Insect multi-tissue expression amount prediction method and system based on multi-modal feature fusion

Abstract

The present application relates to a kind of based on multi-modal feature fusion insect multi-tissue expression quantity prediction method and system, with target gene as basic prediction unit, while inputting the regulatory sequence information, RNA stability characteristic information and protein representation information of target gene, carry out joint modeling by multi-tower deep learning network, output the expression quantity prediction value of the target gene in multiple tissues.Advantage: break through the limitation of traditional method mainly relying on single mode data, first embed uniform into an expression prediction framework for insect gene regulatory sequence, RNA stability characteristic and pre-training protein language model;Using multi-modal prediction model, the multi-modal data features are extracted, fused and predicted, so as to provide accurate and comprehensive gene expression prediction results, while introducing a gating back mechanism for missing protein annotation, avoiding missing values being treated as zero or noise, so as to ensure the stability of prediction.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and artificial intelligence, and in particular to a method and system for predicting the expression levels of multiple tissues in insects based on multimodal feature fusion. Background Technology

[0002] Gene expression level refers to the abundance of RNA produced by a gene transcribed in a specific tissue, cell, or developmental stage. For insects, tissue-resolution gene expression profiles are not only an important foundation for functional gene research, developmental regulation analysis, and comparative transcriptomics, but also directly affect the efficiency of subsequent candidate gene screening, molecular validation experimental design, and cross-species functional inference.

[0003] Currently, insect gene expression levels are primarily determined using RNA-seq, or high-throughput RNA sequencing technology. This technique typically involves steps such as sample collection, RNA extraction, library construction, sequencing, alignment, and quantification, which is costly, time-consuming, and requires high sample quality and specific tissue acquisition conditions. For model insects, relatively complete expression profiles can be obtained from public databases; however, for a large number of agricultural pests, non-model insects, and newly sequenced species, only reference genomes and preliminary annotations are often available, lacking systematic, multi-tissue, and multi-developmental stage transcriptome data. This has made the construction of tissue expression profiles reliant on high-cost experiments for a long time, limiting related research and applications.

[0004] In recent years, artificial intelligence, especially deep learning, has made rapid progress in predicting gene expression from DNA sequences. A representative approach is to use regulatory sequences near the target gene as input, allowing the model to automatically learn the cis-regulatory rules affecting transcriptional activity. Here, "cis-regulation" refers to DNA elements located near the target gene that directly affect its transcription initiation and intensity, such as promoters, enhancers, and repressor elements. Existing models have shown that using only DNA sequences can indeed partially predict expression levels, but the inventors have found that this approach still has significant limitations in predicting expression in multiple tissues of insects.

[0005] First, most existing methods rely heavily on single-modal inputs, often using only DNA sequences and neglecting RNA stability factors closely related to expression levels, as well as cross-species functional conservation information at the protein level. A "modality" refers to data types from different sources and in different forms, such as DNA sequence modalities, RNA structure / stability feature modalities, and protein representation modalities. Single-modal inputs lead to incomplete characterization of the actual expression regulatory mechanisms by the model.

[0006] Second, existing methods are mostly geared towards mammalian, plant, or single-species data, and are not well adapted to insect data, especially cross-species insect data. Insect species vary significantly in gene annotation quality, tissue coverage integrity, and the degree of conservation of homologous genes. Directly applying expression prediction models from other groups often fails to achieve stable results.

[0007] Third, existing methods generally struggle to simultaneously address the issues of "multi-tissue output" and "missing labels." "Multi-tissue output" refers to a model predicting the expression level of a gene in multiple tissues simultaneously, rather than predicting only a single tissue or a single average expression value. For publicly available insect data, tissue coverage varies across species, and many species have only had a few tissues measured. Simply treating missing tissue labels as zero values ​​can misclassify "unmeasured" as "low expression," introducing systematic errors.

[0008] Fourth, many existing schemes do not incorporate modeling of protein functional conservation. In fact, cross-species expression patterns in insects are not solely determined by DNA regulatory sequences; the functional conservation information carried by protein-coded products can also help models understand the commonalities in expression regulation between homologous but different species genes. Recent protein language models can learn high-dimensional representations from massive protein sequences. These representations can reflect protein structure, function, and evolutionary relationships to some extent, but existing expression prediction patents and schemes rarely directly combine them with insect multi-tissue expression prediction.

[0009] Fifth, protein embeddings are not fully available for all genes in real-world data. Some genes lack reliable protein annotations or have missing directly translatable protein sequences. Simply replacing missing protein information with all-zero vectors can cause the model to mistakenly interpret "missing" information as "a special real signal," thus creating a systematic bias.

