A method, system and medium for prokaryotic promoter identification
By combining regional attention mechanisms with evidence branches and spatial grammar constraints, this method solves the problem of the lack of explicit constraints on attention mechanisms in existing promoter recognition methods. It achieves structured recognition of regional features of promoter core elements, improving the reliability and interpretability of recognition results.
Patent Information
- Application Number
- CN202611245764.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
AI Technical Summary
In existing promoter recognition methods based on pre-trained DNA language models, the attention mechanism lacks clear constraints, resulting in insufficient utilization of the regional features of the promoter core elements and making it difficult to verify their true correlation with classification decisions.
We employ a dual verification approach combining regional attention mechanisms and evidence branches, along with spatial grammar constraints. By locating the -35 box, -10 box, and transcription factor binding site regions, we extract location information, attention weights, and evidence strength vectors for structured fusion feature recognition.
This enhances the reliability and interpretability of the association between regional features and classification decisions during promoter identification, thereby improving the reliability and interpretability of the identification results.
Smart Images

Figure CN123120044A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of bioinformatics and artificial intelligence, and specifically relates to a method, system and medium for prokaryotic promoter identification. Background Technology
[0002] Prokaryotic promoters are DNA sequence regions recognized and bound by RNA polymerases, and are core elements regulating the initiation of gene transcription in prokaryotes. In prokaryotic promoter recognition tasks, because promoters contain core elements with specific biological functions, such as -35 boxes and -10 boxes, and the relative positions and combinations of these elements directly affect transcription efficiency, simultaneously utilizing the regional characteristics and interrelationships of these core elements during the discrimination process is of great significance for improving the reliability and interpretability of the recognition results.
[0003] Currently, promoter identification methods based on pre-trained DNA language models have become one of the mainstream technical approaches in this field. For example, the TSSNote-CyaPromBERT model predicts promoters based on the BERT architecture and uses attention visualization to show the model's attention to key functional regions of the promoter. However, the attention mechanism in the above methods is mainly learned automatically by the network according to the classification target as a general feature extraction method. There is a lack of clear correspondence between the regions of attention and the core functional elements of the promoter. Moreover, such visualization results are passive observations after the model has been trained, making it difficult to verify their true correlation with the final classification decision. This results in shortcomings in utilizing regional structured features of the core elements of the promoter. Summary of the Invention
[0004] The purpose of this invention is to provide a method, system and medium for prokaryotic promoter identification, which achieves structured and interpretable identification of prokaryotic promoters through dual verification of region attention mechanism and evidence branch, combined with spatial syntax constraints.
[0005] To address the aforementioned technical problems, embodiments of the present invention provide a prokaryotic promoter identification method, comprising the following steps: Obtain a comprehensive feature representation of the DNA sequence to be tested; The comprehensive feature representation is processed by region focusing to locate the -35 box region, the -10 box region and the transcription factor binding site region. The location information, attention weight distribution and regional context features of each region are obtained, and the corresponding evidence strength vector is extracted for each located region. The sequence of the -35 box region and the -10 box region is determined based on the order of the location information; the distance between the -35 box region and the -10 box region is determined based on the location information as the absolute distance; the distance rationality is determined based on the absolute distance; the pairing rationality is determined based on the peak height of the attention weight distribution of the -35 box region and the -10 box region; the spatial grammatical features are determined based on the sequence indication, absolute distance, distance rationality, and pairing rationality. By fusing spatial grammatical features, contextual features of each region, and evidence strength vectors, a structured fusion feature is obtained. Based on the structured fusion characteristics, the presence of a prokaryotic promoter in the DNA sequence to be tested can be identified.
[0006] In some optional embodiments, after identifying the presence of a prokaryotic promoter in the DNA sequence to be tested based on the structured fusion features, the method further includes: Based on the comprehensive feature representation, the presence of a prokaryotic promoter in the DNA sequence to be tested is identified as a second identification result; The second identification result is fused with the identification result based on structured fusion features to obtain the final identification result of whether a prokaryotic promoter exists in the DNA sequence to be tested.
