The partial statistics of 559 virus genuses and human genomes (C248).
The cost to extract one new biomarker within genomic sequences is very huge. This chapter adopts a scalable approach, developed previously and based on MapReduce programming model, to extract maximal repeats from a huge amount of tagged whole genomic sequences and meanwhile computing the similarities of sequences within the same class and the differences among the other classes, where the types of classes are derived from those tags. The work can be extended to any kind of genomic sequential data if one can have the organisms into several disjoint classes according to one specific phenotype, and then collect the whole genomes of those organisms. Those patterns, for example, biomarkers, if exist in only one class, with distinctive class frequency distribution can provide hints to biologists to dig out the relationship between that phenotype and those genomic patterns. It is expected that this approach may provide a novel direction in the research of biomarker extraction via whole genomic sequence comparison in the era of post genomics.
- comparative genomics
- class frequency distribution
- maximal repeat
- MapReduce programming
It is very attractive and challenging to discover markers  from genomic sequences and then to use these markers for genetic tests  to diagnose diseases and for personalized medicine to adverse drug responses [3, 4]. Nowadays, genome-wide association studies (GWASs)  have already examined single-nucleotide polymorphisms (SNPs) across human genomes to identify specific SNPs related to some diseases, for example, diabetes, heart abnormalities, Parkinson disease, and Crohn disease . Furthermore, GWAS is also used to predict cancer  and to influence human intelligence .
Most of GWASs are achieved with SNP arrays . The “Illumina”  uses the “whole-genome genotyping” to interrogate SNPs across the entire genome to obtain the most comprehensive view of genomic variation; the Affymetrix Genome-Wide Human SNP Array 6.0 features 1.8 million genetic markers which includes more than 906,600 single-nucleotide polymorphisms (SNPs) . The majority of these SNPs are designed to investigate the coding regions of genes in genomic sequences. However, some of the non-coding regions, once being mistaken as “junk DNA,” are believed to contain functions to regulate gene transcription and to account for the genetic differences between individuals . Although, on the one hand, the number of SNPs on one chip may be several hundreds of thousands, on the other hand, its coverage is still not enough  to figure out the relationship between genotypes and phenotypes in humans as given in the database of “dbGaP” .
As the era of post genomics with Next-Generation Sequencing (NGS) is coming, it is expected that the cost of genomic sequencing is decreasing and the availability of complete whole genomes of individual creatures is becoming popular. After using NGS for DNA sequencing , as shown on the right side in Figure 1, for example, one creature, for example, a virus, is supposed to contain three chromosomes with eight genotypes. On the other side of Figure 1, there are three phenotypes, for example, “Drug Resistance” “Envelope,” and “Contents,” inspected and detected by three domain experts, respectively. Under the assumption that these three phenotypes are totally dominated by those eight genotypes, represented as different icons, without considering the epigenetics , as shown in Figure 2, it is difficult for biologists in wet laboratory to analyze aimlessly the relationships among these phenotypes and those genotypes without further bioinformatics information or techniques such as comparative genomics .
With more and more complete whole genomes of distinctive creatures being available and popular in the coming days, it is very interesting and desired to extract common significant subsequences as candidate genomic markers as genotypes via comparing these creatures’ whole DNA sequences according to the classes (or types) of their phenotypes observed and specified by domain experts. Figure 3 shows the conceptual diagram of the corresponding classes for each of these three phenotypes given in Figure 1. With precise observations or experiments (phenotypes), biologists or experts can divide these creatures with complete whole genomes into disjoint classes if possible. Then, it is highly expected for biologists that some distinctive patterns (genotypes) hidden within their DNA sequences can be extracted as the candidates of class markers (phenotypes) if the frequency distributions of these patterns among classes are extremely biased, or some patterns are just in one class solely and appear in all instances belonging to that class ideally. To achieve the earlier-mentioned goal, one needs to extract repeats and to compute class frequency distributions of these repeats from a huge amount of tagged genomic sequences, where the types of classes are derived from the tags.
Due to the availability of genomic sequences in National Center for Biotechnology Information (NCBI) , The Cancer Genome Altas (TCGA) , it is interesting to have class frequency distribution of maximal repeats from these tagged genomic sequences for mining the biomarker or specific patterns. As the age of Next-Generation Sequencing (NGS) is going to be introduced for the project “Cancer Moonshot” in the National Cancer Institute , it is very attractive to identify specific biomarkers from these genomic sequences with tags, such as cancer types or distinctive genotypes. Figure 4 gives the conceptual diagram of how to reduce the gap between phenotypes and genotypes by using the phenotypes as classes to identify those subsequences that appear in unique class only as biomarkers.
The remainder of this chapter is organized as follows. Section 2 gives the review of potential applications with class frequency distributions of maximal repeats. Section 3 shows the scalable approach to extract maximal repeat from tagged sequential data. Section 4 describes the most recent work  that compute co-occurrences of DNA maximal repeat patterns appearing in both humans and viruses. Section 5 concludes and discusses on future works.
