Formula milk powder type selection method for interpretably exploring flora interaction and metabolite relationship based on neural network

Abstract

The invention relates to a formula milk powder type selection method for interpretively exploring flora interaction and metabolite relationship based on a neural network. The method comprises the following steps: establishing a comprehensive database of specific crowd queues; building a global neural network; inputting the key factor and the living habit scale into an environment factor analysis network in a secondary sub-network; main factors are extracted from the multi-cascade interaction network according to the network weight, and bacterial strains associated with a plurality of specific factors are determined; and extracting specific factors, metabolites and main strain genus, inputting the specific factors, the metabolites and the main strain genus into the secondary sub-network for classification, dividing the specific crowd into a plurality of categories, obtaining the intestinal health state of the specific crowd of each category, and carrying out targeted milk powder formula design and model selection according to the associated intestinal strain genus. According to the method, the flora interaction and metabolite relationship thereof are mined through the AI neural network, and in combination with the living environment factors, the metabolite formation process and disease risk thereof are disclosed, and personalized model selection of the milk powder formula is realized.

Description

Technical Field

[0001] The present invention relates to the field of neural networks, and in particular to a method for selecting formula milk powder based on the interpretability of neural networks to explore interactions between microbiota and the relationship between metabolites. Background Art

[0002] Diseases related to the human body are often caused by a variety of factors, including environmental factors, living habits, daily diet, etc. Since human life activities are a metabolic process, food nutrition is acquired, absorbed, and metabolized, and different nutrients have different functions. Therefore, milk powder is also one of the common choices. In addition to meeting its nutritional supply needs, it is also required to have the effect of "diet therapy", trying to improve one's health status based on one's own health cognition.

[0003] The traditional methods for selecting milk powder formula are mainly the following two methods:

[0004] 1. In vitro simulated digestion and absorption experiment: Based on the research of relevant literature, different nutritional combinations are determined, and a combination formula of multiple proteins is pre-designed and configured into corresponding gradient solutions. Protease, lipase and other substances are added to simulate the gastric fluid environment at 37°C to study the formula with the highest protein absorption efficiency. In vitro cell absorption experiments are further designed to observe the protein absorption rate on the culture dish to obtain the best formula.

[0005] 2. Animal experiments: Design a combination formula of multiple proteins in advance, feed young experimental animals for a period of time, and then compare the external indicators such as body length and weight of the experimental animals. At the same time, measure the nutrient intake in the body, and further inject foreign cells to simulate invasion, and measure their phagocytic rate to reflect immunity, so as to determine the best formula of milk powder. Summary of the invention

[0006] The embodiment of the present invention provides a formula milk powder selection method based on the interpretability of neural network to explore the interaction between microflora and the relationship between metabolites, so as to realize the personalized selection of milk powder formula.

[0007] According to an embodiment of the present invention, a method for selecting formula milk powder based on neural network interpretability to explore microbial interactions and metabolite relationships is provided, comprising the following steps:

[0008] Establish a comprehensive database of specific population cohorts;

[0009] Build a global neural network, input multi-omics data of specific population cohorts in the comprehensive database for training, obtain the characteristic relationship between microbial interactions and their metabolites, and extract key factors from them;

[0010] Input the key factors and lifestyle scale into the environmental factor analysis network in the secondary subnetwork, analyze the interaction between environmental factors and metabolites, and obtain the explanatory correlation between specific factors and metabolites, and bacterial strains and genera;

[0011] Extract the main factors in the multi-cascade interaction network according to the network weights, and identify the strain genera associated with several specific factors;

[0012] Specific factors, metabolites, and major bacterial strains are extracted and input into the secondary subnetwork for classification. The specific population is divided into several categories, and the intestinal health status of the specific population in each category is obtained. Targeted milk powder formula design and selection are carried out according to the associated intestinal bacterial strains.

[0013] Furthermore, in the establishment of a comprehensive database for specific population cohorts, user samples are collected and basic information is collected for population objects that meet the conditional combinations. By performing metagenomic sequencing and mass spectrometry analysis on the samples, multi-omics data on the corresponding bacterial species, genera, genes, and metabolism are obtained.

[0014] Furthermore, a Venn diagram intersection comparison was performed on the key data screened out from the association-specific population cohort to confirm that the association screening results were valid.

[0015] Furthermore, the global neural network is designed to train the omics data related to specific populations, including metagenomic data and metabolic data, to obtain their corresponding interaction network probability relationships; wherein the global neural network includes but is not limited to LSTM, RNN, MLP, CNN, RF, and XG commonly used algorithm components.