[0007] In some optional embodiments, the specific steps for obtaining the comprehensive feature representation of the DNA sequence to be tested are as follows: Obtain the target genome sequence, slide a window of a preset length on the target genome sequence with a preset step size, and use the sequence corresponding to each window as the DNA sequence to be tested; The DNA sequence to be tested is segmented according to the 6-mer vocabulary with a step size of 1 and converted into a token sequence; The token sequence is input into the pre-trained DNABERT model, and after processing by the embedding layer and a 12-layer Transformer encoder, global contextual hidden features are extracted. Multi-scale convolution is applied to the global contextual hidden features to enhance local features, resulting in a comprehensive feature representation.
[0008] In some optional embodiments, the multi-scale convolutional local feature enhancement involves the following steps: Four parallel one-dimensional convolutional layers with kernel sizes of 1, 3, 5, and 9 are used to perform convolution processing on the global contextual hidden features; The outputs of the four convolutional branches are concatenated along the feature dimension, and after layer normalization and linear mapping, they are added element-wise with the global context hidden features to obtain a comprehensive feature representation.
[0009] In some optional embodiments, the specific steps of performing region focusing processing on the comprehensive feature representation are as follows: The integrated feature representation was focused on -35 box region, -10 box region and transcription factor binding site region respectively, and the position information, attention weight distribution and regional context features of the -35 box region, -10 box region and transcription factor binding site region were obtained respectively. The -35 box region focusing and the -10 box region focusing are respectively implemented in the following ways: Get the dot product of the query vector and the key vector; Add the learnable position bias matrix to the dot product result to obtain the corrected attention score; The attention weight distribution is determined based on the corrected attention score; The position bias matrix is used to give the attention weights higher prior probabilities at preset sequence positions corresponding to the -35 box region and the -10 box region.
[0010] In some optional embodiments, the sequence indication is: if the location information of the -35 box region is before the location information of the -10 box region, the sequence indication is 1; otherwise, it is 0. The absolute distance is the absolute value of the difference between the location information of the -35 box region and the location information of the -10 box region. The distance reasonableness is: if the absolute distance is between 15 and 25 bp, the value is 1; otherwise, the value is 0. The pairing reasonableness is the product of the peak height of the attention weight distribution of the -35 box region and the peak height of the attention weight distribution of the -10 box region.
[0011] In some optional embodiments, the step of extracting the corresponding evidence strength vector for each located region involves the following steps: Evidence of promoter core elements and evidence of transcription factor binding sites are extracted from the comprehensive feature representation, and corresponding strength vectors of evidence of promoter core elements and evidence of transcription factor binding sites are obtained. The evidence for the promoter core element and the evidence for the transcription factor binding site are extracted by two layers of one-dimensional convolution followed by a sigmoid activation function.
[0012] In some optional embodiments, the specific steps for fusing the second recognition result with the recognition result based on structured fusion features are as follows: The identification results based on structured fusion features and the second identification results are weighted and added together according to a preset ratio, wherein the weight of the identification results based on structured fusion features is greater than the weight of the second identification results; The weighted sum is mapped to the predicted probability of the presence or absence of a prokaryotic promoter using the softmax function.
[0013] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the prokaryotic promoter identification method described above.
[0014] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when run by a processor, is capable of executing the above-described prokaryotic promoter identification method.
[0015] The prokaryotic promoter identification method provided by this invention has at least the following beneficial effects: This invention integrates regional attention mechanisms, evidence extraction, and spatial grammatical relationships into a structured recognition framework. This allows the regional features and their interrelationships of the promoter core elements to no longer be passive observations after training a general model, but to directly participate in the recognition process as structured discrimination criteria. This enhances the reliability and interpretability of the association between regional features and classification decisions while maintaining the promoter recognition and discrimination capabilities. Attached Figure Description
[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0017] Figure 1 This is a flowchart of a prokaryotic promoter identification method provided according to an embodiment of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0019] One embodiment of the present invention relates to a prokaryotic promoter identification method. The implementation details of the prokaryotic promoter identification method of this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.