[0010] Therefore, this case is brought. Summary of the Invention

[0011] The purpose of this invention is to provide a method and system for predicting the expression levels of multiple tissues in insects based on multimodal feature fusion. It aims to address technical issues such as joint prediction of multiple tissues, handling of missing tags, regression of missing protein information, and cross-species generalization by integrating DNA regulatory sequences, RNA stability features, and pre-trained protein language model embeddings in insect scenarios.

[0012] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting multi-tissue expression levels in insects based on multimodal feature fusion includes the following steps: S1. Acquire multimodal input data, which includes reference genome sequences of multiple insect species, gene structure annotation files, protein sequence files, and tissue expression data; S2. Map the publicly available data organization names of different insect species to a standard organization set. If a certain insect species lacks a true measurement value of a standard organization, mark it as a missing value. S3. Determine the transcription start site of each target gene based on the gene structure annotation file, extract a genomic sequence window of a preset length from the reference genome sequence with the transcription start site as the center, use it as the target gene regulatory sequence, and convert the target gene regulatory sequence into a sequence modality input matrix; S4. Based on the gene structure annotation file, extract the RNA stability characteristics of the target gene from the reference genome sequence; S5. For a target gene with a protein sequence file, input its protein amino acid sequence into a pre-trained protein language model to generate a protein embedding vector; For target genes missing in protein sequence files, species embedding vectors are used as protein embedding vectors. The species embedding vector is a fixed-dimensional vector learned for each insect species to represent the overall background features of that insect species. S6. Input the sequence modality input matrix, RNA stability features, and protein embedding vector into a pre-trained multimodal prediction model, and use the feature extraction network and feature fusion network in the multimodal prediction model to extract and fuse features to obtain a fusion vector; S7. The obtained fusion vector is input into the prediction network of the multimodal prediction model, and the output is the predicted expression level of the target gene in the standard tissue set.

[0013] Furthermore, in step S1, the reference genome sequence is in FASTA format; The gene structure annotation file is in GFF or GTF format; The protein sequence file is in FASTA format; The tissue expression data were derived from RNA-seq quantification results and uniformly converted to TPM.

[0014] Furthermore, the tissue expression data can be converted to log2(TPM+1).

[0015] Furthermore, in step S2, the standard tissue set adopts 14 standard tissues, including adult brain, adult head, adult midgut, adult hindgut, adult fat body, adult Malpighian tubules, adult body, adult ovary, adult testis, larval hindgut, larval fat body, larval Malpighian tubules, larval midgut, and larval body.

[0016] Furthermore, in step S3, the genome sequence window adopts an asymmetric window with a length of 20,000 bases, of which 15,000 bases are upstream and 5,000 bases are downstream.

[0017] Furthermore, in step S3, the target gene regulatory sequence is converted into a coding matrix through one-hot encoding, which serves as the sequence modality input matrix.

[0018] Furthermore, in step S4, the RNA stability characteristics are expressed as a 16-dimensional feature vector, including the length of the 5' untranslated region, the length of the 3' untranslated region, the length of the coding region, the GC content, the number of introns, the total length of introns, the number of exons, the codon fitness index, the number of AU-rich elements, the Kozak sequence strength, the GC content at the third codon position, the proportion of rare codons, the GC content of the 5' untranslated region, the GC content of the 3' untranslated region, the total gene length, and the length of mature RNA.

[0019] Furthermore, in step S6, the feature extraction and fusion includes the following steps: S61. Perform multi-stage convolutional downsampling on the sequence modality input matrix to obtain the target gene regulatory sequence features; S62. Add position encoding to the obtained target gene regulatory sequence features, learn the long-range dependencies between different sequence positions, and output the encoded target gene regulatory sequence features containing global context information; S63. Perform global average pooling on the encoded target gene regulatory sequence features to obtain a target gene regulatory sequence representation vector that conforms to the preset dimensions; S64. Input the RNA stability features into a two-layer multilayer perceptron network. The two layers of the multilayer perceptron network are processed by layer normalization and GELU activation function to obtain an RNA stability feature vector that conforms to the preset dimensions. The multilayer perceptron network layer adopts a feedforward neural network composed of several fully connected layers. S65. Input the protein embedding vector into a two-layer multilayer perceptron network. The two layers of the multilayer perceptron network are processed by layer normalization and GELU activation function to obtain a protein embedding vector that conforms to the preset dimensions. S66. The target gene regulatory sequence representation vector, the RNA stability feature vector, and the protein embedding vector that meet the preset dimensions are concatenated to obtain a concatenated vector. The concatenated vector is then input into a two-layer fusion network. Layer normalization, GELU activation function, and dropout network are used to process the two fusion networks to obtain the fusion vector.

