A method, system, device and medium for recognizing prokaryotic open reading frames

Abstract

This invention provides a method, system, device, and medium for identifying open reading frames (ORFs) in prokaryotes, belonging to the field of bioinformatics. The method includes: based on the DeepRibo model, in the DNA sequence branch, capturing local features of codons, the RBS core region, and flanking sequences using three convolutional kernels of different sizes, and equipping each convolutional kernel with a batch normalization layer; in the Ribo-seq time-series branch, adding a self-attention module after the Bi-GRU output layer to dynamically calculate the weights of each time step and generate a global context vector, while retaining the hidden state of the last time step; finally, fusing the multi-scale convolutional features, the attention context vector, and the time-series final state features to complete ORF prediction. This invention solves the problems of incomplete local feature capture, loss of long-range dependencies, and single feature fusion in existing technologies, significantly improving the accuracy of prokaryotic ORF identification.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to a method, system, device, and medium for identifying open reading frames in prokaryotes. Background Technology

[0002] Accurate prediction of open reading frames (ORFs) in prokaryotes using deep learning technology is crucial for genome annotation. Genome annotation involves interpreting and labeling the functional information of the entire nucleotide sequence of a prokaryotic genome, and the ORF is the core region encoding proteins in prokaryotes (a continuous nucleotide sequence from the start codon ATG to the stop codons TAA / TAG / TGA, serving as a direct template for protein translation). Accurate prediction of ORFs is a fundamental prerequisite and core step in genome annotation: without accurately identifying the location, boundaries, and number of ORFs, all subsequent annotation work, including gene function prediction, protein structure analysis, and metabolic pathway construction, is impossible.

[0003] Currently, the closest existing technology to this invention is DeepRibo. DeepRibo primarily employs a single-scale convolutional neural network (CNN) to extract DNA sequence features and a bidirectional gated recurrent unit (Bi-GRU) to extract Ribo-seq temporal features, along with direct feature concatenation, to solve the open reading frame (ORF) recognition problem. While this approach has advanced prokaryotic gene annotation to some extent, its low model accuracy remains a problem due to limitations in its underlying network architecture design.

[0004] Traditional models using single-scale convolutional neural networks to extract features from input DNA sequences suffer from a fixed receptive field due to the fixed kernel size. In contrast, real biological sequences exhibit vastly different regulatory signal lengths (e.g., codon triplets are typically 3 nt, while ribosome binding sites (RBS) and their flanking sequences often span even longer distances). This makes it difficult to achieve multi-resolution feature perception and capture both short motifs and long-range regulatory elements within the same network layer, easily leading to the omission of crucial, weak signals. Existing techniques use bidirectional gated recurrent units (Bi-GRU) to process Ribo-seq coverage data, and in the feature output stage, only the hidden state of the GRU network at the last time step is extracted as the representation vector for the entire sequence. Forcibly compressing all temporal dynamics of long sequences into the last hidden state node inevitably leads to severe information bottlenecks and long-range dependency problems when dealing with long candidate ORF sequences. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method, system, device, and medium for identifying open reading frames in prokaryotes.

[0006] To achieve the above objectives, the present invention provides a method for identifying open reading frames in prokaryotes, comprising: The target prokaryotic genome region to be tested is divided into multiple candidate sequence fragments. Each candidate sequence fragment includes a DNA subsequence and a Ribo-seq subsequence of ribosome sequencing coverage time data in the same genomic coordinate range as the DNA subsequence.

[0007] For each candidate sequence fragment: Local coding features of the short-range codons, mid-range RBS core region, and long-range flanking sequences of the DNA subsequence are captured separately. All local coding features are concatenated to obtain a local sequence coding feature fusion vector. The bidirectional dependencies of the Ribo-seq subsequences of the ribosome sequencing coverage time-series data are captured, and the hidden state sequences of all time steps are output. The attention weights of each time step are dynamically calculated, and all hidden state sequences are weighted and summed to generate a global context feature vector capturing the long-distance temporal dependencies of the Ribo-seq. Simultaneously, the hidden state sequence of the last time step in capturing the bidirectional dependencies is retained as the temporal final state feature vector. The local sequence coding feature fusion vector, the global context feature vector, and the temporal final state feature vector are concatenated along the feature dimension to form a high-dimensional fusion feature vector.

