on the processor. When executed by the processor, the program implements the steps of the deep learning-based bacterial selenoprotein gene recognition method as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program thereon, and the computer-readable storage medium stores a deep learning-based bacterial selenoprotein gene recognition program, which, when executed by a processor, implements the steps of the deep learning-based bacterial selenoprotein gene recognition method as described in any one of claims 1-5.

Images