Methods for classifying biological sequences, computer programs, and computer implementations (identification of unknown genomes and recently known genomes).

Abstract

To provide a method, a computer program and a computer-implemented method for classifying biological sequences (identification of unknown genomes and closest known genomes).SOLUTION: A method for classifying biological sequences comprises: providing a training set related to a biological sequence; dividing the training set and the biological sequence into sequence fragments; extracting feature vectors from the sequence fragments of the training set and from the sequence fragments of the biological sequence; identifying feature vectors extracted from the sequence fragments of the biological sequence that correspond to feature vectors extracted from the sequence fragments of the training set; establishing a threshold value comprising an average degree of divergence of the feature vectors from the sequence fragments of the biological sequence, from the feature vectors from the sequence fragments of the training set; and detecting an anomaly among the extracted features of the sequence fragments of the biological sequence.SELECTED DRAWING: None

Description

[Technical Field]

[0001] This invention relates to general bioinformatics, and more specifically to the use of deep learning artificial intelligence for identifying unknown and recently known genomes. [Background technology]

[0002] The taxonomic classification of viral pathogens is crucial for public safety. For example, the ability to identify a virus or another pathogen often sets the course of action taken by government and public service agencies in terms of food safety, treatment options in clinical cases, and determining potential transmission chains in epidemiological investigations. The diversity of living organisms presents challenges in the ability to detect, combat, and classify pathogens. The decreasing cost and increasing accessibility of high-throughput sequencing have led to a rapid increase in the number of sequenced genomes. Improvements in genome sequencing result in increased data in terms of how to scale the data. While the sheer volume of data (GenBank had >100,000 genomes in November 2017) offers opportunities for the incorporation of diverse viruses, bacteria, fungi, and more, the sheer volume of data also presents challenges for developing methods to computationally derive meaningful features from this data. The use of standard bioinformatics classification software relies on large-scale databases that incorporate all incoming genomic data from organisms of interest. As the amount of data increases, the computational load of these methods becomes increasingly impractical for storage, memory, and computation time reasons. [Overview of the Initiative] [Problems that the invention aims to solve]

[0003] One drawback of deep learning classifiers is that it is generally impossible for the model to call a given observation an outlier in the data on which it was trained. A naive deep learning classifier will make predictions based on the best available candidate. This approach is problematic in the case of species, such as virus species that have not yet been discovered. The ability to detect outlier sequences, along with the ability to determine features that characterize viral genomes and populations in the context of diversity, is beneficial in terms of public health and food safety. [Means for solving the problem]

[0004] In one embodiment, the present invention relates to a method for classifying biological sequences, the method comprising: providing a training set of biological sequences relating to a biological sequence of interest; dividing the training set and the biological sequence of interest into sequence fragments; extracting feature vectors from the sequence fragments of the training set and from the sequence fragments of the biological sequence of interest; identifying feature vectors extracted from the sequence fragments of the biological sequence of interest that correspond to the feature vectors extracted from the sequence fragments of the training set; establishing a threshold including the mean degree of divergence of the feature vectors from the sequence fragments of the biological sequence of interest from the feature vectors from the sequence fragments of the training set, wherein the threshold sets a percentage relating to the classification of the sequence fragments of the biological sequence of interest; and detecting anomalies among the extracted features of the sequence fragments of the biological sequence of interest, wherein the anomalies are sequence fragments that fall outside the threshold.

[0005] In another embodiment, the present invention relates to a computer program product for classifying biological sequences, the computer program product comprising one or more computer-readable storage media and program instructions stored together on one or more computer-readable storage media, the program instructions comprising: program instructions for training a deep learning algorithm having a training set of sequence fragments relating to a biological sequence of interest, the deep learning algorithm comprising: program instructions for extracting feature vectors from the training set; program instructions for applying the trained deep learning algorithm to sequence fragments of a biological sequence of interest and obtaining feature vectors extracted from the sequence fragments; program instructions for clustering the feature vectors extracted from the training set and the feature vectors extracted from the biological sequence of interest and calculating the mean degree of divergence of the extracted feature vectors of the biological sequence of interest from the extracted feature vectors of the training set; program instructions for establishing a threshold based on the mean degree of divergence, the threshold setting a percentage relating to the classification of sequence fragments of a biological sequence of interest; and program instructions for detecting anomalies among the extracted features of sequence fragments of a biological sequence of interest, the anomaly being a sequence fragment outside the threshold.

