Patent Application
20200051667
This disclosure provides a novel method of representation of genome sequencing data which reduces the utilized storage space and improves access performance by providing new functionality that are not available with known prior art methods of representation.
An appropriate representation of genome sequencing data is fundamental to enable efficient genomic analysis applications such as genome variants calling and all other analysis performed with various purposes by processing the sequencing data and metadata. Human genome sequencing has become affordable by the emergence of high-throughput low cost sequencing technologies. Such opportunity opens new perspectives in several fields ranging from the diagnosis and treatment of cancer to the identification of genetic illnesses, from pathogen surveillance for the identification of antibodies to the creation of new vaccines, drugs and the customization of personalized treatments.
Hospitals, genomic analysis providers, bioinformatics and large biological data storage centers are looking for affordable, fast, reliable and interconnected genomic information processing solutions which could enable scaling genomic medicine to a world-wide scale. Since one of the bottleneck in the sequencing process has become data storage, methods for representing genome sequencing data in a compressed form are increasingly investigated.
The most used genome information representations of sequencing data are based on zipping FASTQ and SAM formats. The objective is to compress the traditionally used file formats (respectively FASTQ and SAM for non-aligned and aligned data). Such files are constituted by plain text characters and are compressed, as mentioned above, by using general purpose approaches such as LZ (from Lempel and Ziv, the authors who published the first versions) schemes (the well-known zip, gzip etc). When general purpose compressors such as gzip are used, the result of compression is usually a single blob of binary data. The information in such monolithic form results quite difficult to archive, transfer and elaborate particularly when like in the case of high throughput sequencing the volume of data are extremely large. The BAM format is characterized by poor compression performance due to the focus on compression of the inefficient and redundant SAM format rather than on extracting the actual genomic information conveyed by SAM files and due to the adoption of general purpose text compression algorithms such as gzip rather than exploiting the specific nature of each data source (the genomic data itself).
A more sophisticated approach to genomic data compression that is less used, but more efficient than BAM is CRAM. CRAM provides more efficient compression for the adoption of differential encoding with respect to a reference (it partially exploits the data source redundancy), but it still lacks features such as incremental updates, support for streaming and selective access to specific classes of compressed data.
These approaches generate poor compression ratios and data structures that are difficult to navigate and manipulate once compressed. Downstream analysis can be very slow due to the necessity of handling large and rigid data structures even to perform simple operation or to access selected regions of the genomic dataset. CRAM relies on the concept of the CRAM record. Each CRAM record represents a single mapped or unmapped reads by coding all the elements necessary to reconstruct it.
CRAM presents the following drawbacks and limitations that are solved and removed by the invention described in this disclosure:
1. CRAM does not support data indexing and random access to data subsets sharing specific features. Data indexing is out of the scope of the specification (see section 12 of CRAM specification v 3.0) and it is implemented as a separate file. Conversely the approach of the invention described in this disclosure employs a data indexing method that is integrated with the encoding process and indexes are embedded in the encoded (i.e. compressed) bit stream.
2. CRAM is built by core data blocks that can contain any type of mapped reads (perfectly matching reads, reads with substitutions only, reads with insertions or deletions (also referred to as “indels”)). There is no notion of data classification and grouping of reads in classes according to the result of mapping with respect to a reference sequence. This means that all data need to be inspected even if only reads with specific features are searched. Such limitation is solved by the invention by classifying and partitioning data in classes before coding.
3. CRAM is based on the concept of encapsulating each read into a “CRAM record”. This implies the need to inspect each complete “record” when reads characterized by specific biological features (e.g. reads with substitutions, but without “indels”, or perfectly mapped reads) are searched.
Conversely, in the present invention there is the notion of data classes coded separately in separated information blocks and there is no notion of record encapsulating each read. This enables more efficient access to set of reads with specific biological characteristics (e.g. reads with substitutions, but without “indels”, or perfectly mapped reads) without the need of decoding each (block of) read(s) to inspect its features.
4. In a CRAM record each record field is associated to a specific flag and each flag must always have the same meaning as there is no notion of context since each CRAM record can contain any different type of data. This coding mechanism introduces redundant information and prevents the usage of efficient context based entropy coding. Instead in the present invention, there is no notion of flag denoting data because this is intrinsically defined by the information “block” the data belongs to. This implies a largely reduced number of symbols to be used and a consequent reduction of the information source entropy which results into a more efficient compression. Such improvement is possible because the use of different “blocks” enables the encoder to reuse the same symbol across each block with different meanings according to the context. In CRAM each flag must always have the same meaning as there is no notion of contexts and each CRAM record can contain any type of data.
5. In CRAM substitutions, insertions and deletions are represented by using different syntax elements, option that increases the size of the information source alphabet and yields a higher source entropy. Conversely, the approach of the disclosed invention uses a single alphabet and encoding for substitutions, insertions and deletions. This makes the encoding and decoding process simpler and produces a lower entropy source model which coding yields bitstreams characterized by high compression performance.
The present invention aims at compressing genomic sequences by classifying and partitioning sequencing data so that the redundant information to be coded is minimized and features such as selective access and support for incremental updates are directly enabled in the compressed domain.
One of the aspects of the presented approach is the definition of classes of data and metadata structured in different blocks and encoded separately. The more relevant improvements of such approach with respect to existing methods consist in:
1. the increase of compression performance due to the reduction of the information source entropy constituted by providing an efficient source model for each class of data or metadata;
2. the possibility of performing selective accesses to portions of the compressed data and metadata for any further processing purpose directly in the compressed domain;
3. the possibility to incrementally (i.e. without the need of decoding and re-encoding) update compressed data and metadata with new sequencing data and/or metadata and/or new analysis results associated to specific sets of sequencing reads.
The features of the claims below solve the problem of existing prior art solutions by providing a method for encoding genome sequence data, said genome sequence data comprising reads of sequences of nucleotides, said method comprising the steps of:
aligning said reads to one or more reference sequences thereby creating aligned reads,
classifying said aligned reads according to specified matching rules with said one or more reference sequences, thereby creating classes of aligned reads,
encoding said classified aligned reads as a multiplicity of blocks of syntax elements, structuring said blocks of syntax elements with header information thereby creating successive Access Units,
wherein said encoding further comprises binarizing and entropy coding genomic descriptors.
In another aspect of the coding method said binarizing and entropy coding genomic descriptors is such that the binarization and the entropy coding is different for the different descriptors.
In another aspect of the coding method said descriptors comprise:
pos for signaling the mapping position of a read on a reference sequence,
rcomp for signaling the DNA or RNA strand the reads was mapped on mapping flags for enabling the aligner to further specify the result of the mapping process.
In another aspect the coding method further comprises coding the following descriptors:
mmpos for signaling the position of mismatches in aligned reads with respect to reference sequences,
mmtype for signaling the types of mismatches with respect to reference sequences at the associated positions.
In another aspect the coding method further comprises coding the clips descriptor for signaling soft or hard clipped nucleotides.
In another aspect the coding method further comprises coding the rlen descriptor for signaling the length of each encoded sequence read.
In another aspect the coding method further comprises coding the following descriptors:
mmap for signaling the multiple mapping positions that are associated to a single read or read pair by the mapping procedure,
msar for signaling the identification of the existence of spliced reads (i.e. reads that when split in chunks find mapping positions with higher degrees of matching accuracy than when they are mapped as single contiguous read mapped on a single position on a reference sequence).
