JP5863396B2 - DNA sequence decoding system, DNA sequence decoding method and program - Google Patents

Description

  The present invention relates to a DNA sequence decoding technique using a massively parallel sequencer, and more particularly to a DNA sequence decoding technique that is difficult to read.

  A massively parallel sequencer (hereinafter also referred to as “sequencer”) that is widely used arranges a large number of clusters of DNA fragments amplified from a single molecule on a substrate and simultaneously sequences a large amount of DNA sequences. The sequencing method is based on adding a fluorescently labeled nucleotide probe to each of the DNA fragments and extending the complementary strand. A plurality of types of fluorescent dyes are used, whereby each DNA base is encoded.

  The sequencer excites the fluorescent dye to emit light in each cycle of the extension reaction, and acquires a substrate image of each fluorescent color. Thereafter, the fluorescence intensity of each color of fluorescence in each cycle is measured for each cluster of DNA fragments. The sequence decoding system determines the base of the corresponding position based on the measured fluorescence intensity in each cycle for the DNA sequence of each cluster.

  Here, it is expected that each cluster of DNA fragments ideally has an intensity only in one fluorescent color corresponding to the base at the corresponding position in each cycle of the extension reaction and is not detected in other colors. Is done. However, in reality, noise due to factors such as delay / advance of extension reaction in the cluster and crosstalk of fluorescence occurs, and there are cases where detection is performed in a plurality of colors. This causes deterioration in the reading accuracy of the DNA sequence by the sequence decoding system. Moreover, the influence of noise increases as the extension reaction proceeds. For this reason, when the influence of noise is assumed, it is necessary to limit the readable sequence length.

  As a method of sequence decoding in consideration of the influence of noise, there is a method of estimating the fluorescence color in each cycle by modeling extension reaction delay / advance, fluorescence crosstalk and the like parametrically (Non-patent Document 1). However, each noise factor has a complicated dependency on the cycle and chemical reaction conditions, and it is difficult to model completely. Therefore, a machine learning approach such as support vector machine (SVM) is applied to directly learn the relationship between the fluorescence intensity of four colors and the correct sequence for each cycle obtained from the sequencer based on a known DNA sequence. A method for estimating a fluorescent color in a cycle is also performed (Non-Patent Document 2).

Kao et al., “BayesCall: A model-based base-calling algorithm for high-throughput short-read sequencing” Genome Research, 19, 1884, 2009 Kircher et al., “Improved base calling for the Illuina Genome Analyzer using machine learning strategies” Genome Biology, 10, R83, 2009

  In general, the reading accuracy of DNA sequences is not uniform. For example, it is known that the reading accuracy may deteriorate particularly depending on the arrangement. The occurrence of noise in sequencers based on the chemical extension method is considered to depend greatly on the characteristics of the DNA sequence to be read. For example, (1) a high GC content, (2) a double-base repeat sequence, (3) a palindromic (palindrome) sequence, and the like are more likely to form a higher order structure, which greatly affects the chemical reaction during elongation. It is considered to have an effect.

  However, in a conventional sequence decoding system, a single model constructed using a general reference genome as a control is applied to all DNA sequences. That is, the nature of noise that depends on the difference in the characteristics of DNA sequences is not fully taken into account in conventional sequencing systems.

  The present invention has been made in view of the above situation, and provides a mechanism that can be expected to improve the accuracy of sequence analysis of DNA sequences that are difficult to read (so-called difficult-to-read DNA sequences).

  The present invention decodes sequences based on sequence characteristics of obfuscated DNA sequences. More specifically, the process of classifying obfuscated DNA sequences into feature groups based on the features of the sequences, and learning the criteria for determining the fluorescent color in each cycle using known DNA sequence data for each group feature Provided is a method for executing processing and processing for decoding a sequence by decoding a sequence of the unknown experimental DNA sequence by applying a criterion learned from the feature group on the sequence.

  According to the present invention, it is possible to improve the decoding accuracy of difficult-to-read DNA sequences. Problems, configurations, and effects other than those described above will be clarified by the following description of embodiments.

