Using deep learning for gene mutation detection

Abstract

The present disclosure provides methods, systems and computer program products for classifying candidate mutations detected in sequencing data (especially suboptimal sequencing data) using deep learning models. These methods, systems and programs provide improved efficiency, accuracy and speed in identifying mutations from a wide range of sequencing data.

Description

Technical Field

[0001] The present disclosure generally relates to methods and systems for detecting mutations in low-coverage sequencing data using deep learning. Background Art

[0002] The human genome is a complete set of human nucleic acid sequences, encoded as DNA in 23 pairs of chromosomes in the cell nucleus and a small DNA molecule within a single mitochondria. The human genome consists of 6 billion base pairs, which include DNA genes that encode proteins and non-coding DNA. The first complete individual human genome sequence was published by the Human Genome Project in 2001. Currently, thousands of human genomes have been sequenced and many more have been mapped at lower levels of resolution. The resulting data are used worldwide in biomedicine, anthropology, forensics, and other branches of science. It is widely believed that genome research will lead to advances in disease diagnosis and treatment and new insights in many areas of biology, including human evolution.

[0003] However, understanding the genetic basis of disease requires accurate and rapid determination of an individual's genomic DNA sequence, down to the level of a single base pair. This level of resolution in DNA sequencing allows for the identification of natural variations in sequence that occur between different individuals. These individual sequence variation sites, often referred to as single nucleotide variations (SNVs) or single nucleotide polymorphisms (SNPs), are present in each genome and are potentially critical to the use of genomic sequence information across a whole range of applications. SNPs refer to variations in genomic sequence positions that occur in different populations. For example, at a specific human genomic position, the nucleotide C may occur in most humans, but in a minority of individuals, A may occur at the same genomic position. These two different nucleotides are called alleles at a specific position of the genome. It is estimated that on average one SNP occurs every 300 bp in the human genome, resulting in an average human genome sequence with approximately 10 million SNPs relative to the reference genome.

[0004] SNPs generally refer to variations found in the genome of human germ cells. SNVs are a more general term that can include single nucleotide variations at specific sites that may occur in the genome of somatic cells. Cancer cells represent a highly studied class of somatic cells that include SNVs that are thought to be critical to their pathological phenotype and diagnosis.

[0005] DNA sequencing technology has advanced significantly since the first individual human genome was sequenced in 2000, at an estimated cost of $2.7 billion. Currently, the most advanced high-throughput technology, often referred to as “next-generation sequencing” (NGS), has enabled large-scale sequencing of plant and animal genomes and has enabled the process of determining a whole genome sequence to be achieved in as little as a week at a cost of approximately $1,000.

[0006] NGS technology typically works by performing millions of individual overlapping sequence reactions simultaneously, each of which produces short sequences or "reads" that are hundreds of base pairs in length. Determining the sequence requires multiple sequence reads covering each base position, and typically, it is desirable to have 30 times the read redundancy (i.e., "30X coverage") at each base position. Therefore, NGS generates large data sets consisting of a large number of sequence reads. However, the error rate of sequence reads provided by NGS technology is relatively high, ranging from -0.1-10%. Therefore, processing NGS requires highly complex statistical error analysis for each sample. The complexity of post-reaction processing of NGS sequence reads to account for and minimize errors brings great difficulties to the process of correctly identifying or "calling" mutations or variations in genomic sequences.

[0007] The desire to distinguish true mutations from errors present in NGS data sets has led to the development of methods and software tools for this purpose. Two widely used software tools for identifying mutations (such as SNPs, SNVs, insertions and deletions) from NGS data sets are Genome Analysis ToolKit or "GATK" (available at software.broadinstitute.org / gatk / ) and SAMtools (Li et al., "The SequenceAlignment / Map format and SAMtools," Bioinformatics 25(16): 2078-2079(2009)). These widely used mutation identification software tools use "classical" methods to align sequence reads, and perform bioinformatics analysis and machine learning modeling on the aligned reads to identify mutations. The "classical" bioinformatics and machine learning components of these software tools require labor-intensive "handcrafting" of data features, which greatly limits their ability to be promoted on sequencing data sets obtained from different types of sequencing machines and / or data with different coverage depths. Furthermore, the mutation calling accuracy of classical tools (such as GATK) significantly deteriorates when applied to suboptimal (such as low coverage) sequencing datasets.

[0008] DeepVariant is a deep learning-based software tool developed to improve classical variant calling tools (e.g., see Poplin et al., “Creating auniversal SNP and small indel variant caller with deep neural networks,” bioRxiv 092890; doi: doi.org / 10.1101 / 092890; published Dec. 14, 2016). DeepVariant uses a convolutional neural network (CNN) to call genetic variants in aligned NGS read data by learning the statistical relationship between the cumulative image of reads around putative variants and the ground truth genotype calls. DeepVariant has been shown to outperform GATK on a benchmark whole genome at 30x sequencing depth and generalizes better to data from sequencers that were not used during training. The accuracy of DeepVariant for calling variants using low-coverage sequencing data is unknown.

[0009] Therefore, there remains a need for improved methods that reduce the cost, processing time, and sample requirements for high-precision mutation detection from sequencing data. Summary of the invention

[0010] The present disclosure provides methods, systems and computer program products that classify candidate mutations detected from high-throughput nucleic acid sequencing data using deep learning models. These methods, systems and programs provide improved efficiency, accuracy and speed for classifying candidate mutations (e.g., identifying SNPs) detected in sequencing data from a series of sequencing machines. A surprising result of the present disclosure is that by extracting and encoding selection information representing candidate mutations (e.g., information from SAMtools or GATK software tools), encoding the selection information in a data structure that associates a copy of the reference sequence with each read sequence, and processing the encoded information with a deep learning model, accurate classification of candidate mutations (i.e., whether they are true mutations) can be obtained even from suboptimal (e.g., low coverage) sequencing data. The effect of this process is to greatly reduce the time and cost of sequencing germline and somatic genomes, and accurately classify the presence and identity of mutations found therein.

[0011] In some embodiments, the present disclosure provides a method comprising:

[0012] encoding information representing a candidate mutation, wherein the information includes an absolute position, a reference sequence, a number of reads (N), and a sequence of the reads; and

[0013] The encoded information is processed with a deep learning model, which classifies the candidate mutations.

[0014] In some embodiments, the method is performed, the encoding associates the reference sequence with each read sequence, whereby the deep learning model processes the reference sequence jointly with each read sequence. In some embodiments, the encoding enhances the reference sequence, enhances the read sequence, and / or aligns the enhanced reference sequence with each read sequence. In some embodiments, the encoded information also includes read mapping quality values ​​and / or read base quality values. In some embodiments, the reference sequence and the read sequence are encoded as a 2D Boolean matrix; optionally, wherein the read mapping quality values ​​and / or the read base quality values ​​are associated with the read sequence 2D Boolean matrix.

