Consensus calling with a neural network

Abstract

A method and systems for determining a consensus molecular sequence from sequence data of a sample, the sequence data including base calls and quality scores (q-scores). One or more feature vectors are generated and input into a convolutional neural network (CNN) model, which outputs a consensus call sequence. In order to generate the feature vectors, sequence reads are aligned and clustered. For each cluster, one or more feature vectors is then generated by either concatenating base calls and quality scores from the plurality of sequence reads in the cluster, or calculating aggregation statistics for the plurality of sequence reads and concatenating the aggregation statistics and corresponding quality metrics for the plurality of sequence reads in the cluster.

Description

PATENTClient Reference No.: P39780-WO-1CONSENSUS CALLING WITH A NEURAL NETWORKCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This Application claims priority to United States Provisional Patent Application No. 63 / 722,633, filed on November 20, 2024.BACKGROUND

[0002] Biological samples can be used to identify genetic information including a sequence of ordered sets of nucleic acids in nucleic acid material. Rather recently, single-molecule sequencing has been rapidly advancing in techniques and expanding in applications. These techniques can sequence individual molecules and can be performed in real-time without PCR amplification. Traditional PCR amplification techniques generally involve significant time for reads in the tens of millions of base pairs and may be subject to high error rates.

[0003] Some newer techniques involve chemistry that translates a sequence of deoxyribonucleic acid (DNA) into a measurable surrogate molecule called an Xpandomer. This Xpandomer molecule is then sequenced using a sequencing instrument instead of the original DNA target molecule. Xpandomer synthesis based on the natural function of DNA replication uses expandable nucleotide triphosphates (X-NTPs) that act as substrates for template-dependent, polymerase-based replication. Different X-NTPs may be coded with different reporters used to identify the particular base or combination of bases in the original target molecule complementary to the X-NTP added to the Xpandomer during synthesis. As the Xpandomer molecule transits through nanometer- sized openings in an electrode-resistant membrane, which may be created by embedding a barrel-shaped protein in a lipid bilayer mamebrane (a “nanopore”), a distinct electrical signal is generated for each base reporter in the Xpandomer molecule and identifiable to enable highly accurate and high throughput nanopore-based nucleic acid sequencing (also referred to generally as nanopore sequencing).

[0004] The techniques hold incredible promise with regard to facilitating, at least, the following: building of comprehensive libraries relating genes to diseases, identifying and characterizing new or rare diseases, and / or identifying targeted therapies to treat various diseases like cancer. Although rapid and accurate, such sequencing techniques (e.g., third-generation orPATENTClient Reference No.: P39780-WO-1 next-generation techniques) can result in incomplete reads which, by themselves, fail to adequately identify a sequence and are thus discarded, wasting processing resources. For example, during Xpandomer synthesis, a polymerase that facilitates ligation of X-NTPs into the Xpandomer molecule can detach from the target DNA strand leading to incomplete synthesis of Xpandomers and, subsequently, incomplete sequence reads during sequencing.

[0005] Traditional methods of combining incomplete and / or multiple reads to form a consensus sequence may be performed by analyzing sets of read sequences and applying statistics to determine the most likely / common base call at any particular genomic loci. Some technologies utilize more advanced analysis, including recurrent neural networks or linear regression, for example, in lieu of the statistical analysis. However, these techniques may require high levels of computing power in order to process millions of bases and / or may still suffer high levels of inaccuracy. Improved techniques for utilization and processing of multiple and incomplete reads to determine consensus calls are thus desired.SUMMARY

[0006] The embodiments described herein relate to systems and methods for consensus calling based on analysis of sequencing data generated by a sequencer instrument. More particularly, the embodiments described herein related to computer-implemented methods for analyzing sequencing data using a machine learning model.

[0007] In accordance with a first aspect of the present disclosure, a method is provided for determining a consensus molecular sequence. The method includes: sequencing, using a sequencing device, a biological sample to obtain sequence data; aligning the plurality of sequence reads to a reference genome; sorting the aligned reads into one or more clusters; generating a feature vector for each cluster; and processing the feature vector by a trained convolutional neural network (CNN) model having a plurality of convolutional layers to generate a consensus call sequence for the cluster. The sequence data comprises a plurality sequence reads, each sequence read comprising a plurality of base calls and corresponding quality scores (q-scores) generated by the sequencing device.

[0008] In accordance with at least one embodiment of the first aspect, the feature vector comprises one -hot encoding of the base calls in each of the sequence reads in the cluster.PATENTClient Reference No.: P39780-WO-1

[0009] In accordance with at least one embodiment of the first aspect, the feature vector further comprises a quality metric that indicates q-score for the base call generated by the sequencing device.

[0010] In accordance with at least one embodiment of the first aspect, the feature vector further comprises a strand indicator that indicates whether a particular base call is associated with a forward or reverse strand of the target molecule.

[0011] In accordance with at least one embodiment of the first aspect, the feature vector comprises aggregation statistics for base calls at each genomic position for the aligned sequence of the cluster.

[0012] In accordance with at least one embodiment of the first aspect, the feature vector further comprises quality metrics for each genomic position. The quality metrics comprise an average quality score for each of up to four called bases at the genomic position across all sequence reads in the cluster.

[0013] In accordance with at least one embodiment of the first aspect, sorting the aligned reads into one or more clusters comprises sorting into clusters based on at least one of a unique molecular identifiers (UMIs), starting position, or ending position of each aligned read. All sequence reads in a particular cluster include the same UMI.

[0014] In accordance with at least one embodiment of the first aspect, aligning the sequence reads to the reference genome comprises, for at least one sequence read, replacing one or more base calls in a sequence with a gap or inserting a gap between two adjacent base calls in the sequence read.

[0015] In accordance with at least one embodiment of the first aspect, the CNN includes at least three hidden convolutional layers, each hidden convolutional layer applying a convolution operation to input features to generate filtered output features and an acli vation function to the filtered output features to generate a set of acli vations passed to the next layer of the CNN.

[0016] In accordance with at least one embodiment of the first aspect, the method further comprising training the CNN in accordance with a training data set. The training data set includes sequence data for a plurality of clusters, and the sequence data for each cluster in the training data set includes a plurality of aligned sequence reads and a ground-truth consensus call sequence for the cluster.PATENTClient Reference No.: P39780-WO-1

[0017] In accordance with at least one embodiment of the first aspect, the method further comprising collecting a number of consensus call sequences for a plurality of clusters of the sequence data for the sample, and storing the consensus call sequences in a BAM file.

[0018] In accordance with at least one embodiment of the first aspect, the method further comprising passing the BAM file to a downstream variant caller to idenli ly one or more variants based on the collection of the number of consensus call sequences.

