Formulation bacterial community construction method and device, electronic equipment and storage medium

Abstract

The application provides a formula flora construction method and device, electronic equipment and storage medium, and belongs to the field of bioinformatics, and comprises the following steps: converting a phylogenetic tree constructed by a microbial genome sequence into a weight matrix; inputting a central logarithmic ratio transformed original abundance matrix into an iterative random forest model; correcting the original feature importance to smooth feature importance by using the weight matrix to update the feature sampling weight in the next round; and extracting high-order ecological interaction rules based on the decision path when the model converges to output a candidate core flora, and then determining a target formula flora. The application breaks the independence assumption of traditional abundance data by introducing phylogenetic relationships to smooth correct features, effectively overcomes the sparsity of high-dimensional sequencing data and false positive interference, accurately captures and retains the real interspecies synergistic relationship driven by evolution, and thus guarantees the efficient and stable colonization and precise targeted metabolic regulation of the formula flora in the host microenvironment.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and more particularly to a method, apparatus, electronic device, and storage medium for constructing a formulated microbial community. Background Technology

[0002] As research into the correlation between human gut microbiota and health and disease deepens, precisely customized formulations of gut microbiota (or synthetic microbiota) capable of intervening in specific pathological states have become a core research and development direction for next-generation in vivo biopharmaceuticals and gut microbiota preparations. The development of formulation microbiota aims to precisely extract core strain combinations with key metabolic support capabilities and extremely high ecological stability from massive environmental or gut microbiome pools. The core overall technical requirement for this process is: how to accurately isolate incidental symbiotic noise from highly dimensional and sparse microbial community sequencing data, and to discover a truly biologically significant and ecologically stable synergistic set of candidate strains, thereby providing a highly robust underlying framework for subsequent targeted validation and formulation production.

[0003] To meet the requirements for constructing the aforementioned microbial communities, existing technologies typically rely on conventional bioinformatics statistical analysis or standard machine learning models for feature selection of core microbial communities. Most existing solutions directly acquire relative or absolute quantitative data on the microbial community, then input this abundance data directly into random forests, support vector machines, or univariate statistical test models to rank features by importance, and select the top-ranked microbial taxa as core microbial communities based on a threshold truncation of importance scores. Furthermore, to explore relationships between microbial communities, existing technologies often employ co-occurrence networks based on Pearson or Spearman correlation coefficients to infer interspecies associations.

[0004] However, the aforementioned technologies have significant drawbacks in practical applications. First, existing data processing workflows typically ignore the compositional data attribute inherent in community sequencing data. Directly inputting conventional quantitative features can easily lead to spurious correlations and statistical false positives due to data closure effects. Second, existing machine learning feature selection mechanisms, when dealing with highly complex micro-ecosystems, rely solely on the numerical distribution differences at the data sample level to perform isolated calculations and rankings of variable importance. This completely severs the homology constraints and biological attributes formed between microbial taxa during long-term natural selection. This black-box analysis model, which treats microbial species as absolutely independent variables, is highly prone to overfitting the model to the training set data. Finally, due to the lack of effective ways to capture complex synergistic relationships among multiple species, the key strain combinations screened by existing schemes often only reflect shallow, local, or paired linear correlations. They are unable to truly characterize the high-order ecological interaction networks that maintain system robustness when facing environmental disturbances. Ultimately, this results in the formulated microbial communities screened in this way facing extremely high colonization failure rates and the risk of ecological function decline in real host complex physiological environments. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for constructing a formulated microbial community. These solutions address the shortcomings of existing technologies, such as the closure effect leading to false positives in microbial component data, the tendency for model overfitting due to the fragmentation of homologous evolutionary constraints in feature screening, and the difficulty in uncovering high-order synergistic relationships among multiple species, resulting in high colonization failure rates and easy ecological function decline. The invention effectively overcomes false-positive interference from high-dimensional sparse sequencing data, accurately captures real, evolution-driven high-order interspecies synergistic interaction networks, and significantly improves the biological rationality and model robustness of core microbial community screening. This ensures the efficient and stable colonization of the formulated microbial community within the host microenvironment and precise targeted metabolic regulation.

[0006] This invention provides a method for constructing a formula microbial community, comprising:

[0007] Representative genomic sequences from multiple microbial taxa in the sample set were extracted to construct a phylogenetic tree;

[0008] Calculate the evolutionary distance between each pair of microbial taxonomic units in the phylogenetic tree, and convert the evolutionary distance into a phylogenetic weight matrix. The phylogenetic weight matrix is ​​used to characterize the kinship similarity between the microbial taxonomic units.

[0009] Obtain the original microbial abundance matrix of the sample set, perform a central log ratio transformation on the original microbial abundance matrix to obtain the feature matrix, and use the feature matrix as the feature dimension input to construct an iterative random forest model;

[0010] During the feature selection iteration of the iterative random forest model, the original feature importance of each microbial taxonomic unit is corrected to a smoothed feature importance using the phylogenetic weight matrix, so as to update the feature sampling weights of the iterative random forest model for the microbial taxonomic unit in the next iteration based on the smoothed feature importance;

[0011] When the iterative random forest model reaches the preset convergence condition, high-order ecological interaction rules are extracted based on the decision path and candidate core bacterial communities are output.

[0012] The target formulation microbial community was determined based on the candidate core microbial community.

[0013] According to a method for constructing a formulated microbial community provided by the present invention, the step of performing a central logarithmic ratio transformation on the original microbial abundance matrix to obtain a feature matrix includes:

[0014] Replace the zero-value features in the original microbial abundance matrix with preset non-zero values;

[0015] For the original microbial abundance matrix after replacement processing, determine the geometric mean parameter of the abundance vector of each sample in the sample set;

[0016] Obtain the relative ratio of each abundance value in the abundance vector of each sample to the corresponding geometric mean parameter;

[0017] The feature matrix is ​​obtained by performing a logarithmic scaling on all the relative ratios.

[0018] According to a method for constructing a formulated microbial community provided by the present invention, the step of calculating the evolutionary distance between each pair of microbial taxa in the phylogenetic tree and converting the evolutionary distance into a phylogenetic weight matrix includes:

[0019] Determine the nearest common ancestor node between each pair of microbial taxa in the phylogenetic tree;

[0020] Based on the first path span between each of the microbial taxonomic units and the root node of the phylogenetic tree, and the second path span between the nearest common ancestor node and the root node, the kinship distance between each pair of the microbial taxonomic units is determined;

[0021] The phylogenetic weight matrix, which characterizes the strength of kinship, is obtained by using a preset exponential decay kernel function to perform similarity mapping on the clan distance.

[0022] According to a method for constructing a formulated microbial community provided by the present invention, the phylogenetic weight matrix corrects the original feature importance of each microbial taxonomic unit to a smoothed feature importance, including:

[0023] For any target microbial taxonomic unit among all the aforementioned microbial taxonomic units, based on the phylogenetic similarity represented in the phylogenetic weight matrix, the original feature importance of all other microbial taxonomic units among all the aforementioned microbial taxonomic units is weighted and fused to determine the neighbor feature reference value of the target microbial taxonomic unit;

[0024] A first weight is assigned to the importance of the original features of the target microbial taxonomic unit using a preset smoothing parameter;

[0025] The complementary parameters of the smoothing parameter are used to assign a second weight to the neighbor feature reference value;

[0026] The original feature importance configured with the first weight is weighted and fused with the neighbor feature reference value configured with the second weight to obtain the modified feature importance of the target microbial taxonomic unit.

[0027] According to a method for constructing a formulated microbial community provided by the present invention, the step of extracting high-order ecological interaction rules based on the decision path therein and outputting candidate core microbial communities includes:

[0028] Extract the decision paths of each decision tree in the iterative random forest model to obtain a set of decision paths;

[0029] Identify the microbial taxonomic units commonly included in multiple decision paths within the decision path set to obtain at least one candidate taxonomic unit combination;

[0030] Based on the data resampling mechanism, the stability score of each candidate classification unit combination is determined;

[0031] The candidate core bacterial community is determined based on all combinations of candidate taxa that meet the set threshold for stability scores.

[0032] According to a method for constructing a formula microbial community provided by the present invention, the step of identifying microbial taxonomic units commonly included in multiple decision paths in the decision path set to obtain at least one candidate taxonomic unit combination includes:

[0033] Extract candidate microbial taxonomic units that serve as splitting nodes from each of the aforementioned decision paths;

[0034] Select at least one set of decision paths from the set of decision paths, and perform an intersection operation on the candidate microbial taxonomic units contained in each set of decision paths;

[0035] The set containing multiple candidate microbial taxonomic units obtained by intersecting the decision paths of each group is determined as a candidate taxonomic unit combination.

[0036] According to a method for constructing a formula microbial community provided by the present invention, the step of determining the stability score of each candidate taxonomic unit combination based on a data resampling mechanism includes:

[0037] Multiple random samplings with replacement are performed on the sample set to generate multiple sample subsets;

[0038] For each of the sample subsets, the steps of constructing an iterative random forest model to identify the combination of the candidate classification units are performed to obtain a set of validation combinations corresponding to each of the sample subsets;

[0039] Count the total number of times any candidate classifier combination exists in all the verified combination sets;

[0040] The ratio of the total number of times to the total number of samples in the subset is determined as the stability score of any candidate classification unit combination.

[0041] According to a method for constructing a formulation microbial community provided by the present invention, determining the candidate core microbial community based on all combinations of candidate taxa that satisfy a set threshold in stability scores includes:

[0042] Extract all candidate classification unit combinations whose stability scores are greater than or equal to the set threshold, and use them as candidate ecological functional groups;

[0043] The microbial taxonomic units contained in all the candidate ecological functional groups are aggregated and summarized, and the aggregated set of all the microbial taxonomic units is determined as the candidate core microbial community.

[0044] According to a method for constructing a formulation microbial community provided by the present invention, the step of determining the target formulation microbial community based on the candidate core microbial community includes:

[0045] The entity information of each microbial taxonomic unit in the candidate core bacterial community is input into a pre-trained biological language model for reasoning analysis to obtain the semantic score of the candidate core bacterial community.

[0046] Based on each microbial taxonomic unit in the candidate core bacterial community, a microbial metabolic network model is constructed.

[0047] Based on the microbial metabolic network model, community flux balance simulation is performed to determine the metabolic support index among the microbial taxa in the candidate core microbial community; the metabolic support index is used to characterize the intensity of metabolic mutualism between the microbial taxa acting as donors and the microbial taxa acting as recipients.

[0048] A comprehensive recommendation index is determined based on the semantic score and the metabolic support index.

[0049] The target formulation microbial community is selected from the candidate core microbial communities based on the comprehensive recommendation index.

[0050] According to a method for constructing a formulated microbial community provided by the present invention, the step of inputting the entity information of each microbial taxonomic unit in the candidate core microbial community into a pre-trained biological language model for reasoning analysis to obtain a semantic score of the candidate core microbial community includes:

[0051] Based on the taxonomic identifiers of each microbial taxonomic unit in the candidate core bacterial community, the corresponding standard entity names and biological metadata are obtained to construct structured prompt information;

[0052] The structured prompt information is used to perform feature retrieval in a pre-built external knowledge base to obtain prior knowledge fragments related to the candidate core bacterial community;

[0053] The prior knowledge fragments and the structured prompt information are fused and input into the biological language model to obtain the semantic score of the candidate core bacterial community output by the biological language model;

[0054] The biological language model performs step-by-step reasoning according to the cascade biochemical reaction sequence of substrate utilization analysis, intermediate product transformation analysis, and final product effect assessment to obtain the semantic score.

[0055] According to a method for constructing a formulated microbial community provided by the present invention, the step of constructing a microbial metabolic network model based on each microbial taxonomic unit in the candidate core microbial community includes:

[0056] Obtain the genome sequence of each of the microbial taxa in the candidate core microbial community;

[0057] Identify the genes encoding enzymes in the genome sequence, establish the mapping relationship between genes, proteins and biochemical reactions, and reconstruct the initial metabolic network model;

[0058] By introducing pre-defined pathological environment constraints, metabolic gaps in the initial metabolic network model are filled to obtain the microbial metabolic network model.

[0059] According to a method for constructing a formulated microbial community provided by the present invention, the step of introducing a preset pathological environment constraint and filling metabolic gaps in the initial metabolic network model to obtain the microbial metabolic network model includes:

[0060] Virtual culture medium boundary conditions are configured based on the intestinal microenvironment characteristics of the target disease as the pathological environment constraints. The virtual culture medium boundary conditions include the inflow threshold of specific amino acids and the initial concentration of specific nitrogen sources.

[0061] Identify key housekeeping responses that are missing in the initial metabolic network model due to insufficient sequencing coverage of the genome sequence;

[0062] Under the constraints of the virtual culture medium boundary conditions, the initial metabolic network model is iteratively optimized using a mixed integer linear programming algorithm to supplement the missing key housekeeping reactions into the initial metabolic network model;

[0063] When it is determined that the supplemented metabolic network model can generate a preset biomass in a simulated culture medium that meets the boundary conditions of the virtual culture medium, it is identified as the microbial metabolic network model.

[0064] According to a method for constructing a formulated microbial community provided by the present invention, community flux balance simulation is performed based on the microbial metabolic network model to determine the metabolic support index among the microbial taxonomic units in the candidate core microbial community, including:

[0065] Construct a virtual community model that includes independent strain compartments corresponding to each of the aforementioned microbial taxonomic units and a shared extracellular cavity;

[0066] Configure the material inflow boundary conditions for the shared extracellular chamber, and connect the microbial metabolic network model to the corresponding independent strain compartment;

[0067] Under the constraints of the material inflow boundary conditions, flux balance is solved with the maximization of total community biomass as the joint objective function to obtain the optimal flux vector for each biochemical reaction.

[0068] Based on the optimal flux vector, the generation rate of a specific exchange metabolite released from the donor microbial taxa unit into the shared extracellular cavity is extracted, as well as the consumption rate of the recipient microbial taxa unit absorbing the specific exchange metabolite from the shared extracellular cavity.

[0069] After obtaining the minimum value between the generation rate and the consumption rate, the metabolic support index is obtained by weighting and fusing the index with the preset importance weights of the specific exchange metabolites.

[0070] According to a method for constructing a formulation microbial community provided by the present invention, before screening the target formulation microbial community from the candidate core microbial communities based on the comprehensive recommendation index, the method further includes:

[0071] The genome sequence of each microbial taxonomic unit in the candidate core bacterial community is analyzed to detect whether there are preset pathogenic risk gene features, including virulence factors or antibiotic resistance genes located on mobile genetic elements.

[0072] During the execution of the community flux balance simulation, it is monitored whether the preset flux of harmful metabolites exceeds the safety limit value.

[0073] In response to the detection of the pathogenic risk gene feature or the generation throughput exceeding the safety threshold, the corresponding candidate core microbial community is removed from the candidate pool.

[0074] According to a method for constructing a formula microbial community provided by the present invention, the step of determining a comprehensive recommendation index based on the semantic score and the metabolic support index includes:

[0075] Obtain the stability score corresponding to the candidate core bacterial community;

[0076] Obtain the functional targeting score of the candidate core bacterial community for a specific pathological target in the community flux balance simulation.

[0077] The stability score, the functional targeting score, the semantic score, and the metabolic support index are weighted and fused together to obtain the comprehensive recommendation index.

[0078] According to a method for constructing a formulated microbial community provided by the present invention, after determining the target formulated microbial community, the optimal inoculation stoichiometry is further solved based on the law of mass action, specifically including:

[0079] Obtain the target steady-state relative abundance of each microbial taxa of the target formulation microbial community under simulated steady-state conditions;

[0080] Based on the steady-state relative abundance, a set of community dynamic differential equations including interspecific interaction coefficients, intrinsic growth rates, and environmental carrying capacity is constructed.

[0081] The physiological environmental parameters of the target application scenario are used as initial constraints to input the community dynamic differential equation system, and the inoculation concentration ratio of each microbial taxa that can enable the community to reach the target steady-state relative abundance after a preset colony cycle is derived in reverse as the optimal inoculation stoichiometry.

[0082] Based on the optimal inoculation stoichiometry, industrial preparation formulation data for the target microbial community are generated.

[0083] The present invention also provides a formulation microbial community construction device, comprising the following modules:

[0084] The phylogenetic tree construction module is used to extract representative genomic sequences from multiple microbial taxonomic units in the sample set to construct a phylogenetic tree;

[0085] The phylogenetic weight conversion module is used to calculate the evolutionary distance between each pair of microbial taxonomic units in the phylogenetic tree, and convert the evolutionary distance into a phylogenetic weight matrix, which is used to characterize the phylogenetic similarity between the microbial taxonomic units.

[0086] The abundance feature processing module is used to obtain the original microbial abundance matrix of the sample set, perform a central log ratio transformation on the original microbial abundance matrix to obtain a feature matrix, and use the feature matrix as the feature dimension input to construct an iterative random forest model.

[0087] The feature importance correction module is used to correct the original feature importance of each microbial taxonomic unit to a smoothed feature importance using the phylogenetic weight matrix during the feature selection iteration of the iterative random forest model, so as to update the feature sampling weights of the iterative random forest model for the microbial taxonomic unit in the next iteration based on the smoothed feature importance;

[0088] The higher-order rule extraction module is used to extract higher-order ecological interaction rules based on the decision path of the iterative random forest model when the iterative random forest model reaches the preset convergence condition, and output candidate core bacterial communities.