In another aspect the coding method further comprises coding the mscore descriptor to signal a mapping/alignment score per read as generated by genomic sequence reads aligners.
In another aspect the coding method further comprises coding the pair descriptor to signal, in case of paired end reads, how the reads are paired.
In another aspect the coding method further comprises coding the ureads descriptor to signal reads which could not be aligned at any position of the reference sequence.
In another aspect the coding method further comprises coding the rtype descriptor used to signal the subset of descriptors used to encode sequence reads that cannot be mapped at any position of the reference sequence with specified degrees of matching accuracy.
In another aspect the coding method further comprises coding the rgroup descriptor to signal to which read group the read belongs to.
In another aspect the coding method further comprises coding the following descriptors:
rftp for signaling the position of mismatches between a contig and a reference sequence. Positions of mismatches are terminated by a special terminator character,
rftt for signaling the type of mismatches between a contig and a reference sequence.
In another aspect of the coding method
said pos descriptor is binarized using a double truncated unary code or a single double truncated unary code,
said rcomp descriptor is binarized using a truncated unary code,
said mapping flags descriptors are binarized using binary coding.
In another aspect of the coding method
said mmpos descriptor for signaling the position of mismatches in aligned reads with respect to reference sequences is binarized using a split unit wise truncated unary code,
said mmtype descriptor for signaling the types of mismatches with respect to reference sequences at the associated positions is binarized using a truncated unary code.
In another aspect of the coding method said clips descriptor for signaling soft or hard clipped nucleotides is binarized using a concatenation of Signed Truncated Exponential Golomb, Truncated Unary, Signed Exponential Golomb and Binary Codes.
In another aspect of the coding method said rlen descriptor signaling the length of each encoded sequence read is binarized using a Split Unit-wise Truncated Unary code.
In another aspect of the coding method:
said mmap descriptor for signaling the multiple mapping positions that are associated to a single read or read pair by the mapping procedure is binarized using a Split Unit-wise Truncated Unary code,
said msar descriptor for signaling the identification of the existence of spliced reads is binarized using a Signed Exponential Golomb code.
In another aspect of the coding method said mscore descriptor to signal a mapping/alignment score per read as generated by genomic sequence reads aligners is binarized using a Truncated Unary code,
In another aspect of the coding method said pair descriptor to signal, in case of paired end reads, how the reads are paired is binarized using a concatenation of Binary Coding and Split Unit-wise Truncated Unary code.
In another aspect of the coding method said ureads descriptor to signal reads which could not be aligned at any position of the reference sequence is binarized using a Truncated Unary code.
In another aspect of the coding method said rtype descriptor used to signal the subset of descriptors used to encode sequence reads that cannot be mapped at any position of the reference sequence with specified degrees of matching accuracy is binarized using a Truncated Unary code.
In another aspect of the coding method said rgroup descriptor to signal to which read group the read belongs to is binarized using a Truncated Unary code.
In another aspect of the coding method:
said rftp descriptor for signaling the position of mismatches between a contig and a reference sequence is binarized using a concatenation of Binary Coding and Split Unit-wise Truncated Unary code,
said rftt descriptor for signaling the type of mismatches between a contig and a reference sequence is binarized using a concatenation of Binary Coding and Truncated Unary code.
In another aspect of the coding method the configuration parameters for the coding of said descriptors are contained in a syntax header.
In another aspect of the coding method said configuration parameters are updated by creating updated syntax headers to be added to the encoded genomic file.
In another aspect of the coding method said configuration parameters comprise a dataset type for signaling the type of data encoded in Access Units referring to this encoding parameters.
In another aspect of the coding method said configuration parameters further comprise a reads length for signaling the length in nucleotides of sequence reads in case of constant reads length.
In another aspect of the coding method said configuration parameters further comprise a quality values depth parameter for signaling the number of Quality Values associated to each coded nucleotide.
In another aspect of the coding method said configuration parameters further comprise an alignment score depth for signaling the number of alignments scores associated to each coded alignments.
In another aspect of the coding method said configuration parameters further comprise a terminator size for signaling the size in bytes of the terminator symbol used for the mmpos descriptor.
In another aspect of the coding method said configuration parameters further comprise a terminator value for signaling the value of the terminator symbol used for the mmpos descriptor.
In another aspect of the coding method said configuration parameters further comprise the number of classes for signaling the number of data classes encoded in all Access Units referring to said configuration parameters.
In another aspect of the coding method said configuration parameters further comprise class identifiers to signal the identifiers associated to the data class defined in this disclosure (P, N, M, I, HM, U).
In another aspect of the coding method said configuration parameters further comprise the number of descriptors for signaling the total number of descriptors contained in Access Units referring to said configuration parameters.
In another aspect of the coding method said configuration parameters further comprise coding mode identifiers for signaling the coding modes defined in this disclosure.
In another aspect of the coding method said configuration parameters further comprise a number of groups parameter for signaling the number of different values of the rgroup descriptor present in all Access Units referring to the current encoding parameters.
In another aspect of the coding method said configuration parameters further comprise one or more group name parameters for signaling one or more read group identifiers.
In another aspect of the coding method said configuration parameters further comprise a multiple alignments flag for signaling the presence of multiple alignments in the Access Unit.
In another aspect of the coding method said configuration parameters further comprise a spliced reads flag for signaling the presence of spliced reads in the Access Unit. When set to 0 no spliced reads are present.
In another aspect of the coding method said configuration parameters further comprise a multiple signature base flag for signaling the use of multiple signatures in an Access Unit containing unmapped sequence reads (Class U).
In another aspect of the coding method said configuration parameters further comprise a signature size for signaling the size in bits of each integer representing an encoded signature.
In another aspect of the coding method said configuration parameters further comprise a score exponent parameters for signaling the number of bits used to encode the exponent part of the multiple alignments score encoded in the mscore descriptor.
In another aspect of the coding method said configuration parameters further comprise a score fractional parameter for signaling the number of bits used to encode the fractional part of the multiple alignments score encoded in the mscore descriptor.
The present invention further provides a method for decoding encoded genomic data said genome sequence data comprising reads of sequences of nucleotides, said method comprising the steps of:
parsing Access Units containing said encoded genomic data to extract multiple blocks of genomic descriptors by employing header information,
decoding said multiplicity of blocks,
wherein said decoding of multiplicity of blocks comprise decoding and de-binarizing genomic descriptors to extract aligned reads according to specific matching rules defining their classification with respect to one or more reference sequences.
In another aspect of the decoding method said descriptors comprise:
a pos descriptor for signaling the mapping position of a read on a reference sequence,
a rcomp descriptor for signaling the DNA or RNA strand the reads was mapped on mapping flags for enabling the aligner to further specify the result of the mapping process.
In another aspect the decoding method further comprises decoding the following descriptors:
mmpos for signaling the position of mismatches in aligned reads with respect to reference sequences,
mmtype for signaling the types of mismatches with respect to reference sequences at the associated positions.
In another aspect the decoding method further comprises decoding the clips descriptor for signaling soft or hard clipped nucleotides.
In another aspect the decoding method further comprises decoding the rlen descriptor for signaling the length of each encoded sequence read.
In another aspect the decoding method further comprises decoding the following descriptors:
mmap for signaling the multiple mapping positions that are associated to a single read or read pair by the mapping procedure,
msar for signaling the identification of the existence of spliced reads (i.e. reads that when split in chunks find mapping positions with higher degrees of matching accuracy than when they are mapped as single contiguous read mapped on a single position on a reference sequence).
