Automatic construction method of metabolic model for non-model species based on protein language model

Abstract

A method for automatically constructing metabolic models for non-model species based on a protein language model utilizes a gene homology inference method based on a pre-trained large model, combined with a protein sequence alignment method, to construct gene homology relationships across species. Using model organism GEMs as templates, GEMs for non-model organisms are constructed based on gene homology relationships and non-homologous gene annotations. Gap filling is then performed to optimize the metabolic model structure. This method significantly improves the accuracy of gene homology relationships, the speed and quality of constructing non-model organism GEMs, and the predictive performance of the resulting models.

Description

Technical Field

[0001] The present invention relates to a technology in the field of genetic engineering, specifically a method for automatically constructing high-precision genome-scale metabolic network models (GEMs) of non-model species by using pre-trained large model technology to obtain gene homology relationships. Background Art

[0002] Cellular metabolic network models based on genomic annotation information are important research tools in systems biology and metabolic engineering. Currently, gene homology identification in modeling newly sequenced species is primarily performed using sequence alignment methods, such as blast. However, sequence-based alignment methods are limited in their ability to search for distant homologous relationships and accurately annotate gene functions. Consequently, the metabolic models constructed by most current automated modeling tools are of poor quality and lack sufficient prediction accuracy, failing to meet the demand for accurate metabolic model prediction capabilities. Summary of the Invention

[0003] In response to the shortcomings of existing technologies that cannot effectively identify protein homology relationships brought about by structural similarity, resulting in poor quality of cellular metabolic network models, low prediction accuracy and high cost of model construction, the present invention proposes a method for automatically constructing metabolic models of non-model species based on a protein language model. A method combining blast sequence alignment and a pre-trained deep learning model is used to perform homology alignment of all proteins in the cell, and multiple protein function prediction tools are integrated to annotate non-homologous genes, thereby constructing GEMs. This can significantly improve the accuracy of gene homology relationships, and enhance the speed and quality of constructing GEMs of non-model organisms, as well as the predictive performance of the final model.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a method for automatically constructing a non-model species metabolic model based on a protein language model, comprising:

[0006] Step 1: Determine the gene homology relationship between the target strain and the reference strain based on protein sequence information comparison and gene homology comparison based on the pre-trained model;

[0007] Step 2: Functionally annotate the non-homologous genes in the target strain by integrating multiple protein function annotation methods.

[0008] Step 3: Using the GEMs of model organisms as templates, construct GEMs of non-model organisms through the gene homology obtained in step 1 and the functional annotation of non-homologous genes obtained in step 2, and then optimize the structure of the target model through gap filling to obtain species-specific GEMs.

[0009] The model for gene homology comparison is as follows: after generating sequence embedding vectors using a pre-trained large language model (ESM-1b) through the PLMSearch model, a bilinear mapping neural network is used to obtain the scores of the degree of homology between proteins.

[0010] Cross-species gene homology inference refers to the inference of gene homology relationships by combining protein sequence alignment with homology alignment results from a deep learning model. Specifically, this involves using bidirectional blast comparison (BBH) to compare gene homology relationships between the genomes of the target and reference strains, filtering the comparison results based on identity, coverage, and the presence of a bidirectional best match to obtain sequence-based orthologous relationships. Subsequently, the PLMSearch model is used to perform protein homology alignment based on a pre-trained model, and the alignment results are filtered using a selected threshold to determine the homology relationships between the target and reference strains. Subsequently, pairwise blast comparisons are performed on genes in the homology relationships obtained using the deep learning model to remove homologous genes with low identity and coverage. Finally, the gene homology relationships are integrated with the homology relationships obtained from the deep learning comparison to obtain a complete genetic homology relationship between the target strain and the reference strain.

[0011] The functional annotation of non-homologous genes refers to: providing a variety of protein function annotation methods, including the traditional EggNOG-mapper annotation tool, EC coding annotation tools based on deep learning models such as CLEAN and DeepECTransformer, and the PLMSearch homology search process based on Rhea database protein sequences, which are used to annotate non-homologous genes with metabolic reactions and discover potential metabolic reactions to improve model accuracy.

[0012] Constructing GEMs for non-model organisms involves automatically constructing metabolic models for target strains based on the obtained gene homology relationships and established GEMs for reference strains, combined with functional annotation results for non-homologous genes. Specifically, in the target strain, gene homology relationships are leveraged to add reactions and metabolites corresponding to homologous proteins, modifying gene-protein-reaction mappings. Based on annotation results for non-homologous genes, additional reactions are added to the initially constructed metabolic model using the Rhea metabolic reaction database.