The figure explaining the flow of the whole process in the arrangement | sequence decoding system which concerns on an example. The figure which shows the structural example of the arrangement | sequence decoding system which concerns on an example. The figure explaining the extraction method of an obfuscated DNA sequence. The figure explaining the concept of the feature classification of a DNA sequence. The figure which shows the outline | summary of the learning process of a sequence decoding system. The flowchart explaining the procedure of a learning process. The figure which shows the structural example of a fluorescence color criteria database. The figure which shows the structural example of a fluorescence color likelihood database. The figure showing the outline | summary of the sequence estimation process of a sequence decoding system. The flowchart explaining the procedure of sequence estimation processing.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In addition, this invention is not limited to embodiment mentioned later, Various modifications are included in this invention. For example, an embodiment described later may have an additional configuration, and a part of the configuration may not be included. Moreover, you may substitute the one part structure of the example mentioned later to another structure.

  In addition, each or all of the configurations, functions, processing units, processing units, and the like described below may be realized as part or all of them as, for example, an integrated circuit or other hardware. In addition, each configuration, function, processing unit, processing unit, and the like, which will be described later, may be realized by the processor interpreting and executing a program that realizes each function. That is, each configuration, function, processing unit, processing unit, and the like described later may be realized as software. Information such as programs, tables, and files for realizing each configuration and the like can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or a storage medium such as an IC card, an SD card, or a DVD.

  Note that components having the same function are denoted by the same or related reference symbols throughout the drawings for describing the embodiment, and the repetitive description thereof will be omitted. In the following embodiments, the description of the same or similar parts will not be repeated in principle unless particularly necessary.

[Flow of overall processing]
FIG. 1 shows the overall flow of processing executed in the sequence decoding system according to the embodiment. This sequence decoding system has three stages: an obfuscated DNA sequence search stage 110, a learning stage 120 using known DNA sequence data 121, and an estimation stage 130 for unknown experimental DNA sequence data 131.

When decoding the unknown experimental DNA sequence 131, the DNA sequence is estimated based on the fluorescence intensity sequence 132 for all cycles of the fluorescence intensity acquired from the sequencer for each cycle of the extension reaction, and the estimated DNA is estimated as 133. The array 134 and its reliability 135 are output. For example, for each cycle of the extension reaction, fluorescence intensity (I _a , I _b , I _c , I _d ) of four colors (a, b, c, d) corresponding to four bases is acquired from the sequencer. Then, the fluorescence intensity arrangement 132 for all cycles (ie, (I _a , I _b , I _c , I _d ) ^{cycle 1} , (I _a , I _b , I _c , I _d ) ^{cycle 2} , ... (I _a , I _b , I _c , I _d ) Estimate DNA sequence based on ^cyclen ).

[Details of system configuration and processing operations]
FIG. 2 shows a configuration example of the sequence decoding system according to the embodiment. This sequence decoding system includes an input / output device 210, an analysis device 220 having an obfuscated DNA sequence analysis unit 221, a learning unit 222, and an estimation unit 223, and a storage device 230. In the embodiment, the analysis device 220 realizes a function executed at each stage described later as a processing function of a program executed on a computer.

[Search stage]
The obfuscated DNA sequence analysis unit 221 collects, as an obfuscated DNA sequence, a region in which many errors are detected when sequences are executed on various actual genomes in the obfuscated DNA sequence search stage 110 (FIG. 1). Furthermore, the process which classifies the feature on the sequence which these obfuscated DNA sequences have for every stage is performed.

  FIG. 3 shows an example of a method for determining an obfuscated DNA sequence. First, the read sequence after the sequence is mapped to the reference sequence. When comparing the mapped lead sequence with the reference sequence, it is not decoded within the range of the length of the lead sequence (indicated as base N in FIG. 3), or bases that are not identical for a certain ratio or more. The existing region is determined as an obfuscated DNA sequence.

FIG. 4 shows how DNA sequences determined to be difficult to read are classified (grouped) according to their characteristics. Classification criteria include four-color fluorescence intensity sequences obtained from the sequencer (ie, (I _a , I _b , I _c , I _d ) ^cycle1 , (I _a , I _b , I _c , I _d ) ^cycle2 , ... ( Information obtained by clustering analysis using I _a , I _b , I _c , I _d ) ^cyclen ) as a feature vector is used. In addition, information obtained by performing clustering analysis on a feature space obtained by nonlinearly transforming the above-described feature vector is used as the classification criterion. In this specification, a set of DNA sequences after classification is referred to as feature groups 1, 2,.

