Probiotic screening method based on neural networks and metagenomic data

Abstract

This invention discloses a probiotic screening method based on neural networks and metagenomic data, relating to the field of probiotic screening. It involves obtaining metagenomic sequencing data from NCBI, preprocessing the data, and dividing the processed sequence data into training, validation, and test sets. A deep learning model is constructed based on a Transformer architecture combined with a HyenaOperator hybrid mechanism to capture the internal dependencies of sequence fragments at different distances, performing pooling to obtain better sequence features. These features are then input into a multilayer perceptron (MLP) base model for dimensionality reduction, and the predicted probabilities are output using a softmax function. A voting mechanism is used to summarize and determine the predicted results of the sequence fragments. This invention employs the aforementioned probiotic screening method based on neural networks and metagenomic data, enabling efficient and accurate screening of potential probiotics in large-scale microbial genome data.

Description

Technical Field

[0001] This invention relates to the field of probiotic screening, and in particular to a probiotic screening method based on neural networks and metagenomic data. Background Technology

[0002] With the rapid development of metagenomics technology, genomic data has become easier to obtain. However, effectively screening beneficial microorganisms (probiotics) from massive metagenomic data remains a challenge. Probiotics play a vital role in maintaining gut health. Current methods for screening probiotics mostly rely on traditional laboratory culture techniques and manual analysis. These methods are time-consuming, costly, and have poor reproducibility. With the development of computational and biological data analysis technologies, probiotic screening models based on metagenomic data have gradually become a research hotspot. Most existing technologies are based on specific feature extraction methods and have failed to combine efficient machine learning models for efficient screening and selection. Summary of the Invention

[0003] The purpose of this invention is to provide a probiotic screening method based on neural networks and metagenomic data. By combining deep learning technology with biological data, this method provides an efficient, accurate, and scalable probiotic screening method that can efficiently screen potential probiotics from a large amount of microbial genome data.

[0004] To achieve the above objectives, this invention provides a probiotic screening method based on neural networks and metagenomic data, comprising the following steps: S1. Based on the collected metagenomic sequencing data, perform data preprocessing and divide the processed sequence data into training set, validation set and test set; S2. Construct a deep learning model based on the Transformer architecture combined with the HyenaOperator hybrid convolutional gating mechanism, and train and optimize it. S3. Using the deep learning model constructed in S2, capture the internal dependencies of sequence segments at different distances and perform pooling to obtain better sequence features; S4. Input the sequence features output after pooling in S3 into the multilayer perceptron (MLP) basic model for dimensionality reduction and select the highest score logits among each sequence feature through the self-attention mechanism. Then, use the softmax function to output the prediction probability. S5. Based on the prediction results of the sequence fragment in S4, a voting mechanism is used to summarize and determine the sequence data.

[0005] Preferably, in S1, the collected metagenomic sequencing data is sequence cleaned, the contents of the FASTA format sequence file are traversed, abnormal bases and low-quality sequence fragments already reported in the database are removed, high-quality genomic data are retained, and the genomic data is divided into 4096bp fragments. The segmented data fragments are then divided into training set, validation set and test set.

[0006] Preferably, the specific process of deep learning model training optimization in S2 is as follows: S21. Input the training set data processed in S1 into the constructed deep learning model for training and optimization; S22. Input the test set data processed in S1 into the deep learning model trained in S21 for model evaluation. S23. Input the validation set processed by S1 into the deep learning model trained by S21 to predict sequence segments. Evaluate the prediction results through classification metrics and output the training loss. Input the training loss into the deep learning model to retrain the deep learning model in the direction of reducing the loss. S24. Through continuous iterative training via S23, the final training result is obtained when the loss of adjacent training deep learning models does not decrease significantly, and classification metrics are evaluated.

[0007] Preferably, the deep learning model of S22 is evaluated on the test set using standard evaluation metrics, including accuracy, precision, recall, and F1 score.

[0008] Preferably, during the training process of the deep learning model in S23, the obtained predicted probabilities are updated based on the loss function between the predicted probabilities and the true labels through the backpropagation mechanism. At the same time, the learning rate is dynamically adjusted by combining a preset learning rate adjustment strategy or an adaptive optimizer to promote model convergence.

[0009] Preferably, the specific contents of the classification indicators in S23 and S24 are as follows: Each sequence fragment is assigned a probiotic relevance score, which ranges from 0 to 1. The closer the score is to 1, the greater the probability that the sequence fragment is associated with probiotic characteristics. The further away the score is from 1, the smaller the probability that the sequence fragment is associated with probiotic characteristics.