[0006] In a further embodiment, the present invention relates to a computer implementation method for classifying biological sequences, the computer implementation method comprising: training a deep learning algorithm with a training set of sequence fragments associated with a biological sequence of interest, wherein the deep learning algorithm extracts feature vectors from the training set; applying the trained deep learning algorithm to sequence fragments of the biological sequence of interest to obtain feature vectors extracted from the sequence fragments; clustering the feature vectors extracted from the training set and the feature vectors extracted from the biological sequence of interest, calculating the mean degree of divergence of the extracted feature vectors of the biological sequence of interest from the extracted feature vectors of the training set; establishing a threshold based on the mean degree of divergence, wherein the threshold sets a percentage for classifying the sequence fragments of the biological sequence of interest; and detecting anomalies among the extracted features of the sequence fragments of the biological sequence of interest, wherein the anomalies are sequence fragments that fall outside the threshold.

[0007] Additional aspects, embodiments, or combinations thereof of the present invention will be provided in the detailed description of the invention below, without limitation. [Brief explanation of the drawing]

[0008] [Figure 1] This is a schematic representation of a one-dimensional convolutional neural network (1-D CNN) with maximum pooling capable of identifying and classifying 40 viral genomes from a sample of 150 viral genomes. [Figure 2] This is a schematic representation of the workflow in which the most recent viral genome is identified and classified into 40 viral genomes identified and classified by a 1-D CNN. [Modes for carrying out the invention]

[0009] A description of what is currently considered to be a preferred aspect or embodiment or combination thereof of the present invention is provided below. Any substitutions or modifications in function, purpose, or structure are intended to be covered by the appended claims. Unless otherwise explicitly indicated in the context, the singular forms “a,” “an,” and “the” as used herein and in the appended claims include multiple references. As used herein and in the appended claims, the terms “comprise,” “comprised,” “comprises,” or “comprising,” or any combination thereof, specify the presence of an expressed component, element, feature, or step or combination thereof, but do not exclude the presence or addition of one or more other components, elements, features, or steps or combination thereof.

[0010] As used herein, the term “deep learning” refers to artificial intelligence (AI) capabilities that mimic the workings of the human brain in processing and classifying data. Deep learning-based AI can learn from unstructured and unlabeled data. In operation, a deep learning-based AI program finds the relationships between inputs and outputs by assuming that any input x and any output y are related by a correlation or causal relationship and learning to approximate an unknown function (f(x)=y) between them.

[0011] As used herein, the term “neural network” refers to a deep learning classification algorithm modeled after the human brain, having a set of simulated neurons. Each neuron is a node connected to other nodes via links similar to biological axonal synapse-dendritic connections. Each link has weights that determine the strength of the influence of one node on another node. Neural networks learn (i.e., are trained) by processing examples, with each neural network containing known inputs and outputs, the inputs and outputs forming probabilistically weighted associations between them. During operation, unlabeled neural network groups of input data automatically extract features from the groups based on the similarity between exemplary inputs, cluster the groups by similar features, and classify the output data when a labeled dataset for training is available. Patterns recognized by neural networks are numerical and contained in vectors, which must be transformed. Examples of neural network vectors include, but are not limited to, images, sounds, text, time, or combinations thereof.

