Cross-species expression profile mapping models, methods, systems, devices, media, and products

Abstract

This invention, entitled "Cross-Species Expression Profile Mapping Model, Method, System, Device, Medium, and Product," belongs to the field of biological information technology. The technical problems it aims to solve include the difficulty in accurately converting source species expression profiles to target species expression spaces in existing cross-species transcriptome analyses, and the ineffective utilization of differences in the transferability of different homologous genes. The key technical points are: the model includes an expression input layer, a homologous gene confidence gating module, a biological structure masking module, a mask variational encoder, a pathway constraint decoder, and an expression output layer. The mask variational encoder obtains latent representations through latent variable distribution parameterization and reparameterization sampling. The method includes a model construction phase, a model training phase, and an application prediction phase. The system includes a data acquisition module, a data quality control module, a homologous gene anchoring module, an expression representation construction module, a paired sample construction module, a model training module, and an application prediction module.

Description

Technical Field

[0001] This invention relates to the field of biological information technology, specifically to cross-species expression profile mapping models, methods, systems, devices, media, and products. Background Technology

[0002] Animal models are crucial techniques for drug development, disease mechanism research, efficacy evaluation, safety assessment, and translational medicine research. Mouse, rat, pig, dog, rhesus monkey, and cynomolgus monkey models are widely used to simulate human disease states, evaluate the efficacy of candidate drugs, study drug toxicity, and conduct mechanism validation. However, differences exist between species in genome sequences, the degree of conservation of homologous genes, transcriptional regulatory networks, tissue composition, immune environment, metabolic state, and expression dynamics, making it difficult to directly equate gene expression changes observed in animal models with those in humans.

[0003] Current cross-species transcriptome analyses typically rely on direct comparison of homologous genes, differentially expressed gene overlap analysis, linear regression mapping, or pathway enrichment analysis. While these methods provide some reference, they still have significant limitations. First, raw or conventionally normalized expression levels are insufficient to fully eliminate differences in sequencing depth, gene length, sample source, and expression scale between species. Second, differentially expressed gene overlap analysis simplifies continuous expression changes to a binary set, easily losing information about the magnitude and direction of expression changes. Third, the expression correspondence between the same tissues in different species is usually not a simple linear relationship, but involves complex nonlinear differences in expression regulation.

[0004] In recent years, variational autoencoders, masked autoencoders, and domain adaptation models have been used for learning to represent complex data. However, generalized masked autoencoders typically aim at random feature masking and reconstruction, failing to fully utilize the biological constraints at the level of homologous gene relationships, tissue expression stability, pathway structure, and gene set in cross-species expression profiles. Directly applying generalized models to cross-species expression profile mapping can easily lead to the following problems: different homologous genes are treated with equal weight; low-conservation genes introduce noise interference into the model; the model output only fits at the gene level and lacks consistency at the pathway level; and the prediction results are difficult to interpret.

[0005] Therefore, it is necessary to provide a dedicated model structure for cross-species expression profile mapping tasks. While maintaining the advantages of variational latent variables and mask learning, it introduces homologous gene credibility, biological structure mask, and pathway constraint decoding, so that the model can not only learn the nonlinear expression transformation relationship from the source species to the target species, but also maintain good biological consistency at the level of gene, pathway, and sample state differences. Summary of the Invention

[0006] The purpose of this invention is to provide: A cross-species expression profile mapping model, method, system, device, medium, and product are proposed to address the technical problems in existing cross-species transcriptome analysis, such as the difficulty in accurately converting the source species expression profile to the target species expression space, the inability of direct comparison of homologous genes or gene-by-gene linear mapping to characterize nonlinear cross-species expression relationships, the lack of effective utilization of the transferability differences of different homologous genes, and the lack of pathway-level biological constraints in model output.

[0007] To achieve the above objectives, this invention combines transcriptome sequencing data processing with homologous gene relationships, tissue expression stability, and biological pathway structures. Within a unified homologous gene expression feature space, it sequentially performs housekeeping gene relative expression conversion, homologous gene confidence gating, biological structure masking, mask variational coding, and pathway constraint dual-branch decoding to obtain numerical predicted expression profiles for the target species' corresponding tissue, a specified target tissue, or a specified target expression state.

[0008] Terminology Explanation: Unless otherwise defined, all technical terms herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0009] It should be understood that the above brief description and the following detailed description are exemplary and for illustrative purposes only, and do not limit the subject matter of the invention in any way. In this invention, the singular is used in conjunction with the plural unless otherwise specifically stated. It should also be noted that, unless otherwise stated, the use of “or” or “or” means “and / or”. Furthermore, the use of the term “comprising” and other forms such as “including,” “containing,” and “contains” are not limiting.

[0010] Unless otherwise stated, conventional methods within the scope of the art shall be used.

[0011] Unless specifically defined herein, the use of all commercially available products herein employs standard techniques. For example, it may be carried out using the manufacturer's instructions for use with the kit, or in accordance with methods known in the art or the description of this invention. The techniques and methods described herein can generally be implemented according to conventional methods well known in the art, based on the descriptions in the various summary and more specific documents cited and discussed in this specification.

[0012] The terms “optional / arbitrary” or “optionally / arbitrarily” mean that the event or situation described below may or may not occur, including both the occurrence and non-occurrence of the event or situation.

[0013] In a first aspect, the present invention provides: A cross-species expression profile mapping model includes an expression input layer, a homologous gene confidence gating module, a biological structure masking module, a mask variational encoder, a pathway constraint decoder, and an expression output layer. The expression input layer receives the relative expression profiles of the source species, obtained by tissue-specific stable housekeeping gene transformation within a unified homologous gene expression feature space. The homologous gene confidence gating module assigns differentiated gating weights to different homologous gene expression features. The biological structure masking module performs random gene masking and pathway-level structure masking during model training. The mask variational encoder encodes the masked input expression profiles, generating latent variable distribution parameters including a mean vector and a log-variance vector, and obtains latent representations through reparameterized sampling based on these parameters. The pathway constraint decoder includes a gene-level decoding branch and a pathway activity decoding branch. The gene-level decoding branch generates a gene-level predicted expression profile, and the pathway activity decoding branch generates a pathway activity vector and obtains gene expression correction terms through a pathway-gene association matrix. The expression output layer combines the gene-level predicted expression profiles with the gene expression correction terms to output the target species' predicted expression profile.

[0014] Based on further solutions to the technical problems of the present invention, or simultaneous solutions to multiple technical problems, the preferred solution in the technical solution provided in the first aspect of the present invention includes: First preferred option: The homology gene confidence gating module includes a gating weight initialization unit, a gating parameter learning unit, and an element-wise weighting unit. The gating weight initialization unit generates initial confidence weights based on at least one of expression stability, cross-species expression correlation, tissue stability, and homology correspondence confidence. The gating parameter learning unit updates the gating weights during model training. The element-wise weighting unit multiplies the source species input expression profile with the gating weights element by element.

[0015] Second preferred option: The biological structure masking module includes random gene masking units and pathway-level structure masking units; Random gene masking units are used to randomly mask some gene features; The pathway-level structure masking unit is used to simultaneously mask multiple gene features in the same biological functional module according to a preset pathway-gene relationship.

[0016] Third preferred option: The path constraint decoder is also configured to: The target pathway activity is calculated based on the target expression profile and pathway-gene association matrix of the target species. The predicted pathway activity is calculated based on the predicted expression profile of the target species and the same pathway-gene association matrix. The pathway consistency loss is formed based on the difference between the target pathway activity and the predicted pathway activity.

[0017] Fourth preferred option: The model is jointly trained based on gene-level reconstruction loss, KL divergence loss, pathway consistency loss, pathway activity decoding branch supervision loss, and gating weight regularization loss, so that the predicted expression profile is close to the target expression profile of the target species at both the gene expression numerical structure and pathway activity pattern levels.

[0018] Secondly, the present invention provides: A cross-species expression profile mapping method includes a model building phase, a model training phase, and an application prediction phase; During the model building phase: The transcriptome expression profile matrices and sample annotation information of the source and target species are obtained, and the quality control of the transcriptome expression profile matrices is performed at the sample and gene levels. Based on the homologous gene correspondence between the source and target species, the expression matrices of the source and target species are feature-aligned to construct a unified homologous gene expression feature space. The expression matrix is ​​then converted into a relative expression representation of housekeeping genes based on tissue-specific stable housekeeping genes. Cross-species training sample pairs are constructed based on the same tissue, tissues with preset correspondences, or designated target tissues. Each training sample pair includes at least the source species input expression profile and the target species target expression profile. The source species input expression profile is input into the homologous gene confidence gating module, and the expression features of each homologous gene are weighted to obtain the gated input expression profile. During the model training phase: The gated input expression profile is input into the biological structure masking module, where random gene masking and pathway-level structure masking based on the set of biological functional genes are performed to obtain the masked input expression profile. The masked input expression profile is then input into the mask variational encoder to generate latent variable distribution parameters, and latent variables are obtained through reparameterization. The latent variables are then input into the pathway constraint decoder, where gene-level predicted expression profiles are generated through the gene-level decoding branch, and pathway activity vectors are generated through the pathway activity decoding branch. Gene expression correction terms are obtained based on the pathway-gene association matrix mapping. The gene-level predicted expression profiles and gene expression correction terms are combined to obtain the target species predicted expression profile. The model is jointly trained based on gene-level reconstruction loss, KL divergence loss of latent variable distribution, pathway consistency loss of predicted expression profile, pathway activity decoding branch supervision loss, and gating weight regularization loss. In the application prediction phase: After processing the source species expression profile to be predicted according to the unified homologous gene expression feature space and the relative expression representation of housekeeping genes, the data is input into the trained model to obtain the predicted expression profile of the target species corresponding tissue, the specified target tissue, or the specified target expression state.