[0019] In accordance with a second aspect of the present disclosure, a system is provided for determining a consensus call sequence. The system includes a sequencing device and a computing system, communicatively coupled to the sequencing device. The sequencing device is configured to sequence a biological sample to obtain sequence data. The sequence data comprises a plurality sequence reads, each sequence read comprising a plurality of base calls and corresponding quality scores (q-scores) generated by the sequencing device. The computing system comprises at least one processor executing instructions that cause the computing device to: align the plurality of sequence reads to a reference genome; sort the aligned reads into one or more clusters; generate a feature vector for each cluster; and process the feature vector by a trained convolutional neural network (CNN) model having a plurality of convolutional layers to generate a consensus call sequence for the cluster.

[0020] In accordance with at least one embodiment of the second aspect, the feature vector comprises one -hot encoding of the base calls in each of the sequence reads in the cluster.

[0021] In accordance with at least one embodiment of the second aspect, the feature vector further comprises at least one of: a quality metric that indicates q-score for the base call generated by the sequencing device, or a strand indicator that indicates whether a particular base call is associated with a forward or reverse strand of the target molecule.

[0022] In accordance with at least one embodiment of the second aspect, the feature vector comprises aggregation statistics for base calls at each genomic position for the aligned sequence of the cluster.

[0023] In accordance with at least one embodiment of the second aspect, the feature vector further comprises quality metrics for each genomic position, and wherein the quality metrics comprise an average quality score for each of up to four called bases at the genomic position across all sequence reads in the cluster.PATENTClient Reference No.: P39780-WO-1

[0024] In accordance with at least one embodiment of the second aspect, sorting the aligned reads into one or more clusters comprises sorting into clusters based on at least one of a unique molecular identifiers (UMIs), starting position, or ending position of each aligned read, and wherein all sequence reads in a cluster include the same UMI.

[0025] In accordance with at least one embodiment of the second aspect, aligning the sequence reads to the reference genome comprises, for at least one sequence read, replacing one or more base calls in a sequence with a gap or inserting a gap between two adjacent base calls in the sequence read.

[0026] In accordance with at least one embodiment of the second aspect, the computing device further configured to: collect a number of consensus call sequences for a plurality of clusters of the sequence data for the sample; store the consensus call sequences in a BAM file; and process the BAM file to a downstream variant caller to identify one or more variants based on the collection of the number of consensus call sequences.BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The detailed description is set forth with reference to the accompanying figures.

[0028] FIG. 1 sets forth an illustrative system including a sequencing device communicatively coupled to a computing system, in accordance with at least some embodiments of the present disclosure.

[0029] FIG. 2 illustrates a flow chart of a method 200 for determining a consensus sequence from sequencing data, in accordance with at least some embodiments of the present disclosure.

[0030] FIG. 3 illustrates an architecture of a neural network 300 used to determine a consensus sequence, in accordance with at least some embodiments.

[0031] FIG. 4 illustrates a process for aligning and clustering sequence read data, in accordance with at least some embodiments.

[0032] FIGs. 5A-5B illustrate a process for aligning sequence read data, in accordance with at least some embodiments.

[0033] FIG. 5C illustrates a one-hot encoding for an aligned sequence, in accordance with an embodiment.PATENTClient Reference No.: P39780-WO-1

[0034] FIGs. 6A-6B show a portion of a feature vector generated according to the one-hot encoding scheme, in accordance with at least some embodiments.

[0035] FIGs. 7A-7B show a portion of a feature vector generated according to an aggregation statistics encoding scheme, in accordance with at least some embodiments.

[0036] FIG. 8 illustrates an example computer system that may be utilized to implement techniques disclosed herein.DETAILED DESCRIPTION

[0037] Methods and systems disclosed herein utilize a convolutional neural network (CNN) to process molecular sequence data and determine (or call) consensus sequences. Sequence read data and corresponding quality (q) scores (e.g., Phred scores) are obtained and input into the CNN, wherein q-scores represent the probability of a base call being correct. The sequence read data may be grouped into clusters of sequences (e.g., aligned sequence fragments). The clusters and their size may be utilized by the neural network to assist in determining consensus sequences and corresponding quality scores (and / or the reliability of the output data).

[0038] Prior to input into the CNN, the sequence read data can be transformed into one or more feature vectors that include, in some embodiments, concatenated base calls and q-scores from at least a subset of the aligned sequence reads. The concatenated base calls can be encoded using a one-hot encoding scheme. In other embodiments, the feature vectors represent aggregation statistics of the molecular data and gaps within base calls. For example, the aggregation statistics can include, for each genomic position in the sequence, a percent of the aligned sequence reads having one of the four nucleotide bases at the genomic position. The aggregation statistics can also include a percentage of the aligned sequence reads that include a gap at the genomic position. In some embodiments, aggregation statistics include that of A, T, C, G, and gap calls on forward and / or reverse strands and the respective q-scores.

[0039] In some embodiments, the convolutional layers project logits of a plurality of classes and the output of the CNN is determined based on the class with the maximum logits. In some embodiments, the CNN includes 5 convolutional layers and projects logits of 5 classes.

[0040] The CNN’s convolutional layers can be trained with exemplary sequence data from other biological samples, respective q-scores, and genome and / or other reference data. In some embodiments, the training includes grouping and monitoring (e.g., for accuracy / quality) ofPATENTClient Reference No.: P39780-WO-1 groups of all bases together and one or more groups including bases that fall below particular q- score thresholds. In some embodiments, the groupings include bases with q-scores that fall below about 1% and / or about 2% of all q-scores. The monitoring of these groupings and resulting training based on the monitoring may improve the reliability of subsequent outputs by similarly applying the training to select or filter input sequence data in determining an output consensus (e.g., by filtering confident calls from unconfident calls).

[0041] A reference genome and / or other reference data may be used to provide “ground truth” in training the CNN. For example, information about common or uncommon variants in particular regions may also be used for training the CNN. The training may be adapted to the particular sequencing technology to be used in consensus calling. For example, technology types may include Sanger sequencing, sequencing-by-synthesis (SBS), nanopore sequencing, sequencing-by-expansion (SBX), and / or long- or short-read sequencing to correspond with sequencing data used to make consensus calls from subsequent reads using a particular sequencing instrument. The q-score thresholds / groupings described above may also be adapted based on data received from particular technologies or devices.

[0042] The techniques described herein provide an improved method for performing consensus calling by applying neural networks to analyze sequencing data. The consensus calls may be more accurate than those produced with prior art techniques and can improve technologies in downstream bioinformatics pipelines, particularly in the areas of diagnosis, prognosis, and / or treatment of multiple diseases such as cancer. The techniques disclosed herein may be particularly suited to sequencing data collected using SBX chemistry, such as with a nanopore -based sequencer instrument.Systems and Devices for Identifying a Consensus Sequence

[0043] FIG. 1 sets forth an illustrative system 100 including a sequencing device 110 (which may be alternately referred to as a sequencer instrument) communicatively coupled to a computing system 102. Sequencing device 110 can be coupled to coinpuling system 102 either directly (e.g., through one or more communication cables) or through network 130, which may be the Internet or any other combination of wide-area, local area, wired, and / or wireless networks. In some embodiments, computing system 102 may be included in or integrated with the sequencing device 110. In some embodiments, sequencing device 110 may sequence (e.g., perform a biochemical assay) a sample containing genetic material and produce resul lingPATENTClient Reference No.: P39780-WO-1 sequencing data. The sequencing data can be sent to computing system 102 (e.g., through network 130) or stored on a storage device and at a later stage transferred to computing system 102 (e.g., through network 130). In some embodiments, computing system 102 may or may not include a display 108 and one or more input devices (not illustrated) for receiving commands from a user or operator (e.g. a technician or a geneticist). In some embodiments, computing system 102 and / or sequencing device 110 can be accessed by users or other devices remotely through network 130. Thus, in some embodiments various methods discussed herein may be run remotely on computing system 102.

