A prokaryotic overlapping reading frame recognition method, system, device and medium
By introducing multi-view feature fusion and conditional random field decoding, the boundary conflict problem caused by overlapping genes and operon structures in prokaryotic ORF prediction is solved, achieving high accuracy and high recall of ORF annotation and outputting biologically consistent results.
Patent Information
- Application Number
- CN202611212909.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-03
AI Technical Summary
Existing technologies cannot effectively handle overlapping genes and operon structures in prokaryotic ORF prediction, leading to boundary conflicts and annotation inconsistencies.
Parallel multi-scale convolutional neural networks and bidirectional gated recurrent networks are used to extract multi-view features. Combined with conditional random field decoding, global joint decoding is performed using the Viterbi dynamic programming algorithm to generate emission fraction matrices and transition fraction matrices. This addresses the spatial constraints and physical transition probabilities of candidate ORFs and outputs the globally optimal open reading frame position.
It achieves high accuracy and high recall in complex genomic regions, resolves spatial conflicts in overlapping gene regions, and ensures that the annotation results are biologically consistent and reasonable.
Smart Images
Figure CN123033583A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary fields of deep learning and biology, and in particular to a method, system, device and medium for recognizing overlapping reading frames in prokaryotes. Background Technology
[0002] Precise annotation of prokaryotic (bacterial and archaea) genomes is a core foundation for microbial functional genomics, evolutionary research, and synthetic biology. Prokaryotic genomes are extremely compact, with high coding density, and widely exhibit complex organizational forms such as operon structures (multiple functionally related open reading frames tandemly transcribed) and overlapping genes (stop and start codons of adjacent ORFs sharing or overlapping several bases). Because traditional homology alignment and sequence motif searches are insensitive to short ORFs and species-specific ORFs, a large number of small open reading frames (sORFs) are missed.
[0003] Current mainstream models generally construct prokaryotic ORF prediction as an "independent scoring-threshold filtering or greedy selection" process. Specifically, the model first generates a large number of candidate ORF fragments on the genome through methods such as six-frame translation. Using architectures such as convolutional neural networks, recurrent neural networks, or Transformers, features are extracted from DNA sequences (codon usage, sequence motifs) and Ribo-seq read distributions (tribase periodicity, temporal variation of read density), respectively. A confidence score (i.e., emission score) characterizing the translational activity is independently calculated for each candidate ORF. Representative methods such as DeepRibo, RiboTISH, and Reparation treat each candidate ORF as an independent classification or regression sample. During decoding, they are sorted according to the score and a greedy strategy is used to retain the candidate with the highest score that does not have serious physical overlap with the retained ORFs, or a global threshold is used for truncation.
[0004] However, existing models rely on the independent "emission score" (local confidence) of each candidate ORF for isolated scoring and greedy filtering. They mechanically retain one or both candidates based on their scores, which creates irreconcilable boundary conflicts for the candidate ORFs. This prevents the model from jointly searching for the optimal path that meets physical constraints across the entire candidate sequence. When encountering complex regions such as overlapping genes, irresolvable boundary conflicts will inevitably arise, and the final output ORF annotations will be spatially inconsistent. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a method, system, device, and medium for recognizing overlapping reading frames in prokaryotes, thereby solving the problems in the prior art.
[0006] The present invention specifically provides the following technical solution: A method for recognizing overlapping reading frames in prokaryotes, comprising the following steps: Obtain the genome sequence of the prokaryote to be tested and its corresponding ribosome sequencing time-series data; Multi-view features are extracted from the genome sequence and ribosome sequencing time-series data, including local sequence features, time-series features and global context features; After fusing the multi-view features, a linear mapping is performed to generate the emission score matrix required for each candidate open reading frame in the candidate open reading frame sequence, and a transition score matrix representing the physical transition probability and spatial boundary constraints between adjacent candidate open reading frames is obtained; wherein the emission score matrix represents the independent confidence of each candidate open reading frame segment without considering the context; based on the emission score matrix and the transition score matrix, the Viterbi dynamic programming algorithm is used to perform global joint decoding on the candidate open reading frame sequence, search for the candidate open reading frame sequence path with the highest comprehensive score, and output the globally optimal open reading frame position annotation result.