[0019] Based on further solutions to the technical problems of the present invention, or simultaneous solutions to multiple technical problems, the preferred solution in the technical solution provided in the second aspect of the present invention includes: First preferred option: The homologous gene correspondence includes a one-to-one homologous gene correspondence; the source species expression matrix and the target species expression matrix are aligned according to homologous gene identifiers and a unified feature order, so that the model input features and output features correspond to the same homologous gene pair respectively.

[0020] When a single tissue-specific stable housekeeping gene is used as a reference gene, the first In the nth sample Relative expression effect size of housekeeping genes for each gene for: ,in, For the first In the nth sample The expression values ​​of each gene after logarithmic transformation The expression value of the reference gene in the same sample after logarithmic transformation; When using multiple tissue-specific stable housekeeping genes, a reference expression baseline is constructed using the mean, median, or weighted statistic of the expression values ​​of the multiple housekeeping genes.

[0021] Second preferred option: The homologous gene credibility gating module generates an initial credibility score based on at least one of the following: expression stability of each homologous gene in the source and target species, expression correlation in cross-species training sample pairs, tissue stability, and credibility of homologous correspondence. The initial credibility score is then set as the initial value of a fixed gating weight or a learnable gating weight.

[0022] Let the input expression profile of the source species be... The gating weight vector is Then the gated input expression spectrum Among them, the gating weight is limited to a preset range. Internally, and composed of trainable parameters pass get.

[0023] Third preferred option: The biological structure masking module generates a mask vector. and in accordance with The input representation spectrum after masking is obtained; Random gene masking randomly masks some gene features according to a preset masking ratio. Pathway-level structure masking is based on gene sets from Reactome, KEGG, Gene Ontology, MSigDB, or other biological functional databases, while simultaneously masking at least some member genes in the selected functional gene set.

[0024] Fourth priority option: The masked variational encoder comprises a fully connected layer, a normalization layer, a nonlinear activation layer, and a random deactivation layer connected in sequence, and outputs the latent variable mean vector μ and the log-variance vector, respectively. ; Reparameterization operation according to Obtain latent variables ,in, ε follows a standard normal distribution.

[0025] Fifth priority option: Let the output of the gene-level decoding branch be The output of the active decoding branch of the pathway is The pathway-gene association matrix is The predicted expression profile of the target species ,in, Adjust the weights for preset or learnable pathways.

[0026] 13. The cross-species expression profile mapping method according to claim 5, characterized in that the total model loss... according to Calculate; where, The gene-level reconstruction loss between the predicted expression profile of the target species and the target expression profile of the target species is... The KL divergence loss is the difference between the latent variable posterior distribution and the pre-defined prior distribution. The loss of consistency between the predicted expression profile of the target species and the target expression profile of the target species in the pathway activity space. The monitoring loss between the output of the pathway activity decoding branch and the target pathway activity is... The loss is the deviation between the gating weights and the initial confidence weights after training. β , , and These are the weight parameters for the corresponding loss terms.

[0027] Sixth priority option: In the early stages of model training, the weights of the KL divergence loss are increased round by round, and model parameters are selected based on the validation set loss. The training set is used to update model parameters, the validation set is used for model selection and early stopping judgment, and the independent test set does not participate in model parameter updates or model selection.

[0028] Seventh priority option: The source species and target species are any combination of species with matching sets of homologous genes; the transcriptome expression profile includes a reference tissue expression profile, a disease tissue expression profile, a drug treatment expression profile, or a genetically modified model expression profile; the predicted expression profile is used for target tissue expression feature recovery, tissue-specific gene identification, drug toxicity risk assessment, disease mechanism analysis, or animal model transformation evaluation.

[0029] Thirdly, the present invention provides: A cross-species expression profile mapping system includes a data acquisition module, a data quality control module, a homologous gene anchoring module, an expression representation construction module, a paired sample construction module, a model training module, and an application prediction module. The data acquisition module acquires transcriptome expression profile matrices and sample annotation information for both the source and target species. The data quality control module performs quality control at both the sample and gene levels. The homologous gene anchoring module constructs a unified homologous gene expression feature space. The expression representation construction module constructs relative expression representations of housekeeping genes based on tissue-specific stable housekeeping genes. The paired sample construction module constructs training sample pairs consisting of the input expression profile of the source species and the target expression profile of the target species. The model training module includes a cross-species expression profile mapping model, and is trained based on gene-level reconstruction loss, KL divergence loss, pathway consistency loss, pathway activity decoding branch supervision loss, and gate weight regularization loss; the application prediction module is used to input the expression profile of the source species to be predicted into the trained cross-species expression profile mapping model and output the predicted expression profile of the target species.

[0030] Based on further solutions to the technical problems of the present invention, or simultaneous solutions to multiple technical problems, the preferred solution in the technical solution provided in the third aspect of the present invention includes: First preferred option: It also includes a model evaluation module and a downstream analysis module; the model evaluation module is used to evaluate the model performance based on at least one of the following: reconstruction error between the predicted expression profile and the actual target species expression profile, Pearson correlation coefficient, Spearman correlation coefficient, cosine similarity, pathway activity consistency, sensitivity or specificity; the downstream analysis module is used to perform target tissue expression feature recovery, tissue-specific gene identification, drug toxicity risk classification, disease differential expression analysis or pathway enrichment analysis based on the predicted expression profile of the target species.

[0031] Fourthly, the present invention provides: An electronic device is characterized by comprising a processor and a memory, wherein the memory stores program instructions, and when the processor executes the program instructions, it implements a cross-species expression profile mapping method.

[0032] Fifthly, the present invention provides: A computer-readable storage medium having program instructions stored thereon, which, when executed by a processor, implement a cross-species expression profile mapping method.

[0033] Sixthly, the present invention provides: A computer program product includes a computer program that, when executed by a processor, implements a cross-species expression profile mapping method.

[0034] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention provides a cross-species expression profile mapping model, method, system, and medium based on homology gating, biological structure masking, and pathway constraint decoding. This model focuses on cross-species expression profile mapping, addresses the issue of differences in the transferability of different homologous genes through homology gene confidence gating, enhances the model's robustness through pathway gene set-driven biological structure masking, and improves the biological consistency and interpretability of predicted expression profiles at the pathway level through pathway constraint decoders.

[0035] Second, this invention primarily focuses on cross-species, same-tissue expression profile mapping. Based on homologous gene correspondences, it anchors and aligns the expression features of source and target species, constructing a unified homologous gene expression feature space. It then employs a relative expression representation based on tissue-specific stable housekeeping genes, thereby transforming the expression profile of a specific tissue in the source species to the expression space of the corresponding tissue in the target species. This technical solution enables a clear homologous gene correspondence between model input and output, and helps reduce the impact of gene length, sequencing library size, sequencing batch size, inter-species expression scale differences, and tissue expression baseline differences on model training and prediction results. It provides a unified, quantifiable, and biologically correlated expression reference for the transformation of animal model transcriptome results to the expression space of target species, especially humans.

[0036] Third, this invention uses a homology gene confidence gating module to assign differentiated gating weights to different homology gene features, reducing the interference of low-confidence homology relationships, unstable gene expression, and high-noise features on the model, avoiding equal weighting of all homology genes, thereby improving the accuracy and stability of cross-species expression profile mapping.

[0037] Fourth, this invention constructs pathway-level structured masks based on the Reactome database or other biological functional gene sets, and combines gene-level expression reconstruction with pathway-level activity constraints through a pathway-constrained decoder. The pathway-level structured mask enables the model to recover the target expression profile using unmasked genes and their biological associations even when some biological functional modules are missing, thereby enhancing the model's robustness and generalization ability to local feature loss, measurement noise, and sample perturbations. The pathway-constrained decoder ensures that the model output not only approximates the target expression profile at the single-gene expression numerical level but also maintains consistency at the biological pathway activity pattern level, thereby improving the biological rationality and mechanistic explanation ability of the predicted expression profile.

[0038] Fifth, this invention does not limit the model to specific sample sources, disease states, or treatment conditions, but rather provides a reusable expression profile mapping framework. This framework can be trained, transferred, or fine-tuned based on transcriptome expression data from different species, tissues, biological states, and treatment conditions. It can be applied to tasks such as mapping expression space from model animals to human sources, restoring expression characteristics of target tissues, identifying tissue-specific genes, assessing drug-induced toxicity risks, interpreting disease mechanisms, evaluating animal model translation, and mapping expression profiles across tissues within the same species. This provides technical support at the expression profile level for translational medicine research, drug development, and animal model evaluation. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall structure of the cross-species expression profile mapping model based on homology gene confidence gating, biological structure masking, and pathway constraint decoding of the present invention. Figure 2 This is a graph showing the changes in training set loss and validation set loss during the training process of the cross-species expression profile mapping model from mouse to human expression space in Example 7 of the present invention. Figure 3 This is a diagram showing the comparison results of the direct extrapolation method of mouse homologous genes, the FIT-HKG model, and the model of this invention in the mouse-to-human expression space mapping task in Example 7 of this invention; Figure 4 The ROC curve obtained by mapping the expression profile of rat drug-treated liver tissue to the human expression space in Example 9 of the present invention and classifying the risk of human drug-induced liver injury based on the mapped expression profile is shown. Figure 5 This figure shows the comparison results of direct extrapolation results from mice in Example 10 of this invention, prediction results from the model of this invention, and actual human expression results in the enrichment analysis of upregulated pathways related to the mechanism of alveolar soft tissue sarcoma. Detailed Implementation

[0040] The following non-limiting embodiments are intended to enable those skilled in the art to gain a more comprehensive understanding of the present invention, but do not limit the invention in any way. The following content is merely an exemplary description of the scope of protection claimed by the present invention, and those skilled in the art can make various changes and modifications to the present invention based on the disclosed content, and such changes should also fall within the scope of protection claimed by the present invention.