[0008] Based on the high-dimensional fusion feature vector, the probability that the candidate sequence fragment is an open reading frame is generated. Based on the probability of all candidate sequence fragments and their corresponding DNA subsequence coordinate intervals, the positions of each open reading frame in the target prokaryotic genome region to be tested are identified.

[0009] Preferably, the candidate sequence fragment is input into the DeepORF model, and the probability that the candidate sequence fragment is an open reading frame is output. The DeepORF model is based on the traditional DeepRibo model, replacing the single-scale fixed convolutional kernel of the DNA sequence branch in the traditional DeepRibo model with a parallel multi-scale convolutional module. A self-attention module is added after the Bi-GRU output layer of the Ribo-seq temporal branch of the traditional DeepRibo model. The parallel multi-scale convolutional module uses three convolutional kernels of different scales set in parallel, and a batch normalization (BN) layer is equipped after each convolutional kernel.

[0010] Preferably, after dividing the candidate sequence fragments, the method further includes preprocessing the Ribo-seq subsequence of the ribosome sequencing coverage time series data, specifically including: using a four-parameter S-curve to filter and reduce noise in the Ribo-seq ribosome sequencing coverage time series data, and then normalizing the noise-reduced Ribo-seq data; capturing the bidirectional dependency relationship of the normalized Ribo-seq data.

[0011] Preferably, after dividing the candidate sequence fragments, the process further includes preprocessing the DNA subsequence, specifically including: performing one-hot encoding on the DNA subsequence to obtain a binary feature tensor; and capturing each local encoded feature of the binary feature tensor.

[0012] Preferably, the DeepORF model is trained using the AdamW optimizer, which decouples the model's weight decay from gradient updates.

[0013] Preferably, based on the probability of all candidate sequence fragments and their corresponding DNA subsequence coordinate intervals, the positions of each open reading frame in the target prokaryotic genome region to be tested are identified. Specifically, this includes: setting a probability threshold, identifying candidate sequence fragments with a probability greater than or equal to the probability threshold as candidate open reading frames, and the DNA subsequence coordinate interval of each candidate open reading frame being the candidate position of that open reading frame; if the coordinate intervals of multiple candidate open reading frames are adjacent or overlap, they are merged to determine the precise start and end positions of an open reading frame.

[0014] This invention also provides a system for recognizing open reading frames in prokaryotes, comprising: The data acquisition module is used to divide the target prokaryotic genome region to be tested into multiple candidate sequence fragments. Each candidate sequence fragment includes a DNA subsequence and a Ribo-seq subsequence of ribosome sequencing coverage time series data in the same genomic coordinate range as the DNA subsequence.

[0015] The feature extraction module is used to: capture the local coding features of short-range codons, mid-range RBS core regions, and long-range flanking sequences of the DNA subsequence for each candidate sequence fragment; concatenate all local coding features to obtain a local sequence coding feature fusion vector; capture the bidirectional dependencies of the Ribo-seq subsequences of the ribosome sequencing coverage time series data; output the hidden state sequences of all time steps; dynamically calculate the attention weights of each time step; and sum all hidden state sequences by weight to generate a global context feature vector that captures the long-distance temporal dependencies of the Ribo-seq; simultaneously, retain the hidden state sequence of the last time step in the process of capturing bidirectional dependencies as the temporal final state feature vector; and concatenate the local sequence coding feature fusion vector, the global context feature vector, and the temporal final state feature vector in the feature dimension to form a high-dimensional fusion feature vector.

[0016] The identification module is used to generate the probability that the candidate sequence fragment is an open reading frame based on the high-dimensional fusion feature vector, and to identify the position of each open reading frame in the target prokaryotic genome region based on the probability of all candidate sequence fragments and their corresponding DNA subsequence coordinate intervals.