[0044] Computing system 102 may include one computing device or a combination of a number of computing devices of any type, such as personal computers, laptops, network servers (e.g., local servers or servers included on a public / private / hybrid cloud), mobile devices, etc., where some or all of the devices can be interconnected. Computing system 102 may include one or more processors (not illustrated), each of which can have one or more logic cores. In some embodiments, computing system 102 can include one or more general-purpose processors (e.g., CPUs), special-purpose processors such as graphics processors (GPUs), digital signal processors, or any combination of these and other types of processors. In some embodiments, some or all processors in computing system can be implemented using customized or customizable circuitry, such as application specific integrated circuits (ASICs) or field programmable gate arrays (FPGAs). Computing system 102 can also in some embodiments retrieve and execute non- transitory computer-readable instructions stored in one or more memories or storage devices (not illustrated) integrated into or otherwise communicatively coupled to computing system 102. The memory / storage devices can include any combination of non-transitory computer readable storage media including semiconductor memory chips of various types (e.g., dynamic random access memory (DRAM), static random access memory (SRAM), synchronous dynamic random access memory (SDRAM), flash memory, programmable read-only memory, etc.) and so on. Magnetic and / or optical disks can also be used. The memories / storage devices can also include removable storage media that can be readable and / or writeable; examples of such media include compact disc (CD), read-only digital versatile disc (e.g., DVD-ROM, dual -layer DVD-ROM), read-only and recordable Blu-ray® disks, ultra-density optical disks, flash memory cards (e.g., SD cards, mini-SD cards, micro-SD cards, etc.), and so on. In some embodiments, data and other information (e.g. sequencing data) can be stored in one or more remote locations, e.g., cloud storage, and synchronized with other the components of system 100.PATENT Client Reference No.: P39780-WO-1

[0045] In some embodiments, the sequencing device 110 can generate sequencing data by a sequencing by expansion (SBX) process. Examples of the SBX process include those described in U.S. Patent Application No. 17 / 456,342 (U.S. Publication No. US20220411458A1), entitled “Translocation control elements, reporter codes, and further means for translocation control for use in nanopore sequencing,” filed November 23, 2021, which is herein incorporated by reference in its entirety. During library preparation in the SBX process, a number of surrogate molecules are derived from and characterize nucleic acid material provided in a sample.

[0046] More particularly, the SBX process may translate a sequence of DNA into a measurable surrogate molecule called an Xpandomer. Xpandomer synthesis based on the natural function of DNA replication uses expandable nucleotide triphosphates (X-NTPs) that act as substrates for template-dependent, polymerase-based replication. These Xpandomer molecules are then processed by a sequencer instrument (e.g., the sequencing device 110) to measure the sequence in the original DNA template. As the Xpandomer molecule transits through nanometersized openings in an electrode -resistant membrane (a “nanopore”), each nanopore corresponding to a selective channel, a distinct electrical signal is generated for each base reporter and identifiable to enable highly accurate and high throughput nanopore-based nucleic acid sequencing (also referred to generally as nanopore sequencing).

[0047] Sequencing device 110 can generate a plurality of sequence reads corresponding to a genetic sample (e.g., a sample comprising a patient’s DNA or RNA material). For example, a sequence may be identified by processing a sample, which may include (for example) a blood, saliva, or tissue biopsy collected from a subject. Sequence reads can be obtained either directly from sequencing device 110, or from one or more local or remote volatile or non-volatile memories, storage devices, or databases communicatively coupled to computing system 102. Sequence reads can be pre-processed (e.g., pre-aligned) or they can be “raw,” in which case a downstream method may include a preprocessing (e.g., pre-aligning) step. Also, while in some embodiments, entire sequence reads (as generated by sequencing device 110) can be obtained, in other embodiments only sections of the sequence reads can be obtained. Thus, “obtaining a sequence read,” as used herein, refers generally to obtaining one or more sections of one or more (e.g., adjacent) sequence reads.Method for Determining a Consensus SequencePATENTClient Reference No.: P39780-WO-1

[0048] FIG. 2 illustrates a flow chart of a method 200 for determining a consensus sequence from sequencing data, in accordance with at least some embodiments of the present disclosure. The consensus sequence may be determined based on a subset of aligned sequence reads that overlap a particular genomic loci (position). The sequence data can comprise at least one of intermolecular sequence reads or intramolecular sequence reads.

[0049] As used herein, intermolecular sequence reads refer to a subset of overlapping sequence reads that include at least one of the following: similar unique molecular identifiers (UMIs), similar genomic starting positions, and / or similar genomic ending positions. UMIs are codes added to adapters ligated to target molecules being sequenced that can be used to identify separate molecules derived from a particular target molecule during an amp I i fi cat i on step of library preparation. Thus, sequence reads sharing the same UMI can be identified as being associated with the same molecule prior to amplification, allowing a plurality of sequence reads sharing the same UMI to be clustered and used to form a consensus that can improve accuracy in distinguishing true variants versus sequencing errors.

[0050] As used herein, intramolecular sequence reads refer to a subset of sequence reads from an individual molecule. In certain embodiments, during library preparation, a doublestranded DNA molecule can be ligated to a y adapter on one end and a hairpin adapter on the other. The DNA molecule can then be separated to form a single strand with both sense and an ( i - sense parent strands of the original dsDNA strand in a single-stranded DNA template molecule. Additional steps can be performed to synthesize a new complementary daughter strand to the ssDNA template. A new set of adapters are ligated to the synthesized dsDNA molecule, and this dsDNA molecule can then be separated to form a ssDNA molecule with both sense and antisense parent and daughter strands. This process can be repeated up to N limes such that a single ssDNA molecule includes N sets of scnsc / anli-scnsc daughter strands in addition to the original sense / anti-sense parent strands. During sequencing, this single molecule can produce a number of intramolecular reads, each intramolecular read corresponding to either the sense or anti-sense strand of one of the parent strand or daughter strands located between portions of an adapter added to the ends of the molecule during library preparation. Because the reads are all included in a single molecule, and the adapters used during library preparation all include the same UMI, it is possible to identify these reads as intramolecular reads, which all should include the same base calls (ore reverse complements thereof). Eike intermolecular sequence reads, intramolecular sequence reads can be used to determine a consensus sequence using a plurality of sequence reads that can improve accuracy in distinguishing true variants versus sequencing errors.PATENTClient Reference No.: P39780-WO-1

[0051] At 210, sequencing data of a biological sample is obtained, such as utilizing a sequencing instrument based on SBS or SBX chemistry. In some embodiments, a technology known as Xpandomer synthesis is utilized, and sequencing of the Xpandomer molecules is performed instead of the original target nucleic acid material. As the Xpandomer molecule transits through nano-sized openings in an electrode-resistant membrane (a “nanopore”), a distinct electrical signal is generated for each base reporter and iden (i liable to enable highly accurate and high throughput nanopore-based nucleic acid sequencing (also referred to generally as nanopore sequencing). A q-score (Phred quality score) is provided with a base call from a sequencer instrument.