[0015] In some embodiments, the method is performed, wherein processing with a deep learning model comprises: transforming the encoded information in parallel with multiple CNNs with different kernel sizes, and combining the outputs into a tensor. In some embodiments, combining the outputs comprises: generating a ranked score and using the score to generate a tensor for selecting TopK. In some embodiments, processing with a deep learning model further comprises: transforming the tensor with a CNN comprising a one-dimensional convolutional layer followed by multiple fully connected layers, wherein the last fully connected layer outputs a classification of the candidate mutation.

[0016] In some embodiments, the method further comprises: training the deep learning model, wherein the training comprises:

[0017] encoding information representing a suboptimal candidate mutation, wherein the encoded information includes an absolute position, a model reference sequence, a number of reads (N), and a read sequence;

[0018] processing the encoded information with a deep learning model, the deep learning model classifying the suboptimal candidate mutations; and

[0019] By adjusting the parameters of the deep learning model, the error in classification of the suboptimal candidate mutation relative to the model reference sequence and the ground truth candidate mutation is minimized.

[0020] In some embodiments of the method, the method further comprises: performing training using a loss function, minimizing errors in classification. In some embodiments, the suboptimal candidate mutations are from low coverage sequencing data, optionally from simulated low coverage sequencing data generated from a model sequencing data set.

[0021] In some embodiments, the present invention provides a system comprising:

[0022] A processor, a memory device, and a classification engine executable on the processor according to software instructions stored in the memory device, wherein the classification engine is configured to:

[0023] encoding information representing the candidate mutation, wherein the encoded information includes an absolute position, a reference sequence, a number of reads (N), and a read sequence; and

[0024] The encoded information is processed with a deep learning model, which classifies the candidate mutations.

[0025] In some embodiments of the system, the encoded information has a reference sequence associated with each read sequence, whereby the deep learning model jointly processes the reference sequence with each read sequence. In some embodiments, the encoding enhances the reference sequence, enhances the read sequence, and / or aligns the enhanced reference sequence with each read sequence. In some embodiments, the encoded information also includes a read mapping quality value and / or a read base quality value. In some embodiments, the reference sequence and the read sequence are encoded as a 2D Boolean matrix; optionally, wherein the read mapping quality value and / or the read base quality value are associated with the read sequence 2D Boolean matrix.

[0026] In some embodiments of the system, the deep learning model includes transforming the encoded information in parallel with multiple CNNs with different kernel sizes, and combining the outputs into a tensor. In some embodiments, combining the outputs includes: generating a ranked score and using the score to generate a tensor for selecting TopK. In some embodiments, processing with the deep learning model also includes: transforming the tensor with a CNN, the CNN including a one-dimensional convolutional layer followed by multiple fully connected layers, wherein the last fully connected layer outputs a classification of the candidate mutation.

[0027] In some embodiments of the system, the system further comprises a training engine executable on the processor according to software instructions stored in the memory device, wherein the training engine is configured to:

[0028] Encoding information representing a suboptimal candidate mutation, wherein the encoded information includes an absolute position, a model reference sequence, a number of reads (N), and a suboptimal read sequence;

[0029] processing the encoded information with a deep learning model, the deep learning model classifying the suboptimal candidate mutations; and

[0030] The error in classification of the suboptimal candidate mutations relative to the ground truth candidate mutations of the model reference sequence is minimized by adjusting the parameters of the deep learning model.

[0031] In some embodiments of the system, the training engine is in communication with the classification engine.

[0032] In some embodiments, the present disclosure further provides a non-transitory computer-readable medium, comprising instructions for classifying candidate mutations, wherein when the instructions are executed by a processor, the processor performs one or more steps, comprising:

[0033] encoding information representing the candidate mutation, wherein the encoded information includes an absolute position, a reference sequence, a number of reads (N), and a read sequence; and

[0034] The encoded information is processed with a deep learning model, which classifies the candidate mutations.

[0035] In some embodiments, the non-transitory computer-readable medium further comprises instructions for training the deep learning model, wherein the training comprises:

[0036] encoding information representing a suboptimal candidate mutation, wherein the encoded information includes an absolute position, a model reference sequence, a number of reads (N), and a read sequence;

[0037] processing the encoded information with a deep learning model, the deep learning model classifying the suboptimal candidate mutations; and

[0038] By adjusting the parameters of the deep learning model, the error in classification of the suboptimal candidate mutation relative to the model reference sequence and the ground truth candidate mutation is minimized.

[0039] In various embodiments of the non-transitory computer-readable medium, various features useful in the methods and systems for classifying candidate mutations described above and elsewhere herein, including deep learning architecture features and training features, are contemplated as also being features in the non-transitory computer-readable medium. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Depicted is a flow chart overview of exemplary operations for classifying candidate mutations identified in sequencing data, according to an embodiment of the present invention.

[0041] Figure 2 Depicted is a block diagram of exemplary operations for encoding information representing candidate mutations in accordance with an embodiment of the present invention.

[0042] Figure 3 A schematic diagram depicting an exemplary process for encoding information representing an exemplary single candidate mutation.

[0043] Figure 4A schematic diagram depicting exemplary information encoding a set of candidate mutations (denoted as Candidate Mutation 1 to Candidate Mutation Z) obtained from a high-throughput sequencing dataset.

[0044] Figure 5A and Figure 5B Depicted is a block diagram of an exemplary deep learning model architecture for processing encoded information representing candidate mutations in accordance with an embodiment of the present invention. Figure 5A A first neural network module is described, which includes multiple parallel CNNs with different kernel sizes, whose outputs are combined into a single tensor. Figure 5B A second neural network module is depicted, comprising a CNN and a fully connected layer, capable of processing the output tensor of the first module, which represents the processed features of the encoded information, and producing an output classification of candidate mutations.

[0045] Figure 6 A block diagram of an exemplary system for classifying candidate mutations identified in sequencing data using a deep learning model and training the deep learning model according to an embodiment of the present disclosure is described.

[0046] Figure 7 A system is described in which processes, architectures, and / or functionality for classifying candidate mutations identified in sequencing data using a deep learning model according to embodiments of the present invention may be implemented. DETAILED DESCRIPTION

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention relates. In general, the terms used to describe the techniques and procedures described herein are those understood and commonly used by those skilled in the art, such as common techniques and methodologies for high-throughput (or next-generation) sequencing and identifying mutations (e.g., SNPs or SNVs) from such types of sequencing data. Such common techniques and methods for sequencing data acquisition and mutation analysis are described, for example, in Goodwin et al., "Coming of Age: A Decade of Next-Generation Sequencing Technology" Nat Rev Genet. June 2016; 17(6): 333-51; and Nielsen et al., "Genotype and SNP Identification from Next-Generation Sequencing Data," Nat. Rev. Genet. June 2011; 12(6): 443-451.