[0007] Preferably, multi-view features are extracted from the genome sequence and ribosome sequencing time-series data, specifically as follows: The genome sequence is input into a parallel multi-scale convolutional neural network to extract local sequence multi-scale fusion features; wherein, the parallel multi-scale convolutional neural network contains three parallel convolutional branches, each branch is followed by a batch normalization layer and an activation function, and the output is spliced and flattened in the channel dimension. Ribosome sequencing time-series data are input into a bidirectional gated recurrent network to extract time-series features. These features are then weighted globally by an attention mechanism to obtain global context features. The bidirectional gated recurrent network employs a two-layer bidirectional gated recurrent unit structure, outputting the hidden state sequence matrix of all time steps to the attention mechanism module and outputting the time-series final state features of the last time step to the splicing and fusion layer.
[0008] Preferably, the attention mechanism adopts an additive attention mechanism based on nonlinear mapping, dynamically calculates the weights at each time step, performs a weighted summation on the hidden state sequence matrix, and outputs global context features.
[0009] Preferably, the emission score matrix is generated using a fully connected linear mapping layer, which directly linearly maps the emission scores of each candidate segment.
[0010] Preferably, the transition score matrix is automatically learned and updated through backpropagation during model training, wherein: for cases that conform to the biological law of reasonable overlap between the stop codon of the previous open reading frame and the start codon of the next open reading frame, a positive transition reward weight is assigned; for cases of range overlap that violate biological common sense, a negative penalty is applied.
[0011] Preferably, the physical transfer probabilities encoded by the transfer score matrix include: operon co-transcriptional dependence of adjacent open reading frames, start-stop codon spatial coupling of overlapping genes, and positional mutual exclusion constraints between candidate open reading frames.
[0012] Preferably, the construction process of candidate open reading frame sequences is as follows: all candidate open reading frames on the same genome sequence to be tested are arranged into an observation sequence according to their physical location to obtain the candidate open reading frame sequence.
[0013] This invention provides a prokaryotic overlapping reading frame recognition system, comprising: The data acquisition module is used to acquire the genome sequence of the prokaryotic organism to be tested and its corresponding ribosome sequencing time-series data; The feature extraction module is used to extract multi-view features from the genome sequence and ribosome sequencing time-series data, the multi-view features including local sequence features, time-series features and global context features; The identification module is used to fuse the multi-view features and perform linear mapping to generate the emission score matrix required for each candidate open reading frame in the candidate open reading frame sequence, and to obtain the transition score matrix representing the physical transition probability and spatial boundary constraints between adjacent candidate open reading frames; wherein the emission score matrix represents the independent confidence of each candidate open reading frame segment without considering the context; based on the emission score matrix and the transition score matrix, the Viterbi dynamic programming algorithm is used to perform global joint decoding on the candidate open reading frame sequence, search for the candidate open reading frame sequence path with the highest comprehensive score, and output the globally optimal open reading frame position annotation result.
[0014] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described method for recognizing overlapping reading frames of prokaryotes.
[0015] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for recognizing overlapping reading frames of prokaryotes.
