Biological classification character related gene identification method, device and equipment

Abstract

The invention provides a biological classification character related gene identification method, device and equipment, and relates to the technical field of biological gene identification, and the method comprises the following steps: obtaining gene sequence data and image data of a target organism; a classification model is constructed, a bilinear convolutional neural network is adopted to perform feature extraction on image data to obtain macroscopic image features, a hierarchical attention mechanism model is applied to gene sequence data, and a word-level and sentence-level attention mechanism is used to extract microscopic gene features in a gene sequence; the macroscopic image features and the microscopic gene features are fused, and an incidence matrix is generated through an attention mechanism; performing convolution pooling processing on the incidence matrix through a convolutional neural network, extracting joint features, and completing a species classification task by using a multi-layer perceptron; and outputting a final species classification result, and identifying a gene most related to species classification. According to the method, the genetic characteristics of the organism can be more comprehensively understood, and the model is helped to more accurately position and analyze key genes.

Description

Technical Field

[0001] The present invention relates to the field of biological gene recognition technology, and in particular to a method, device and equipment for identifying genes related to biological classification traits. Background Art

[0002] Traditional molecular bioinformatics methods are widely used in the study of biological gene functions. For example, through RNA-Seq (RNA Sequencing) technology, scientists can comprehensively and quickly obtain transcriptome information of organisms at different developmental stages or tissues. This information helps to reveal the expression patterns of genes under different conditions and thus infer their possible functions. Once the genes of interest are discovered through technologies such as RNA-Seq, scientists usually use technologies such as RNA (Ribonucleic Acid) interference, gene knockout, or CRISPR-Cas9 (Clustered Regularly Interspaced Short Palindromic Repeats-CRISPR-associated protein 9, CRISPR-Cas9 gene editing technology) to verify the functions of these genes. These technologies can directly change the expression or structure of genes, thereby observing their effects on biological growth and development, behavioral habits, etc. Scientists can also verify which genes play a key role in the macroscopic biological processes of this species, such as growth and development, metabolic regulation, immune response, etc. These studies not only help to reveal the mysteries of biology, but also provide a scientific basis for the protection and utilization of biological resources. Take fish as an example. In the long process of evolution, freshwater fish and marine fish have emerged. They have adapted to completely different ecological environments. This adaptive evolution is not only reflected in morphological and physiological characteristics, but also has a profound impact on their genome structure and function. Due to the complex regulatory relationship between genes and the difficulty in assessing their importance in evolution, the response mechanism of different genes under environmental pressure also shows a high degree of complexity. Traditional molecular biology is difficult to fully reveal the synergistic effects of multi-gene systems under complex environmental conditions, so identifying the importance distribution of these genes is a major challenge facing traditional molecular biology.

[0003] Traditional molecular bioinformatics methods usually use RNA-Seq and other gene expression analysis to analyze gene expression in different developmental stages or tissues, and then verify the function of genes through RNA interference, gene knockout, CRISPR-Cas9 and other technologies, and then verify the genes that work on the macro level of fish; moreover, traditional functional genomics methods precisely knock out or edit target genes, observe their effects on development, and evaluate gene function by specifically inhibiting gene expression (RNA interference: RNAInterference, RNAi).

[0004] However, traditional molecular bioinformatics methods require a lot of experimental operations and data analysis, which are time-consuming and expensive. There is insufficient data support for non-model fish species, and it is difficult to fully capture the complex regulatory network of multiple genes. Multiple genes may have similar functions, and it is difficult to reveal the actual role of a single gene knockout. RNAi may cause non-specific off-target effects, affecting the accuracy of the results. Therefore, how to effectively identify key genes in different parts of organisms has become an urgent problem to be solved. Summary of the invention

[0005] In view of this, the present invention provides a method, device and equipment for identifying genes related to biological classification traits, so as to solve the defects of time-consuming, labor-intensive and high cost in the prior art of identifying the gene distribution of organisms. The identification method of the present invention, on the basis of reducing the cost of traditional molecular bioinformatics experiments, takes the association between macroscopic characteristics of biological species classification and microscopic key genes as the research object, and searches for key genes related to biological classification macroscopic traits through attention weights, so as to achieve the purpose of screening out macroscopic image blocks of appearance parts with strong correlation with biological classification based on fine-grained image classification, marking and identifying and extracting important features, and forming macroscopic feature encoding vectors; achieving the purpose of using the deep attention mechanism to screen out the genes with the largest weights related to biological classification, and using them as key genes (microscopic features) related to classification; and achieving the purpose of using the attention mechanism to establish the correspondence between the biological macroscopic feature encoding vector and the microscopic key gene encoding vector, and finding the microscopic key genes that align and match the macroscopic classification characteristics of fish.