In another aspect the decoding method further comprises decoding the mscore descriptor to signal a mapping/alignment score per read as generated by genomic sequence reads aligners.
In another aspect the decoding method further comprises decoding the pair descriptor to signal, in case of paired end reads, how the reads are paired.
In another aspect the decoding method further comprises decoding the ureads descriptor to signal reads which could not be aligned at any position of the reference sequence.
In another aspect the decoding method further comprises decoding the rtype descriptor used to signal the subset of descriptors used to encode sequence reads that cannot be mapped at any position of the reference sequence with specified degrees of matching accuracy.
In another aspect the decoding method further comprises decoding the rgroup descriptor to signal to which read group the read belongs to.
In another aspect the decoding method further comprises decoding the following descriptors:
rftp for signaling the position of mismatches between a contig and a reference sequence. Positions of mismatches are terminated by a special terminator character.
rftt for signaling the type of mismatches between a contig and a reference sequence.
In another aspect of the decoding method the configuration parameters for the decoding of said descriptors are extracted from a syntax header.
In another aspect of the decoding method said configuration parameters comprise a dataset type for signaling the type of data encoded in Access Units referring to this encoding parameters.
In another aspect of the decoding method said configuration parameters further comprise a reads length for signaling the length in nucleotides of sequence reads in case of constant reads length.
In another aspect of the decoding method said configuration parameters further comprise a quality values depth parameter for signaling the number of Quality Values associated to each coded nucleotide.
In another aspect of the decoding method said configuration parameters further comprise an alignment score depth for signaling the number of alignments scores associated to each coded alignments.
In another aspect of the decoding method said configuration parameters further comprise a terminator size for signaling the size in bytes of the terminator symbol used for the mmpos descriptor.
In another aspect of the decoding method said configuration parameters further comprise a terminator value for signaling the value of the terminator symbol used for the mmpos descriptor.
In another aspect of the decoding method said configuration parameters further comprise the number of classes for signaling the number of data classes encoded in all Access Units referring to said configuration parameters.
In another aspect of the decoding method said configuration parameters further comprise class identifiers to signal the identifiers associated to the data class defined in this disclosure (P, N, M, I, HM, U).
In another aspect of the decoding method said configuration parameters further comprise the number of descriptors for signaling the total number of descriptors contained in Access Units referring to said configuration parameters.
In another aspect of the decoding method said configuration parameters further comprise coding mode identifiers for signaling the coding modes defined in this disclosure.
In another aspect of the decoding method said configuration parameters further comprise a number of groups parameter for signaling the number of different values of the rgroup descriptor present in all Access Units referring to the current encoding parameters.
In another aspect of the decoding method said configuration parameters further comprise one or more group name parameters for signaling one or more read group identifiers.
In another aspect of the decoding method said configuration parameters further comprise a multiple alignments flag for signaling the presence of multiple alignments in the Access Unit.
In another aspect of the decoding method said configuration parameters further comprise a spliced reads flag for signaling the presence of spliced reads in the Access Unit. When set to 0 no spliced reads are present.
In another aspect of the decoding method said configuration parameters further comprise a multiple signature base flag for signaling the use of multiple signatures in an Access Unit containing unmapped sequence reads (Class U).
In another aspect of the decoding method said configuration parameters further comprise a signature size for signaling the size in bits of each integer representing an encoded signature.
In another aspect of the decoding method said configuration parameters further comprise a score exponent parameters for signaling the number of bits used to encode the exponent part of the multiple alignments score encoded in the mscore descriptor.
In another aspect of the decoding method said configuration parameters further comprise a score fractional parameter for signaling the number of bits used to encode the fractional part of the multiple alignments score encoded in the mscore descriptor.
The present invention further provides an encoding apparatus comprising encoding means for implementing all the aspects of the previously mentioned encoding methods.
The present invention further provides a decoding apparatus for implementing all the aspects of the previously mentioned decoding methods.
The present invention further provides a file format comprising the previously defined genomic descriptors
The present invention further provides a computer-readable medium comprising instructions that when executed cause at least one processor to perform all the aspects of the previously mentioned coding methods.
The present invention further provides a computer-readable medium comprising instructions that when executed cause at least one processor to perform all the aspects of the previously mentioned decoding methods.
The present invention further provides a support data storing genomic encoded according perform all the aspects of the previously mentioned coding methods.
The genomic or proteomic sequences referred to in this invention include, for example, and not as a limitation, nucleotide sequences, Deoxyribonucleic acid (DNA) sequences, Ribonucleic acid (RNA), and amino acid sequences. Although the description herein is in considerable detail with respect to genomic information in the form of a nucleotide sequence, it will be understood that the methods and systems for compression can be implemented for other genomic or proteomic sequences as well, albeit with a few variations, as will be understood by a person skilled in the art.
Genome sequencing information is generated by High Throughput Sequencing (HTS) machines in the form of sequences of nucleotides (a. k. a. “bases”) represented by strings of letters from a defined vocabulary. The smallest vocabulary is represented by five symbols: {A, C, G, T, N} representing the 4 types of nucleotides present in DNA namely Adenine, Cytosine, Guanine, and Thymine. In RNA Thymine is replaced by Uracil (U). N indicates that the sequencing machine was not able to call any base and so the real nature of the position is undetermined. In case the IUPAC ambiguity codes are adopted by the sequencing machine, the alphabet used for the symbols is (A, C, G, T, U, W, S, M, K, R, Y, B, D, H, V, N or -).
The nucleotides sequences produced by sequencing machines are called “reads”. Sequence reads can be between a few dozens to several thousand nucleotides long. Some technologies produce sequence reads in “pairs” where one read is from one DNA strand and the second is from the other strand. A read associated to another read in a sequencing process producing pairs is said to be its mate.
The process of arranging sequence reads to identify regions of similarity with segments of a reference genome according to a set of matching rules is called “alignment” or “mapping”.
Throughout this disclosure, a reference sequence is a sequence of nucleotides associated to a mono-dimensional integer coordinate system for which each integer coordinate is associated to a single nucleotide. Coordinate values can only be equal or larger than zero. This coordinate system in the context of this invention is zero-based (i.e. the first nucleotide has coordinate 0 and it is said to be at position 0) and linearly increasing from left to right. A reference sequence is any sequence on which the nucleotides sequences produced by sequencing machines are aligned/mapped. One example of sequence could actually be a “reference genome”, a sequence assembled by scientists as a representative example of a species' set of genes. For example GRCh37, the Genome Reference Consortium human genome (build 37) is derived from thirteen anonymous volunteers from Buffalo, N.Y. However, a reference sequence could also consist of a synthetic sequence conceived and constructed to merely improve the compressibility of the reads in view of their further processing.
A common element of efficient approaches to genomic sequence reads compression is the exploitation of the correlation of sequence data with respect to a reference sequence. Even if the somatic profile of the human population is extremely diversified, the actual portion of the number of nucleotides that differs from person to person is about only 0.1% of the total number of nucleotides composing an entire genome. Therefore, the specific genomic information characterizing each individual is very limited with respect to the entire information carried by an entire genome. When a pre-existing reference genome is available, be it for previous sequencing or as a published “average” consensus reference, the most common way as of today to encode the information is to identify and encode only the differences with respect to the reference genome.