[0010] Preferably, the specific process of S3 is as follows: S31. Capture mid-range dependencies of sequence segments through self-attention mechanism; S32. Capture short-range dependencies of sequence segments through convolution operations; S33. Introduce frequency domain representation through fast Fourier transform to capture long-range correlation of sequence segments; S34. Pool the internal dependencies of sequences at different distances obtained in S31~S33 to obtain better sequence features and output them in 512-dimensional form.

[0011] Preferably, in S5, the complete metagenomic sequence is summarized and judged based on the probiotic relevance score of the sequence fragment. The sequence fragments are divided into different priorities by setting a threshold, and the set threshold can be optimized and calibrated based on the performance index of the validation set.

[0012] Preferably, the voting mechanism in S5 is based on a weighted voting system using the probability scores of all sequence fragments. If the proportion of positive votes exceeds a set threshold, the sequence is determined to be a potential probiotic sequence; if the proportion of positive votes does not exceed the set threshold, the sequence is determined to be a non-potential probiotic sequence.

[0013] Therefore, the probiotic screening method based on neural networks and metagenomic data described above has the following advantages compared with the prior art: 1. This invention combines deep learning methods to efficiently screen potential probiotics from massive metagenomic datasets, offering strong interpretability and stable results. By combining three types of neural networks, the model can accurately identify dependencies between sequences of different lengths, assisting in the screening of microorganisms with probiotic potential and distinguishing non-probiotic samples. It can automatically extract effective features from massive metagenomic data, achieving automated and high-throughput probiotic screening. Compared to traditional manual methods, this invention offers faster screening speeds, lower costs, and excellent scalability and adaptability across multiple samples and scenarios.

[0014] 2. Innovative model structure: The proposed neural network integrates self-attention, convolution, and Fast Fourier Transform (FFT) mechanisms, enabling it to handle both short and long sequence dependencies and achieve multi-scale feature learning. The introduction of the Hyena operator ensures the model remains efficient when processing ultra-long sequences. Experimental results show that this mechanism achieves performance comparable to or even better than traditional attention when inputs exceed 8,000 base pairs, while significantly reducing computational resource consumption.

[0015] 3. Improved algorithm performance: Based on pre-trained and deep architecture-based deep learning models, this invention demonstrates significant advantages over traditional sequence feature methods or CNN baseline models in typical genome classification tasks. The method of this invention achieves an accuracy of approximately 97% on test data, significantly outperforming existing shallow methods, thereby improving the reliability and stability of the screening process.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1This is an overall flowchart of a probiotic screening method based on neural networks and metagenomic data according to the present invention; Figure 2 This is the ROC result of a probiotic screening method based on neural networks and metagenomic data trained at the document level according to the present invention; Figure 3 This is the PRC result of a probiotic screening method based on neural networks and metagenomic data trained at the document level according to the present invention; Figure 4 This is the ROC result of sequence-level training for a probiotic screening method based on neural networks and metagenomic data according to the present invention; Figure 5 This is the PRC result of sequence-level training of a probiotic screening method based on neural networks and metagenomic data according to the present invention. Detailed Implementation