[0089] The target microbial community selection module is used to determine the target formulation microbial community based on the candidate core microbial communities.

[0090] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the formulation microbial community construction method as described above.

[0091] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the formulation microbial community construction method as described above.

[0092] The method, apparatus, electronic device, and storage medium for constructing formulated microbial communities provided by this invention introduce phylogenetic relationships to smooth the characteristics, breaking the independence assumption of traditional abundance data, effectively overcoming the sparsity and false positive interference of high-dimensional sequencing data, accurately capturing and preserving the real interspecies synergistic relationships driven by evolution, thereby ensuring the efficient homeostatic colonization and precise targeted metabolic regulation of the formulated microbial communities in the host microenvironment. Attached Figure Description

[0093] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0094] Figure 1 This is a flowchart illustrating a method for constructing a formula microbial community provided by the present invention.

[0095] Figure 2 This is an implementation logic diagram of a formula microbial community construction method provided by the present invention.

[0096] Figure 3 This is a schematic diagram of a technical route for multimodal data construction and phylogenetic tensor generation provided by the present invention.

[0097] Figure 4 This is a schematic diagram of a process for performing central logarithmic ratio transformation on the original microbial abundance matrix provided by the present invention.

[0098] Figure 5 This is a schematic diagram illustrating the process of determining a phylogenetic weight matrix provided by the present invention.

[0099] Figure 6 This is a schematic diagram of the process for determining phylogenetic distance provided by the present invention.

[0100] Figure 7 This is a schematic diagram of a process for correcting the importance of original features of microbial taxonomic units provided by the present invention.

[0101] Figure 8 This is an implementation logic diagram of a high-order ecological interaction mining technology driven by the Phy-iRF algorithm provided by the present invention.

[0102] Figure 9 This is a schematic diagram of the process for determining a candidate core bacterial community provided by the present invention.

[0103] Figure 10 This is a flowchart illustrating an iterative random forest algorithm provided by the present invention.

[0104] Figure 11 This is a flowchart illustrating a process for determining the stability score of candidate classification unit combinations provided by the present invention.

[0105] Figure 12 This is a schematic diagram of a process for determining the target formulation microbial community based on candidate core microbial communities, provided by the present invention.

[0106] Figure 13This is an implementation logic diagram of a semantic-physical dual-function decoding technology based on a biological language model and a microbial metabolic network model provided by the present invention.

[0107] Figure 14 This is a schematic diagram of a process for obtaining semantic scores of candidate core bacterial communities provided by the present invention.

[0108] Figure 15 This is one of the flowcharts for constructing a microbial metabolic network model provided by the present invention.

[0109] Figure 16 This is the second schematic diagram of a process for constructing a microbial metabolic network model provided by the present invention.

[0110] Figure 17 This is a schematic diagram of a process for determining the metabolic support index between microbial taxa in a candidate core microbial community, provided by the present invention.

[0111] Figure 18 This is a schematic diagram of the structure of the formula microbial community construction device provided by the present invention.

[0112] Figure 19 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0113] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0114] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Those skilled in the art will understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0115] The terms "first," "second," etc., used in this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more.

[0116] Before introducing the method, apparatus, electronic device, and storage medium for constructing the formulated microbial community provided by this invention, a brief explanation of some technical terms that may appear in subsequent embodiments of this invention will be provided, including:

[0117] 1) Core algorithm and data processing terminology.

[0118] 1) Phylogenetic Enhanced Iterative Random Forest (Phy-iRF) is a feature selection algorithm that introduces a phylogenetic tensor as a prior constraint. In this invention, the algorithm uses phylogenetic clan distance to smoothly correct the feature importance during the random forest iteration process, thereby accurately extracting a high-order ecological interaction skeleton with evolutionary conservation from sparse data.

[0119] 2) The Central Log-Ratio (CLR) transformation is a mathematical transformation method that projects microbial component data from simplex space to Euclidean space. In this invention, it is used to eliminate statistical artifacts (spurious negative correlations) caused by the limitation of summation in microbial abundance data and output a standardized feature matrix.

[0120] 3) Patristic distance refers to the path length between any two microbial nodes calculated based on a phylogenetic tree. This distance reflects not only the topological structure between species but also evolutionary time or genetic variation. In this invention, it is converted into prior similarity weights to correct the feature selection process in machine learning.

[0121] 4) Random Intersection Trees (RIT) is an algorithm for efficiently searching frequently co-occurring feature subsets in a decision tree path set. In this invention, it is used to transform the implicit "black box" model of random forests into interpretable high-order microbial interaction rules (such as third-order co-occurrence patterns).

[0122] 5) Compositional data refers to relative abundance data with a fixed sum (e.g., total abundance of 1 or 100%). This invention points out that directly using such data for network analysis can lead to spurious negative correlations between species that were not originally related.

[0123] (ii) Metagenomics and Bioinformatics Terminology.

[0124] 1) Live Biotherapeutic Products (LBPs) refer to synthetic microbial communities composed of well-defined live bacterial strains. This invention aims to construct optimally formulated microbial communities with high colonization rates and high safety for the treatment of liver cirrhosis.

[0125] 2) Metagenome-Assembled Genomes (MAGs) refer to microbial genomes obtained by clustering and binning longer contigs assembled from raw metagenomic sequencing fragments. In this invention, the system prioritizes extracting high-quality, patient-specific MAGs that meet the requirements for integrity and contamination level as sequence input to preserve the strain variation characteristics in patients to the greatest extent possible.

[0126] 3) Metagenome-Wide Association Study (MWAS) refers to an existing method that uses high-throughput sequencing to obtain community abundance and statistical tests to identify species with differential abundance between healthy and disease groups. The background of this invention points out that this existing technology lacks causal inference capabilities and easily misidentifies strains that are merely "passengers" accompanying changes in the disease environment as "drivers," leading to an extremely high false positive rate in strain screening.

[0127] 4) Operational Taxonomic Units (OTUs) or Amplicon Sequence Variants (ASVs) refer to taxonomic unit markers in microbial community analysis, defined based on sequence similarity. This invention points out that most traditional deep learning models treat them as unordered feature inputs, ignoring the underlying evolutionary distances.

[0128] (iii) Semantic large model and artificial intelligence terminology.

[0129] 1) A Biological Large Language Model (Bio-LLM) is a large-scale natural language processing model fine-tuned using biomedical proprietary knowledge (such as unannotated genes and specific metabolic pathways). In this invention, it is used to perform functional "decoding" and semantic reasoning on candidate core bacterial communities, qualitatively excluding bacterial combinations with metabolic conflicts or potential risks.

[0130] 2) Retrieval-Augmented Generation (RAG) refers to a technique that combines external knowledge bases (such as the metabolic pathway library KEGG and the virulence factor library VFDB) to enhance the accuracy of large language models. In this invention, the system dynamically loads external knowledge bases for retrieval and matching when the model receives prompt words, forcing the model to generate answers based on objective facts, thereby overcoming the hallucination problem inherent in large language models.

[0131] 3) Chain of Thought (CoT) refers to the process of guiding a large language model to perform step-by-step reasoning according to preset logical steps. In this invention, this mechanism is used to guide Bio-LLM to simulate the cascade process of biochemical reactions, sequentially deducing substrate utilization, intermediate product transformation, and final product effects.

[0132] 4) Graph Neural Network (GNN) refers to a deep learning architecture used to process graph-structured data. As mentioned in the background of this invention, existing solutions use GNNs to treat microbial taxonomic units as nodes and learn node embeddings in the graph through message passing to infer interactions. However, due to the lack of phylogenetic constraints and functional semantic understanding, they face black-box defects in formulation design tasks.

[0133] IV) Terminology related to microbial metabolic network models and quantitative decision-making (solving for the optimal inoculation stoichiometry).

[0134] 1) Genome-scale metabolic network models (GEMs) refer to systems biology mathematical models based on microbial genome sequence reconstruction that contain gene-protein-response (GPR) logical relationships. In this invention, the system automatically reconstructs this model and rigorously verifies the qualitative conclusions of the large model through quantitative physical simulation.

[0135] 2) Community Flux Balance Analysis (cFBA) is an optimization method for calculating the synergistic metabolic flux of microbial communities in a virtual culture medium under specific boundary conditions and constraints. In this invention, the system uses maximizing the total community biomass as the joint objective function and iteratively solves the model using a mixed integer linear programming (MILP) algorithm.

[0136] 3) The Metabolic Support Index (MSI) is an evaluation index used to quantitatively calculate the flux of metabolites produced by donor strains consumed by recipient strains. This invention uses it as a core quantitative indicator for determining the "mutualistic effectiveness" of a bacterial community; a higher MSI value indicates a higher degree of symbiotic dependence between strains.

[0137] 4) Antibiotic resistance genes located on mobile genetic elements (MGE-ARGs) are antibiotic resistance genes located on genetic elements such as plasmids and transposons that are prone to horizontal transfer between different strains. This invention uses these as a "pre-emptive veto line" for static safety assessment; once detected, an immediate veto is applied.

[0138] 5) Colony-Forming Unit (CFU) is a unit used to measure the number of viable bacteria in a microecological drug or culture. The final output of this invention utilizes a limiting substrate supply-demand balance equation to calculate the optimal inoculation stoichiometry ratio (CFU Ratio) that maximizes the total community biomass, thereby providing precise digital production parameters for microecological drugs.

[0139] It should be noted that the target diseases, specific pathological targets, or pathological environmental constraints mentioned in this invention are not limited to a single specific disease, but rather refer to a variety of complex diseases closely related to the host's gut / microecological imbalance, such as, but not limited to, inflammatory bowel disease, metabolic syndrome, various liver metabolic diseases, and autoimmune diseases. This invention provides a universal, underlying framework for formula microbiota calculation and construction.

[0140] To more clearly and concretely illustrate the core technical details of this invention, such as the extraction of high-order ecological interaction rules, metabolic gap filling under pathological environmental constraints, and functional targeting verification based on semantic and physical dual modalities, the following specific embodiments will uniformly use liver cirrhosis, a typical disease of microecological imbalance, as a specific application scenario for illustrative purposes.

[0141] Those skilled in the art should understand that using liver cirrhosis as a starting point is merely for the purpose of fully demonstrating the closed loop of technical data flow in this invention, and does not constitute any physical limitation on the overall scope of protection of this invention. Based on the general computing architecture and multimodal fusion mechanism disclosed in this invention, those skilled in the art, when facing other target diseases, only need to make equivalent substitutions between the corresponding physiological environmental parameters (such as the inflow of specific amino acids, pH value, etc.) and the prior knowledge base, and can directly apply this method to the formulation and research of microbial communities for other microecological diseases without creative effort.

[0142] As the end stage of chronic liver disease, cirrhosis's pathophysiological mechanisms are not limited to fibrosis and regenerative nodule formation within the liver itself, but also profoundly involve systemic dysregulation of the gut-liver axis. With the widespread adoption of high-throughput sequencing technology and the deepening of multi-omics analysis, gut microbiota dysbiosis has been established as a key driver of cirrhosis progression. This dysbiosis manifests not only as changes in abundance at the taxonomic level, but also as functional remodeling of the microbial metabolic network, leading to endotoxemia, elevated blood ammonia levels, and abnormal bile acid metabolism, which in turn induces serious complications such as hepatic encephalopathy (HE) and spontaneous bacterial peritonitis (SBP).

[0143] Traditional gut microbiota interventions, such as fecal microbiota transplantation (FMT), have shown potential in reconstructing the gut microbiota, but they face barriers such as high donor heterogeneity, uncontrollable safety (e.g., passive transfer of drug-resistant bacteria or pathogenic viruses), and difficulties in standardized production. Therefore, rationally designed gut microbiota formulations (LBP) are becoming a core direction for next-generation gut microbiota therapy. However, screening for the optimal target microbiota formulation from thousands of gut symbiotic bacteria that can stably colonize, synergistically exert therapeutic effects, and possess ecological robustness presents a significant combinatorial optimization challenge.

[0144] Traditional trial-and-error methods or simple co-occurrence network analyses often fail to capture the ubiquitous high-order interactions and nonlinear dynamics in microbial communities, and are easily misled by spurious correlations. Furthermore, before constructing therapeutic microbiota formulations, it is essential to accurately decode the specific alterations in the gut microbiota of patients with cirrhosis. These alterations are not random chaos, but rather follow specific ecological succession patterns, forming highly reproducible cirrhotic microecological characteristics.

[0145] Therefore, to overcome this bottleneck, this invention proposes a method for constructing therapeutic microbial biosynthetic formulas (LBPs) for liver cirrhosis based on phylogenetic constraints and semantic-physical dual-modal fusion. In one optional implementation, a closed-loop system of "evolutionary mining—dual verification—quantitative design" is constructed. The Phy-iRF algorithm, incorporating phylogenetic tensor constraints, accurately extracts high-order ecological interaction frameworks from high-dimensional data. Furthermore, Bio-LLM (semantic reasoning) and genome-wide metabolic models (GEMs quantitative simulation) are innovatively combined to confirm the mechanisms of candidate microbiota. Based on this, a pre-implementation safety circuit breaker mechanism (MGE-ARGs / metabolic toxicity) and a limiting substrate supply-demand balance equation can be used to ultimately output a set of interpretable quantitative microecological drug formulations, thus obtaining the final target microbial biosynthetic formula.

[0146] Despite significant progress made in the combination of microbiome and artificial intelligence in the diagnosis and biomarker screening of cirrhosis, existing technologies still have the following core shortcomings in the rational design and construction of therapeutic microbial formulations for cirrhosis and its complications (such as hepatic encephalopathy and acute-on-chronic liver failure):

[0147] First, the widely used metagenomic association analysis (MWAS) lacks causal inference capabilities, leading to a high false positive rate in strain screening. Current mainstream techniques primarily rely on metagenomic association analysis to identify differentially abundant species between healthy controls and patients with cirrhosis. However, this statistically correlated approach has significant limitations, mainly manifested in:

[0148] 1) Inability to distinguish between causal and concomitant relationships (i.e., prone to spurious correlations): Many strains with significantly altered abundance in cirrhosis patients are merely "passengers" resulting from changes in the disease environment (such as increased intestinal pH and abnormal bile acid metabolism), rather than "drivers" of the disease. For example, the colonization of oral flora (such as Veillonella and Streptococcus) in the gut of cirrhotic patients is often due to impaired gastric acid barrier. Simply intervening with these strains with altered abundance (or supplementing them with probiotics) is often ineffective and may even worsen the condition.

[0149] 2) Ignoring ecological network effects: Traditional differential analysis often treats microorganisms as isolated variables, neglecting the complex cross-feeding and competitive relationships among them. In the gut of cirrhotic patients, the loss of butyric acid-producing bacteria (such as *Faecalibacterium prausnitzii*) is often due to the absence of upstream substrate providers (such as *Bifidobacterium*) or an increase in competitors. Current techniques struggle to capture these dependencies, leading to graft failure of selected single strains in vivo due to a lack of niche support.

[0150] On the other hand, co-occurrence network analysis used in existing technologies cannot resolve higher-order and nonlinear interactions. To infer interactions between microorganisms, existing technologies often employ co-occurrence network analysis based on Pearson or Spearman correlation coefficients. However, this method faces significant challenges in guiding the design of formulated microbial communities, specifically:

[0151] 1) Statistical artifacts in compositional data: Microbiome data is a typical example of compositional data. The constraint of a total abundance of 1 can lead to spurious negative correlations between species that are not originally related. Although algorithms such as SparCC attempt to correct this problem, they are still not robust enough in metagenomic data with extremely high sparsity.

[0152] 2) Inability to capture high-order nonlinear interactions: The function of microbial communities is often determined by the synergistic effects of three or more species (for example, species A can only function to detoxify ammonia if species B is present and species C is absent). Traditional pairwise network analysis cannot express such complex logical relationships (such as XOR relationships and threshold effects) through simple edge connections, resulting in the designed formulation microbial communities lacking necessary synergistic modules and exhibiting poor ecological stability.

[0153] Furthermore, existing deep learning models lack biological constraints and functional semantic understanding. In recent years, although graph neural networks (GNNs) and deep learning models have been applied to microbiome classification, they suffer from black-box and semantic gaps in recipe design tasks, primarily manifested in the following ways:

[0154] 1) Lack of Phylogenetic Constraints: Most deep learning models (such as multilayer perceptrons and standard CNNs) treat microbial taxa as unordered feature inputs, ignoring evolutionary distances between species. However, closely related species often share similar metabolic functions (phylogenetic signals). Ignoring this prior knowledge makes deep learning models prone to overfitting when faced with data noise and makes it difficult to use known species patterns to predict the behavior of unclassified species (microbial dark matter).

[0155] 2) Lack of functional semantics in genomes: Traditional machine learning methods only utilize species abundance values ​​without understanding the functional semantics behind their genome sequences (such as whether they encode specific bile salt hydrolases or urease genes). Although genome language models such as DNABERT have emerged recently, they are mostly used for functional prediction of single sequences and have not yet been deeply integrated with community-level ecological interaction models. Existing LBP designs often cannot directly predict the metabolic adaptability of strains under specific pathological environments (such as high ammonia, high inflammation) at the sequence level.