2. Potential applications with class frequency distribution of maximal repeats extracted from tagged sequential data
The previous work in  was a scalable approach based on Hadoop MapReduce programming model to overcome the computational bottleneck of using single computer with external memory [23, 24]. Furthermore, it had been applied for a USA patent (US-2017-0255634-A1)  whose publication data is as “Sep. 7, 2017” . Recently, in these 2 years, many novel and potential applications, derived from that work, were launched in diverse fields successfully, due to its scalability being able to handle a huge amount of sequential data. There were many experiments in diverse applications with a huge amount of tagged sequential data, such as textual data for trend analysis [26, 27, 28], genomic sequences for biomarker identification [21, 29, 30], time-stamped gantry sequences for significant travel time intervals  and, most recently, the sequences of product traceability for quality control .
The scalable approach of maximal repeat extraction adopted in this chapter is based on Hadoop MapReduce programming model, and the details can be found in . To illustrate the concept of the earlier approach clearly, as shown in Figure 5, there are 20 creatures generated manually. Each of them is with three phenotypes, “Drug Resistance,” “Envelope,” and “Contents,” as given in Figure 3, and all of its chromosomes are concatenated into one line which may contain genotypes including motifs, domains, or unknown DNA segments that are represented as icons for simplicity. Even though with the conceptual diagram as shown in Figure 5, it is still very difficult for users to catch the hidden connection (or relationship) among these three phenotypes and those icons (genotypes) at first glance, let alone each of these icons (genotypes) presents one continuous subsequence whose length is not fixed and its location is unknown within chromosomes.
To reveal the possible mapping of phenotype “Drug Resistance,” for example, to genotypes on purpose, Figure 6 presents the rearrangement in the order of these 20 chromosomes which may contain icons as hidden or unknown DNA segments. The mapping of different types of phenotype “Drug Resistance” to the corresponding genotypes (icons) can be observed. Similarly, one can have the mapping of different types of phenotype “Envelope” and “Contents” to the corresponding genotypes (icons). Due to the page limitation, the corresponding mapping of figures for “Envelope” and “Contents” are given in the supplements. Focusing on the repeats whose class frequency distributions are biased, as shown in Figure 7, one can estimate these repeats as candidate class markers which can be the clues for further experiments of analyzing the mapping of phenotypes and genotypes derived from 20 creatures in Figure 5.
4. Case study: mining for the co-occurrences of DNA maximal repeat patterns in both human and viruses
There were three studies with a huge amount of genomic sequences [21, 29, 30] based on the scalable approach of maximal repeat extraction with class frequency distribution mentioned in this chapter. This chapter only describes the most recent work  that the co-occurrences of DNA maximal repeat patterns appearing in both humans and viruses are extracted via a scalable approach that is based on Hadoop distributed computing ; that work aimed to mine for specific DNA patterns within human genomes via observing class frequency distribution of DNA maximal repeats extracted from the whole genomic DNA sequences of humans and 559 virus genuses. The detail in  is described for reference in the following.
4.1. Genome resources
In , Wang et al. extracted significant DNA sequences appearing in both the genomes of humans and viruses. In this study, the taxonomic level of viruses is “genus” and is selected as the classes (tags) for further experiments. Experimental resources included the complete whole genomes of humans (GRCh38.p7 Primary Assembly) downloaded from the NCBI FTP  and that of 559 virus genuses, including 2712 viruses that had genus name and were selected from the total of 4388 viruses download from in NCBI FTP  on January 14, 2017. Table 1 shows the partial statistics of 560 classes, including 559 virus genuses and the humans as “C248.” Note that each of the 24 human chromosomes is estimated as one individual instance for observing the frequency distribution among human chromosomes. This chapter, for simplicity, only takes the positive-strand DNA sequences of humans and viruses for further experiments.
|Class ID||Human and virus genuses||No of Instances|
|C247||Human mastadenovirus E||1|
4.2. Computational time and environment
To show the scalability of this approach from a practical view of point, as shown in Figure 8, the computational platform was the Hadoop cluster with eight computing nodes, two name (master) nodes, and six data (slave) nodes; Table 2 showed the specifications of hardware and software of one computing node; the computational time was about 37.5 h when the maximum length of maximal repeat patterns was limited to 500 bp (base pair).
|Hardware||CPU||Intel® Xeon® Processor E5-2630 v3 (8 cores)|
|RAM||128 GB (16GB*8, ECC/REG DDR4 2133)|
|Hard Disk||6 TB (SATA3 2 TB*3, 7200 rpm 3.5 inch)|
|Network Card||Intel Ethernet X540 10GBASE-T RJ45 DualPort *4|
|Hadoop||Hadoop 2.6 (“Cloudera Express 5.4.5”)|
4.3. The length distribution of DNA maximal repeats in both the genomes of human and 559 virus genuses
Comparing the maximal repeats that appear only in virus (Virus only), only in humans (Human only) or in both human and virus (Human and virus), Table 3 shows the partial frequency distribution of maximal repeats whose lengths are from 5 to 500 bp. It is observed that the majority of those maximal repeats whose length range from 7 to 11 almost belong to the “Human and Virus.” Note that there may exist extra nucleic acid codes, for example, “N,” within these DNA sequences such that the number of maximal repeat (length = 5) appearing in both humans and viruses in Table 3 is and that is great than (= 1024).