[0048] For purposes of interpreting the present disclosure, where appropriate, terms used in the singular will also include the plural, and vice versa. For the descriptions provided herein and in the appended claims, the singular forms "a" and "an" include plural references unless the context clearly indicates otherwise. The use of "comprise", "comprises", "comprising", "include", "includes" and "including" are interchangeable and are not meant to be limiting. It should be further understood that when the description of various embodiments uses the term "comprising", those skilled in the art will understand that in certain specific cases, the language "consisting essentially of" or "consisting of" may be used to describe an embodiment alternatively. When describing a numerical range, unless the context clearly dictates otherwise, it should be understood that each intermediate integer of the numerical value between the upper and lower limits of the range and each intermediate integer of the numerical value, unless the context clearly dictates otherwise, as well as any other specified value or intermediate numerical value within the specified range, are included in the present invention. The upper and lower limits of these smaller ranges can be independently included in the smaller ranges and are also included in the present invention, but are subject to any specific exclusive restrictions in the specified range. Therefore, when the specified range includes one or two of these restrictions, the range that excludes any one or two of these included restrictions is also included in the present invention. For example, "1 to 50" includes "2 to 25", "5 to 20", "25 to 50", "1 to 10" and the like.

[0049] It will be appreciated by those of ordinary skill that the methods, systems, and computer-readable media products described herein provide many advantageous technical effects, including improving the accuracy, speed, compactness, and overall efficiency of detecting mutations in sequencing data using deep learning models, even for suboptimal low coverage data. It will also be appreciated that the following specification is not intended as an extensive overview, and therefore, concepts may be simplified for the sake of clarity and brevity.

[0050] All publications, patents, patent applications, and other documents cited in this disclosure are hereby incorporated by reference in their entirety for all purposes, to the same extent as if each individual publication, patent, patent application, or other document was individually indicated to be incorporated by reference for all purposes.

[0051] Classifying candidate mutations using deep learning

[0052] The present disclosure describes methods and systems for using deep learning disclosed herein, which can be used to classify mutation identification from any type of nucleic acid sequencing data. However, it is expected that these methods and systems will find the greatest advantage when they are applied to classify candidate mutations in a large number of overlapping sequence read data sets obtained from NGS experiments. A series of NGS technologies are known in the art, and it is expected that the methods and systems based on deep learning disclosed in the present disclosure can be used to classify mutations from sequencing data obtained from any of them. For a review of NGS technology, see, for example, Goodwin et al., "The Coming of Age: A Decade of Next Generation Sequencing Technology", Nat Rev Genet. June 2016; 17(6): 333-51.

[0053] As mentioned above, NGS technology works by performing millions of sequencing reactions in parallel, each of which results in millions of overlapping sequences or "reads", which are usually several hundred base pairs (bp) in length. In NGS experiments, overlapping sequence reads lead to useful data redundancy. Currently, the standard for redundancy required for accurate sequences is "30-fold coverage", which means that an average of 30 separate sequence reads cover each base position of the sampled genome. Unfortunately, the error rate of a large number of parallel sequence reads is relatively high, about 0.1% to 10%. Therefore, further post-processing analysis of the errors is required to provide an accurate overall sequence for each sample. In addition, the 30-fold coverage is the average of the entire sequencing data set. There are many positions in the data set with much fewer reads (e.g., only 10), while other positions have much more reads (e.g., 1000). When trying to use NGS sequencing data to accurately detect and confirm whether there are mutations in nucleic acid sequences, the high error rate in a large number of sequence reads and the uneven coverage level create a particularly challenging problem. As used herein, "mutation" refers to any variation, change or difference in a nucleic acid sequence relative to a reference sequence, including, for example, single nucleotide variation (SNV), single nucleotide polymorphism (SNP), insertion, deletion or any other difference.

[0054] Typically, in high-throughput sequencing data, differences or changes in base (e.g., A, C, G, or T) identities will occur in some but not all of the multiple sequence reads that overlap or cover a specific position. These differences detected at a certain position in a set of sequence reads may be due to the presence of actual mutations in the sample sequence (i.e., true positives), or simply due to sequence read errors (i.e., false positives) caused by noise or some other experimental errors. Differences detected but not confirmed at sequence positions relative to a reference sequence are called "candidate mutations." "Candidate mutation classification" includes determining whether a mutation observed at a specific position relative to a reference sequence is a true mutation, and optionally, also determining the mutation type (e.g., insertion, deletion, SNP). In a typical sequencing data set spanning a genome, up to millions of candidate mutations can be observed ("genome" refers to a complete genome or a portion of a genome, such as an exome). When using a suboptimal sequencing data set, the problem of classifying these numerous candidate mutations is amplified. "Suboptimal" refers to quality below optimal quality. Therefore, suboptimal candidate mutations are candidate mutations obtained from sequencing data with quality below optimal. Such data may come from any source that may compromise the quality of the sequencing data, including the use of fewer sequencing reads (e.g., low coverage datasets), a limited number of samples, and / or any source that can introduce noise or errors. However, the ability to use suboptimal sequencing data is desirable because such data can be obtained rapidly, allowing for faster and cheaper sequencing of the complete genomes of human individuals. Therefore, much work remains to be done on methods and systems for improving candidate mutation classification.

[0055] The present disclosure provides methods and systems for classifying candidate mutation information from high-throughput sequencing data using a deep learning model. Specifically, the method and system can be used to classify candidate mutation information obtained from a high-throughput sequencing data set, and the high-throughput sequencing data set includes suboptimal sequencing data, such as low coverage data obtained with less than 30X sequencing depth. By allowing accurate candidate mutation classification with low coverage data, the deep learning-based methods and systems disclosed in the present disclosure can promote faster, more efficient and more cost-effective nucleic acid sequencing resources. Therefore, faster and more efficient sequencing can make nucleic acid sequencing more widely used in medical applications such as personalized medicine and diagnosis.

[0056] The deep learning models and related techniques used in the disclosed methods are based on the use of neural networks, such as convolutional neural networks (CNNs), in computer-based systems. Such deep learning models have been successfully used to extract features and classify data in a wide range of applications, most notably image, language, and speech recognition. In order to effectively process candidate mutations from high-throughput sequencing data sets using deep learning models, appropriate sequencing data is encoded into a data structure that facilitates the use of neural networks.

[0057] Figure 1 An overview flow chart of exemplary operations for classifying candidate mutations from sequencing data (especially suboptimal sequencing data) is shown, as applied using the deep learning methods and systems described in the present disclosure. Candidate mutation information 101 is generated from standard sequencing data using any software tool known in the art for processing sequencing data and performing variant identification (call). The two most widely used software tools are GATK (i.e., "Genome Analysis Toolkit"), which is available from the Broad Institute (software.broadinstitute.org / GATK / ) and SAMtools (Li et al., "Sequence Alignment / Mapping Format and SAMtools," Bioinformatics 25 (16): 2078-2079 (2009)). These software tools perform "classic" variant identification by aligning numerous sequence reads from a target sample to a reference genome, and then using bioinformatics analysis and machine learning modeling of aligned reads to generate candidate mutations.

[0058] Candidate mutation information useful in the methods and systems of the present invention can be generated from standard sequencing data using standard well-known software for identifying mutations. For example, a data set of high-throughput sequence reads can be analyzed using corresponding reference sequences and classical variant identification software tools (such as GATK and SAMtools) to generate candidate mutations. In some embodiments, the sequencing data used in the method is a low-coverage data set with less than the optimal number of sequence reads at each candidate mutation position. Low-coverage data sets useful in the candidate mutation classification method disclosed herein may include data sets with a coverage of the target genome sequence of less than 30 times, less than 20 times, less than 15 times or even lower.

