A deep learning-based early warning method for respiratory virus new variants

Abstract

The present application relates to the technical field of virus data processing, and particularly relates to a respiratory virus new variant early warning method based on deep learning, through variation feature extraction, three nodes most prone to variation of respiratory viruses are extracted as variation features, namely variation caused by instability of RNA itself, variation caused by errors in translation of nucleotides into amino acids, and variation caused by gene fragment recombination, namely quantifying variation activity of viruses at the nucleotide level, functional changes at the protein level, evaluating the risk of new variants caused by gene recombination, and an increase in recombination events means an increase in the probability of variants; and the propagation feature is added to deep learning, the model can automatically learn the variation-propagation correlation mode of the new variant, and then predict the outbreak trend of the respiratory virus variant, and then realize early warning; various data are updated in real time, and quick early warning can be realized according to the latest prediction effect.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of virus data processing, and in particular to a respiratory virus new variant early warning method based on deep learning. BACKGROUND

[0002] The frequent emergence of respiratory virus new variants has brought great challenges to global public health. Timely and accurate early warning of the outbreak of new variants is crucial for taking effective prevention and control measures and reducing the impact of the epidemic. However, the existing technology has many shortcomings in dealing with this problem. Traditional monitoring methods still mainly rely on laboratory detection and case reporting, which have obvious lag. For example, in the early stage of the emergence of a new virus variant, due to the limitations of detection methods, it may not be possible to discover it in time, resulting in the spread of the epidemic without being detected. At the same time, prediction methods based on search engine data or simple models are difficult to fully capture the complex characteristics and transmission rules of virus variation. Some influenza prediction models based on search engine data, although they can use big data, are still passive predictions and have a lag. Therefore, a deep learning-based respiratory virus new variant early warning method with a certain foresight and reduced passivity is needed. SUMMARY

[0003] The technical problem to be solved by the application is to provide a deep learning-based respiratory virus new variant early warning method with a certain foresight and reduced passivity.

[0004] To solve the above technical problems, the technical scheme adopted by the application is:

[0005] A deep learning-based respiratory virus new variant early warning method, comprising:

[0006] Data collection and preprocessing, collecting whole genome sequence data of respiratory viruses, daily new case number and transmission range of respiratory viruses;

[0007] Variation feature extraction, calculating nucleotide variation index , = 1 - (1 - x) / x ; translating nucleotide sequences into amino acid sequences, calculating amino acid variation coefficient , = a x ; calculating gene fragment recombination risk value , = beta x (1 - x) ; defining variation transmission acceleration factor = gamma x (x + y + z) ​​​

[0008] wherein, is the length of the viral sequence, is the maximum variant frequency for each site, n is the number of nucleotide types, is the number of times the jth nucleotide type appears at the ith site; is the total number of all nucleotide types at the ith site in the same viral sequence population; and a is an adjustment coefficient, is the length of the amino acid sequence, is the number of sites where substitution occurs; and b is an adjustment coefficient, is the number of sites where substitution occurs; and b is an adjustment coefficient, is the number of identical bases, is the length of the recombinant fragment,

[0009] The transmission characteristics are extracted by calculating the transmission speed and the range expansion rate from the daily new cases and the transmission range. and the range expansion rate , to construct a range expansion index = × ;

[0010] A deep learning model is constructed, which adopts a combined architecture of CNN and LSTM, takes , , , , , as input, and takes whether a new variant outbreak occurs as output label; the constructed model inputs new data and outputs predicted probability values.

[0011] Preferably, the transmission speed = , wherein is the number of new cases on the tth day, is the number of new cases on the (t-1)th day.

[0012] Preferably, the range expansion rate = , wherein is the number of regions where respiratory viruses spread on the tth day, is the number of regions where respiratory viruses spread on the (t-1)th day.

[0013] Preferably, a combined architecture of CNN and LSTM is adopted, CNN extracts local patterns of viral sequence variation characteristics, and LSTM processes time series dynamic trends of transmission data.

[0014] Preferably, when constructing the deep learning model, a cross-entropy loss function L is used to quantify the difference between the predicted probability of the model and the true label, L=- ;

[0015] wherein N is the number of samples, is the true label of the kth sample; is the predicted probability of the model for the kth sample, with a value range of 0-1.

[0016] Preferably, when constructing the deep learning model, model optimization is also included, L1 and L2 regularization and Dropout technology are adopted to prevent overfitting, and the loss function after optimization is =L+ + ;

[0017] wherein, , is a regularization parameter, is a model parameter, is the number of parameters, and the generalization ability of the model is improved.