[0016] Furthermore, the environmental analysis network of the secondary subnetwork is used to analyze the relationship between environmental factors and metabolites, thereby obtaining the multi-level interaction network association relationship of environmental factors, which is then associated with metabolites and further associated with the corresponding bacterial strains and genera.

[0017] Furthermore, the classification network of the secondary subnetwork automatically divides the main factors into several types based on the extracted main factors, including user information, metabolites, and strain genera, and analyzes the data characteristics of specific strain genera in each type to guide the selection of nutrients in milk powder formulas and regulate intestinal health status in a targeted manner. The classification calculation methods include but are not limited to manual grouping, automatic classification algorithms, and automatic clustering algorithms.

[0018] Furthermore, the comprehensive database is divided into four layers. The first layer is user information, living habits and environment, and eating habits. The second layer is macro-gene data. By collecting user fecal samples and obtaining intestinal-related bacterial genes, bacterial abundance, bacterial genus abundance and other data reflecting the user's intestinal health status through sequencing experiments, the third layer is metabolomics data, which further obtains metabolic substance composition and gene expression metabolic data through mass spectrometry analysis; the fourth layer is multi-layer interaction network data. Through specific factors, the causal process of related metabolites and strains and genera is obtained from the interaction network diagram.

[0019] Furthermore, the neural network framework is a logical container, which includes a global neural network and a secondary sub-network. It receives cohort subject data, metagenetic data, and metabolic data. The cohort subject data and metagenetic data use themselves as input and each other as output to calculate the interaction relationship with each other, and explore the multi-level interaction relationship of specific populations.

[0020] Furthermore, the global neural network uses a nonlinear neural network algorithm to fit the interaction relationship between metabolites and intestinal flora, performs bidirectional training on metabolites and intestinal flora, predicts intestinal flora with metabolites as input in the forward direction, and predicts metabolites with intestinal flora as input in the reverse direction. After multiple calculations, the weights are extracted to obtain the interaction network relationship between mutual metabolites and intestinal flora.

[0021] Furthermore, according to the network relationship weights, any metabolite or environmental factor can obtain an accurate explanation of the action process of the whole process, its degree of action and expression level. By drawing a multi-level interaction network diagram and identifying key paths, the causal relationship of the causes of related factors can be explained.

[0022] The formula milk powder selection method based on the interpretable mining of microbial interactions and metabolite relationships by neural networks in the embodiments of the present invention mines microbial interactions and their metabolite relationships through AI neural networks, combines their living environment factors, reveals the metabolite formation process and their disease risks, and realizes personalized selection of milk powder formula. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0024] Figure 1 Customized milk powder formula based on the interaction between bacterial species and their metabolism;

[0025] Figure 2 Screenshots of 5% of the basic information of 712 users;

[0026] Figure 35% screenshot for metagenomic sequencing experiment data;

[0027] Figure 4 5% screenshot for metabolic mass spectrometry analysis experimental data;

[0028] Figure 5 It is a four-layer data structure diagram of the specific cohort;

[0029] Figure 6 It is a calculation scheme diagram of milk powder formula selection based on nonlinear multi-level interaction network relationship;

[0030] Figure 7 A comparison chart of metabolic characteristics of interactions between different pairs of global neural networks;

[0031] Figure 8 10% screenshot for the validation data of Bifidobacterium longum;

[0032] Fig. 9 This is a diagram of the screening effect based on the validation model of Bifidobacterium longum;

[0033] Fig.10 10% screenshot for the validation data of Bifidobacterium longum;

[0034] Fig.11 This is a diagram of the screening effect based on the validation model of Bifidobacterium longum;

[0035] Fig.12 Output the relationship pair graph with mutual weights;

[0036] Fig.13 It is a multi-level action network diagram of a single factor;

[0037] Fig.14 It is a multi-factor multi-level interaction network diagram;

[0038] Fig.15 This is an explanatory diagram of the living environment factors;

[0039] Fig.16 Provide an explanatory diagram of the multi-level causal process of dietary factors;

[0040] Fig.17 It is the overall interaction network and the coverage diagram of the main factors;

[0041] Fig.18 Comparison diagram of the main factor interaction network diagrams of different groups;

[0042] Fig.19 This is a comparison chart of the normal distribution of the theoretical population;

[0043] Fig. 20 This is the P value verification chart of the normal distribution of the theoretical population;

[0044] Fig.21Design flow charts for application examples;

[0045] Fig. 22 406 sample images were collected for the application example;

[0046] Fig.23 This is a comparative analysis chart of feeding methods;