[0006] The present invention is implemented by the following technical solutions:

[0007] In a first aspect, the present invention provides a method for identifying genes related to biological taxonomic traits, comprising the following steps:

[0008] Obtain the gene sequence data and image data of the target organism and perform preprocessing;

[0009] Constructing a classification model, using a bilinear convolutional neural network to extract features from the image data to obtain macroscopic image features, applying a hierarchical attention mechanism model to the gene sequence data, and using word-level and sentence-level attention mechanisms to extract microscopic gene features in the gene sequence;

[0010] The macroscopic image features are integrated with the microscopic gene features to generate a correlation matrix through an attention mechanism;

[0011] The association matrix is ​​subjected to convolution pooling processing by a convolutional neural network to extract joint features, and a multi-layer perceptron is used to complete the species classification task;

[0012] Output the final species classification results and identify the genes most relevant to the species classification.

[0013] As a further solution of the present invention, the target organisms are marine fish and freshwater fish with downloadable protein sequences selected from the class Actinopterygii as experimental subjects; the acquired gene sequence data and image data form a preliminary experimental data set; the gene sequence data is used to characterize the genetic information of the species, and the image data is used to reflect the macroscopic morphological characteristics of the species.

[0014] As a further solution of the present invention, when the gene sequence data is preprocessed, the genome integrity is quality assessed by Busco (Benchmarking Universal Single-Copy Orthologs), and the genes are normalized using Orthofinder (Ortholog Identification), the text data in the gene sequence data is segmented using BPE (Byte Pair Encoding), and word vectors are generated using Word2Vec (word vectors), and a multi-level word structure is constructed based on the word vectors.

[0015] As a further solution of the present invention, when the image data is preprocessed, the image data is an acquired appearance image of the target organism, and the appearance image is subjected to data cleaning, normalization and data enhancement, and is standardized, and the training data set is expanded by data enhancement technology, and the data enhancement technology includes rotation, cropping and flipping transformations.

[0016] As a further solution of the present invention, constructing a classification model includes the following steps:

[0017] Based on the acquired gene sequence data and image data of the target organism, using fine-grained image classification to filter out macroscopic image blocks from the image data;

[0018] Based on the macro image block, a macro feature coding vector is formed, wherein the macro image block is an appearance part image having a correlation with the target biological classification higher than a certain threshold value;

[0019] Using a deep attention mechanism, a key gene is selected from the gene sequence data, and the key gene is encoded to obtain a key gene vector; the key gene is a gene with the largest weight related to the target biological classification;

[0020] The attention mechanism is used to establish the corresponding relationship between the macro-feature encoding vector and the key gene vector, and determine the key gene matching the macro-feature encoding vector.

[0021] As a further solution of the present invention, based on the acquired gene sequence data and image data of the target organism, macroscopic image blocks are screened out from the image data using fine-grained image classification, and macroscopic feature encoding vectors are formed based on the macroscopic image blocks, including the following steps:

[0022] Use clustering to detect significant areas related to classification and remove the interference of irrelevant areas on the classification results;

[0023] ResNet (Residual Network) pre-trained on ImageNet and fine-tuned on a fine-grained dataset is used as a feature extractor to characterize salient areas and fuse high-level semantic features with low-level image features at multiple levels.

[0024] A bilinear convolutional neural network (B-CNN) model is used to extract and classify features of discriminative regions and associate global and local features.

[0025] As a further solution of the present invention, a deep attention mechanism is used to select key genes from the gene sequence data, and the key genes are encoded to obtain key gene vectors; the key genes are genes with the largest weights related to the target biological classification, including the following steps:

[0026] Using the word-level attention mechanism, the weight matrix a of the gene information is obtained i,t , a i,t Represents the weight of the tth word in sentence i;

[0027] Using the sentence-level attention mechanism, the weight matrix a i,t A weighted sum is performed to finally obtain the key gene vector.