[0018] Preferably, the cross-entropy loss function L adjusts the parameters through the back propagation algorithm, minimizes the loss function, and improves the prediction accuracy.

[0019] Preferably, the constructed model inputs new data and outputs a predicted probability value, and a warning threshold T is set through historical data and / or expert experience, and a warning signal is issued when the predicted probability is greater than T.

[0020] Preferably, alpha is a weight coefficient obtained by regression analysis according to historical data, which is used to quantify the influence degree of amino acid variation on the transmission ability of the virus.

[0021] Preferably, beta is an adjustment coefficient set according to different virus types.

[0022] The beneficial effects of the present application are as follows: by extracting variation characteristics, the three nodes most likely to vary of respiratory viruses are extracted as variation characteristics, i.e. variation caused by instability of RNA itself, variation caused by errors in translation of nucleotides into amino acids, and variation caused by recombination of gene fragments, quantify the variation activity of the virus at the nucleotide level, the functional change at the protein level, the greater the probability of change, the more likely the protein to increase the transmission force, evaluate the risk of new variants caused by gene recombination, and an increase in recombination events means an increase in the probability of variants; and add the transmission characteristics to the deep learning, the model can automatically learn the variation-transmission correlation pattern of new variants, and then predict the outbreak trend of respiratory virus variants, and then realize early warning; and various data are updated in real time, so that the prediction model can realize rapid warning and response according to the latest prediction effect, and reduce passivity. DETAILED DESCRIPTION

[0023] To explain the technical content of the present application, the purposes achieved and the effects, the following will be described in conjunction with the embodiments.

[0024] A deep learning-based early warning method for respiratory virus new variants, comprising:

[0025] Data collection and preprocessing, collecting whole genome sequence data of respiratory viruses, daily new case number and transmission range of respiratory viruses;

[0026] Variation feature extraction, calculating nucleotide variation index , = ; translating nucleotide sequences into amino acid sequences, calculating amino acid variation coefficient , =α× ; calculating gene fragment recombination risk value , =β×(1- ); defining variation transmission acceleration factor =γ×( + + );

[0027] wherein, is the length of the virus sequence, is the maximum variation frequency for each site, n is the number of nucleotide types, is the number of times the jth nucleotide appears at the ith site; is the total number of all nucleotide types at the ith site in the same virus sequence group; α is the adjustment coefficient, is the length of the amino acid sequence, is the number of sites replaced; β is the adjustment coefficient, is the number of same bases, is the length of the recombination fragment; Transmission feature extraction, calculating transmission speed

[0028] and range expansion rate from the daily new case number and transmission range, constructing range expansion index = × ;

[0029] Constructing a deep learning model, adopting the architecture combining CNN and LSTM, and inputting , , , , ​、 As input, whether a new variant outbreak occurs as output label; the constructed model inputs new data and outputs predicted probability values.

[0030] From the above description, by extracting variation characteristics, the three nodes most susceptible to variation of respiratory viruses are extracted as variation characteristics, i.e., variation caused by instability of RNA itself, variation caused by errors in translation of nucleotides into amino acids, and variation caused by gene segment recombination, Quantify the variation activity of viruses at the nucleotide level, The functional change at the protein level is the probability of change, and the greater the probability of change, the more likely the protein will increase the transmission capacity, Assess the risk of new variants caused by gene recombination, and an increase in recombination events means an increase in the probability of variants; and add the transmission characteristics to the deep learning, the model can automatically learn the variation-transmission correlation pattern of new variants, and then predict the outbreak trend of respiratory virus variants, and then realize early warning; and various data are updated in real time, so the prediction model can realize rapid warning and response according to the latest prediction effect, and reduce passivity.

[0031] Further, the transmission speed = , wherein is the number of new cases on the tth day, is the number of new cases on the (t-1)th day.

[0032] Further, the range expansion rate = , wherein is the number of regions where respiratory viruses are transmitted on the tth day, is the number of regions where respiratory viruses are transmitted on the (t-1)th day.

[0033] Further, the architecture combining CNN and LSTM is adopted, CNN extracts local patterns of virus sequence variation characteristics, and LSTM processes the time sequence dynamic change trend of transmission data.

[0034] Further, when constructing the deep learning model, the cross-entropy loss function L is used to quantify the difference between the predicted probability of the model and the true label, L=- ;

[0035] , wherein N is the number of samples, is the true label of the kth sample; is the predicted probability of the model for the k samples, with a value range of 0-1.