[0052] At 220, sequence reads from the sequencing data are aligned and sorted into clusters of aligned (overlapping) sequence reads, such as illustrated in FIG. 4. The sequencing data may comprise a set of reads, each read comprising a series of base calls and corresponding q-scores associated with the calls (e.g., as shown in FIGs. 6A-6B and 7A-7B). In some embodiments, the base calls may comprise one of the nucleotide bases (A, T, C, G) or a gap (-) in the case that a call cannot be made. The corresponding q-scores may be included in each sequence read, such as determined by a primary base calling algorithm and stored in a FASTQ file, as is known to those of skill in the art.

[0053] In an embodiment, a subset of aligned reads are idenli lied. The subset may comprise a set of intermolecular sequence reads or intramolecular sequence reads selected during clustering. In an embodiment, the raw sequence reads may be aligned to a reference genome using, e.g., a Burrows-Wheeler Alignment (BWA) algorithm or other well-known alignment algorithm. After alignment, the sequence reads can be sorted into clusters. During clustering, the location of adapters in each sequence read can be identified. UMIs can be read from the adapters, and genomic starting and ending positions of the sequence of the target molecule can be determined based on their alignment with the reference genome. Sequence reads sharing at least one of the same UMI, starting position, and / or ending position can be added into a cluster. The sequence reads in a given cluster are then used to determine, via a neural network, a consensus sequence.

[0054] At block 230, one or more feature vectors are generated for the sequence reads in a cluster. In some embodiments, the sequence read data is concatenated to form feature vectors having base calls and corresponding q-scores for each of the sequence reads in the cluster. The concatenation may include all data from all reads in the cluster. In some embodiments, the number of sequence reads in a cluster can be capped, for example, to be no more than 5, 10, 15,PATENTClient Reference No.: P39780-WO-120, 25, or more reads. If there are more sequence reads in the cluster than the threshold number of sequence reads, then a number of reads having lowest average q-score for all bases in the read can be filtered from the cluster until the remaining number of reads in the cluster is equal to the threshold number. Alternatively, if the number of reads in the cluster is less than the threshold number of sequence reads, then elements of the feature vector corresponding to missing reads can be zero padded to maintain a consistent size of the feature vector.

[0055] In an embodiment, particular calls can be filtered out of the average q-score calculation if the q-score for a particular call falls below a threshold. For example, the threshold can be calculated by finding a distribution of q-scores across all reads in a population (e.g., the cluster of reads, the sequence data for the sample, sequence data for a plurality of samples, etc.) and selecting the threshold corresponding to the bottom 1%, 2%, 3%, etc. of all q-scores. This filtering step can help to remove a small number of outlier calls from the reads when selecting the reads in the cluster for generating a feature vector.

[0056] In an embodiment, the feature vector can be formed, starling at a first position of the consensus sequence, by concatenating a one-hot vector for each read in the cluster. For example, the feature vector for a given genomic position may include 7 rows and N columns, where N is the number of sequence reads in the cluster. The feature vector for each position is then concatenated with the feature vector for the next position in the sequence. In an embodiment, the seven rows correspond to the four base calls and gap, a strand indicator (0 / 1 for forward / reverse), and a q-score for that call.

[0057] In another embodiment, the feature vector is formed by concatenating aggregation statistics of the sequence data in the cluster. Instead of using one-hot encoding, the feature vector for a particular genomic posilion in the sequence can include two vectors: a first vector that encodes the percentages of each base or gap called for a particular genomic posilion, and a second vector that encodes the average quality score (q-score) for each of the calls corresponding to a particular nucleotide at the genomic posilion. In an embodiment, the aggregation statistics may include both forward and reverse strand data, which can be either encoded cumulatively or separately. For example, for a given posilion, the feature vector can include 11 rows, corresponding to 5 rows for percentage of base calls or gap calls in the forward strand, 5 rows for percentage of base calls or gap calls in the reverse strand, and one row for average quality score for all sequence reads for both forward and reverse strand. Alternatively, the feature vector can include 20 rows, 10 rows for percentage of base calls or gap calls in both forward and reversePATENTClient Reference No.: P39780-WO-1 strand, respectively, and 10 rows for average q-score by corresponding base or gap, for both forward and reverse strands, respectively. The use of aggregation statistics rather than one-hot encoding may be preferrable when the number of sequence reads in a cluster is above a given threshold, thereby reducing the complexity of the CNN. If the number of sequence reads in the cluster is too low, then percentage calculations may be too quantized and the accuracy of the CNN may be improved by using a feature vector generated with one-hot encoding.

[0058] Again, the use of feature vectors based on aggregation statistics may lower the complexity of the CNN and the computational load of inference without significantly impacting the quality of the output for particular use cases (e.g., depending on the error rate relative to sample types / sequencing technology and / or required quality parameters for a particular use case). The selection of a number of layers in the CNN can provide a balance between accuracy and speed in some embodiments, and increasing or decreasing the number of layers may adjust the speed or accuracy depending on a particular use case. Utilizing aggregation statistic features may yield faster processing time with a minor sacrifice in accuracy, which may be desirable in some cases.

[0059] At block 240, the generated feature vectors are input into the first layer of a CNN (e.g., layer 320A of FIG. 3). Prior to input of data for a sample under test, the CNN is trained with sequence data from other samples, similarly structured as feature vectors. At block 250, after the data is processed by layers of the CNN, an output from the CNN representing a consensus sequence is received. In some embodiments, the output includes a consensus base call for each position of a sequence and a corresponding quality measure, such as a quality score based on a likelihood of a logit output (predicted base call) predicting the accuracy of the output. The quality of the prediction may depend on accuracies of predictions generated from the training data, number of cumulative runs / reads, and / or quality scores of the data.

[0060] In an embodiment, the blocks 230 through 250 can be repeated for each cluster of the sequence data, and the consensus call sequence for each cluster can be collected and stored in a BAM file, as is well-known to those of skill in the art. The consensus call sequence information can then be processed by downstream elements of the bioinformatics pipeline. For example, the BAM file can be provided as input to a variant caller in order to call variants associated with the sequence data generated by the sequencer. The use of the consensus call data from the CNN may be more accurate than conventional methods such as majority vote. Furthermore, the structure of the CNN and limits to size made possible by the information collected to generate the featurePATENTClient Reference No.: P39780-WO-1 vectors can improve computational complexity allowing for much faster run times than other types of machine learning models.