[0018] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0019] Example like Figures 1-5 As shown, the present invention provides a probiotic screening method based on neural networks and metagenomic data, comprising the following steps: S1. Based on the collected metagenomic sequencing data, perform data preprocessing and divide the processed sequence data into training set, validation set and test set; The preprocessing process is as follows: The collected metagenomic sequencing data is sequence cleaned, the contents of the FASTA format sequence file are traversed, abnormal bases and low-quality sequence fragments already reported in the database are removed, high-quality genomic data are retained and the genomic data is divided into 4096bp fragments, and then the segmented data fragments are divided into training set, validation set and test set. The complete metagenomic data of a single microorganism is divided into corresponding sequences of 4096bp. Each 4096bp corresponds to a tag that is consistent with the microorganism it belongs to. The numbers 0-9 are used to represent ATCG and its task types such as classification, reasoning, and clustering, respectively, and the biological sequence data is converted into digital sequence information. S2. A deep learning model is constructed and trained and optimized based on the Transformer architecture combined with the HyenaOperator hybrid convolutional gating mechanism. The FFT algorithm and attention mechanism are alternated in each training round, with 20 layers alternating. The attention mechanism is used in layers 2, 7, 11 and 16 to strengthen the frequency domain correlation after transformation. The complex biological sequence is converted into a visual digital feature by using time-domain frequency domain transformation and random linearized Q, K and V attention interaction respectively. Q is the query, K is the key and V is the value. S21. Input the training set data processed by S1 into the deep learning model constructed by S2 for training and optimization. After each training, the downstream sample is weighted by the similarity score of each feature and then subjected to softmax processing to obtain the probability corresponding to each sample. At the same time, the samples of the same type of microorganism are weighted and averaged to obtain the probability corresponding to the complete microorganism. S22. Input the test set data processed in S1 into the deep learning model trained in S21 for model evaluation. S23. Input the validation set processed by S1 into the deep learning model trained by S21 to predict sequence segments. Evaluate the prediction results through classification metrics and output the training loss. Input the training loss into the deep learning model to retrain the deep learning model in the direction of reducing the loss. S24. Through continuous iterative training in S23, the final training result is obtained when the loss of adjacent training deep learning models no longer decreases significantly (i.e., the reduction is less than 0.1% of the previous loss), and the classification index is evaluated. The S22 deep learning model evaluates its performance on the test set using standard evaluation metrics, including accuracy, precision, recall, and F1 score. During the training process of the deep learning model in S23, the obtained prediction probability is updated based on the loss function between the prediction probability and the true label through the backpropagation mechanism. At the same time, the learning rate is dynamically adjusted by combining the preset learning rate adjustment strategy or the adaptive optimizer to promote model convergence. The specific details of the classification indicators in S23 and S24 are as follows: Each sequence fragment is assigned a probiotic relevance score, which ranges from 0 to 1. The closer the score is to 1, the greater the probability that the sequence fragment is associated with probiotic characteristics. The further away the score is from 1, the smaller the probability that the sequence fragment is associated with probiotic characteristics. S3. Using the deep learning model constructed in S2, capture the internal dependencies of sequence segments at different distances and perform pooling to obtain better sequence features; S31. Capture mid-range dependencies of sequence segments through self-attention mechanism; S32. Capture short-range dependencies of sequence segments through convolution operations; S33. Introduce frequency domain representation through fast Fourier transform to capture long-range correlation of sequence segments; S34. Pool the internal dependencies of sequences at different distances obtained in S31~S33 to obtain better feature information and output it in 512-dimensional form. S35. Input the 512-dimensional features output after pooling in S24 into the multilayer perceptron (MLP) basic model to reduce the feature dimension to 128 dimensions. Process each sequence feature by using a self-attention mechanism to treat the attention score as logits. Select the highest logits among each sequence feature and use the softmax function to predict the probability. When processing sequences ranging from thousands to tens of thousands of bases in length, the Hyena operator, through alternating hybrid convolution and gating operations, can significantly reduce computational complexity while maintaining recognition accuracy comparable to traditional attention mechanisms. For example, the Hyena operator can operate twice as fast as traditional attention mechanisms for 8K-length sequences and up to 100 times faster for 64K-length sequences. S4. Input the sequence features output after pooling in S3 into the multilayer perceptron (MLP) basic model for dimensionality reduction and select the highest score logits among each sequence feature through the self-attention mechanism. Then, use the softmax function to output the prediction probability. S5. Based on the prediction results of the sequence fragment in S4, a voting mechanism is used to summarize and determine the sequence data; The complete metagenomic sequences are aggregated and judged based on the probiotic relevance score of the sequence fragments using a voting mechanism. Sequence fragments are divided into different priorities by setting thresholds, and the set thresholds can be optimized and calibrated based on the performance indicators of the validation set. The voting mechanism is based on a weighted voting system that calculates the probability scores of all sequence fragments. If the proportion of positive votes exceeds a set threshold, the sequence is identified as a potential probiotic sequence. If the proportion of positive votes does not exceed the set threshold, the sequence is identified as a non-potential probiotic sequence.

[0020] The model boasts multi-scenario adaptability, drawing data from publicly available microbial genome databases (such as NCBI) and published probiotic research literature. Through literature mining and manual review, complete genomes of experimentally validated probiotic strains are selected as positive samples. Simultaneously, genomes of known non-probiotic or conditionally pathogenic microorganisms are chosen as negative samples. Each complete genome sequence corresponds to a category label, enabling the model to be applied to microbial metagenomic data from various sources, including the gut, food, and environment. During training, the model learns the general characteristics of microbial sequences, allowing for cross-sample and cross-species generalization and rapid adaptation to diverse application scenarios. Currently, models developed for soil samples have also achieved similar high accuracy, demonstrating that this type of deep learning framework can be universally applied to probiotic discovery in various environmental samples.