[0036] From the above description, by minimizing L, the model output probability is forced to approach or focus on the true label, i.e. the prediction probability of positive samples (outbreak) tends to 1, and the negative sample tends to 0.

[0037] Further, when constructing the deep learning model, model optimization is also included, L1 and L2 regularization and Dropout technology are used to prevent overfitting, and the optimized loss function =L+ + ;

[0038] wherein, 、 is a regularization parameter, is a model parameter, is the number of parameters, and the generalization ability of the model is improved.

[0039] Further, the cross-entropy loss function L adjusts the parameters through the back propagation algorithm to minimize the loss function and improve the prediction accuracy.

[0040] Further, the constructed model inputs new data to output prediction probability values, and sets a warning threshold T through historical data and / or expert experience. When the prediction probability is greater than T, a warning signal is issued.

[0041] Further, α is a weight coefficient obtained by regression analysis according to historical data, which is used to quantify the influence degree of amino acid variation on virus transmission ability.

[0042] Further, β is an adjustment coefficient set according to different virus types.

[0043] Embodiment one

[0044] A respiratory virus new variant early warning method based on deep learning, comprising:

[0045] Data collection:

[0046] Whole genome sequence data: The whole genome sequence data of respiratory viruses can be collected from the China National Genomics Science Data Center, GISAID (Global Initiative to Share Influenza Data), NCBI (National Center for Biotechnology Information), and the like. These data contain the genetic information of viruses and are the basis for analyzing virus variation. For each virus sequence, record its collection time, place, host, and other metadata.

[0047] Transmission data: Collect respiratory virus-related transmission data, including case numbers, incidence rates, geographic transmission ranges, and the like. These data can be obtained from official channels such as national and municipal disease prevention and control centers and the World Health Organization.

[0048] Preprocessing: For the collected viral sequence data, use sequence alignment algorithms such as BLAST (Basic Local Alignment Search Tool) to remove duplicate sequences and low-quality sequences. For transmission data, perform data cleaning, fill in missing values, and handle outliers. For example, for an unusually high number of cases in a certain area, verify the data source or use statistical methods to correct it. If the data is processed, there is no need to process it again.

[0049] Variation feature extraction:

[0050] Nucleotide variation comprehensive index: Calculate the nucleotide variation frequency of each site in the viral sequence , = , and establish the nucleotide variation comprehensive index , = ;

[0051] where is the length of the viral sequence, is the maximum variation frequency for each site, n is the number of nucleotide types (RNA viruses commonly have four types (AGCU, adenine (A), guanine (G), cytosine (C), and uracil (U) four bases)), is the number of times the jth nucleotide appears at the ith site; is the total number of all nucleotide types at the ith site in the same viral sequence group;

[0052] Amino acid variation driving coefficient: Translate the nucleotide sequence into an amino acid sequence and calculate the amino acid substitution rate , = ; is the length of the amino acid sequence, is the number of sites that have been replaced; this substitution rate reflects the degree of variation at the protein level. Since amino acid variation can change the function of viral proteins and affect transmission, define the amino acid variation coefficient , =α× ; α is the adjustment coefficient, which is a weight coefficient obtained by regression analysis of historical data, used to quantify the impact of amino acid variation on viral transmission ability.

[0053] Gene fragment recombination risk value: Detect gene fragment recombination and calculate the similarity s of suspected recombination fragments to the reference sequence, s= ; is the number of identical bases, ​Recombination fragment length; construct gene fragment recombination risk value , = β × (1- ); β is the adjustment coefficient, which is set according to different virus types. The lower the similarity s, the higher the recombination risk value, which means the greater the possibility of new characteristics of the virus due to recombination.

[0054] Transmission feature extraction:

[0055] Mutation transmission acceleration factor: calculate the transmission speed by the daily new cases and the transmission range , = , where is the number of new cases on the tth day, is the number of new cases on the (t-1)th day.

[0056] Mutation range expansion index: calculate the range expansion rate , = , where is the number of regions where respiratory viruses are transmitted on the tth day, is the number of regions where respiratory viruses are transmitted on the (t-1)th day.

[0057] Define the mutation transmission acceleration factor = γ × ( + + ); this factor reflects the promoting effect of the degree of virus sequence variation on the transmission speed (i.e. the more easily the virus mutates, the more likely it is to produce a contagious strain, and thus the stronger the infectivity and the faster the transmission speed), the more significant the variation, the larger the acceleration factor value, and the more obvious the promotion of transmission speed.