[0012] As used herein, the term “convolutional neural network” or “CNN” refers to a classification algorithm to which a neural network is applied to analyze data. A CNN may be one-dimensional (1-D), two-dimensional (2-D), or three-dimensional (3-D). Generally, 1-D CNNs are used for speech and text data, 2-D CNNs for image data, and 3-D CNNs for 3-D image data. A CNN uses a convolutional kernel, which is a learnable filter matrix used to extract features from input data and make predictions from the extracted features. The kernel matrix moves along the input data by a stride value. For example, if the stride value is 2, the kernel moves the data in the input matrix by 2 columns. In a 1-D CNN, the input data is one-dimensional, and the convolutional kernel slides one-dimensionally along the time axis. In a 2-D CNN, the input data is two-dimensional, and the kernel slides two-dimensionally along the height and width axes. In a 3-D CNN, the input data is three-dimensional, and the kernel slides in three dimensions along the height, width, and volume axes.

[0013] Generally, a convolutional layer convolves the input and passes the result to the next layer. While sufficiently connected, a feedforward neural network can be used to learn features and classify data. A CNN regularizes a multi-layered, sufficiently connected network (also known as a perceptron) where each neuron in one layer is connected to all neurons on the next layer. The nature of the sufficiently connected layers in a CNN leads to overfitted data, which must be regularized. A CNN regularizes by leveraging the hierarchical patterns of the data, building complex patterns using smaller, simpler patterns. A CNN consists of an input layer, hidden hidden layers, and an output layer. The hidden hidden layers perform convolutions. After passing through the convolutional layers, the data becomes abstracted into a feature map, which is the output of the convolutional kernel applied to the previous layer.

[0014] All references to deep learning made in this specification are to be understood to include neural networks and CNNs as a subset of deep learning-based AI.

[0015] As used herein, the term "max pooling" refers to an operation added to a CNN following an individual convolutional layer. Since max pooling is added after the convolutional layer, the output from the convolutional operation is the input to the max pooling operation. When added to a CNN, max pooling reduces the data output from the previous convolutional layer. When the max pooling output is input to a convolutional layer, the CNN is analyzing a broader area of the original input image, and thus reducing the number of parameters for analysis and the computational load on the CNN system. Max pooling also reduces overfitting inherent in CNN analysis by maintaining high data values while discarding lower values.

[0016] As used herein, the term "Y percentile" refers to a deep learning regression variable. In deep learning, the input variable is identified as X, the output value is identified as Y such that Y = f(X), where f(X) is the predicted value and Y is the parameter to be adjusted. For example, when the Y percentage for the predicted value is set to 75%, there is a 75% probability that the output value will be less than the predicted value and a 25% probability that the output value will be greater than the predicted value.

[0017] As used herein, the term "k-mer" refers to a string of nucleotides (i.e., ATGC) of length k contained within a DNA sequence. For illustrative purposes, the sequence "AATCGCC" has 7 monomers (A, A, T, C, G, C, C), 6 dimers (AA, AT, TC, CG, GC, CC), 5 trimers (AAT, ATC, TCG, CGC, GCC), 4 tetramers (AATC, ATCG, TCGC, CGCC), 3 pentamers (AATCG, ATCGC, TCGCC), 2 hexamers (AATCGC, ATCGCC), and 1 heptamer (AATCGCC).

[0018] As used herein, the term “one-hot encoding” refers to a computational method for pre-processed categorical variables such that, within a set of bits, one variable is identified as a single high (1) ("one-hot"), and all others are identified as low (0). One-hot encoding allows non-numerical input variables, such as nucleotides, to become numerical values ​​suitable for computational analysis in AI systems. Table 1 provides an example of a one-hot coded binary matrix for 7-MAR AATCGCC. As shown in the table, the nucleotide being analyzed is represented by the value (1), and the remaining nucleotides have the value (0). Table 1 [Table 1]

[0019] As used herein, the term “genome sequence” refers to the sequence of nucleotide bases present in the entire genome of any living organism. Examples of genomes expected in the methods described herein include, but are not limited to, animals, insects, plants, and microorganisms. Examples of microorganisms include, but are not limited to, viruses, bacteria, archaea, fungi, molds, protists, helminths, and prions. The term “gene sequence” refers to the sequence of nucleotide bases that completely or partially control the expression of one or more traits of any living organism. The term “protein sequence” refers to the amino acid sequence of all or part of a protein or peptide. The term “protein domain sequence” refers to a conserved portion of a protein sequence that is stable and folds independently of the rest of the protein sequence.

