Neural network-based maternal transmission of key bacteria prediction and screening method and system

Abstract

The application belongs to the technical field of computers and discloses a sow-source transmission key bacteria prediction and screening method based on a neural network, which comprises the following steps: step 1: obtaining sow and offspring microbial data, forming a data set, and dividing the data set into a training set and a test set; step 2: constructing a neural network model based on an attention mechanism, training and adjusting the neural network model by using the training set and the test set, and obtaining a neural network model for predicting offspring microbial data according to input sow microbial data; and step 3: performing a feature ablation experiment on the neural network model trained in step 2 by removing the genera of the sow microbial data in the data set one by one, and obtaining key bacteria that have a significant impact on sow-source transmission. The method uses an advanced deep learning algorithm in the field of artificial intelligence, analyzes sow and offspring microbial sequencing data, and constructs a neural network model capable of predicting key sow-source bacteria. The model can not only extract genera that have a key impact on growth performance from a large-scale microbial community, but also can screen strains that affect key phenotypes through directional culture groups of predicted bacteria. Meanwhile, the application also provides a system based on the method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a mother-source transmission key bacteria prediction and screening method based on neural network. BACKGROUND

[0002] In recent years, the research on the correlation between animal microbiome and host health and growth performance has attracted more and more attention, especially in the livestock industry, the composition and dynamic change of microbial community have a profound impact on the growth and development, immune regulation, feed utilization rate and other aspects of animals. The microbial community colonization in the early growth stage is largely influenced by the mother, and this mother-source transmitted flora plays a crucial role in the maturation of the intestinal microenvironment, nutrient absorption and immune system of the offspring.

[0003] Traditional research methods mainly evaluate the influence of mother-source microorganisms on the growth of offspring through observation and experimental means, usually relying on macroscopic biological statistical analysis, focusing on the simple correlation between the richness, diversity of microbial population and the growth performance of the host. This kind of method often ignores the complex interaction of microbial community, and it is difficult to accurately identify the specific key bacteria, and it is also difficult to effectively predict how these key bacteria affect the growth and development of offspring through mother-source transmission. When facing large-scale high-dimensional microbiome data, traditional research methods are easily disturbed by noise, and cannot fully utilize the potential information in the data, resulting in unstable prediction results and low accuracy. In addition, the traditional screening of key bacteria has certain blindness, and the present method provides a direction for the screening of mother-source key bacteria.

[0004] With the development of high-throughput sequencing technology, research methods that can analyze the composition and function of microbial community at the molecular level have gradually become the mainstream. However, how to extract the flora closely related to the growth performance of the host from the vast microbiome data and further predict the mother-source transmission mechanism of these flora is still a difficult problem that has not been fully solved. In addition, although existing researches have shown that mother-source microorganisms are crucial to the early growth of piglets, and the type of piglet intestinal flora will affect its feed utilization rate, there is still a significant technical gap in the prediction and verification of specific key bacteria.

[0005] The technical problem to be solved by the present application is how to improve the prediction and screening efficiency of mother-source transmission key bacteria. SUMMARY

[0006] One of the purposes of the present application is to provide a neural network-based maternal transmission key bacteria prediction and screening method, which utilizes advanced deep learning algorithms in the field of artificial intelligence, constructs a neural network model capable of predicting maternal key bacteria by analyzing maternal and child microbial sequencing data. This model not only can extract bacteria genera that have a key impact on growth performance from large-scale microbial communities, but also can screen strains that affect key phenotypes through directional culture groups of predicted bacteria.

[0007] Meanwhile, the present application also provides a system based on the method.

[0008] To achieve the above-mentioned purpose, the present application provides a neural network-based maternal transmission key bacteria prediction and screening method, comprising the following steps:

[0009] Step 1: Obtain maternal and child microbial data to form a data set, and divide the data set into a training set and a test set; the maternal and child microbial data includes sow microbial data and piglet microbial data; the sow microbial data and piglet microbial data are genus-level feature information of sows and piglets;

[0010] Step 2: Construct a neural network model based on an attention mechanism, train and adjust the neural network model using the training set and the test set, and obtain a neural network model for predicting piglet microbial data based on input sow microbial data;

[0011] Step 3: Perform feature ablation experiments on the neural network model trained in step 2 by removing the genera of the sow microbial data in the data set one by one, and obtain key bacteria that have a significant impact on maternal transmission.