[0041] The present invention will be further described below by way of specific embodiments. Unless otherwise specified, all instruments, devices, equipment, products, etc., used in the embodiments of the present invention are obtained through conventional commercial means.

[0042] Example 1 This embodiment provides a cross-species expression profile mapping model, which includes an expression input layer, a homology gene confidence gating module, a biological structure masking module, a mask variational encoder, a pathway constraint decoder, and an expression output layer. This cross-species expression profile mapping model is jointly trained based on gene-level reconstruction loss, KL divergence loss, pathway consistency loss, pathway activity decoding branch supervision loss, and gating weight regularization loss, enabling the predicted expression profile to closely approximate the target species' target expression profile at both the gene expression numerical structure and pathway activity patterns levels.

[0043] Specifically, the expression input layer receives the relative expression profile of the source species, obtained by tissue-specific stable housekeeping gene transformation, located in a unified homologous gene expression feature space; the homologous gene confidence gating module assigns differentiated gating weights to different homologous gene expression features; the biological structure masking module performs random gene masking and pathway-level structure masking during model training; the mask variational encoder encodes the masked input expression profile to generate latent variable distribution parameters, including a mean vector and a log-variance vector, and obtains a latent representation based on the latent variable distribution parameters through reparameterized sampling; the pathway constraint decoder includes a gene-level decoding branch and a pathway activity decoding branch, the gene-level decoding branch generates a gene-level predicted expression profile, and the pathway activity decoding branch generates a pathway activity vector and obtains gene expression correction terms through a pathway-gene association matrix; the expression output layer combines the gene-level predicted expression profile and the gene expression correction terms to output the target species predicted expression profile.

[0044] The homologous gene confidence gating module includes a gating weight initialization unit, a gating parameter learning unit, and an element-wise weighting unit. The gating weight initialization unit generates initial confidence weights based on at least one of expression stability, cross-species expression correlation, tissue stability, and homologous correspondence confidence. The gating parameter learning unit updates the gating weights during model training. The element-wise weighting unit multiplies the source species input expression profile with the gating weights element-wise.

[0045] Furthermore, the biological structure masking module includes a random gene masking unit and a pathway-level structure masking unit; the random gene masking unit is used to randomly mask some gene features; the pathway-level structure masking unit is used to simultaneously mask multiple gene features in the same biological functional module according to a preset pathway-gene relationship.

[0046] Furthermore, the pathway constraint decoder is also configured to: calculate the target pathway activity based on the target expression profile of the target species and the pathway-gene association matrix, calculate the predicted pathway activity based on the predicted expression profile of the target species and the same pathway-gene association matrix, and form a pathway consistency loss based on the difference between the target pathway activity and the predicted pathway activity.

[0047] Example 2 Based on the cross-species expression profile mapping model of Example 1, this example provides a cross-species expression profile mapping method, including a model building stage, a model training stage, and an application prediction stage.

[0048] Specifically, in the model construction phase: the transcriptome expression profile matrices and sample annotation information of the source and target species are obtained, and the transcriptome expression profile matrices undergo quality control at both the sample and gene levels; based on the homologous gene correspondence between the source and target species, the expression matrices of the source and target species are feature-aligned to construct a unified homologous gene expression feature space, and the expression matrix is ​​converted into a relative expression representation of housekeeping genes based on tissue-specific stable housekeeping genes; cross-species training sample pairs are constructed based on the same tissue, tissues with a preset correspondence, or designated target tissues, each training sample pair including at least the source species input expression profile and the target species target expression profile, both located in the homologous gene expression feature space with consistent feature order; the source species input expression profile is input into the homologous gene confidence gating module, and the expression features of each homologous gene are weighted to obtain the gated input expression profile; Among them, when a single tissue-specific stable housekeeping gene is used as a reference gene, the first In the nth sample Relative expression effect size of housekeeping genes for each gene for: ,in, For the first In the nth sample The expression values ​​of each gene after logarithmic transformation The expression value of the reference gene in the same sample after logarithmic transformation; When using multiple tissue-specific stable housekeeping genes, a reference expression baseline is constructed using the mean, median, or weighted statistic of the expression values ​​of the multiple housekeeping genes.

[0049] During the model training phase: the gated input expression profile is input into the biological structure masking module, which performs random gene masking and pathway-level structure masking based on the set of biological functional genes to obtain the masked input expression profile; the masked input expression profile is input into the mask variational encoder to generate latent variable distribution parameters, and latent variables are obtained through reparameterization; the latent variables are input into the pathway constraint decoder, which generates gene-level predicted expression profiles through the gene-level decoding branch, generates pathway activity vectors through the pathway activity decoding branch, and obtains gene expression correction terms based on the pathway-gene association matrix mapping; the gene-level predicted expression profiles and gene expression correction terms are combined to obtain the target species predicted expression profiles; the model is jointly trained based on gene-level reconstruction loss, KL divergence loss of latent variable distribution, pathway consistency loss of predicted expression profiles, supervision loss of pathway activity decoding branch, and gating weight regularization loss; In the application prediction phase: the expression profile of the source species to be predicted is processed according to a unified homologous gene expression feature space and the relative expression representation of housekeeping genes, and then input into the trained model to obtain the predicted expression profile of the target species' corresponding tissue, a specified target tissue, or a specified target expression state. The homologous gene correspondence includes a one-to-one homologous gene correspondence; the source species expression matrix and the target species expression matrix are aligned according to homologous gene identifiers and a unified feature order, so that the input and output features of the cross-species expression profile mapping model correspond to the same homologous gene pair.

[0050] Specifically, the homologous gene confidence gating module generates an initial confidence score based on at least one of the following: expression stability of each homologous gene in the source and target species, expression correlation in cross-species training sample pairs, tissue stability, and confidence of homologous correspondence. This initial confidence score is then set as the initial value of a fixed gating weight or a learnable gating weight. The source and target species are any combination of species with matching sets of homologous genes; the source species include mice, rats, pigs, dogs, rhesus monkeys, or cynomolgus monkeys; the target species includes humans; the transcriptome expression profile includes a reference tissue expression profile, a disease tissue expression profile, a drug-treated expression profile, or a genetically modified model expression profile; the predicted expression profile is used for target tissue expression feature recovery, tissue-specific gene identification, drug toxicity risk assessment, disease mechanism analysis, or animal model translation evaluation.

[0051] Specifically, let the input expression profile of the source species be... The gating weight vector is Then the gated input expression spectrum Among them, the gating weight is limited to a preset range. Internally, and composed of trainable parameters pass get; The gated input expression spectrum is input into the biological structure masking module; this module generates a mask vector. and in accordance with The input expression profile is obtained after masking; among them, the random gene mask randomly masks some gene features according to a preset mask ratio, and the pathway-level structure mask is based on gene sets from Reactome, KEGG, Gene Ontology, MSigDB or other biological functional databases, while masking at least some member genes in the selected functional gene set, so that the cross-species expression profile mapping model can recover the expression profile of the target species using the unmasked genes and biological association structures.

[0052] The masked input representation spectrum is then input into a masked variational encoder. The masked variational encoder comprises a fully connected layer, a normalization layer, a nonlinear activation layer, and a random deactivation layer connected in sequence, and outputs the latent variable mean vector μ and the log-variance vector, respectively. ; Reparameterization operation according to Obtain latent variables ,in, , ε follows a standard normal distribution.

[0053] The latent variable z is input into the pathway constraint decoder, which includes a gene-level decoding branch and a pathway activity decoding branch. Let the output of the gene-level decoding branch be... The output of the pathway activity decoding branch is The pathway-gene association matrix is Gene expression correction terms were obtained through pathway-gene association matrix P mapping. The predicted expression profile of the target species is represented as follows: ,in, Weights are adjusted for preset or learnable pathways. Through the aforementioned dual-branch decoding structure, the model output is simultaneously constrained by gene-level expression reconstruction and pathway-level biological structure.

[0054] In this embodiment, the cross-species expression profile mapping model is trained using a joint loss function. Specifically, the total model loss... Expressed as: ;in, The gene-level reconstruction loss between the predicted expression profile of the target species and the target expression profile of the target species is... The KL divergence loss is the difference between the latent variable posterior distribution and the pre-defined prior distribution. The loss of consistency between the predicted expression profile of the target species and the target expression profile of the target species in the pathway activity space. The monitoring loss between the output of the pathway activity decoding branch and the target pathway activity is... The loss is the deviation between the gating weights and the initial confidence weights after training. β , , and These are the weight parameters for the corresponding loss terms.

[0055] In one implementation, the weight of the KL divergence loss is incrementally increased round by round during the initial stage of model training, and the total validation set loss is used as a monitoring metric for model parameter selection and training termination. When the total validation set loss in a certain training round is lower than the previously obtained minimum total validation set loss, the model parameters corresponding to that training round are saved, and the minimum total validation set loss is updated. When the total validation set loss does not fall below the previously obtained minimum total validation set loss for 20 consecutive training rounds, model training is terminated early. After model training is completed, the model parameters corresponding to the lowest total validation set loss are determined as the final model parameters. That is, the training set is used to update model parameters, the validation set is used for model selection and early stopping based on the loss, and the independent test set does not participate in model parameter updates or model selection.