[0020] Furthermore, the multimodal prediction model also includes a species-tissue bias term, which learns and stores a bias parameter of a preset size for each insect species and each standard tissue to compensate for the overall expression baseline differences of different insect species in different standard tissues. Step S7 specifically includes the following processes: S71. Using the prediction network of a multimodal prediction model, generate a pre-correction expression level prediction for each standard tissue; S72. Based on the predicted expression level before correction, obtain the bias parameters corresponding to the insect species and standard tissue in the species-tissue bias term; S73. The corresponding pre-correction expression level prediction value is superimposed with the bias parameter to obtain the post-correction expression level prediction value; S74. Aggregate all corrected expression level predictions and output the target gene's expression level prediction in the standard tissue set.

[0021] A multi-tissue expression level prediction system for insects based on multimodal feature fusion for use with the above method includes: The multi-species data acquisition module is used to acquire reference genome sequences, gene structure annotation files, protein sequence files, and tissue expression data for multiple insect species. The organization standardization module is used to uniformly map the publicly available data organization names of different insect species to a standard organization set. If a certain insect species lacks the actual measurement value of the standard organization, it is marked as a missing value. The transcription start site identification module is used to determine the transcription start site of each target gene based on the gene structure annotation file; The DNA sequence construction module is used to extract a genomic sequence window of a preset length from the reference genomic sequence as the target gene regulatory sequence, and to convert the target gene regulatory sequence into a sequence mode input matrix. The RNA stability feature construction module is used to extract RNA stability features of target genes based on gene structure annotation files. The protein embedding construction module is used to generate protein embedding vectors based on the protein sequence file using a pre-trained protein language model, and to use the species embedding vector as the protein embedding vector when the target gene is missing from the protein sequence file. The feature extraction network module includes a sequence tower, an RNA feature tower, and a protein embedding tower. The sequence tower comprises a convolutional downsampling module, a Transformer encoding module, and a global average pooling layer connected in sequence. The convolutional downsampling module performs multi-stage convolutional downsampling on the sequence modality input matrix to obtain target gene regulatory sequence features. The Transformer encoding module learns the long-range dependencies between different sequence positions on the position-encoded target gene regulatory sequence features and outputs encoded target gene regulatory sequence features containing global context information. The global average pooling layer performs global average pooling on the encoded target gene regulatory sequence features to obtain a target gene regulatory sequence representation vector conforming to a preset dimension. The fusion network module is used to splice and fuse the target gene regulatory sequence representation vector that conforms to the preset dimensions, the RNA stability feature vector that conforms to the preset dimensions, and the protein embedding vector that conforms to the preset dimensions to obtain the fusion vector; The prediction network module is used to output the predicted expression level of the target gene in a standard tissue set based on the fusion vector.

[0022] Compared with the prior art, the present invention has the following advantages: 1. This method overcomes the limitations of traditional prediction methods that primarily rely on single-modal data. For the first time, it integrates insect gene regulatory sequences, RNA stability features, and pre-trained protein language models into a unified expression prediction framework. Utilizing a multimodal prediction model, it extracts, fuses, and predicts multimodal data features, providing more accurate and comprehensive gene expression prediction results. Furthermore, this method can simultaneously output expression levels from multiple tissues, making it suitable for insect tissue-resolution expression profile reconstruction, rather than being limited to single tissue or overall average expression. The multimodal prediction model demonstrated in the examples significantly outperforms the ablation model, indicating that multimodal fusion is not a simple splicing but a collaborative modeling scheme with clearly improved technical performance.

[0023] 2. A gating backoff mechanism for missing protein annotations is introduced. When faced with target genes with missing protein information, the system can introduce species embedding vectors as protein embedding vectors, avoiding the mistreatment of missing values ​​as zero values ​​or noise, thereby ensuring the stability of predictions.

[0024] 3. In the feature extraction stage of the target gene regulatory sequence, a strategy combining multi-stage convolutional downsampling and Transformer encoding is adopted to learn the long-range dependencies between different sequence positions. This enables the model to not only identify local regulatory motifs, but also learn the interactions between long-distance regulatory fragments, making the proposed solution more suitable for real-world application scenarios involving insects across species, multiple tissues, and incomplete tags. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of the architecture of the insect multi-tissue expression level prediction system based on multimodal feature fusion in the embodiment. Figure 2 This is a flowchart illustrating the insect multi-tissue expression level prediction method based on multimodal feature fusion in the embodiments. Figure 3 This is a schematic diagram of the protein embedding gating backoff mechanism in the embodiment; Figure 4 This is a schematic diagram illustrating the training process of the multimodal prediction model in the embodiment; Figure 5 This is a schematic diagram comparing the ablation effects after using the prediction system and method of the examples, used to show the differences in prediction indicators between the complete model, RNA+ESM only, ESM removal, RNA removal, and DNA only models. Detailed Implementation

[0027] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] This embodiment proposes a method for predicting the expression levels of insects in multiple tissues based on multimodal feature fusion. This method uses the target gene as the basic prediction unit and simultaneously inputs the regulatory sequence information, RNA stability feature information, and protein representation information of the target gene. It performs joint modeling through a multi-tower deep learning network and outputs the predicted expression levels of the target gene in multiple tissues.