[0012] In the above-mentioned neural network-based maternal transmission key bacteria prediction and screening method, step 1 is specifically:

[0013] Step 1: Obtain maternal and child microbial data;

[0014] Step 2: Preprocess the maternal and child microbial data, which includes removing missing values and performing zero substitution processing;

[0015] Step 3: Perform center log ratio conversion and time point padding on the data set to generate the data set.

[0016] In the above-mentioned neural network-based maternal transmission key bacteria prediction and screening method, the feature information is one or a combination of more of microbial community composition, genus abundance, and functional annotation.

[0017] In the above-mentioned neural network-based maternal transmission key bacteria prediction and screening method, the neural network is composed of an encoder, a decoder, and an attention mechanism.

[0018] The encoder adopts a long short-term memory network to process the sow microbial data, extract time sequence features, and convert them into a hidden state vector.

[0019] The decoder adopts a long short-term memory network to extract information related to the piglet microbiome from the hidden state vector to form a prediction result.

[0020] The attention mechanism is used to calculate the attention weight α t to highlight the contribution of microbial characteristics at each different time point to the prediction result.

[0021]

[0022] where W a and b a are learnable parameters, is the hidden state vector at time point t.

[0023] In the above neural network-based sow-derived key bacteria prediction and screening method, the neural network model is optimized using a loss function during the training process.

[0024] In the above neural network-based sow-derived key bacteria prediction and screening method, the loss function is a combination of Euclidean distance loss, mean square error algorithm, and L1 regularization term.

[0025] The formula of the loss function is:

[0026] L = α × EuclideanLoss + β × MSE + λ × L1Regularization

[0027] Where: EuclideanLoss is the Euclidean distance loss algorithm, which is used to measure the root mean square difference between the predicted piglet microbiome composition and the actual value.

[0028] MSE is the mean square error algorithm, which is used to calculate the sum of the mean square errors at each time point t.

[0029] L1Regularization is the L1 regularization term, which is used to encourage the sparsity of model parameters to prevent overfitting.

[0030] α, β, λ are constants.

[0031] In the above neural network-based sow-derived key bacteria prediction and screening method, the step 3 is specifically:

[0032] Step 31: Remove each genus of sow microbial data in the data set to obtain multiple new sow microbial data.

[0033] Step 32: Use the new sow microbial data and a neural network model to predict the new piglet microbial data;

[0034] Step 33: Compare the differences between the piglet microbial data and the corresponding new piglet microbial data to obtain the degree of influence of each bacterial genera on maternal transmission, and then screen out the key bacteria that have a significant impact on maternal transmission.

[0035] In the above-mentioned neural network-based method for predicting and screening key bacteria for maternal transmission, step 33 involves calculating the Euclidean distance between piglet microbial data and corresponding new piglet microbial data to determine the degree of influence of each bacterial genus on maternal transmission.

[0036] Furthermore, the present invention also provides a system for implementing the above-described method, comprising:

[0037] Data preprocessing module: used to acquire mother and child microbial data, form a dataset, and divide the dataset into training and testing sets;

[0038] Model building and training module: used to build a neural network model based on the attention mechanism. The neural network model is trained and adjusted using training and test sets to obtain a neural network model that predicts piglet microbial data based on the input sow microbial data.

[0039] Ablation Experiment Module: This module is used to perform feature ablation experiments on the neural network model trained in step 2 by removing bacterial genera from the sow microbial data in the dataset one by one, so as to obtain the key bacteria that have a significant impact on maternal transmission.