[0028] As a further solution of the present invention, a bilinear convolutional neural network is used to extract features of the image data to obtain macroscopic image features, including using ResNet34 as a basic network to extract global features and local features of the image through a bilinear convolutional neural network and a basic network respectively.

[0029] As a further solution of the present invention, when using word-level and sentence-level attention mechanisms to extract micro-gene features in gene sequences, the word-level attention mechanism is used to capture the features of important positions in the gene sequence, and the sentence-level attention mechanism is used to extract overall information with classification significance in the gene sequence.

[0030] As a further solution of the present invention, the attention mechanism uses the image feature vector as the query vector (Q) and the gene feature vector as the key (K), calculates the dot product to generate an association matrix, and obtains a weighted vector (V) of the gene features through Softmax weighting.

[0031] As a further solution of the present invention, the multi-layer perceptron (MLP) is used to perform nonlinear classification on the extracted features, thereby achieving species classification, outputting species classification results, and identifying gene features related to the target biological classification task.

[0032] In a second aspect, the present invention also provides a device for identifying genes related to biological classification traits, comprising the following components:

[0033] A data acquisition module, used to acquire gene sequence data and image data of the target organism;

[0034] A feature extraction module is used to extract features from the image data based on the constructed classification model using a bilinear convolutional neural network to obtain macroscopic image features, and to apply a hierarchical attention mechanism model to the gene sequence data to extract microscopic gene features in the gene sequence using word-level and sentence-level attention mechanisms;

[0035] A feature fusion module, used to fuse the macroscopic image features with the microscopic gene features, and generate a correlation matrix through an attention mechanism;

[0036] A task classification module, used to perform convolution pooling processing on the association matrix through a convolutional neural network, extract joint features, and use a multi-layer perceptron to complete the species classification task;

[0037] The classification recognition module is used to output the final species classification results and identify the genes most relevant to the species classification.

[0038] In a third aspect, the present invention also includes a device for identifying genes related to biological classification traits, comprising: 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, and the instructions are executed by the at least one processor so that the at least one processor executes the method for identifying genes related to biological classification traits.

[0039] Compared with the prior art, the method, device and equipment for identifying genes related to biological classification traits proposed in the present invention have the following advantages:

[0040] The method, device and equipment for identifying genes related to biological classification traits provided by the present invention are suitable for biological image and gene correlation analysis, especially combining image classification with key genes and correlation matching, realizing effective alignment of image block features with key gene vectors, and finally realizing the weight relationship between the two by calculating the attention mechanism, and being able to identify the image block features and key genes with the closest correlation, thereby realizing semantic alignment between macro and micro, and finding the key genes most related to macro features, such as when a user uploads a picture of a certain part of a biological body, the model can identify the key genes related to the part. The present invention combines the image information and gene information of the organism, so that the model can understand the genetic characteristics of the organism more comprehensively, which helps the model to locate and analyze key genes more accurately, thereby reducing the dependence on traditional experimental verification, which not only reduces the experimental cost, but also shortens the research cycle.

[0041] These and other aspects of the present application will be more concise and understandable in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory and cannot limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 It is a flow chart of a method for identifying genes related to biological classification traits provided by the present invention;

[0044] Figure 2 It is a technical circuit diagram of a method for identifying genes related to biological classification traits provided by the present invention;

[0045] Figure 3 It is a flow chart of constructing a classification model in a method for identifying genes related to biological classification traits provided by the present invention;

[0046] Figure 4 It is a flow chart of forming a macro-feature coding vector in a method for identifying genes related to biological classification traits provided by the present invention;

[0047] Figure 5 It is a schematic diagram of the overall framework of BCNN in a method for identifying genes related to biological classification traits provided by the present invention;

[0048] Figure 6It is a schematic diagram of a label tree corresponding to BCNN in a method for identifying genes related to biological classification traits provided by the present invention;

[0049] Figure 7 It is a schematic diagram of a classification model in a method for identifying genes related to biological classification traits provided by the present invention;