[0059] In some embodiments, the candidate mutation information 101 used in the methods and systems of the present disclosure includes a subset of standard information found in a file format generated by a classic variant calling software tool (such as GATK and SAMtools, such as VCF and / or BAM files). For example, candidate mutation information can be extracted from a variant calling format (VCF) file. A VCF file includes information about the position and base recognition of a specific candidate mutation, the absolute position in the genome (e.g., detailing the chromosome site), and the quality of the base recognition in the read sequence (BQ). The VCF file also includes a compressed sequence alignment / mapping format (BAM) file, which contains information about read mapping / alignment and its quality, also referred to herein as read mapping quality (MQ). VCF and BAM file formats are well known in the art, and a complete description of each format is publicly available online, for example, at github.com / samtools / hts-specs.

[0060] As mentioned above, not all candidate mutation information generated by standard variant calling software tools is used in the deep learning method of the present disclosure. One advantage of the method of the present disclosure is that selected data representing candidate mutation information 101 is used to generate encoded information 102, which can then be processed by the deep learning model 103 to generate an output of candidate mutation classification 104. The selection and structure of the encoded information 102 is Figure 2 The architecture of the deep learning model is described in more detail in Figure 3 Detailed description is given in .

[0061] Figure 2 A block diagram of exemplary candidate mutation information 101 is shown, which is extracted and then encoded into a data structure of encoded information 102 so that it can be processed by a deep learning model for the method and system of the present disclosure. In some embodiments, the extracted and encoded information includes a reference sequence 201, a read sequence 202 (which includes the number of reads N), and an absolute position 205 of the candidate mutation. The absolute position is usually represented by a digital "address" or "index", which indicates a clear position in the genome (or other reference sequence) in the candidate mutation set (e.g., "chromosome 1:10009"). The set of N selected read sequences 202 is those that overlap with the candidate mutation position. Similarly, the reference sequence 201 corresponds to a genomic sequence that is equidistant from upstream and downstream around the mutation site. Therefore, the absolute position 205 defines a set of N read sequences 202 and corresponding reference sequences 201.

[0062] The extraction and encoding of the absolute position 205 , the reference sequence 201 , and the N read sequences 202 constitute a basic set of candidate mutation information 101 to be included in the encoded information 102 representing the candidate mutation.

[0063] Due to the high incidence of sequencing errors (e.g., insertions and / or deletions) in NGS, sequence reads 202 obtained from candidate mutation generation software (e.g., GATK) are also often not aligned with reference sequences 201. In order to encode candidate mutation information for more accurate and efficient processing by a deep learning model, in some embodiments, encoding includes preparing an enhanced reference sequence 206 and an enhanced read sequence 207. The sequence is enhanced by adding insertions at the positions of the reference sequence and the N read sequences, resulting in the best alignment of the candidate mutation position and all sequences of equal length. The process and software for preparing the best alignment of read sequences with reference sequences are well known in the art. The alignment process will typically result in each of the enhanced read sequence 207 and / or the enhanced reference sequence 206, which includes one or more insertions relative to the input read and reference sequences 202 and 202. The process of enhancing the sequence results in an enhanced reference sequence and an enhanced read sequence of the same length that are optimally aligned with the candidate mutation site of the site. This encoding of the enhanced reference sequence and read sequence further facilitates the processing of the encoded information 102 by the deep learning model 103, especially the efficiency of feature learning and extraction of the neural network.

[0064] Reference sequences and read sequences extracted from candidate mutation information generated by software tools such as GATK are represented as strings of A, C, G, and T, or integer numbers representing A, C, G, and T. In order to facilitate more efficient processing of the deep learning model of the present disclosure, in some embodiments, the enhanced reference sequence and the enhanced read sequence are encoded as 2D Boolean matrices 208 and 209. More specifically, the Boolean matrix includes five rows, representing A, C, G, T, or insertion, and a number of columns corresponding to the length of the sequence, and each base position is represented by a column. Therefore, the reference sequence and the N read sequences are represented by a two-dimensional Boolean matrix of the same dimension: five rows represent the number of columns of the sequence bp length. Whether there is A, C, T, G, or insertion at each sequence position is represented by a grid of "1" or "0". Therefore, the 5 rows and columns of each Boolean matrix will include a single "1" and four "0". Such a two-dimensional Boolean matrix is ​​quickly and efficiently converted by a neural network such as a CNN, which greatly improves the feature learning and extraction process for classifying mutation candidates.

[0065] It is also contemplated that certain quality features associated with each of the N individual sequence reads, as well as the individual base calls in each read, may optionally be included in the encoded information 102. Thus, in some embodiments, a read mapping quality score (MQ) 204, and / or a base quality score (BQ) 203 for each base in the read may also be encoded in the encoded information along with the read sequence 202. Including the quality scores in the encoded information along with the read sequence provides additional parameters useful for processing by the deep learning model 103, as described below.

[0066] Figure 3 A schematic diagram is described that illustrates an exemplary process of encoding information representing a candidate mutation. The reference sequence is a 15 bp sequence centered on the C base of the candidate mutation site at the absolute position "chr1:10009". Below the reference sequence, a stack of six read sequences is described, which are 13 to 15 bp and have C, T, or G at the candidate mutation site, depending on the alignment. The six read sequences were optimally aligned using the sequence standard alignment tool and enhanced with insertions (indicated by a dash "-") so that all six read sequences were 18 bp in length and the candidate mutation site was located at the 10th position of the sequence. The reference sequence was also enhanced with insertions and similarly aligned. In the enhanced alignment of the read sequences, three have C at the mutation site and three have T at the mutation site. The reference sequence has C at the site. Therefore, the enhanced alignment illustrates the related mutation classification problem: whether the read indicates the presence of a T mutation at the site, or whether the three Ts detected are errors or noise. In order to provide more convenient processing for deep learning models, the aligned sequences can be further encoded into a two-dimensional Boolean matrix. As Figure 3 As shown, the two-dimensional Boolean matrix has five rows to allow for four possible bases and insertions at each position. The number of columns in the matrix corresponds to the length of each sequence (eg, 18 bp). Figure 3 The cells of the two-dimensional matrix encoded in are described as black or white, but can also be represented as 1 or 0. Since read 1 and read 2 shown have different bases at the 8th position, the two-dimensional Boolean matrix patterns they encode are also different at the cells corresponding to the 1st and 2nd rows of the 8th column. Figure 3 The two-dimensional Boolean matrix for reads 3-6 is not described, but will have similarly different patterns corresponding to their different coding sequences. Figure 3 The black and white two-dimensional Boolean encoding of the sequences described in highlights how this type of encoding can facilitate the feature extraction process using neural networks.