[0156] Finally, due to the high trial-and-error costs and lack of safety assessments in formula microbial community design, current formulation microbial community development mainly relies on an empirical "mix-cultivate-test" process, resulting in a huge experimental space (the combinatorial explosion problem). Currently, there is a lack of effective computational simulation methods to pre-assess the ecological stability and safety of strain combinations. Especially in the treatment of cirrhosis, if strains carrying virulence factors or drug resistance genes are mistakenly selected, or if the constructed microbial community produces excessive ammonia or endogenous ethanol in vivo, it will pose a fatal risk to patients. Existing technologies lack an end-to-end computational framework based on rational design to mitigate these risks.

[0157] In view of this, the present invention creatively proposes a method, apparatus, electronic device, and storage medium for constructing a formulation microbial community, aiming to solve, or at least partially solve, the four core technical challenges in the development of existing therapeutic probiotics for liver cirrhosis (LBP): spurious correlations in interaction discovery, black-box metabolic mechanisms, lagging safety evaluation, and empirical reliance on dosage form ratios. Specifically, it represents a technological breakthrough in at least one of the following aspects:

[0158] (1) The Phy-iRF algorithm, which introduces phylogenetic tensor constraints, achieves a leap from data fitting to pattern mining, solving the problem that traditional statistical correlation analysis cannot identify real ecological interactions. It innovatively constructs a phylogenetic distance tensor as a prior constraint and uses a normalized smoothing formula to correct the feature selection process of random forest. When dealing with high-dimensional sparse metagenomic data, it can effectively suppress sequencing noise and force the model to focus on "clan functional groups" that are evolutionarily conserved and functionally complementary. This makes the high-order interaction rules (such as the third-order co-occurrence patterns extracted by RIT) have extremely high biological robustness and significantly improves the ecological stability of the core framework.

[0159] (2) A two-dimensional verification funnel was constructed to transform probabilistic predictions into biological confirmation. This breaks through the limitations of a single AI model and establishes a dual verification mechanism of semantic reasoning Bio-LLM and physical simulation GEMs, solving the problems of lack of physical constraints and mechanistic interpretability in AI prediction models. On the one hand, the chain reasoning CoT of the large model is used to eliminate logical paradoxes; on the other hand, the whole genome metabolic network is reconstructed for the pathological environment of liver cirrhosis, and community flux balance analysis (cFBA) is performed. By calculating the metabolic support index (MSI), the mutualistic flux between strains is quantitatively confirmed at the material flow level for the first time in the "dry experiment" stage, which greatly improves the success rate of wet experiments.

[0160] (3) A pre-emptive tiered circuit breaker mechanism was established, achieving industrial-grade safety control. Safety evaluation was shifted from post-detection to pre-screening through computation, constructing a stringent "one-vote veto" system to address the problem of delayed and incomplete safety evaluation of probiotic drugs. At the genomic level, high-risk drug-resistant genes MGE-ARGs located on plasmids / transposons were specifically identified to avoid the risk of drug resistance spread; at the metabolite level, metabolic toxicity fluxes (such as H2S and ethanol) during the simulation process were dynamically monitored to avoid the potential for exacerbating liver damage. This mechanism ensures that the output formulation meets safety standards.

[0161] (4) Quantitative formulation design based on supply and demand balance equations has achieved a breakthrough from qualitative formulation to quantitative process. GEMs models are used to identify limiting substrates, and a supply and demand balance equation is established based on the law of mass action, solving the problem of lacking quantitative calculation basis for the inoculation ratio of the formulation microbial community. Through gradient ascent iterative solution, the optimal inoculation stoichiometry that maximizes the total biomass of the community is output, solving the resource competition problem of LBP drugs in industrial fermentation and in vivo colonization, and providing accurate digital parameters for subsequent process development.

[0162] The following will combine Figures 1 to 19 This invention provides a detailed description of the specific implementation methods, apparatus, electronic devices, and storage media for constructing formulated microbial communities.

[0163] The method provided in this invention can be performed by an electronic device, which can be a server, computing cluster, or local computing terminal with powerful data processing capabilities, and internally includes core hardware such as a processor and memory. For ease of description, in the following embodiments, the electronic device will be consistently described as the executing entity, or the executing entity will be omitted in the description.

[0164] Figure 1 This is a flowchart illustrating a method for constructing a formula microbial community provided by the present invention, as shown below. Figure 1 As shown, it mainly includes:

[0165] Step 11: Extract representative genomic sequences from multiple microbial taxa in the sample set to construct a phylogenetic tree.

[0166] In the context of microbial ecology research and formulation microbial community development, the sample set processed in this embodiment refers to a collection of biological sample data collected from target subjects, such as patient groups suffering from specific target diseases and healthy control groups, including metagenomic sequencing data of intestinal contents or mucosal tissue. The microbial taxonomic unit refers to a set of microbial communities with clear taxonomic definition, which can be specified down to phylum, class, order, family, genus, species, or even specific metagenomic assembly genomes (MAGs) or operational taxonomic units (OTUs). The representative genome sequence refers to a set of nucleic acid sequences that uniquely and accurately reflect the genetic essence of the microbial taxonomic unit.

[0167] In practice, the first step is to acquire the aforementioned sample set and extract representative genomic sequences corresponding to each microbial taxonomic unit. Subsequently, sequence alignment algorithms, such as multiple sequence alignment (MSA), can be used to identify homologous regions and variant sites among these representative genomic sequences. Based on the degree of sequence differences, phylogenetic analysis methods such as clustering or maximum likelihood methods are employed to construct a phylogenetic tree. The phylogenetic tree visually reconstructs the evolutionary trajectory and differentiation levels of each microbial taxonomic unit throughout the long course of nature in a tree-like format.

[0168] Step 12: Calculate the evolutionary distance between each pair of microbial taxa in the phylogenetic tree, and convert the evolutionary distance into a phylogenetic weight matrix. The phylogenetic weight matrix is ​​used to characterize the kinship similarity between the microbial taxa.

[0169] Traditional analyses often use species abundance tables directly, ignoring evolutionary distances between species. This embodiment aims to integrate microbial abundance information, clinical phenotype information, and biological evolutionary distance information to construct a feature input with evolutionary coordinates.

[0170] Specifically, after obtaining the phylogenetic tree, this embodiment considers that a simple tree-like topology is difficult to directly use as numerical input for subsequent machine learning models, and therefore requires quantization and dimensionality reduction. The evolutionary distance between any two microbial taxa (i.e., leaf nodes on the tree) refers to the total path length or number of nodes required to connect any two microbial taxa (i.e., leaf nodes on the tree) in the topology of the phylogenetic tree, which intuitively reflects the degree of divergence between the two in their evolutionary history.

[0171] After calculating the evolutionary distances between all pairs of microbial taxa using graph theory algorithms or tree traversal algorithms, the evolutionary distances can be converted into phylogenetic weight matrices using a preset mathematical mapping function, such as, but not limited to, the inverse distance method or normalized kernel function.

[0172] The values ​​of the elements in the phylogenetic weight matrix can be directly used to characterize the phylogenetic similarity between the microbial taxa. That is, the shorter the evolutionary distance between two taxa, the higher their corresponding weight values ​​in the phylogenetic weight matrix, which means that their phylogenetic similarity is stronger and they are more likely to have homologous substitution in ecological functions.

[0173] Step 13: Obtain the original microbial abundance matrix of the sample set, perform a central logarithmic ratio transformation on the original microbial abundance matrix to obtain the feature matrix, and use the feature matrix as the feature dimension input to construct an iterative random forest model.

[0174] In addition to evolutionary information, this embodiment also processes ecological abundance information simultaneously. The original microbial abundance matrix refers to the absolute counts or relative proportions of each microbial taxa in each sample, obtained directly from sequencing data. Since high-throughput sequencing data is inherently limited by the fixed depth of the sequencer, it is compositional data with a "sum of a fixed value (e.g., 100%)" constraint; direct use of this data could lead to statistically spurious correlations.

[0175] To overcome this data closure effect, this embodiment performs a Centered Log-Ratio Transformation (CLR) on the original microbial abundance matrix. Through logarithmic scaling and centering, the bias in sample sequencing depth is effectively eliminated, projecting the data from the simplex space to the real number space to obtain the feature matrix. Subsequently, this feature matrix can be used as the feature dimension input to construct an iterative random forest model. The iterative random forest model is a nonlinear machine learning architecture based on the ensemble of multiple decision trees. It can update the probability of feature selection multiple times during its training cycle, thereby deeply exploring the interaction effects between variables.

[0176] Step 14: During the feature selection iteration of the iterative random forest model, the original feature importance of each microbial taxonomic unit is corrected to a smoothed feature importance using the phylogenetic weight matrix, so as to update the feature sampling weights of the iterative random forest model for the microbial taxonomic unit in the next iteration based on the smoothed feature importance.

[0177] In conventional random forest training, iterative random forest models evaluate feature contributions solely based on purely mathematical information (such as Gini impurity or information gain), a concept known as original feature importance. However, micro-ecosystems contain a large number of "functionally redundant strains" due to evolutionary homology. Traditional purely mathematical evaluation is prone to incorrectly identifying strains with homologous functions as collinear redundant features and removing them.

[0178] To address this issue, this embodiment creatively introduces phylogenetic constraints during the feature selection iteration process of the iterative random forest model.

[0179] Specifically, this embodiment utilizes a phylogenetic weight matrix to correct the original feature importance of each microbial taxonomic unit to a smoothed feature importance. This is equivalent to allowing each microbial taxonomic unit to absorb or share the importance scores of other related taxonomic units with high phylogenetic similarity when assessing its own importance. Through this evolutionary-based smoothing correction, the feature sampling weights for each microbial taxonomic unit in the next iteration of the iterative random forest model can be updated based on the smoothed feature importance. This ensures that in the subsequent tree construction process, microbial taxonomic units that have group synergistic value in evolutionary branches but have weak individual abundance can obtain a higher sampling probability and thus be fully preserved by the iterative random forest model.

[0180] Step 15: When the iterative random forest model reaches the preset convergence condition, extract high-order ecological interaction rules based on the decision path and output candidate core bacterial communities.

[0181] This embodiment continuously executes the aforementioned iterative training with phylogenetic smoothing. When the iterative random forest model reaches a preset convergence condition, such as when the variation in the feature sampling weight distribution is below a minimum threshold or when the maximum preset number of iterations is reached, it indicates that the iterative random forest model has fully learned the complex steady-state characteristics within the micro-ecological environment. At this point, its internal decision tree structure is analyzed. The decision path refers to the sequence of variable discriminations traversed from the root node to each leaf node in the multiple decision trees within the iterative random forest model.

[0182] This embodiment extracts higher-order ecological interaction rules based on the decision paths. Since multiple microbial taxonomic unit nodes that cooperate or restrict each other inevitably appear on a single decision path, these frequently co-selected node combinations represent complex cooperative patterns that go beyond simple pairwise associations, i.e., higher-order ecological interaction rules.

[0183] Furthermore, this embodiment will perform a combined evaluation based on these extracted high-order co-occurrence relationships to output candidate core bacterial communities. These candidate core bacterial communities represent a basic set of species that initially possess extremely strong ecological synergistic stability, forming the data foundation for subsequent formulation optimization.

[0184] Step 16: Determine the target formulation microbial community based on the candidate core microbial community.

[0185] The target formulation microbial community refers to the preferred set of microorganisms that can ultimately be directly used to intervene in the host pathological environment and meet the needs of industrial preparation or clinical treatment.

[0186] In this embodiment, after obtaining a highly robust candidate core microbial community, downstream verification and screening logic will be further carried out. For example, by introducing external biological function tests, safety screening, or virtual simulation analysis and other general optimization methods, strains that do not have specific metabolic intervention capabilities or contain risk genes can be eliminated. Finally, the target formulation microbial community is screened from the candidate core microbial community, thereby completing the entire closed loop of digital formulation construction.

[0187] The method for constructing microbial communities provided in this invention effectively addresses the core pain points of existing technologies, such as the tendency for microbial feature screening to overfit and the lack of biological significance due to reliance solely on mathematical statistics. Technically, this embodiment performs a central logarithmic ratio transformation on the original abundance matrix, completely eliminating false-positive interference from the closure effect of component data. More importantly, this embodiment breaks with the conventional understanding of treating microorganisms as absolutely independent variables in traditional machine learning, creatively abstracting evolutionary distance related to biological phylogenetic evolution into a spatial topological penalty matrix, and smoothly correcting feature importance during the dynamic training of iterative random forests. This mechanism cleverly utilizes evolutionary homology information to combat the extreme sparsity of high-dimensional sequencing data, ensuring that the extracted high-order ecological interaction rules and candidate core microbial communities are not only mathematically robust but also possess strong interpretability and real ecological synergy support at the biological evolutionary scale.

[0188] To further elaborate on the underlying logic and data tracing process of obtaining the original microbial abundance matrix of the sample set using the above embodiments, and performing a central logarithmic ratio transformation on the original microbial abundance matrix to obtain the feature matrix, the following will combine... Figure 2 and Figure 3The pre-construction multimodal data architecture of the present invention is described in detail.

[0189] Figure 2 This is an implementation logic diagram of a method for constructing a formula microbial community provided by the present invention. Figure 3 This is a schematic diagram of a technical route for multimodal data construction and phylogenetic tensor generation provided by the present invention, such as... Figure 2 and Figure 3 As shown, in actual research scenarios targeting complex gut microbiota dysbiosis diseases such as cirrhosis, this embodiment can first obtain raw metagenomic sequencing data from clinical subjects (including disease groups and healthy control groups).

[0190] In practice, the raw reads are first rigorously filtered using quality control tools (such as KneadData and Trimmomatic workflows) to remove low-quality and adapter sequences, resulting in clean reads. Considering the inevitable presence of host DNA in clinical samples, sequence alignment algorithms (such as Bowtie2) can be used to align the clean reads to a human reference genome (such as the hg38 or CHM13 database) to precisely remove host contamination sequences.

[0191] Subsequently, high-precision species sequence annotation tools (such as MetaPhlAn 4) can be used to extract representative genome sequences of each microbial taxonomic unit, and the relative abundance or absolute read count of each microbial taxonomic unit in each sample can be calculated, thereby constructing the original microbial abundance matrix. X raw .

[0192] Due to the physical upper limit of sequencing depth in next-generation sequencing (NGS) technology, the extracted original microbial abundance matrix... X raw Essentially subject to the Constant Sum Constraint, it is a typical example of compositional data. If directly applied... X raw In machine learning modeling, an increase in the abundance of one dominant microorganism inevitably leads to a spurious decrease in the relative abundance of other microorganisms, resulting in significant statistical false positive associations. To completely break this closure effect, this embodiment introduces a central logarithmic ratio transformation process.

[0193] Figure 4 This is a schematic diagram of a process for performing a central logarithmic ratio transformation on the original microbial abundance matrix provided by the present invention, as shown below. Figure 4 As shown, it can specifically include:

[0194] Step 41: Replace the zero-value features in the original microbial abundance matrix with preset non-zero values.

[0195] In highly sparse microbial metagenomic sequencing data, the original microbial abundance matrix X raw The original abundance matrix often contains a large number of zero values. These zero values ​​may originate from the true absolute absence of the microorganism in the host environment (structural zero), or they may be due to the microorganism's low abundance, which failed to be captured by the current sequencing depth (sampling zero). Since subsequent logarithmic transformations cannot operate on mathematical zeros, this embodiment performs pseudo-counting smoothing on the original microbial abundance matrix.

[0196] In practical implementation, the zero-value features in the original microbial abundance matrix can be replaced with preset non-zero values ​​in the following two ways:

[0197] The first approach is to directly use a preset, extremely small non-zero value, such as an empirical constant. This is then uniformly added to all elements of the original microbial abundance matrix.

[0198] The second approach is to use the Bayesian Multiplicative Replacement algorithm. While keeping the total relative abundance of the samples unchanged, the zero values ​​are replaced with tiny non-zero values ​​by inferring from prior probability. This avoids the collapse of mathematical operations and preserves the true physical structure of the sparse distribution of species in the original matrix to the greatest extent.

[0199] Step 42: For the original microbial abundance matrix after replacement processing, determine the geometric mean parameter of the abundance vector of each sample in the sample set.

[0200] After zero-value smoothing, the abundance distribution characteristics of each sample in the original microbial abundance matrix are extracted. Let... For the first i The abundance vector of each sample, M This represents the total number of microbial taxonomic units (i.e., the total number of feature dimensions). For the first i In the nth sample k The abundance value of each microbial taxonomic unit is calculated by multiplying the products of all feature components in this abundance vector. M The geometric mean parameter of the abundance vector of each sample in the sample set is determined by the power root.

[0201] Specifically, the electronic device calculates the geometric mean parameter strictly according to the following formula. :

[0202] .

[0203] In the Aitchison geometric space, this geometric mean parameter represents the absolute reference centroid of the micro-ecological community within the sample.

[0204] Step 43: Obtain the relative ratio of each abundance value in the abundance vector of each sample to the corresponding geometric mean parameter.

[0205] Optionally, after establishing the baseline centroid of the sample, this embodiment will assign the specific abundance value of each microbial taxonomic unit in the abundance vector. Divide by the geometric mean parameter corresponding to the sample respectively The process involves obtaining the relative ratio of each abundance value in the abundance vector of each sample to the corresponding geometric mean parameter. This division operation essentially uses the overall internal ecological capacity of the sample itself as an anchor to measure whether each strain is in a state of relative enrichment or relative decay.

[0206] Step 44: Perform logarithmic scaling on all the relative ratios to obtain the feature matrix.