[0040] Beneficial effects

[0041] Compared with the prior art, the present invention has at least the following advantages:

[0042] This technology can accurately identify key bacterial genera in maternal transmission and effectively correlate related phenotypes, enabling the screening of target strains based on prediction results and improving screening efficiency. It provides a new tool for discovering scientific mechanisms and developing microbial resources in animal husbandry, and has broad application prospects. Attached Figure Description

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments;

[0044] Figure 1 This is a flowchart of the present invention;

[0045] Figure 2 Flowchart for calculating the core degree of maternal microorganisms in deep learning.

[0046] Figure 3A prediction model architecture based on an LSTM encoder-decoder architecture is provided for an embodiment of the present application

[0047] Composition.

[0048] Figure 4 The model training set training process index change, ten-fold cross-validation, true value and predicted value PC1 and PC2 principal component correlation.

[0049] Figure 5 The model calculation result.

[0050] Figure 6 The piglet intestinal microorganism prediction flora network diagram

[0051] Figure 7 According to the prediction result, the Christensenella genus is screened.

[0052] Figure 8 The full gene sequencing of the screened C.minuta bacteria.

[0053] Figure 9 The C.minuta gavage pregnant mice affect the intestinal door level of the mother and offspring mice, and increase the intestinal propionic acid concentration of the mother and offspring.

[0054] Figure 10 The C.minuta gavage pregnant mice increase the mother and offspring colon crypt depth and improve the intestinal structure.

[0055] Figure 11 The absolute quantification of C.minuta. DETAILED DESCRIPTION

[0056] The present application will be further described below in conjunction with the embodiments, but does not constitute any limitation on the present application, and any limited number of modifications made within the scope of the claims of the present application are still within the scope of the claims of the present application.

[0057] Embodiment 1

[0058] Reference Figure 1 and Figure 2 A neural network-based maternal transmission key bacteria prediction and screening method, comprising the following steps:

[0059] Step 1: Obtain maternal and offspring microbial data to form a data set, and divide the data set into a training set and a test set; the maternal and offspring microbial data includes sow microbial data and piglet microbial data; the sow microbial data and piglet microbial data are genus-level feature information of sows and piglets;

[0060] The acquisition of the mother-child microbial data can be achieved through various ways, such as NCBI retrieval, domestic database retrieval, and self-collection of data for gene analysis and sequencing.

[0061] The specific practice of the present application is based on the acquisition of mother-child microbial data through NCBI retrieval. First, the keywords such as "Sow", "piglet", "microbiome", "gut" are searched in NCBI, and the metadata file therein is used to screen the microbial data meeting the requirements in the database. Then, the sequence number information is sorted according to the metadata, and the original sequencing file of the microorganism is downloaded to process the original file of the microorganism to generate the microbial data with ASV table. More specifically, it is found whether the sequence number information of the microbial data is uploaded on the National Center for Biotechnology Information (NCBI) in the United States, the original information in FASTQ format of each sample is obtained according to the sequence number, and the qiime2 and DADA2 methods are used for quality filtering, trimming, denoising, merging and sequence chimerism, etc. Finally, the ASV table is obtained.

[0062] The ASV table can be seen in Table 1 below;

[0063] Table 1 ASV table

[0064]

[0065]

[0066] In Table 1, the characteristic information is the abundance, and the numbers in Table 1 are the abundance of each genus. In the actual operation process, the characteristic information is not limited to the abundance, but can also be the microbial community composition and functional annotation.

[0067] There is also a sample information table existing simultaneously with Table 1, and the sample information table marks the growth stage of the biological object of the sample and the mother-child relationship between the samples. For example, Table 2; in Table 2, the stage is the number of days, and the same pair_id represents the same group of sows and piglets.

[0068] Table 2 Sample information table

[0069] Sample ID Stage pair_id Generation Sample 1 D3 1 piglet Sample 2 D7 1 piglet Sample 3 D14 1 piglet Sample 4 S_D1 1 sow Sample 5 D3 2 piglet Sample 6 D7 2 piglet Sample 7 D14 2 piglet Sample 8 D21 2 piglet Sample 9 S_D3 2 sow Sample 10 D3 3 piglet Sample 11 D21 3 piglet Sample 12 S_D1 3 sow

[0070] The processing method of the mother-child microbial data includes the following operations: removing the missing values, performing zero substitution processing, center logarithmic ratio conversion and time point filling.