[0020] In the following discussion, viral sequences are used as illustrative sequences, but it should be understood that the present invention has applications to non-viral gene sequences when described herein.

[0021] Deep learning has the ability to reduce the computational load required for taxonomic classification. By learning predictive features of data, deep learning approaches can incorporate complex latent features that can only be approximated by the data width in simplified bioinformatics approaches. One drawback of deep learning classifiers is that it is generally impossible for the model to call a given observation an outlier in the data on which it was trained. A naive deep learning classifier will make predictions based on the best available candidate. This approach is problematic in the case of species, such as virus species that have not yet been discovered. The ability to detect outlier sequences, along with the ability to determine features that characterize viral genomes and populations in the context of diversity, is beneficial in terms of public health and food safety.

[0022] Modern methods of sequence-based classification for taxonomic identification are performed by comparing a query sequence to a reference database of ground-true-labeled sequences, either by locating or pseudo-locating. Locating methods are generally considered the most accurate and interpretable, while pseudo-locating methods offer greater speed at the expense of some interpretability and variability due to parameter selection. Both locating and pseudo-locating methods work well when comparing sequences of high similarity. When the query belongs to a species different from any species in the reference database, there is no mechanism for recognizing this situation. Machine learning classification (MLC), performed by deep learning models, offers an alternative to both locating and pseudo-locating methods that is faster and lighter in terms of memory and memory usage. Several deep learning-based classification models have been developed for bacteria; however, all of these models are trained to classify only a single region of the genome encoding the 16S gene. Because deep learning models rely on data structures that are invariant formats, such as matrices or tensors of fixed dimensions, the use of single genes that maintain a relatively narrow size distribution is useful for training deep learning models. However, single-gene classification is ineffective and unsuitable for general taxonomic classification.

[0023] For general taxonomic classification, selected genes must (i) be universally present in the genome of interest, (ii) be of similar length among all examples, and (iii) be representative of the diversity of taxonomic labels for training purposes (e.g., each taxonomic group must have a sequence that uniquely defines it with respect to genes, otherwise it would be indistinguishable from other taxonomic groups). The differences observed in gene sequences are the accumulation of mutation rates (which differ among organisms), natural selection (which is environment-dependent), and time, all of which are not constant among different genes. Genes with slow mutations provide poor resolution for closely related organisms, while genes with fast mutations provide poor resolution for distantly related organisms.

[0024] The deep learning workflow described herein implements a method for analyzing biological sequence fragments for genome classification and outlier detection. For outlier detection, the input is a biological sequence fragment of unknown origin, and the output is the nearest known biological genome that may share phenotypic characteristics with the biological species of unknown origin. The workflow has applications for genome classification, identification of variations within known genomes, and identification of the nearest class of an unknown species. The use of deep learning algorithms in the workflow makes it scalable for large datasets.

[0025] The first step in implementing the workflow is training a deep learning algorithm to perform deep feature extraction. In Figures 1 and 2, the deep learning algorithm is a 1-D CNN deep feature extractor 1 with two convolutional layers and two max pooling layers, which are inserted after each convolutional layer to reduce CNN overfitting. The 1-D CNN must be trained on sequence data so that it can extract features from biological sequence fragments in the form of feature vectors. For preprocessing, the sequence fragments to be analyzed are split into k-mers consisting of four bases, which are then structured into a binary matrix by one-hot coding. In Figures 1 and 2, 150-mer viral sequence fragments are used to train the 1-D CNN to identify and classify viral genomes. After structuring the 150-mer fragments into a binary matrix for preprocessing by one-hot coding, the 150-mer sequences are input to the 1-D CNN. After processing with two convolutional and max pooling layers, the 1-D CNN output becomes a 40-mar viral sequence fragment.