In order to do so with raw sequence reads, generally expressed in the form of FASTQ data files, a preliminary pre-processing step the mapping on a reference genome. In case an appropriate reference genome is not available, or if the bias introduced by the usage of a specific reference is not desirable, the construction of a new reference sequence by means of assembling the sequence reads at hand into longer sequences called contigs, is a possible alternative. When mapping sequence reads on a reference sequence, said reference sequence is used as axis of a mono-dimensional coordinate system in which the left-most position is denoted as position 0. For each sequence read, mapped to a reference sequence, the nucleotide mapped at the reference sequence position identified by the smallest coordinate number is usually referred to as the “left-most” nucleotide, whereas the nucleotide mapped at the reference sequence position identified by the largest coordinate number is referred to as the “right-most” nucleotide. This is illustrated in
When a sequence read is mapped to a reference sequence, the coordinate of the left-most mapped base is said to represent the mapping position of the read on the reference sequence.
A base present in the aligned read and not present in the reference sequence (a.k.a. insertion) and bases preserved by the alignment process but not mapped on the reference sequence (a.k.a. soft clips) do not have mapping positions.
When a sequence read cannot be mapped to any mapped position of the used reference sequences according to the specified matching rules, it is said to be unmapped.
The process of building longer genomic sequences by looking for overlapping regions among sequence reads is called assembly.
A longer genomic sequence built assembling shorter reads is called contig (see https://en.wikipedia.org/wiki/Contig).
Sequence reads that fail to build any contig during an assembly process are said to be unaligned.
A reference genome is composed by one or more reference sequences and it is assembled by scientists as a representative example of a species' set of genes. For example GRCh37, the Genome Reference Consortium human genome (build 37) is derived from thirteen anonymous volunteers from Buffalo, N.Y. However, a reference sequence could also consist of a synthetic sequence conceived and merely constructed to improve the compressibility of the reads in view of their further processing.
In this disclosure the read composing a read pair with a base mapping on the smallest coordinate on a reference sequence is referred to as “Read 1” whereas its mate is referred to as “Read 2”.
The distance, expressed as number of nucleotides (or bases), separating two reads generated as a pair, by a sequencing machine using current state of the art sequencing technology, is unknown, and it is determined by mapping both reads composing the pair (i.e. minimizing appropriate matching functions) to a reference sequence.
Throughout this disclosure, a genomic record is a data structure encoding either a single sequence read or a paired sequence read optionally associated with alignment information, read identifier and quality values.
Throughout this disclosure, an Access Unit (AU) is defined as a logical data structure containing a coded representation of genomic information or related metadata to facilitate the bit stream access and manipulation. It is the smallest data organization that can be decoded by a decoding device implementing the invention described in this disclosure.
According to the type of coded information, an AU can be decoded either independently of any other AU or using information contained in other AUs.
AUs can be classified into a multiplicity of types according to the nature of the coded sequence data. An Access Unit contains either a reference sequence, or a portion thereof, or encoded reads or read pairs belonging to a single class of data. Any single AU cannot contain two or more types of sequence data. For example an Access Unit may contain the entire chromosome 1 of GRCh37, the Genome Reference Consortium human genome (build 37). Another Access Unit may contain the coded representation of nucleotides of chromosome 1 of GRCh37 that are located between coordinates 50′000 and 150′000. Another Access Unit may contain only reads or read pairs that perfectly map on the reference sequence without any mismatch. Another Access Unit may contain reads or read pairs that only contain “N” symbols as mismatches with respect to the reference sequence. Another Access Unit may contain reads or read pairs that contain any type of substitutions (e.g. one base present in the read or read pair is different from the base at the corresponding mapping position in the reference sequence). Another Access Unit may contain reads or read pairs that contain mismatches, insertions, deletions and soft clipped bases. Another Access Unit may contain only read or read pairs that do not map on the reference sequence. Another Access Unit may contain only read pairs in which one read is mapped and the other is not mapped on the reference sequence. Another type of Access Unit may contain only encoded segments of a reference genome composed by one or more reference sequences (for example chromosomes).
The essential characteristic of an Access Unit is that it contains in compressed form all elements needed to reconstruct the genomic information (sequence reads or read pairs, reference sequences), associated alignment information and metadata of reads or read pairs it represents. In other words, to fully reconstruct the reads, read pairs or reference sequence and associated information carried by an Access Unit it is only necessary to retrieve the Access Unit itself and, when applicable, the Access Units containing the reference sequence it refers to.
In each Access Unit the descriptors listed in the next section representing the encoded read or read pairs are aggregated in separate data blocks—one per type—in order to exploit their homogeneous statistical properties for achieving high performance entropy coding.
Each Access Unit contains the compressed sub-set of descriptors representing sequence reads or read pairs belonging to the same class mapped to a genomic region on a reference sequence. Such genomic region on the reference sequence is defined by a start coordinate (or start position) and an end coordinate (or end position).
An example of Access Unit is illustrated in
In the context of this disclosure the following definitions apply:
access unit start position: left-most genomic record position among all genomic records contained in the access unit.
access unit end position: right-most base position among all mapped bases of all genomic records contained in the access unit.
access unit range: the genomic range comprised between the access unit start position and the right-most genomic record position among all genomic records contained in the access unit.
access unit size: number of genomic records contained in an access unit.
access unit covered region: genomic range comprised between the Access Unit start position and the Access Unit end position.
In the context of this disclosure, one or more access unit are organized in a structure called genomic dataset. A genomic dataset is a compression unit containing headers and access units. The set of access units composing the genomic dataset constitutes the genomic dataset payload.
A collection of one or more genomic datasets is called dataset group.
In the context of this disclosure, genomic descriptors are syntax elements representing part of the information (and also elements of a syntax structure of a file format and/or a bitstream) necessary to reconstruct (i.e. decode) coded reference sequences, sequence reads and the associated mapping information. The genomic descriptors disclosed in this invention are listed in Table 3.
According to the method disclosed in this invention, reference sequences or portion thereof, sequence reads and the associated alignment information are coded using a sub-set of the descriptors listed above which are then entropy coded using a multiplicity of entropy coders according to each descriptor specific statistical properties. Blocks of compressed descriptors with homogeneous statistical properties are structured in Access Units which represent the smallest coded representation of one or more genomic sequence that can be manipulated by a device implementing the invention described in this disclosure.
Genomic descriptors are organized in blocks and streams as defined below.
A block is defined as a data unit composed by a header and a payload, which is composed by portions of compressed descriptors of the same type.
A descriptor stream is defined as a sequence of encoded descriptor blocks used to decode a descriptor of a specific Data Class.
Sequencing devices can introduce errors in the sequence reads such as:
In genome sequencing the term “coverage” is used to express the level of redundancy of the sequence data with respect to a “reference sequence”. The average coverage of aligned genome sequencing data is the average number of time each base at each position of the reference genome is present in the aligned data. For example, to reach a coverage of 30× on a human genome (3.2 billion bases long) a sequencing machine shall produce a total of 30×3.2 billion bases so that in average each position in the reference is “covered” 30 times.
The coverage is said to be:
This invention aims at defining a genomic information representation format in which the relevant information is efficiently accessible and transportable and the weight of the redundant information is reduced.
The main innovative aspects of the disclosed invention are the following.
1 Sequence reads are classified and partitioned into data classes according to the results of the alignment with respect to reference sequences. Such classification and partitioning enables the selective access to encoded data according to criteria related to the alignment results and to the matching accuracy.