[0017] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in the prokaryotic open reading frame recognition method.

[0018] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, is capable of executing any of the steps in the prokaryotic open reading frame recognition method.

[0019] The method for identifying open reading frames in prokaryotes provided by this invention has the following beneficial effects: This invention captures local coding features of short-range codons, mid-range RBS core regions, and long-range flanking sequences of DNA subsequences in parallel, and then splices and fuses them to achieve multi-resolution feature perception, taking into account the capture of regulatory signals of different lengths and avoiding the omission of local key weak signals. For Ribo-seq time series data processing, instead of only extracting the hidden state of the last time step, it first captures the bidirectional dependencies of the data and outputs the hidden state sequences of all time steps, then calculates the weights and sums them to generate a global context feature vector, while retaining the hidden state of the last time step as the time series final state feature vector. This not only mines the long-distance time series dependency information of long sequences, but also retains the key features of the time series final state, breaking the information bottleneck of single final state features and effectively solving the long-range dependency problem. Finally, by splicing and fusing the three-way features, the feature information of both types of data is fully utilized.

[0020] This invention independently predicts each candidate DNA subsequence and its corresponding Ribo-seq data, generating the coding probability of each subsequence. Finally, based on the probability distribution and positional information of all subsequences, it achieves accurate identification of open reading frames in the complete genome. It realizes multi-resolution feature perception of DNA sequences, simultaneously capturing biological regulatory signals of different lengths, avoiding the omission of key weak features, and improving the comprehensiveness and accuracy of DNA sequence feature extraction. It overcomes the information bottleneck and long-range dependency problem in Ribo-seq long sequence processing, mining global temporal dependency information while preserving temporal final-state features, fully capturing the temporal features of Ribo-seq data; and it improves the model's recognition capability. Attached Figure Description

[0021] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a method for identifying open reading frames in prokaryotes according to an embodiment of the present invention; Figure 2 This is a diagram of the DeepORF model architecture according to an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0024] This invention proposes a neural network for predicting open reading boxes (DeepORF) based on multi-scale convolution and attention mechanisms, building upon the baseline model DeepRibo. Figure 2 As shown, the DeepORF model is based on the traditional DeepRibo model, replacing the single-scale fixed convolution kernel of the DNA sequence branch in the traditional DeepRibo model with a parallel multi-scale convolution module; after the Bi-GRU output layer of the Ribo-seq temporal branch of the traditional DeepRibo model, a self-attention module is added; the parallel multi-scale convolution module uses three convolution kernels of different scales set in parallel, and a batch normalized (BN) layer is equipped after each convolution kernel.

[0025] Based on this, the present invention provides a method for recognizing open reading frames in prokaryotes, specifically a neural network method for predicting open reading frames based on multi-scale convolution and attention mechanisms, such as... Figure 1 As shown, it includes the following steps: S1. Divide the target prokaryotic genome region into multiple candidate sequence fragments. Each candidate sequence fragment includes a DNA subsequence and a Ribo-seq subsequence of ribosome sequencing coverage time data in the same genomic coordinate range as the DNA subsequence.

[0026] The DeepORF model takes as input data the complete gene sequence of a prokaryote. The complete gene sequence is divided into equal-length candidate sequence fragments using a fixed sliding window. Each candidate sequence fragment includes a DNA subsequence and a Ribo-seq (ribosomal imprinting) subsequence representing the ribosomal sequencing coverage within the same genomic coordinate region. The complete gene sequence includes multiple candidate ORF fragments. One-hot encoding is performed on the input candidate sequence fragments (e.g., ACGCTCG...TCCCAGA) to obtain a binary feature tensor. To address the high noise issue in Ribo-seq sequencing data, a four-parameter S-curve method is introduced to filter and normalize the data, removing noise and establishing a standard training dataset.