[0071] Because there are a large number of zero values in the microbial data, these zero values need to be replaced to reduce the influence of sparsity on model training. The cmultRepl method in the zCompositions library is used for zero replacement operation, and after removing the missing values and performing zero replacement processing, the ASV table becomes Table 3.

[0072] Table 3 ASV table

[0073]

[0074]

[0075] Then, the center logarithmic ratio conversion is performed on Table 3, and the converted ASV table becomes Table 4; the center logarithmic ratio (CLR) conversion formula is as follows, wherein P1 to Pn are the relative abundances of the first to nth features of a certain sample (a certain ASV): N is the geometric mean of all features in the sample.

[0076]

[0077] wherein is the geometric mean of all features in the sample.

[0078] The data set is divided into sow and piglet groups, and further divided into lactation and weaning stages according to the number of days of development.

[0079] Table 4 ASV table

[0080]

[0081] Time point filling is to keep the number of input data consistent when inputting the encoder and decoder; in specific practice, there may be missing samples, so it is necessary to fill the code, fill the sample number to the sample number required by the LSTM, and the filling information is the last time point of the sample; at the same time, the data is grouped according to different attributes of the sample, including the separation of sow and piglet data, and the division of piglet data according to the number of days, as shown in Table 5.

[0082] Table 5 ASV table

[0083]

[0084] Through the above operation, the data set as shown in Table 5 can be obtained; of course, the above is only an example, and in the actual operation process, the amount of data is much larger than the number of 12 samples listed above.

[0085] Step 2: Construct a neural network model based on attention mechanism, use the training set and test set to train and adjust the neural network model, and get the neural network model for predicting piglet microbial data according to the input sow microbial data;

[0086] In this embodiment, an encoder-decoder model based on long short-term memory (LSTM) network is constructed, and an attention mechanism is combined to predict the transmission of maternal bacteria to piglet flora. The model mainly consists of three parts: encoder, decoder and attention mechanism. Encoder design: the encoder extracts the microbiome features of sows at each stage through the LSTM network, and outputs the hidden state and cell state at each stage. Decoder design: the decoder also adopts the LSTM structure, and combines the feature value attention mechanism to extract key information from the output of the encoder, and decode into the microbiome data of piglets at each stage. Attention mechanism: the attention mechanism is used to calculate the importance of each time step in the decoding process, and highlights the microorganisms in the sow microbiome that have a greater impact on the piglet flora. Training process: the model is trained using the Adam optimizer, and a composite loss function is used, which combines the Euclidean distance, mean square error (MSE) and positive correlation weight of maternal bacteria. Model performance evaluation: the mean square error (MSE), mean absolute error (MAE) and determination coefficient (R 2 ) are used as evaluation indexes, and the model is verified for 10 times, and the correlation of PC1 and PC2 principal components of the predicted value and the true value is compared Figure 4

[0087] By encoding and decoding the sow data at multiple stages, the piglet flora changes are predicted, and the parameters of the model are optimized according to the real data.

[0088] The model consists of an encoder, a decoder and an attention mechanism, which can effectively capture the dynamic transmission process of maternal bacteria to piglet microbiome. Encoder: the encoder uses long short-term memory (LSTM) network, inputs the microbiome data of sows at each stage, and extracts key time series features.

[0089] LSTM (Long Short-Term Memory Network) is a kind of recurrent neural network (RNN), which introduces "memory gating unit" on the basis of traditional RNN, including input gate, forget gate and output gate. These gating units can effectively avoid the long-term dependence problem by selectively retaining or updating past information, thus showing excellent performance in modeling time series data. In this invention, LSTM encodes the time series features of sow microbiome data, which can capture the dynamic characteristics of flora at different stages, thus providing key support for the prediction of piglet flora.