[0026] Features extracted from the training set are mapped to the genome feature space 2 for each fragment, while the corresponding centroid 3 is calculated. The genome feature map in Figure 2 shows 10 classes of clusters, each with its own calculated centroid. For each extracted feature mapped to the genome feature space, the mean degree of the intra-cluster density is calculated by equation (1).

number

[0027] Following the training of the convolution algorithm and mapping of the extracted features, a new unknown sample 4 may be introduced into the workflow for classification. Similar to the training samples, the new sample is pre-processed into a binary matrix, executed through a deep feature extractor, and mapped as such to a genomic feature map like a cluster. The intra-cluster density of the new sample is also calculated by Equation (1). Unlike the unlabeled training data, the new unknown sample is labeled.

[0028] To classify the new sample, an output score 5 is calculated as the difference between each of the input features in the new sample and the average value of the training variances of each feature. For an input by the total number M of features, the nearest class X to the input is each feature s j having an output score regarding. (2)s j =(d j -μ j ) Output s0, s1,... s M regarding Here, d is the Euclidean distance calculated among features j, and μ j is the average value of the distances regarding the training samples.

[0029] A threshold 6 is heuristically determined from the average degree of the variance of the new sample from the training samples to identify anomalies, and the average degree of divergence is the difference between the degree of the score and the degree of the inter-cluster density 6. The threshold θ is the percentile value of the training samples above the Y percentile (e.g., 75%), and Y is an adjustable parameter. For a sample input by feature M, the degree of density deg j is calculated by Equation (3). (3)deg j =s j -σ j Output deg0, deg1,... deg M regarding Here, s j σ is the output score for each feature j. j is the degree of intra-cluster density for each feature j. The score for anomalies is calculated by equation (4). (4)s j >θ*deg j * Output deg0 * ,deg1 * ...deg M * Regarding Here, divergence deg j * The degree of the score s j If the value is higher, the feature is then considered an anomalous fragment. (Values ​​θ and M) * The anomaly is heuristically determined by a linear search under the constraint that a specific number of k-mers in the training genome are outliers. For example, referring to the 40-mer output in Figures 1 and 2, setting the threshold for the convolution algorithm to 95% will allow the algorithm to run until the number of samples classified into anomalies is 95:5. The result of the 95% threshold is the classification of 38 different viral genomes (95%) from the 40-mer output by two viral genomes (5%) representing an unknown viral genome (Example 1). * If a genome is abnormal in more than a certain number of features, it is labeled as abnormal, and the most recent genome is determined to be the most recent class X (Example 2). Workflow output 7 is either (i) a viral genome species if the genome is recognized as belonging to the training (known) set, or (ii) the most recent known genome if it is an anomaly.

[0030] The present invention may be a system, method, or computer program product, or a combination thereof, at any possible level of technical detail of integration. The computer program product may include a computer-readable storage medium (or more mediums) storing computer-readable program instructions for causing a processor to perform aspects of the present invention.

[0031] A computer-readable storage medium can be a tangible device capable of holding and storing instructions for use by an instruction execution device. A computer-readable storage medium may, for example, be an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. A non-exclusive list of more specific examples of computer-readable storage media may also include, the following: portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital multipurpose disks (DVDs), memory sticks, floppy disks, mechanical encoding devices such as punched cards or grooved structures on which instructions are recorded, and any suitable combination of the above. As used herein, computer-readable storage media should not be interpreted as transient signals in themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses passing through optical fiber cables), or electrical signals transmitted through wires.

[0032] The computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to each computing / processing device or to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, or a wireless network or a combination thereof. The network may include copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers, or edge servers or a combination thereof. A network adapter card or network interface in each computing / processing device receives computer-readable program instructions from the network and transfers those computer-readable program instructions for storage in a computer-readable storage medium within the respective computing / processing device.