[0027] S2. For each candidate sequence fragment: capture the local coding features of the short-range codons, mid-range RBS core region, and long-range flanking sequences of the DNA subsequence, and concatenate all local coding features to obtain a local sequence coding feature fusion vector; capture the bidirectional dependencies of Ribo-seq subsequences in ribosome sequencing coverage time-series data, output the hidden state sequences of all time steps, dynamically calculate the attention weights of each time step, and sum all hidden state sequences by weight to generate a global context feature vector that captures long-distance temporal dependencies in Ribo-seq; at the same time, retain the hidden state sequence of the last time step in the process of capturing bidirectional dependencies as the temporal final state feature vector; concatenate the local sequence coding feature fusion vector, the global context feature vector, and the temporal final state feature vector in the feature dimension to form a high-dimensional fusion feature vector.

[0028] One-hot encoded DNA sequences are input into a parallel multi-scale convolution module. The model performs prediction processing on each sub-sequence to ultimately confirm the presence of ORF fragments. The parallel multi-scale convolution module contains three parallel convolution kernels with sizes of 3, 7, and 15, respectively used to capture local coding features of short-range codons, the mid-range RBS core region, and the long-range flanking sequences. Each convolution kernel is followed by a batch normalization layer (BN) to align the distribution of features at different scales, accelerate convergence, and prevent feature masking. The outputs of the three branches are concatenated along the channel dimension to obtain a fused vector of local sequence coding features.

[0029] A complete genome sequence contains both coding and non-coding genes. Due to the lack of an attention allocation mechanism, the model cannot dynamically focus on the truly biologically significant regions (such as the translation initiation site TIS or ribosome pause sites) in the Ribo-seq signal during processing. This results in the effective signal being diluted by background noise from a large number of non-coding regions, reducing the accuracy of gene boundary recognition. This invention designs multi-scale convolutions, enabling the network to target the same subsequence region of the input at the same level, simultaneously capturing short motifs (such as 3nt codon triplets), medium-length signals (such as the 7nt prokaryotic ribosome binding site RBS core), and longer-range regulatory elements (such as the 15nt RBS and its flanking spacers), achieving multi-resolution feature perception of the DNA sequence.

[0030] The preprocessed Ribo-seq data is input into a Bidirectional Gated Recurrent Unit (Bi-GRU), which captures the temporal bidirectional dependencies through a two-layer bidirectional network, outputting a sequence of hidden states at each time step. This sequence is then input into a self-attention module, where the attention weights for each time step are dynamically calculated, and all hidden states are weighted and summed to generate a global context feature vector that captures the long-distance temporal dependencies of Ribo-seq. Simultaneously, the hidden state of the last time step of the Bi-GRU is retained as the temporal final state feature vector.

[0031] This invention uses a two-layer bidirectional gated recurrent unit (Bi-GRU, with a hidden size of 128) to process sequence temporal information, effectively capturing the bidirectional temporal dependencies of ribosome movement on mRNA. A self-attention module is added above the Bi-GRU output layer. This mechanism dynamically calculates the attention weights at each time step, generating a weighted summation context vector. This allows the model to automatically focus on key biological regions such as the upstream regulatory region of the TIS (Translation Initiation Site), suppressing background noise in non-coding regions, solving long-range dependencies, and automatically assigning higher weights to significant biological regions, effectively resisting interference from background noise in non-coding regions and resolving long-range dependencies.

[0032] The three feature vectors are concatenated along their respective feature dimensions to form a high-dimensional fused feature vector. This vector integrates multi-scale local information from the DNA sequence, key region focusing information from Ribo-seq, and global temporal information, constructing a more complete and complementary data view. This solves the problems of single feature extraction, loss of long-range dependencies, and simplistic fusion methods in traditional DeepRibo models.

[0033] S3. Based on the high-dimensional fusion feature vector, generate the probability that the candidate sequence fragment is an open reading frame. Based on the probability of all candidate sequence fragments and their corresponding DNA subsequence coordinate intervals, identify the position of each open reading frame in the target prokaryotic genome region to be tested.