[0029] The following is an explanation of some technical terms that may be involved in the embodiments.

[0030] 1. Transcription start site: The TSS is the starting point where a gene begins to be transcribed into RNA.

[0031] 2. One-hot encoding: This is a method to convert discrete base sequences into matrices. For example, A is encoded as [1,0,0,0], C as [0,1,0,0], G as [0,0,1,0], and T as [0,0,0,1].

[0032] 3. RNA stability characteristics: These are numerical features that reflect the degradation rate, translation efficiency, and structural properties of RNA molecules, such as the length of the untranslated region, GC content, number of introns, and codon bias.

[0033] 4. Protein Language Model: This refers to a neural network model that pre-trains using the amino acid sequence of a protein as a "language," enabling it to learn a high-dimensional vector representation of the protein from the amino acid sequence. ESM (Evolutionary Scale Modeling) is a representative protein language model; in the examples below, ESM-2 is used to generate protein embeddings.

[0034] 5. Embedding: refers to converting the original sequence into a continuous numerical vector representation that can be further processed by a neural network.

[0035] 6. Gating mechanism: refers to the mechanism by which the model selects or weights and switches between two or more representations based on the input conditions.

[0036] 7. Mask loss: This refers to calculating the loss only for locations with true labels during training, and not calculating the loss for locations with missing labels.

[0037] 8. Multi-tissue expression prediction: This refers to the model outputting the expression level of the target gene in multiple tissues at once, rather than training multiple single-tissue models separately.

[0038] like Figure 2 As shown, the prediction method in this embodiment includes the following steps.

[0039] Step S1 involves acquiring multimodal input data, including reference genome sequences, gene structure annotation files, protein sequence files, and tissue expression data for multiple insect species. The reference genome sequences are in FASTA format, the gene structure annotation files are in GFF or GTF format, and the protein sequence files are from NCBI RefSeq / GenBank proteomics data, in FASTA format (.faa files). By using these formats, the input data is compatible with the standard data interfaces of mainstream bioinformatics databases, ensuring that the model can directly utilize existing public genomic resources for efficient and accurate multimodal feature extraction and modeling.

[0040] Tissue expression data were preferably derived from RNA-seq quantification results and uniformly converted to TPM, which stands for TPM (Transcriptional Expression Per Million Transcripts). To compress extremely high expression values ​​and improve model training stability, tissue expression levels can be further converted to log2(TPM+1).

[0041] Step S2: Since the names of publicly available data organizations for different insect species are inconsistent, this embodiment maps the names of publicly available data organizations for different insect species to a standard organization set. This eliminates semantic ambiguity caused by differences in naming habits in multi-species data sources, enabling the subsequent model prediction steps to map similar organization features of different insect species to the same set of output nodes, thereby laying the foundation for cross-species feature sharing and parameter learning.

[0042] In this embodiment, the standard tissue set uses 14 standard tissues, including adult brain, adult head, adult midgut, adult hindgut, adult fat body, adult Malpighian tubules, adult body, adult ovary, adult testis, larval hindgut, larval fat body, larval Malpighian tubules, larval midgut, and larval body.

[0043] In this step, if a certain insect species lacks a true measurement value of a standard tissue, it is marked as a missing value instead of a zero value, thus avoiding the introduction of systematic bias.

[0044] Step S3: Determine the transcription start site of each target gene based on the gene structure annotation file. Using the transcription start site as the center, extract a pre-defined genomic sequence window from the reference genome sequence as the target gene regulatory sequence. Preferably, this embodiment uses an asymmetric window of 20,000 bases, with 15,000 bases upstream and 5,000 bases downstream. The purpose of setting the asymmetric window is to allow the model to more fully cover the cis-regulatory region upstream of the transcription start site while preserving local sequence information downstream of the start site.

[0045] The extracted target gene regulatory sequences are converted into a 4×20000 matrix using one-hot encoding, which serves as the sequence modality input matrix.

[0046] Step S4: Based on the gene structure annotation file, extract the RNA stability features of the target gene from the reference genome sequence. The RNA stability features are presented as a 16-dimensional feature vector, including the length of the 5' untranslated region, the length of the 3' untranslated region, the length of the coding region, GC content, number of introns, total intron length, number of exons, codon fitness index, number of AU-rich elements, Kozak sequence strength, GC content at the third codon position, proportion of rare codons, GC content in the 5' untranslated region, GC content in the 3' untranslated region, total gene length, and mature RNA length.