[0013] The gap filling process involves selecting glucose as the sole carbon source and using the mature model of a reference strain as a reference to fill the gap, ensuring that the target species GEMs can simulate normal cell growth on the selected culture medium, thereby further improving the model quality and predictive performance.

[0014] The present invention relates to a system for implementing the above-mentioned method, comprising: a sequence alignment unit, a non-homologous gene annotation unit, and GEMS preliminary and final construction units, wherein: the sequence alignment unit performs gene (protein) homology alignment based on the proteome sequence information of the target strain and the reference species using the traditional blastp sequence alignment algorithm and the deep learning algorithm PLMSearch based on a pre-trained model, respectively, to integrate and obtain gene homology relationship results. The non-homologous gene unit integrates multiple protein function annotation methods to perform metabolic reaction annotation for the non-homologous genes of the target strain. The GEMs preliminary construction unit adds reaction and metabolite information corresponding to homologous proteins and gene-protein-reaction correspondences to the target strain based on the obtained gene homology relationship and the model information of the reference model strain, thereby obtaining initial GEMs of the target strain. The GEMs final construction unit uses the obtained metabolic reaction annotation results of the non-homologous genes based on the initial GEMs of the target strain to add additional reactions to the initial GEMs, then uses the mature model of the reference strain as a reference to fill in the reaction gaps using glucose as the sole carbon source, ultimately obtaining high-quality GEMs of the target strain.

[0015] Technical Effects

[0016] This invention combines a gene homology annotation method based on a pre-trained model-based gene (protein) homology alignment algorithm (i.e., PLMSearch) with traditional sequence alignment methods. It also develops an enzyme catalytic reaction prediction process that integrates multiple protein function prediction tools. Furthermore, it integrates these two methods to form an automatic algorithm for constructing high-precision GEMs for non-model organisms, using GEMs from model species (including but not limited to human cells, Saccharomyces cerevisiae, and Escherichia coli) as templates and relying on information from open-source biological reaction databases. Compared with existing technologies, this invention enables high-throughput GEM construction, automates the entire process, and significantly reduces the time required to build high-quality models. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the method flow for the gene homology inference step of the present invention

[0018] Figure 2 Schematic diagram of the method flow for the GEMs construction steps of the present invention

[0019] Figure 3Schematic diagram of the non-homologous gene annotation tool of the present invention;

[0020] Figure 4 A schematic diagram of the scale increase of the GEMs construction process model of the present invention;

[0021] Figure 5-Figure 8 Schematic diagram of the embodiment effect; DETAILED DESCRIPTION

[0022] like Figure 1 As shown, this embodiment relates to a method for implementing an automatic construction algorithm of a high-precision metabolic model based on a protein language model, including:

[0023] Step 1: Figure 1 As shown, a bidirectional blast comparison is used to find the gene homology relationship between the target strain genome and the reference strain genome. The specific steps are: use blastp (protein to protein) to use all gene sequences of the target strain as query sequences, compare them to the genome database of the reference strain, and record the best match of each query sequence in the reference strain genome; then use all gene sequences of the reference strain as query sequences, compare them to the genome database of the target strain, and record the best match of each query sequence in the target strain genome; through the comparison results, the best mutual match between the target strain gene and the reference species gene is screened; finally, the obtained homologous gene pairs are screened based on the sequence alignment similarity Identity and sequence coverage to obtain the gene homology relationship based on sequence alignment.

[0024] Step 2: Figure 1 As shown, the PLMSearch model was used to compare the homology relationships of genes in the target genome and the reference genome one by one, and screened according to a certain threshold; then, blastp was used to compare the genes in the screened homologous gene pairs, and the homologous gene pairs were screened according to Identity and Coverage, filtering out homology relationships with extremely low sequence similarity; the homology relationships obtained based on sequence and deep learning models were integrated to obtain the complete gene homology relationship between the target strain and the reference strain.

[0025] Step 3: Figure 2 As shown, based on the obtained gene homology relationship and the mature model of the reference strain, a metabolic model of the target strain was constructed, and the reactions and metabolites corresponding to the homologous proteins were added using the gene homology relationship, and the gene-protein-reaction correspondence was modified.