2 The classified sequence reads and the associated metadata are represented by genomic descriptors organized in blocks with homogeneous statistical properties enabling the definition of distinct information sources characterized by a low information entropy.
3 The possibility of modeling each separated information source with distinct source model adapted to the statistical characteristics of each class and the possibility of changing the source model within each class of reads and within each descriptor block for each separately accessible data units (Access Units). The adoption of the appropriate transformation, binarization and context adaptive probability models and associated entropy coders according to the statistical properties of each source model.
4 The definition of correspondences and dependencies among the descriptors blocks to enable the selective access to the sequencing data and associated metadata without the need to decode all the descriptors blocks if not all information is required.
5 The coding of each sequence data class and associated metadata blocks with respect to “pre-existing” (also denoted as “external”) reference sequences or with respect to “transformed” reference sequences obtained by applying appropriate transformations to “pre-existing” reference sequences so as to reduce the entropy of the descriptors blocks information sources. Said descriptors represent the reads partitioned into the different data classes. Following any encoding of reads using the corresponding descriptors with reference to a “pre-existing” reference or “transformed” “pre-existing” reference sequence, the occurrence of the various mismatches can be used to define the appropriate transformations to the reference sequences in order to find a final coded representation with low entropy and achieve higher compression efficiency.
6 The construction of one or more reference sequences (also referred to as “internal” references to distinguish from the “pre-existing” also referred here as “external” reference sequences) used to encode the class of reads that present a degree of matching accuracy with respect to the pre-existing reference sequences not satisfying a set of constraints. Such constraints are set with the objective that the coding costs of representing in compressed form the class of reads aligned with respect to the “internal” reference sequences and the cost of representing the “internal” reference sequences themselves, is lower than encoding the unaligned class of reads verbatim, or using the “external” reference sequences without or with transformations.
7 The transmission of the configuration parameters governing the process of both encoding and decoding by means of data structures embedded in the compressed genomic data in the form of header information. Such configuration parameters can be updated during the encoding process in order to improve the compression performance. Such updates are conveyed in the compressed content in the form of updated configuration data structures. In the following, each of the above aspects will be further described in details.
The sequence reads generated by sequencing machines are classified by the disclosed invention into six different “classes” according to the matching results of the alignment with respect to one or more “pre-existing” reference sequences.
When aligning a DNA sequence of nucleotides with respect to a reference sequence the following cases can be identified:
In this disclosure this type of mismatch are referred to as “n type” mismatch. Such sequences belong to “Class N” reads. Once the read is classified to belong to “Class N” it is useful to limit the degree of matching inaccuracy to a given upper bound and set a boundary between what is considered a valid matching and what it is not. Therefore, the reads assigned to Class N are also constrained by setting a threshold (MAXN) that defines the maximum number of undefined bases (i.e. bases called as “N”) that a read can contain. Such classification implicitly defines the required minimum matching accuracy (or maximum degree of mismatch) that all reads belonging to Class N share when referred to the corresponding reference sequence, which constitutes an useful criterion for applying selective data searches to the compressed data.
The classification specified in the previous section concerns single sequence reads. In the case of sequencing technologies that generates read in pairs (i.e. Illumina Inc.) in which two reads are known to be separated by an unknown sequence of variable length, it is appropriate to consider the classification of the entire pair to a single data class. A read that is coupled with another is said to be its “mate”.
If both paired reads belong to the same class the assignment to a class of the entire pair is the following: the entire pair is assigned to the same class for any class (i.e. P, N, M, I, U). In the case the two reads belong to a different class, but none of them belongs to the “Class U”, then the entire pair is assigned to the class with the highest priority defined according to the following expression:
P<N<M<I
in which “Class P” has the lowest priority and “Class I” has the highest priority.
In case only one of the reads belongs to “Class U” and its mate to any of the Classes P, N, M, I a sixth class is defined as “Class HM” which stands for “Half Mapped”.
The definition of such specific class of reads and the associated assignment rules are motivated by the fact that such data is used for attempting to determine gaps or unknown regions of reference genomes (a.k.a. little known or unknown regions). Such regions are reconstructed by mapping pairs at the edges using the pair read that can be mapped on the known regions. The unmapped mate is then used to build the so called “contigs” of the unknown region as it is shown in
The table below summarizes the matching rules applied to reads in order to define the class of data each read belongs to. The rules are defined in the first five columns of the table in terms of presence or absence of type of mismatches (n, s, d, i and c type mismatches). The sixth column provide rules in terms of maximum threshold for each mismatch type and any function f(n,s) and w(n,s,d,i,c) of the possible mismatch types.
The data classes of type N, M and I as defined in the previous sections can be further decomposed into an arbitrary number of distinct sub-classes with different degrees of matching accuracy. Such option is an important technical advantage in providing a finer granularity and as consequence a much more efficient selective access to each data class. As an example and not as a limitation, to partition the Class N into a number k of subclasses (Sub-Class N1, . . . , Sub-Class Nk) it is necessary to define a vector with the corresponding components MAXN1, MAXN2, . . . , MAXN(k-1), MAXN(k), with the condition that MAXN1<MAXN2< . . . <MAXN(k-1_<MAXN and assign each read to the lowest ranked sub-class that satisfy the constrains specified in Table 1 when evaluated for each element of the vector. This is shown in
In the case of the classes of type M and I the same principle is applied by defining a vector with the same properties for MAXM and MAXTOT respectively and use each vector components as threshold for checking if the functions f(n,s) and w(n,s,d,l,c) satisfy the constraint. Like in the case of sub-classes of type N, the assignment is given to the lowest sub-class for which the constraint is satisfied. The number of sub-classes for each class type is independent and any combination of subdivisions is admissible. This is shown in
When two reads in a pair are classified in the same sub-class, then the pair belongs to the same sub-class.
When two reads in a pair are classified into sub-classes of different classes, then the pair belongs to the sub-class of the class of higher priority according to the following expression:
N<M<I
where N has the lowest priority and I has the highest priority.
When two reads belong to different sub-classes of one of classes N or M or I, then the pair belongs to the sub-class with the highest priority according to the following expressions:
N1<N2< . . . <Nk
M1<M2< . . . <Mj
I1<I2< . . . <Ih
where the highest index has the highest priority.
The mismatches found for the reads classified in the classes N, M and I can be used to create “transformed” references to be used to compress more efficiently the read representation. Reads classified as belonging to the Classes N, M or I (with respect to the “pre-existing” (i.e. “external”) reference sequence denoted as RS0) can be coded with respect to the “transformed” reference sequence RS, according to the occurrence of the actual mismatches with the “transformed” reference. For example if readMin belonging to Class M (denoted as the ith read of class M) containing mismatches with respect to the reference sequence RSn, then after “transformation” readMin=readPi(n+1) can be obtained with A(Refn)=Refn+1 where A is the transformation from reference sequence RSn to reference sequence RSn+1.
RS
2
=A(RS1)
When the representation of the transformation A which generates RS2when applied to RS1 plus the representation of the reads versus RS2 corresponds to a lower entropy than the representation of the reads of class M versus RS1, it is advantageous to transmit the representation of the transformation A and the corresponding representation of the read versus RS2 because an higher compression of the data representation is achieved.
The coding of the transformation A for transmission in the compressed bitstream requires the definition of two additional syntax elements as defined in the table below.