[0047] Fig.24 This is a comparative analysis chart of birth modes;

[0048] Fig.25 It is an interpretable multi-omics network interaction feature map;

[0049] Fig.26 A comparison diagram of the interpretable multi-omics network features of feeding methods;

[0050] Fig. 27 Process explanatory diagram for multi-omics signature factors of breastfeeding;

[0051] Fig.28 Extract the main health factor graphs for the feature graphs based on different combinations of limiting domain condition values. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0053] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0054] The technical problem solved by the present invention is to mine the interaction between flora and their metabolites through AI neural network for the health status of human intestinal microecology, and to reveal the metabolite formation process and disease risk in combination with living environment factors, so as to realize the personalized selection of milk powder formula (such as Figure 1 ).

[0055] This method has open-ended limited domain conditions, including open conditions such as age, gender, region, altitude, and medical history. Combined with the target population in the market information, several limited domain conditions and their conditional values ​​are proposed. Multiple limited domain conditions are combined into a specificity, and the target population that meets these limited domain conditions is described as a specific population. For example, if the design theme is "chronic disease patients", the limited domain conditions are "three highs", "diabetes", "excessive BMI", etc.

[0056] In this method embodiment, the limited domain conditions proposed are "carrying 14 diseases", "0-30 years old, grouped into infants, children, teenagers, young people, and middle-aged and young people", and the causes of their metabolites are learned by studying the living habits, daily routines, and dietary structures of specific populations, and further finding key bacterial strains and genera, and classifying specific populations, comprehensively evaluating the intestinal health status of each classification, and then specifically determining several milk powder formulas, and selecting key nutrients in the formulas, to achieve "diet therapy" intervention for human health. The process of this method embodiment includes the following steps:

[0057] A comprehensive database of specific population queues was established, and basic information of 712 users who met the specific conditions was collected, including name, gender, height, age, weight, daily routine, and eating habits. The users were grouped by age 1-6, 7-12, 13-18, 19-24, and 25-30 (e.g. Figure 2 ), collect stool samples from users, and perform metagenomic sequencing (such as Figure 3 ) and metabolic mass spectrometry experiments (e.g. Figure 4 ) to obtain multi-omics data including bacterial abundance, bacterial genus abundance, bacterial species genes, and metabolite data.

[0058] Establish a data model N-Tier. This embodiment sets a 4-Tier, where L1 is the cohort subject data, collected user basic information, including user basic information, living habits Q&A, dietary structure, etc., and use the authoritative literature PubMed in the field of biomedicine and health to annotate the relevant information of specific populations, such as smoking and drinking, irregular work and rest, dietary characteristics, etc. in living habits, and annotate the associated health status information such as diseases and risks. L2 is the metagenomic colony abundance data collected by the cohort, and the strain species and genus are annotated using the microbial metabolite database MimeDB to obtain the names of the strains and genus. L3 is the metabolite abundance data collected by the cohort, which is annotated using the chemical database PubChem and the biologically related chemical entity database ChEBI to obtain the metabolite identifier InChIKey. L4 is a multi-level interaction network relationship based on L1 / L2 and L2 / L3, and further selects different characteristics for specific populations for classification, and designs related functions or applications for the corresponding classifications, such as the milk powder formula in the embodiment of this method (such as Figure 5 ).

[0059] Based on the above data model, an overall calculation scheme for milk powder formula selection based on a multi-level interaction network relationship of a nonlinear model is designed (such as Figure 6 ).

[0060] Among them, S10 is a specific population cohort after the limited domain conditions are selected. The samples of the S10 cohort are analyzed by mass spectrometry in S101 to obtain L3 metabolic data S102, and then annotated using PubChem / ChEBI in S103 to obtain the InChiKey of the metabolite S104. The metabolomics data dimension itself is very large. The commonly used machine learning algorithm Random Forest is used to perform feature mining on the data in S105, and then the linear Pearson algorithm is used to obtain the internal correlation of the metabolite components, extract its weight and filter the data, only retaining the metabolites that can be annotated by HMDB, removing other metabolites, and then using them as the input of S301 in the global network.

[0061] At the same time, the samples of the S10 queue were sequenced through S201 metagenomic sequencing to obtain L2 metagenomic data S202, which were annotated using metaphla4 in S203 to obtain S204 bacterial abundance, including strains, genera and their expression levels. The metagenomic data dimension is also very large. In S205, the machine learning algorithm random forest was used to perform feature mining on the data, and then the linear Pearson algorithm was used to obtain the internal correlation between strains and genera, extract their weights and filter the data, remove non-zero values ​​less than 25%, remove genera with an average abundance of less than 0.1%, retain the relevant main factors, and then use them as the input of S302 in the global network.