[0021] In the specific implementation process, firstly, the metagenomic sequencing data is preprocessed to remove low-quality sequences and retain high-quality sequences, and then relevant features are embedded. Secondly, a composite neural network model is constructed using a deep learning framework (such as PyTorch). The model integrates the Transformer architecture, self-attention mechanism, convolutional neural network (CNN), and fast Fourier transform (FFT) module to achieve comprehensive modeling of sequence features at different scales, and combines HyenaOperator to enhance the ability to extract local features. The model hyperparameters are optimized through methods such as cross-validation to ensure its generalization ability. Finally, the trained model is used to predict new metagenomic samples, calculate the probiotic relevance score of each sequence fragment, and screen potential probiotics through a weighted voting mechanism.

[0022] The specific implementation method is as follows: A laboratory has completed metagenomic sequencing of gut microbiota samples, obtaining a batch of high-quality genomic sequence data. Traditional probiotic screening relies on culture and phenotypic detection, which is time-consuming and has poor reproducibility. In order to achieve efficient identification of potential probiotics in experimental samples, the laboratory applied the deep learning model for probiotic screening based on metagenomic data of this invention. The model is used to extract features and make intelligent predictions on the sequencing data, thereby quickly screening out strain sequences that may have probiotic characteristics.

[0023] Step 1: The metagenomic sequences obtained in the laboratory are structured and the sequencing data is imported into the model input pipeline. The sequences are standardized and cleaned (abnormal bases and repetitive fragments are removed). High-quality sequences are divided into 4096bp fragments. Each fragment serves as the basic input unit of the model. The embedding layer is used to convert the DNA sequence fragments into high-dimensional vector representations and embed sequence feature information (such as base frequency distribution, GC content, etc.). This step ensures that the input data has consistency and feature integrity before entering the deep model.

[0024] The second step involves capturing medium- to long-range dependencies in metagenomic sequences using a self-attention module. This enables the model to understand the contextual structure across segments and improves its ability to identify functionally related genes. The model incorporates convolutional neural networks (CNNs) and fast Fourier transform (FFT) modules. CNNs are responsible for extracting short-range local dependencies, while FFTs are responsible for capturing long-term or frequency-domain features in the sequence. The combination of these two modules enhances the model's sensitivity to features at different scales in the genomic sequence.

[0025] The third step involves multi-layer pooling of features from the self-attention layer, convolutional layer, and Fourier layer. The training error is calculated using a feed-forward layer and a loss function. Hyperparameters are adjusted using cross-validation to optimize the model's generalization performance. The model is then input into a multilayer perceptron (MLP) for comprehensive learning to achieve a unified representation of macroscopic and microscopic information.

[0026] Step 4: The model output is converted into a probability distribution through the softmax function. Each sequence fragment obtains a probiotic relevance score, which ranges from 0 to 1. The closer the score is to 1, the greater the probability that the sequence fragment is related to probiotic features. The further away the score is from 1, the smaller the probability that the sequence fragment is related to probiotic features.

[0027] Step 5: For a complete metagenomic sequence, the model performs weighted voting based on the probability scores of all sequence fragments. If the proportion of positive votes exceeds a set threshold (e.g., 0.7), the sequence is determined to be a potential probiotic sequence. If the proportion of positive votes does not exceed the set threshold, the sequence is determined to be a non-potential probiotic sequence.

[0028] Therefore, this invention employs a probiotic screening method based on neural networks and metagenomic data, which has significant innovation and advantages compared with traditional culture and shallow analysis methods. It utilizes an advanced neural network model to achieve automated feature extraction and classification, eliminating the need for manually designed complex features, thus significantly improving screening speed and flexibility. Secondly, in terms of model structure, it integrates the global self-attention mechanism of Transformer with the convolutional gating mechanism of Hyena operators, enabling the model to simultaneously capture short-range, medium-range, and long-range dependencies in microbial sequences. This architectural innovation gives the method a natural advantage in processing ultra-large-scale metagenomic data, effectively alleviating the quadratic computation bottleneck of the attention mechanism and achieving efficient utilization of ultra-long sequence information.

[0029] Compared with existing technologies, the deep learning screening method of this invention significantly improves algorithm performance and application effectiveness. Traditional probiotic screening relies on artificial culture and feature design, which is time-consuming, costly, and difficult to scale. In contrast, this invention achieves intelligent discrimination in one step by constructing an end-to-end learning model. Related studies have shown that artificial intelligence technology exhibits extremely high accuracy and the ability to process massive amounts of data in microbial identification and screening. The neural network model constructed in this invention achieves an accuracy of approximately 97% on publicly available genomic datasets, far exceeding the results based on shallow machine learning or traditional genomics analysis. This means that this invention can more stably identify potential probiotics, thereby significantly improving the sensitivity and accuracy of screening. Furthermore, this invention adopts a voting-weighted result aggregation strategy and adjusts the threshold through the validation set, effectively balancing the false positive and false negative rates, and improving the reliability and adjustability of the screening results.