[0034] Fully connected network and final prediction output. The fused feature vector is non-linearly mapped and dimensionality reduced through a multi-layer fully connected network (dimension changes sequentially: 1500 -> 1024 -> 512 -> 2). Finally, the softmax activation function is used to output the ORF sequence (e.g., ATGTCC…GCGTAG) and its corresponding encoding probability score (Predicted Score, range 0-1). The high-dimensional fused feature vector is then input into the multi-layer fully connected network for non-linear mapping and dimensionality reduction. The softmax function outputs the probability of the existence of open reading frames (OPFs) in the target prokaryotic genome region. Based on this probability, the ORFs are identified. If the coordinate intervals of multiple candidate ORFs are adjacent or overlap, they are merged to determine the precise start and end positions of a single ORF.

[0035] To address the shortcomings of existing technologies (DeepRibo) in sequence feature extraction, temporal feature compression, and multimodal fusion, this invention employs a parallel multi-scale convolutional neural network (kernel sizes of 3, 7, and 15), which can simultaneously capture features of the codon triplet (3nt), the prokaryotic ribosome binding site RBS core (7nt), and the region covering the RBS and its flanking space (15nt), achieving multi-resolution capture of biological sequence features. Hierarchical batch normalization is employed, equipping each CNN branch at multiple scales with an independent BN layer. This forces alignment of feature distributions at different scales, preventing strong features such as long convolutional kernel responses from masking other weak features during fusion, thus significantly accelerating model convergence. This invention introduces a self-attention mechanism above the GRU output layer. This mechanism dynamically calculates attention weights and generates context vectors, enabling the model to automatically focus on biologically significant regions in the Ribo-seq data (such as the translation initiation site TIS and key sequences in the coding region), effectively solving the problems of information redundancy and background noise interference in non-coding regions during data fusion. This invention provides a three-way multi-view fusion that combines local sequence features, global attention features, and temporal final state features, offering the model a more complete and complementary data view. It also employs the AdamW optimizer, correctly decoupling weight decay and gradient updates, resulting in better generalization performance compared to the standard Adam.

[0036] The main differences between this invention and existing solutions (DeepRibo) are: (1) The scale of sequence feature extraction differs from the receptive field. Existing methods (DeepRibo) use a single-scale convolutional neural network, employing only a fixed-size convolutional kernel, resulting in a fixed receptive field that makes it difficult to accommodate regulatory signals of different lengths. This invention traverses the complete genome sequence through a sliding window, independently predicting each candidate subsequence and its corresponding Ribo-seq data within the window to generate the coding probability of each subsequence. Finally, based on the probability distribution and location information of all subsequences, it achieves accurate identification of open reading frames in the complete genome. This invention (DeepORF) also innovatively employs a parallel multi-scale convolutional neural network, using three different sizes of convolutional kernels (3nt, 7nt, and 15nt) in parallel, achieving simultaneous capture of multi-resolution biological features such as codons, RBS cores, and their flanking spacers, significantly enhancing perception capabilities.

[0037] (2) The attention focusing mechanism for ribosome mapping (Ribo-seq) feature extraction differs. Existing schemes rely solely on bidirectional gated recurrent units (Bi-GRU) and extract only the hidden state of the last time step as the output, resulting in severe long-range dependency issues. This invention introduces a self-attention module above the GRU output layer. By calculating the attention weights at each time step to generate a context vector, the model can automatically and dynamically focus on key regions with significant biological significance (such as the translation initiation site TIS), thus solving the feature compression bottleneck of long sequences.

[0038] (3) The fusion strategies for multimodal features are different. Existing solutions use dual-path concatenation, i.e., Concat(CNN_Output, GRU_Last_Hidden), which has a single fusion method and is difficult to provide a complete and complementary multimodal data view. Concat(CNN_Output, GRU_Last_Hidden) is used in this invention. The present invention adopts a three-path multi-view fusion strategy, which performs Concat(multi-scale CNN output, attention context vector, GRU terminal hidden state), which fuses local sequence features, global attention features and temporal terminal state features, and provides a more complete and complementary data view.