[0050] Figure 8 It is a schematic diagram of the architecture of a hierarchical attention network in a method for identifying genes related to biological classification traits provided by the present invention;

[0051] Fig. 9 It is a schematic diagram of the architecture of the fusion of macroscopic and microscopic features in a method for identifying genes related to biological classification traits provided by the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] Since experiments and data analysis in traditional molecular bioinformatics methods are time-consuming and expensive, data support for non-model fish species is insufficient, it is difficult to fully capture the complex regulatory network of multiple genes, and traditional functional genomics methods observe the effects on development by precisely knocking out or editing target genes. Gene function is evaluated by specifically inhibiting gene expression (RNA interference (RNAi)). In view of this, the present invention provides a method, device and equipment for identifying genes related to biological classification traits, combining macroscopic (image) and microscopic (gene) to mine key genes that affect the macroscopic characteristics of organisms, analyzing and mining key genes that determine the macroscopic traits of fish classification, and converting them into attention weight problems in deep learning.

[0054] The method, device and equipment for identifying genes related to biological classification traits of the present invention use natural language processing and attention technology, combined with fine-grained image positioning and classification technology in deep learning, and regard the task of identifying key genes that affect biological classification as cross-modal retrieval and association analysis. With the powerful computing power of deep neural networks, the present invention can screen out key genes related to fish macro-evolution from massive biological genes. Biological classification can be based on morphological structure and reflected by the difference in macroscopic features. Fine-grained image classification and positioning mainly achieves classification by identifying distinguishable biological part features or macroscopic trait differences, and fine-grained image classification can generally be divided into two methods: strong supervised learning and weak supervised learning. Strong supervised learning uses bounding boxes and local annotation information to obtain the location information and size of the target, thereby improving the classification accuracy and identifying the significant features of the object in the picture. Weakly supervised learning, on the other hand, only relies on the category annotation information of the image, locates the discriminative parts through clustering and other methods, and uses auxiliary features for classification.

[0055] On the basis of reducing the cost of traditional molecular bioinformatics experiments, the present invention takes the association between macroscopic characteristics of biological species classification and microscopic key genes as the research object, and studies the key genes related to macroscopic traits of biological classification through attention weights, which can achieve the following goals:

[0056] (1) Based on fine-grained image classification, macro-image blocks of appearance parts with strong correlation with biological classification are screened out, and important features are marked and extracted to form macro-feature encoding vectors.

[0057] (2) The deep attention mechanism is used to screen out the genes with the largest weights related to biological classification, and use them as the key genes (micro features) related to classification.

[0058] (3) The attention mechanism is used to establish the correspondence between the biological macro-feature encoding vector and the micro-key gene encoding vector, and to find the micro-key genes that align and match the macro-classification characteristics of fish.

[0059] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0060] See also Figure 1 and Figure 2 As shown, one embodiment of the present invention provides a method for identifying genes related to biological classification traits, comprising the following steps:

[0061] Step S10, obtaining gene sequence data and image data of the target organism, and performing preprocessing;

[0062] Step S20, constructing a classification model, using a bilinear convolutional neural network to extract features from the image data to obtain macroscopic image features, applying a hierarchical attention mechanism model to the gene sequence data, and using word-level and sentence-level attention mechanisms to extract microscopic gene features in the gene sequence;

[0063] Step S30, fusing the macroscopic image features with the microscopic gene features, and generating a correlation matrix through an attention mechanism;

[0064] Step S40, performing convolution pooling processing on the association matrix through a convolutional neural network, extracting joint features, and using a multi-layer perceptron to complete the species classification task;

[0065] Step S50: output the final species classification result and identify the gene most relevant to the species classification.

[0066] The method for identifying genes related to biological classification traits of this embodiment, by combining the image information and gene information of the organism, enables the model to more comprehensively understand the genetic characteristics of the organism, helps the model to more accurately locate and analyze key genes, thereby reducing dependence on traditional experimental verification, which not only reduces experimental costs but also shortens the research cycle.

[0067] In this embodiment, the target organism in step S10 is selected marine fish and freshwater fish with downloadable protein sequences under the class Actinopterygii as experimental subjects; the acquired gene sequence data and image data form a preliminary experimental data set; the gene sequence data is used to characterize the genetic information of the species, and the image data is used to reflect the macroscopic morphological characteristics of the species.