[0026] Step 4: Since gene homology correspondence may not necessarily cover all metabolic reactions of the target strain, gene-protein-reaction annotation is performed on genes that do not have homology correspondence in the target strain, such as Figure 3 As shown, this patent provides a variety of protein function annotation tools, including: predicting the biological reaction information catalyzed by these proteins based on EggNOG-mapperv2, such as the KEGG number (i.e., KO number) of the reaction; predicting the EC code of the biological reactions catalyzed by these proteins based on the deep learning model of CLEAN and DeepECtransformer; using PLMSearch to search for homologous proteins in the reaction-related protein sequences in the Rhea database to predict possible metabolic reactions; in the modeling process, only the EggNOG-mapper tool can be used for rapid annotation, or the predicted information can be integrated to obtain reaction information based on the Rhea database, and then the corresponding metabolic reactions of non-homologous genes can be added to the model.

[0027] Step 5: Glucose is selected as the sole carbon source, and the mature model of the reference strain (such as the GEMs-Yeast9 model of Saccharomyces cerevisiae) is used as a reference to fill the reaction gaps, thereby further improving the model.

[0028] In a specific experiment, using Saccharomyces cerevisiae GEMs-Yeast9 as a reference strain model and Candida albicans as an example, rapid construction of GEMs for non-model species was carried out. The results are as follows: In the first step of gene homology inference, using thresholds of PID>=0.6 and Coverage>=0.25, 808 homology relationships were identified through bidirectional BLAST sequence alignment and screening. PLMSearch model structure alignment and screening using thresholds of similarity>=0.8, PID>=0.2, and Coverage>=0.5 yielded 5642 homology relationships, totaling 5698 after integration. In the second step of metabolic model construction, metabolic reactions were added based on gene homology, non-homologous gene annotation, and reaction gap filling. Here, only non-homologous gene annotation was performed using EggNOG-mapper.

[0029] like Figure 4 and Figure 5As shown in the figure, after the three steps were completed, the number of reactions, genes, and metabolites in the model were [3813, 3912, 3929], [1030, 1079, 1100], and [2690, 3110, 3110], respectively. Using the C. albicans GEMs generated using the above method to predict the number of available carbon sources, the prediction precision, sensitivity, specificity, and accuracy were 75%, 81.8%, 75%, and 78.3%, respectively, for 24 available carbon sources. The constructed metabolic model was phenotypically comparable to the published manually corrected model in terms of carbon source utilization prediction accuracy. The model that had undergone extensive manual correction had a prediction accuracy of 82.6% for these 24 available carbon sources.

[0030] In addition, the prediction performance of the high-throughput automatically constructed model of the present invention on 24 available carbon sources is also comparable to that of the manually calibrated model on another 210 specific yeast species. Figure 6 The figure shows a comparison of the average values ​​of the five indicators (Precision, Specificity, Sensitivity, Accuracy, and F1-score) of carbon source utilization prediction performance between the metabolic models of 210 yeast species automatically constructed by the high-throughput method of the present invention and the published manually corrected models. The above results show that the species-specific GEMs established by the present invention are very close to the published metabolic models that have undergone extensive manual modifications in terms of prediction accuracy.

[0031] In the two yeast species C.albicans and S.pombe, the non-homologous gene functional annotation results at the EC coding level were more accurate than those of a single method, using the species EC coding information in the UniProt database as a reference and integrating multiple functional annotation tools such as EggNOG-mapper, CLEAN, DeepECTransformer and Rhea database protein search. Figure 7 As shown in FIG, a comparison of three prediction performance indices (Recall, Precision, F1-Score) of the non-homologous gene annotation results obtained by using various non-homologous gene function annotation tools used in the present invention in Candida albicans (C.albicans) separately and in combination, the indices are calculated with reference to the EC code annotation results obtained by sorting out the UniProt database; Figure 8As shown, a comparison chart of three prediction performance indicators (Recall, Precision, F1-Score) of the non-homologous gene annotation results obtained by using various non-homologous gene function annotation tools used in the present invention separately and in combination in fission yeast (S. pombe). The indicators are calculated with reference to the EC code annotation results obtained by sorting out the UniProt database.

[0032] Compared to conventional GEM construction, which often requires lengthy manual corrections and repeated iterative updates, the present invention can efficiently and automatically construct GEMs from scratch for a large number of non-model species, significantly saving time and effort. Compared to metabolic modeling based solely on sequence alignment, the present invention introduces a gene homology inference algorithm based on pre-trained models and a multi-tool integrated functional annotation process for non-homologous proteins, effectively optimizing model structure and significantly improving GEM quality and prediction accuracy.