[0047] The untranslated region (UTR) refers to the region of RNA that is not directly translated into protein. The coding region length refers to the total number of nucleotides (bases) in the DNA or RNA sequence of a gene that are responsible for encoding the amino acid sequence of a protein. GC content refers to the percentage of guanine (G) and cytosine (C) bases in a DNA or RNA sequence; it is a core indicator of nucleic acid sequence composition and has a significant impact on the physical properties and biological functions of DNA. Exons are the parts of a gene that are retained in mature mRNA and are ultimately translated into protein. Introns are non-coding sequences sandwiched between two exons. The codon fitness index is an indicator reflecting codon usage preference. The number of AU-rich elements refers to the number of times a specific adenine (A) and uracil (U)-rich repetitive sequence motif appears in the mRNA sequence of a target gene. The Kozak sequence is a conserved sequence near translation initiation and affects translation initiation efficiency.

[0048] Among the above dimensions, logarithmic transformation is preferred for length-type features to make features of different dimensions comparable.

[0049] Step S5: For the target gene with a protein sequence file, input its protein amino acid sequence into a pre-trained protein language model to generate a protein embedding vector. Preferably, the ESM-2 model is used to generate a 320-dimensional protein embedding vector.

[0050] As mentioned in the background section, some genes lack reliable protein annotations or have missing directly translatable protein sequences. Simply replacing missing protein information with an all-zero vector can cause the model to mistakenly interpret "missing" information as "a specific, real signal," leading to systematic bias. For target genes with missing protein sequence files, a species embedding vector is used as the protein embedding vector. This species embedding vector is a fixed-dimensional vector learned for each insect species, representing the overall background characteristics of that species. Figure 3 As shown, for samples lacking protein embeddings, a gating mechanism is used to switch from protein representation to species embedding representation instead of directly using all-zero vectors, which can reduce systematic bias caused by missing data.

[0051] Step S6: Input the sequence modality input matrix, RNA stability features, and protein embedding vector into the pre-trained multimodal prediction model, and use the feature extraction network and feature fusion network in the multimodal prediction model to extract and fuse features to obtain a fusion vector.

[0052] In this embodiment, as Figure 1 and Figure 2As shown, the multimodal prediction model adopts a three-tower multimodal prediction model, that is, a three-tower deep learning architecture, including a sequence tower, an RNA feature tower, and a protein embedding tower.

[0053] Step S6 includes the following specific steps.

[0054] S61. The sequence tower receives a 4×20000 one-hot sequence modality input matrix as input. In a preferred embodiment, the sequence tower first passes through a seven-stage convolutional downsampling module. The first stage uses four parallel one-dimensional convolutional branches with kernel sizes preferably 7, 11, 15, and 21, each branch outputting 32 channels. The four branches are concatenated to obtain 128 channels, which are then subjected to 2x max pooling. The second stage uses a one-dimensional convolutional layer with a kernel size of 7 and performs 2x max pooling. The third stage uses five dilated residual blocks with dilation rates of 1, 2, 4, 8, and 16, respectively, to expand the receptive field without significantly increasing the number of parameters and to learn longer-range local regulatory combinations. The fourth to sixth stages use convolutional and pooling structures, with the fourth stage increasing the number of channels from 128 to 256. The seventh stage performs the final pooling operation. After the above processing, the original 20000-byte sequence is compressed into a target gene regulatory sequence feature of approximately 312 bytes in length and 256 channels.

[0055] S62. Positional encoding is added to the obtained target gene regulatory sequence features, and the results are input into a multi-layer Transformer encoding module to learn the long-range dependencies between different sequence positions. The output is the encoded target gene regulatory sequence features containing global context information, enabling the multimodal prediction model to not only identify local regulatory motifs but also learn the interactions between long-distance regulatory fragments. In this embodiment, the Transformer encoding module includes 4 layers of encoding blocks, each with 8 attention heads, a model dimension of 256, and a feedforward network intermediate dimension of 1024.

[0056] S63. Global average pooling is performed on the characteristics of the encoded target gene regulatory sequence to obtain a 256-dimensional target gene regulatory sequence representation vector.

[0057] In S64. The RNA feature tower, 16-dimensional RNA stability features are input into a two-layer multilayer perceptron network, with a preferred structure of 16→128→64. Layer normalization and GELU activation function are used in the middle to process the features, resulting in a 64-dimensional RNA stability feature vector. In this embodiment, the multilayer perceptron is a feedforward neural network composed of several fully connected layers, used to process fixed-length numerical feature vectors.

[0058] S65. In the protein embedding tower, the 320-dimensional protein embedding is input into a two-layer multilayer perceptron network, with the preferred structure being 320→128→64. Layer normalization and GELU activation function are used for processing in between, and a 64-dimensional protein embedding vector is output.

[0059] S66. The 256-dimensional vector output from the sequence tower, the 64-dimensional vector output from the RNA feature tower, and the 64-dimensional vector output from the protein embedding tower are concatenated to form a 384-dimensional concatenated vector. The concatenated vector is then input into a two-layer fusion network. The preferred structure is 384→512→256. Layer normalization, GELU activation function, and dropout network are used in the middle to process the vector and obtain the fusion vector.