It has to be observed that, in some cases the transformation applied to the reference:
Once the classification of reads is completed with the definition of the Classes, further processing consists in defining a set of distinct syntax elements which represent the remaining information enabling the reconstruction of the read sequence when represented as being mapped on a given reference sequence. The data structure of these syntax elements requires the storage of global parameters and metadata to be used by the decoding engine. These data are structured in a Genomic Dataset Header described in the table below. A dataset is defined as the ensemble of coding elements needed to reconstruct the genomic information related to a single genomic sequencing run and all the following analysis. If the same genomic sample is sequenced twice in two distinct runs, the obtained data will be encoded in two distinct datasets.
A sequence read (i.e. a DNA segment) referred to a given reference sequence can be fully expressed optionally using a subset formed of various combinations of the following descriptors:
For class U, the descriptor clips identifies those parts of the reads (typically the edges) that do not match, with a specified set of matching accuracy constraints, with the “internal” references.
Descriptor ureads is used to encode verbatim the reads that cannot be mapped on any available reference being it a pre-existing (i.e. “external” like an actual reference genome) or an “internal” reference sequence.
This classification creates groups of descriptors (syntax elements) that can be used to univocally represent genome sequence reads. The table below summarizes the syntax elements needed for each class of reads aligned with “external” (i.e. “pre-existing”) or “internal” (i.e. “constructed”) references. The asterisk “*” indicates mandatory descriptors always present in all encoded read for each class.
Reads belonging to class P are characterized and can be perfectly reconstructed by only a position, a reverse complement information and an offset between mates in case they have been obtained by a sequencing technology yielding mated pairs, some flags and a read length.
The next section further details how these descriptors are defined for classes P, N, M and I while for class U they are described in a following section Class HM is applied to read pairs only and it is a special case for which one read belongs to class P, N, M or I and the other to class U.
The pos descriptor is used to calculate the absolute mapping position on a Reference Sequence of the left-most mapped base of a Genomic Record. The value of each pos descriptor represents the difference between the coordinates, on the Reference Sequence, of the left-most mapping base of a Genomic Record and the previous one.
The first value of the pos descriptor in each coded Block is always 0 since no differential coding is possible for the first mapped read or read pair coded in an Access Unit. The absolute position of the first mapped read or read pair coded in an Access Unit is contained in the Access Unit Header.
The absolute position on the Reference Sequence of the left-most mapped base of the nth Genomic Record is therefore calculated as:
where p0 is the mapping value retrieved from the Access Unit Header for the first Genomic Record in the Access Unit.
In order to calculate the absolute position on a Reference Sequence of a base the following formula applies:
p=p
start
+n
del
−n
ins
+d
start+delta
where
Position
NOTE In case of paired end reads, when calculating the offset dstart from the Genomic Record Position, the two reads are considered contiguous. The relative reads position is taken into account when adding delta.
An example of calculation of the mapping position p for one base on a Reference Sequence is shown in
The rcomp descriptor conveys information about the strandedness of reads. Each bit of a decoded rcomp descriptor is a flag indicating if the read is on the forward (bit set to 0) or reverse (bit set to 1) strand.
The flag descriptor is a set of flaas as described in Table 6.
The mmpos descriptor represents the position, within a read or read pairs, of a mismatch with respect to the reference sequence. The position is represented as a distance from the position of the previous mismatch in the Genomic Record. The position of the first mismatch is represented as the distance from the left-most mapped base in the Genomic Record.
In case of paired reads, or in general records containing more than one genomic segment, the gaps between consecutive segments are not considered in the calculation of distances between consecutive mismatches.
If the encoded pair contains both read 1 and read 2, the positions of mismatches in read 2 are offset by the length of read 1. For example in case of reads with constant length equal to 100, if the first mismatches in the pair is in read 2 at position 44, the first mmpos descriptor decoded for this Genomic Record assumes the value 144.
If the described pair is missing read 1 (either because it is encoded in another block, or because read 2 is unpaired), the mismatch positions are not offset by the length of read 1. For example in case of fixed read length 100, if the first mutation in read 2 is at position 44, but read 2 is unpaired, the first mmpos descriptor decoded for this Genomic Record assumes the value 44.
Each mmpos descriptor is associated to an mmtype descriptor representing the type of mismatch occurring in the decoded read or read pair at the position calculated using mmpos.
An example of how to calculate mismatches positions in a read pair is provided in
The absolute position on a Reference Sequence of the ith mismatch in a Genomic Record shall be calculated as shown in Table 7.
In Table 7 the following variables are defined:
Sequences of mmpos descriptors referring to a Genomic Record are terminated with a reserved terminator value which cannot be interpreted as mismatch position.
The mmtype descriptor specifies the type of mismatch occurring in the decoded read at the position calculated using the associated mmpos descriptor.
The mmtype descriptor does not have a reserved value for the terminator since each Genomic Record shall contain the same number of mmtype and mmpos descriptors.
Table 8 lists the values and corresponding semantics of the mmtype descriptor according to the used alphabet.
The clips descriptor is used to represent clipped bases (a.k.a. soft or hard clips) in mapped reads or read pairs. This descriptor encodes soft clips as sequences of ASCII characters with additional elements to identify the position of the clipped bases in the read or read pair. In case of hard clips only the position and the number of clipped bases is encoded. Each descriptor contains a Genomic Record identifier followed by information related to the clipped bases position in the Genomic Record and the actual clipped bases in case of soft clips.
Syntax and semantics of the clips descriptor is provided in Table 9 and Table 10.
record_id is a counter of Genomic Records encoded in the current Access Unit.
clips_pos represents the position of the next clipped bases in the read or read pair. Values for positions have the following meaning:
soft_clipped_base is one of the symbols of the alphabet identified by alphabet_id.
hard_clipped_bases represent the number of hard clipped bases at the position indicated by the corresponding clips_pos;
The ureads descriptor represents verbatim sequence reads as a sequence of ASCII characters belonging to the currently used alphabet identified by alphabet_id.
The rlen descriptor is used only in case of variable length reads when reads_length=0 in the Parameter Set defined in this disclosure.
A decoded rlen descriptor represents the length of the current sequence read as number of bases including soft clips.
The information required to reconstruct paired reads is encoded using the pair descriptor. The pairing information associating one genomic segment to another can be expressed in three ways:
The pairing information is encoded as described in points 2 and 3 above when the first two bytes of a decoded pair descriptor have one of the values listed in Table 11.
When aligning reads to a reference sequence, read 2 can be mapped to a position that is smaller (e.g. to the left) than the mapping position of read 1; in this case, the pairing distance used in case 1 above will be negative. This implies that the information about reads strandedness is encoded in the sign of the pairing distance descriptor.
The reads distance is encoded as a 2-bytes signed integer, where:
This approach enables representing pairing distances in a range from −32766 to 32766. In case the reads are separated by a larger gap, the absolute position shall be encoded in the pair descriptor after the special value 0x7ffd or 0x8003 as defined in Table 11 and the two reads are encoded in two separate Genomic Records (i.e. the pair is “split”).
The decoding process of the reads distance is shown below:
sign=ReadsDistance & 0x0001;
ReadsDistance=ReadsDistance»1;
if (sign) ReadsDistance=-ReadsDistance;
The mscore descriptor provides a score per alignment. It shall be used to represent mapping/alignment score per read generated by genomic sequence reads aligners. The score shall be expressed using an exponent and fractional part. The number of bits used to represent the exponent and the fractional part are specified in the encoding parameters (see Parameter Set below). Table 12 shows how this is specified in IEEE RFC 754 for a 11-bits exponent and a 52-bits fractional part.