[0061] FIG. 3 illustrates an architecture of a neural network 300 used to determine a consensus sequence, in accordance with at least some embodiments. The neural network 300 shown in FIG. 3 is a convolutional neural network (CNN) that includes a number of layers. At an input layer 310, the input feature vector(s) are received. The feature vectors can include, e.g., an encoding of the base calls and q-scores of a subset of sequence reads, as described above. Allernali vely, the feature vectors can include the aggregation statistics. In an embodiment, the input layer 310 can apply a convolution kernel to the elements of the feature vectors to increase a number of channels of the input feature vectors for further processing by lower layers of the CNN 300. For example, a 1 x 1 convolution kernel can be applied to each feature vector that increases or decreases the importance of certain elements of the input feature vector to the output, and different kernel weights can be stored for each of a number of channels (e.g., 128 channels). In other embodiments, the input layer applies a 2D convolution filter to the input feature vector, which may be used to adjust the dimensions of the input feature map to dimensions used by the hidden layers 320. In yet other embodiments, the input layer 310 does not apply a convolution filter and simply passes the input feature vector to the first hidden layer 320A.

[0062] The CNN 300 includes hidden layers 320A, 320B , 320C, 320D, each layer corresponding with a convolution operation using a trained filter kernel and bias values. As shown, the convolution operations may each be one-dimensional convolution operations. However, depending on the structure of the feature vectors, in some embodiments, the convolution operations may be two-dimensional convolution operations. Each intermediate feature map output of the hidden layers 320A, 320B, and 320C is further processed by an activation function (e.g., Sigmoid function, rectified linear unit (ReEU), etc.) to manage nonlinearities introduced by the data. Although the hidden layer 320D is shown without a corresponding activation function, in some embodiments, the hidden layer 320D may also include an activation function. Finally, an output layer 330 at block 330 produces a consensus sequence output comprising, at least, a consensus base call and corresponding quality metric associated with the base call for each position of the consensus sequence. In some embodiments, the output layer 300 may also include a one-dimensional convolution operation. In other embodiments, the output layer 300 may be a fully connected layer.PATENTClient Reference No.: P39780-WO-1

[0063] It will be appreciated that the CNN 300 includes five layers. However, it will be appreciated by one of skill in the art that greater or fewer hidden layers 320 can be included in the CNN 300 to balance accuracy and / or processing efficiency as suited for a particular use case. Experimentation determined that the use of four layers provided a reasonable balance between speed of the computations and quality of the output (e.g., see Table 1 and FIGs. 9A-9D).

[0064] The CNN 300 may be trained using a training dataset. The training dataset may include sequence data, which may be aligned and sorted into a plurality of clusters. For each cluster of aligned sequences in the training dataset, a corresponding ground-truth consensus call sequence may be defined and included in the training dataset. Training may be performed according to any number of well-known supervised learning techniques, such as using back- propagation with gradient descent. During training, the accuracy of the model can be monitored using, e.g., a validation dataset, which may be, e.g. a subset of sequence data in the training dataset. The model weights may be adjusted during a number of training passes, with the model parameters being updated during each pass. Training may be stopped once the accuracy converges or once the accuracy satisfies some other criteria.

[0065] FIG. 4 illustrates a process for aligning and clustering sequence read data, in accordance with at least some embodiments. A set of sequence reads 400 (e.g., obtained from a sequencer instrument 102) can represent strands having 5’ and 3’ ends with a series of base calls (A,T,G,C, -) read in forward (e.g., from 5’ to 3’ end) and / or reverse direction (e.g., from 3’ to 5’ end). The reads 400 are aligned and sorted into clusters 410 by UMI and / or starting / ending positions by methods known to those of ordinary skill in the art. The alignment technique may include multiple sequence alignment, which may use (for example) a progressive, iterative, Hidden-Markov Model (HMM) and / or intramolecular or intermolecular consensus technique to align and cluster the sequences. The alignment may include introducing gaps between at least some sequential nucleotides to account for insertion and / or deletion type errors. Using a consensus of multiple reads of the same DNA sample with the CNN can lead to a more reliable consensus sequence prediction, particularly where a sequencing technology frequently generates sequencing errors.

[0066] FIGs. 5A-5B illustrate a process for aligning sequence read data, in accordance with at least some embodiments. As shown in FIG. 5A, seven exemplary sequences, S1-S7, are associated with a cluster 410. Each of si-s? is a sequence read having, e.g., the same UMI and / or starting / ending position. Each of si -s? may be generated by sequencing different moleculesPATENTClient Reference No.: P39780-WO-1 amplified from the same starting molecule during library preparation, or generated by sequencing a single molecule that includes multiple copies of the same strand of the target molecule.

[0067] Each of the exemplary sequences ,v / -.V7 inay have been generated (for example) using one or more sequencing techniques, using one or more different sequencing instruments, and / or under the control of one or more different entities. Each of exemplary sequences si -s? includes an ordered set of base calls corresponding to a nucleic acid sequence of the target molecule. In some instances, each of one or more of the exemplary sequences si -s? may have been generated using a different portion of one or more amplified or cloned sequences.

[0068] As shown in FIG. 5B, the exemplary sequences ,v / -.V7 inay be converted into an alignment result where the exemplary sequences si-s?are aligned to each other. The alignment result includes multiple gaps (-) between sequential bases to account for potential deletions and / or insertions in some of the sequences. More specifically, the exemplary sequences si -s? of Fig. 5A may have different lengths. The alignment process, therefore, may introduce gaps into the sequences such that the alignment result comprises sequences of the same length, which better aligns with the reference sequence.

[0069] The alignment may have been performed using one or more alignment techniques disclosed herein and / or well-known to those of skill in the art. In the illustrated instance, generally, for each position ( / ?,) of the alignment result, the position includes (across all sequences) only a single nucleotide or gap. It will be appreciated that the length of the alignment result may not match a corresponding length of the aligned sequences. For example, the alignment result can be longer than the corresponding aligned sequence if a number of gaps are inserted into the sequence. As another example, the alignment result can be shorter than the aligned sequence if a number of bases were deleted from the sequence or if two or more bases were replaced by an inserted gap.

[0070] The alignment technique may insert gaps to shift one or more nucleotides of a particular sequence to better align those nucleotides with a corresponding portion of the reference sequence. For example, where a particular sequence includes a base call at a position which is inconsistent with the base of the reference sequence, such a discordant base may indicate the presence of a deletion in that sequence that shifts one or more nucleotides of that sequence one or more positions relative to the reference sequence. Of course, a discordant base can also be the result of a mutation. Therefore, an alignment function can be used to calculate a likelihood forPATENTClient Reference No.: P39780-WO-1 whether a gap should be inserted into the sequence based on the context of nucleotides surrounding that position in the sequence, other sequences in the cluster, and the corresponding reference sequence. For example, as shown in FIG. 5B, with respect to one posilion (ps), a particular sequence (52) includes a different base than is present in the other sequences. It will be appreciated that configurations and / or hyperparameters of an alignment function may influence whether to accept nucleotide inconsistency across sequences (e.g., to a particular degree) rather than introducing one or more additional gaps to allow for the nucleotides to be shifted to different positions.