[0207] Specifically, the logarithmic scaling can be achieved by taking the natural logarithm ln, i.e., the formula for performing logarithmic scaling on all the relative ratios is:

[0208] ;

[0209] in, Abundance value Perform CLR transformation.

[0210] Through the rigorous scaling and logarithmic transformation described above, this embodiment successfully eliminates combinatorial bias, projecting the nonlinear component data confined to the simplex space onto the unconstrained Euclidean real space, ultimately obtaining the characteristic matrix (i.e., Figure 3 The microbial relative abundance matrix obtained after eliminating combinatorial bias and projecting to Euclidean space It follows the assumption of a multivariate normal distribution, which is suitable for subsequent random forest modeling.

[0211] This invention, through the rigorous introduction of zero-value Bayesian substitution and a formula-based establishment of a central logarithmic ratio (CLR) transformation feature chain for the baseline centroid, fundamentally breaks the independence assumptions and constant sum constraints of high-throughput metagenomic sequencing data at the underlying mathematical mechanism level. In practical microbial ecosystem AI modeling, directly using untransformed abundance data easily leads the model to learn pseudo-competitive relationships caused by sequencing depth limitations. This invention uses the geometric mean parameter calculated at the sample level as a dynamic benchmark for logarithmic space projection, restoring the zero-sum game of relative abundance to Euclidean geometric features that reflect real microbial ecosystem fluctuations. This design fundamentally eliminates false-positive noise caused by sequencer system errors, giving the subsequent input feature matrix reasonable covariance properties. This not only greatly improves the stability of machine learning models when extracting high-order ecological rules but also endows the finally selected candidate core bacterial communities with true biological fidelity.

[0212] In order to give these features evolutionary dimensions vitality in subsequent machine learning, as an optional embodiment, this method provides an implementation of extracting the phylogenetic relationships between microbial taxonomic units through phylogenetic trees.

[0213] Figure 5 This is a schematic diagram illustrating the process of determining a phylogenetic weight matrix provided by the present invention. Figure 6 This is a schematic diagram of a process for determining phylogenetic distance provided by the present invention, combined with... Figure 5 and Figure 6 As shown, the calculation of the evolutionary distance between each pair of microbial taxa in the phylogenetic tree, and the conversion of the evolutionary distance into a phylogenetic weight matrix, mainly includes, but is not limited to, the following steps:

[0214] Step 51: Determine the nearest common ancestor node between each pair of microbial taxonomic units in the phylogenetic tree.

[0215] Step 52: Determine the kinship distance between each microbial taxonomic unit based on the first path span between each microbial taxonomic unit and the root node of the phylogenetic tree and the second path span between the nearest common ancestor node and the root node.

[0216] Step 53: Use a preset exponential decay kernel function to perform similarity mapping on the clan distance to obtain the phylogenetic weight matrix that characterizes the strength of kinship.

[0217] In practice, representative genomic sequences of all detected microbial taxa are first extracted, and a maximum likelihood phylogenetic tree is constructed using a phylogenetic tree construction process. In this phylogenetic tree, leaf nodes represent specific microbial taxa. To measure the evolutionary distance between any two microbial taxa, the electronic device traverses the topology of the phylogenetic tree to determine the Most Recent Common Ancestor (MRCA) node between each pair of taxa. The MRCA node represents the historical intersection point where the two taxa last shared a common starting point in their evolutionary history.

[0218] After locating the nearest common ancestor node, the phylistic distance, including branch length, is calculated. Let L(n, root) represent the path span from node n in the phylogenetic tree to the root node of the phylogenetic tree. The first path span from the first microbial taxonomic unit u to the root node is calculated. And the first path span from the second microbial taxonomic unit v to the root node. Simultaneously, calculate the second path span from the nearest common ancestor node MRCA(u,v) of these two microbial taxa to the root node. .

[0219] Subsequently, the kinship distance between each microbial taxonomic unit can be determined based on the first path span from each microbial taxonomic unit to the root node of the phylogenetic tree and the second path span from the nearest common ancestor node to the root node. The clan distance can be calculated using the following formula:

[0220] .

[0221] Using the above formula, this embodiment calculates the kinship distance between all pairs of microbial taxonomic units and finally outputs a symmetric matrix containing the kinship distances of all the microbial taxonomic units.

[0222] Since clan distance is a measure of dissimilarity, in order to convert it into weights that can be used for subsequent machine learning, this embodiment will convert the obtained clan distance into similarity weights.

[0223] In practical implementation, a preset exponential decay kernel function can be used to perform similarity mapping on the clan distance, and the conversion formula is as follows:

[0224] ;

[0225] in, To adjust the hyperparameter of the evolutionary distance on the weight constraint strength (e.g., taking a value of 0.5), when the clan distance... The smaller the value, the higher the weight after mapping. The closer the similarity is to 1, the closer it is to 0; conversely, the closer it is to 0. After the above similarity mapping, the phylogenetic weight matrix representing the strength of kinship is finally obtained, and this phylogenetic weight matrix is ​​stored as prior knowledge.

[0226] Traditional machine learning typically treats microbial taxa as unordered, independent inputs, ignoring the evolutionary distances between them, making the model highly susceptible to overfitting to noise. This invention introduces quantitative indicators from evolutionary biology as feature constraints. By analyzing the phylogenetic tree and accurately calculating kinship distances, and then using an exponentially decaying kernel function for nonlinear transformation, the phylogenetic tree is successfully converted into a mathematical distance tensor. This approach forcibly preserves similar metabolic functional signals between microbial taxa at the underlying data structure level, enabling the model to possess strong noise resistance and prioritize the capture of evolutionarily conserved functional groups when facing high-dimensional sparse data.

[0227] As a specific embodiment, the following is combined with Figure 6 The above-mentioned calculation process of clan distance is illustrated by example.

[0228] exist Figure 6 The phylogenetic tree shown includes taxonomic units A, B, C, D, E, and F. Taking the calculation of the kinship distance between taxonomic units A and D as an example, we first backtrack upwards in the phylogenetic tree to determine the internal node where taxonomic units A and D intersect. Figure 6 The node marked as MRCA(A,D) is taken as the nearest common ancestor node.

[0229] Subsequently, the first path span L(A,ROOT) from microbial taxonomic unit A to the top ROOT node of the phylogenetic tree and the first path span L(D,ROOT) from microbial taxonomic unit D to the ROOT node are calculated; at the same time, the second path span L(MRCA(A,D),ROOT) from the nearest common ancestor node MRCA(A,D) to the ROOT node is calculated.

[0230] Finally, the electronic device adds the first path span L(A,ROOT) to the first path span L(D,ROOT) and subtracts twice the second path span L(MRCA(A,D),ROOT) to accurately calculate the kinship distance between microbial taxonomic unit A and microbial taxonomic unit D, thereby providing an accurate data basis for the subsequent exponential decay kernel function mapping.

[0231] Figure 7 This is a schematic diagram of a process for correcting the importance of original features of microbial taxonomic units, as provided by the present invention. Figure 7 As shown, during the feature selection iteration process of the iterative random forest model, the original feature importance of each microbial taxonomic unit is corrected to a smoothed feature importance using the phylogenetic weight matrix, specifically including but not limited to:

[0232] Step 71: For any target microbial taxonomic unit among all the said microbial taxonomic units, based on the phylogenetic similarity represented in the phylogenetic weight matrix, the original feature importance of all other microbial taxonomic units among all the said microbial taxonomic units is weighted and fused to determine the neighbor feature reference value of the target microbial taxonomic unit.

[0233] In traditional random forest models, feature sampling typically assumes that features are independent of each other, thus ignoring the evolutionary relationships between different microbial taxa. In this step, during the t-th iteration of the iterative random forest training, the electronic device first calculates the Mean DecreaseGini for each microbial taxa and denotes it as the importance of the original feature. ,in ，i This indicates the target microbial taxonomic unit.

[0234] Specifically, let j For other microbial taxonomic units, The phylogenetic weight matrix represents the target microbial taxa. i Other microbial taxonomic units j The weighted elements of the kinship similarity between neighbors can be obtained by the following formula to obtain the neighbor feature reference value. .

[0235] The aforementioned neighbor feature reference values ​​essentially reflect the overall importance of neighboring microbial taxa that are similar to the target microbial taxa on the phylogenetic tree.

[0236] Step 72: Assign a first weight to the original feature importance of the target microbial taxonomic unit using a preset smoothing parameter.

[0237] The preset smoothing parameter is denoted as (For example, a value of 0.7) is used to balance the performance of the target microbial taxonomic unit itself with that of its evolutionary neighbors. In this case, the original feature importance configured with the first weight is expressed as: .

[0238] Step 73: Configure a second weight for the neighbor feature reference value using the complementary parameter of the smoothing parameter.

[0239] Complementary parameters can be expressed as Then, the neighbor feature reference value configured with the second weight is expressed as:

[0240] .

[0241] Step 74: The original feature importance configured with the first weight is weighted and fused with the neighbor feature reference value configured with the second weight to obtain the modified feature importance of the target microbial taxonomic unit.

[0242] Specifically, the importance of modified features of target microbial taxa. The smoothing correction can be calculated using the following formula:

[0243] .

[0244] The process of smoothing the importance of original features provided in the above embodiments means that if closely related microbial taxa that are very close in evolutionary distance to the target microbial taxa have similarity weights, then the weight of the similarity is adjusted accordingly. If other larger microbial taxa also have high original feature importance, the overall weight of the target microbial taxa will be significantly improved through smoothing. In this embodiment, the corrected feature importance is normalized and then used as the feature sampling weight for the next iteration. Typically, after multiple iterations (e.g., 5 to 10 iterations), the feature sampling weight distribution will stably converge to a set of key microbial taxa with phylogenetic aggregation characteristics.

[0245] This embodiment creatively introduces an iterative weighted mechanism for phylogenetic smoothing, addressing the core deficiency of traditional machine learning in treating microbial taxonomic units as isolated variables when processing microbial ecological data. By setting smoothing and complementarity parameters for convolutional smoothing of Gini importance, the machine learning model is forced to focus on evolutionarily conserved and functionally complementary kinship functional groups when selecting features. When processing high-dimensional and sparse metagenomic sequencing data, information from closely related microbial taxonomic units can be used to enhance the true ecological signal, effectively suppressing false positive selection caused by sequencing noise, thereby significantly improving the biological robustness and statistical reliability of the model when extracting combinations of core microbial taxonomic units.

[0246] Figure 8 This is an implementation logic diagram of a high-order ecological interaction mining technology based on the Phy-iRF algorithm provided by the present invention, such as... Figure 8 As shown, in the implementation scenario of constructing the formula microbial community, this embodiment first obtains a multimodal structured data object. The multimodal structured data object uses the microbial relative abundance matrix after central log ratio transformation as the numerical feature, and uses the phylogenetic weight matrix as the prior tensor constraint reflecting the evolutionary distance.

[0247] First, initializing the feature sampling weights involves setting an initial feature sampling weight vector. Before any evolutionary feedback is obtained, all detected microbial taxa are equally probable in round 0. Subsequently, the electronic device initiates an iterative weighted loop based on phylogenetic smoothing. t =1... K ).

[0248] In the t In each iteration, the feature space is sampled with replacement based on the current feature sampling weights, and a weighted random forest containing multiple decision trees is trained. After training, the mean decrease in Gini impurity for each microbial taxonomic unit is calculated and used as the original feature importance for the current iteration.

[0249] At this point, the smoothing correction step provided in the above embodiment can be used in the iterative weighted loop. That is, for any target microbial taxonomic unit, its original feature importance is no longer considered in isolation, but the phylogenetic similarity between the target microbial taxonomic unit and all other microbial taxonomic units in the phylogenetic weight matrix is ​​extracted simultaneously. Then, a preset smoothing parameter is used. Assign a first weight to the original feature importance of the target microbial taxonomic unit itself, calculate the weighted mean importance of all its evolutionary neighbors, and utilize complementary parameters (1- A second weight is assigned to the neighbor feature reference value, and the isolated original feature importance is transformed into a smoothed feature importance that incorporates the clan background by performing a smoothing correction formula.

[0250] After obtaining the importance of the smoothed features for all microbial taxa, they can be further normalized, and the random forest model can be updated accordingly in the next round (the ). t The feature sampling weights in the +1 round of iterations. Through continuous feedback and weight updates, the iterative random forest model will be forced to prioritize sampling those kinship functional groups that are highly conserved in evolutionary lineage and have complementary functions when building the phylogenetic tree in subsequent iterations.

[0251] Furthermore, the distribution of feature sampling weights can be continuously monitored to determine whether the preset convergence condition for feature set stability has been met. When it is determined that the model weight update magnitude is extremely small and convergence has been achieved, the iteration loop can be exited, and the decision path parsing and random cross-tree (RIT) analysis stage can be entered.

[0252] As an alternative implementation, all decision tree paths in the converged forest can be extracted, and the RIT algorithm can be used to mine higher-order interaction patterns, thereby explicitly transforming the nonlinear classification rules hidden within the smoothed and corrected random forest into a specific set of candidate higher-order relations.

[0253] To further verify the reliability of the discovered high-order ecological interaction patterns in real-world environments, an optional implementation can introduce an interaction stability screening mechanism. This mechanism uses a Bootstrap resampling loop to calculate the stability score of each candidate high-order relationship set under different data subsets. After eliminating unstable combinations, a final list of candidate microbial functional groups that meet the set stability threshold is output, thus identifying them as candidate core microbial communities.

[0254] Figure 9 This is a schematic diagram of the process for determining candidate core bacterial communities provided by the present invention, as shown below. Figure 9 As shown, when the iterative random forest model reaches the preset convergence condition, in order to explicitly transform the implicit nonlinear classification rules within the model into specific microbial combinations, the following specific operation steps can be performed:

[0255] Step 91: Extract the decision paths of each decision tree in the iterative random forest model to obtain a set of decision paths.

[0256] Step 92: Identify the microbial taxonomic units commonly included in multiple decision paths in the decision path set, and obtain at least one candidate taxonomic unit combination.

[0257] Step 93: Based on the data resampling mechanism, determine the stability score of each candidate classification unit combination.

[0258] Step 94: Determine the candidate core microbial community based on all combinations of candidate taxonomic units whose stability scores meet the set threshold.

[0259] After the iterative random forest model converges, each decision tree within it contains complex micro-ecological classification rules. In this embodiment, for each decision tree in the iterative random forest model, all decision paths extending from the root node to the leaf node are extracted. Each decision path represents a specific combination of ecological conditions.

[0260] For example, a decision path can be represented as a specific sequence of conditions: {abundance of microbial taxonomic unit A > low threshold, abundance of microbial taxonomic unit B < high threshold, abundance of microbial taxonomic unit C > medium threshold}. This means that when these three ecological conditions are simultaneously met, the corresponding sample is determined to be in a specific health or disease state. This embodiment extracts all these decision paths and aggregates them to obtain the decision path set, which will serve as a potential high-order interaction candidate pool.

[0261] Because individual decision trees have a high degree of randomness, in order to find synergistic effects with universal biological significance, this embodiment will perform high-frequency pattern mining in the massive path data.

[0262] Specifically, a specific search algorithm can be used to find feature subsets that co-occur repeatedly in multiple decision trees and multiple decision paths, thereby identifying the microbial taxonomic units commonly contained in the set of decision paths.

[0263] Through this identification process, the electronic device successfully transforms the black-box model of the random forest into an interpretable set of rules, identifies high-order interaction patterns such as "microbial taxonomic unit A and microbial taxonomic unit B must coexist", and thus obtains at least one candidate taxonomic unit combination.

[0264] To eliminate false positive combinations caused by random noise, this embodiment further introduces a stability selection mechanism. Specifically, the original dataset can be perturbed to generate multiple resampled data subsets, and the aforementioned feature extraction and path parsing process can be repeatedly executed on these data subsets.

[0265] Subsequently, based on the data resampling mechanism, the stability score of each candidate classification unit combination is determined. The stability score essentially reflects the probability that a particular candidate classification unit combination can still be robustly identified by the model when faced with small perturbations in the data samples.

[0266] Finally, the candidate core microbial community obtained in this embodiment includes a list of one or more candidate ecological functional groups. Its data structure records in detail the combined members (such as specific gut symbiotic bacteria), stability scores (such as 0.85), interaction order (such as 3rd order interaction) and associated phenotypes, thus providing an extremely reliable high-order interaction framework for subsequent biological function verification.

[0267] This invention addresses the core challenge of traditional microbial co-occurrence network analysis in capturing high-order nonlinear interactions. It innovatively makes the black-box algorithm transparent by analyzing decision tree paths. By extracting decision paths and identifying co-occurring microbial taxa across the tree, it can accurately capture complex ecological logic (such as XOR relationships and threshold effects) determined collaboratively by three or more species. Simultaneously, it introduces a stability score evaluation based on a data resampling mechanism, effectively eliminating spurious correlations caused by sequencing sparsity, ensuring that the final output candidate core microbiota possesses strong ecological robustness and co-colonization potential in the real complex gut microbiota environment.

[0268] Figure 10 This is a flowchart illustrating an iterative random forest algorithm provided by the present invention. This embodiment can be combined with... Figure 10 The architecture shown performs precise extraction from high-dimensional abundance features to a high-order ecological interaction framework, which will be explained in detail below.