[0068] In the experimental sample data collection phase of step S10, marine and freshwater fish with downloadable protein sequences under the class Actinopterygii were first selected as experimental subjects. The genome data and image data of these fish were screened and collected to form a preliminary experimental data set. The image data was mainly used to extract visual features, while the gene sequence data was used to characterize the genetic information of the species. The combination of multimodal data laid the foundation for the classification task.

[0069] In the preprocessing stage of step S10, the gene sequence data and image data need to undergo a series of cleaning and standardization processes. For gene sequence data, Busco is first used to evaluate the quality of genome integrity, and then Orthofinder is used to normalize the genes. BPE segmentation (windows = 20) is used to segment the text data in the gene sequence data, and word embedding (Word2Vec) is used to generate word vectors, and a multi-level word structure is constructed based on the word vectors. For image data, the image data is the appearance image of the acquired target organism. The appearance image is cleaned, normalized, and enhanced, and standardized. The training data set is expanded by data enhancement technology, which includes rotation, cropping, and flipping transformations. Among them, data enhancement technology is used to further expand the training data set to ensure the generalization ability of the model.

[0070] In this embodiment, see Figure 3 As shown, building a classification model includes the following steps:

[0071] Step S201: based on the acquired gene sequence data and image data of the target organism, use fine-grained image classification to filter out macroscopic image blocks from the image data;

[0072] Step S202, forming a macro feature coding vector based on the macro image block, wherein the macro image block is an appearance part image with a correlation with the target biological classification higher than a certain threshold;

[0073] Step S203: using a deep attention mechanism, selecting key genes from the gene sequence data, and encoding the key genes to obtain key gene vectors; the key genes are genes with the largest weights related to the target biological classification;

[0074] Step S204: Use the attention mechanism to establish a corresponding relationship between the macro-feature encoding vector and the key gene vector, and determine the key gene matching the macro-feature encoding vector.

[0075] By building a classification model, the macro-evolutionary information of fish is combined with micro-protein differences. For image data, B-CNN is used as the basic network, and the two branches use residual networks (ResNet34) to extract global features and local features respectively to enhance the model's ability to capture complex visual information. For gene sequence data, a hierarchical attention mechanism is introduced, including word-level and sentence-level attention modules, to capture important features in gene sequences. Through these steps, a multimodal classification model that can process both image and gene data is constructed.

[0076] In this embodiment, see Figure 4As shown, based on the acquired gene sequence data and image data of the target organism, a macroscopic image block is screened out from the image data using fine-grained image classification, and a macroscopic feature encoding vector is formed based on the macroscopic image block, including the following steps:

[0077] Step S2021: Detect significant areas related to classification by clustering, and remove interference of irrelevant areas on classification results;

[0078] Step S2022: Use ResNet pre-trained on ImageNet and fine-tuned on a fine-grained dataset as a feature extractor to characterize the salient regions and perform multi-level fusion of high-level semantic features and low-level image features;

[0079] Step S2023: Use a bilinear convolutional neural network (B-CNN) model to extract and classify features of the discriminative region, and associate global and local features.

[0080] Among them, the present invention is a weakly supervised fine-grained classification method based on salient region detection and multi-level, multi-scale feature fusion. It uses clustering to detect salient regions related to classification, and then optimizes the detection results of object-level salient regions through a saliency detection model, focusing on discriminative regions and reducing the interference of irrelevant regions on classification results. At the same time, ResNet, which is pre-trained on ImageNet and fine-tuned on a fine-grained dataset, is used as a feature extractor to characterize the salient regions, and high-level semantic features are multi-level fused with underlying image features to increase feature diversity. Finally, a bilinear convolutional neural network (B-CNN) model is used to extract and classify features of discriminative regions, such as Figure 5 and Figure 6 As shown in , global and local features are associated to improve the classification accuracy of fine-grained images, such as Figure 7 shown.