[0039] (4) Different Normalization Processes. Existing solutions do not include a Batch Normalization (BN) layer. Due to the significant difference in numerical dimensions between DNA sequences and Ribo-seq, model training is prone to oscillations and convergence difficulties. This invention introduces a hierarchical batch normalization mechanism, equipping each scale branch (3, 7, 15) of the multi-scale CNN with an independent BN layer, forcibly aligning the distribution of features at different scales. This effectively prevents strong response features from masking weak features and significantly accelerates model convergence. Existing solutions use the standard Adam optimizer. This invention employs the AdamW optimizer, which correctly decouples weight decay and gradient updates, resulting in better generalization performance compared to the standard Adam.

[0040] Based on the same inventive concept, this invention also provides a system for identifying open reading frames in prokaryotes, comprising: The data acquisition module is used to divide the target prokaryotic genome region to be tested into multiple candidate sequence fragments. Each candidate sequence fragment includes a DNA subsequence and a Ribo-seq subsequence of ribosome sequencing coverage time series data in the same genomic coordinate range as the DNA subsequence.

[0041] The feature extraction module is used to: capture the local coding features of short-range codons, mid-range RBS core regions, and long-range flanking sequences of the DNA subsequence for each candidate sequence fragment; concatenate all local coding features to obtain a local sequence coding feature fusion vector; capture the bidirectional dependencies of the Ribo-seq subsequences of the ribosome sequencing coverage time series data; output the hidden state sequences of all time steps; dynamically calculate the attention weights of each time step; and sum all hidden state sequences by weight to generate a global context feature vector that captures the long-distance temporal dependencies of the Ribo-seq; simultaneously, retain the hidden state sequence of the last time step in the process of capturing bidirectional dependencies as the temporal final state feature vector; and concatenate the local sequence coding feature fusion vector, the global context feature vector, and the temporal final state feature vector in the feature dimension to form a high-dimensional fusion feature vector.

[0042] The identification module is used to generate the probability that the candidate sequence fragment is an open reading frame based on the high-dimensional fusion feature vector, and to identify the position of each open reading frame in the target prokaryotic genome region based on the probability of all candidate sequence fragments and their corresponding DNA subsequence coordinate intervals.

[0043] This invention also provides a computer device, which, at the hardware level, includes a processor, an internal bus, a network interface, memory, and non-volatile storage, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile storage into the memory and then runs it to implement the above-mentioned method for recognizing prokaryotic open reading frames.

[0044] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described method for identifying prokaryotic open reading frames.

[0045] For specific limitations on the computational system for identifying prokaryotic open reading frames, please refer to the limitations on the identification method for prokaryotic open reading frames mentioned above, which will not be repeated here. Each module in the aforementioned prokaryotic open reading frame identification system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0046] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for identifying open reading frames in prokaryotes, characterized in that, Includes the following steps: The target prokaryotic genome region to be tested is divided into multiple candidate sequence fragments. Each candidate sequence fragment includes a DNA subsequence and a Ribo-seq subsequence of ribosome sequencing coverage time data in the same genomic coordinate range as the DNA subsequence. For each candidate sequence fragment: Local coding features of the short-range codons, mid-range RBS core region, and long-range flanking sequences of the DNA subsequence are captured separately. All local coding features are concatenated to obtain a local sequence coding feature fusion vector. The bidirectional dependencies of the Ribo-seq subsequences of the ribosome sequencing coverage time-series data are captured, and the hidden state sequences of all time steps are output. The attention weights of each time step are dynamically calculated, and all hidden state sequences are weighted and summed to generate a global context feature vector capturing the long-distance temporal dependencies of the Ribo-seq. Simultaneously, the hidden state sequence of the last time step in capturing the bidirectional dependencies is retained as the temporal final state feature vector. The local sequence coding feature fusion vector, the global context feature vector, and the temporal final state feature vector are concatenated along the feature dimension to form a high-dimensional fusion feature vector. Based on the high-dimensional fusion feature vector, the probability that the candidate sequence fragment is an open reading frame is generated. Based on the probability of all candidate sequence fragments and their corresponding DNA subsequence coordinate intervals, the positions of each open reading frame in the target prokaryotic genome region to be tested are identified.