[0062] The neural network framework S30 is a logical container, which includes a global neural network S300 and a secondary sub-network S600. It receives L1 cohort subject data, L2 macrogene data, and L3 metabolic data. L1 and L2 use themselves as input and each other as output to calculate the interaction relationship with each other, and then explore the multi-level interaction relationship of specific populations.

[0063] The global neural network S300 uses a nonlinear neural network algorithm to fit the interaction between metabolite S301 and intestinal flora S302. Unlike the traditional linear Pearson method, the nonlinear neural network algorithm can more accurately mine the correlation between them. Here, the global neural network performs bidirectional training on S301 and S302, predicting S302 with S301 as input in the forward direction and predicting S301 with S302 as input in the reverse direction. After multiple calculations, the weights are extracted to obtain the interaction network relationship between S301 and S302.

[0064] To further explain the global neural S300 design, S300 is a logical component container that integrates different algorithm components. It can load time series-related algorithm components as well as traditional feature representation algorithm components. The standard neural network types all belong to standard algorithm components, such as LSTM, RandomForest, XGBoost, CNN, MLP, etc., which are briefly described as LM, RF, XG, CNN, and MLP.

[0065] In terms of the ability to characterize metabolic features, experimental data from public literature were used to compare the ability of different algorithms to characterize the interaction relationship between metabolites. Based on the indicators in the binary classification evaluation model, including F1 score, precision, and recall, the performance of several algorithm components was compared, including RF, XG, MLP-model1, MLP-model2, RF / XG combination, xSwiGLU / ReTa combination, etc. (e.g. Figure 7 ), select the models RF and XG with the smallest residual as the algorithm components for metabolic characterization in the embodiment of this method.

[0066] The global neural network S300 is an algorithm combination selected in this embodiment, wherein the timeline causal network S303 uses LM, and the traditional feature network S304 uses RF and XG. S301 and S302 are respectively input into LM, RF, and XG to obtain LM, RF, and XG outputs, and the weight matrix of the hidden layer is extracted to obtain the interaction network relationship I between the metabolites and the bacterial strains. lm ,I rf ,I xg .

[0067] For interactive network relationships I lm,I rf ,I xg , calculate the weight of each factor and sort them according to S301 and S302 respectively. On the S301 side, extract the associated strain genus and its weight coefficient according to the interaction network relationship, and take the LM subnet Sort the LM outputs, where coef i is the weight corresponding to metabolite i. Similarly, for the output of RF and XG subnets, we take The output of the RF or XG subnetwork is sorted, and then the Top M main metabolites are extracted, thus in the interaction network relationship I lm ,I rf ,I xg Get three outputs R of metabolites lm , R rf , R xg The same sorting operation was applied on the S302 side to extract the top N main strain genera. lm ,I rf ,I xg The three outputs R′ of the strain genus are obtained lm , R′ rf , R′ xg .

[0068] The top N factors calculated by the global network S300 are summarized to obtain the comprehensive result S400, where the calculation process of the comprehensive result is to take the intersection of LM, RF, and XG, including the metabolite R lm ∩R rf ∩R xg and strain group R′ lm ∩R′ rf ∩R′ xg , using the intersection result R of metabolites lm ∩R rf ∩R xg From the interaction network relationship I lm ,I rf ,I xg Obtain the set S' of related strains, and use the intersection result R' of strain genera lm ∩R′ rf ∩R′ xg From the interaction network relationship I lm ,I rf ,I xg Obtain a set S of associated metabolites, thereby obtaining the screened metabolites (R lm ∩R rf ∩R xg )∪S and strain population (R′ lm ∩R ′ rf∩R ′ xg )∪S′, further in the interaction network relationship between metabolites and strains and genus I xg Extract the correlation between the main factors And according to the Top N factors, the input of metabolites and the input of bacterial strains are obtained from S301 and S302 respectively to form a comprehensive result output S400 of the global network.

[0069] In this embodiment, Bifidobacterium is used as an example for explanation. S400 takes the top 100 strains, including strains and metabolites, and uses MiMeDB to associate and annotate the strains. 489 items are obtained based on the strain species name matching, and 826 items are obtained by using the metabolite InChIKey matching of the metabolites. The data is further deduplicated to obtain 385 annotated metabolites and 194 strain genus items. Because MiMeDB is a public authoritative data set, the range that can be annotated is known knowledge. The feature range obtained by LM, RF, and XG learning may exceed the annotation range of MiMeDB and belong to currently unknown knowledge. The intersection of the MiMeDB annotation range and the results of model learning is also used to confirm that the results of model learning are valid.