[0056] In the application prediction stage, the expression profile of the source species to be predicted is processed according to the unified homologous gene expression feature space and the relative expression representation of housekeeping genes, and input into the trained cross-species expression profile mapping model to obtain the predicted expression profile of the target species corresponding tissue, the specified target tissue, or the specified target expression state.

[0057] Example 3 This embodiment provides a cross-species expression profile mapping system, specifically including a data acquisition module, a data quality control module, a homologous gene anchoring module, an expression representation construction module, a paired sample construction module, a model training module, and an application prediction module. The data acquisition module acquires transcriptome expression profile matrices and sample annotation information for the source and target species. The data quality control module performs quality control at the sample and gene levels. The homologous gene anchoring module constructs a unified homologous gene expression feature space. The expression representation construction module constructs relative expression representations of housekeeping genes based on tissue-specific stable housekeeping genes. The paired sample construction module constructs training sample pairs consisting of the input expression profile of the source species and the target expression profile of the target species. The model training module uses a cross-species expression profile mapping model as described in Embodiment 1, trained based on gene-level reconstruction loss, KL divergence loss, pathway consistency loss, pathway activity decoding branch supervision loss, and gate weight regularization loss. The application prediction module inputs the expression profile of the source species to be predicted into the trained cross-species expression profile mapping model and outputs the predicted expression profile of the target species.

[0058] In other embodiments, a model evaluation module and a downstream analysis module are also included; the model evaluation module is used to evaluate the model performance based on at least one of the following: reconstruction error between the predicted expression profile and the actual target species expression profile, Pearson correlation coefficient, Spearman correlation coefficient, cosine similarity, pathway activity consistency, sensitivity, or specificity; the downstream analysis module is used to perform target tissue expression feature recovery, tissue-specific gene identification, drug toxicity risk classification, disease differential expression analysis, or pathway enrichment analysis based on the predicted expression profile of the target species.

[0059] Example 4 This embodiment provides an electronic device, including a processor and a memory. The memory stores program instructions, and when the processor executes the program instructions, it implements the cross-species expression profile mapping method in Embodiment 2.

[0060] Example 5 This embodiment provides a computer-readable storage medium storing program instructions, which, when executed by a processor, implement the cross-species expression profile mapping method in Embodiment 2.

[0061] Example 6 This embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the cross-species expression profile mapping method in Embodiment 2.

[0062] Example 7 This embodiment illustrates the data processing, model construction, training, and performance evaluation process of the cross-species expression profile mapping model of the present invention in the task of mapping the same tissue expression profile from model animals to human expression space. By comparing it with the direct extrapolation method of homologous genes of the source species and the FIT-HKG model based on gene-by-gene linear correction, the technical effect of the model of the present invention in the reconstruction of the expression profile of the target species is verified.

[0063] The model of this invention can be used for cross-species expression profile mapping from model animals to human expression spaces. These model animals include, but are not limited to, mice, rats, pigs, dogs, rhesus monkeys, and cynomolgus monkeys. Furthermore, the model is also applicable to any combination of source and target species with matching sets of homologous genes. This embodiment uses the mapping of mouse expression profiles to corresponding human tissue expression spaces as an example to illustrate the construction, training, and performance evaluation process of the model of this invention.

[0064] To evaluate the ability of different methods to recover the expression profile of the target species, this embodiment sets up the following three expression profile mapping methods: The first method is direct extrapolation from the source species. In this method, within a pair of homologous gene feature spaces, the expression effect size of the source species is directly used as the derived result of the expression effect size of the corresponding homologous gene in the target species. This method does not perform cross-species parameter learning and is used to evaluate the consistency between the expression profiles of the source species without cross-species correction and the true expression profiles of the target species.

[0065] The second method is the FIT-HKG model, which is a gene-by-gene linear mapping benchmark model based on the relative expression effect sizes of housekeeping genes, constructed with reference to the FIT-type linear mapping idea. FIT-HKG learns the linear correspondence between the expression effect sizes of the source and target species for each homologous gene in a one-to-one homologous gene feature space. For the first... i The th training sample pair g There are 1 homologous genes, and the expression effect size of the source species is 1. The actual target species expression effect size is Then the predicted value of the FIT-HKG model is expressed as: ； in, For the first g Intercept parameters corresponding to each homologous gene For the first g The slope parameters corresponding to each homologous gene. g The model training parameters for each gene are obtained by minimizing the following objective function: ； in, N This represents the number of sample pairs used for training. This is the regularization coefficient. The regularization term is used to constrain the intercept. Approaching 0, slope A value close to 1 allows the model to approximate the direct extrapolation results of the source species when sufficient correction evidence is lacking; when the training data supports a stable systematic difference between the source and target species, the corresponding linear correction parameter is learned. The regularization parameter can be determined through cross-validation.

[0066] The third method is the cross-species expression profile mapping model proposed in this invention, based on homologous gene confidence gating, biological structure masking, variational latent representation, and pathway constraint decoding. Unlike FIT-HKG, which models independently for each gene, the model in this invention jointly models all homologous gene features, learns the nonlinear transformation relationship between the expression state of the source species and the expression space of the target species, and uses homologous gene confidence and biological pathway structure to constrain the model training process.

[0067] First, transcriptome expression profile matrices, sample annotation information, and homologous gene correspondences between source and target species were obtained from public databases such as GTEx and GEO. In this embodiment, expression profiles of 29 corresponding tissues from mice and humans were collected, including brain (11, 19), pituitary gland (9, 313), eyeball and optic nerve (9, 30), thyroid gland (15, 684), esophagus (7, 561), aorta (9, 472), lymph nodes (8, 10), thymus (9, 10), bone marrow (8, 11), heart (22, 461), liver (18, 262), lung (19, 604), kidney (27, 104), adrenal gland (11, 295), spleen (14, 277), stomach (10, 407), and pancreas (19, 19, 10 ... The tissues included (362), large intestine (9,419), small intestine (22,207), bladder (12,77), skin (9,651), skeletal muscle (9,818), prostate (male) (9,282), testis (male) (13,414), epididymis (male) (32,14), mammary gland (female) (16,514), ovary (female) (11,193), uterus (female) (10,153), and peripheral blood (10,803). The values ​​in parentheses after each tissue name represent the number of mouse samples and the number of human samples, respectively, totaling 387 mouse samples and 9427 human samples.

[0068] The sample annotation information includes at least sample number and tissue information, and may also include species, sex, age, treatment conditions, disease status, and other information that can be used for sample screening, grouping, and model construction. The expression profile data undergoes sequential sample quality control, low-quality gene filtering, tissue name standardization, duplicate gene processing, abnormal sample identification, and one-to-one homologous gene screening. The expression matrices of the source and target species are aligned according to a unified homologous gene order, thereby constructing a homologous gene expression space shared by the source and target species and with consistent feature order.

[0069] In one specific implementation, the expression profiles of the source and target species are converted into relative expression effect sizes of housekeeping genes. For each tissue, a housekeeping gene that is relatively stably expressed within that tissue is selected as a reference feature, and a tissue-specific expression baseline is constructed based on the housekeeping gene. i In the nth sample g The relative expression effect size of housekeeping genes for each gene can be expressed as: ； in, No. i In the nth sample g The expression values ​​of each gene after logarithmic transformation The expression value of the reference housekeeping gene in the same sample after logarithmic transformation is given.

[0070] The expression values ​​of each homologous gene are converted relative to the corresponding reference baseline. This conversion can reduce the impact of sequencing depth, library size, overall expression scale, sample composition, species background differences, and tissue expression baseline differences on model training.

[0071] Subsequently, cross-species training sample pairs were constructed within the same tissue. Each training sample pair included at least the source species input expression vector, the target species target expression vector, and the tissue label. The source species input and the target species target were located in the same pair of homologous gene feature spaces, with each feature corresponding to a homologous gene pair consisting of a source species gene and a target species gene.

[0072] The model of this invention first performs gene-by-gene weighting on the input expression profile of the source species through a homologous gene credibility gating module. This module sets trainable gating weights for each homologous gene feature. The gating weights can be initialized based on the expression stability in the source and target species, cross-species pairing correlation, and the credibility of homologous correspondence, and are further optimized during model training.

[0073] The gated expression vector of the source species can be represented as: ； in, For the first g The gating weights of homologous genes. In this embodiment, the gating weights can be limited to a preset range. Internally, and composed of trainable parameters The following transformation is used to obtain: ； The gating weights can be initialized based on the expression stability in the source and target species, the expression correlation in cross-species training sample pairs, tissue stability, and the reliability of homologous gene correspondences, and further optimized during model training. The homologous gene reliability gating module is used to reduce the interference of low-reliability homologous relationships, genes with unstable expression relationships, and high-noise genes on the cross-species mapping results, while retaining gene features that contribute significantly to the reconstruction of the target species' expression profile.

[0074] Subsequently, the model further performs biological structure masking on the gated source species expression profile. Let the mask vector be:

[0075] The input after the mask is represented as: ； The biological structure mask includes a random gene mask and a pathway-level structure mask. The random gene mask randomly maps some gene features to a predetermined probability. Setting it to 0 improves the model's robustness to local expression loss, measurement noise, and sample perturbations. Pathway-level structure masking, based on pathway gene sets provided by biological functional databases such as the Reactome database, simultaneously masks multiple member genes in one or more pathways. For the selected [specific gene]... k Each pathway, whose member gene set is denoted as ,but:

[0076] By using pathway-level structured masking, the model is forced to recover the expression profile of the target species by utilizing the unmasked genes and the association structures between different biological functional modules, rather than relying solely on the direct correspondence between individual homologous genes.