[0071] FIG. 5C illustrates a one-hot encoding for an aligned sequence (s ), in accordance with an embodiment. Specifically, for each position, a set of binary values are determined - each corresponding to a particular nucleotide base (A,C,G,T) or a gap (-). The value can be set to 1 if the called nucleotide at the position for the particular aligned sequence is equal to the corresponding base (or gap) and otherwise can be set to 0. Thus, in some instances, each position is associated with one, and only one, value of 1. These binary values may be determined for each of the exemplary sequences S1-S7 of the alignment result of FIG. 5B.

[0072] It will be appreciated that the exemplary sequences of FIGs. 5A-5C are not annotated as either forward (sense) or reverse (anti- sense) strands. However, in some sequencing technologies, a pair of sequence reads can be associated with each other as paired forward and reverse reads. It will be appreciated that, in these cases, the one-hot encoding of a forward strand can be concatenated with the one-hot encoding for a corresponding reverse strand. For example, instead of the encoding data structure having 5 rows for a single strand, the data structure can have 10 rows, with a first 5 rows used for the forward strand and a second 5 rows used for the reverse strand. Furthermore, it will be appreciated that any arrangement of the one -hot encoding data, such as by interleaving data for the forward and reverse strands every other row of the matrix, or transposing the matrix such that rows correspond to positions and columns correspond to bases / gaps, are contemplated as within the scope of the present disclosure.

[0073] FIGs. 6A-6B show a portion of a feature vector generated according to the one-hot encoding scheme, in accordance with at least some embodiments. In particular, the first portion (e.g., the first 7 columns) of the feature vector 600 depicted in FIG. 6A corresponds to the first position, pi, in the genome. The depicted representations of the called bases, using one-hot encoding, correspond to the called bases in the alignment result at the first position pi across all sequences (si-s?) in the cluster. Each column of the feature vector 600 also includes a strandPATENTClient Reference No.: P39780-WO-1 direction indicator (f / r) using binary encoding to indicate whether each read is forward (0) or reverse (1). Finally, each column of the feature vector 600 also includes a corresponding quality metric (e.g., q-score) for the called base.

[0074] It will be appreciated that the quality metric reflects a degree of confidence that a portion of raw data (e.g., electrical signal, intensity signal, etc.) corresponds to a given type of nucleotide. For example, each of the four nucleotides may be associated with a particular wavelength and / or electrical signal signature. If a signal clearly included a peak at a particular wavelength associated with cytosine and / or clearly included an electrical signal signature for cytosine, a quality metric associated with cytosine may be high so as to indicate a high confidence in a prediction that the nucleotide was cytosine. Alternatively, if a signal clearly lacked a peak at the particular wavelength associated with cytosine and / or clearly lacked an electrical signal signature of cytosine, a quality metric associated with cytosine may be lower so as to indicate a lower confidence in a prediction that the nucleotide was cytosine. One or more quality metrics may be output by one or more devices that generate the corresponding sequence data. A quality metric may be (for example) a binary, integer, or real-number value. In some embodiments, the quality score may be a Phred score.

[0075] A second portion of the feature vector 600 includes the one -hot encoded data for the second position (p2)~ of the sequence. As shown in FIG. 6A, all seven sequences have a called cytosine base at the second position. The feature vector 600 may be generated by concalenaling new portions of the feature vector for each position of the aligned sequence.

[0076] While the called bases of the first and second position were concordant across all seven exemplary sequences, it will be appreciated that at position ps there is a discordant base call in the second sequence S2. Six of the exemplary sequences have a gap base call at posilion p5, but the second sequence S2 includes a called thymine base. The feature vector 600 will be generated until all portions corresponding to all positions of the aligned sequence have been concatenated to the portions shown in FIG. 6A.

[0077] It will be appreciated that, in some cases, a quality metric is not provided for called gaps (-), as the gap does not represent a called base but may instead represent a potential insertion or deletion in the sequence read.

[0078] While the feature vector 600 has been shown as a 2D matrix or array, in some other embodiments, the feature vector 600 may be structured differently. In some instances, the featurePATENTClient Reference No.: P39780-WO-1 vector 600 may be transposed. In other words, the feature vector 600 may be 7 columns wide and contain a number of rows corresponding to the number of sequences in the cluster times the number of positions in the aligned sequence. In another instance, as shown in FIG. 6B, the data shown in the feature vector 600 may be structured as a row vector of ^-elements. The vector can be generated similarly to the feature vector 600, but where each row of each portion of the feature vector 600 is simply concatenated with the row below. This one-dimensional structure may be preferable as input to a CNN that is structured to use a ID convolution kernel.

[0079] In particular, in at least one embodiment, the vector-form of feature vector 600 can include the elements for the adenine base followed by the elements for a thymine base, followed by the elements for a cytosine base, followed by the elements for a guanine base, followed by the elements for gaps, followed by the elements for strand indicators, followed by the elements for quality metrics, for each position of the aligned sequence. Thus, the first element of the vector corresponds to a binary value for the called base at the first posilion pi in the first sequence si indicating whether said called base is an adenine base; the second element of the vector corresponds to a binary value for the called base at the first posilion pi in the second sequence si indicating whether said called base is an adenine base; and so forth.

[0080] Of course, it will be appreciated that the order of elements in the vector can be rearranged, such as by switching the order of called bases, or having the first row of feature vector 600 correspond to quality metrics, without departing from the scope of the present disclosure. The order used, however, should remain fixed during both training and inference.

[0081] FIGs. 7A-7B show a portion of a feature vector 700 generated according to an aggregation statistics encoding scheme, in accordance with at least some embodiments. As shown in FIG. 7A, the feature vector 700 is structured similarly to feature vector 600 (i.e., as a matrix or array with a number of rows), but instead of indicating the called base in each of the exemplary sequences independently using one-hot encoding, the feature vector 700 includes both aggregation statistics for each posilion across all aligned sequences, and corresponding quality metrics for the position.

[0082] For example, a first column of feature vector 700 includes aggregation statistics for position pi. In an embodiment, the aggregation statistics comprise percentage values that indicate a percentage of aligned sequences in the cluster that include a particular called base at a given position. Thus, for position pi, the percentage is zero for all bases except the guanine base, whichPATENT Client Reference No.: P39780-WO-1 has a percentage of 100 (percentage given as a value between 0 and 1, inclusive). In particular, it will be appreciated that the aggregation statistics may be given as a floating point or fixed point value rather than as an integer or binary value (in the case of one-hot encoding. Alternatively, in some embodiments, the percentage can be given as an integer (0-100), with the percentage rounded to the nearest 1%.