[0070] In this embodiment, taking s__Bifidobacterium_longum as an example, MiMeDB associates 113 metabolites (such as Figure 8 ), the top 100 results of LM, RF, and XG output by the global network were taken for intersection analysis, among which the intersection results of LM, RF, and XG were 11 items, the intersection of MimeDB and RF was 26 items, the intersection of MimeDB and XG was 25 items, and the intersection of MimeDB and LM was 18 items (such as Fig. 9 ), proving that the model is effective in screening Bifidobacterium longum.

[0071] Example Taking s__Bifidobacterium_breve as an example, MiMeDB associates 100 metabolites (such as Fig.10 ), the top 100 results of LM, RF, and XG output by the global network were taken for intersection analysis, among which the intersection results of LM, RF, and XG were 15 items, the intersection of MimeDB and RF was 24 items, the intersection of MimeDB and XG was 27 items, and the intersection of MimeDB and LM was 14 items (such as Fig.11 ), proving that the model is effective in screening Bifidobacterium breve.

[0072] Then the comprehensive result S400 is input to the input end S602 of the second-layer sub-network. The second-layer sub-network S600 is also a logical component container that can integrate different neural network algorithm components. This method embodiment inherits the environmental factor analysis network S603 and the classification sub-network S604.

[0073] In S501, the L1 queue topic data, in this embodiment, includes user basic information and lifestyle Q&A. In S502, PubMeb is used for monitoring and annotation to obtain the user's health status evaluation, including disease-related bad lifestyle habits and disease risks, such as smoking, drinking, etc., which will cause changes in related intestinal flora and possible related diseases, and then the data is input into the secondary subnetwork S601.

[0074] In the second-layer analysis network, the environmental factor analysis subnetwork S603 first uses S601 including living habits, food, environmental factors and other scales as input, and uses S602 bacterial abundance as output. Then, S602 is used as input and S601 as output. After two-way training, the interaction network relationship between living habits and bacterial strains is obtained. env .

[0075] In the second-layer analysis network, the classification subnetwork S604 aggregates the input of S601 and the input of S602. There are four grouping methods. The first method uses a neural network algorithm to implement sample grouping. The second method uses a clustering algorithm to form several natural categories according to the intrinsic characteristics of the data. Due to the large number of categories, the adjacent categories are merged into several groups according to the Euclidean distance. The third method uses a classification algorithm to specify the number of categories. By continuously adjusting the position of the center point of each category in the multidimensional space of the data, each data point is closest to the center point of its own category and the Euclidean distance between the center points of each category is the largest. The fourth method specifies the grouping label when the sample is collected. The above four methods can be replaced in the implementation path according to the specific research objectives. In this embodiment, the fourth method is used, using manually specified sample grouping, that is, in the 5 age-end groups, the correlation relationship between the bacterial strains, genera, metabolites, and living environment factors associated with 14 diseases is comprehensively observed, and each group is analyzed to see whether it has obvious characteristics, thereby determining different milk powder formulas.

[0076] Based on the above calculation results, the weight relationships among environmental factors, metabolites, and strains were summarized, and the interaction network between metabolites and strains was constructed using I' xg In this embodiment, the best XG network weights for metabolic characterization are used to obtain the mutually weighted output S700 as I env ∪I′ xg (like Fig.12). In this embodiment, the relationship between strain genes and environmental factors was calculated to be 2073600 pairs, the relationship between bacterial genus and environmental factors was 398161 pairs, the relationship between bacterial genus, metabolites, and environmental factors was 36216324 pairs, and the relationship between strain genes, metabolites, and environmental factors was 27133681 pairs.

[0077] Based on S700, pairs of different omics dimensions are exhaustively enumerated, their correlations are calculated, Top N′ is set, and its weight W is extracted. n , extract multi-level interaction network relationship weights S800, set indicators Used in the exhaustive process Determine the pairs Whether it is within the weight range of Top N, and the weight of pairs is extracted from S700 to obtain the multi-level interaction network relationship weight S800.

[0078] According to the network relationship weight S800, any metabolite or environmental factor can obtain an accurate explanation of the action process of the whole process, its degree of action and expression level. By drawing a multi-level interaction network diagram and identifying key paths, the cause of the relevant factors can be explained by the causal relationship S900.