[0077] The source species expression vector, after gating and path-level structure masking, is input to the variational encoder. In this specific implementation, the encoder first performs fully connected transformation, normalization, nonlinear activation, and random deactivation on the input: ； in, and Here, LN represents the encoder weight matrix and bias vector, respectively, and GELU represents the nonlinear activation function. The encoder further generates the mean vector and log-variance vector of the latent distribution: ； ； Subsequently, latent variables are obtained through reparameterization: ； in: ; The reparameterization operation enables the model to perform gradient backpropagation while maintaining random latent variable sampling.

[0078] The model decoder of this invention includes a gene-level decoding branch and a pathway-constrained decoding branch. The gene-level decoding branch directly generates the gene expression reconstruction values ​​of the target species; the pathway-constrained decoding branch predicts the pathway activity of the target species based on the Reactome pathway-gene relationship matrix and maps the pathway activity back to the gene-level expression correction terms. The final output of the model is formed by the combined gene-level reconstruction results and pathway-level correction results. ； in, To obtain latent variables for encoder reparameterization operations, For the gene decoding branch output, Gene correction terms obtained from pathway activity mapping. The pathway-gene association matrix contains, K For the number of pathways, G The number of homologous gene characteristics. These are trainable or preset path correction coefficients.

[0079] The training loss of the model in this invention includes target species gene expression reconstruction loss, latent space KL divergence loss, predicted expression profile pathway consistency loss, pathway decoding branch supervision loss, and homologous gene gating regularization loss. The total loss can be expressed as:

[0080] in: 、 、 and These are the weight parameters for the corresponding loss terms; Used to constrain gene-level errors between the predicted expression profile of the target species and the actual expression profile of the target species; Used to constrain the distribution of latent variables; Used to constrain the consistency of pathway activity between predicted expression profiles and actual expression profiles; Used to monitor the decoding branch of the pathway; Used to limit the gating weights from deviating excessively from their initial confidence level.

[0081] In this specific implementation, the model is trained for 200 epochs, with the KL divergence loss weight gradually increased in the first 20 epochs to reduce reconstruction instability caused by excessively strong potential spatial constraints in the early stages of training. During training, the total loss, gene reconstruction error, KL divergence, pathway consistency error, and gating regularization error of the training and validation sets are recorded. The training results are as follows: Figure 2The results show that during the first 50 training rounds, the total loss of the training set decreased from 4.6011 in round 1 to 0.4768 in round 50, and the total loss of the validation set decreased from 2.3412 to 0.4287, with the validation set loss decreasing to 0.4246 in round 48. The mean squared error of gene reconstruction in the training set decreased from 4.4409 to 0.3976, and the mean squared error of gene reconstruction in the validation set decreased from 2.2638 to 0.3526; the mean squared error of pathway reconstruction in the training set decreased from 0.6750 to 0.0458, and the mean squared error of pathway reconstruction in the validation set decreased from 0.3151 to 0.0420. These results indicate that the prediction errors of the model at both the gene and pathway levels continuously decrease during the training process and tend to stabilize in the later stages of training, without significant loss divergence or continuous deterioration of validation performance.

[0082] After model training, evaluation was conducted on an independent test set covering 29 tissues. For each test sample, the human expression profile predicted by the model was compared with the corresponding real human expression profile, and Pearson correlation coefficient, Spearman correlation coefficient, cosine similarity, and pathway cosine similarity were calculated. The average test results for the 29 tissues are as follows: Pearson correlation coefficient ranged from 0.9185 to 0.9950 across the 29 tissues; Spearman correlation coefficient ranged from 0.9325 to 0.9955 across the 29 tissues; cosine similarity ranged from 0.9201 to 0.9951 across the 29 tissues; and pathway cosine similarity ranged from 0.9546 to 0.9979 across the 29 tissues.

[0083] The results show that the model of this invention can maintain high consistency in gene expression profiles and pathway activity across different tissues. Even in tissues with relatively low predictive performance, the Pearson correlation coefficient still reaches 0.9185; in the best-performing tissue, the Pearson correlation coefficient reaches 0.9950, indicating that the model has good cross-tissue applicability and mapping stability.

[0084] Further, all gene observations from 29 tissues were merged, resulting in 383,918 tissue-gene observations. The overall correlation and cosine similarity between the source species direct projection, FIT-HKG, and the predictions of this invention's model and the actual human expression profile were calculated. Compared to the source species direct projection, FIT-HKG significantly improved the consistency of the expression profile through gene-by-gene linear correction, indicating that the source species expression profile, after cross-species correction, is closer to the actual human expression space. This invention's model further improves the overall Pearson correlation coefficient and cosine similarity based on FIT-HKG, with the Pearson correlation coefficient increasing from 0.9353 to 0.9782 and the cosine similarity increasing from 0.9368 to 0.9788.

[0085] The Spearman correlation coefficients of the model of this invention and FIT-HKG are both at a high level, at 0.9791 and 0.9780 respectively, indicating that both methods can maintain the relative gene order well; however, the model of this invention has a higher Pearson correlation coefficient and cosine similarity, indicating that it can not only maintain the gene order, but also more accurately recover the overall numerical structure of the expression effect size of the target species and the direction of the expression vector.

[0086] To evaluate the ability of different methods to recover expression characteristics of the target species, real human expression profiles were used as a reference, and TP, FP, FN, and TN were defined based on the expression effect size threshold and expression direction. Specifically: TP indicates that both the predicted result and the real human expression result exceed the preset effect size threshold, and the expression change direction is consistent; FP indicates that the predicted result exceeds the threshold, but the real human result does not exceed the threshold, or the expression change directions of the two are inconsistent; FN indicates that the real human result exceeds the threshold, but the predicted result does not reach the threshold; TN indicates that neither the predicted result nor the real human result reaches the threshold.

[0087] Sensitivity and Specificity are defined as follows: ; ； Sensitivity evaluates the model's ability to recover genes expressing the true target species' characteristics, while Specificity evaluates the model's ability to exclude genes that do not conform to the target species' expression characteristics. Using both together avoids achieving a high detection rate solely by expanding the predicted positive set.

[0088] In this embodiment, the proportion of top genes is used as the threshold for screening expression features. For each tissue, genes in the predicted expression profile and the real human expression profile are sorted from largest to smallest according to the absolute value of their expression effect size, and genes in the top proportion are selected as the feature gene set of the corresponding expression profile. The proportion of top genes is set sequentially as 0.01, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, and 0.40, that is, the top 1%, 5%, 10%, 15%, 20%, 25%, 30%, 35%, and 40% of genes are selected, respectively. Under each top gene proportion threshold, true positive, false positive, false negative, and true negative are defined according to the set relationship and expression change direction of the feature genes in the predicted results and the real human results. Sensitivity and specificity of 29 tissues are calculated, and the median and quartile range of the results of each tissue are summarized.

[0089] like Figure 3As shown, at a threshold of 0.05, the median Sensitivity of the model of this invention is 0.8072, the median Specificity is 0.9898, the FIT-HKG values ​​are 0.7834 and 0.9887, respectively, and the direct extrapolation values ​​for mice are 0.2975 and 0.9630, respectively.

[0090] At a threshold of 0.40, the median Sensitivity of the model of this invention was 0.9164, the median Specificity was 0.9443, the FIT-HKG values ​​were 0.9115 and 0.9411, respectively, and the direct extrapolation values ​​for mice were 0.6197 and 0.7617, respectively.

[0091] Of the nine effect size thresholds evaluated, the median sensitivity and median specificity of the model in this invention were numerically higher than those of FIT-HKG at eight thresholds. At all thresholds, both the model in this invention and FIT-HKG significantly outperformed direct extrapolation from the source species. These results indicate that while the model in this invention improves the recovery rate of spatially consistent genes in the target area, it does not achieve increased sensitivity by significantly increasing non-target genes, but rather maintains a high level of specificity.

[0092] This embodiment demonstrates that the source species direct extrapolation method does not correct for systematic expression differences between the source and target species, and its correlation with the real human expression profile and its ability to recover target expression features are relatively limited. FIT-HKG can significantly improve cross-species expression profile consistency through gene-by-gene regularized linear mapping, but it models each gene separately, making it difficult to fully characterize the nonlinear synergistic relationships and pathway structural dependencies between genes.

[0093] This invention's model learns the transformation relationship from the source species to the human expression space at the full expression profile level through homology gene confidence gating, biological structure masking, variational latent space, and pathway-constrained dual-branch decoding. Independent test results show that this invention's model achieves the highest values ​​among the three methods in terms of overall Pearson correlation coefficient, Spearman correlation coefficient, and cosine similarity; at most effect size thresholds, its median sensitivity and specificity are also higher than FIT-HKG.

[0094] Therefore, compared with direct extrapolation of the source species and gene-by-gene linear mapping models, the model of this invention can more accurately reconstruct the numerical structure, gene sequence, expression direction and pathway activity characteristics of human expression profiles, thereby improving the accuracy, stability and biological consistency of the transformation of model animal expression profiles to human expression space.

[0095] Example 8 This embodiment illustrates that, compared with the direct extrapolation method of homologous genes in the source species, the cross-species expression profile mapping model of the present invention can more effectively recover the gene expression direction and tissue expression characteristics in the expression space of the target species and target tissue.