[0058] Construct the range expansion index = × ; Combine the changes in transmission speed and range expansion to reflect the comprehensive expansion ability of the virus due to variation in the transmission scale, providing a more comprehensive basis for early warning.

[0059] Model construction: adopt the architecture combining CNN and LSTM, CNN extracts the local pattern of virus sequence variation characteristics, and LSTM processes the time series dynamic trend of transmission data.

[0060] Model training: extract the virus sequence features ( , , ), transmission features ( , ), and comprehensive features ( ) as input, and whether a new variant outbreak occurs as output label (1 for occurrence, 0 for non-occurrence). The difference between the model prediction probability and the true label is quantified using the cross-entropy loss function L, L = - [y log (p) + (1 - y) log (1 - p)], where y is the true label and p is the prediction probability. The parameters are adjusted through a back propagation algorithm to minimize the loss function and improve the prediction accuracy.

[0061] where N is the number of samples, is the true label of the kth sample; is the prediction probability of the model for the kth sample, with a value range of 0-1.

[0062] Model optimization: L1 and L2 regularization and Dropout technology are used to prevent overfitting, and the optimized loss function is = L + L1 + L2 + ;

[0063] where , is the regularization parameter, is the model parameter, is the number of parameters, and the generalization ability of the model is improved.

[0064] Early warning judgment: the trained model inputs new data to output prediction probability value, and sets the early warning threshold T through historical data and / or expert experience. When the prediction probability is greater than T, an early warning signal is issued.

[0065] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent transformation or direct or indirect application in related technical fields based on the content of the present application is also included in the patent protection scope of the present application.

Claims

1. A deep learning-based early warning method for new variants of respiratory viruses, characterized in that, Comprise: Data collection and preprocessing, collecting whole genome sequence data of respiratory viruses, daily new case number and spread range of respiratory viruses; Variant feature extraction, calculating nucleotide variant index , ; translating nucleotide sequence to amino acid sequence, calculating amino acid variant coefficient , = α × ; Computing gene segment recombination risk values , = β × (1- ); defining mutation propagation acceleration factor = γ × ( + + ); in, is the length of the viral sequence, To take the maximum variation frequency for each site, n is the number of nucleotide types, is the number of occurrences of the jth nucleotide at the i-th site; represents the total number of occurrences of all nucleotide types at the i-th site in the same viral sequence population; α is the adjustment coefficient, is the length of the amino acid sequence, for The number of sites replaced; β is the adjustment coefficient, For the same number of bases, is the length of the recombinant fragment; Transmission feature extraction, transmission speed is calculated by the daily new cases and the transmission range and the range expansion rate , the range expansion index is constructed = × ; A deep learning model is constructed, which adopts the architecture combining CNN and LSTM, takes , , , , , as input and whether a new variant outbreak occurs as output label; the constructed model inputs new data and outputs prediction probability value; the propagation speed = wherein is the number of new cases on day t, is the number of new cases on day t-1; The range expansion rate = wherein is the number of regions with respiratory virus transmission on day t, is the number of regions with respiratory virus transmission on day t-1; Using the architecture combining CNN and LSTM, CNN extracts local patterns of virus sequence variation characteristics, and LSTM processes the time series dynamic trend of spread data; α is the weight coefficient obtained by regression analysis according to historical data, which is used to quantify the influence degree of amino acid variation on virus transmission ability; β is the adjustment coefficient set according to different virus types. 2.The deep learning-based respiratory viral new variant early warning method of claim 1, wherein, When constructing a deep learning model, a cross-entropy loss function L is used to quantify the difference between the model's predicted probabilities and the true labels, L = -∑ylog(p) - (1 - y)log(l - p) ; wherein N is the number of samples, is the true label of the kth sample; is the predicted probability of the model for the kth sample, with a value range of 0-1. 3.The deep learning-based respiratory viral new variant early warning method of claim 2, wherein, Model optimization was also included when building the deep learning model, L1 and L2 regularization and Dropout technique were adopted to prevent overfitting, and the optimized loss function = L + 0.5 * L2 + ; wherein, , is a regularization parameter, is a model parameter, then the number of parameters is increased, which improves the generalization ability of the model. 4.The deep learning-based respiratory viral new variant early warning method of claim 2, wherein The cross entropy loss function L adjusts the parameters through the back propagation algorithm, minimizes the loss function and improves the prediction accuracy. 5.The deep learning-based respiratory viral new variant early warning method of claim 1, wherein The constructed model inputs new data and outputs prediction probability value, and sets the warning threshold T through historical data and / or expert experience. When the prediction probability is greater than T, the warning signal is issued.

Images