[0079] Explanatory S900 can be represented by a multi-level interaction network diagram, using line colors to distinguish different levels and line thickness to distinguish different weights according to S800. Starting from a single environmental factor, it is possible to associate metabolites, strains, and bacterial genera to draw an interaction network (e.g. Fig.13 ), when multiple environmental factors interact, the metabolites and bacterial strains are related to draw a multi-factor interaction network diagram (such as Fig.14 ). For example, taking a specific living environment factor as an example, such as the effect of smoking on metabolism and intestinal flora, the correlation and explanation of the main metabolites and the genus status of intestinal strains can be obtained from the interaction network relationship diagram (e.g. Fig.15 ), further focusing on the explanatory power of causal processes, such as meat in dietary habits, Sangkey diagrams were drawn based on the relevant main metabolites and intestinal bacterial strains extracted by S800, and the multi-level causal process explanatory power of dietary factors (such as Fig.16 ).

[0080] In the present method embodiment, the top 30000 are selected, and a multi-level interaction network diagram is drawn to obtain an intuitive overall interaction network. Then, the metabolites are gradually selected according to the weights {Top n, n=1, 2…n} and drawn in the same way. The points and lines involving the associated life factors, metabolites, and strain flora are drawn in red to observe whether the environmental factors and strains associated with the circled metabolites are the main factors, thereby determining that the interaction network relationship range associated with the metabolites marked with n=17 occupies the main part of the overall network (such as Fig.17 ).

[0081] Based on the 17 major metabolites discovered above, the multi-level main influencing factors of the specific population cohort were further clarified. By drawing corresponding multi-level interaction network relationship diagrams for different age groups, including daily routines, eating habits, gender, age, metabolites, strains and flora, it was further found that the network diagrams of different age groups had significantly different characteristics (such as Fig.18 ). After clarifying that different age groups in the selected specific population cohorts involving 14 diseases have different intestinal flora characteristics, this embodiment designs several milk powder formulas by evaluating the health status of the intestinal flora involved in different groups, selects nutrient molecules that have a targeted intervention effect on intestinal bacterial strains and genera, and improves the health status of users by regulating the health status of the intestinal flora.

[0082] Perform a P value check S1000 for the explanatory power of health causes to confirm whether the explanatory power is credible. Use a random method to shuffle the index data extracted in S900 to generate theoretical population data that conforms to the normal distribution for each factor index. Statistic the mean square error, mean value, and CV value of each factor index to compare whether the S900 sample data conforms to the theoretical distribution. At the same time, judge whether the model-generated screening result S900 is credible according to the confidence interval ≤ 0.05. This embodiment uses 200 randomizations and 10,000 randomizations respectively, with the x-axis being the theoretical value and the y-axis being the ranking value of the Top 100 indexes. Compare the probability graphs of the S900 sample data and the theoretical population data (such as Fig.19 ), the results show that the selected S900 samples are in accordance with the normal distribution and are relatively consistent in the middle section. Through the standard deviation std, mean mean, and coefficient of variation CV, it can be known that most of the S900 samples are within the confidence interval of ≤0.05 (such as Fig. 20 ). Therefore, it is confirmed that the nonlinear model of this method is consistent with the real world, and the above calculation process is integrated to obtain the nonlinear model S1200.

[0083] In a specific application, stool samples are collected from consumers S20 for mass spectrometry analysis S1301 and macrogene sequencing S1302, and corresponding cohort subject data are collected, such as the basic information and lifestyle questionnaire Q&A used in this embodiment, and predictive analysis is performed on user data S1400, and the user's multi-level interaction network characteristics and cohort similarity scores are calculated S1500, so as to select milk powder formula based on molecules according to the user's intestinal flora characteristics S1600.

[0084] Based on the above method, a verification example was designed. The subject queue T1001 was set as "infant formula", and the limited domain condition T1002 included "feeding method, birth method". Each limited domain condition was used as a control group through the condition value, and its genetic, metabolic, and microbial interaction characteristics were compared and analyzed. Several feature classifications were mined through this method. Each classification represented a health condition. The health status in this case was identified through health annotations, and the nutritional formula was determined to regulate the intestinal "microbial flora" and promote infant health (such as Fig.21 ).

[0085] Fecal samples from 406 infants were collected, and cohort T1003 registered their feeding method, gender, birth method, birth day, and other information (such as Fig. 22 ), where BF was breast-fed, FF was formula-fed, and birth methods included cesarean section and natural birth. Metagenomic sequencing and mass spectrometry metabolic analysis were performed on each sample, and T1004 used this method to calculate the multi-omics multi-level interaction network features.