[Learning stage]
In the learning stage 120 (FIG. 1), the learning unit 222 detects the determination tendency of the fluorescent color that frequently appears in each cycle of the DNA sequence classified into the feature group, and uses these to detect the fluorescent color unique to each feature group. Learning is performed as the criterion 126 (FIG. 1).

FIG. 5 shows an outline of a processing procedure executed by the learning unit 222 in the learning stage 120. As a premise of the learning process, known DNA sequence data 121 is prepared for each feature group of the obfuscated DNA sequence. The known DNA sequence data 121 includes four-color fluorescence intensity sequences 124 (that is, (I _a , I _b , I _c , I _d ) ^cycle1 , (I _a , I _b ) in each cycle of the extension reaction obtained from the sequencer. , I _c , I _d ) ^{cycle 2} ,... (I _a , I _b , I _c , I _d ) ^{cycle n} ) and a correct fluorescent color array 125 determined from the correct DNA sequence. The learning unit 222 uses these array data and performs the following learning process for each feature group. A part of the array data belonging to each feature group is used as training data 122 and the rest is used as fluorescence color likelihood calculation data 123.

  The learning unit 222 includes a fluorescent color determination reference learning unit 501, a fluorescent color determination unit 502, and a fluorescent color likelihood calculation unit 503. The fluorescent color determination criterion learning unit 501 refers to the four-color fluorescent intensity array 124 and the correct fluorescent color array 125 constituting the training data 122 for each feature group, and determines the fluorescent color for determining the fluorescent color in each cycle. The reference 126 is learned and stored in the fluorescent color determination reference database 231. Details of this learning process will be described later. The fluorescent color determination unit 502 and the fluorescent color likelihood calculation unit 503 use the fluorescent color likelihood calculation data 123 of each feature group, and for each cycle, the fluorescent color likelihood 127 for each of the four fluorescent colors (FIG. 1). ) And is stored in the fluorescence color likelihood database 232.

  FIG. 6 shows a procedure example of learning processing executed for a certain feature group in the learning stage 120 (FIG. 1). In step S <b> 1, the learning unit 222 acquires training data 122 belonging to a feature group to be learned from a storage area (not shown).

In step S <b> 2, the fluorescent color determination reference learning unit 501 of the learning unit 222 reads the four-color fluorescence intensity array 124 and the correct fluorescent color array 125 from the training data 122. Next, the fluorescence color determination reference learning unit 501 learns the relationship between the arrangement of all the fluorescence intensities of the four colors or the arrangement of a partial cycle and the correct fluorescence color arrangement, and determines the fluorescence color in each cycle. A fluorescent color criterion 126 is derived. For example, a support vector machine (SVM) is used for learning. For example, a fluorescent color determination criterion 126 for determining the fluorescent color in the cycle i from the fluorescent intensity of the cycle i and the preceding and subsequent cycles i−1 and i + 1 is derived. In this case, the fluorescent color determination reference learning unit 501 uses the fluorescence intensity array (that is, (I _a , I _b , I _c , I _d ) ^{cycle (i-1)} , (I _a , I _b , I _c , I _d ) ^cyclei , (I _a , I _b , I _c , I _d ) ^{cycle (i + 1)} ) and correct fluorescent color ^{x i} are input to SVM and learned to determine the fluorescent color in cycle i A fluorescent color criterion 126 is derived.

  In step S <b> 3, the learning unit 222 stores the fluorescent color determination criterion 126 in each cycle of the extension reaction derived by the fluorescent color determination criterion learning unit 501 in the fluorescent color determination criterion database 231 of the storage device 230. FIG. 7 shows a configuration example of a database that stores support vectors as fluorescent color determination criteria 126 learned for each cycle of each feature group. Note that the number of stored support vectors is arbitrary, and one or a plurality of fluorescent color criteria 126 are stored for one cycle.