[0090] ​The attention mechanism is introduced to enhance the model's focus on the specific time period of the microbial characteristics. Specifically, the attention mechanism assigns a weight to each time point's feature value, measuring the importance of different feature values in the decoding stage at that time step, allowing the decoder to focus on the sow microbial characteristics that have a greater impact on the piglet microbiome characteristics. This mechanism effectively improves the modeling accuracy of the maternal microbial transmission relationship, facilitating accurate prediction of key maternal microbes. The algorithm design of the model in this invention refers to the attention mechanism proposed by Bahdanau et al. (Bahdanau et al., 2015) and draws on the LSTM network proposed by Hochreiter and Schmidhuber (Hochreiter & Schmidhuber, 1997). These methods have been widely used in biological data analysis and sequence prediction. Time series features refer to the microbiome characteristics of sows and piglets at different periods, reflecting the dynamic changes of these microbial communities over time, such as changes in characteristics during the lactation period, weaning period, etc. Multi-stage features refer to microbiome characteristics collected at different growth stages (e.g., sow during pregnancy, lactation, weaning) or different time points of piglets (e.g., 7th day, 14th day, 28th day), which can include information such as microbial community composition, abundance, functional annotation, etc. In the neural network model, the encoder part uses the LSTM network to extract the time series microbiome characteristics of the sow, capturing the trends of microbial dynamic changes. Multi-stage features are input into the decoder, combined with the attention mechanism, allowing the model to effectively focus on the microbial transmission relationship between sows and piglets at different stages, thereby predicting the microbial communities of piglets at different stages.

[0091] The encoder is used to process the time series of sow microbiome data to capture the temporal dynamic characteristics and important features in microbial transmission. In the code, the input of the encoder is the sow's microbiome data, which has a shape of (batch_size, time_steps, feature_size), corresponding to the input sequences P1, P2, P3, P4, etc. in the figure. Figure 3 ) Formalized representation as follows, for each sow, the time series input of the encoder is where Pm is the final ASV table (Table 5) corresponding to the t = 1 to t = T time period of a certain sow, which contains feature_size number of features, where T is the input time_steps, where the input T needs to be consistent with the output piglet prediction time points, such as needing to output D3, D7, D14, D21 piglet data, then the input sow data T length needs to be 4, if the sow data collection point is missing, it can be filled (biologically, the sow's microorganisms are relatively stable during the lactation period).

[0092] which is mapped to

[0093]

[0094] The encoder is implemented using an LSTM network, with an initial hidden state h0, which is then updated sequentially to h1, h2, h3, etc. The encoder transforms the input sequence into multiple hidden states through multiple LSTM units. The output of the encoder is a set of hidden states at all time steps, rather than compressed into a single context vector. Each hidden state contains abstract information about the input features at that time step.

[0095] The decoder is used to predict the composition of the piglet microbiome at future time points, using the latent representation Hm and incorporating an attention mechanism that focuses on the most relevant time points and microbial features,

[0096]

[0097] where, is the predicted piglet microbiome composition at time point t.

[0098] Decoder: The decoder also uses an LSTM structure and incorporates an attention mechanism to extract key information related to the piglet microbiome from the hidden states output by the encoder, for decoding and prediction. Attention mechanism: The attention mechanism calculates the importance of each feature at the current time step. Specifically, the attention mechanism does not weight different time steps of the encoder output, but rather calculates the importance of each feature. This allows the model to focus more on key bacteria in the transmission process. The attention mechanism is used to calculate the attention weight a t to highlight the contribution of each time point and microbial feature to the prediction result:

[0099]

[0100] where W a and b a are learnable parameters, and is the hidden state at time point t

[0101] Through this structure, the model can model the relationship between the sow microbiome and the piglet microbiome at different stages and predict changes in the piglet microbiome. The LSTM output of the decoder is passed through a fully connected layer (fc) to map the hidden state to the target feature space, generating the final prediction value (i.e., the abundance value of the piglet microbiome feature).