2. The method for identifying open reading frames in prokaryotes according to claim 1, characterized in that, The candidate sequence fragment is input into the DeepORF model, and the probability that the candidate sequence fragment is an open reading frame is output. The DeepORF model is based on the traditional DeepRibo model, but replaces the single-scale fixed convolution kernel of DNA sequence branching in the traditional DeepRibo model with a parallel multi-scale convolution module. A self-attention module is added after the Bi-GRU output layer of the Ribo-seq temporal branch in the traditional DeepRibo model; The parallel multi-scale convolution module uses three convolution kernels of different scales set in parallel, and is equipped with a batch normalized (BN) layer after each convolution kernel.

3. The method for identifying open reading frames in prokaryotes according to claim 1, characterized in that, After dividing the candidate sequence fragments, the process further includes preprocessing the Ribo-seq subsequences of the ribosome sequencing coverage time series data. Specifically, this includes: using a four-parameter S-curve to filter and reduce noise in the Ribo-seq ribosome sequencing coverage time series data, and then normalizing the noise-reduced Ribo-seq data; capturing the bidirectional dependencies of the normalized Ribo-seq data.

4. The method for identifying open reading frames in prokaryotes according to claim 1, characterized in that, After dividing the candidate sequence fragments, the process further includes preprocessing the DNA subsequence, specifically including: performing one-hot encoding on the DNA subsequence to obtain a binary feature tensor; and capturing each local encoded feature of the binary feature tensor.

5. The method for identifying open reading frames in prokaryotes according to claim 2, characterized in that, The DeepORF model is trained using the AdamW optimizer, which decouples the model's weight decay from gradient updates.

6. The method for identifying open reading frames in prokaryotes according to claim 1, characterized in that, The method of identifying the location of each open reading frame (OPF) in the target prokaryotic genome region based on the probability of all candidate sequence fragments and their corresponding DNA subsequence coordinate intervals specifically includes: setting a probability threshold, identifying candidate sequence fragments with a probability greater than or equal to the probability threshold as candidate OPFs, and the DNA subsequence coordinate interval of each candidate OPF is the candidate location of the OPF; if the coordinate intervals of multiple candidate OPFs are adjacent or overlap, they are merged to determine the precise start and end locations of an OPF.

7. A system for recognizing open reading frames in prokaryotes, characterized in that, include: The data acquisition module is used to divide the target prokaryotic genome region to be tested into multiple candidate sequence fragments. Each candidate sequence fragment includes a DNA subsequence and a Ribo-seq subsequence of ribosome sequencing coverage time series data in the same genomic coordinate interval as the DNA subsequence. The feature extraction module is used to: capture the local coding features of short-range codons, mid-range RBS core regions, and long-range flanking sequences of each candidate sequence fragment; concatenate all local coding features to obtain a local sequence coding feature fusion vector; capture the bidirectional dependencies of the Ribo-seq subsequences of the ribosome sequencing coverage time series data; output the hidden state sequences of all time steps; dynamically calculate the attention weights of each time step; and weight and sum all hidden state sequences to generate a global context feature vector that captures the long-distance temporal dependencies of the Ribo-seq; simultaneously, retain the hidden state sequence of the last time step in the process of capturing bidirectional dependencies as the temporal final state feature vector; and concatenate the local sequence coding feature fusion vector, the global context feature vector, and the temporal final state feature vector in the feature dimension to form a high-dimensional fusion feature vector. The identification module is used to generate the probability that the candidate sequence fragment is an open reading frame based on the high-dimensional fusion feature vector, and to identify the position of each open reading frame in the target prokaryotic genome region based on the probability of all candidate sequence fragments and their corresponding DNA subsequence coordinate intervals.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 6.

Images