[0086] Feeding methods include breastfeeding and formula feeding. The top 20 of bacterial genera, bacterial species and metabolites were extracted and analyzed. Through comparison, it can be found that, for example, the higher the abundance of Shannon species, the greater the probability of formula feeding, and vice versa, the lower the abundance, the greater the probability of breastfeeding. In terms of bacterial species genes, for example, the higher the abundance of Clostridium B, the greater the probability of formula feeding, and vice versa, the lower the abundance, the greater the probability of breastfeeding. By extracting metabolites based on metabolic pathways through the top 20 genes, it can also be found that there are obvious differences between the two groups (such as Fig.23 ).

[0087] Feeding methods included cesarean section and natural birth. The top 20 were extracted and analyzed for bacterial genera, bacterial species, and metabolites. For example, the higher the abundance of the bacterial species Phocaeicola merdigallinarumShannon, the greater the probability of natural birth, and vice versa, the lower the abundance, the greater the probability of cesarean section. Another example is similar to the bacterial species Bacteroides intestinigallinarum. The higher the abundance, the greater the probability of natural birth, and vice versa, the lower the abundance, the greater the probability of cesarean section. The higher the abundance of bacterial species genes such as Ruminiclostridium, the greater the probability of natural birth, and vice versa, the lower the abundance, the greater the probability of cesarean section. By extracting metabolites based on metabolic pathways through the top 20 genes, it can also be found that there are obvious differences between the two groups (such as Fig.24 ).

[0088] In T1005, this method was used to further generate interpretable multi-omics network interaction diagrams (such as Fig.25 ), corresponding multi-omics network interaction graphs can also be formed based on different limiting domain condition values ​​(such as Fig.26 ), compared with breastfeeding and formula feeding, the corresponding network interaction feature graphs have obvious differences. The interpretable multi-omics network interaction graph is used to obtain the interpretability of the specified limited domain condition value, such as the bacterial species, bacterial species and their metabolic processes affected by breastfeeding (such as Fig. 27 ).

[0089] Based on the combination of the limited domain condition values, four network feature graphs are formed, and the main health factors are extracted, and health annotation and evaluation are performed on them (such as Fig.28 ). If there are too many combined categories, you can specify the number of categories, use the classification algorithm to classify all samples, and then summarize the sample features of each class, extract their common features, and then perform health labeling and evaluation. After understanding the characteristics of each class, several standard model libraries T1006 are formed, such as A, B, C, etc., and milk powder formula is formulated based on health labeling and evaluation to improve infant intestinal flora through nutrient absorption. Here, 4 limited domain condition value combinations are used.

[0090] Subsequently, any infant sample is sampled T1007, and after metagenomic sequencing and mass spectrometry metabolic analysis using this method, the multi-omics multi-level interaction network characteristics of the sample are calculated T1008, and then based on the standard model library T1006, the distance between networks is calculated T1009 to identify whether the sample matches a specific standard template, such as A.

[0091] In T1010, according to the multi-omics multi-level interaction network characteristics of the sample, based on the health factors of the main factors and the matched limited domain condition values, an explanatory report on the infant's intestinal health is provided, and milk formula A is selected according to the infant's specific individual health status.

[0092] Compared with the prior art, the present invention has the following advantages and positive effects:

[0093] 1. The innovation of this program is to use AI neural networks to analyze the causes of metabolites related to specific factors. By sampling, experimenting and analyzing the metabolites of the target population, combined with external living environment factors such as dietary structure and living habits, we can deeply explore the health status and possible risk factors of the target population, and then customize the milk powder formula in a targeted manner to be closer to the individual's actual nutritional needs.

[0094] 2. The innovation of this program is to deeply discover the intestinal bacterial strains by analyzing the metabolites of specific populations, and then comprehensively evaluate the health status to determine the formula selection. Compared with the positive selection method of traditional milk powder formula, this method aims to improve immunity or the high absorbability of specific nutrients, and is based on the milk powder formula obtained by designed animal experiments. It has a more targeted and personalized health improvement effect.

[0095] The invention points of this application are as follows:

[0096] 1. The innovation of this program is the interpretability of AI neural networks. For specific population cohorts, metabolic samples are collected for experiments to obtain corresponding multi-omics data such as bacterial flora, genes, and metabolism. AI neural networks are used for training to obtain a network diagram of specific factor interactions, thereby understanding the key processes related to specific factors, and further using secondary sub-networks to analyze the interactions between environmental factors and metabolites, ultimately obtaining the interpretability of metabolites.