[0030] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A probiotic screening method based on neural networks and metagenomic data, characterized in that, Includes the following steps: S1. Based on the collected metagenomic sequencing data, perform data preprocessing and divide the processed sequence data into training set, validation set and test set; S2. Construct a deep learning model based on the Transformer architecture combined with the HyenaOperator hybrid convolutional gating mechanism and train and optimize it; S3. Using the deep learning model constructed in S2, capture the internal dependencies of sequence segments at different distances and perform pooling to obtain better sequence features; S4. Input the sequence features output after pooling in S3 into the multilayer perceptron (MLP) basic model for dimensionality reduction and select the highest score logits among each sequence feature through the self-attention mechanism. Then, use the softmax function to output the prediction probability. S5. Based on the prediction results of the sequence fragment in S4, a voting mechanism is used to summarize and determine the entire sequence data.

2. The probiotic screening method based on neural network and metagenomic data according to claim 1, characterized in that: In S1, the collected metagenomic sequencing data is cleaned by traversing the contents of the FASTA format sequence files, removing abnormal bases and low-quality sequence fragments already reported in the database, retaining high-quality genomic data, and dividing the genomic data into 4096bp fragments. The segmented data fragments are then divided into training set, validation set, and test set.

3. The probiotic screening method based on neural network and metagenomic data according to claim 2, characterized in that: The specific process of deep learning model training and optimization in S2 is as follows: S21. Input the training set data processed in S1 into the constructed deep learning model for training and optimization; S22. Input the test set data processed in S1 into the deep learning model trained in S21 for model evaluation. S23. Input the validation set processed by S1 into the deep learning model trained by S21 to predict sequence segments. Evaluate the prediction results through classification metrics and output the training loss. Input the training loss into the deep learning model to retrain the deep learning model in the direction of reducing the loss. S24. Through continuous iterative training via S23, the final training result is obtained when the loss of adjacent training deep learning models does not decrease significantly, and classification metrics are evaluated.

4. The probiotic screening method based on neural network and metagenomic data according to claim 3, characterized in that: The S22 deep learning model uses standard evaluation metrics on the test set to evaluate model performance. Standard evaluation metrics include accuracy, precision, recall, and F1 score.

5. The probiotic screening method based on neural network and metagenomic data according to claim 3, characterized in that: During the training process of the deep learning model in S23, the obtained predicted probabilities are used to update the model parameters based on the loss function between the predicted probabilities and the true labels through the backpropagation mechanism. At the same time, the learning rate is dynamically adjusted by combining a preset learning rate adjustment strategy or an adaptive optimizer to promote model convergence.

6. The probiotic screening method based on neural network and metagenomic data according to claim 5, characterized in that: The specific details of the classification indicators in S23 and S24 are as follows: Each sequence fragment is assigned a probiotic relevance score, which ranges from 0 to 1. The closer the score is to 1, the greater the probability that the sequence fragment is associated with probiotic characteristics. The further away the score is from 1, the smaller the probability that the sequence fragment is associated with probiotic characteristics.

7. The probiotic screening method based on neural network and metagenomic data according to claim 1, characterized in that: The specific process of S3 is as follows: S31. Capture mid-range dependencies of sequence segments through self-attention mechanism; S32. Capture short-range dependencies of sequence segments through convolution operations; S33. Introduce frequency domain representation through fast Fourier transform to capture long-range correlation of sequence segments; S34. Pool the internal dependencies of sequences at different distances obtained in S31~S33 to obtain better sequence features and output them in 512-dimensional form.

8. The probiotic screening method based on neural network and metagenomic data according to claim 7, characterized in that: In S5, the complete metagenomic sequence is aggregated and judged based on the probiotic relevance score of the sequence fragments. The sequence fragments are divided into different priorities by setting thresholds, and the set thresholds can be optimized and calibrated based on the performance indicators of the validation set.

9. The probiotic screening method based on neural network and metagenomic data according to claim 8, characterized in that: The voting mechanism in S5 is based on a weighted voting system that calculates the probability scores of all sequence fragments. If the proportion of positive votes exceeds a set threshold, the sequence is identified as a potential probiotic sequence. If the proportion of positive votes does not exceed the set threshold, the sequence is identified as a non-potential probiotic sequence.

Images