  In step S <b> 4, the learning unit 222 acquires the remainder of the known DNA sequence data 121 belonging to the feature group as fluorescence color likelihood calculation data 123. At this time, the fluorescence color likelihood calculation data 123 is provided to the fluorescence color determination unit 502 and the fluorescence color likelihood calculation unit 503.

  In step S <b> 5, the learning unit 222 searches the fluorescent color determination criterion database 231, and extracts the fluorescent color determination criterion 126 using the DNA sequence having the characteristic as training data.

In step S <b> 6, the fluorescence color determination unit 502 acquires the fluorescence intensity array 124 of the four colors from the fluorescence color likelihood calculation data 123 for each feature group, and uses the fluorescence color determination reference 126 to indicate the fluorescence color in each cycle. Determine. For example, when learning the fluorescent color determination standard 126 for determining the fluorescent color in the cycle i using the fluorescent intensity of the cycle i and the preceding and subsequent cycles i-1, i + 1, the fluorescent color determining unit 502 From the intensity array 124 {(I _a , I _b , I _c , I _d ) ^{cycle (i-1)} , (I _a , I _b , I _c , I _d ) ^cyclei , (I _a , I _b , I _c , I _d ) ^{cycle (i + 1)} } is input and the fluorescent color judgment criterion 126 is applied to judge the fluorescent color in cycle i.

At this time, the fluorescent color likelihood calculating unit 503 compares the fluorescent color array determined by the fluorescent color determining unit 502 with the correct fluorescent color array 125, and the four color fluorescent color likelihoods P (x ′ _i) in each cycle. | x _i ) is derived. Here, x _i is the correct fluorescence color in cycle i, and x ′ _i is the determination fluorescence color in cycle i.

In step S <b> 7, the learning unit 222 stores the fluorescence color likelihood 127 of the four colors calculated for each cycle of the extension reaction by the fluorescence color likelihood calculation unit 503 in the fluorescence color likelihood database 232. FIG. 8 shows a configuration example of a database that stores four colors of color likelihood P (x ′ _i | x _i ) for each cycle of each feature group.

[Estimation stage]
In the estimation stage 130 (FIG. 1), the estimation unit 223 uses the fluorescence color determination reference database 231 and the fluorescence color likelihood database 232 to estimate the fluorescence color in each stage of the unknown experimental DNA sequence data 131. FIG. 9 shows an outline of a processing procedure executed by the estimation unit 223 in the estimation stage 130.

The unknown experimental DNA sequence data 131 that is the premise of the estimation process is the fluorescence intensity sequence 132 (ie, (I _a , I _b , I _c , I _d ) ^cycle1 in each cycle of the extension reaction obtained from the sequencer. , (I _a , I _b , I _c , I _d ) ^cycle2 ,... (I _a , I _b , I _c , I _d ) ^cyclen ). The estimation unit 223 outputs the estimated DNA sequence 134 and its reliability 135 as the estimation result 133.

  The estimation unit 223 includes a sequence feature determination unit 901, a fluorescent color determination unit 902, and a DNA sequence likelihood calculation unit 903. The sequence feature discriminating unit 901 discriminates the features of the DNA sequence included in the unknown experimental DNA sequence data 131 and discriminates which of the feature groups generated for the known DNA sequence data 121 belongs. The fluorescent color determination unit 902 and the DNA sequence likelihood calculation unit 903 use the fluorescent color determination standard 126 and the emission color likelihood 127 of the feature group obtained from the determination result, and estimate the DNA sequence of the fluorescence intensity sequence 132 to be estimated. 134 and confidence 135 (ie, DNA sequence likelihood) are calculated.

FIG. 10 shows a procedure example of the estimation process executed in the estimation stage 130 (FIG. 1).
In step S11, the sequence feature determination unit 901 and the fluorescent color determination unit 902 acquire unknown experimental DNA sequence data 131 from the sequencer. The unknown experimental DNA sequence data 131 may be acquired from a storage area (not shown).

In step S12, the sequence feature discriminating unit 901 obtains the fluorescence intensity sequences 132 of four colors for all cycles obtained from the sequencer (that is, (I _a , I _b , I _c , I _d ) ^cycle1 , (I _a , I _b , I _c , I _d ) ^cycle2 ,… (I _a , I _b , I _c , I _d ) ^cyclen ) as a feature vector, or a clustering analysis on a feature space obtained by nonlinear transformation of this feature vector As described in FIG. 4, the cluster is classified into one of the clusters constructed by the obfuscated DNA sequence analysis unit 221 (FIG. 2).