[0067] Figure 3Also depicted are the read mapping quality (MQ) values ​​for each of read 1 (MQ: 30) and read 2 (MQ: 35), as well as the read base quality (BQ) values ​​(e.g., read 1 BQ: 27, 31, 24, 20, 7, 16, 18, 9, 12, 16, 15, 20 ...). These scalar integer values ​​can be selectively associated with an encoded two-dimensional Boolean matrix of read sequences. Thus, candidate mutation information can be encoded to include a 2D Boolean matrix of BQ values ​​with additional rows and associated MQ values. In some embodiments, each of the multiple N read sequences encoded as a 2D Boolean matrix is ​​associated with a copy of a corresponding enhanced reference sequence, which is also encoded as a two-dimensional Boolean matrix.

[0068] Figure 4 A schematic diagram depicting exemplary information encoded for a set of candidate mutations (denoted as candidate mutation 1 to candidate mutation Z) obtained from a high-throughput sequencing dataset. Each candidate mutation is encoded as a set of 2D Boolean matrices corresponding to multiple sequence reads overlapping the candidate mutation position. There are a total of N1 2D Boolean matrices encoding candidate mutation 1 and Nz 2D Boolean matrices encoding candidate mutation Z. In this example, N1 is not equal to Nz, and typically the number of reads N for each candidate mutation on the dataset is not the same. Each 2D Boolean matrix representing a sequence read overlapping a candidate mutation position has been enhanced and associated with a copy of the reference sequence overlapping that position.

[0069] Associating a reference sequence with each individual sequence read for a candidate mutation provides significant advantages in speed and efficiency in processing the encoded information using a deep learning model, as described in the methods of the invention. For example, as described elsewhere herein, it provides efficiency by allowing the processing of candidate mutations at positions with a large number of redundant overlapping sequence reads to be shortened or otherwise shortened.

[0070] also, Figure 4 It shows how each 2D Boolean matrix representing a read associates the scalar BQ value of the read as an array of integer values ​​(depicted as the bottom row). Figure 4 Not depicted in , but indicating a scalar MQ value associated with each read sequence may also be associated with each matrix.

[0071] Figure 4 It is also shown how 1 to N 2D matrices (each representing a different read segment, but each associated with a copy of the reference sequence matrix) can be further combined into a 3D matrix (i.e., a tensor) that represents all the encoded information for a particular candidate mutation. In addition, Figure 4It illustrates how a large number of different candidate mutations (ie, "candidate mutation 1, candidate mutation 2 ... candidate mutation Y, candidate mutation Z") across a sequencing dataset (e.g., covering a complete genome) can be represented by a set of Z-encoded 3D matrices.

[0072] It is also important that Figure 4 As described, each three-dimensional matrix may include two-dimensional Boolean matrices of different numbers (e.g., N1...Nz) because in real high-throughput sequencing datasets, the number of sequence reads representing each candidate mutation can vary greatly, for example, from N=10 to N=10,000. This inconsistency in multiple reads on the numerous candidate mutations that need to be classified in large sequencing datasets poses difficulties for processing using neural networks. As described elsewhere in this article, one advantage of the deep learning model of the present disclosure is that such variable-sized datasets are efficiently classified by encoding candidate mutation information and processing in parallel with multiple convolutional neural network (CNN) layers with different kernel sizes, and the CNN combines their outputs in a single TopK selected tensor.

[0073] The data structure of the included specific sequence data and information and the encoded information 102 provides advantages for a deep learning model for classifying candidate mutations of a genome. In particular, the encoded information associates (i.e., links) the reference sequence 201 with each of the N read sequences 202 that overlap with the position of the candidate mutation. The association of the reference sequence with each of the N read sequences allows the deep learning model 103 to process these jointly. This joint processing allows the neural network of the model to learn and extract features in real time, which can effectively shorten the processing of sequencing reads. For example, if the neural network can extract the learned features using only 12 of the 15 read sequences, then this particular candidate mutation classification can be terminated 3 reads in advance and start processing the next candidate mutation, thereby saving computational costs and time.

[0074] In general, a deep learning model has an "architecture" that includes a sequence of neural networks, such as a CNN, each of which includes multiple neural layers. The architecture can therefore be described in terms of a sequence of layers, the operations performed by each layer, and the connections between the layers. The key features of a neural network layer are the "kernels" or "filters," which perform computational transformations on the data input to that layer. Each filter has associated adjustable weight parameters, and the combination of the filters and their weights determines the precise computational transformation performed. During the "training" process of a deep learning model, the weight parameters of the filters are repeatedly adjusted to minimize the loss function, which constitutes "learning." The final learned values ​​of the weight parameters set at the end of training determine the structure of a deep learning model that can be used for the task of classifying candidate mutations.

[0075] As described further below, in some embodiments, the training process of a deep learning model can be simplified by using a method known as "transfer learning." Typically, transfer learning is used to train a neural network with training data that is different from the data originally used to initially train the neural network. Therefore, in transfer learning, adjustments to weighting parameters are limited. For example, selected weights and / or selected kernels can remain the same while others are adjusted. In this way, transfer learning can train a neural network to work better at classifying different types of data sets without completely retraining the neural network.

[0076] Figure 5A and Figure 5B A block diagram illustrating an exemplary deep learning model architecture 103 for processing encoded information 102 representing candidate mutations and outputting candidate mutation classifications 104 is depicted. Figure 5A and Figure 5B As depicted, in some embodiments of the invention, the deep learning model comprises a sequence of two neural network modules, wherein the first neural network module comprises multiple parallel CNNs with different kernel sizes, whose outputs are combined into a single tensor representing extracted features of the encoded information, and the second neural network module comprises a 1D convolutional layer and a fully connected layer capable of processing the tensors of the first module and outputting a classification of the candidate mutation 104.

[0077] Figure 5A An exemplary first neural network module is depicted in which an input encoded data structure 102 is processed by a plurality of parallel one-dimensional convolutional layers 501, 502 and 503 with different kernel sizes. Figure 5A As shown, there are three input one-dimensional convolutional layers with relative sizes of X, 2X, and 4X. However, it is contemplated that the difference in kernel size can vary in the range of about X to about 10X. In some embodiments, for example, the kernel sizes of the three layers can be X, 1.5X-3X, and 4X-10X. In some embodiments, multiple CNNs can include kernel sizes described in terms of base pairs (bp). For example, multiple three layers 501, 502, and 503 can have kernel sizes of 4bp, 8bp, and 16bp, respectively. In addition, as Figure 5A As shown, it is contemplated that multiple parallel 1D convolutional layers 501, 502, and 503 with different kernel sizes may each include additional neural network layers that are processed in parallel. Figure 5A The output of each 1D convolutional layer 501, 502, and 503 depicted in is then processed by a ReLU and MaxPool layer 504. Typically, the size of the layer is equal to or less than X, the kernel size of the smallest input convolutional layer. Again, as Figure 5A As shown, it is contemplated that each of the parallel ReLU and MaxPool layers may include additional parallel layers. Figure 5A As shown in the exemplary neural network of FIG, the parallel processed outputs of the convolutional layers 501, 502, and 503 and the subsequent ReLU and MaxPool layers 504 are then combined or concatenated into a single tensor 505.