[0097] 2. The innovation of this program is the multi-level specific factor interaction network. Through the interaction network diagram, the related metabolites and their intestinal bacterial strains can be clearly found for specific factors, and the main key factors can be extracted for classification. The intestinal health status and causal process of these categories can be comprehensively evaluated, and then the milk powder formula can be customized in a targeted manner, and the nutrients in the milk powder formula can be used to effectively intervene in the health of specific populations.

[0098] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0099] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0100] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the system embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0101] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0102] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0103] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it 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 all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0104] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships, characterized in that: The following steps are involved: Establish a comprehensive database of specific population cohorts; Build a global neural network, input multi-omics data of specific population cohorts in the comprehensive database for training, obtain the characteristic relationship between microbial interactions and their metabolites, and extract key factors from them; Input the key factors and lifestyle scale into the environmental factor analysis network in the secondary subnetwork, analyze the interaction between environmental factors and metabolites, and obtain the explanatory correlation between specific factors and metabolites, and bacterial strains and genera; Extract the main factors in the multi-cascade interaction network according to the network weights, and identify the strain genera associated with several specific factors; Specific factors, metabolites, and major bacterial strains are extracted and input into the secondary subnetwork for classification. The specific population is divided into several categories, and the intestinal health status of the specific population in each category is obtained. Targeted milk powder formula design and selection are carried out according to the associated intestinal bacterial strains.

2. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 1, characterized in that: In the comprehensive database of specific population cohorts, user samples are collected and basic information is collected for the population subjects that meet the conditional combinations. By performing metagenomic sequencing and mass spectrometry analysis on the samples, multi-omics data of corresponding bacterial species, genera, genes, and metabolism are obtained.

3. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 1, characterized in that: A Venn diagram intersection comparison was performed on the key data screened out from the association-specific population cohort to confirm that the association screening results were valid.

4. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 1, characterized in that: The global neural network is designed to train the omics data related to specific populations, including metagenomic data and metabolic data, to obtain the corresponding interaction network probability relationship; The global neural network includes but is not limited to LSTM, RNN, MLP, CNN, RF, and XG common algorithm components.

5. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 1, characterized in that: The environmental analysis network of the secondary subnetwork is used to analyze the relationship between environmental factors and metabolites, thereby obtaining the multi-level interaction network association relationship of environmental factors, which is then associated with metabolites and further associated with the corresponding bacterial strains and genera.

6. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 1, characterized in that: The classification network of the secondary subnetwork automatically divides the main factors into several types based on the extracted main factors, including user information, metabolites, and strain genera, and analyzes the data characteristics of specific strain genera in each type to guide the selection of nutrients in milk powder formulas and regulate intestinal health status in a targeted manner. The classification calculation methods include but are not limited to manual grouping, automatic classification algorithms, and automatic clustering algorithms.

7. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 1, characterized in that: The comprehensive database is divided into four layers. The first layer is user information, living habits and environment, and eating habits. The second layer is metagenetic data. By collecting user fecal samples and obtaining intestinal-related bacterial genes, bacterial abundance, and bacterial genus abundance through sequencing experiments, the data reflecting the user's intestinal health status are obtained. The third layer is metabolomics data, which further obtains metabolic substance composition and gene expression metabolic data through mass spectrometry analysis. The fourth layer is multi-layer interaction network data. Through specific factors, the causal process of related metabolites and strains and genera is obtained from the interaction network diagram.

8. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 1, characterized in that: The neural network framework is a logical container, which includes a global neural network and secondary sub-networks. It receives cohort subject data, metagenetic data, and metabolic data. The cohort subject data and metagenetic data use themselves as input and each other as output to calculate the interaction relationship with each other, and explore the multi-level interaction relationship of specific populations.

9. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 8, characterized in that: The global neural network uses a nonlinear neural network algorithm to fit the interaction between metabolites and intestinal flora, performs bidirectional training on metabolites and intestinal flora, predicts intestinal flora with metabolites as input in the forward direction, and predicts metabolites with intestinal flora as input in the reverse direction. After multiple calculations, the weights are extracted to obtain the interaction network relationship between mutual metabolites and intestinal flora.

10. The method for selecting formula milk powder based on neural network interpretability to explore bacterial flora interactions and metabolite relationships according to claim 1, characterized in that: According to the network relationship weights, any metabolite or environmental factor can obtain an accurate explanation of the action process of the whole process, its degree of action and expression level. By drawing a multi-level interaction network diagram and identifying key paths, the causal relationship of the causes of related factors can be explained.

Images