  In step S13, the fluorescent color determination unit 902 searches the fluorescent color determination reference database 231 and extracts the fluorescent color determination reference 126 according to the determined feature group.

  In step S14, the DNA sequence likelihood calculating unit 903 searches the fluorescent color likelihood database 232, and extracts the four color fluorescent color likelihoods 127 corresponding to the identified feature groups.

In step S15, the fluorescent color determination unit 902 determines the fluorescent color in each cycle based on the fluorescent color determination criterion 126 unique to the feature group to which the four-color fluorescent intensity array to be estimated belongs. For example, when the criterion for determining the fluorescence color in cycle i is learned from the fluorescence intensity of cycle i and cycles i-1, i + 1 before and after that cycle i, fluorescence color determination unit 902 uses {(I _a , I _b , I _c , I _d ) ^{cycle (i-1)} , (I _a , I _b , I _c , I _d ) ^cyclei , (I _a , I _b , I _c , I _d ) ^{cycle (i + 1)} }, The fluorescent color criterion 126 is applied to the fluorescent color of cycle i.

Further, the DNA sequence likelihood calculating unit 903 calls each cycle called from the fluorescence color likelihood database 232 for the fluorescence color x ′ _i of each cycle determined (estimated) by the fluorescence color determining unit 902 and the identified feature group. The DNA sequence likelihood P (u _i | x ′) is calculated based on the fluorescence color likelihood P (x ′ _i | x _i ) of the four colors in FIG.

  After these processes, the estimation unit 223 outputs the estimated DNA sequence 134 and the reliability 135 as the estimation result 133 for the fluorescence intensity sequence to be estimated.

Incidentally, the DNA sequence likelihood P (u _i | x ′), the estimated DNA sequence b _i , and the reliability R _i are given by the following equations, respectively. Here, x ′ is a determination fluorescent color sequence, and u is a DNA sequence.
P (u _i | x ') ∝ Σ P (x' | u) * P (u)
ΣP (x ′ | u) * P (u) = ΣP (x ′ | x) * P (u)
However, u _i ∈ {A, G, C, T}. When P (u) is known, a known value is used as it is for P (u), and when P (u) is unknown, 1/4 is used for P (u). Also, b _i = argmax _ui P (u _i | x ′) and R _i = −10 log ₁₀ P (u _i = b _i | x ′).

[Summary]
As described above, the sequence decoding system according to the embodiment classifies the obfuscated DNA sequences for each feature group, and learns in advance the fluorescent color criterion 126 and the fluorescent color likelihood 127 that are specific to each feature group. At the time of decoding the DNA sequence data, it is first determined to which feature group the fluorescence intensity sequence 132 of the unknown experimental DNA sequence data 131 belongs, and then the learned fluorescence color criterion 126 for the determined feature group. And the fluorescent color likelihood 127 are applied to estimate the DNA sequence to be decoded. By applying this processing technique, it is possible to increase the accuracy of decoding obfuscated DNA sequences. Further, when the decoding accuracy is improved, the sequence information that can be acquired by one execution of the sequencer can be increased. As a result, it is possible to improve the detection power of sequence variation and improve the accuracy of sequence analysis (for example, mapping, assembly, etc.).

110 ... Searching stage 120 for obfuscated DNA sequences ... Learning stage 121 ... Known DNA sequence data 122 ... Training data 123 ... Fluorescence color intensity calculation data 124 ... Fluorescence intensity array 125 ... Correct fluorescence color array 126 ... Fluorescence color judgment standard 127 ... Fluorescent color likelihood 130 ... estimated stage 131 ... unknown experimental DNA sequence data 132 ... fluorescent intensity sequence 133 ... estimated result 134 ... estimated DNA sequence 135 ... estimated DNA sequence reliability 210 ... input / output device 220 ... analyzer 221 ... obfuscated DNA sequence analysis unit 222 ... learning unit 223 ... estimation unit 230 ... storage device 231 ... fluorescent color judgment reference database 232 ... fluorescent color likelihood database 501 ... fluorescent color judgment reference learning unit 502 ... fluorescent color judgment unit 503 ... fluorescent color likelihood Calculation unit 901 ... Sequence feature determination unit 902 ... Fluorescent color determination unit 903 ... DNA sequence likelihood calculation unit