[0078] As mentioned above, high-throughput sequencing datasets may generate millions of candidate mutations, each located at a different genomic location, each with different characteristics of multiple sequence reads, with different mapping quality characteristics, and base quality associated with each read. The large number and different characteristics of each candidate mutation make it very difficult to use deep learning to accurately and effectively classify candidate mutations. Figure 5A The surprising technical effect and advantage of the deep learning model architecture shown in is that parallel processing of the encoded information using multiple CNNs with different kernel sizes that combine their outputs in a single tensor leads to accurate and effective classification of candidate mutations. In addition, the architecture provides accurate and effective classification even starting from suboptimal (e.g., low coverage) sequencing datasets (e.g., average coverage less than 30x, less than 20x, less than 15x, or less than 10x).

[0079] In addition, Figure 5A In some embodiments of the first neural network module of the present disclosure illustrated in FIG. 1 , simultaneously with the concatenation process, optional ranking scores are also generated (not shown). Optionally, these ranking scores are then used to generate a TopK selected tensor 506 from tensor 505. Typically, the TopK selected K values ​​useful in the methods of the present disclosure range from about 10 to about 100. The result of the parallelization of multiple CNN outputs, the TopK selected tensor 506 output has a significantly reduced size, which can be obtained by Figure 5B The second neural network module described in is more efficient. In fact, Figure 5A Another surprising technical effect and advantage of the deep learning model architecture shown in is that even starting from a suboptimal (e.g., low coverage) sequencing dataset (e.g., average coverage less than 30X, less than 20X, less than 15X, or less than 10X), the reduced-size TopK-selected tensors 506 generated with K of 10-100 can be processed to provide accurate and efficient candidate mutation classification.

[0080] like Figure 5BAs shown in the example of, in some embodiments, the deep learning model includes a second neural network module, which includes an initial 1D convolutional layer 507, followed by at least two fully connected layers 508 and 509. This second neural network module is capable of accepting the tensor output of the first module as input and processing it. The output of the last fully connected layer of the second network is a classification of the candidate mutation 104. The output candidate mutation classification may include an output value that confirms whether a true mutation is observed at a specific position relative to the reference sequence based on whether it meets some predetermined thresholds. In addition, the candidate mutation classification can provide further information, including the type of mutation, for example, whether it is a SNP, insertion, deletion, etc.

[0081] Before using a deep learning model to classify candidate mutations, a training process is performed in which the neural network "learns" an optimal set of parameters (e.g., adjustable weighting factors) that results in the best fit of a training dataset (e.g., a suboptimal sequencing dataset) to a model reference dataset (e.g., a standard reference genome sequence). This learning process typically involves iterative stochastic gradient descent-based minimization through the space of adjustable weighting factors for the neural network. Thus, the training process typically includes error calculations, which are then back-propagated through the network to adjust the weighting parameters.

[0082] As contemplated herein, the training of a deep learning model for classifying candidate mutations may include a more time-consuming initial training of the model, and also include a less time-consuming "transfer learning" training of a previously trained model, but on a different data set. For example, transfer learning may involve further training of a neural network without relearning the entire set of adjustable weighting parameters. Transfer learning may be performed by training a neural network (e.g., CNN) that has been trained with a suboptimal data set using a new and / or slightly different type of data set. For example, a deep learning model initially trained with candidate mutation information and sequencing data obtained from one NGS instrument may be partially retrained with information and data obtained from different types of NGS instruments. Therefore, the learning from the initial training of the first class of NGS instructions will be transferred in training the neural network with model data from the second class of NGS instruments. Because there is no need to adjust the weighting parameters from scratch, the training process involved in transfer learning is greatly reduced. Therefore, it is contemplated in the method disclosed herein for classifying candidate mutations from sequencing data that the method further includes training a deep learning model. In some embodiments, the training corresponds to transfer learning using candidate mutation information and model sequencing data from different types of samples, sequencing instruments, and / or candidate mutation identification software, rather than being used in the original training of the deep learning model.

[0083] Figure 6An exemplary system and operational flow for training and then using a trained deep learning model to classify candidate mutations is described. Suboptimal candidate mutation information 601 and model reference sequence data 602 are input to a system including a deep learning model training system 603. In order to train a deep learning model to classify mutations from suboptimal (e.g., low coverage) NGS data. Ideally, the model reference sequence 602 used in training should be a reference sequence that provides ground truth candidate mutation classifications at each position. Therefore, the model reference sequence 602 can be used as a "ground truth" set for training a deep learning model. Typically, the model reference sequence is a fully characterized sequence that has been determined using a high-coverage sequencing data set and confirmed by independently conducted experiments, so that it provides the true position and identity of each base in the sequence with an acceptable level of confidence.

[0084] In some embodiments, the model reference sequence useful in the training method is obtained from high-quality (e.g., 30 times or greater coverage) sequencing data, in which the true sequence of each position has been confirmed by repeated experiments. This high-quality sequence useful for training (as a model reference sequence) can be prepared by performing high-quality NGS experiments and analysis on well-known samples. In some embodiments, the model reference sequence 602 can be a "gold standard" genome sequence from a source, such as a publicly available benchmark human genome from the "Bottle Genome Alliance (Genome in a Bottle Consortium)" (accessible at jimb.stanford.edu / giab) or a mouse genome in the "Mouse Genome Project" (available from the following website www.broadinstitute.org / mouse / mouse-genome-project).

[0085] Suboptimal candidate mutation information 601 may be obtained from sequencing datasets that are of lower quality in some aspects relative to the sequencing data used for model reference sequence 602. For example, while a model reference sequence will typically be generated from a sequencing dataset having an average read sequence coverage of 30x or greater throughout the sequence, suboptimal candidate mutation information 601 may be generated from a sequencing dataset obtained with an average sequence read coverage of less than 30x, less than 20x, less than 15x, less than 12x, or even less than 10x.

[0086] The final accuracy of the trained deep learning model depends at least in part on the quality level of the model reference sequence 602. In some embodiments, for training, the model reference sequence 602 and the suboptimal candidate mutation information 601 are obtained from the same experiment or using the same experimental conditions. In some embodiments, the suboptimal candidate mutation information 601 can be prepared from the sequencing data of the model reference sequence 602 by excluding certain portions of the read sequence, thereby effectively creating a low-quality data set. For example, by randomly excluding half of the sequence reads of a high-quality 30X coverage NGS data set, it can be effectively downgraded to a suboptimal 15X coverage data set. Other methods known in the art can also be used to simulate low-quality data from a high-quality data set, for example, by introducing noise.

[0087] Generally speaking, suboptimal candidate information 601 is obtained from low-coverage sequencing datasets using standard software tools (such as GATK) for generating candidate mutations, whether experimentally obtained or simulated.