[0016] Compared with the prior art, the present invention has the following significant advantages: This invention introduces a transition score matrix that characterizes the physical transition probability and spatial boundary constraints between adjacent candidate ORFs, and integrates independent confidence (emission score) and spatial compatibility (transition score) into a Viterbi dynamic programming framework for global joint decoding. This allows for the automatic adjudication and resolution of false positive conflicts using the Viterbi globally optimal path, fundamentally changing the decision-making paradigm of ORF prediction. Instead of mechanically making isolated greedy choices for individual candidate ORFs, it searches for globally optimal paths that meet physical constraints across the entire candidate sequence. This automatically and systematically resolves spatial conflicts in overlapping gene regions, completely avoiding the contradictory situation of assigning high confidence to physically incompatible candidate ORFs simultaneously. This ensures that the final output ORF annotation is strictly self-consistent in physical space, achieving higher annotation accuracy and recall in complex genomic regions, and providing biologically reasonable and globally optimal annotation results. Attached Figure Description
[0017] Figure 1 The flowchart illustrates a method for recognizing overlapping reading frames in prokaryotes provided by this invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0019] In the field of prokaryotic overlapping reading frame (ORF) identification, significant progress has been made in predicting prokaryotic ORFs using deep learning techniques combined with ribosome mapping (Ribo-seq) and DNA sequence analysis. However, existing models primarily treat ORF prediction as an "independent binary classification or scoring problem." In real-world biological scenarios, prokaryotic genomes are extremely compact. A key characteristic is the widespread presence of "operon" structures, where multiple adjacent ORFs are tandemly transcribed together. Furthermore, prokaryotes contain a large number of "overlapping genes," where the stop and start codons of adjacent ORFs may share or overlap several bases. Treating each candidate ORF in isolation not only violates the physical spatial continuity of the prokaryotic genome but also leads to severe boundary determination conflicts in complex overlapping regions.
[0020] To address the shortcomings of existing deep learning schemes, such as boundary conflicts and poor biological interpretability caused by "isolated predictions," this invention proposes a prokaryotic ORF identification method based on joint decoding of conditional random fields (CRF).
[0021] The purpose of this invention is to completely transform the network post-processing / decoding mechanism of the model while retaining the multi-scale feature extraction capability of the underlying layer. By introducing a sequence joint decoding algorithm, the "co-transcriptional characteristics" of prokaryotic operons and the "spatial constraint transfer probability" of overlapping genes are encoded into the prediction network, thereby automatically resolving boundary conflicts between adjacent ORFs and outputting globally optimal and biologically rigorous self-consistent reading frame sequences.
[0022] This invention, based on existing deep multi-scale feature extraction networks, introduces Conditional Random Fields (CRF) to construct an end-to-end joint decoding system. The implementation scheme is divided into two parts: "Basic Physical Structure of the Model" and "Implementation Steps for Joint Decoding Based on CRF," specifically as follows:
[0023] Part 1: The Complete Structure and Connections of the Basic Model of this Invention. The network system architecture of this invention includes a dual-branch input terminal, a parallel multi-scale convolutional network (CNN), a bidirectional gated recurrent network (Bi-GRU), an attention module, a multi-view feature concatenation layer, an emission fraction mapping layer, and a conditional random field decoding module (CRFModule). The specific inputs, outputs, and connections of each module are as follows: Module 1, Overall Model Input Terminal: Includes a DNA sequence input terminal and a Ribo-seq time-series data input terminal. The DNA sequence is input to the multi-scale CNN module after one-hot encoding; the Ribo-seq coverage data is input to the Bi-GRU module after S-curve denoising and normalization. Module 2, Local Sequence Feature Extraction Module (Multi-scale CNN): Internally contains three parallel convolutional branches (with kernel sizes of 3, 7, and 15 respectively), each branch followed by a batch normalization (BN) layer and an activation function. The output terminal directly concatenates and flattens along the channel dimension, outputting a "local sequence encoded feature fusion vector" to the multi-view feature concatenation layer. Module 3, Temporal Feature Extraction Module (Bi-GRU): Employs a two-layer Bi-GRU to capture bidirectional dependencies. The output has two dimensions: first, the "hidden state sequence matrix" for all time steps (connected to the next-level attention mechanism module); second, the "temporal final state feature vector" for the last time step (directly connected to the multi-view feature concatenation layer). Module 4, Attention Mechanism Module: Receives the hidden state sequence matrix from the Bi-GRU. Internally, it employs an additive attention mechanism structure based on nonlinear mapping. It dynamically calculates the weights of each time step, performs a weighted summation of the hidden states, and outputs a "global context feature vector" to the multi-view feature concatenation layer. Module 5, Multi-view Feature Concatenation Layer: Receives features from the three independent branches mentioned above, performs first-to-last concatenation and fusion along the feature dimension, and outputs a single "high-dimensional fused feature vector." This output is connected to the emission fraction mapping layer. Module 6, Emission Mapping Layer: Contains a fully connected linear mapping layer and receives the high-dimensional fused feature vector from the multi-view feature concatenation layer. Instead of using traditional Softmax activation for independent probability classification, it directly linearly reduces the dimensionality and maps it to the "Emission Score" matrix required by the CRF module. Module 7, Conditional Random Field Decoding Module (CRF Module): Receives the above emission score matrix and internally initializes and maintains a learnable "Transition Score" matrix (used to characterize the physical transition probabilities and boundary constraints between adjacent candidate ORFs).The module integrates a Viterbi decoding algorithm unit at its end, and outputs the globally optimal ORF sequence path.