[0096] In one comparative scheme, the expression profile of the source species is aligned with a pair of homologous genes and directly used as the derivation result of the expression trend of the target species, which is denoted as the direct derivation result of the source species; at the same time, the expression profile of the same source species is input into the model trained by this invention to obtain the predicted expression profile of the target species, which is denoted as the prediction result of the model of this invention.

[0097] For a target species and target tissue, the set of characteristic genes for target tissue expression is defined based on the expression profile of the actual target species as a reference, and on gene expression effect size, direction of expression change, and preset thresholds. For example, genes in the target species whose absolute expression effect size exceeds a set threshold and show a clear up-regulation or down-regulation trend relative to the reference expression background can be defined as characteristic genes for target tissue expression. The reference expression background can be other tissues of the target species, reference samples from the same tissue, the expression baseline after housekeeper gene correction, or a pre-defined tissue expression reference space.

[0098] The directional consistency and threshold consistency between the direct extrapolation results of the source species and the prediction results of the model of this invention and the expression profiles of the actual target species are compared. Specifically, genes that simultaneously meet the following conditions can be defined as target space consistent genes: the predicted results and the expression change direction of the actual target species expression profile are consistent, and both exceed the preset expression effect size threshold. For the direct extrapolation results of the source species, the number of genes with the same directional consistency between the direct extrapolation results of the source species and the expression profiles of the actual target species can be counted; for the prediction results of the model of this invention, the number of genes with the same directional consistency between the predicted expression profile of the model and the expression profiles of the actual target species can be counted.

[0099] Furthermore, genes that are additionally recovered in the model prediction results of this invention compared to the direct extrapolation results of the source species can be defined as model-recovered genes. These genes represent those that failed to maintain the same direction or meet the threshold requirements as the expression profile of the true target species during direct extrapolation from the source species, but after mapping by the model of this invention, can exhibit consistent expression direction and intensity characteristics with the expression profile of the true target species. This result demonstrates that the model of this invention does not merely perform direct projection at the homologous gene level, but rather learns and corrects the nonlinear differences between the expression profile of the source species and the expression space of the target species through modules such as homologous gene confidence gating, biological structure masking, and pathway constraint decoding.

[0100] Taking the expression profile of mouse brain tissue mapped to human brain tissue as an example, under the condition that the absolute value of the expression effect size is greater than the threshold of 0.25, a total of 13,448 homologous genes were included. In the direct extrapolation results of the source species, 9,792 genes were consistent with the expression profile of real human brain tissue and exceeded the threshold, accounting for 72.81% of all homologous genes; in the prediction results of the model of this invention, 11,647 genes were consistent with the expression profile of real human brain tissue and exceeded the threshold, accounting for 86.61% of all homologous genes. Compared with the direct extrapolation results of the source species, the prediction results of the model of this invention have a net increase of 1,855 target spatially consistent genes, an increase of 13.80 percentage points, and an improvement of about 18.95% compared with the direct extrapolation results of the source species. This result shows that the model of this invention can recover more gene signals consistent with the expression pattern of real human brain tissue, rather than simply extrapolating the expression profile of the source species at the homologous gene level.

[0101] This embodiment demonstrates that, compared to direct extrapolation methods from source species, the model of this invention can recover more genes consistent with the actual tissue expression patterns of the target species, and reduce the number of genes with inconsistent orientations and missed genes caused by direct extrapolation from source species. Therefore, the model of this invention can improve the ability to reconstruct the tissue expression characteristics of the target species during cross-species expression profile transformation, providing more reliable expression profile mapping results for cross-species biological extrapolation, animal model evaluation, and target species expression response prediction.

[0102] Example 9 This embodiment illustrates the application of the model of the present invention in the risk assessment of drug-induced liver injury. Specifically, the RNA-seq expression profile of rat liver tissue after drug treatment is mapped to the human expression space, and the mapped human predicted expression profile is input into a pre-constructed human drug-induced liver injury binary classification model, thereby achieving cross-species translational judgment of the hepatotoxicity risk of candidate drugs.

[0103] First, RNA-seq data were collected from rat liver tissue after drug administration. This data included rat liver tissue expression matrices, sample processing information, compound names, dosage, administration time, control group information, and optional hepatotoxicity tags.

[0104] Subsequently, the RNA-seq expression matrix of rat liver tissue treated with drugs was preprocessed. The preprocessing steps included sample quality control, low-quality gene filtering, expression level standardization, homologous gene matching, expression effect size calculation, and gene feature order unification. After processing, the rat drug-treated expression profile was consistent with the input features of the trained model of this invention, and a rat drug-treated expression effect size matrix was formed in a one-to-one homologous gene feature space.

[0105] To avoid data leakage due to duplicate drug information during model evaluation, the rat drug set used for external validation in this embodiment does not overlap with the drug set used in training the human drug-induced liver injury binary classification model, and drugs with no overlap between the two are selected as external test drugs. Next, a human drug-induced liver injury binary classification model is constructed. The human data comes from large-scale human toxicology transcriptome resources, and the binary classification model uses drug-induced liver injury-related characteristic genes as classification features.

[0106] In one specific implementation, the characteristic genes are the most common differentially regulated genes in drug-induced liver injury, selected from literature screening, with the top 100 differentially regulated genes ranked by absolute fold change. Based on these characteristic genes, a corresponding feature matrix is ​​extracted from the expression profile of human drug treatment, and a binary training set consisting of positive and negative samples of drug-induced liver injury is constructed. The classification model can employ a stable feature selection method and be trained in conjunction with a logistic regression classifier. The model training process may include steps such as missing value imputation, low-variance feature filtering, data standardization, stable feature selection, and final classifier fitting. After training, the classification model outputs the predicted probability of each sample belonging to the positive category of drug-induced liver injury, and the classification performance is evaluated using ROC curves and AUC.

[0107] Then, the pretreated rat drug-treated liver tissue expression profile is input into the trained cross-species expression profile mapping model of this invention to obtain the corresponding human expression space prediction results. The predicted expression profile does not require the existence of human drug-treated samples that are one-to-one paired with the rat drug-treated samples in the experiment. Instead, it is a humanized predicted expression representation generated by the model of this invention in the human liver tissue expression space based on the expression state of the rat drug-treated liver tissue. This process realizes the conversion of the transcriptomic response of liver tissue in animals into an expression profile representation that is closer to the human expression background.

[0108] Subsequently, the human expression profile predicted by the model of this invention is input into a pre-trained binary classification model for human drug-induced liver injury. For each rat drug-treated sample, the classification model outputs its corresponding positive prediction probability of drug-induced liver injury. Based on the predicted probability and the true hepatotoxicity label of the sample, an ROC curve is plotted and the AUC is calculated to evaluate whether the human predicted expression profile mapped by the model of this invention can support the identification of human hepatotoxicity risk.

[0109] In one specific implementation, rat liver tissue RNA-seq samples treated with drugs that were not included in the training drug set of the human drug-induced liver injury classification model are selected as an external validation set. After mapping by the model of this invention, the obtained human expression spatial prediction spectrum is input into the human drug-induced liver injury binary classification model to obtain the corresponding hepatotoxicity risk prediction results.

[0110] External validation results such as Figure 4 The results show that the joint analysis pipeline consisting of the cross-species expression profile mapping model and the human hepatotoxicity classification model of this invention has an AUC of 0.768 on the rat external drug test set. This indicates that the model of this invention can convert the transcriptomic response of rat liver tissue after drug treatment into an expression representation that can be recognized by the human hepatotoxicity classification model, and retains a certain ability to discriminate the risk of human drug-induced liver injury in external drug test scenarios where the classification model was not trained.

[0111] This embodiment demonstrates that the model of the present invention can serve as an expression profile conversion and bridging module between animal toxicology studies and human toxicity risk assessment. By mapping the RNA-seq expression profile of rat liver tissue after drug treatment to the human expression space, and combining it with a binary classification model of human drug-induced liver injury, the present invention can achieve the computational conversion from animal liver tissue transcriptome response to human hepatotoxicity risk. This method can be used for early safety evaluation of candidate drugs, human extrapolation of animal toxicology results, screening of compounds with potential drug-induced liver injury risk, and cross-species toxicological mechanism analysis.

[0112] Example 10 This embodiment illustrates the application value of the model of the present invention in interpreting disease mechanisms. Specifically, taking the skeletal muscle expression profile analysis related to alveolar soft part sarcoma (ASPS) as an example, it evaluates whether the model of the present invention can more accurately map the expression profile of mouse disease states to the human disease expression space, and further recover gene and pathway signals related to human disease mechanisms.

[0113] First, we obtained human / mouse RNA-seq expression profiles of ASPS skeletal muscle and healthy skeletal muscle from the GSE54729 dataset in the GEO database. Based on sample origin and disease status, the samples were divided into four categories: healthy mice, ASPS mice, healthy humans, and ASPS humans. We then performed sample quality control, homologous gene alignment, expression representation transformation, and grouping on the expression matrices to ensure that the mouse expression profiles were consistent with the input features of the trained model of this invention.

[0114] Then, the RNA-seq expression profiles of healthy mouse skeletal muscle and ASPS mouse skeletal muscle are input into the trained model of this invention to obtain the predicted expression profiles of human healthy skeletal muscle and the predicted expression profiles of human ASPS skeletal muscle.