[0088] like Figure 6 As shown, suboptimal candidate mutation information 601 is used together with a model reference sequence 602 as input to a deep learning model training system 603. The system 603 is configured to train a deep learning model to accurately classify candidate mutations obtained from suboptimal sequencing data. The process performed by the deep learning model training system 603 includes: using suboptimal candidate mutation information as input to an untrained deep learning model, obtaining an output candidate mutation classification from the untrained model, and determining a classification error of the ground truth candidate mutation classification relative to the model reference sequence. In some embodiments, error determination is performed using a loss function or a similar type of calculation that quantifies the error between the suboptimal candidate mutation classification and the true mutation classification known from the model reference sequence.

[0089] As described above, training is an iterative process, and the error of each iteration is used to adjust the model parameters. When the calculated error at the end of the iteration drops below a certain threshold, it is determined that the error is minimized and the deep learning model is optimized. Therefore, the training process involves minimizing the error in the classification of ground truth candidate mutations and suboptimal candidate mutations relative to the model reference sequence by adjusting the parameters of the deep learning model.

[0090] In some embodiments, error minimization is performed by stochastic gradient descent involving iterative adjustment of weight parameters within a deep learning model, which produces a trained deep learning model 103. A range of loss functions for training neural networks, such as CNNs, are known in the art and can be used in the candidate mutation classification methods and systems of the present disclosure.

[0091] The computations and data manipulations performed in deep learning models to classify candidate mutations from sequencing data are beyond human capabilities. The training and use of neural networks typically require high-performance, energy-efficient computer processors and are typically performed using graphics processing units (GPUs) or GPU farms. Therefore, in some embodiments, the present disclosure provides a computing device configured to apply a deep learning model in the process of classifying candidate mutations from large amounts of digital data associated with genomic sequences.

[0092] Figure 6 A block diagram 600 of an exemplary computer-based system for training and using a trained neural network system to classify candidate mutations from an obtained sequencing data set according to an embodiment of the present invention is further shown. The main computer-based components include a training engine 610 and a classification engine 620. As described above, the training engine is configured to accept inputs of suboptimal candidate mutation information 601 and inputs of a model reference sequence 602, and is connected to a memory device 630, which may include a temporary memory and / or a persistent storage device. The training engine 610 executes computer-based instructions for configuring a neural network of a deep learning model training system 603 based on a training process. The deep learning model parameters during training can be stored in the memory device 630. The main function of the training engine 610 in conjunction with the memory device 630 is to train and provide a final optimized deep learning model that can be accessed and used by the classification engine 620.

[0093] The computer-based implementation of the classification engine 620 is configured to accept an input of encoded information 102 representing candidate mutation information 101 and process it using the trained deep learning model 103 and output a candidate mutation classification 104. The input of encoded information 102 and the output of classification 104 may be stored in a memory device 630.

[0094] It is contemplated that in some embodiments, the classification engine 620 may provide further computer-based processing prior to input, such as encoding candidate mutation information 101 obtained from other software tools (e.g., GATK or SAMtools). It is also contemplated that the classification engine may provide further processing of the output candidate mutation classifications, such as recombining the classified mutations to provide a sequence including the mutations as output, such as a full-length sequence based on a complete reference sequence.

[0095] For ease of understanding, Figure 6The various computer-based elements illustrated in the figures and the functions attributed to them are generally described. Those skilled in the art will recognize that one or more functions assigned to each element may be performed by any of the other elements and / or elements not shown in the figures. In addition, it is contemplated that these elements may be configured to perform a combination of the various functions described above, as supported by the various embodiments described elsewhere herein. Therefore, the description of the training engine 610, the classification engine 620, and the memory device 630 is intended to include any suitable combination of computing devices, including servers, interfaces, systems, databases, agents, peers, engines, controllers, modules, or other types of computing devices that operate individually or collectively to perform the functions assigned to the various elements. In addition, those skilled in the art will recognize that the description herein Figure 6 One or more functions of the system may be performed in the context of a client-server relationship, such as by one or more servers, one or more client devices (e.g., one or more user devices), and / or by a combination of one or more servers and client devices.

[0096] Figure 7An exemplary system 700 is described in which embodiments of the general candidate mutation classification process 100, candidate mutation data encoding 200, deep learning model architecture 400, and / or training engine and classification engine functionality 500 may be implemented. The system 700 includes at least one central processor 701 connected to a communication bus 702. The communication bus 702 may be implemented using any suitable protocol, such as PCI (Peripheral Component Interconnect), PCI-Express, AGP (Accelerated Graphics Port), HyperTransport, or any other bus or point-to-point communication protocol. The system 700 also includes a main memory 704. Control logic (software) and data are stored in the main memory 704, which may include random access memory (RAM). The system 700 also includes an input device 712, a graphics processor 706, and a display 708, such as a conventional CRT (cathode ray tube), LCD (liquid crystal display), LED (light emitting diode) display, etc. User input may be received from the input device 712, which may include, but is not limited to, a keyboard, a mouse, a touchpad, a microphone, etc. In one embodiment, the graphics processor 706 may include a plurality of shader modules, a rasterization module, etc. Each of the above modules can even be located on a single semiconductor platform to form a graphics processing unit (GPU). As used herein, a single semiconductor platform can refer to a unique integrated circuit or chip based on a single semiconductor, or to a multi-chip module with increased connectivity that simulates on-chip operations. The individual modules can also be located on a semiconductor platform or various combinations thereof, depending on the user's wishes. The system 700 can also include auxiliary storage 710, such as a hard drive and / or a removable storage drive, which represents a floppy disk drive, a tape drive, a compact disk drive, a digital versatile disk (DVD) drive, a recording device, a universal serial bus (USB) flash memory.

[0097] Software instructions, computer programs, and / or computer control logic algorithms may be stored in the system's main memory 704 and / or secondary storage 710. Such software instructions, when executed, allow the system 700 to perform various functions. The main memory 704, secondary storage 710, and / or any other storage are examples of computer-readable media.

[0098] In one embodiment, each of the above Figure 1-6 The functionality and / or architecture may be implemented in the context of a central processor 701, a graphics processor 706, an integrated circuit (not shown) capable of having at least a portion of the capabilities of both the central processor 701 and the graphics processor 706, a chipset (i.e., a group of integrated circuits designed to operate and sold as a unit that performs related functions, etc.), and / or any other integrated circuit.

[0099] In some embodiments, Figure 1-6 The functionality and / or architecture of the system 700 can be implemented in the context of a general purpose computer system, a circuit board system, an application specific system, and / or any other desired system. For example, the system 700 can take the form of a desktop computer, a laptop computer, a server, a workstation, an embedded system, and / or any other type of logic. In some embodiments, the system 700 can take the form of various other devices, including but not limited to a personal digital assistant (PDA) device, a mobile phone device, a television, etc.

[0100] It is also contemplated that in some embodiments, system 700 can be coupled to a network for communication purposes, such as a telecommunications network, a local area network (LAN), a wireless network, a wide area network (WAN) such as the Internet, a peer-to-peer network, a cable network, etc.

[0101] In some embodiments, methods and systems for classifying candidate mutations using deep learning, as described herein and by Figure 1-6 As shown, the system 700 can be remotely implemented, controlled, and / or otherwise used via the Internet or other network systems. For example, the system 700 can be hosted on one or more servers that can be accessed by remote users and used to classify candidate mutations from user sequencing datasets. Therefore, in some embodiments, users can upload their own sequencing datasets and / or candidate mutation information for classification using a fully trained version of the deep learning model 103 hosted on the server.