The score of each alignment shall be represented by:
If the base (radix) to be used for the calculation of scores is 10, the score is calculated as:
score=−1s×10E×M
The rgroup descriptor identifies the read group the Genomic Record belongs to. It is an unsigned 8-bit integer with values going from 0 to num_groups −1. The presence of read groups in an Access Unit is signaled by num_groups >0 in the Parameter Set as defined in the Parameter Set defined below.
The msar (Multiple Segments Alignment Record) descriptor supports spliced reads and alternative secondary alignments which contain indels or soft clips. msar is intended to convey information on:
msar can is used to represent mismatches, insertions, deletions, strandedness and clipped bases of secondary alignments of an aligned read
The following descriptors are defined for the support of multiple alignments.
The mmap descriptor is used to signal at how many positions the read or the leftmost read of a pair has been aligned. A Genomic Record containing multiple alignments is associated with one multi-byte mmap descriptor. The first two bytes of a mmap descriptor represent an unsigned integer N which refers to the read as a single segment (if spliced_reads_flag=0 as defined in this disclosure) or instead to all the segments into which the read has been spliced for the several possible alignments (if spliced_reads_flag=1). The value of N represents the number of values of the pos descriptor which are coded for the template in the current record. N is followed by one or more 8-bit unsigned integers Mi as described in this disclosure.
In case of multiple alignments, the rcomp descriptor defined in this disclosure is used to specify the strandedness of each read alignment using the same syntax specified above.
In case of multiple alignments, one mscore as defined in this disclosure is assigned to each alignment.
If no splices are present in the Access Unit, spliced_reads_flag is unset.
In paired-end sequencing, the mmap descriptor is composed by a 16-bit unsigned integer N followed by one or more 8-bit unsigned integers Mi, with i assuming values from 1 to the number of complete first (here, the left-most) read alignments. For each first read alignment, spliced or not, Mi is used to signal how many segments are used to align the second read (in this case, without splices, this is equal to the number of alignments), and then how many values of the pair descriptor are coded for that alignment of the first read.
The values of Mi shall be used to calculate P=Σi=1NMi, which indicates the number of alignments of the second read.
A special value of Mi (Mi=0) indicates that the alignment of the left-most read is paired with an alignment of the right-most read which is already paired with a kth alignment of the left-most read with k<i (then there is no new alignment detected, which is consistent with the equation above).
As an example, in the simplest cases:
When Mi is 0, the associated value of pair shall link to an existing second read alignment; a syntax error will be raised otherwise and the alignment considered broken.
Example: if the first read has two mapping positions and the second read only one, N is 2, M1 is 1 and M2 is 0 as said earlier. If this is followed by another alternative secondary mapping for the entire template, N assumes the value 3, and M3 assumes the value 1.
With respect to 10 the following applies:
If the dataset is encoded with spliced reads, the msar descriptor enables representaiton of splices length and strandedness as defined in this disclosure.
After having decoded the mmap and the msar descriptors, the decoder knows how many reads or read pairs have been encoded to represent the multiple mappings and how many segments are composing each read or read pair mapping. This is shown in
With reference to
With reference to
The mscore descriptor allows signaling the mapping score of an alignment. In single-end sequencing it will have N1 values per template; in paired-end sequencing it will have a value for each alignment of the entire template (number of different alignments of the first read possibly +the number of further second read alignments, i.e. when Mi−1>0).
Number of scores=MAX(N1, N2)+M0
where M0 represent the total number of Mi=0.
The number of scores associated to each alignment is signalled by the encoding parameter
as_depth as defined in this disclosure.
Table 14 shows how to calculate the number of descriptors needed to represent multiple alignments in one Genomic Record in case of multiple alignments with splices.
It may happen that the alignment process finds alternative mappings to another reference sequence than the one where the primary mapping is positioned.
For read pairs which are uniquely aligned, a pair descriptor shall be used to represent the absolute read positions when there is for example a chimeric alignment with the mate on another chromosome. The pair descriptor shall be used to signal the reference and the position of the next record containing further alignments for the same template. The last record (e.g. the third if alternative mappings are coded in three different AUs) shall contain the reference and position of the first record.
In case one or more alignments for the leftmost read in a pair are present on a different reference sequence than the one related to the currently encoded AU, a reserved value of the pair descriptor shall be used (not the same as the one used for alignments present to another reference in case of unique alignment). The reserved value shall be followed by the reference sequence identifier and the position of the leftmost alignment among all those contained in the next AU (i.e. the first decoded value of the pos descriptor for that record).
When an alternative secondary mapping does not preserve the contiguity of the reference region where the sequence is aligned, it may be impossible to reconstruct the exact mapping generated by the aligner because the actual sequence (and then the descriptors related to mismatches such as substitutions or indels) is only coded for the primary alignment. The msar descriptor shall be used to represent how secondary alignments map on the reference sequence in case they contain indels and/or soft clips. If msar is represented by the special symbol “*” for a secondary alignment, the decoder shall reconstruct the secondary alignment from the primary alignment and the secondary alignment mapping position.
Raw reads belong to class U only. They are encoded as unmapped reads in aligned datasets.
Some of the descriptors defined for reads aligned to an external or internal reference are used to encode raw reads. This is motivated by the fact that raw reads are encoded using reference sequences built from the data to be encoded.
The ureads descriptor represents verbatim sequence reads as a sequence of ASCII characters belonging to the currently used alphabet.
The rtype descriptor is used to signal the subset of descriptors used to encode one unmapped read or read pair in a Genomic Record as shown in Table 15.
The rtype descriptor also enables mixing reference-based and reference-less compression in the same Dataset. In this scenario rtype=0 signals reference based encoded records, while rtype>0 signals the set of descriptors to be used for reference less compression(in this case descriptors refer to the computed reference. when needed).
In an embodiment, the present invention uses context-adaptive binary arithmetic coding (CABAC) for the compression of the genomic descriptors. CABAC first converts to a binary representation all symbols to be encoded. The process of binarization converts a non-binary-valued symbol (e.g. a mapping position, a mapped read length or a mismatch type) into a binary code prior to arithmetic coding.
The selection of appropriate binarizations adapted to the statistical properties of each descriptor provides better compression ratios than existing formats based on general purpose compressor applied on blocks of heterogeneous elements. In the following sections these variables are defined:
While the following binarization tables are calculated for fixed values of these variables, it has to be appreciated that the present principles are not limited to these values, and thus other values can also be used in accordance with the present principles, while maintaining the spirit of the present principles.
Each binarization algorithm used in this disclosure is identified by an identifier as shown in Table 16.
This is a standard binary representation whereby each numerical value is coded in its binary representation. The variable cLength—shown in Table 28 when binarization_id =0—represents the numbers of bits with which the value will be represented.
A TU binary string is the concatenation of sym Val ones followed by one zero. If sym Val==cMax, the trailing 0-bit is discarded. Table 17 illustrates the bin strings of this truncated unary binarization with cMax=3.
The syntax for this binarization process along with arithmetic decoding is described below.
binValue is the binarized value which can be either 0 or 1.
The parsing process for genomic descriptors binarized using this technique begins with reading the bits starting at the current location in the bitstream up to and including the first non-zero bit, and counting the number of leading bits that are equal to 0.