[0033] The computer-readable program instructions for performing the operations of the present invention may be either source code or object code written in any combination of one or more programming languages, including assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, configuration data for integrated circuits, or object-oriented programming languages ​​such as Smalltalk®, C++, and procedural programming languages ​​such as the "C" programming language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or wide area network (WAN), or the connection may be made to an external computer (for example, via the Internet using an Internet service provider). In some embodiments, for example, an electronic circuit including a graphics processing unit (GPU), a programmable logic circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA) may execute computer-readable program instructions by personalizing the electronic circuit using state information of the computer-readable program instructions in order to perform an aspect of the present invention.

[0034] Aspects of the present invention are described herein with reference to flowcharts or block diagrams or combinations thereof of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block in a flowchart or block diagram or combination thereof, and each combination of blocks in a flowchart or block diagram or combination thereof, can be implemented in computer-readable program instructions.

[0035] These computer-readable program instructions may be provided to a computer processor or other programmable data processing device in order to generate a machine. The instructions, executed via the processor of the computer or other programmable data processing device, then form means for implementing functions / operations specified in blocks or more of a flowchart or block diagram or a combination thereof. These computer-readable program instructions may be stored in a computer-readable storage medium that can instruct a computer, programmable data processing device, or other device or combination thereof to function in a particular manner, such that the computer-readable storage medium storing the instructions provides a product containing instructions that implements modes of functions / operations specified in blocks or more of a flowchart or block diagram or a combination thereof.

[0036] Computer-readable program instructions may be loaded onto a computer, another programmable device, or another device to perform a series of operational steps on that computer, another programmable device, or another device to generate a computer implementation process, so that the instructions executed on the computer, another programmable device, or another device implement functions / operations specified in one or more blocks of a flowchart or block diagram or a combination thereof.

[0037] The flowcharts and block diagrams in the drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of instructions comprising one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions described in a block may occur out of the order shown in the diagram. For example, two consecutively shown blocks may actually be implemented as a single step and executed simultaneously or substantially simultaneously in a manner that partially or entirely overlaps in time, or, depending on the function, the blocks may be executed in reverse order. It should also be noted that each block in a block diagram or flowchart or a combination thereof, and combinations of blocks in a block diagram or flowchart or a combination thereof, may be implementable by a special-purpose hardware-based system that performs the specified function or operation, or can perform a combination of special-purpose hardware and computer instructions.

[0038] The descriptions of various aspects, embodiments, or combinations thereof of the present invention are presented for illustrative purposes only and are not intended to be comprehensive or to limit the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the embodiments described. The terminology used herein has been selected to best describe the principles, practical applications, or technical improvements to the technologies available on the market of the aspects, embodiments, or combinations thereof, or to enable those skilled in the art to understand the aspects, embodiments, or combinations thereof disclosed herein.

[0039] The following examples are provided to those skilled in the art to provide a complete disclosure of how to use the aspects and embodiments of the invention described herein. While efforts have been made to ensure accuracy with respect to the variables, experimental errors and deviations should be taken into consideration. All components are commercially obtained unless otherwise indicated. Example 1 Classification of an Isolated Unknown Virus

[0040] Unknown viruses isolated (e.g., either novel viruses or examples of previously identified virus species) were classified by the biological classification processes described herein. For pretreatment, i.e., to construct a known viral feature space, 100 viral genomes belonging to 40 virus species were collected. The 100 viral genomes were represented using k-mers. The genome k-mers were divided into 150-base sequences with 25 base overlaps. Each sequence was converted into a binary image of size 4 (representing the number of bases A, C, G, and T) × 150 (number of k-mers, i.e., sequence length).