[0115] At the level of overall similarity of expression profiles, the consistency between the direct extrapolation results from mice and the mapping results of the model of this invention and the real human expression profiles was compared. For the expression profile of healthy skeletal muscle, the Pearson correlation coefficient between the direct extrapolation results from mice and the real healthy human skeletal muscle expression profile was 0.761, the Spearman correlation coefficient was 0.763, and the cosine similarity was 0.761. After mapping by the model of this invention, the Pearson correlation coefficient between the predicted human healthy skeletal muscle expression profile and the real healthy human skeletal muscle expression profile increased to 0.838, the Spearman correlation coefficient increased to 0.822, and the cosine similarity increased to 0.834. For the ASPS disease state expression profile, the Pearson correlation coefficient between the direct extrapolation results from mice and the real human ASPS skeletal muscle expression profile was 0.597, the Spearman correlation coefficient was 0.617, and the cosine similarity was 0.598. After mapping using the model of this invention, the Pearson correlation coefficient between the predicted human ASPS expression profile and the real human ASPS expression profile increased to 0.688, the Spearman correlation coefficient increased to 0.714, and the cosine similarity increased to 0.690. These results demonstrate that the model of this invention can not only improve the cross-species consistency of expression profiles under healthy conditions but also enhance the mapping accuracy of mouse expression profiles to the human disease expression space under disease conditions.

[0116] Furthermore, at the level of differential expression in diseases, differentially expressed genes in ASPS mice relative to healthy mice, and differentially expressed genes in ASPS humans relative to healthy humans, were calculated. For the model results of this invention, the human expression profiles of ASPS mice and healthy mice mapped by the model of this invention were first obtained, and then the predicted human differential expression effect size of ASPS predicted by the model of this invention relative to the predicted healthy control was calculated. Subsequently, using the differential expression results of real human ASPS relative to healthy humans as a reference, the recovery ability of the direct extrapolation results from mice and the prediction results of the model of this invention for human ASPS differential expression signals was compared.

[0117] In one specific implementation, 1.5 was used as the absolute value threshold for the differential expression effect size, and it was required that the predicted results of mice or the model of this invention showed the same direction of change as the differential expression results of real human ASPS. The results showed that among 11,933 intersecting homologous genes, 1,765 differentially expressed genes from mice, whose direct extrapolation results were consistent with the differential expression profile of real human ASPS, were in the same direction; the model of this invention predicted 1,870 differentially expressed genes from mice, whose prediction results were consistent with the differential expression profile of real human ASPS. The model of this invention added a net 105 differentially expressed genes consistent with human ASPS compared to direct extrapolation from mice. Simultaneously, the predicted results of the model of this invention included 695 human ASPS-consistent differentially expressed genes that were not captured by direct extrapolation from mice but were recovered after model mapping, suggesting that the model of this invention can recover some human disease-related signals masked by cross-species expression differences from the mouse disease expression profile.

[0118] At the level of genes related to disease mechanisms, genes newly captured in the model of this invention and whose differential expression direction is consistent with that of human ASPS were labeled and interpreted. The results showed that some genes restored by the model of this invention can be classified into functional modules related to ASPS pathological features, including epithelial-mesenchymal transition, extracellular matrix remodeling, inflammatory and immune microenvironment, complement activation, and tumor-microenvironment interactions. For example, FGF2, ITGA2, PLOD2, SFRP1, FLNA, and FAS can be associated with interstitial state, ECM remodeling, tissue invasion, or perivascular microenvironment; CD83 and PTX3 can be associated with TNFα / NF-κB-mediated inflammation and immune activation; CFB, LYN, PHEX, and PRCP can be associated with complement, myeloid immunity, or changes in the inflammatory microenvironment.

[0119] At the pathway level, GSEA analysis was further performed on the differential expression results predicted by human, mouse direct extrapolation, and the model of this invention. The GSEA can use the Hallmark gene set in the MSigDB database; in one implementation of this embodiment, significantly differentially expressed genes with an absolute effect size greater than 1.5 and a significance q-value < 0.05 were used for pre-sorting enrichment analysis, and human gene markers were used as unified gene markers. By comparing the enriched pathways of the three types of results—human, mouse direct extrapolation, and the model of this invention—the pathways newly captured by the model of this invention compared to mouse direct extrapolation and related to the human ASPS mechanism were identified.

[0120] It should be noted that, Figure 5The English names shown on the left are the standard pathway identifiers for the Hallmark gene set in the MSigDB database, which are professional terms used in bioinformatics pathway enrichment analysis. For ease of understanding, their Chinese definitions include: E2F_TARGETS (E2F target genes), G2M_CHECKPOINT (G2 / M phase checkpoint), ALLOGRAFT_REJECTION (allogeneic transplant rejection), INTERFERON_GAMMA_RESPONSE (interferon-gamma response), COMPLEMENT (complement pathway), TNFA_SIGNALING_VIA_NFKB (tumor necrosis factor α-mediated NF-κB signaling pathway), ANGIOGENESIS (angiogenesis), COAGULATIO (coagulation). N represents coagulation, XENOBIOTIC_METABOLISM represents the metabolism of exogenous compounds, KRAS_SIGNALING_UP represents KRAS signal upregulation, APOPTOSIS represents apoptosis, MITOTIC_SPINDLE represents the mitotic spindle, INTERFERON_ALPHA_RESPONSE represents the interferon-alpha response, p53_PATHWAY represents the p53 signaling pathway, APICAL_SURFACE represents the apical surface, ESTROGEN_RESPONSE_LATE represents the late estrogen response, and INFL... AMMATORY_RESPONSE represents the inflammatory response, GLYCOLYSIS represents glycolysis, APICAL_JUNCTION represents apical junctions, EPITHELIAL_MESENCHYMAL_TRANSITION represents the epithelial-mesenchymal transition pathway, SPERMATOGENESIS represents spermatogenesis, CHOLESTEROL_HOMEOSTASIS represents cholesterol homeostasis, UV_RESPONSE_UP represents upregulation of the ultraviolet radiation response, and ESTROGEN_RESPONSE_EARLY represents the early estrogen response. MTORC1_SIGNALING is the mTORC1 signaling pathway, UNFOLDED_PROTEIN_RESPONSE is the unfolded protein response, HEDGEHOG_SIGNALING is the Hedgehog signaling pathway, MYC_TARGETS_V1 is the MYC target gene V1, UV_RESPONSE_DN is the downregulation of ultraviolet radiation response, ANDROGEN_RESPONSE is the androgen response, and IL6_JAK_STAT3_SIGNALING is the interleukin-6 / JAK / STAT3 signaling pathway.

[0121] The results are as follows Figure 5The results show that the model prediction results of this invention can newly capture multiple human upregulated pathways related to the ASPS mechanism, including EPITHELIAL_MESENCHYMAL_TRANSITION (epithelial-mesenchymal transition pathway), TNFA_SIGNALING_VIA_NFKB (tumor necrosis factor α-mediated NF-κB signaling pathway) and COMPLEMENT (complement pathway). Among them, EPITHELIAL_MESENCHYMAL_TRANSITION (epithelial-mesenchymal transition pathway) is associated with mesenchymal state, ECM remodeling, invasion, perivascular microenvironment, and tissue infiltration, and can be linked to ASPS characteristics such as abundant vascularity, metastatic tendency, and local tissue infiltration; TNFA_SIGNALING_VIA_NFKB (tumor necrosis factor α-mediated NF-κB signaling pathway) is associated with inflammatory microenvironment, immune activation, and tumor-immune interaction, and can be linked to ASPS immune microenvironment and the background of immunotherapy response; COMPLEMENT (complement pathway) is associated with complement system, inflammatory response, and myeloid immune microenvironment, and can serve as auxiliary evidence of changes in ASPS immune microenvironment.

[0122] In the aforementioned pathways, the genes predicted by the model of this invention and those co-located in the enriched pathways with the actual human ASPS results include: FGF2, ITGA2, PLOD2, SFRP1, FLNA, and FAS in the EPITHELIAL_MESENCHYMAL_TRANSITION pathway; CD83 and PTX3 in the TNFA_SIGNALING_VIA_NFKB pathway (tumor necrosis factor α-mediated NF-κB signaling pathway); and CFB, LYN, PHEX, and PRCP in the COMPLEMENT pathway. The co-enrichment of these genes and pathways indicates that the model of this invention not only improves the overall similarity of expression profiles and the consistency of differentially expressed gene directions, but also recovers more mechanistically valuable human disease-related functional modules in the expression space of the target species.

[0123] This embodiment demonstrates that the model of this invention can be used not only for cross-species expression profile prediction but also for interpreting disease mechanisms. Compared to direct extrapolation of homologous genes from the source species, the model of this invention can more accurately map the expression profile of mouse disease states to the human disease expression space and recover more gene and pathway signals consistent with the expression patterns of real human diseases. Through homology gating, biological structure masking, and pathway constraint decoding, the model of this invention can enhance the ability to translate animal disease models into explanations of human disease mechanisms, providing more reliable human expression profile evidence for disease mechanism research, animal model evaluation, target discovery, and drug action mechanism analysis.

[0124] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A cross-species expression profile mapping model, characterized in that, It includes an expression input layer, a homology gene confidence gating module, a biological structure masking module, a mask variational encoder, a pathway constraint decoder, and an expression output layer; The expression input layer is used to receive the relative expression profiles of the source species, which are located in a unified homologous gene expression feature space and have been converted by tissue-specific stable housekeeping genes. The homologous gene confidence gating module is used to assign differential gating weights to the expression characteristics of different homologous genes; The biological structure masking module is used to perform random gene masking and pathway-level structure masking during the model training phase; The masked variational encoder is used to encode the masked input representation spectrum, generate latent variable distribution parameters including mean vector and log-variance vector, and obtain the latent representation based on the latent variable distribution parameters through reparameterized sampling; The pathway constraint decoder includes a gene-level decoding branch and a pathway activity decoding branch. The gene-level decoding branch is used to generate gene-level predicted expression profiles, and the pathway activity decoding branch is used to generate pathway activity vectors and obtain gene expression correction terms through the pathway-gene association matrix. The expression output layer is used to combine gene-level predicted expression profiles with gene expression correction terms to output the predicted expression profile of the target species.