[0081] In this embodiment, a deep attention mechanism is used to select key genes from the gene sequence data, and the key genes are encoded to obtain key gene vectors; the key genes are genes with the largest weights related to the target biological classification, including the following steps:

[0082] Using the word-level attention mechanism, the weight matrix a of the gene information is obtained i,t , a i,t Represents the weight of the tth word in sentence i;

[0083] Using the sentence-level attention mechanism, the weight matrix a i,t A weighted sum is performed to finally obtain the key gene vector.

[0084] In this embodiment, a bilinear convolutional neural network is used to extract features from the image data to obtain macroscopic image features, including using ResNet34 as a basic network to extract global features and local features of the image through a bilinear convolutional neural network and a basic network respectively.

[0085] In step S20, when word-level and sentence-level attention mechanisms are used to extract micro-gene features in gene sequences, the word-level attention mechanism is used to capture features of important positions in gene sequences, and the sentence-level attention mechanism is used to extract overall information with classification significance in gene sequences.

[0086] In the feature extraction and fusion stages of step S20 and step S30, feature vectors extracted from image data and gene data are processed respectively. In order to fully utilize the complementarity of the two types of data, a macroscopic and microscopic fusion strategy is adopted. Macroscopic features are compared at a high level through the overall representation of images and genes, while microscopic features are used to deeply explore the intrinsic connection between the two types of data through fine-grained feature matching.

[0087] Among them, the attention mechanism uses the image feature vector as the query vector (Q) and the gene feature vector as the key (K), calculates the dot product to generate the association matrix, and obtains the weighted vector (V) of the gene features through Softmax weighting.

[0088] See also Figure 8 As shown, this embodiment uses a multi-level attention mechanism to construct a hierarchical attention classification network, effectively extracts important information at each level in the gene sequence, and achieves accurate classification. Among them, the input gene sequence is converted into a word embedding representation, and then a high-order hidden representation is obtained through a recurrent neural network, and then the mining of the sequence semantic information is gradually deepened through a series of attention layers. Each attention layer processes the sequence in layers and assigns corresponding attention weights to each layer, so that the model can focus on key information in different layers. Through this hierarchical attention mechanism, the semantic structure of the gene sequence can be captured at different detail granularities, so as to understand the meaning of the sequence at a deeper level and complete the classification task. After extracting the key information, the information is integrated through pooling, and finally classified prediction is performed through the fully connected layer. The design of this hierarchical attention network enables it to efficiently process sequences of different types and lengths and capture complex semantic information.

[0089] The hierarchical attention network (HAN) includes a word encoder layer, a word attention layer, a sentence encoder layer, and a sentence attention layer. The word encoder layer passes data through BiLSTM to obtain an abstract hidden layer representation of each data segment. In order to measure the importance of words, the present invention uses u i,tand a randomly initialized context vector u w The similarity is expressed as , and then a normalized attention weight matrix a is obtained through softmax operation i,t , represents the weight of the tth word in sentence i, the attention mechanism of the word vector. After obtaining the Wordattention weight matrix, the sentence vector can be regarded as the weighted sum of these word vectors. The BiLSTM neural network is used to combine the forward and reverse context information to obtain the sentence context hidden layer output. Similar to the word-level attention, for the document in the Sentence Attention layer, the importance of a sentence relative to the document is also measured by the sentence-level context vector.

[0090] In this embodiment, the generated association matrix can reflect the association between fish gene sequences and image features, and the joint classification of fish images and their corresponding gene sequences can be achieved through this association matrix. For example, when a user uploads a picture of a fish eye, the model can identify the type of fish and associate it with its protein gene sequence information, thus achieving multi-dimensional classification and analysis of fish images and gene data.