[0041] The sequences were processed by a 1D CNN (hereinafter referred to as “Neural Network”) as described herein, which mapped the genomes by the neural network, generating a set of 40 features for each sequence, the sequence of 40 features corresponding to 40 dimension vectors, (i) each viral genome belonging to each virus species was mapped to the 40 features, and (ii) the centroid was calculated for each known genome in the mapping space.

[0042] For each known virus species, the standard deviation of the distance from the centroid to the corresponding class was calculated. This was done by calculating the Euclidean distance between the centroid for a species and the k-mer extracted from the genome belonging to that species (i.e., from 40 mapped features).

[0043] Next, for the unknown virus species, the genome of the unknown virus species was split into k-mers and converted into a binary image. For each k-mer, a neural network was used to classify each k-mer as belonging to one of the known species or to the unknown species, and the output was the most likely species for the unknown virus species. The unknown sequences were then mapped to the 40 feature spaces described above. For each of the 40 features, a score was calculated, which represents the difference between the unknown genome feature and the mean of the training sequences belonging to the predicted class. The degree of variance was also calculated for each of the 40 features, and the degree of variance is the difference between the score and the degree of density, the latter of which is calculated by equation (3). The calculated degree of divergence for each of the 40 features was then compared to a threshold for the predicted class. The threshold was heuristically calculated based on the mean degree of divergence for the training samples. For the training samples, the distance from the mean was set to 95% (i.e., above the 75th percentile of the Y percentile), and the convolution algorithm was run until the number of samples classified as anomalies was 95:5, resulting in 38 viral genomes from a feature space of 40, each containing two viral genomes representing an unknown genome (and possible anomalies) (Figures 1 and 2). Example 2: Determining whether an unknown viral genome is a viral anomaly.

[0044] If, with respect to two unknown viral genomes from Example 1, one (or both) has a higher degree of variance than the score (calculated as previously described herein), then the unknown viral genome is considered an anomalous fragment. If the unknown viral genome is similar to the predicted species but significantly different from the known viral genomes included in the training set, then the unknown viral genome may be a mutation of a known virus, or a virus that shares certain characteristics with the predicted species.

[0045] As described in Example 1, once training is performed, a set of thresholds for known genomes is determined. For each viral genome sample, the trained neural network is used to classify each genome by assigning it to the most likely class. The variance score and degree are then calculated as previously described herein. Finally, the degree of divergence is compared to the corresponding threshold to assess whether the unknown viral genome is an anomaly.

Claims

1. A method for classifying biological sequences, performed by a computer, The computer provides a training set containing sequence fragments of known biological sequences, The computer trains a deep learning model using the training set, The computer divides the biological sequence of interest into sequence fragments, The computer inputs a sequence fragment of the biological sequence of interest into the deep learning model and extracts a feature vector of the sequence fragment of the biological sequence of interest. The computer maps the extracted feature vectors to the genome feature space and identifies clusters of biological sequences of interest based on the distance from the centroid calculated based on the feature vectors extracted from the training set. The computer establishes a threshold based on the mean degree of divergence of the feature vector from the sequence fragment of the biological sequence of interest from the feature vector from the sequence fragment of the training set. The computer detects anomalies from the extracted feature vectors of the sequence fragments of the biological sequence of interest, wherein the computer detects sequence fragments as anomalies if the score calculated as the difference between the distance between the feature vector and the centroid of the sequence fragment of the biological sequence of interest and the average value of the distances between the feature vector and the centroid extracted from the training set is greater than the threshold. A method that includes [a certain feature].

2. The method according to claim 1, wherein the sequence fragment of the biological sequence of interest whose score is less than or equal to the threshold is assigned to a known biological sequence.

3. The method according to claim 1, wherein the anomaly is assigned to the most recent known biological sequence.

4. The method according to claim 1, wherein the biological sequence of interest is selected from the group consisting of genome sequences, gene sequences, protein sequences, and protein domain sequences.

5. The method according to claim 4, wherein the genome sequence is a microbial genome sequence.

6. The method according to any one of claims 1 to 5, wherein the feature vector is extracted from sequence fragments of the biological sequence of interest by the deep learning model of the convolutional neural network.