[0102] 3. Loss function design

[0103] In order to effectively deal with the prediction problem of key bacteria transmitted by the mother source, the present application proposes a composite loss function combining mean square error (MSE) and Euclidean loss, and adjusting according to the positive proportion of the bacterial flora, so as to emphasize more common microorganisms:

[0104] L = a x EuclideanLoss + b x MSE + l x L1Regularization

[0105] Wherein: EuclideanLoss: used to measure the root mean square difference between the predicted piglet microbial flora composition and the actual value. MSE: the sum of the mean square error calculated at each time point t. L1Regularization: encourages the sparsity of model parameters to prevent overfitting.

[0106] Step 3: Perform feature ablation experiment on the neural network model trained in step 2 by removing each genus of maternal microbial data in the data set, and obtain key bacteria that have a significant impact on maternal transmission.

[0107] The step 3 is specifically:

[0108] Step 31: Remove each genus of maternal microbial data in the data set to obtain a plurality of new maternal microbial data;

[0109] Step 32: Predict new piglet microbial data using the neural network model with new maternal microbial data;

[0110] Step 33: Compare the differences between the piglet microbial data and the corresponding new piglet microbial data to obtain the degree of influence of each genus on maternal transmission, and then filter to obtain key bacteria that have a significant impact on maternal transmission. More specifically, in step 33, the degree of influence of each genus on maternal transmission is determined by calculating the Euclidean distance between the piglet microbial data and the corresponding new piglet microbial data.

[0111] The experimental results show that certain specific genera have a key role in the maternal transmission process, and removing these genera will significantly affect the composition of the piglet microbiome. For each piglet, the trained model is used to predict the composition of the piglet's microbial flora Then systematically remove each specific microbial group from the maternal microbial flora data:

[0112] For all t, set

[0113] The maternal microbial flora vector is renormalized to ensure that it still belongs to the probability simplex.

[0114] Generating new predictions using modified sow data

[0115] The impact of removing group i is quantified by calculating the Transmission Centrality (TC) between the original and modified predictions:

[0116]

[0117] To verify the actual prediction performance of the model, the prediction results of the model were evaluated in combination with the growth performance data of piglets. Mean Squared Error (MSE), Mean Absolute Error (MAE) and R 2 ) were used as evaluation indicators to test the prediction accuracy of the model. Through comparative analysis of the validation dataset, it was found that the model could effectively predict the transmission of key maternal bacterial genera and was closely related to the growth performance of piglets. The distribution difference between the predicted results and the true values was visualized by PCA (Principal Component Analysis) Figure 4 ), which further verified the prediction effect of the model; based on the prediction results, the predicted key microorganisms were subjected to directional culture screening and functional verification.

[0118] In one case of the present application, Christensenellaceae R-7 group was predicted by the model to have a higher maternal genetic score Figure 5 ), and Christensenellaceae R-7 group showed a strong centrality in the network analysis of bacterial genera Figure 6 ). Therefore, we screened for PYG medium against Christensenellaceae Figure 7 ), and obtained a key species C. minuta, and performed whole genome determination Figure 8 ). Female ICR mice (4-5 weeks old) were raised under specific pathogen-free (SPF) conditions at Hunan Agricultural University. After one week of adaptation, the mice were mated, and after successful pairing, the mice in each group were orally administered 5x10 8 colony forming units (CFU) suspended in sterile phosphate buffered saline (PBS). Gavage was continued from mating to lactation until weaning. Maternal fecal samples were collected during late pregnancy and after weaning, and offspring fecal samples were collected after weaning. Before gavage, C. minuta was cultured in a 37°C peptone-yeast-glucose (PYG) medium. The results showed that after gavage of the mother mice, compared with the control group, the gavage bacteria group showed an increase in the abundance of Bacillus and a decrease in the abundance of Bacteroides in both the mother mice and the offspring, and the intestinal propionic acid concentration and crypt depth of the offspring were increased Figure 9 , 10). And the absolute quantification of bacteria in feces also showed an increase in the content of C. minutaFigure 11 ). N 母鼠 (control) = 9, N 子鼠 (control) = 11, N 母鼠 (inoculated) = 8, N 子鼠 (inoculated) = 10.

[0119] The embodiments presented herein are only selected embodiments according to a combination of all possible embodiments. The appended claims should not be limited by the embodiments illustrating the present application. Some numerical ranges used in the claims include subranges within them, and variations of these ranges should also be covered by the appended claims.