[0024] Part Two: Specific Implementation Steps of CRF-Based Joint Decoding. Based on the above network architecture, this invention proposes a method for recognizing overlapping reading frames in prokaryotes during prediction, such as... Figure 1 As shown, the specific steps include the following:
[0025] Step 1: Obtain the genome sequence of the prokaryotic organism to be tested and its corresponding ribosome sequencing time series data.
[0026] Step 2: Extract multi-view features from genome sequence and ribosome sequencing time-series data. Multi-view features include local sequence features, time-series features, and global context features.
[0027] High-dimensional feature extraction at the front end based on multi-scale networks. The genome sequence of the prokaryote to be tested and the Ribo-seq ribosome sequencing time series data are input into the front-end deep neural network (modules 1 to 5), and the output is a "high-dimensional fusion feature vector" of all candidate ORF fragments on the entire sequence to be tested.
[0028] The genome sequence is input into a parallel multi-scale convolutional neural network to extract local sequence multi-scale fusion features. The parallel multi-scale convolutional neural network contains three parallel convolutional branches, each followed by a batch normalization layer and an activation function. The output is spliced and flattened in the channel dimension.
[0029] Ribosome sequencing time-series data is input into a bidirectional gated recurrent network (BRNN) to extract temporal features. These features are then weighted globally using an attention mechanism to obtain global context features. The BRNN employs a two-layer bidirectional gated recurrent unit structure, outputting the hidden state sequence matrix of all time steps to the attention mechanism module and the final temporal state features of the last time step to the splicing and fusion layer. The attention mechanism uses an additive attention mechanism based on nonlinear mapping, dynamically calculating the weights of each time step and performing a weighted summation of the hidden state sequence matrix to output the global context features.
[0030] Step 3: After fusing the multi-view features, perform linear mapping to generate the emission score matrix required for each candidate open reading frame in the candidate open reading frame sequence, and obtain the transition score matrix representing the physical transition probability and spatial boundary constraints between adjacent candidate open reading frames; where the emission score matrix represents the independent confidence of each candidate open reading frame segment without considering the context; based on the emission score matrix and the transition score matrix, use the Viterbi dynamic programming algorithm to perform global joint decoding of the candidate open reading frame sequence, search for the candidate open reading frame sequence path with the highest comprehensive score, and output the globally optimal open reading frame position annotation result.
[0031] The local sequence multi-scale fusion features, the temporal final state features of the last time step of the bidirectional gated recurrent network, and the global context features are concatenated and fused to obtain the fused features. The fused features are then directly linearly mapped through a fully connected linear mapping layer to obtain the emission score matrix corresponding to each candidate segment, which is the conditional random field required. The emission score matrix is used to characterize the independent confidence of each candidate open reading frame segment without considering the context.
[0032] Construct candidate ORF sequences and emission score matrices. The construction process for candidate open reading frame (ORF) sequences is as follows: all candidate ORFs on the same genome sequence are arranged according to their physical position (5' to 3') into an observation sequence, resulting in candidate ORF sequences. The feature vectors generated in step one are then converted into an emission score matrix for the CRF layer through an emission score mapping layer (module 6). This matrix represents the independent confidence that each fragment itself constitutes a true ORF, without considering the context.