[0020] The specific process of a prokaryotic promoter identification method in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Obtain the comprehensive feature representation of the DNA sequence to be tested; (1) Data Preprocessing and Sequence Encoding: Using prokaryotic genome sequences as the data source, known annotated promoter sequences were collected as positive samples, and equal-length random sequences from non-promoter regions were collected as negative samples. All sequences were uniformly truncated to a fixed length, for example, −60 bp upstream of the transcription start site (TSS) to +40 bp downstream, totaling 100 bp. The sequences were segmented using a 6-mer vocabulary with a step size of 1, converting each DNA sequence into a token sequence. The pre-trained DNABERT model was used as the global encoding backbone. This model had been pre-trained on multiple prokaryotic genomes for masked language modeling. The input token sequences were fed into the 12-layer Transformer encoder of DNABERT after the embedding layer, and the output was a global contextual hidden feature with a dimension of 768. The output layer of DNABERT was not used in this embodiment.
[0021] (2) Multi-scale convolutional local feature enhancement: Based on the hidden features output by DNABERT, a multi-scale convolutional local feature enhancement module is constructed. This module contains four parallel one-dimensional convolutional layers with kernel sizes of 1, 3, 5, and 9, and a stride of 1. Same padding is used to maintain the sequence length, and the number of output channels of each convolutional branch is set to 192. The outputs of the four branches are concatenated along the feature dimension to obtain a local enhanced feature with a dimension of 768. Subsequently, it passes through layer normalization (LayerNorm) and a linear mapping layer to compress the dimension back to 768. Finally, the result is added element-wise with the global features of DNABERT to form a comprehensive feature representation.
[0022] (3) Sliding window scanning strategy: In the actual inference stage, facing an unknown target genome sequence, a whole-genome sliding window scanning strategy is adopted. A window of a preset length is slid across the target genome sequence with a preset step size, and the sequence corresponding to each window is taken as the DNA sequence to be tested; wherein, the center position of the window of the preset length is taken as the candidate transcription start site (TSS), and the sequence of each window is a DNA sequence of a preset length truncated with the candidate TSS as the center. Subsequent steps are performed on the DNA sequences to be tested corresponding to multiple windows respectively to obtain the confidence level of the prediction result corresponding to each window; based on the confidence level of the prediction result corresponding to each window, the prokaryotic promoter position in the target genome sequence is determined. Optionally, the preset step size is 1 base and the preset length is 100 bp. Optionally, the center position of the window with a confidence level greater than a preset threshold is determined as the candidate position of the transcription start site.
[0023] It should be understood that the above sliding window scanning strategy solves the problem of unknown TSS location in practical applications -- by exhaustively listing every base site as a candidate TSS, the model can scan the whole genome and accurately locate the promoter.
[0024] In some optional embodiments, the specific steps for obtaining a comprehensive feature representation of the DNA sequence to be tested are as follows: Obtain the target genome sequence and divide it into several sub-sequences as DNA sequences to be tested; The DNA sequence to be tested is segmented and converted into a token sequence; Input the token sequence into a pre-trained DNA language model to extract global contextual hidden features; Multi-scale convolution is applied to the global contextual hidden features to enhance local features, resulting in a comprehensive feature representation.
[0025] In some optional embodiments, the specific steps of multi-scale convolutional local feature enhancement are as follows: Four parallel one-dimensional convolutional layers with kernel sizes of 1, 3, 5, and 9 are used to perform convolution processing on the global contextual hidden features; The outputs of the four convolutional branches are concatenated along the feature dimension, and after layer normalization and linear mapping, they are added element-wise with the global context hidden features to obtain a comprehensive feature representation.