[0269] In the characteristic screening stage of formula microbial community construction, the following steps are first performed: Figure 10 The upper part shows the iterative reweighted RFs process. From iteration 1 to iteration K-1, the feature sampling weights are continuously corrected using the phylogenetic weight matrix provided in the previous embodiment. For example, in iteration 1, the input feature set... X and initial equal sampling weights Output after training Importance score; in iteration 2, the weights updated in the previous round are used as input. X , Train again to output new results. Importance scores are assigned, and so on, iteratively updated until convergence. .

[0270] When the weight distribution converges to a set of microbial taxa with phylogenetic aggregation characteristics, the electronic device stops iterating and enters a new state. Figure 10 The lower half shows the generalized RIT algorithm and the stable relation aggregation module with sampling with replacement.

[0271] First, this embodiment will start from the converged iterative random forest model (whose input is at this point) X , This process extracts all decision tree paths. For each decision tree in the iterative random forest model, all decision paths from the root node to the leaf node are extracted, resulting in a set of decision paths. In the context of microecology, each decision path represents a specific combination of ecological conditions, such as decision paths... This means that when the abundance conditions of these three microorganisms in the host gut are simultaneously met, the microecological environment will exhibit a specific health or disease phenotype. These massive decision-making paths constitute a potential pool of high-order interaction candidates.

[0272] Subsequently, to address the issue that standard random forests struggle to intuitively represent the nonlinear combined effect of "microorganism A + microorganism B - microorganism C", this embodiment will input the decision path set into... Figure 10 The RandomIntersection Trees (RIT) module is shown. Due to the extremely large number of decision paths, direct exhaustive search would be computationally intensive. This embodiment utilizes the RIT algorithm, constructing a deep binary tree structure and performing bitwise AND operations on the paths. Through this efficient search mechanism, it identifies subsets of microbial features that repeatedly appear across multiple decision trees and paths, thereby identifying microbial taxonomic units commonly contained across multiple decision paths, obtaining at least one candidate taxonomic unit combination, and outputting its interactions. This process successfully transforms the black-box model into an interpretable rule set, accurately capturing real cross-feeding relationships such as "butyric acid-producing bacteria must coexist with specific acetic acid-producing bacteria to survive stably."

[0273] Furthermore, in order to eliminate false positive combinations caused by data sparsity or random noise, this embodiment can further initiate the operation of, for example... Figure 10 The stability shown involves a mechanism for sampling B with replacement. Multiple subsets (e.g., ...) are generated by repeatedly resampling the original dataset with replacement. Figure 10 In to By combining the final weights on each subset of the dataset. The above-described random cross-tree path cross-process is run independently, and the interaction set of each group is output separately. For any identified candidate classifier combination, the frequency of its successful reproduction in all resampling results is counted, thereby determining the stability score of each candidate classifier combination based on the data resampling mechanism.

[0274] Finally, a strict stability threshold can be set. ( The candidate core microbial community is determined by combining all candidate taxa that meet the set stability threshold based on their stability scores.

[0275] The output candidate core microbial community not only includes a list of specific strain members, but also their interaction order (e.g., 3rd order synergy) and stability score. This refined set of microorganisms (i.e. Figure 10The final output, "aggregating all interactions based on stability scores," means that it can be robustly identified under more than half of the data perturbations, forming the optimal core framework of the microbial community for subsequent treatment of specific diseases (such as liver cirrhosis and other microecological disorders).

[0276] Based on the above embodiments, as an optional embodiment, the step of identifying microbial taxonomic units commonly included in multiple decision paths within the decision path set to obtain at least one candidate taxonomic unit combination may specifically include, but is not limited to, the following implementation steps:

[0277] Step 1: Extract candidate microbial taxonomic units that serve as splitting nodes in each decision path;

[0278] Step 2: Select at least one set of decision paths from the set of decision paths, and perform an intersection operation on the candidate microbial taxonomic units contained in each set of decision paths.

[0279] Step 3: The set containing multiple candidate microbial taxonomic units obtained after the intersection operation of each group of decision paths is determined as a candidate taxonomic unit combination.

[0280] In the set of decision paths obtained by the method provided in any of the above embodiments, each decision path is actually composed of a series of splitting conditions from the root node to the leaf node, which represent a specific combination of ecological conditions. This embodiment extracts candidate microbial taxa units as splitting nodes in each decision path.

[0281] For example, a decision path can be represented as a specific sequence of conditions: {microbial taxonomic unit A > low threshold, microbial taxonomic unit B < high threshold, microbial taxonomic unit C > medium threshold}. This means that when all three conditions are met simultaneously, the sample is classified as a specific state. At this point, microbial taxonomic units A, B, and C can be extracted as candidate microbial taxonomic units under this decision path. All these decision paths will form what is considered a potential pool of higher-order interaction candidates.

[0282] Considering that standard random forests cannot intuitively represent complex combination effects such as "microbial taxonomic unit A + microbial taxonomic unit B - microbial taxonomic unit C", and given the huge number of decision paths, directly exhaustively searching would be too computationally intensive, this embodiment creatively introduces the Random Intersection Trees (RIT) algorithm. By constructing a deep binary tree, it efficiently searches for frequently co-occurring feature subsets in the path set.

[0283] Specifically, at least one set of decision paths is selected from the set of decision paths, and the intersection operation is performed on the candidate microbial taxonomic units contained in each set of decision paths. In the actual calculation execution, a bitwise AND operation can be performed on the selected decision paths to quickly find the set of microbial taxonomic units that repeatedly appear in multiple trees and multiple decision paths.

[0284] Through the efficient intersection search and bitwise AND operation described above, the implicit classification rules hidden deep within the iterative random forest model are explicitly expressed as specific species combinations. The set containing multiple candidate microbial taxa units obtained after intersecting each set of decision paths is determined as a candidate taxa unit combination. For example, identifying a specific set of higher-order interactions. .

[0285] Traditional pairwise network analysis cannot express complex logical relationships such as XOR relationships and threshold effects through simple edge connections. This invention creatively utilizes a random cross-tree algorithm to perform intersection operations, perfectly solving the technical bottlenecks of combinatorial explosion and excessive computational load in complex microbial networks. By extracting split nodes in the decision path and performing positional and layer-wise intersection operations, nonlinear collaborative functional modules containing multiple species can be directly extracted. This not only greatly improves the search efficiency of high-order nonlinear interactions but also completely breaks the functional mechanism black box of artificial intelligence models in microbial drug design, providing a highly biologically interpretable core combinatorial framework for the rational design of subsequent formulation microbial communities.

[0286] Figure 11 This is a flowchart illustrating a process for determining the stability score of candidate classification unit combinations provided by the present invention, as shown below. Figure 11 As shown, in order to eliminate false positive combinations caused by random noise, this embodiment introduces a stability selection and bagging mechanism. The stability score of each candidate classification unit combination is determined based on the data resampling mechanism, which specifically includes the following steps:

[0287] Step 111: Perform multiple random samplings with replacement on the sample set to generate multiple sample subsets.

[0288] Step 112: For each of the sample subsets, perform the step of constructing an iterative random forest model to identify the combination of the candidate classification units, and obtain a set of verification combinations corresponding to each of the sample subsets.

[0289] Step 113: Count the total number of times any candidate classification unit combination exists in all the verified combination sets.

[0290] Step 114: The relative ratio of the total number of times to the total number of sample subsets is determined as the stability score of any candidate classification unit combination.

[0291] In practice, the original sample set containing multimodal features is first subjected to multiple (let's say B times, e.g., B=50) Bootstrap resampling operations, i.e., random sampling with replacement. Because it's sampling with replacement, each sampling generates a dataset of the same size as the original sample set, but with slight differences in sample composition (some samples are duplicated, some are missing). Through this resampling operation, multiple sample subsets can be generated. This step, by artificially introducing data perturbation, provides a robust data foundation for subsequent evaluation of the anti-interference ability of microbial assemblages.

[0292] After generating multiple sample subsets, the steps of building an iterative random forest model to identify the combination of candidate classification units can be run independently for each sample subset.

[0293] Specifically, on each independent sample subset, the aforementioned iterative weighted feature selection based on phylogenetic smoothing, and the decision path parsing and intersection operation based on random cross-trees are repeatedly executed. By independently running the complete ecological interaction mining process on B subsets, B sets of higher-order interactions can be extracted, thus obtaining a set of validation combinations corresponding to each sample subset, denoted as . ,in b Representing the b Secondary resampling.

[0294] For any combination of candidate classification units identified in the early stage, let it be denoted as The process involves iterating through the B validation combination sets generated above and counting the total number of times any candidate classification unit combination exists in all validation combination sets. During this process, an indicator function is introduced. When a specific candidate classifier combination exists in the validation combination set generated in a certain resampling process... When the condition is met, the value of the indicator function is converted to 1; otherwise, it is converted to 0.

[0295] Then, by summing up the indicator function values ​​from all B resampling, the total number of successful reproductions of the combination can be obtained.

[0296] In the specific computational implementation, this embodiment can calculate the stability score of any candidate classification unit combination according to the following formula:

[0297] ;

[0298] in, B This represents the total number of sample subsets (i.e., the number of Bootstrap resampling attempts). is the indicator function. The stability score threshold is [0, 1], which intuitively reflects the reproducibility probability of a specific microbial combination under data perturbation. For example, if the stability score of a candidate taxonomic unit combination is 0.85, it means that the candidate taxonomic unit combination can be robustly identified under 85% of data perturbation conditions.

[0299] Metagenomic data are highly susceptible to sample heterogeneity, and traditional single-model training often outputs a large number of spurious associations. This embodiment perfectly solves the problem of false positive combinations caused by random noise in high-dimensional sparse sequencing data by introducing a bootstrap resampling and bagging mechanism. Through multiple sampling with replacement and independent repetition of the ecological interaction mining process on subsets, stability scores are calculated, and combinations that only appear occasionally in specific data slices are forcibly removed. This quantification mechanism ensures that the final output microbial taxonomic unit combinations have extremely high ecological reproducibility and real synergistic potential, significantly improving the colonization success rate of subsequent formulated microbial communities in complex physiological environments.

[0300] Based on the above embodiments, as an optional embodiment, an implementation method is provided to determine the candidate core microbial community based on all combinations of candidate taxonomic units whose stability scores meet a set threshold, mainly including but not limited to:

[0301] Extract all candidate classification unit combinations whose stability scores are greater than or equal to the set threshold, and use them as candidate ecological functional groups;

[0302] The microbial taxonomic units contained in all the candidate ecological functional groups are aggregated and summarized, and the aggregated set of all the microbial taxonomic units is determined as the candidate core microbial community.

[0303] After completing multiple rounds of resampling and score calculation, this embodiment pre-sets a preset stability threshold in the system. (For example, setting) Subsequently, all microbial assemblies with calculated scores are traversed, and all candidate taxonomic unit assemblies with stability scores greater than or equal to the set threshold are extracted as candidate core guilds. This means that the retained microbial community assemblies can be robustly identified under more than half of the data perturbations, exhibiting extremely high ecological reproducibility.

[0304] In the specific implementation data structure, each data entry in the output list of one or more candidate ecological functional groups contains the following multi-dimensional information:

[0305] (1) Combinatorial members, which consist of specific microbial taxonomic units, such as: {Faecalibacterium prausnitzii, Eubacterium hallii, Roseburia intestinalis}.

[0306] (2) Stability score, for example, 0.85 as calculated.

[0307] (3) Interaction order, for example, identified as 3-order cooperative interaction.

[0308] (4) Associated phenotypes, such as a negative correlation with the target disease (e.g., cirrhosis), indicate that it is a potential probiotic combination.

[0309] After successfully extracting one or more candidate ecological functional groups with strong ecological associations, this embodiment will further perform deduplication and merging operations, mainly by summarizing the microbial taxonomic units contained in all the candidate ecological functional groups.

[0310] Specifically, all microbial taxonomic entities in each independent high-order interaction module (e.g., three bacteria in module A and two bacteria in module B) are merged into the same macro data pool, and duplicate species are removed. Finally, the set of all the aforementioned microbial taxonomic entities after aggregation can be determined as the candidate core microbial community. This candidate core microbial community constitutes a high-fidelity candidate species library that is subsequently input into large-scale semantic reasoning verification and whole-genome metabolic network physical simulation.

[0311] Traditional differential abundance analysis can only output a single list of candidate species, which is insufficient to guide the rational design of complex microbial ecosystem formulations. This invention, by strictly setting stability thresholds and outputting multi-dimensional ecological functional groups, successfully achieves a leap from isolated species-level analysis to system-level analysis of community ecological functions. It not only effectively filters out low-frequency random false positive combinations caused by sequencing noise through a threshold truncation mechanism, but also, by defining and extracting candidate ecological functional groups containing multi-dimensional information such as membership, interaction order, and phenotypic associations, materializes abstract mathematical combinations into microbial communities with clear biological synergistic significance.

[0312] As mentioned above Figure 1 As shown in the overall technical architecture, Figure 1The final step, "determining the target formulation microbiota based on the candidate core microbiota," is the ultimate goal of the entire microecological drug design provided by this invention. After obtaining candidate core microbiota with strong ecological associations through iterative weighting and stability screening, in order to address the ecological functional mechanism question of "what these microorganisms do together" and ensure the clinical efficacy and safety of the final formulation, this invention also constructs a "semantic-physical" bimodal validation funnel.

[0313] Figure 12 This is a schematic diagram of a process for determining the target formulation microbial community based on candidate core microbial communities provided by the present invention, such as... Figure 12 As shown, the determination of the target formulation microbial community based on the candidate core microbial community specifically includes, but is not limited to:

[0314] Step 121: Input the entity information of each microbial taxonomic unit in the candidate core bacterial community into a pre-trained biological language model for reasoning analysis to obtain the semantic score of the candidate core bacterial community.

[0315] Step 122: Construct a microbial metabolic network model based on each microbial taxonomic unit in the candidate core microbial community.

[0316] Step 123: Perform community flux balance simulation based on the microbial metabolic network model to determine the metabolic support index among the microbial taxonomic units in the candidate core bacterial community.

[0317] The metabolic support index is used to characterize the intensity of metabolic intertrophic interaction between microbial taxa acting as donors and microbial taxa acting as recipients.

[0318] Step 124: Determine the comprehensive recommendation index based on the semantic score and the metabolic support index.

[0319] Step 125: Select the target formulation microbial community from the candidate core microbial communities based on the comprehensive recommendation index.

[0320] In practice, the entity information of each microbial taxonomic unit in the candidate core bacterial community is first extracted. The entity information may be the taxonomic ID (such as NCBI TaxID), Latin name and related biological metadata (such as Gram staining attributes, oxygen requirements, etc.) of the microbial taxonomic unit.

[0321] Subsequently, the aforementioned entity information is input into a pre-trained biological language model (such as Bio-LLM) for inference analysis. The biological language model refers to a large-scale natural language or sequence model (such as variants of DNABERT) that has been fine-tuned from massive amounts of biomedical literature, genome sequences, or metabolic pathway data.

[0322] By simulating the process of human experts consulting literature and metabolic pathway databases, this embodiment utilizes the biological language model to conduct in-silico co-culture analysis on candidate bacterial communities, thereby obtaining semantic scores for the candidate core bacterial communities. These semantic scores quantify the probability that existing scientific literature or genomic functional data supports the existence of a synergistic effect in this combination, representing a verifiable qualitative confidence level.

[0323] Furthermore, in order to compensate for the generative illusion problem that may exist in biological language models, this embodiment sets up a physical mechanism simulation stage, that is, to construct microbial metabolic network models (GEMs) based on each microbial taxonomic unit in the candidate core bacterial community.

[0324] Specifically, the genomic sequences of candidate strains are obtained, for example, from metagenomic assembly genomes (MAGs) or public reference genomes. The genes encoding enzymes are identified through sequence alignment, and a logical association of "gene-protein-reaction" is established. This allows for the automatic reconstruction of a quantitative mathematical network for each microbial taxonomic unit, containing all biochemical reactions and the relationship between metabolite consumption / generation.

[0325] After reconstructing the microbial metabolic network model, it was placed under virtual boundary conditions simulating a specific disease (such as the intestinal microenvironment of cirrhosis) to perform community flux balance simulation based on the microbial metabolic network model. When the simulation reached steady state, the metabolic support index (MSI) among the microbial taxa in the candidate core microbiota was calculated and determined.

[0326] The metabolic support index is primarily used to characterize the intensity of metabolic intertrophic interaction between the donor and recipient microbial taxa. Specifically, it quantitatively calculates the flux of specific exchange metabolites (such as acetic acid and amino acids) produced by strain A (donor) actually consumed by strain B (recipient). A higher metabolic support index indicates a greater degree of symbiotic dependence between the two at the underlying material flow level.

[0327] After completing the bimodal validation, qualitative probabilistic inference can be integrated with quantitative physical simulation to determine the comprehensive recommendation index (RI) based on semantic scores and metabolic support indices. In the actual mathematical calculation, a multi-objective utility function can be used to perform multi-dimensional weighted calculations of the semantic scores (representing semantic significance) and the metabolic support indices (representing physical significance), thereby obtaining a final comprehensive score that reflects the overall application value and colonization stability of the microbial community assemblage.

[0328] Finally, all candidate core microbiota were sorted in descending order according to the calculated comprehensive recommendation index, and the target formulation microbiota was selected from the candidate core microbiota based on the comprehensive recommendation index. Specifically, the best combination with the highest ranking (Top-K) was selected as the final live microbial biopharmaceutical (LBP) formulation for intervention against the target disease (such as cirrhosis).