[0033] A transition score matrix encoding biological prior laws is constructed. This matrix is continuously learned and updated in the Conditional Random Field Decoding module (Module 7). This matrix is specifically used to capture the physical transition probabilities between adjacent candidate ORFs. During model training, the transition score matrix is automatically learned and updated through backpropagation. During training, the model automatically assigns high transition reward weights to typical biological phenomena such as "reasonable overlap between the stop codon of ORF_A and the start codon of ORF_B," i.e., cases that conform to the biological law of reasonable overlap between the stop codon of the preceding open reading frame and the start codon of the following open reading frame; simultaneously, it imposes a large negative penalty on extremely large overlaps that violate biological norms. The physical transition probabilities encoded by the transition score matrix include: operon co-transcriptional dependence between adjacent open reading frames, spatial coupling between the start and stop codons of overlapping genes, and positional mutual exclusion constraints between candidate open reading frames.
[0034] Global joint decoding based on the Viterbi algorithm. In the prediction and inference phase, the CRF decoding module uses the Viterbi dynamic programming algorithm to comprehensively calculate the sum of "emission score" and "transition score". Instead of locally and independently selecting the segment with the highest emission score, the algorithm searches the solution space of the entire input sequence for a globally optimal ORF sequence path that has the highest overall score and completely conforms to the arrangement rules of the prokaryotic operative subspace. Finally, it outputs the precise ORF position that excludes boundary conflicts.
[0035] The beneficial effects (advantages) achieved by this invention are as follows: (1) Perfectly solves the boundary conflict problem of overlapping genes: It abandons local greedy and isolated scoring and introduces physical space constraints through the CRF transition matrix. When encountering overlapping genes or dense manipulator regions, the model can automatically adjudicate and resolve false positive conflicts using the global optimal path, and output biologically consistent prediction results.
[0036] (2) Significantly improve the prediction recall and accuracy of the operative subregion: The joint decoding mechanism simulates the Markov chain characteristics of "translation coupling" in prokaryotes, and uses the contextual dependency relationship between upstream and downstream ORFs to assist in the determination of extremely weak signal regions, effectively finding short sequence ORFs that were previously easy to miss.
[0037] Based on the above method, this invention proposes a prokaryotic overlapping reading frame recognition system, comprising: The system comprises the following modules: a data acquisition module for acquiring the genome sequence of the prokaryotic organism to be tested and its corresponding ribosome sequencing time-series data; a feature extraction module for extracting multi-view features from the genome sequence and ribosome sequencing time-series data, including local sequence features, time-series features, and global context features; an identification module for fusing the multi-view features and performing linear mapping to generate the emission score matrix required for each candidate open reading frame (OPG) in the candidate OPG sequence, and obtaining the transition score matrix representing the physical transition probability and spatial boundary constraints between adjacent candidate OPGs; the emission score matrix represents the independent confidence of each candidate OPG fragment without considering the context; and a Viterbi dynamic programming algorithm for globally joint decoding of the candidate OPG sequences based on the emission score matrix and the transition score matrix, searching for the candidate OPG sequence path with the highest comprehensive score, and outputting the globally optimal OPG position annotation result.
[0038] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described method for recognizing overlapping reading frames of prokaryotes.
[0039] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).
[0040] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for recognizing overlapping reading frames of prokaryotes.
[0041] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
[0042] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.
Claims
1. A method for recognizing overlapping reading frames in prokaryotes, characterized in that, Includes the following steps: Obtain the genome sequence of the prokaryote to be tested and its corresponding ribosome sequencing time-series data; Multi-view features are extracted from the genome sequence and ribosome sequencing time-series data, including local sequence features, time-series features and global context features; After fusing the multi-view features, a linear mapping is performed to generate the emission score matrix required for each candidate open reading frame in the candidate open reading frame sequence, and a transition score matrix representing the physical transition probability and spatial boundary constraints between adjacent candidate open reading frames is obtained; wherein the emission score matrix represents the independent confidence of each candidate open reading frame segment without considering the context; based on the emission score matrix and the transition score matrix, the Viterbi dynamic programming algorithm is used to perform global joint decoding on the candidate open reading frame sequence, search for the candidate open reading frame sequence path with the highest comprehensive score, and output the globally optimal open reading frame position annotation result.