[0083] It will be appreciated that the second column of the feature vector 700 includes quality metrics that, like feature vector 600, represent the average quality score for a particular base called in any of the aligned sequences. Thus, instead of a q-score for a particular base call in a particular read, the quality metrics used in FIG. 7A will be average q-scores over all clustered sequences having a particular called base. While positions with concordant bases have a value of 100 percent in one base and values of 0 in all other bases, it will be appreciated that positions with discordant bases will have percentages greater than zero but less than 100 in two or more bases (or gaps). For example, position ps includes a percentage of 0.1 for thymine and 0.9 for gap. The actual percentages are 1 / 7= 14% and 6 / 7=86%, respectively, which have been rounded to the nearest tenth in FIG. 7A to fit in the Figure. For these discordant base calls, you may have two (or more) different average q-scores for a given genomic posi lion corresponding to each of the different discordant bases at that posi lion.

[0084] Of course, like feature vector 600, feature vector 700 may be rearranged in different form without deparling from the scope of the present disclosure. For example, the quality metrics in every other column may be stacked immediately below the base calls and gap information in the preceding column. In some cases, the element for the quality metric corresponding to a gap call can be omitted, thus a column may include nine elements.

[0085] Furthermore, in some embodiments, aggregation statistics can be computed, for each position, independently for forward and reverse strand directions. In an embodiment, aggregation statistics for all forward strands in a cluster at a given position are included in a first column and aggregation slalislics for all reverse strands in the cluster at the given position are included in a second column immediately adjacent the first column. In such embodiments, quality metrics can be included in the same column as the corresponding aggregation slalislics. In another example, aggregation slalislics for both the forward and reverse strand can be included in a single column, such as in the first 10 elements of each column. In such embodiments, quality metrics can be included in the same column as the corresponding aggregation statistics, or in the column next to the aggregation slalislics.PATENTClient Reference No.: P39780-WO-1

[0086] Table 1 shown below shows a comparison between processing input data with different machine learning models according to various embodiments herein and other models (e.g., recurrent neural network (RNN), Long Short Term Memory (LSTM), etc.). The first column of Table 1 indicates the model used and the hyperparameters for the model. The second and third columns of Table 1 indicate the accuracy of the model using a set of validation data. Accuracy is computed by comparing the consensus call output by the model to the ground-truth calls for a number of clusters in the validation data. “Hard Data” refers to a subset of clusters within the validation data where the ground truth consensus call at any position is different from a result obtained using a majority voting (or if majority voting would result in a tie between two base calls). The accuracy is reported in Table 1 using a Phred Quality score. The last column in Table 1 indicates run time for the model.

[0087] As shown in Table 1, the running time of the CNN 300, compared to a prior art LSTM model, is vastly improved. Lor example, limiting the number of channels to 128 and using a feature dimension of up to 20 (i.e., the number of rows of each feature vector), it is shown that running time can be approximate 1 / 5 of a corresponding LSTM model with half the number of channels and a much larger number of features, while maintaining similar accuracy.Table 1PATENTClient Reference No.: P39780-WO-1Exemplary Computer System

[0088] FIG. 8 illustrates an example computer system that may be utilized to implement techniques disclosed herein. Any of the computer systems mentioned herein, such as for hosting the systems and implementing the processes described for calculating consensus sequences, may utilize any suitable number of subsystems. Examples of such subsystems are shown in FIG. 8 in computer system 800. In some embodiments, a computer system includes a single computer apparatus, where the subsystems can be the components of the computer apparatus. In other embodiments, a computer system can include multiple computer apparatuses, each being a subsystem, with internal components. A computer system can include desktop and laptop computers, tablets, mobile phones, telecommunication devices or other mobile devices. In some embodiments, a cloud infrastructure (e.g., Amazon Web Services), a graphical processing unit (GPU), etc., can be used to implement the disclosed techniques.

[0089] The subsystems shown in FIG. 8 are interconnected via a system bus 75. Additional subsystems such as a printer 74, keyboard 78, storage device(s) 79, monitor 76, which is coupled to display adapter 82, and others are shown. Peripherals and input / output (I / O) devices, which couple to I / O controller 71, can be connected to the computer system by any number of means known in the art such as input / output (I / O) port 77 (e.g., USB, FireWire®). For example, I / O port 77 or external interface 81 (e.g. Ethernet, Wi-Fi, etc.) can be used to connect computer system 800 to a wide area network such as the Internet, a mouse input device, or a scanner.

[0090] The interconnection via system bus 75 allows the central processor 73 to communicate with each subsystem and to control the execution of a plurality of instructions from system memory 72 or the storage device(s) 79 (e.g., a fixed disk, such as a hard drive, or ophcal disk), as well as the exchange of information between subsystems. The system memory 72 and / or the storage device(s) 79 may embody a computer readable medium. Another subsystem is a data collection device 85, such as a camera, microphone, accelerometer, and the like. Any of the dataPATENT Client Reference No.: P39780-WO-1 mentioned herein can be output from one component to another component and can be output to the user.

[0091] A sequencer 90 (e.g., a nanopore sequencer) is connected through external interfance 81 for providing sequencing data to a data collection device 85 and / or storage devices 79.

[0092] A computer system can include a plurality of the same components or subsystems, e.g., connected together by external interface 81 or by an internal interface. In some embodiments, computer systems, subsystem, or apparatuses can communicate over a network. In such instances, one computer can be considered a client and another computer a server, where each can be part of a same computer system. A client and a server can each include multiple systems, subsystems, or components.

[0093] Aspects of embodiments can be implemented in the form of control logic using hardware (e.g. an application specific integrated circuit or field programmable gate array) and / or using computer software with a generally programmable processor in a modular or integrated manner. As used herein, a processor includes a single-core processor, multi-core processor on a same integrated chip, or multiple processing units on a single circuit board or networked. Based on the disclosure and teachings provided herein, a person of ordinary skill in the art will know and appreciate other ways and / or methods to implement embodiments of the present invention using hardware and a combi nation of hardware and software.

[0094] Machine learning models utilized herein may include one or more of a Naive Bayes (NB) model, a logistic regression (LR) model, a random forest (RF) model, a support vector machine (SVM) model, an artificial neural network model, a rnul (i layer perceptron (MLP) model, a convolutional neural network (CNN), a Large Language model (LLM), and / or other machine learning or deep leaning models, etc. The machine learning models can be updated / trained using a supervised learning technique, an unsupervised learning technique, etc.

[0095] Any of the software components or functions described in this application may be implemented as software code to be executed by a processor using any suitable computer language such as, for example, Java, C, C++, C#, Objective-C, Swift, or scripting language such as Perl or Python using, for example, conventional or object-oriented techniques. The software code may be stored as a series of instructions or commands on a computer readable medium for storage and / or transmission. A suitable non-transitory computer readable medium can include random access memory (RAM), a read only memory (ROM), a magnetic medium such as a hard-PATENTClient Reference No.: P39780-WO-1 drive or a floppy disk, or an optical medium such as a compact disk (CD) or DVD (digital versatile disk), flash memory, and the like. The computer readable medium may be any combination of such storage or transmission devices.