2. The cross-species expression profile mapping model according to claim 1, characterized in that, The homologous gene confidence gating module includes a gating weight initialization unit, a gating parameter learning unit, and an element-wise weighting unit; The gating weight initialization unit generates initial confidence weights based on at least one of expression stability, cross-species expression correlation, tissue stability, and confidence of homology correspondence. The gating parameter learning unit updates the gating weights during model training. The element-wise weighted unit multiplies the source species input expression profile with the gating weights element-wise.

3. The cross-species expression profile mapping model according to claim 1, characterized in that, The biological structure masking module includes random gene masking units and pathway-level structure masking units; Random gene masking units are used to randomly mask some gene features; The pathway-level structure masking unit is used to simultaneously mask multiple gene features in the same biological functional module according to a preset pathway-gene relationship.

4. The cross-species expression profile mapping model according to claim 1, characterized in that, The path constraint decoder is also configured to: The target pathway activity is calculated based on the target expression profile and pathway-gene association matrix of the target species. The predicted pathway activity is calculated based on the predicted expression profile of the target species and the same pathway-gene association matrix. The pathway consistency loss is formed based on the difference between the target pathway activity and the predicted pathway activity.

5. A cross-species expression profile mapping method, characterized in that, Execution is based on a cross-species expression profile mapping model as described in any one of claims 1-4; The methodology includes the model building phase, the model training phase, and the application prediction phase. During the model building phase: The transcriptome expression profile matrices and sample annotation information of the source and target species are obtained, and the quality control of the transcriptome expression profile matrices is performed at the sample and gene levels. Based on the homologous gene correspondence between the source and target species, the expression matrices of the source and target species are feature-aligned to construct a unified homologous gene expression feature space. The expression matrix is ​​then converted into a relative expression representation of housekeeping genes based on tissue-specific stable housekeeping genes. Cross-species training sample pairs are constructed based on the same tissue, tissues with preset correspondences, or designated target tissues. Each training sample pair includes at least the source species input expression profile and the target species target expression profile. The source species input expression profile is input into the homologous gene confidence gating module, and the expression features of each homologous gene are weighted to obtain the gated input expression profile. During the model training phase: The gated input expression profile is input into the biological structure masking module, where random gene masking and pathway-level structure masking based on biological functional gene sets are performed to obtain the masked input expression profile. The masked input expression profile is then input into the mask variational encoder to generate latent variable distribution parameters, and latent variables are obtained through reparameterization. The latent variables are input into the pathway constraint decoder. Gene-level predicted expression profiles are generated through the gene-level decoding branch, and pathway activity vectors are generated through the pathway activity decoding branch. Gene expression correction terms are obtained based on the pathway-gene association matrix mapping. The gene-level predicted expression profiles and gene expression correction terms are combined to obtain the predicted expression profiles of the target species. The model is jointly trained based on gene-level reconstruction loss, KL divergence loss of latent variable distribution, pathway consistency loss of predicted expression profiles, supervision loss of pathway activity decoding branch, and gating weight regularization loss. In the application prediction phase: After processing the source species expression profile to be predicted according to the unified homologous gene expression feature space and the relative expression representation of housekeeping genes, the data is input into the trained model to obtain the predicted expression profile of the target species corresponding tissue, the specified target tissue, or the specified target expression state.

6. The cross-species expression profile mapping method according to claim 5, characterized in that, The homologous gene correspondence includes a one-to-one homologous gene correspondence; the source species expression matrix and the target species expression matrix are aligned according to homologous gene identifiers and a unified feature order, so that the model input features and output features correspond to the same homologous gene pair respectively.

7. The cross-species expression profile mapping method according to claim 5, characterized in that, When a single tissue-specific stable housekeeping gene is used as a reference gene, the first In the nth sample Relative expression effect size of housekeeping genes for each gene for: ,in, For the first In the nth sample The expression values ​​of each gene after logarithmic transformation The expression value of the reference gene in the same sample after logarithmic transformation; When using multiple tissue-specific stable housekeeping genes, a reference expression baseline is constructed using the mean, median, or weighted statistic of the expression values ​​of the multiple housekeeping genes.

8. The cross-species expression profile mapping method according to claim 5, characterized in that, The homologous gene credibility gating module generates an initial credibility score based on at least one of the following: expression stability of each homologous gene in the source and target species, expression correlation in cross-species training sample pairs, tissue stability, and credibility of homologous correspondence. The initial credibility score is then set as the initial value of a fixed gating weight or a learnable gating weight.

9. The cross-species expression profile mapping method according to claim 8, characterized in that, Let the input expression profile of the source species be... The gating weight vector is Then the gated input expression spectrum ; Among them, the gating weights are limited to a preset range. Internally, and composed of trainable parameters pass get.

10. The cross-species expression profile mapping method according to claim 5, characterized in that, The biological structure masking module generates a mask vector. and in accordance with The input representation spectrum after masking is obtained; Random gene masking randomly masks some gene features according to a preset masking ratio. Pathway-level structure masking is based on gene sets from Reactome, KEGG, Gene Ontology, MSigDB, or other biological functional databases, while simultaneously masking at least some member genes in the selected functional gene set.

11. The cross-species expression profile mapping method according to claim 5, characterized in that, The masked variational encoder comprises a fully connected layer, a normalization layer, a nonlinear activation layer, and a random deactivation layer connected in sequence, and outputs the latent variable mean vector μ and the log-variance vector, respectively. ; Reparameterization operation according to Obtain latent variables ,in, ε follows a standard normal distribution.

12. The cross-species expression profile mapping method according to claim 5, characterized in that, Let the output of the gene-level decoding branch be The output of the active decoding branch of the pathway is The pathway-gene association matrix is The predicted expression profile of the target species ,in, Adjust the weights for preset or learnable pathways.

13. The cross-species expression profile mapping method according to claim 5, characterized in that, Total model loss according to Calculate; where, The gene-level reconstruction loss between the predicted expression profile of the target species and the target expression profile of the target species is... The KL divergence loss is the difference between the latent variable posterior distribution and the pre-defined prior distribution. The loss of consistency between the predicted expression profile of the target species and the target expression profile of the target species in the pathway activity space. The monitoring loss between the output of the pathway activity decoding branch and the target pathway activity is... The loss is the deviation between the gating weights and the initial confidence weights after training. β , , and These are the weight parameters for the corresponding loss terms.

14. The cross-species expression profile mapping method according to claim 5, characterized in that, In the early stages of model training, the weights of the KL divergence loss are increased round by round, and model parameters are selected based on the validation set loss. The training set is used to update model parameters, the validation set is used for model selection and early stopping judgment, and the independent test set does not participate in model parameter updates or model selection.

15. The cross-species expression profile mapping method according to claim 5, characterized in that, The source species and target species are any combination of species with matching sets of homologous genes; the transcriptome expression profile includes a reference tissue expression profile, a disease tissue expression profile, a drug treatment expression profile, or a genetically modified model expression profile; the predicted expression profile is used for target tissue expression feature recovery, tissue-specific gene identification, drug toxicity risk assessment, disease mechanism analysis, or animal model transformation evaluation.

16. A cross-species expression profile mapping system, characterized in that, A cross-species expression profile mapping method according to any one of claims 5-15, the system comprising a data acquisition module, a data quality control module, a homologous gene anchoring module, an expression representation construction module, a paired sample construction module, a model training module, and an application prediction module; the data acquisition module is used to acquire transcriptome expression profile matrices and sample annotation information of the source species and the target species; the data quality control module is used to perform quality control at the sample level and the gene level; the homologous gene anchoring module is used to construct a unified homologous gene expression feature space; the expression representation construction module is used to construct relative expression representations of housekeeping genes based on tissue-specific stable housekeeping genes; the paired sample construction module is used to construct training sample pairs consisting of the input expression profile of the source species and the target expression profile of the target species; The model training module includes a cross-species expression profile mapping model as described in any one of claims 1 to 5, and is trained based on gene-level reconstruction loss, KL divergence loss, pathway consistency loss, pathway activity decoding branch supervision loss, and gate weight regularization loss; the application prediction module is used to input the expression profile of the source species to be predicted into the trained cross-species expression profile mapping model and output the predicted expression profile of the target species.

17. A cross-species expression profile mapping system according to claim 16, characterized in that, It also includes a model evaluation module and a downstream analysis module; the model evaluation module is used to evaluate the model performance based on at least one of the following: reconstruction error between the predicted expression profile and the actual target species expression profile, Pearson correlation coefficient, Spearman correlation coefficient, cosine similarity, pathway activity consistency, sensitivity or specificity; the downstream analysis module is used to perform target tissue expression feature recovery, tissue-specific gene identification, drug toxicity risk classification, disease differential expression analysis or pathway enrichment analysis based on the predicted expression profile of the target species.

18. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores program instructions, and when the processor executes the program instructions, it implements the cross-species expression profile mapping method according to any one of claims 5 to 15.

19. A computer-readable storage medium having program instructions stored thereon, characterized in that, When the program instructions are executed by the processor, the cross-species expression profile mapping method according to any one of claims 5 to 15 is implemented.

20. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the cross-species expression profile mapping method according to any one of claims 5 to 15.

Images