2. The method for recognizing overlapping reading frames in prokaryotes as described in claim 1, characterized in that, Multi-view features are extracted from the genome sequence and ribosome sequencing time-series data, specifically as follows: The genome sequence is input into a parallel multi-scale convolutional neural network to extract local sequence multi-scale fusion features; wherein, the parallel multi-scale convolutional neural network contains three parallel convolutional branches, each branch is followed by a batch normalization layer and an activation function, and the output is spliced and flattened in the channel dimension. Ribosome sequencing time-series data are input into a bidirectional gated recurrent network to extract time-series features. These features are then weighted globally by an attention mechanism to obtain global context features. The bidirectional gated recurrent network employs a two-layer bidirectional gated recurrent unit structure, outputting the hidden state sequence matrix of all time steps to the attention mechanism module and outputting the time-series final state features of the last time step to the splicing and fusion layer.
3. The method for recognizing overlapping reading frames in prokaryotes as described in claim 2, characterized in that, The attention mechanism adopts an additive attention mechanism based on nonlinear mapping, dynamically calculates the weights at each time step, performs weighted summation on the hidden state sequence matrix, and outputs global context features.
4. The method for recognizing overlapping reading frames in prokaryotes as described in claim 1, characterized in that, The emission score matrix is generated using a fully connected linear mapping layer, which directly obtains the emission score corresponding to each candidate segment through linear mapping.
5. The method for recognizing overlapping reading frames in prokaryotes as described in claim 1, characterized in that, The transition score matrix is automatically learned and updated through backpropagation during model training. Specifically, for cases that conform to the biological principle of reasonable overlap between the stop codon of the previous open reading frame and the start codon of the next open reading frame, a positive transition reward weight is assigned; for cases of overlap that violate biological norms, a negative penalty is applied.
6. The method for recognizing overlapping reading frames in prokaryotes as described in claim 1, characterized in that, The physical transfer probabilities encoded by the transfer score matrix include: operon co-transcriptional dependence of adjacent open reading frames, start-stop codon spatial coupling of overlapping genes, and positional mutual exclusion constraints between candidate open reading frames.
7. The method for recognizing overlapping reading frames in prokaryotes as described in claim 1, characterized in that, The process of constructing candidate open reading frame sequences is as follows: all candidate open reading frames on the same genome sequence to be tested are arranged into an observation sequence according to their physical location to obtain the candidate open reading frame sequence.
8. A prokaryotic overlapping reading frame recognition system, characterized in that, include: The data acquisition module is used to acquire the genome sequence of the prokaryotic organism to be tested and its corresponding ribosome sequencing time-series data; The feature extraction module is used to extract multi-view features from the genome sequence and ribosome sequencing time-series data, the multi-view features including local sequence features, time-series features and global context features; The identification module is used to fuse the multi-view features and perform linear mapping to generate the emission score matrix required for each candidate open reading frame in the candidate open reading frame sequence, and to obtain the transition score matrix representing the physical transition probability and spatial boundary constraints between adjacent candidate open reading frames; wherein the emission score matrix represents the independent confidence of each candidate open reading frame segment without considering the context; based on the emission score matrix and the transition score matrix, the Viterbi dynamic programming algorithm is used to perform global joint decoding of the candidate open reading frame sequence, search for the candidate open reading frame sequence path with the highest comprehensive score, and output the globally optimal open reading frame position annotation result.
9. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a program that, when executed by the processor, causes the processor to perform the steps of the method for recognizing overlapping reading frames of prokaryotes as described in any one of claims 1 to 7.
10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for recognizing overlapping reading frames of prokaryotes according to any one of claims 1 to 7.