Claims (5)

In the DNA sequencing system in the massively parallel sequencer,
About the obfuscated DNA sequence data, the fluorescence intensity sequence data acquired from the massively parallel sequencer is classified into one or a plurality of feature groups according to the features on the sequence, and the known prepared for each classified feature group Using the fluorescence intensity sequence data of the DNA sequence data of, and learning the fluorescence color criteria for judging the fluorescence color, the function to calculate the fluorescence color likelihood that gives its reliability, and the unknown experimental DNA sequence data A corresponding feature group is discriminated from the one or more feature groups based on the features on the sequence, and the estimated DNA sequence data of the unknown experimental DNA sequence data and its reliability are given based on the discrimination result An analysis device having a function of calculating a fluorescent color likelihood;
A storage device for storing the fluorescent color criteria and the fluorescent color likelihood learned for each feature group on the array;
The inputs the fluorescence intensity sequence data, have a input-output device for outputting a fluorescent color likelihood that gives the reliability and the estimated DNA sequence data,
The function of classifying into the feature groups is
By applying a clustering analysis using the fluorescence intensity sequence data acquired from the massively parallel sequencer as a feature vector to the DNA sequence data determined to be the obfuscated DNA sequence data, or nonlinearly transforming the feature vector By applying clustering analysis on the feature space, the DNA sequence data is classified into the one or more feature groups . In the DNA sequence decoding system according to claim 1,
A DNA sequencing system that extracts difficult-to-read DNA sequence data that is difficult to read based on the mapping accuracy when mapping the actual sequence of various genomes to the reference sequence. In the DNA sequencing method in a massively parallel sequencer,
Processing to classify fluorescence intensity sequence data acquired from the massively parallel sequencer for obfuscated DNA sequence data into one or more feature groups according to the features on the sequence;
Using fluorescence intensity sequence data of known DNA sequence data prepared for each classified feature group, learn fluorescence color criteria for determining fluorescence color and calculate fluorescence color likelihood that gives its reliability Processing to
A corresponding feature group is determined from the one or more feature groups based on the sequence characteristics of the unknown experimental DNA sequence data, and the estimated DNA sequence of the unknown experimental DNA sequence data is determined based on the determination result possess a process of calculating a fluorescence color likelihood providing data and its reliability,
The process of classifying into the feature groups is
By applying a clustering analysis using the fluorescence intensity sequence data acquired from the massively parallel sequencer as a feature vector to the DNA sequence data determined to be the obfuscated DNA sequence data, or nonlinearly transforming the feature vector A DNA sequence decoding method , wherein the DNA sequence data is classified into the one or more feature groups by applying clustering analysis on the feature space . In the DNA sequence decoding method according to claim 3 ,
A DNA sequencing method characterized by extracting difficult-to-read DNA sequence data based on mapping accuracy when mapping the actual sequence of various genomes to a reference sequence. On the computer,
Processing to classify the fluorescence intensity sequence data obtained from the massively parallel sequencer for the obfuscated DNA sequence data into one or more feature groups according to the features on the sequence;
Using fluorescence intensity sequence data of known DNA sequence data prepared for each classified feature group, learn fluorescence color criteria for determining fluorescence color and calculate fluorescence color likelihood that gives its reliability Processing to
A corresponding feature group is determined from the one or more feature groups based on the sequence characteristics of the unknown experimental DNA sequence data, and the estimated DNA sequence of the unknown experimental DNA sequence data is determined based on the determination result A program that executes data and a process for calculating a fluorescent color likelihood that gives the reliability of the data ,
The process of classifying into the feature groups is
By applying a clustering analysis using the fluorescence intensity sequence data acquired from the massively parallel sequencer as a feature vector to the DNA sequence data determined to be the obfuscated DNA sequence data, or nonlinearly transforming the feature vector Classifying the DNA sequence data into the one or more feature groups by applying a clustering analysis on the feature space
A program characterized by that .

Images