7. The method according to claim 6, wherein the deep learning model of the convolutional neural network comprises a maximum pooling algorithm.

8. A computer program for classifying biological sequences, wherein the computer program comprises one or more computer-readable storage media and program instructions stored together on one or more computer-readable storage media, and the program instructions are A program instruction for training a deep learning model using a training set of sequence fragments relating to known biological sequences, wherein the deep learning model includes a program instruction for extracting feature vectors from the training set. Program instructions for applying the trained deep learning model to a sequence fragment of a biological sequence of interest and obtaining feature vectors extracted from the sequence fragment, A program instruction for mapping feature vectors extracted from the biological sequence of interest to the genome feature space, clustering the feature vectors extracted from the biological sequence of interest based on the distance from the centroid calculated based on the feature vectors extracted from the training set, and calculating the mean degree of divergence of the extracted feature vectors of the biological sequence of interest from the extracted feature vectors of the training set, A program instruction for establishing a threshold based on the aforementioned average degree of divergence, A program instruction for detecting anomalies from the extracted feature vectors of the sequence fragments of a biological sequence of interest, comprising: a program instruction for detecting sequence fragments as anomalies if the score calculated as the difference between the distance between the feature vector and the centroid of the sequence fragment of the biological sequence of interest and the average value of the distances between the feature vector and the centroid extracted from the training set is greater than the threshold; Computer program.

9. The computer program according to claim 8, wherein a sequence fragment of the biological sequence of interest whose score is less than or equal to the threshold is assigned to a known biological sequence.

10. The computer program according to claim 8, wherein the anomaly is assigned to the most recent known biological sequence.

11. The computer program according to claim 8, wherein the biological sequence of interest is selected from the group consisting of genome sequences, gene sequences, protein sequences, protein domain sequences, and combinations thereof.

12. The computer program according to claim 11, wherein the genome sequence is a microbial genome sequence.

13. The computer program according to any one of claims 8 to 12, wherein the deep learning model comprises a deep learning model of a convolutional neural network.

14. The computer program according to claim 13, wherein the deep learning model of the convolutional neural network comprises a maximum pooling algorithm.

15. A computer implementation method for classifying biological sequences, wherein the computer implementation method is The steps include establishing a training set of sequence fragments related to the sequences of known virus species, The step of training a deep learning model with the aforementioned training set, wherein the deep learning model extracts feature vectors from the sequence fragments of the training set. The step of applying the trained deep learning model to a sequence fragment of an unknown virus species, wherein the deep learning model extracts feature vectors from the sequence fragment of the unknown virus species. The steps include: mapping the feature vectors extracted from the sequence fragments of the unknown virus species to the genome feature space; clustering the feature vectors extracted from the unknown virus species based on the distance from the centroid calculated based on the feature vectors extracted from the training set; and calculating the average degree of divergence of the extracted feature vectors of the unknown virus species from the extracted feature vectors of the training set. A step of establishing a threshold based on the aforementioned average degree of divergence, A step of detecting anomalies from the extracted feature vectors of the sequence fragments of the unknown virus species, wherein sequence fragments whose score, calculated as the difference between the distance between the feature vector and the centroid of the sequence fragment of the unknown virus species and the average value of the distances between the feature vector and the centroid extracted from the training set, is greater than the threshold, are detected as anomalies. A computer implementation method comprising the above.

16. The computer implementation method according to claim 15, wherein the sequence fragment of the unknown virus species whose score is less than or equal to the threshold is assigned to a known virus species sequence.

17. The computer implementation method according to claim 15, wherein the anomaly is assigned to the most recent known virus species sequence.

18. The computer implementation method according to claim 15, wherein the deep learning model is a deep learning model of a convolutional neural network.

19. The computer implementation method according to any one of claims 15 to 18, wherein the deep learning model comprises a maximum pooling algorithm.

Images