[0033] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principles and purpose of the present invention. The scope of protection of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. All implementation schemes within its scope shall be subject to the constraints of the present invention.

Claims

1. A method for automatically constructing a metabolic model of a non-model species based on a protein language model, characterized in that: A gene homology inference method based on a pre-trained large model was combined with a protein sequence alignment method to construct gene homology relationships across species. Using GEMs of model organisms as templates, GEMs of non-model organisms were constructed based on gene homology relationships and non-homologous gene annotation results, and then gap filling was performed to optimize the metabolic model structure. The protein sequence alignment method is to use bidirectional blast comparison (BBH) to compare the gene homology relationship between the genomes of the target strain and the reference strain, and screen the comparison results based on whether there is a bidirectional best match based on sequence alignment similarity and sequence coverage to obtain sequence-based orthologous relationships; The gene homology inference method based on a pre-trained large model is: the published deep learning model based on the ESM-1b large model - PLMSearch, which uses a bilinear mapping neural network to fine-tune the output of the pre-trained large model. The model is trained on a public homologous protein database. The model accepts two protein sequences as input and outputs the similarity between the two proteins to assess the closeness of gene homology relationships; The combination of the gene homology inference method based on the pre-trained large model and the protein sequence alignment method includes: using the PLMSearch model to perform protein homology alignment based on sequence information, and selecting a threshold to screen the alignment results; then performing pairwise blast alignments on the proteins in the homology relationships obtained based on deep learning, screening and removing homology relationships with too low sequence alignment similarity and sequence coverage; then integrating the homology relationships obtained based on deep learning with the gene homology relationships obtained based on bidirectional blast alignment to obtain a complete gene homology relationship of the target strain relative to the reference strain; The non-homologous gene annotation results are: prediction of possible metabolic reactions by integrating multiple protein function annotation tools, including EggNOG-Mapper, CLEAN, DeepECtransformer, and PLMSearch homology search based on the protein sequence of the Rhea database; in the prediction process, EggNOG-mapper v2 is used for annotation or the prediction information of the four methods is integrated to obtain reaction information based on the Rhea database and add the corresponding metabolic reactions of non-homologous genes to the model.

2. The method for automatically constructing a non-model species metabolic model based on a protein language model according to claim 1, characterized in that: The construction of GEMs of non-model organisms refers to: constructing a model of the target strain based on the obtained gene homology relationship and the mature model of the reference strain, adding reactions and metabolites corresponding to homologous proteins using the gene homology relationship, and modifying the gene-protein-reaction correspondence; Based on the functional annotation results of non-homologous proteins, corresponding reactions were added to the initially constructed model.

3. The method for automatically constructing a non-model species metabolic model based on a protein language model according to claim 1, characterized in that: The gap filling process involves selecting glucose as the sole carbon source and using the mature model of a reference strain as a reference to fill the gap, ensuring that the target species GEMs can simulate normal cell growth on the selected culture medium, thereby further improving the model quality and predictive performance.

4. A system for constructing a model based on sequence alignment and deep learning homology alignment to implement the method described in any one of claims 1 to 3, characterized in that: include: The sequence alignment unit, the deep learning-based homology alignment unit, the non-homologous gene function annotation unit, and the GEMS preliminary and final construction units include: the sequence alignment unit performs sequence alignment processing based on the proteomic sequence information of the target strain and the reference strain to obtain preliminary gene homology relationship results; the deep learning homology alignment unit uses a pre-trained deep learning model to perform structural homology alignment based on the proteomic sequence information of the target strain and the reference strain to obtain final gene homology relationship results; the non-homologous gene function annotation unit integrates multiple annotation tools to annotate the reaction information catalyzed by non-homologous proteins; the GEMs preliminary construction unit adds the reaction and metabolite information corresponding to the homologous proteins and the gene-protein-reaction correspondence to the target strain based on the obtained gene homology relationship and the model information of the reference model strain to obtain the initial GEMs of the target strain; the GEMs final construction unit uses the non-homologous protein function annotation results to add additional reactions to the initial GEMs based on the initial GEMs of the target strain, and then uses the mature model of the reference strain as a reference to fill the reaction gaps with glucose as the sole carbon source, and finally obtains high-quality GEMs of the target strain.

Images