[0026] Step 102: Perform region focusing processing on the comprehensive feature representation to locate the -35 box region, -10 box region and transcription factor binding site region, obtain the location information, attention weight distribution and regional context features of each region, and extract the corresponding evidence strength vector for each located region. Three independent region evidence attention heads are defined on top of the comprehensive features: the box35 attention head, the box10 attention head, and the TFBS attention head. Among them, the box35 and box10 attention heads introduce learnable positional bias vectors, which give the attention weights higher prior probabilities in the early part of the sequence (corresponding to the -35 box region, approximately positions 20~30) or the lower middle part (corresponding to the -10 box region, approximately positions 45~55); the TFBS attention head does not introduce positional bias and is completely data-driven.
[0027] In some optional embodiments, the -35 box region attention operation and the -10 box region attention operation specifically include: Get the dot product of the query vector and the key vector; Add the learnable position bias matrix to the dot product result to obtain the corrected attention score; The attention weight distribution is determined based on the corrected attention score; The position bias matrix is used to give the attention weights a higher prior probability at the preset sequence positions corresponding to the -35 box region and / or the -10 box region.
[0028] The position bias matrix is a trainable parameter matrix, whose values are updated through backpropagation during model training. In the initial training phase, the position bias matrix is configured to have higher values at the preset sequence positions. During the inference phase, a fixed-length DNA sequence is extracted from the target genome sequence using a sliding window, centered at each base site, with the center of each window serving as a candidate transcription start site. For each window, the position bias matrix in the -35-box and -10-box region attention operations is used to verify whether the center position of the window possesses the sequence features flanking the transcription start site: if the center position of the current window is a true transcription start site, the upstream -35-box and -10-box regions fall into the preset sequence positions corresponding to the position bias matrix, and the model outputs a high-confidence prediction result; if the center position of the current window is not a transcription start site, the -35-box and -10-box regions cannot fall into the preset sequence positions corresponding to the position bias matrix, and the model outputs a low-confidence prediction result. By traversing the entire target genome sequence through a sliding window, prokaryotic promoter scanning and localization across the entire genome are achieved.
[0029] During the attention calculation process, a downstream bias matrix (with higher values in the first half of the sequence) is added to the query-key dot product result of the attention head for box35, and a downstream bias matrix (with higher values in the middle to later part of the sequence) is added to the attention head for box10. After obtaining the attention weights, the response regions of box35, box10, and TFBS are weighted and aggregated respectively: the comprehensive features are weighted and summed using the attention weights as coefficients to obtain the respective region context features (each region context is 768-dimensional).
[0030] We set up a promoter evidence branch and a TFBS evidence branch. Each branch consists of two one-dimensional convolutional layers (kernel size 3, output channels 64) followed by a sigmoid activation function, outputting an evidence strength vector of the same length as the input sequence. Through loss function constraints, we improve the consistency between the box35 attention distribution and promoter evidence in the -35 region, improve the consistency between the box10 attention distribution and promoter evidence in the -10 region, and ensure that the TFBS attention distribution remains consistent with the TFBSevidence distribution.
[0031] Meanwhile, highly responsive local fragments (taking the top-k peak positions, where k=5) are extracted from the outputs of the promoter evidence branch and the TFBS evidence branch, and the average value thereof is taken as the local evidence strength vector.
[0032] In some optional embodiments, the promoter core element evidence branch processing and the transcription factor binding site evidence branch processing are respectively implemented by two layers of one-dimensional convolution followed by a sigmoid activation function.
[0033] Step 103: determining a sequence indicator of the -35 box region and the -10 box region according to the sequence of the position information; determining the distance between the -35 box region and the -10 box region according to the position information, as an absolute distance; determining distance rationality according to the absolute distance; determining pairing rationality according to the peak heights of the attention weight distributions of the -35 box region and the -10 box region; determining a spatial grammar feature according to the sequence indicator, the absolute distance, the distance rationality and the pairing rationality; pos35 and pos10 are calculated according to the weighted center positions of the box35 and box10 attention response distributions. Then the following features are calculated: sequence indicator (taking 1 if pos35 < pos10, otherwise taking 0), absolute distance (|pos10-pos35|, in the unit of base number), distance rationality (taking 1 if the distance is between 15 and 25 bp, otherwise taking 0), and pairing rationality (the product of the attention peak heights of the two regions). The above four features are concatenated to form a 4-dimensional vector, which is input into a small MLP (two layers, hidden layer 32-dimensional, output 16-dimensional) in the interpretable branch for feature encoding.