Claims

1. A method for predicting and screening key maternally transmitted bacteria based on neural networks, characterized in that, Includes the following steps: Step 1: Obtain maternal and piglet microbial data to form a dataset, and divide the dataset into a training set and a test set; the maternal and piglet microbial data includes sow microbial data and piglet microbial data; the sow microbial data and piglet microbial data are genus-level feature information of sows and piglets; the feature information is one or more combinations of microbial community composition, genus abundance, and functional annotation; Step 2: Construct a neural network model based on the attention mechanism. Use training and test sets to train and adjust the neural network model to obtain a neural network model that predicts piglet microbial data based on the input sow microbial data. Step 3: The neural network model trained in Step 2 was subjected to feature ablation experiments by removing bacterial genera from the sow microbial data in the dataset one by one to obtain the key bacteria that have a significant impact on maternal transmission. The neural network model consists of an encoder, a decoder, and an attention mechanism. The encoder employs a long short-term memory network to process sow microbial data, extract time-series features, and convert them into hidden state vectors. The decoder employs a long short-term memory network to extract information related to the piglet microbiome from the hidden state vectors and form prediction results. The attention mechanism is used to calculate attention weights. To highlight the contribution of each time point and microbial characteristics to the prediction results; ；； in and For learnable parameters, It is the hidden state vector at time t.

2. The method for predicting and screening key maternally transmitted bacteria based on neural networks according to claim 1, characterized in that, Step 1 specifically involves: Step 1: Obtain mother and offspring microbial data; Step 2: Preprocess the mother and child microbial data, including removing missing values ​​and performing zero substitution. Step 3: Perform center-log ratio transformation and time point filling on the dataset to generate the dataset.

3. The method for predicting and screening key maternally transmitted bacteria based on neural networks according to claim 1, characterized in that, During the training process of the neural network model, a loss function is used to optimize the neural network model.

4. The method for predicting and screening key maternally transmitted bacteria based on neural networks according to claim 3, characterized in that, The loss function is a combination of three methods: Euclidean distance loss, mean squared error algorithm, and L1 regularization term; The formula for the loss function is: L=α×EuclideanLoss+β×MSE+λ×L1Regularization Wherein: EuclideanLoss is the Euclidean distance loss algorithm, used to measure the root mean square difference between the predicted piglet microbiome composition and the actual value; MSE is the mean squared error algorithm, used to calculate the sum of mean squared errors at each time point t; L1 Regularization is an L1 regularization term used to encourage the sparsity of model parameters to prevent overfitting. α, β, and λ are constants.

5. The method for predicting and screening key maternally transmitted bacteria based on neural networks according to claim 1, characterized in that, Step 3 specifically involves: Step 31: Remove the bacterial genera from the sow microbial data in the dataset one by one to obtain multiple new sow microbial data; Step 32: Use the new sow microbial data and a neural network model to predict the new piglet microbial data; Step 33: Compare the differences between the piglet microbial data and the corresponding new piglet microbial data to obtain the degree of influence of each bacterial genera on maternal transmission, and then screen out the key bacteria that have a significant impact on maternal transmission.

6. The method for predicting and screening key maternally transmitted bacteria based on neural networks according to claim 5, characterized in that, In step 33, the degree of influence of each genera on maternal transmission is determined by calculating the Euclidean distance between the piglet microbial data and the corresponding new piglet microbial data.

7. A system for implementing the method of any one of claims 1 to 6, characterized in that, include: Data preprocessing module: used to acquire mother and child microbial data, form a dataset, and divide the dataset into training and testing sets; Model building and training module: used to build a neural network model based on the attention mechanism. The neural network model is trained and adjusted using training and test sets to obtain a neural network model that predicts piglet microbial data based on the input sow microbial data. Ablation Experiment Module: This module is used to perform feature ablation experiments on the neural network model trained in step 2 by removing bacterial genera from the sow microbial data in the dataset one by one, so as to obtain the key bacteria that have a significant impact on maternal transmission.

Images