[0091] Among them, macroscopic images and microscopic gene expression are two different ways of expressing information, each with unique characteristics, and there is a close connection between the two. Microscopic genes, as internal driving factors, determine the phenotype of organisms, while macroscopic images are manifested as external phenotypic results. Genes control the classification and specific phenotypes of organisms. The present invention can determine the key genes most relevant to the classification of the image by giving fish eye pictures. The method adopts an LSTM model based on CNN and hierarchical attention to fuse macroscopic image features with microscopic gene information. Specifically, the bilinear convolutional neural network of ResNet34 is used to extract the regional features of the image, and normalize and vectorize them. At the same time, LSTM is used to extract gene sequence features and generate corresponding vectors. Then, the macroscopic image feature vector is used as the query (Q), the gene feature vector is used as the key (K), the product matrix is ​​generated by dot product, and the weighted gene vector (V) is obtained by Softmax processing. The association between macroscopic features and microscopic genes is realized, and these associations are quantified through the attention mechanism. Next, CNN is used to perform convolution pooling on the correlation matrix to further fuse macro and micro information, and finally the classification output is performed through the multi-layer perceptron (MLP), and the classification result is generated through the Softmax layer. However, the purpose of this project is not a simple biological classification, but to find the micro key genes that are most related to the macro image features through the attention mechanism. Finally, the attention fusion of the macro and micro levels is achieved, and the technical roadmap of the project is finally formed, such as Fig. 9 shown.

[0092] In this embodiment, the multi-layer perceptron (MLP) is used to perform nonlinear classification on the extracted features, thereby achieving species classification, outputting species classification results, and identifying gene features related to the target biological classification task.

[0093] The method for identifying genes related to biological classification traits of the present invention is suitable for biological image and gene correlation analysis, especially combining image classification with key genes and correlation matching, realizing effective alignment of image block features and key gene vectors, and finally realizing the weight relationship between the two by calculating the attention mechanism, and being able to identify the image block features and key genes with the closest correlation, thereby realizing semantic alignment between macro and micro, and finding the key genes most related to the macro features. For example, when a user uploads a picture of a certain part of a biological body, the model can identify the key genes related to the part.

[0094] It should be understood that, although described in a certain order, these steps are not necessarily performed in sequence in the above order. Unless there is clear explanation in this article, the execution of these steps does not have strict order restriction, and these steps can be performed in other orders. Moreover, a part of the steps of the present embodiment may include a plurality of steps or a plurality of stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a part of the steps or stages in other steps or other steps.

[0095] In one embodiment, the present invention provides a device for identifying genes related to biological classification traits, comprising the following components:

[0096] A data acquisition module, used to acquire gene sequence data and image data of the target organism;

[0097] A feature extraction module is used to extract features from the image data based on the constructed classification model using a bilinear convolutional neural network to obtain macroscopic image features, and to apply a hierarchical attention mechanism model to the gene sequence data to extract microscopic gene features in the gene sequence using word-level and sentence-level attention mechanisms;

[0098] A feature fusion module, used to fuse the macroscopic image features with the microscopic gene features, and generate a correlation matrix through an attention mechanism;

[0099] A task classification module, used to perform convolution pooling processing on the association matrix through a convolutional neural network, extract joint features, and use a multi-layer perceptron to complete the species classification task;

[0100] The classification recognition module is used to output the final species classification results and identify the genes most relevant to the species classification.

[0101] In this embodiment, the biological classification trait-related gene identification device of the present invention adopts the steps of the aforementioned biological classification trait-related gene identification method during execution. Therefore, the operation process of the biological classification trait-related gene identification device in this embodiment will not be described in detail.

[0102] In one embodiment, a device for identifying genes related to biological classification traits is also provided in an embodiment of the present invention, comprising at least one processor and a memory communicatively connected to the at least one processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the at least one processor so that the at least one processor executes a method for identifying genes related to biological classification traits, and the processor implementing the steps of the above-mentioned method for identifying genes related to biological classification traits when executing the instructions.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying genes related to biological taxonomic traits, characterized in that: include: Obtain the gene sequence data and image data of the target organism and perform preprocessing; Constructing a classification model, using a bilinear convolutional neural network to extract features from the image data to obtain macroscopic image features, applying a hierarchical attention mechanism model to the gene sequence data, and using word-level and sentence-level attention mechanisms to extract microscopic gene features in the gene sequence; The macroscopic image features are integrated with the microscopic gene features to generate a correlation matrix through an attention mechanism; The association matrix is ​​subjected to convolution pooling processing by a convolutional neural network to extract joint features, and a multi-layer perceptron is used to complete the species classification task; Output the final species classification results and identify the genes most relevant to the species classification.