[0034] In some optional embodiments, the sequence indicator is: taking a value of 1 if the first position information is less than the second position information, otherwise taking a value of 0; the absolute distance is the absolute value of the difference between the first position information and the second position information; the distance rationality is: taking a value of 1 if the absolute distance is between 15 and 25 bp, otherwise taking a value of 0; the pairing rationality is the product of the peak height of the attention weight distribution of the -35 box region and the peak height of the attention weight distribution of the -10 box region.
[0035] Step 104: fusing the spatial grammar feature, the context features of each region and the evidence strength vector to obtain a structured fused feature; Step 105: identifying whether a prokaryotic promoter exists in the DNA sequence to be tested according to the structured fused feature.
[0036] The interpretable branch concatenates the following features: box35 region context (768-dimensional), box10 region context (768-dimensional), TFBS region context (768-dimensional), promoter evidence summary vector (average of high-response fragments, 64-dimensional), TFBS evidence summary vector (64-dimensional), and 16-dimensional features obtained after spatial grammar relation encoding, totaling approximately 2448 dimensions. This concatenated vector is input into a two-layer MLP (128-dimensional hidden layer, 2-dimensional logits output).
[0037] The global discriminant branch inputs the comprehensive feature representation into a linear classification layer after global average pooling (outputting 768 dimensions), outputting 2-dimensional logits. Finally, the logits of the two branches are added at a fixed ratio, with the fusion ratio of the interpretable branch set to 0.85 and the fusion ratio of the global branch set to 0.15. The sum is then passed through a softmax function to obtain the final classification probability (promoter / non-promoter).
[0038] In some optional embodiments, the specific steps for fusing the first classification result and the second classification result are as follows: The first classification result and the second classification result are weighted and added together according to a preset ratio, wherein the weight of the first classification result is greater than the weight of the second classification result; The weighted sum is mapped to the predicted probability of the presence or absence of a prokaryotic promoter using the softmax function.
[0039] In some optional embodiments, the weight of the first classification result is 0.85, and the weight of the second classification result is 0.15.
[0040] The model is trained using a structured constraint loss function: The total loss includes the classification cross-entropy loss and multiple structured constraint losses. The specific settings are as follows:
[0041] Classification loss: , in, For predicting labels, This is a real label.
[0042] Sequence constraints: That is, punishment pos35 Greater than pos10 The situation.
[0043] Spacing constraints: ,in .
[0044] Overlap suppression: ,in, and They are respectively box35Attention distribution and box10 Attention distribution in the Attention weights for each sequence position.
[0045] Consistency of evidence: ,in, , and These are the normalized attention weight distributions; , and These represent the evidence strength distributions of the corresponding regions after channel mapping and normalization.
[0046] High-response enhancement: ,in, The number of high-response positions is set to 5; Represents attention distribution The highest weight One position.
[0047] Negative sample anomaly suppression: The loss is calculated on negative samples; and These are the location masks for the -35 box region and the -10 box region, respectively. This indicates element-wise multiplication.
[0048] Stability constraints: ,in, For attention distribution entropy, Let the target entropy value be 1.5.
[0049] Perturbation Causal Alignment Loss: ,in, Preset interval; To determine the decrease in logit for the corresponding subcategory after masking the critical area. This is the decrease in logit for the corresponding subclass after masking a random, equal-length region.
[0050] The balance coefficient λ for each loss is set to 0.5, and the total loss is... .