[0329] This invention creatively constructs a dual-modal functional decoding architecture combining semantic reasoning from a large language model with physical simulation of a whole-genome metabolic network, completely solving the core pain points of uninterpretable AI predictions and black-box mechanisms in microbial formulation design. First, semantic scoring using a biological language model achieves literature-level logical verification. Then, metabolic support indices are calculated through metabolic network simulation, quantitatively confirming the actual cross-feeding flux between strains at the material flow level during the dry experiment stage. This dual verification mechanism, which transforms the probabilistic predictions of a large model into deterministic conclusions consistent with molecular biology principles, significantly eliminates invalid strain combinations with substrate competition and metabolic conflicts, providing a technical closed loop for accurately outputting target formulation microbial communities with high colonization rates and high synergistic efficacy.

[0330] Figure 13 This is an implementation logic diagram of a semantic-physical dual-function decoding technology based on a biological language model and a microbial metabolic network model provided by the present invention. To completely break the black-box state of artificial intelligence in microbial ecosystem analysis and transform probabilistic predictions into deterministic conclusions that conform to biological laws, this embodiment... Figure 13 The "two-dimensional verification funnel" architecture shown is analyzed in detail throughout the entire process. Based on Figure 13 The specific implementation logic can be divided into the following four core stages:

[0331] Phase 1: Hybrid sequence acquisition and static security circuit breaking (source control layer).

[0332] After receiving the candidate microbial functional group combination output by the front-end algorithm, this embodiment will first perform fine preparation of the underlying sequence and the first security check.

[0333] Specifically, in order to balance patient specificity and data integrity, a hybrid sequence acquisition strategy is adopted: patient-specific high-quality assembled genome MAGs are extracted first; if data is missing, an external database is called to match the best reference genome, and finally a standardized genome sequence (FASTA sequence) is output.

[0334] Following this, before proceeding with complex simulation calculations, a static safety circuit breaker is executed. This includes performing a specific scan at the sequence level to rigorously detect whether the combination carries virulence factors (such as tcdA / tcdB) or antibiotic resistance genes (MGE-ARGs) located on mobile genetic elements. Once this genomic red line is triggered, the assembled genome will face high-risk elimination, thereby avoiding the risk of iatrogenic infection and drug resistance spread from the source of probiotics and significantly saving subsequent computational resources.

[0335] Phase 2: Bio-LLM semantic function decoding and verification (qualitative logic verification layer).

[0336] For candidate genomes that pass the static security review, they are input into the Bio-LLM inference engine to perform semantic verification for silicon-based preclinical experiments.

[0337] First, entity semantic alignment and cue word construction are performed to transform rigid taxonomic IDs into structured cue words containing context such as Gram staining attributes and oxygen requirements. To overcome the generative illusion of large models, this embodiment utilizes an external knowledge base (containing KEGG metabolic pathways, VFDB virulence factor database, and the latest literature) to extract prior knowledge fragments through retrieval enhancement (RAG).

[0338] Subsequently, the engine initiates chain-like metabolic reasoning (CoT logic chain), judging the logical rationality of the combinations according to the cascade biochemical reaction sequence of substrate utilization, intermediate product transformation, and final product effect assessment. If the biological language model reasoning finds severe niche mutualism between strains (e.g., all are competitors for the same basic carbon source and lack cross-feeding enzymes), logical elimination is triggered to ensure that the combinations entering the next stage have a rigorous mutualistic symbiotic logic within the existing scientific literature system.

[0339] Phase 3: Quantitative simulation verification of metabolic networks (quantitative physical verification layer).

[0340] Semantic reasoning addresses the theoretical feasibility issue, while physical simulation addresses the actual potential output. This embodiment will reconstruct microbial genome-scale metabolic networks (GEMs) containing gene-protein-response mappings based on the aforementioned FASTA sequences.

[0341] To recreate the real host's internal environment, pathological environmental constraints (such as virtual media with high ammonia and high bile acid) are introduced, and a mixed integer linear programming algorithm is used under these constraints to perform gap filling and complete the missing key home-keeping responses.

[0342] Subsequently, community flux balance analysis (cFBA) is performed. During this dynamic simulation, a dynamic safety circuit breaker (hazard elimination) is executed simultaneously: the metabolic flux distribution under steady-state conditions is monitored in real time, and once the generation flux of harmful metabolites (such as hydrogen sulfide or ethanol) exceeds the preset safety threshold, the combination will be immediately eliminated, thereby avoiding the risk of aggravating damage to host organs such as the liver.

[0343] Phase 4: Quantitative indicator output and formula combination generation (decision output layer).

[0344] After passing through the aforementioned funnel filtering process, the surviving assembled genomes are considered the optimal combination, supported by literature and conforming to physical mass conservation. At this point, based on steady-state simulation results, two core quantitative indicators are simultaneously output:

[0345] First, the microbial metabolic mutualistic index (MSI) is used to precisely quantify the synergistic strength of cross-feeding between strains at the material flow level.

[0346] Secondly, specific functional targeting flux (Func) is used to evaluate the efficacy of targeted therapy in specific pathological environments (e.g., ammonia-lowering flux in cirrhosis).

[0347] Finally, by combining the above quantitative indicators, a validated high-confidence formulation microbial strain combination is output, namely the target formulation microbial community.

[0348] Figure 14 This is a schematic diagram of a process for obtaining semantic scores of candidate core bacterial communities provided by the present invention, such as... Figure 14 As shown, the step of inputting the entity information of each microbial taxonomic unit in the candidate core bacterial community into a pre-trained biological language model for reasoning analysis to obtain the semantic score of the candidate core bacterial community specifically includes, but is not limited to, the following steps:

[0349] Step 141: Based on the taxonomic identifiers of each microbial taxonomic unit in the candidate core bacterial community, obtain the corresponding standard entity names and biological metadata to construct structured prompt information.

[0350] Step 142: Use the structured prompt information to perform feature retrieval in a pre-built external knowledge base to obtain prior knowledge fragments related to the candidate core bacterial community.

[0351] Step 143: The prior knowledge fragment and the structured prompt information are fused and then input into the biological language model to obtain the semantic score of the candidate core bacterial community output by the biological language model.

[0352] In practice, the taxonomic identifiers (e.g., NCBI TaxID) of the candidate core bacterial groups output in the early stage are converted into standard entity names (e.g., Faecalibacterium prausnitzii) commonly used in the pre-training corpus of the large model. At the same time, the system automatically extracts the biological metadata of the microbial taxonomic unit (e.g., Gram staining attributes, oxygen requirements, etc.) and generates a structured text description.

[0353] Furthermore, this captured information can be integrated to obtain corresponding standard entity names and biological metadata based on the taxonomic identifiers of each microbial taxonomic unit in the candidate core microbial community, thereby constructing structured prompts. For example, prompts can be constructed that include role settings, task objects, and core tasks (such as predicting synergistic effects based on the pathological background of liver cirrhosis), ensuring that the large model can understand the biological meaning behind the numbers.

[0354] To overcome the lag and generation illusion of language model training data, this embodiment further introduces a retrieval-enhanced generation (RAG) mechanism. Data such as metabolic pathways, metabolic responses, virulence factors, and recent authoritative journal abstracts are pre-integrated, vectorized, and stored in an external knowledge base. When structured prompts are received, feature retrieval can be performed in the pre-built external knowledge base using these prompts to obtain prior knowledge fragments related to the candidate core bacterial community. For example, when querying a specific strain, the system automatically retrieves key attributes such as "major butyric acid-producing bacteria" and "anti-inflammatory" as prior knowledge fragments (Context).

[0355] Furthermore, in this embodiment, the search results will be appended to the prompt information, forcing the biological language model to generate answers based on facts.

[0356] During the reasoning process, the biological language model performs step-by-step reasoning according to the cascade biochemical reaction sequence of substrate utilization analysis, intermediate product transformation analysis, and final product effect assessment to obtain the semantic score.

[0357] Specifically, the biological language model can simulate chain reasoning (CoT) of biochemical reactions: first, it analyzes whether strain A can utilize undigested complex carbohydrates from the host to produce acetic acid; second, it verifies whether strain B possesses the enzyme system to utilize the products of strain A to confirm cross-feeding symbiosis; finally, it assesses the total amount and proportion of short-chain fatty acids (SCFA) produced, and evaluates tolerance in conjunction with pathological environments (such as high bile acid), ultimately outputting a semantic score representing the probability that existing scientific literature supports the effectiveness of this combination.

[0358] This invention, through the organic combination of structured prompts, RAG knowledge mounting, and CoT chain reasoning, enables biological language models to go beyond shallow text generation and perform rigorous biochemical cascade reasoning like domain experts, achieving white-box analysis of candidate bacterial community synergy mechanisms at the literature corpus level.

[0359] Figure 15 This is one of the flowcharts illustrating the construction of a microbial metabolic network model provided by the present invention, such as... Figure 15 As shown, based on the above embodiments, to verify the physical authenticity of semantic prediction, the physical simulation stage of the whole-genome metabolic network is described in detail below. Specifically, an implementation method for constructing a microbial metabolic network model based on each microbial taxonomic unit in the candidate core microbial community is provided, including but not limited to:

[0360] Step 151: Obtain the genome sequence of each of the microbial taxonomic units in the candidate core bacterial community.

[0361] Step 152: Identify the genes encoding enzymes in the genome sequence, establish the mapping relationship between genes, proteins and biochemical reactions, and reconstruct the initial metabolic network model.

[0362] Step 153: Introduce preset pathological environment constraints to fill metabolic gaps in the initial metabolic network model to obtain the microbial metabolic network model.

[0363] In a specific implementation, a hybrid acquisition strategy of "MAGs first, reference genome as a fallback" is adopted. If the quality of the metagenomic assembly genome MAGs of the sample meets the standard, the genome sequence is directly extracted; if high-quality MAGs are not obtained, the best matching reference genome sequence is retrieved from a public database as a substitute, thereby obtaining the genome sequences of each of the microbial taxa in the candidate core bacterial community.

[0364] Furthermore, genome-scale metabolic network reconstruction tools can be used to analyze the genome sequence, identify genes encoding enzymes within the sequence, establish mapping relationships between genes, proteins, and biochemical reactions, and reconstruct an initial metabolic network model. At the underlying data structure, a stoichiometric matrix is ​​automatically generated. .in, m Indicates the amount of metabolites. n Represents the number of biochemical reactions, matrix elements Representing the j In the first reaction i The stoichiometric coefficients of each metabolite. If This indicates that the metabolite has been consumed (reactant). This indicates that the metabolite has been generated (product). This means that the metabolite does not participate in the reaction.

[0365] In practical metagenomic sequencing, especially for genome sequences extracted from metagenomically assembled genomes (MAGs), insufficient sequencing coverage often results in numerous "open circuits" in the directly reconstructed metabolic network. To eliminate simulation bias caused by differences in model scale, this invention provides a method for filling metabolic gaps in the initial metabolic network model by introducing pre-set pathological environment constraints, thereby obtaining a microbial metabolic network model.

[0366] Figure 16 This is the second schematic diagram of a process for constructing a microbial metabolic network model provided by the present invention, as shown below. Figure 16 As shown, the introduction of preset pathological environment constraints to fill metabolic gaps in the initial metabolic network model yields the microbial metabolic network model, specifically including but not limited to:

[0367] Step 161: Configure virtual culture medium boundary conditions as pathological environment constraints based on the intestinal microenvironment characteristics of the target disease. The virtual culture medium boundary conditions include the inflow threshold of specific amino acids and the initial concentration of specific nitrogen sources.

[0368] Step 162: Identify key housekeeping responses that are missing in the initial metabolic network model due to insufficient sequencing coverage of the genome sequence.

[0369] Step 163: Under the constraints of the virtual culture medium boundary conditions, the initial metabolic network model is iteratively optimized using a mixed integer linear programming algorithm to supplement the missing key homeostatic reactions into the initial metabolic network model.

[0370] Step 164: When it is determined that the supplemented metabolic network model can generate the preset biomass in a simulated culture medium that meets the boundary conditions of the virtual culture medium, it is identified as the microbial metabolic network model.

[0371] In specific implementation plans, simulations are not conducted in an ideal, general nutritional environment, but rather strict boundaries must be set according to the target application scenario.

[0372] Taking intestinal microecological dysbiosis in liver cirrhosis as an example, virtual culture medium boundary conditions are configured as pathological environmental constraints based on the intestinal microenvironment characteristics of the target disease. Specifically, the virtual culture medium boundary conditions include inflow thresholds for specific amino acids and initial concentrations of specific nitrogen sources.

[0373] In practice, the lower limit of the inflow of branched-chain amino acids (BCAAs) can be set to an extremely low value, while the upper limit of the inflow of aromatic amino acids (AAAs) can be set to an extremely high value, so as to accurately simulate the typical amino acid imbalance in the intestines of patients with cirrhosis. At the same time, a high concentration of ammonia is set as the only initial nitrogen source for the model, which is used to test the ammonia-reducing function of the bacterial community in subsequent tests.

[0374] After setting the boundary conditions, a connectivity scan is performed on the initial metabolic network to identify key housekeeping reactions that are missing in the initial metabolic network model due to insufficient sequencing coverage of the genome sequence. These housekeeping reactions are biochemical steps necessary to maintain the basic metabolism of microbial cells (such as core carbon metabolism and nucleotide synthesis), and their absence will cause the microbial metabolic network model to fail in computational simulation.

[0375] To repair the network, this embodiment invokes a global database containing universal biochemical reactions. Under the constraints of the virtual culture medium boundary conditions, a mixed-integer linear programming (MILP) algorithm is used to iteratively optimize the initial metabolic network model, supplementing the missing key biochemical reactions into the initial metabolic network model. The mathematical objective of this optimization is to introduce the minimum number of non-native biochemical reactions from the global database to reconnect broken circuits while ensuring network connectivity.

[0376] In practice, the repaired model is continuously tested. When it is determined that the supplemented metabolic network model can generate the preset biomass in a simulated culture medium that meets the boundary conditions of the virtual culture medium, it is identified as the microbial metabolic network model. This step ensures that the network connectivity of all strain models built based on sequencing data of different quality is aligned to the same order of magnitude as models built based on the reference genome, thereby completely eliminating the deviations in subsequent physical simulations caused by differences in data quality.

[0377] After the metabolic network models of all candidate microbial taxa are completed, this embodiment can begin to perform community-level metabolic interaction calculations.

[0378] Figure 17 This is a flowchart illustrating a process for determining the metabolic support index between microbial taxa in a candidate core microbial community, as provided by the present invention. Figure 17 As shown, community flux balance simulation is performed based on the microbial metabolic network model to determine the metabolic support index among the microbial taxa in the candidate core bacterial community, specifically including but not limited to:

[0379] Step 171: Construct a virtual community model containing independent strain compartments corresponding to each of the microbial taxonomic units and shared extracellular compartments.

[0380] Step 172: Configure the material inflow boundary conditions for the shared extracellular chamber and connect the microbial metabolic network model to the corresponding independent strain compartment.

[0381] Step 173: Under the constraint of the material inflow boundary conditions, the flux balance is solved with the maximization of total community biomass as the joint objective function to obtain the optimal flux vector for each biochemical reaction.

[0382] Step 174: Based on the optimal flux vector, extract the generation rate of a specific exchange metabolite released from the donor microbial taxa unit into the shared extracellular cavity, and the consumption rate of the recipient microbial taxa unit absorbing the specific exchange metabolite from the shared extracellular cavity.

[0383] Step 175: After obtaining the minimum value between the generation rate and the consumption rate, the metabolic support index is obtained by weighted fusion in combination with the preset importance weight of the specific exchange metabolite.

[0384] In practical implementation, to accurately simulate the synergistic interactions within the microbial community, this embodiment constructs a virtual community model comprising independent strain compartments corresponding to each microbial taxonomic unit and a shared extracellular space (lumen compartment). In the underlying data architecture, each candidate microbial taxonomic unit, such as... Each is modeled as an independent and closed compartment, each possessing its own pre-reconstructed, independent enzyme system and metabolic network matrix.

[0385] Furthermore, in the simulation rules, metabolites (such as acetic acid, ammonia, and glucose) are allowed to exchange substances between individual strain compartments and the shared extracellular cavity through specific transport reactions, but direct exchange of enzymes or large protein molecules between different strain compartments is strictly prohibited. Simultaneously, an upper limit on the inflow of substances into the shared cavity is set to simulate a typical patient diet (such as a low-protein diet and the inflow of undigested complex carbohydrates), and a high bile acid inflow flux is set to simulate pathological inhibitory pressure.

[0386] The growth and colonization of microorganisms in the gut are abstracted as a mathematical optimization problem under strict mass conservation constraints. Electronic devices employ community flux balance analysis (cFBA), using the maximization of total community biomass as the joint objective function to solve for flux balance under the constraints of the aforementioned material inflow boundary conditions.

[0387] Specifically, the joint objective function can be:

[0388] ;

[0389] in, For the first k The relative abundance coefficient of each microbial taxonomic unit; This represents the growth rate of the microbial taxonomic unit.