[0060] In step S7, the obtained fusion vector is input into the prediction network of the multimodal prediction model, and the predicted expression level of the target gene in the standard tissue set is output through the linear output layer.

[0061] Preferably, the multimodal prediction model also includes a species-tissue bias term, wherein the species-tissue bias term learns and stores a bias parameter of a preset size for each insect species and each standard tissue, which is used to compensate for the overall expression baseline differences of different insect species in different standard tissues.

[0062] This step specifically includes the following process.

[0063] S71. Using the prediction network of the multimodal prediction model, a pre-correction expression level prediction value is generated for each standard tissue. That is, in an insect species, the output is 14 real values ​​corresponding to 14 standard tissues. S72. Based on the predicted expression level before correction, obtain the bias parameters corresponding to the insect species and standard tissue in the species-tissue bias term; S73. The corresponding pre-correction expression level prediction value is superimposed with the bias parameter to obtain the post-correction expression level prediction value; S74. Aggregate all corrected expression level predictions and output the target gene's expression level prediction in the standard tissue set.

[0064] Using the above prediction method, for any target insect gene, as long as its reference genome sequence, gene structure annotation file, and protein sequence file are obtained, the predicted expression level of the gene in multiple tissues can be output through a multimodal prediction model.

[0065] like Figure 1 As shown, this embodiment also proposes an insect multi-tissue expression level prediction system based on multimodal feature fusion for the above method, including: The multi-species data acquisition module is used to acquire reference genome sequences, gene structure annotation files, protein sequence files, and tissue expression data for multiple insect species. The organization standardization module is used to uniformly map the publicly available data organization names of different insect species to a standard organization set. If a certain insect species lacks the actual measurement value of the standard organization, it is marked as a missing value. The transcription start site identification module is used to determine the transcription start site of each target gene based on the gene structure annotation file; The DNA sequence construction module is used to extract a genomic sequence window of a preset length from the reference genomic sequence as the target gene regulatory sequence, and to convert the target gene regulatory sequence into a sequence mode input matrix. The RNA stability feature construction module is used to extract RNA stability features of target genes based on gene structure annotation files. The protein embedding construction module is used to generate protein embedding vectors based on the protein sequence file using a pre-trained protein language model, and to use the species embedding vector as the protein embedding vector when the target gene is missing from the protein sequence file. The feature extraction network module includes a sequence tower, an RNA feature tower, and a protein embedding tower. The sequence tower comprises a convolutional downsampling module, a Transformer encoding module, and a global average pooling layer connected in sequence. The convolutional downsampling module performs multi-stage convolutional downsampling on the sequence modality input matrix to obtain target gene regulatory sequence features. The Transformer encoding module learns the long-range dependencies between different sequence positions on the position-encoded target gene regulatory sequence features and outputs encoded target gene regulatory sequence features containing global context information. The global average pooling layer performs global average pooling on the encoded target gene regulatory sequence features to obtain a target gene regulatory sequence representation vector conforming to a preset dimension. The fusion network module is used to splice and fuse the target gene regulatory sequence representation vector that conforms to the preset dimensions, the RNA stability feature vector that conforms to the preset dimensions, and the protein embedding vector that conforms to the preset dimensions to obtain the fusion vector; A prediction network module is used to output the predicted expression level of the target gene in a standard tissue set based on the fusion vector; The feature extraction network module, fusion network module, and prediction network module are included in the multimodal prediction model.

[0066] The training method of the multimodal prediction model in this embodiment is as follows: Figure 4As shown, the training process includes seven steps: training sample construction, expression label generation, missing tissue mask construction, multimodal model forward computation, loss function calculation, parameter update, and optimal model saving. During training, a masked regression loss function is used to train the model. This masked loss addresses the issue of inconsistent tissue coverage among species, enabling stable training even in scenarios with missing labels commonly found in publicly available insect data. Preferably, the loss function includes at least a mean squared error term and is calculated only at locations where true tissue expression labels are present. More preferably, a correlation constraint term can be added to improve the model's fitting ability to the overall expression trend. The AdamW optimizer, early stopping strategy, and learning rate warm-up strategy can be employed during training. Early stopping refers to terminating training prematurely when validation set performance no longer improves over a long period to prevent overfitting; learning rate warm-up refers to gradually increasing the learning rate in the early stages of training to improve training stability.

[0067] The tissue expression data in step S1 plays a central role in the construction, training, and data preprocessing of the multimodal prediction model. Specifically, it plays a key role in the following four aspects: 1. As a label for model training: Organizational expression data serves as a supervisory signal when training multimodal prediction models. Loss functions (mean squared error, correlation constraints, etc.) require real expression values ​​to calculate gradients and update parameters. Without this real expression data, the model cannot be trained.