[0051] A two-stage training strategy is used to train the model: Phase 1: Training for 8 epochs, using AdamW as the optimizer, with the learning rate of the pre-trained DNABERT backbone set to 1×10⁻ 5 The learning rate for the head modules (attention head, evidence branch, MLP classifier) is set to 3×10⁻ 4 The weight decays to 1×10⁻ 4The classification loss, order constraint, and margin constraint are enabled, and the remaining constraints are introduced in a stepwise manner, with the weights increasing linearly to the maximum value starting from round 3.
[0052] Phase 2: Training for 80 epochs, freezing the parameters of the pre-trained DNABERT backbone and the linear layers after global average pooling in the global discriminant branch; enabling all structured constraint losses, with each loss weight maintained at 0.5; adding explicit region alignment training, applying a weak supervised bias every 10 steps based on the current attention center position (if the true -35 region of the positive sample is known, then increasing the attention weight of that region). The fusion ratio of the interpretable branch is fixed at 0.85. An early stopping strategy is adopted: every 5 epochs, the mean absolute error between the box35 attention center and the true -35 region is calculated on the validation set; if there is no improvement after 10 consecutive epochs, training is stopped. The learning rate uses cosine annealing scheduling, with the initial learning rate maintained at the above values, and the minimum learning rate reduced to 1×10⁻. 6 .
[0053] Output the model recognition results and interpretation results.
[0054] For any input DNA sequence, this invention outputs the following: Binary classification results and confidence levels (probability values) for promoter / non-promoter; Attention response distribution map of box35, box10, and TFBS regions (can be output as a vector or heatmap of length 100). The specific location of the high-response local window (e.g., the range of 3 bases before and after the attention peak); The distance between boxes 35 and 10 (expressed as the number of bases) and the pairing rationality score (between 0 and 1). Optional: The amount of change in prediction before and after partial occlusion (occluding high-response areas), used to indicate the causal contribution of that area to the classification decision.
[0055] The above output results can be saved as text files or visualizations for subsequent biological validation and regulatory element design.
[0056] In some optional embodiments, obtaining the DNA sequence to be tested includes: A window of a preset length is slid across the target genome sequence with a preset step size, and the sequence corresponding to each window is taken as the DNA sequence to be tested; The above method is performed on the DNA sequences to be tested corresponding to multiple windows respectively to obtain the prediction results corresponding to each window; Based on the confidence level of the prediction results corresponding to each window, the prokaryotic promoter position in the target genome sequence is determined.
[0057] In some optional embodiments, the preset step size is 1 base and the preset length is 100 bp.
[0058] In some optional embodiments, determining the prokaryotic promoter position in the target genome sequence based on the confidence level of the prediction results corresponding to each window includes: determining the center position of the window with a confidence level greater than a preset threshold as a candidate position for transcription start site.
[0059] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the protection scope of this invention.
[0060] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.
[0061] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0062] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.
Claims
1. A method of prokaryotic promoter recognition, characterized by, The method includes: Obtain a comprehensive feature representation of the DNA sequence to be tested; The comprehensive feature representation is processed by region focusing to locate the -35 box region, the -10 box region and the transcription factor binding site region. The location information, attention weight distribution and regional context features of each region are obtained, and the corresponding evidence strength vector is extracted for each located region. The sequence of the -35 box region and the -10 box region is determined based on the order of the location information; the distance between the -35 box region and the -10 box region is determined based on the location information as the absolute distance; the distance rationality is determined based on the absolute distance; the pairing rationality is determined based on the peak height of the attention weight distribution of the -35 box region and the -10 box region; the spatial grammatical features are determined based on the sequence indication, absolute distance, distance rationality, and pairing rationality. By fusing spatial grammatical features, contextual features of each region, and evidence strength vectors, a structured fusion feature is obtained. Based on the structured fusion characteristics, the presence of a prokaryotic promoter in the DNA sequence to be tested can be identified.