[0096] Such programs may also be encoded and transmitted using carrier signals adapted for transmission via wired, optical, and / or wireless networks conforming to a variety of protocols, including the Internet. As such, a computer readable medium may be created using a data signal encoded with such programs. Computer readable media encoded with the program code may be packaged with a compatible device or provided separately from other devices (e.g., via Internet download). Any such computer readable medium may reside on or within a single computer product (e.g. a hard drive, a CD, or an entire computer system), and may be present on or within different computer products within a system or network. A computer system may include a monitor, printer, or other suitable display for providing any of the results mentioned herein to a user.

[0097] Any of the methods described herein may be totally or partially performed with a computer system including one or more processors, which can be configured to perform the steps. Thus, embodiments can be directed to computer systems configured to perform the steps of any of the methods described herein, potentially with different components performing a respective steps or a respective group of steps. Although presented as numbered steps, steps of methods herein can be performed at a same time or in a different order. Additionally, portions of these steps may be used with portions of other steps from other methods. Also, all or portions of a step may be optional. Additionally, any of the steps of any of the methods can be performed with modules, units, circuits, or other means for performing these steps.

[0098] The specific details of particular embodiments may be combined in any suitable manner without departing from the spirit and scope of embodiments of the invention. However, other embodiments of the invention may be directed to specific embodiments relating to each individual aspect, or specific combinations of these individual aspects.

[0099] The above description of example embodiments of the invention has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the in vention to the precise form described, and many modifications and variations are possible in light of the teaching above.PATENT Client Reference No.: P39780-WO-1

[0100] A recitation of "a", "an" or "the" is intended to mean "one or more" unless specifically indicated to the contrary. The use of “or” is intended to mean an “inclusive or,” and not an “exclusive or” unless specifically indicated to the contrary. Reference to a “first” component does not necessarily require that a second component be provided. Moreover reference to a “first” or a “second” component does not limit the referenced component to a particular location unless expressly stated.

[0101] All patents, patent applications, publications, and descriptions mentioned herein are incorporated by reference in their entirety for all purposes. None is admitted to be prior art.

Claims

PATENTClient Reference No.: P39780-WO-1CLAIMS1. A method of determining a consensus molecular sequence, the method comprising: sequencing, using a sequencing device, a biological sample to obtain sequence data, wherein the sequence data comprises a plurality sequence reads, each sequence read comprising a plurality of base calls and corresponding quality scores (q-scores) generated by the sequencing device; aligning the plurality of sequence reads to a reference genome; sorting the aligned reads into one or more clusters; generating a feature vector for each cluster; and processing the feature vector by a trained convolutional neural network (CNN) model having a plurality of convolutional layers to generate a consensus call sequence for the cluster.

2. The method of claim 1, wherein the feature vector comprises one -hot encoding of the base calls in each of the sequence reads in the cluster.

3. The method of claim 2, wherein the feature vector further comprises a quality metric that indicates q-score for the base call generated by the sequencing device.

4. The method of claim 3, wherein the feature vector further comprises a strand indicator that indicates whether a particular base call is associated with a forward or reverse strand of the target molecule.

5. The method of claim 1, wherein the feature vector comprises aggregation statistics for base calls at each genomic position for the aligned sequence of the cluster.

6. The method of claim 5, wherein the feature vector further comprises quality metrics for each genomic position, and wherein the quality metrics comprise an average quality score for each of up to four called bases at the genomic position across all sequence reads in the cluster.

7. The method of claim 1, wherein sorting the aligned reads into one or more clusters comprises sorting into clusters based on at least one of a unique molecular i den tillers (UMIs), starting position, or ending position of each aligned read, and wherein all sequence reads in a cluster include the same UMI.PATENTClient Reference No.: P39780-WO-18. The method of claim 7, wherein aligning the sequence reads to the reference genome comprises, for at least one sequence read, replacing one or more base calls in a sequence with a gap or inserting a gap between two adjacent base calls in the sequence read.

9. The method of claim 1, wherein the CNN includes at least three hidden convolutional layers, each hidden convolutional layer applying a convolution operation to input features to generate filtered output features and an activation function to the filtered output features to generate a set of activations passed to the next layer of the CNN.

10. The method of claim 1, further comprising training the CNN in accordance with a training data set, wherein the training data set includes sequence data for a plurality of clusters, and the sequence data for each cluster in the training data set includes a plurality of aligned sequence reads and a ground-truth consensus call sequence for the cluster.

11. The method of claim 1, further comprising collecting a number of consensus call sequences for a plurality of clusters of the sequence data for the sample, and storing the consensus call sequences in a BAM file.

12. The method of claim 11, further comprising passing the BAM file to a downstream variant caller to identify one or more variants based on the col lection of the number of consensus call sequences.

13. A system for dsetermining a consensus call sequence, the system comprising: a sequencing device configured to sequence a biological sample to obtain sequence data, wherein the sequence data comprises a plurality sequence reads, each sequence read comprising a plurality of base calls and corresponding quality scores (q-scores) generated by the sequencing device; and a computing system, communicatively coupled to the sequencing device, the coinpuling system comprising at least one processor executing instructions that cause the computing device to: align the plurality of sequence reads to a reference genome; sort the aligned reads into one or more clusters;PATENTClient Reference No.: P39780-WO-1 generate a feature vector for each cluster; and process the feature vector by a trained convolutional neural network (CNN) model having a plurality of convolutional layers to generate a consensus call sequence for the cluster.

14. The system of claim 13, wherein the feature vector comprises one-hot encoding of the base calls in each of the sequence reads in the cluster.

15. The system of claim 14, wherein the feature vector further comprises at least one of: a quality metric that indicates q-score for the base call generated by the sequencing device, or a strand indicator that indicates whether a particular base call is associated with a forward or reverse strand of the target molecule.

16. The system of claim 13, wherein the feature vector comprises aggregation statistics for base calls at each genomic position for the aligned sequence of the cluster.

17. The system of claim 16, wherein the feature vector further comprises quality metrics for each genomic position, and wherein the quality metrics comprise an average quality score for each of up to four called bases at the genomic position across all sequence reads in the cluster.

18. The system of claim 13, wherein sorting the aligned reads into one or more clusters comprises sorting into clusters based on at least one of a unique molecular identifiers (UMIs), starting position, or ending position of each aligned read, and wherein all sequence reads in a cluster include the same UMI.

19. The system of claim 13, wherein aligning the sequence reads to the reference genome comprises, for at least one sequence read, replacing one or more base calls in a sequence with a gap or inserting a gap between two adjacent base calls in the sequence read.

20. The system of claim 13, the computing device further configured to: collect a number of consensus call sequences for a plurality of clusters of the sequence data for the sample; store the consensus call sequences in a BAM file; and process the BAM file to a downstream variant caller to identify one or more variants based on the collection of the number of consensus call sequences.

Images