[0102] In addition, in some embodiments, users can also upload their own suboptimal sequencing data sets and / or suboptimal candidate mutations to further train the deep learning model hosted on the remote server (e.g., reference learning). The user will then classify the candidate mutations generated by their own sequencing data set using the further trained deep learning model hosted on the server. Typically, the user will download the output mutation classification for further local use, however, in some embodiments, the hosting system for mutation classification may include other tools for analysis, such as a database of model or suboptimal candidate mutation information, model reference sequences, and / or other data useful in the various sequence analysis methods described herein.

[0103] In general, computing devices used with the deep learning-based mutation detection and classification processes and systems of the present disclosure may include any appropriate combination of computing devices, including servers, interfaces, systems, databases, agents, peers, engines, controllers, modules, or other types of computing devices operating individually or collectively. At least one processor (e.g., GPU, CPU, ASIC, FPGA, DSP, x88, ARM, etc.) of the computing device is configured (or programmed) to execute software instructions stored on a computer-readable tangible non-transitory medium (e.g., hard drive, solid-state drive, RAM, flash memory, ROM, etc.), thereby performing the processes involved in the training method and classifying mutations using a deep learning model.

[0104] Software instructions for configuring or programming a computing device to provide candidate mutation classification functionality can be prepared using standard programming tools. For example, extracting and encoding candidate mutation information from high-throughput sequencing data can be programmed using specifications for VCF and BAM file formats and software tools available online, such as the SAMtools repository at github.com / samtools. Deep learning model architectures and neural network configurations can be programmed using software tools such as Keras (v.2.1.3) and Tensorflow (v.1.4.0). Typically, software instructions are implemented as a computer program product that includes a non-temporary tangible computer-readable medium storing instructions that cause a processor of a computing device to perform the steps of the denoising process disclosed herein.

[0105] In addition, the methods of the present disclosure may be implemented using standard protocols and algorithms for data exchange between servers, systems, databases, or interfaces during the implementation of each process. For example, the data exchange used in implementing the methods and systems of the present disclosure may be performed using protocols based on HTTP, HTTPS, AES, public-private key exchange, web service APIs, financial transaction protocols, or other electronic information exchange methods through packet switching networks, circuit switching networks, the Internet, local area networks, wide area networks, VPNs (or other types of networks).

[0106] Although the foregoing disclosure of the present invention has been described in some detail by way of example and illustration for purposes of clarity and understanding, the present disclosure, including the examples, descriptions, and embodiments described herein, is intended to be exemplary and should not be construed as limiting the present disclosure. It will be clear to those skilled in the art that various modifications or changes may be made to the examples, descriptions, and embodiments described herein and incorporated into the spirit and scope of the present disclosure and the appended claims. In addition, those skilled in the art will recognize that there are many methods and procedures equivalent to the methods and procedures described herein. All such equivalents are understood to be within the scope of the present disclosure and are covered by the appended claims.

[0107] Further embodiments of the invention are set forth in the following claims.

Claims

1. A method for detecting gene mutations, comprising: encoding information representing a candidate mutation, wherein the encoded information comprises an absolute position, a reference sequence, a number of reads (N), and a read sequence, wherein the reference sequence and the read sequence are encoded as a 2D Boolean matrix, wherein each possible base identity or insertion at each position in each sequence of the reference sequence and the read sequence is represented as being present or absent by a Boolean element in the 2D Boolean matrix encoding the sequence, wherein the encoding associates the reference sequence with each read sequence; and The encoded information is processed using a deep learning model, wherein the deep learning model jointly processes the reference sequence with each read sequence and classifies whether the candidate mutation is a true mutation, wherein the processing of the encoded information using the deep learning model includes: processing the encoded information using multiple parallel CNNs with different kernel sizes, and combining the outputs into a tensor.

2. The method of claim 1, wherein the encoding enhances the reference sequence.

3. The method of claim 1, wherein the encoded information further comprises a read mapping quality value and / or a read base quality value.

4. The method of claim 1 , wherein combining the outputs comprises: Generate the sorted scores and use the sorted scores to generate a tensor for selecting TopK.

5. The method of claim 1, wherein processing the encoded information using a deep learning model further comprises: The tensor is transformed using another CNN, wherein the other CNN includes a one-dimensional convolutional layer followed by multiple fully connected layers, wherein the last fully connected layer outputs a classification of the candidate mutation.

6. The method according to claim 1, wherein the method further comprises: Training the deep learning model, wherein the training comprises: encoding information representing a suboptimal candidate mutation, wherein the encoded information for the suboptimal candidate mutation includes an absolute position, a model reference sequence, a number of reads (N), and a read sequence; processing the encoded information for the suboptimal candidate mutations with a deep learning model, the deep learning model classifying the suboptimal candidate mutations; and By adjusting the parameters of the deep learning model, the error in classification of the suboptimal candidate mutation relative to the ground truth candidate mutation of the model reference sequence is minimized.

7. A system for detecting gene mutations, comprising a processor, a memory device, and a classification engine executable on the processor according to software instructions stored in the memory device, wherein the classification engine is configured to: Encoding information representing a candidate mutation, wherein the encoded information includes an absolute position, a reference sequence, a number of reads (N), and a read sequence, wherein the reference sequence and the read sequence are encoded as a 2D Boolean matrix, wherein, each possible base identity or insertion at each position in the reference sequence and each sequence of the read sequences is represented as present or absent by a Boolean element in the 2D Boolean matrix encoding the sequence, wherein the encoding relates the reference sequence to each read sequence; as well as The encoded information is processed using a deep learning model, wherein the deep learning model jointly processes the reference sequence with each read sequence and classifies whether the candidate mutation is a true mutation, wherein the processing of the encoded information using the deep learning model includes: processing the encoded information using multiple parallel CNNs with different kernel sizes, and combining the outputs into a tensor.

8. The system of claim 7, wherein the encoded information further comprises a read mapping quality value and / or a read base quality value.

9. A non-transitory computer-readable medium comprising instructions for classifying a candidate mutation, the instructions, when executed by a processor, causing the processor to perform steps comprising: Encoding information representing a candidate mutation, wherein the encoded information includes an absolute position, a reference sequence, a number of reads (N), and a read sequence, wherein the reference sequence and the read sequence are encoded as a 2D Boolean matrix, wherein, each possible base identity or insertion at each position in the reference sequence and each sequence of the read sequences is represented as present or absent by a Boolean element in the 2D Boolean matrix encoding the sequence, wherein the encoding relates the reference sequence to each read sequence; as well as The encoded information is processed using a deep learning model, wherein the deep learning model jointly processes the reference sequence with each read sequence and classifies whether the candidate mutation is a true mutation, wherein the processing of the encoded information using the deep learning model includes: processing the encoded information using multiple parallel CNNs with different kernel sizes, and combining the outputs into a tensor.

Images