2. The prokaryotic promoter recognition method of claim 1, wherein, After identifying the presence of a prokaryotic promoter in the DNA sequence to be tested based on the structured fusion features, the results also include: Based on the comprehensive feature representation, the presence of a prokaryotic promoter in the DNA sequence to be tested is identified as a second identification result; The second identification result is fused with the identification result based on structured fusion features to obtain the final identification result of whether a prokaryotic promoter exists in the DNA sequence to be tested.
3. The prokaryotic promoter recognition method of claim 1, wherein, The specific steps for obtaining the comprehensive feature representation of the DNA sequence to be tested are as follows: Obtain the target genome sequence, slide a window of a preset length on the target genome sequence with a preset step size, and use the sequence corresponding to each window as the DNA sequence to be tested; The DNA sequence to be tested is segmented according to the 6-mer vocabulary with a step size of 1 and converted into a token sequence; The token sequence is input into the pre-trained DNABERT model, and after processing by the embedding layer and a 12-layer Transformer encoder, global contextual hidden features are extracted. Multi-scale convolution is applied to the global contextual hidden features to enhance local features, resulting in a comprehensive feature representation.
4. The prokaryotic promoter recognition method of claim 3, wherein, The multi-scale convolutional local feature enhancement involves the following steps: Four parallel one-dimensional convolutional layers with kernel sizes of 1, 3, 5, and 9 are used to perform convolution processing on the global contextual hidden features; The outputs of the four convolutional branches are concatenated along the feature dimension, and after layer normalization and linear mapping, they are added element-wise with the global context hidden features to obtain a comprehensive feature representation.
5. The prokaryotic promoter recognition method of claim 1, wherein, The specific steps for performing region focusing processing on the comprehensive feature representation are as follows: The integrated feature representation was focused on -35 box region, -10 box region and transcription factor binding site region respectively, and the position information, attention weight distribution and regional context features of the -35 box region, -10 box region and transcription factor binding site region were obtained respectively. The -35 box region focusing and the -10 box region focusing are respectively implemented in the following ways: Get the dot product of the query vector and the key vector; Add the learnable position bias matrix to the dot product result to obtain the corrected attention score; The attention weight distribution is determined based on the corrected attention score; The position bias matrix is used to give the attention weights higher prior probabilities at preset sequence positions corresponding to the -35 box region and the -10 box region.
6. The prokaryotic promoter recognition method of claim 1, wherein, The sequence indication is as follows: if the position information of the -35 box region is before the position information of the -10 box region, the sequence indication value is 1; otherwise, the value is 0. The absolute distance is the absolute value of the difference between the position information of the -35 box region and the position information of the -10 box region. The distance reasonableness is as follows: if the absolute distance is between 15 and 25 bp, the value is 1; otherwise, the value is 0. The pairing reasonableness is the product of the peak height of the attention weight distribution of the -35 box region and the peak height of the attention weight distribution of the -10 box region.
7. The prokaryotic promoter recognition method of claim 1, wherein, The specific steps for extracting the corresponding evidence strength vector for each located region are as follows: Evidence of promoter core elements and evidence of transcription factor binding sites are extracted from the comprehensive feature representation, and corresponding strength vectors of evidence of promoter core elements and evidence of transcription factor binding sites are obtained. The evidence for the promoter core element and the evidence for the transcription factor binding site are extracted by two layers of one-dimensional convolution followed by a sigmoid activation function.
8. The prokaryotic promoter identification method as described in claim 2, characterized in that, The specific steps for fusing the second recognition result with the recognition result based on structured fusion features are as follows: The identification results based on structured fusion features and the second identification results are weighted and added together according to a preset ratio, wherein the weight of the identification results based on structured fusion features is greater than the weight of the second identification results; The weighted sum is mapped to the predicted probability of the presence or absence of a prokaryotic promoter using the softmax function.
9. A computer system, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the prokaryotic promoter identification method as described in any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, is capable of performing the prokaryotic promoter identification method as defined in any one of claims 1 to 8.