[0390] Meanwhile, the linear programming solver is used to solve the following system of equations under hard constraints:

[0391] ;

[0392] ；

[0393] ;

[0394] in, Representing microbial taxonomic units k In the reaction j Metabolites i stoichiometric coefficients; Indicates the reaction rate; and These represent the lower and upper limits of flux, respectively. Representing microbial taxonomic units k Metabolites are ingested from shared environments. m The rate; Represents metabolites in a simulated gut environment m The maximum available flow rate.

[0395] When the objective function converges, the optimal flux vector for each biochemical reaction can be obtained. .

[0396] After reaching steady state, to quantitatively assess the synergistic effect, this embodiment analyzes the flow of substances within the shared extracellular compartment. For any specific exchange metabolite m that may undergo crossfeeding (e.g., short-chain fatty acids, vitamins, amino acids, etc., typically excluding non-specific substances such as water and protons), based on the optimal flux vector, the generation rate of the specific exchange metabolite released from the donor microbial taxa into the shared extracellular compartment is extracted (denoted as ). ), and the consumption rate (denoted as ) of the microbial taxonomic unit acting as the receptor absorbing the specific exchange metabolite from the shared extracellular compartment. ).

[0397] Since the actual mutual-support flux is inevitably limited by the bottleneck between the "supply side" and the "demand side," the electronic device obtains the minimum value between the generation rate and the consumption rate, i.e. .

[0398] When the optimal state is reached, the Metabolic Support Index (MSI) is calculated, which is a quantitative calculation of the flux of metabolites produced by strain A being consumed by strain B, denoted as [missing information]. This is used to quantitatively assess the synergistic effect, and can be calculated using the following summation formula:

[0399] ;

[0400] in, m To exchange metabolites, such as acetic acid, vitamin B12, and amino acids; This indicates the metabolic contribution of strain A to strain B. The larger the value, the more metabolites produced by A are utilized by B, and the higher the symbiotic dependence between the two. This indicates a list of all metabolites that may be exchanged between strains, typically excluding non-specific substances such as water and protons, and focusing on carbon sources, nitrogen sources, and cofactors; This indicates that strain A releases metabolites into the shared environment. m The rate; This indicates the rate at which strain B consumes metabolite m from the shared environment; This indicates that the mutual support flux is limited by shortcomings on both the supply and demand sides; The weights representing the importance of metabolites are coefficients pre-defined based on pathological characteristics (for example, for the treatment of cirrhosis, the weight of short-chain fatty acids (SCFA) is set to 2.0, and that of common carbohydrates is set to 0.5).

[0401] In order to meet the extremely high safety standards of in vivo biological drugs in industrial production and clinical application, and to ultimately output the best formulation that takes into account both colonization stability and targeted therapeutic effect, after performing the aforementioned semantic scoring based on a large language model and metabolic support index calculation based on community flux balance simulation, this embodiment further sets up a stringent pre-safety circuit breaker mechanism and a multi-dimensional comprehensive scoring fusion.

[0402] Specifically, before selecting the target formulation microbial community from the candidate core microbial communities based on the comprehensive recommendation index, a tiered safety circuit breaker mechanism (i.e., "one-vote veto") is implemented to prevent the misselection of microbial combinations with fatal risks. This mechanism includes, but is not limited to:

[0403] The genome sequence of each microbial taxonomic unit in the candidate core bacterial community is analyzed to detect whether there are preset pathogenic risk gene features, including virulence factors or antibiotic resistance genes located on mobile genetic elements.

[0404] During the execution of the community flux balance simulation, it is monitored whether the preset flux of harmful metabolites exceeds the safety limit value.

[0405] In response to the detection of the pathogenic risk gene feature or the generation throughput exceeding the safety threshold, the corresponding candidate core microbial community is removed from the candidate pool.

[0406] In practice, a static safety pre-circuit breaker is first implemented at the genomic level, including: scanning and parsing the genomic sequence of each microbial taxonomic unit in the candidate core bacterial community, and detecting whether there are preset pathogenic risk gene features by comparing with a professional external pathogen feature database.

[0407] The pathogenic risk gene characteristics mainly include virulence factors or antibiotic resistance genes located on mobile genetic elements. For example, this embodiment specifically identifies the presence of virulence factors (such as the tcdA / B gene) and resistance genes (MGE-ARGs) located on mobile elements such as plasmids or transposons. This static genomic redline screening blocks the risk of highly dangerous pathogens and drug resistance spread at the source, and can eliminate high-risk combinations at an early stage to save the extremely computationally intensive physical simulation resources in the future.

[0408] Microbial assemblies that pass the static safety screening will enter the dynamic safety post-interception stage. Since some symbiotic bacteria may produce toxic substances that aggravate the host's condition under specific pathological conditions, it is necessary to monitor whether the pre-set flux of harmful metabolites exceeds the safety limit during the execution of the community flux balance simulation. For example, for liver damage diseases such as cirrhosis, dynamic monitoring is needed to determine whether excessive hydrogen sulfide or endogenous ethanol is produced under the simulated steady state. Specifically, when the flux of hydrogen sulfide production is detected... Or the flux of ethanol production At that time, it is determined that the dynamic metabolism red line has been crossed.

[0409] This embodiment further establishes a stringent "one-vote veto" logic gate in the aforementioned multidimensional security detection. In response to the detection of the pathogenic risk gene feature or the generation throughput exceeding the security threshold, the recommendation index of the combination will be directly forced to zero and marked as "high-risk elimination," thereby removing the corresponding candidate core bacterial community from the candidate pool.

[0410] This invention replaces the traditional, delayed "post-experimental testing" in in vivo biopharmaceutical development with "computational screening and circuit breaking," constructing static genomic red lines (screening virulence factors and MGE-ARGs) and dynamic metabolic red lines (monitoring metabolic toxicity fluxes such as hydrogen sulfide and ethanol). This allows for strict avoidance of iatrogenic infections and organ damage caused by microecological drugs during the digital simulation stage, ensuring that the final microbial combination fully meets the stringent safety requirements for clinical trial applications from the underlying design level.

[0411] As an optional embodiment, for candidate core microbial communities that successfully pass the above-mentioned dynamic and static safety checks, a comprehensive recommendation index will be calculated based on a multi-objective utility function to perform the final formulation optimization.

[0412] Specifically, determining the comprehensive recommendation index based on the semantic score and the metabolic support index includes:

[0413] Obtain the stability score corresponding to the candidate core bacterial community;

[0414] Obtain the functional targeting score of the candidate core bacterial community for a specific pathological target in the community flux balance simulation.

[0415] The stability score, the functional targeting score, the semantic score, and the metabolic support index are weighted and fused together to obtain the comprehensive recommendation index.

[0416] In order to comprehensively evaluate the overall value of the microecological formula, this embodiment not only requires the semantic score, which represents the verifiable evidence (i.e., the probability that the combination is effective, supported by existing scientific literature), and the metabolic support index, which represents true symbiosis (i.e., the amount of material exchange that actually occurs in the simulation calculation), but also requires the introduction of quantitative indicators of ecological stability and specific therapeutic effects.

[0417] First, obtain the stability score corresponding to the candidate core bacterial community, denoted as . The value is derived from the stable reproducibility calculated earlier based on the Bootstrap resampling mechanism, representing the ecological robustness of the bacterial community in coexisting under data perturbation.

[0418] Simultaneously, this embodiment also obtains the functional targeting score of the candidate core bacterial community for a specific pathological target in the community flux balance simulation, denoted as... This score refers to the specific flux calculation results for the target application scenario (such as the pathological characteristics of cirrhosis), such as calculating the total ammonia uptake flux of a virtual community, thereby representing the targeted therapeutic potential of the combination (such as ammonia-lowering function).

[0419] Finally, after aggregating the core indicators from the above four dimensions, the electronic device performs a multi-objective weighted fusion of the stability score, the functional targeting score, the semantic score, and the metabolic support index to obtain the comprehensive recommendation index, the specific calculation formula of which can be expressed as:

[0420] ;

[0421] in, This represents the stable reproducibility rate of the Bootstrap algorithm in the Phy-iRF algorithm, indicating that they can coexist. This indicates that the score given by the large model based on literature retrieval represents the probability that existing scientific literature supports the validity of this combination, i.e., it is verifiable. For example, if the literature clearly states that bacteria A and bacteria B have a mutualistic relationship, then... Set as high score. This indicates that the metabolic support index calculated based on GEMs is hard physical evidence, representing the actual amount of material exchange that occurs in the simulation calculation, i.e., true mutualism exists. The higher the value, the stronger the dependency between strains and the higher the colonization stability. This represents a specific flux calculation for the liver cirrhosis scenario, namely the functional target score, which indicates therapeutic efficacy, such as ammonia-lowering effects. , , and The importance of the four dimensions mentioned above is pre-defined with weights. Furthermore, in practical applications, it is recommended to set... and A high weight (e.g., 0.35) is given to emphasize the importance of physical simulation evidence.

[0422] Once the calculations are complete, all candidate core bacterial communities can be ranked according to the calculated comprehensive recommendation index. RI The microbiota are sorted in descending order, and the best combination of the top-ranked (e.g., Top-K) microbiota is selected as the therapeutic target microbiota for formal export.

[0423] This invention provides a scoring mechanism based on a rigorous logic funnel, which ensures that the final selected formulation microbiota can not only coexist stably in the human population, but also have real mutual metabolic support at the physical level, providing the most valuable digital design scheme for the subsequent clinical-grade precision development of in vivo biological drugs.

[0424] Based on the above embodiments, as an optional embodiment, after determining the target formulation microbial community, the optimal inoculation stoichiometry is further solved based on the law of mass action, specifically including:

[0425] Obtain the target steady-state relative abundance of each microbial taxa of the target formulation microbial community under simulated steady-state conditions;

[0426] Based on the steady-state relative abundance, a set of community dynamic differential equations including interspecific interaction coefficients, intrinsic growth rates, and environmental carrying capacity is constructed.

[0427] The physiological environmental parameters of the target application scenario are used as initial constraints to input the community dynamic differential equation system, and the inoculation concentration ratio of each microbial taxa that can enable the community to reach the target steady-state relative abundance after a preset colony cycle is derived in reverse as the optimal inoculation stoichiometry.

[0428] Based on the optimal inoculation stoichiometry, industrial preparation formulation data for the target microbial community are generated.

[0429] After determining "which bacteria to use", this embodiment further addresses the question of "how much to use".

[0430] In practice, the proportion of naturally occurring selected strains in the healthy gut is first extracted from the metagenomic data of the healthy control group (e.g., the natural ratio of strains A, B, and C is 6:3:1), and this proportion is obtained as the target steady-state relative abundance of each microbial taxa of the target formulation flora under simulated steady-state conditions. This value will serve as the initial reference value for subsequent formulation optimization.

[0431] To accurately quantify the dynamic proliferation and substrate consumption processes of microorganisms in the real gut environment, the electronic device constructed a set of community dynamic differential equations based on the steady-state relative abundance, including interspecific interaction coefficients, intrinsic growth rates, and carrying capacity. During the construction of this equation set, the electronic device performed flux balance correction analysis, i.e., using genome-wide metabolic models (GEMs) to perform limiting substrate scanning and identify limiting cross-feeding metabolites (such as acetic acid, specific amino acids, etc.) in the community. To incorporate the physical laws of metabolic mutualism into the differential equations, the electronic device embedded supply-demand balance constraint equations into the community dynamic differential equation set based on the law of mass action.

[0432] ;

[0433] in, and These represent the colony-forming units (CFU) or concentrations of the donor and recipient bacteria, respectively. The flux rate at which donor bacteria produce limiting metabolites; This indicates the flux rate at which the recipient bacteria consume the metabolite; k (Dimensionless) represents the conversion efficiency coefficient of various biochemical reactions, reflecting the loss of metabolites during intercellular transport and non-ideal stoichiometry, typically ranging from 0.8 to 1.0. Assuming ideal transport conditions, it can be set... k =1; or corrected based on the substrate conversion rate determined by in vitro co-culture experiments.

[0434] After establishing the dynamic model, physiological environmental parameters of the target application scenario can be used as initial constraints input into the community dynamic differential equation system. For example, if physical simulations reveal that the recipient bacteria are highly dependent on acetic acid produced by the donor bacteria, and insufficient acetic acid supply during inoculation at a completely natural ratio will lead to limited growth of the recipient bacteria, then the gradient ascent method can be used to... and The ratio is iteratively adjusted.

[0435] The objective function is to maximize the overall community growth rate or the production rate of key functional products (such as butyric acid). When the objective function converges, the inoculation concentration ratio of each microbial taxa that enables the community to reach the target steady-state relative abundance after a predetermined colony cycle is derived in reverse, serving as the optimal inoculation stoichiometry. For example, the initial 6:3:1 ratio might be dynamically adjusted to 4:1:2. By calculating the optimal stoichiometry to maintain the maximum growth rate, the relative abundance of donor bacteria can be reasonably increased to meet the substrate requirements of recipient bacteria.

[0436] Finally, based on the optimal inoculation stoichiometry, industrial preparation formulation data for the target microbial community can be generated. The output will be formatted digital formulation information, such as a structured result containing "Formulation ID (e.g., LC-LBP-01)," "Strain Composition (e.g., including specific species and preservation numbers such as Faecalibacterium prausnitzii)," and "Optimal Ratio (CFU ratio)," directly guiding downstream industrial fermenter feeding and capsule encapsulation processes. The output format can be:

[0437] {

[0438] "Formula ID": "LC-LBP-01";

[0439] "Strain composition": "Faecalibacterium prausnitzii (DSM 17677), Blautia" producta (DSM 2950), Roseburia intestinalis (DSM 14610)”;

[0440] "Optimal ratio (CFU)": "4:1:2"

[0441] }。

[0442] This invention utilizes a whole-genome metabolic model to identify limiting substrates and establishes a rigorous supply-demand balance equation based on the law of mass action, completely solving the core pain point of traditional microecological preparation development: the lack of quantitative calculation basis for inoculation ratios and heavy reliance on experimental trial and error. By iteratively solving the optimal inoculation stoichiometry ratio that maximizes the total biomass of the community using the gradient ascent method, it not only achieves fully digital design of microecological drug formulations but also preemptively resolves the resource competition problem in the process of industrial fermentation and colonization in the human body by addressing the underlying material flow.

[0443] Figure 18 This is a schematic diagram of the structure of the formulation microbial community construction device provided by the present invention, as shown below. Figure 18 As shown, the present invention also provides a formulation microbial community construction device, which mainly includes, but is not limited to, the following modules:

[0444] The phylogenetic tree construction module 181 is used to extract representative genomic sequences from multiple microbial taxonomic units in the sample set to construct a phylogenetic tree;

[0445] The phylogenetic weight conversion module 182 is used to calculate the evolutionary distance between each pair of microbial taxonomic units in the phylogenetic tree, and convert the evolutionary distance into a phylogenetic weight matrix, which is used to characterize the phylogenetic similarity between the microbial taxonomic units.

[0446] Abundance feature processing module 183 is used to obtain the original microbial abundance matrix of the sample set, perform a central log ratio transformation on the original microbial abundance matrix to obtain a feature matrix, and use the feature matrix as the feature dimension input to construct an iterative random forest model.

[0447] The feature importance correction module 184 is used to correct the original feature importance of each microbial taxonomic unit to a smoothed feature importance using the phylogenetic weight matrix during the feature selection iteration of the iterative random forest model, so as to update the feature sampling weights of the iterative random forest model for the microbial taxonomic unit in the next iteration based on the smoothed feature importance.

[0448] The higher-order rule extraction module 185 is used to extract higher-order ecological interaction rules based on the decision path of the iterative random forest model when the iterative random forest model reaches the preset convergence condition, and output candidate core bacterial communities.

[0449] The target microbial community selection module 186 is used to determine the target formulation microbial community based on the candidate core microbial community.

[0450] It should be noted that the formulation microbial community construction device provided by the present invention can execute the formulation microbial community construction method described in any of the above embodiments during specific operation, and will not be elaborated on in this embodiment.

[0451] The formulation microbial community construction device provided by this invention introduces phylogenetic relationships to smooth the characteristics, breaking the independence assumption of traditional abundance data, effectively overcoming the sparsity and false positive interference of high-dimensional sequencing data, accurately capturing and retaining the real interspecies synergistic relationships driven by evolution, thereby ensuring the efficient homeostatic colonization and precise targeted metabolic regulation of the formulation microbial community in the host microenvironment.

[0452] Specifically, the formulation microbial community construction device provided by this invention constructs a closed-loop microecological drug design method of "evolutionary constraint mining - semantic and physical dual verification - safe quantitative output", which solves the three major pain points of the existing technology: "correlation is not causal", "AI illusion is uncontrollable" and "formulation ratio is based on experience".

[0453] At the algorithm level, a novel approach is taken to introduce the quantitative indicator of evolutionary biology (Patristic Distance) as a prior constraint. During the feature selection process, the Gini Importance is smoothed by using kinship distance, which forces the model to prioritize capturing functional groups with evolutionary conservation under sparse data conditions. The black-box model is then transformed into a white-box system from which explicit interaction rules (RIT paths) can be extracted.

[0454] At the validation level, a dual-modal validation process combining semantic reasoning and metabolic simulation is implemented. By constructing a dual validation funnel that combines Bio-LLM (semantic knowledge analysis) and GEMs / cFBA (physical law verification), the biological logic rationality is verified (qualitatively) using CoT logical chain reasoning; flux balance analysis (FBA) is performed using a whole-genome metabolic model to calculate the metabolic support index MSI, transforming the probabilistic predictions of AI into deterministic conclusions that conform to the laws of molecular biology.