[0068] 2. It is the basis for tissue standardization in step S2: In step S2, "the publicly available tissue names of different insect species are uniformly mapped to the standard tissue set" - which is the tissue label in the tissue expression data.

[0069] 3. As a basis for missing value masking: In step S2, "if a certain insect species lacks the true measurement value of standard tissue, it is marked as a missing value" - the basis for missing value judgment is tissue expression data.

[0070] 4. Learning objective as species-tissue bias: The bias parameter in step S7 is learned from the training data. It requires real tissue expression data as the training objective so that the model can capture the systematic bias of a specific species or a specific tissue in the overall expression level, thereby calibrating the output results in the prediction stage and improving the accuracy of cross-species prediction.

[0071] Furthermore, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0072] like Figure 5 The diagram shown illustrates the comparison of ablation effects after using the aforementioned prediction system and method, illustrating the differences in prediction metrics between the full model, RNA+ESM only, ESM removal, RNA removal, and DNA only models. Figure 5 As can be seen, the full model, which simultaneously uses DNA sequences, RNA features, and ESM protein embeddings, achieves an overall Pearson correlation coefficient (Pearson product-moment correlation coefficient) of 0.673. When only RNA and ESM are used, the Pearson coefficient is 0.599. Removing ESM reduces the Pearson coefficient to 0.517; removing RNA features further reduces the Pearson coefficient to 0.370; and using only DNA reduces the Pearson coefficient to 0.120. These results demonstrate the significant advantages of the trimodal joint modeling of this invention, indicating that the functionally conserved information carried by DNA regulatory information, RNA stability information, and protein language model embeddings are complementary.

[0073] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting the expression levels of multiple tissues in insects based on multimodal feature fusion, characterized in that, Includes the following steps: S1. Acquire multimodal input data, including reference genome sequences of multiple insect species, gene structure annotation files, protein sequence files, and tissue expression data; S2. Map the publicly available data organization names of different insect species to a standard organization set. If a certain insect species lacks a true measurement value of a standard organization, mark it as a missing value rather than a zero value. S3. Determine the transcription start site of each target gene based on the gene structure annotation file, extract a genomic sequence window of a preset length from the reference genome sequence with the transcription start site as the center, use it as the target gene regulatory sequence, and convert the target gene regulatory sequence into a sequence modality input matrix; The genome sequence window uses an asymmetric window of 20,000 bases in length, with 15,000 bases upstream and 5,000 bases downstream; S4. Based on the gene structure annotation file, extract the RNA stability characteristics of the target gene from the reference genome sequence; S5. For a target gene with a protein sequence file, input its protein amino acid sequence into a pre-trained protein language model to generate a protein embedding vector; For target genes missing in protein sequence files, a gating mechanism is used to switch from protein representation to species embedding representation instead of directly using all-zero vectors. The species embedding vector is used as the protein embedding vector. The species embedding vector is a fixed-dimensional vector learned for each insect species to represent the overall background features of that insect species. S6. Input the sequence modality input matrix, RNA stability features, and protein embedding vector into a pre-trained multimodal prediction model, and use the feature extraction network and feature fusion network in the multimodal prediction model to extract and fuse features to obtain a fusion vector; S7. The obtained fusion vector is input into the prediction network of the multimodal prediction model, and the output is the predicted expression level of the target gene in the standard tissue set; The multimodal prediction model also includes a species-tissue bias term, which learns and stores a bias parameter of a preset size for each insect species and each standard tissue to compensate for the overall expression baseline differences of different insect species in different standard tissues. Step S7 specifically includes the following processes: S71. Using the prediction network of a multimodal prediction model, generate a pre-correction expression level prediction for each standard tissue; S72. Based on the predicted expression level before correction, obtain the bias parameters corresponding to the insect species and standard tissue in the species-tissue bias term; S73. The corresponding pre-correction expression level prediction value is superimposed with the bias parameter to obtain the post-correction expression level prediction value; S74. Aggregate all corrected expression level predictions and output the target gene's expression level prediction in the standard tissue set; During the training process of the multimodal prediction model, a masked regression loss function is used to train the model. The masked loss addresses the problem of inconsistent tissue coverage among species. The loss function includes at least a mean squared error term and is calculated only at locations where true tissue expression labels are present.

2. The method for predicting multi-tissue expression levels in insects based on multimodal feature fusion as described in claim 1, characterized in that, In step S1, the reference genome sequence is in FASTA format; The gene structure annotation file is in GFF or GTF format; The protein sequence file is in FASTA format; The tissue expression data were derived from RNA-seq quantification results and uniformly converted to TPM.

3. The method for predicting insect multi-tissue expression levels based on multimodal feature fusion as described in claim 2, characterized in that, The tissue expression data can be converted to log2(TPM+1).