This process is specified as follows:
The variable sym Val is then assigned as follows:
symVal=2leadingZeroBits−1+read_bits(leadingZeroBits)
where the function call read_bits reads a number of bits from a storage medium equal to the parameter passed as input. The value returned from read_bits(leadingZeroBits) is interpreted as a binary representation of an unsigned integer with the most significant bit written first.
Table 18 illustrates the structure of the Exp-Golomb code by separating the bit string into “prefix” and “suffix” bits. The “prefix” bits are those bits that are parsed as specified above for the computation of leadingZeroBits, and are shown as either 0 or 1 in the bit string column of Table 18. The “suffix” bits are those bits that are parsed in the computation of symVal and are shown as xi in Table 18, with i in the range of 0 to leadingZeroBits −1, inclusive. Each xi is equal to either 0 or 1.
Table 19 illustrates the explicit assignments of bit strings to symVal values.
Depending on the genomic descriptor, the value of a binarized syntax element is decoded using one of the following methods:
According to this binarization method the genomic descriptor is associated to the symVal by ordering the syntax element by its absolute value in increasing order and representing the positive value for a given absolute value with the lower symVal. Table 20 shows the assignment rule.
This binarization process requires the use of an additional input parameter tegParam which defines how the binarization is calculated.
Output of this process is the TEG binarization of the syntax element.
A TEG bin string is the concatenation of 1 (in case of symVal==0) or 2 (in case of symVal>0) types of binarization:
Table 21 illustrates the bin strings of this Truncated Exponential Golomb binarization with tegParam==2.
This binarization process requires the use of an additional input parameter stegParam. A STEG binary string is the concatenation of 1 (in case of symVal ==0) or 2 (for other cases) binarizations:
This binarization process requires the use of two input parameters splitUnitSize and outputSymSize. outputSymSize must always be a multiple of splitUnitSize.
The SUTU binary string is a concatenation of repeated TU binarizations, where each TU binarization is applied to portions of symVal which are splitUnitSize bits long. In other words, symVal is represented by x binary string obtained with the TU binarization, where x =outputSymSize/splitUnitSize. The cMax parameter for each binary string is defined as cMax=(1«splitUnitSize)−1.
Table 23 illustrates the binary strings of split unit-wise truncated unary binarizations with splitUnitSize=2 and outputSymbSize=8.
The bitstream syntax for this binarization process is described below.
This binarization process requires the use of two input parameters splitUnitSize and outputSymSize.
The SSUTU binary string is obtained by an extension of the SUTU binarization process with the sign of symVal coded as a separate flag.
Table 25 illustrates the binary strings of the Signed Split Unit-wise Truncated Unary binarization with splitUnitSize=2, outputSymbSize=8.
The syntax for this binarization process is described below.
sign_flag represents the cabac decoding of a bit on context variable identified by ctxldx.
decode_cabac_SUTU( ) represents the cabac decoding process for the SUTU binarization.
This binarization process requires the use of two input parameters splitUnitSize and outputSymSize.
The DTU binary string is a concatenation of two binarizations, namely the TU binarization and the SUTU binarization. The parameter cMax is used for the TU binarization, and parameters splitUnitSize and outputSymSize are used for the SUTU binarization (where its cMax is derived internally).
Table 26 illustrates the binary strings of the Double Truncated Unary binarization with cMax =1, splitUnitSize=2, outputSymSize=8.
The binarization process is described below.
decode_cabac_TU( ) represents the cabac decoding process for TU binarization.
decode_cabac_SUTU( ) represents the cabac decoding process for SUTU binarization.
This binarization process requires the use of two additional input parameters splitUnitSize and outputSymSize.
The SDTU binary string is obtained by an extension of the DTU binarization process with the sign of symVal coded as a flag.
Table 27 illustrates the bin strings of double truncated unary binarization with with cMax=1, splitUnitSize=2, outputSymSize=8.
The syntax for this binarization process is described below.
sign_flag represents the cabac decoding of a bit on context variable identified by ctxIdx.
decode_cabac_DTU( ) represents the cabac decoding with DTU binarization.
Each binarization algorithm introduced in the previous sections requires configuration parameters at the encoding and decoding ends. In an embodiment, said configuration parameters are encapsulated in a data structure described in Table 28. Each binarization algorithm is identified by an identifier as listed in Table 16.
In Table 28 the following semantics applies:
cMax represents the largest value to be binarized. Larger values will be truncated.
cLength represents the numbers of bits with which the value is binarized.
tegParam represents the tegParam variable defined in this disclosure for TEG binarization.
stegParam represents the stegParam variable defined in this disclosure for STEG binarization.
splitUnitSize represents the splitUnitSize variable defined in this disclosure for SUTU, SSUTU and DTU binarizations.
outputSymSize represents the outputSymSize variable defined in this disclosure for SUTU, SSUTU DTU and SDTU binarizations.
By applying the indicated CABAC binarization to the respective genomic descriptors as indicated in Table 29, the compression performance reported in * no additional information is necessary since it is already available in the compressed representation according to the principles of this disclosure.
Table 30 can be obtained. The improvement in compression performance of the method described in this disclosure can be appreciated by comparison with the corresponding file sizes of BAM and CRAM approaches and one of the best compressors in literature known as DeeZ (see Numanagic, I., et al “Comparison of high-throughput sequencing data compression tools”, Nature Methods (ISSN: 1548-7091), vol. 13, p. 1005-1008 London: Nature Publishing Group, 2016). It has to be appreciated that the DeeZ, BAM and CRAM compression performance are calculated by adding the size of the compressed reference genome used for alignment to the sizes of the compressed genome sequence data. According to the principles of the present disclosure, the reference genome is embedded in the compressed file. In today practice said compressed reference genome is a FASTA (ASCII text) file compressed using general purpose compressors such as GZIP, LZMA, Bzip2. In the proposed comparison the reference genome hs37d5.fa was compressed using the xz Linux command with the option of maximum compression (-9).
Table 30 shows the binarization applied to the genomic descriptors defined in this disclosure. When a concatenation of several binarizations is indicated, the different binarizations are applied to the different elements composing each descriptor as defined in this disclosure.
An example of binarization of rftp and rftt is provided in this section and illustrated in
Each nucleotide symbol is associated to an integer code:
After transformation the values become:
The binarized values for rftp are calculates as follows:
The binarized values for rftt are calculates as follows:
3. In the given example the five mismatches would be binarized using the TU binarization as:
With the binarization approach shown above the following compression results are achieved:
In an embodiment, the parameters needed to encode and decode each Access Unit are encapsulated in a data structure named Parameter Set as defined in Table 31.
The inventive techniques herewith disclosed may be implemented in hardware, software, firmware or any combination thereof. When implemented in software, these may be stored on a computer medium and executed by a hardware processing unit. The hardware processing unit may comprise one or more processors, digital signal processors, general purpose microprocessors, application specific integrated circuits or other discrete logic circuitry. The techniques of this disclosure may be implemented in a variety of devices or apparatuses, including mobile phones, desktop computers, servers, tablets and similar devices.
| Number | Date | Country | Kind |
|---|---|---|---|
| PCT/US2017/017842 | Feb 2017 | US | national |
| PCT/US2017/041579 | Jul 2017 | US | national |
This application claims priority to and the benefit of Patent Applications PCT/US2017/041579 filed on Jul. 11, 2017 and PCT/US17/17842 filed on Feb. 14, 2017.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/US2017/066863 | 12/15/2017 | WO | 00 |