[0455] At the decision-making level, a tiered "one-vote veto" logic gate was established. This involved specific detection of antibiotic resistance genes (MGE-ARGs) located on mobile genetic elements and simulation calculations of metabolic toxicity fluxes (such as H2S and ethanol). This approach mitigated the risks of iatrogenic infections and drug resistance spread associated with probiotics from the source, meeting the stringent safety requirements for IND applications.

[0456] At the output level, a computational method based on the supply and demand balance of limiting substrates is proposed. An equation is established using the law of mass action, and the optimal inoculum stoichiometry (CFU Ratio) is iteratively calculated, enabling the digital design of drug formulations and ensuring that the bacterial community reaches its maximum growth rate and functional output under specific conditions.

[0457] Figure 19 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 19As shown, the electronic device may include: a processor 1910, a communications interface 1920, a memory 1930, and a communications bus 1940, wherein the processor 1910, the communications interface 1920, and the memory 1930 communicate with each other through the communications bus 1940. Processor 1910 can call logic instructions in memory 1930 to execute a formula microbial community construction method. This method includes: extracting representative genomic sequences from multiple microbial taxa in a sample set to construct a phylogenetic tree; calculating the evolutionary distance between each pair of microbial taxa in the phylogenetic tree, converting the evolutionary distance into a phylogenetic weight matrix, the phylogenetic weight matrix being used to characterize the phylogenetic similarity between the microbial taxa; obtaining the original microbial abundance matrix of the sample set, performing a central logarithmic ratio transformation on the original microbial abundance matrix to obtain a feature matrix, and using the feature matrix as a feature dimension input to construct an iterative random forest model; during the feature selection iteration of the iterative random forest model, using the phylogenetic weight matrix to correct the original feature importance of each microbial taxa to a smoothed feature importance, so as to update the feature sampling weights of the iterative random forest model for the microbial taxa in the next iteration based on the smoothed feature importance; when the iterative random forest model reaches a preset convergence condition, extracting high-order ecological interaction rules based on the decision paths therein, and outputting candidate core microbial communities; and determining the target formula microbial community based on the candidate core microbial communities.

[0458] Furthermore, the logical instructions in the aforementioned memory 1930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0459] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the formulation microbial community construction method provided in the above embodiments, the method comprising: extracting representative genomic sequences of multiple microbial taxonomic units in a sample set to construct a phylogenetic tree; calculating the evolutionary distance between each pair of microbial taxonomic units in the phylogenetic tree, converting the evolutionary distance into a phylogenetic weight matrix, the phylogenetic weight matrix being used to characterize the phylogenetic similarity between the microbial taxonomic units; and obtaining the original abundance of microorganisms in the sample set. The original microbial abundance matrix is ​​transformed using a central logarithmic ratio to obtain a feature matrix. This feature matrix is ​​then used as the feature dimension input to construct an iterative random forest model. During the feature selection iteration of the iterative random forest model, the original feature importance of each microbial taxonomic unit is corrected to a smoothed feature importance using the phylogenetic weight matrix. This smoothed feature importance is then used to update the feature sampling weights of the iterative random forest model for the microbial taxonomic unit in the next iteration. When the iterative random forest model reaches a preset convergence condition, higher-order ecological interaction rules are extracted based on the decision paths within the model, and candidate core microbial communities are output. The target formulation microbial community is then determined based on the candidate core microbial communities.

[0460] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for constructing a formulated microbial community provided in the above embodiments. The method includes: extracting representative genomic sequences of multiple microbial taxonomic units from a sample set to construct a phylogenetic tree; calculating the evolutionary distance between each pair of microbial taxonomic units in the phylogenetic tree, converting the evolutionary distance into a phylogenetic weight matrix, the phylogenetic weight matrix being used to characterize the phylogenetic similarity between the microbial taxonomic units; obtaining the original microbial abundance matrix of the sample set, and performing... The feature matrix is ​​obtained by central logarithmic ratio transformation, and the feature matrix is ​​used as the feature dimension input to construct an iterative random forest model. During the feature selection iteration of the iterative random forest model, the original feature importance of each microbial taxonomic unit is corrected to a smoothed feature importance using the phylogenetic weight matrix, so as to update the feature sampling weights of the iterative random forest model for the microbial taxonomic unit in the next iteration based on the smoothed feature importance. When the iterative random forest model reaches the preset convergence condition, higher-order ecological interaction rules are extracted based on the decision path, and candidate core bacterial groups are output. The target formulation bacterial group is determined based on the candidate core bacterial groups.

[0461] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0462] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0463] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a formula microbial community, characterized in that, include: Representative genomic sequences from multiple microbial taxa in the sample set were extracted to construct a phylogenetic tree; Calculate the evolutionary distance between each pair of microbial taxonomic units in the phylogenetic tree, and convert the evolutionary distance into a phylogenetic weight matrix. The phylogenetic weight matrix is ​​used to characterize the kinship similarity between the microbial taxonomic units. Obtain the original microbial abundance matrix of the sample set, perform a central log ratio transformation on the original microbial abundance matrix to obtain the feature matrix, and use the feature matrix as the feature dimension input to construct an iterative random forest model; During the feature selection iteration of the iterative random forest model, the original feature importance of each microbial taxonomic unit is corrected to a smoothed feature importance using the phylogenetic weight matrix, so as to update the feature sampling weights of the iterative random forest model for the microbial taxonomic unit in the next iteration based on the smoothed feature importance; When the iterative random forest model reaches the preset convergence condition, high-order ecological interaction rules are extracted based on the decision path and candidate core bacterial communities are output. The target formulation microbial community was determined based on the candidate core microbial community.

2. The method for constructing a formulated microbial community according to claim 1, characterized in that, The process of performing a central logarithmic ratio transformation on the original microbial abundance matrix to obtain the feature matrix includes: Replace the zero-value features in the original microbial abundance matrix with preset non-zero values; For the original microbial abundance matrix after replacement processing, determine the geometric mean parameter of the abundance vector of each sample in the sample set; Obtain the relative ratio of each abundance value in the abundance vector of each sample to the corresponding geometric mean parameter; The feature matrix is ​​obtained by performing a logarithmic scaling on all the relative ratios.

3. The method for constructing a formulated microbial community according to claim 1, characterized in that, The calculation of the evolutionary distance between each pair of microbial taxa in the phylogenetic tree, and the conversion of the evolutionary distance into a phylogenetic weight matrix, includes: Determine the nearest common ancestor node between each pair of microbial taxa in the phylogenetic tree; Based on the first path span between each of the microbial taxonomic units and the root node of the phylogenetic tree, and the second path span between the nearest common ancestor node and the root node, the kinship distance between each pair of the microbial taxonomic units is determined; The phylogenetic weight matrix, which characterizes the strength of kinship, is obtained by using a preset exponential decay kernel function to perform similarity mapping on the clan distance.

4. The method for constructing a formula microbial community according to claim 1, characterized in that, The phylogenetic weight matrix corrects the original feature importance of each microbial taxonomic unit to a smoothed feature importance, including: For any target microbial taxonomic unit among all the aforementioned microbial taxonomic units, based on the phylogenetic similarity represented in the phylogenetic weight matrix, the original feature importance of all other microbial taxonomic units among all the aforementioned microbial taxonomic units is weighted and fused to determine the neighbor feature reference value of the target microbial taxonomic unit; A first weight is assigned to the importance of the original features of the target microbial taxonomic unit using a preset smoothing parameter; The complementary parameters of the smoothing parameter are used to assign a second weight to the neighbor feature reference value; The original feature importance configured with the first weight is weighted and fused with the neighbor feature reference value configured with the second weight to obtain the modified feature importance of the target microbial taxonomic unit.

5. The method for constructing a formulated microbial community according to claim 1, characterized in that, The process of extracting high-order ecological interaction rules based on the decision paths and outputting candidate core bacterial communities includes: Extract the decision paths of each decision tree in the iterative random forest model to obtain a set of decision paths; Identify the microbial taxonomic units commonly included in multiple decision paths within the decision path set to obtain at least one candidate taxonomic unit combination; Based on the data resampling mechanism, the stability score of each candidate classification unit combination is determined; The candidate core bacterial community is determined based on all combinations of candidate taxa that meet the set threshold for stability scores.

6. The method for constructing a formulated microbial community according to claim 5, characterized in that, The step of identifying microbial taxonomic units commonly included in multiple decision paths within the decision path set, and obtaining at least one candidate taxonomic unit combination, includes: Extract candidate microbial taxonomic units that serve as splitting nodes from each of the aforementioned decision paths; Select at least one set of decision paths from the set of decision paths, and perform an intersection operation on the candidate microbial taxonomic units contained in each set of decision paths; The set containing multiple candidate microbial taxonomic units obtained by intersecting the decision paths of each group is determined as a candidate taxonomic unit combination.

7. The method for constructing a formulated microbial community according to claim 5, characterized in that, The stability score of each candidate classification unit combination is determined based on the data resampling mechanism, including: Multiple random samplings with replacement are performed on the sample set to generate multiple sample subsets; For each of the sample subsets, the steps of constructing an iterative random forest model to identify the combination of the candidate classification units are performed to obtain a set of validation combinations corresponding to each of the sample subsets; Count the total number of times any candidate classifier combination exists in all the verified combination sets; The ratio of the total number of times to the total number of samples in the subset is determined as the stability score of any candidate classification unit combination.

8. The method for constructing a formulated microbial community according to any one of claims 5-7, characterized in that, The process of determining the candidate core microbial community based on all combinations of candidate taxa that satisfy a set threshold in terms of stability scores includes: Extract all candidate classification unit combinations whose stability scores are greater than or equal to the set threshold, and use them as candidate ecological functional groups; The microbial taxonomic units contained in all the candidate ecological functional groups are aggregated and summarized, and the aggregated set of all the microbial taxonomic units is determined as the candidate core microbial community.

9. The method for constructing a formulated microbial community according to claim 1, characterized in that, The step of determining the target formulation microbial community based on the candidate core microbial community includes: The entity information of each microbial taxonomic unit in the candidate core bacterial community is input into a pre-trained biological language model for reasoning analysis to obtain the semantic score of the candidate core bacterial community. Based on each microbial taxonomic unit in the candidate core bacterial community, a microbial metabolic network model is constructed. Based on the microbial metabolic network model, community flux balance simulation is performed to determine the metabolic support index among the microbial taxa in the candidate core microbial community; the metabolic support index is used to characterize the intensity of metabolic mutualism between the microbial taxa acting as donors and the microbial taxa acting as recipients. A comprehensive recommendation index is determined based on the semantic score and the metabolic support index. The target formulation microbial community is selected from the candidate core microbial communities based on the comprehensive recommendation index.

10. The method for constructing a formulated microbial community according to claim 9, characterized in that, The step of inputting the entity information of each microbial taxonomic unit in the candidate core bacterial community into a pre-trained biological language model for reasoning analysis to obtain a semantic score for the candidate core bacterial community includes: Based on the taxonomic identifiers of each microbial taxonomic unit in the candidate core bacterial community, the corresponding standard entity names and biological metadata are obtained to construct structured prompt information; The structured prompt information is used to perform feature retrieval in a pre-built external knowledge base to obtain prior knowledge fragments related to the candidate core bacterial community; The prior knowledge fragments and the structured prompt information are fused and input into the biological language model to obtain the semantic score of the candidate core bacterial community output by the biological language model; The biological language model performs step-by-step reasoning according to the cascade biochemical reaction sequence of substrate utilization analysis, intermediate product transformation analysis, and final product effect assessment to obtain the semantic score.

11. The method for constructing a formulated microbial community according to claim 9, characterized in that, The construction of a microbial metabolic network model based on each microbial taxonomic unit in the candidate core microbial community includes: Obtain the genome sequence of each of the microbial taxa in the candidate core microbial community; Identify the genes encoding enzymes in the genome sequence, establish the mapping relationship between genes, proteins and biochemical reactions, and reconstruct the initial metabolic network model; By introducing pre-defined pathological environment constraints, metabolic gaps in the initial metabolic network model are filled to obtain the microbial metabolic network model.

12. The method for constructing a formulated microbial community according to claim 11, characterized in that, The process of introducing pre-defined pathological environmental constraints to fill metabolic gaps in the initial metabolic network model, resulting in the microbial metabolic network model, includes: Virtual culture medium boundary conditions are configured based on the intestinal microenvironment characteristics of the target disease as the pathological environment constraints. The virtual culture medium boundary conditions include the inflow threshold of specific amino acids and the initial concentration of specific nitrogen sources. Identify key housekeeping responses that are missing in the initial metabolic network model due to insufficient sequencing coverage of the genome sequence; Under the constraints of the virtual culture medium boundary conditions, the initial metabolic network model is iteratively optimized using a mixed integer linear programming algorithm to supplement the missing key housekeeping reactions into the initial metabolic network model; When it is determined that the supplemented metabolic network model can generate a preset biomass in a simulated culture medium that meets the boundary conditions of the virtual culture medium, it is identified as the microbial metabolic network model.

13. The method for constructing a formulated microbial community according to claim 9, characterized in that, Based on the microbial metabolic network model, community flux balance simulation is performed to determine the metabolic support index among the microbial taxa in the candidate core bacterial community, including: Construct a virtual community model that includes independent strain compartments corresponding to each of the aforementioned microbial taxonomic units and a shared extracellular cavity; Configure the material inflow boundary conditions for the shared extracellular chamber, and connect the microbial metabolic network model to the corresponding independent strain compartment; Under the constraints of the material inflow boundary conditions, flux balance is solved with the maximization of total community biomass as the joint objective function to obtain the optimal flux vector for each biochemical reaction. Based on the optimal flux vector, the generation rate of a specific exchange metabolite released from the donor microbial taxa unit into the shared extracellular cavity is extracted, as well as the consumption rate of the recipient microbial taxa unit absorbing the specific exchange metabolite from the shared extracellular cavity. After obtaining the minimum value between the generation rate and the consumption rate, the metabolic support index is obtained by weighting and fusing the index with the preset importance weights of the specific exchange metabolites.

14. The method for constructing a formulated microbial community according to claim 9, characterized in that, Before selecting the target formulation microbial community from the candidate core microbial communities based on the comprehensive recommendation index, the method further includes: The genome sequence of each microbial taxonomic unit in the candidate core bacterial community is analyzed to detect whether there are preset pathogenic risk gene features, including virulence factors or antibiotic resistance genes located on mobile genetic elements. During the execution of the community flux balance simulation, it is monitored whether the preset flux of harmful metabolites exceeds the safety limit value. In response to the detection of the pathogenic risk gene feature or the generation throughput exceeding the safety threshold, the corresponding candidate core microbial community is removed from the candidate pool.

15. The method for constructing a formulated microbial community according to claim 9, characterized in that, The determination of the comprehensive recommendation index based on the semantic score and the metabolic support index includes: Obtain the stability score corresponding to the candidate core bacterial community; Obtain the functional targeting score of the candidate core bacterial community for a specific pathological target in the community flux balance simulation. The stability score, the functional targeting score, the semantic score, and the metabolic support index are weighted and fused together to obtain the comprehensive recommendation index.

16. The method for constructing a formulated microbial community according to any one of claims 9-15, characterized in that, After determining the target microbial community, the optimal inoculum stoichiometry is also calculated based on the law of mass action, specifically including: Obtain the target steady-state relative abundance of each microbial taxa of the target formulation microbial community under simulated steady-state conditions; Based on the steady-state relative abundance, a set of community dynamic differential equations including interspecific interaction coefficients, intrinsic growth rates, and environmental carrying capacity is constructed. The physiological environmental parameters of the target application scenario are used as initial constraints to input the community dynamic differential equation system, and the inoculation concentration ratio of each microbial taxa that can enable the community to reach the target steady-state relative abundance after a preset colony cycle is derived in reverse as the optimal inoculation stoichiometry. Based on the optimal inoculation stoichiometry, industrial preparation formulation data for the target microbial community are generated.

17. A device for constructing a formula microbial community, characterized in that, include: The phylogenetic tree construction module is used to extract representative genomic sequences from multiple microbial taxonomic units in the sample set to construct a phylogenetic tree; The phylogenetic weight conversion module is used to calculate the evolutionary distance between each pair of microbial taxonomic units in the phylogenetic tree, and convert the evolutionary distance into a phylogenetic weight matrix, which is used to characterize the phylogenetic similarity between the microbial taxonomic units. The abundance feature processing module is used to obtain the original microbial abundance matrix of the sample set, perform a central log ratio transformation on the original microbial abundance matrix to obtain a feature matrix, and use the feature matrix as the feature dimension input to construct an iterative random forest model. The feature importance correction module is used to correct the original feature importance of each microbial taxonomic unit to a smoothed feature importance using the phylogenetic weight matrix during the feature selection iteration of the iterative random forest model, so as to update the feature sampling weights of the iterative random forest model for the microbial taxonomic unit in the next iteration based on the smoothed feature importance; The higher-order rule extraction module is used to extract higher-order ecological interaction rules based on the decision path of the iterative random forest model when the iterative random forest model reaches the preset convergence condition, and output candidate core bacterial communities. The target microbial community selection module is used to determine the target formulation microbial community based on the candidate core microbial communities.

18. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for constructing a formula microbial community as described in any one of claims 1 to 16.

19. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method for constructing a formula microbial community as described in any one of claims 1 to 16.

Images