2. The method for identifying genes related to biological taxonomy traits according to claim 1, characterized in that: The target organisms are marine and freshwater fish with downloadable protein sequences selected from the class Actinopterygii as experimental subjects; the acquired gene sequence data and image data form a preliminary experimental data set; the gene sequence data is used to characterize the genetic information of the species, and the image data is used to reflect the macroscopic morphological characteristics of the species.

3. The method for identifying genes related to biological taxonomy traits according to claim 2, characterized in that: When preprocessing the gene sequence data, Busco is used to evaluate the quality of genome integrity, Orthofinder is used to normalize genes, BPE word segmentation is used to segment text data in the gene sequence data, Word2Vec is used to generate word vectors, and a multi-level word structure is constructed based on the word vectors.

4. The method for identifying genes related to biological taxonomy traits according to claim 3, characterized in that: Building a classification model includes the following steps: Based on the acquired gene sequence data and image data of the target organism, using fine-grained image classification to filter out macroscopic image blocks from the image data; Based on the macro image block, a macro feature coding vector is formed, wherein the macro image block is an appearance part image having a correlation with the target biological classification higher than a certain threshold value; Using a deep attention mechanism, a key gene is selected from the gene sequence data, and the key gene is encoded to obtain a key gene vector; the key gene is a gene with the largest weight related to the target biological classification; The attention mechanism is used to establish the corresponding relationship between the macro-feature encoding vector and the key gene vector, and determine the key gene matching the macro-feature encoding vector.

5. The method for identifying genes related to biological taxonomy traits according to claim 4, characterized in that: Based on the acquired gene sequence data and image data of the target organism, a macroscopic image block is screened out from the image data using fine-grained image classification, and a macroscopic feature encoding vector is formed based on the macroscopic image block, including the following steps: Use clustering to detect significant areas related to classification and remove the interference of irrelevant areas on the classification results; ResNet, which is pre-trained on ImageNet and fine-tuned on a fine-grained dataset, is used as a feature extractor to represent the salient regions and fuse high-level semantic features with low-level image features at multiple levels. A bilinear convolutional neural network model is used to extract and classify features of discriminative regions, correlating global and local features.

6. The method for identifying genes related to biological taxonomy traits according to claim 4, characterized in that: Using a deep attention mechanism, key genes are selected from the gene sequence data, and the key genes are encoded to obtain a key gene vector; the key gene is a gene with the largest weight related to the target biological classification, including the following steps: Using the word-level attention mechanism, the weight matrix a of the gene information is obtained i,t , a i,t represents the weight of the tth word in sentence i; Using the sentence-level attention mechanism, the weight matrix a i,t A weighted sum is performed to finally obtain the key gene vector.

7. The method for identifying genes related to biological taxonomy traits according to claim 1, characterized in that: A bilinear convolutional neural network is used to extract features from the image data to obtain macroscopic image features, including using ResNet34 as a basic network to extract global features and local features of the image through a bilinear convolutional neural network and a basic network respectively.

8. The method for identifying genes related to biological taxonomy traits according to claim 7, characterized in that: When using word-level and sentence-level attention mechanisms to extract microgene features in gene sequences, the word-level attention mechanism is used to capture the features of important positions in the gene sequence, and the sentence-level attention mechanism is used to extract overall information with classification significance in the gene sequence.

9. A device for identifying genes related to biological classification traits, characterized in that: The device is used to perform the method for identifying genes related to biological classification traits as described in any one of claims 1 to 8, and the identification device comprises: A data acquisition module, used to acquire gene sequence data and image data of the target organism; A feature extraction module is used to extract features from the image data based on the constructed classification model using a bilinear convolutional neural network to obtain macroscopic image features, and to apply a hierarchical attention mechanism model to the gene sequence data to extract microscopic gene features in the gene sequence using word-level and sentence-level attention mechanisms; A feature fusion module, used to fuse the macroscopic image features with the microscopic gene features, and generate a correlation matrix through an attention mechanism; A task classification module, used to perform convolution pooling processing on the association matrix through a convolutional neural network, extract joint features, and use a multi-layer perceptron to complete the species classification task; The classification recognition module is used to output the final species classification results and identify the genes most relevant to the species classification.

10. A device for identifying genes related to biological classification traits, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method for identifying genes related to biological classification traits as described in any one of claims 1 to 8 is implemented.

Images