4. The method for predicting insect multi-tissue expression levels based on multimodal feature fusion as described in claim 1, characterized in that, In step S2, the standard tissue set uses 14 standard tissues, including adult brain, adult head, adult midgut, adult hindgut, adult fat body, adult Malpighian tubules, adult body, adult ovary, adult testis, larval hindgut, larval fat body, larval Malpighian tubules, larval midgut, and larval body.

5. The method for predicting insect multi-tissue expression levels based on multimodal feature fusion as described in claim 1, characterized in that, In step S3, the target gene regulatory sequence is converted into a coding matrix through one-hot encoding, which serves as the sequence modality input matrix.

6. The method for predicting multi-tissue expression levels in insects based on multimodal feature fusion as described in claim 1, characterized in that, In step S4, the RNA stability characteristics are expressed as a 16-dimensional feature vector, including the length of the 5' untranslated region, the length of the 3' untranslated region, the length of the coding region, the GC content, the number of introns, the total length of introns, the number of exons, the codon fitness index, the number of AU-rich elements, the Kozak sequence strength, the GC content at the third codon position, the proportion of rare codons, the GC content of the 5' untranslated region, the GC content of the 3' untranslated region, the total gene length, and the length of mature RNA.

7. The method for predicting insect multi-tissue expression levels based on multimodal feature fusion as described in claim 1, characterized in that, In step S6, the feature extraction and fusion includes the following steps: S61. Perform multi-stage convolutional downsampling on the sequence modality input matrix to obtain the target gene regulatory sequence features; S62. Add position encoding to the obtained target gene regulatory sequence features, learn the long-range dependencies between different sequence positions, and output the encoded target gene regulatory sequence features containing global context information; S63. Perform global average pooling on the encoded target gene regulatory sequence features to obtain a target gene regulatory sequence representation vector that conforms to the preset dimensions; S64. Input the RNA stability features into a two-layer multilayer perceptron network. The two layers of the multilayer perceptron network are processed by layer normalization and GELU activation function to obtain an RNA stability feature vector that conforms to the preset dimensions. The multilayer perceptron network adopts a feedforward neural network composed of several fully connected layers. S65. Input the protein embedding vector into a two-layer multilayer perceptron network. The two layers of the multilayer perceptron network are processed by layer normalization and GELU activation function to obtain a protein embedding vector that conforms to the preset dimensions. S66. The target gene regulatory sequence representation vector, the RNA stability feature vector, and the protein embedding vector that meet the preset dimensions are concatenated to obtain a concatenated vector. The concatenated vector is then input into a two-layer fusion network. Layer normalization, GELU activation function, and dropout network are used to process the two fusion networks to obtain the fusion vector.

8. A multimodal feature fusion-based insect multi-tissue expression level prediction system for the method of any one of claims 1 to 7, characterized in that, include: The multi-species data acquisition module is used to acquire reference genome sequences, gene structure annotation files, protein sequence files, and tissue expression data for multiple insect species. The organization standardization module is used to uniformly map the publicly available data organization names of different insect species to a standard organization set. If a certain insect species lacks the actual measurement value of the standard organization, it is marked as a missing value. The transcription start site identification module is used to determine the transcription start site of each target gene based on the gene structure annotation file; The DNA sequence construction module is used to extract a genomic sequence window of a preset length from the reference genomic sequence as the target gene regulatory sequence, and to convert the target gene regulatory sequence into a sequence mode input matrix. The RNA stability feature construction module is used to extract RNA stability features of target genes based on gene structure annotation files. The protein embedding construction module is used to generate protein embedding vectors based on the protein sequence file using a pre-trained protein language model, and to use the species embedding vector as the protein embedding vector when the target gene is missing from the protein sequence file. The feature extraction network module includes a sequence tower, an RNA feature tower, and a protein embedding tower. The sequence tower comprises a convolutional downsampling module, a Transformer encoding module, and a global average pooling layer connected in sequence. The convolutional downsampling module performs multi-stage convolutional downsampling on the sequence modality input matrix to obtain target gene regulatory sequence features. The Transformer encoding module learns the long-range dependencies between different sequence positions on the position-encoded target gene regulatory sequence features and outputs encoded target gene regulatory sequence features containing global context information. The global average pooling layer performs global average pooling on the encoded target gene regulatory sequence features to obtain a target gene regulatory sequence representation vector conforming to a preset dimension. The fusion network module is used to splice and fuse the target gene regulatory sequence representation vector that conforms to the preset dimensions, the RNA stability feature vector that conforms to the preset dimensions, and the protein embedding vector that conforms to the preset dimensions to obtain the fusion vector; The prediction network module is used to output the predicted expression level of the target gene in a standard tissue set based on the fusion vector.

Images