|Length||Virus (only)||Human (only)||Human and virus|
4.4. The longest DNA maximal repeat (length = 463 bp) appearing in both the genomes of human and 559 virus genuses
Table 3 shows the length of the longest maximal repeat extracted in both the genomes of humans and selected viruses of 559 virus genuses is 463 bp. In , the result of blasting two sequences, “Homo sapiens chromosome 5” (NC_000005.10) and “Human herpesvirus 6B” (NC_000898.1), as shown in Table 4, that longest repeat appears 109 times within human chromosome 5 and two times within virus “Human herpesvirus 6B.” To further inspect the longest maximal repeat, as show in Figure 9, one can find that the longest one is a tandem repeat (TAACCC) and appears within virus “Human herpesvirus 6B” at two intervals, the front (8249–8711 bp) and tail (161570–162,032 bp), that are located within the regions of direct repeats (DR) . Figure 10 gives one of two longest patterns aligned within “Human herpesvirus 6B” (8249–8711 bp) in Figure 9.
|Maximal repeat patterns||DF||TF||Length||Class frequency distribution (ClassID#DF#TF)||Regular expression||Human chromosome (GRCh38.p7 Primary assembly)||Viruses|
|ctaaccctaaccctaaccctaaccctaac||2||111||463||(C248#1#109)(C442#1#2)||(TAACCC)n||5||Human herpesvirus 6B|
4.5. The statistics of DNA maximal repeat patterns (length = 100 bp) appearing in both human and 559 virus genuses
Table 5, for example, shows the statistics of 13 DNA maximal repeat patterns (length = 100 bp) appearing in both human and 559 virus genuses. It is observed that the three repeats as the 1st, the 6th, and the11th, for example, have the similar regular expression as “(AACCCT)n”, “(CTAACC)n,” and “(TAACCC)n”, respectively, and all of them appear in human chromosomes “1,” “5,” “10” and “12”; all of these three repeats appear in these viruses, “Cyprinid herpesvirus 1,” “Falconid herpesvirus 1,” “Gallid herpesvirus 2,” “Human herpesvirus 6A,” “Human herpesvirus 6B,” and “Equid herpesvirus 3.” It is very interesting to investigate the relationship between these human chromosomes and those viruses for further research. On the other hand, from the biological viewpoint, furthermore, (AACCCT)n, (CCCTAA)n, and (CTAACC)n may comprise the same maximal repeat pattern with different repeat frame; (GGGTTA)n, and (AGGGTT)n can also comprise the same maximal repeat pattern in complementary sequence. Moreover, the (GGGTTA)n is expected to be targeted by cisplatin .
|Total||Length||Class frequency distribution (ClassID#DF#TF)||Regular expression||Human chromosome (GRCh38.p7 primary assembly)||Viruses|
|Maximal Repeat Patterns||DF||TF||C248||C5,C14,C149,C284,C305,C357,C442C541|
|10||591||100||(Cl 49# 1 # 18)(C248#4#299) (C305#2#41) (C442#2#208) (C541#l#25)||(AACCCT)n||1, 5, 10, 12||Cyprinid herpesvirus 1, Falconid herpesvirus 1, Gallid herpesvirus 2, Human herpesvirus 6A, Human herpesvirus 6B, Equid herpesvirus 3|
|3||42||100||(C248#1#18)(C284#1#11)(C357#1#13)||(AATAG)n||X||Rabbit fibroma virus, Taterapox virus,|
|9||188||100||(C248#7#152)(C305#2#36)||(AGGGTT)n||2,12,13,18,22, X, Y||Falconid herpesvirus 1, Gallid herpesvirus 2|
|6||533||100||(C14#1#34)(C248#5#499)||(AT)n||2, 3, 7, 19, X||Gryllus bimaculatus nudivirus|
|3||4||100||(C248#1#1)(C305#1#2)(C541#1#1)||(CCCTAA)n||5||Meleleagrid herpesvirus 1, Equid herpesvirus 3|
|11||591||100||(C149#1#18)(C248#4#298)(C305#2#41)(C442#3#210) (C541#l#24)||(CCCTAA)n||1, 5,10,12||Cyprinid herpesvirus 1, Falconid herpesvirus 1, Gallid herpesvirus 2, Human herpesvirus 6A, Human herpesvirus 6B, Human herpesvirus 7, Equid herpesvirus 3|
|3||55||100||(C149#1#32)(C248#2#23)||(GA)n||6, 11||Cyprinid herpesvirus 3|
|3||3||100||(C248#2#2)(C305#1#1)||G(GGGTTA)n||13, 18||Meleleagrid herpesvirus 1|
|9||188||100||(C248#7#152)(C305#2#36)||(GGGTTA)n||2, 12,13,18,22, X, Y||Falconid herpesvirus 1, Gallid herpesvirus 2|
|5||78||100||(C248#4#38)(C5#1#40)||(GT)n||2,10, 16, 19||Orgyia pseudotsugata MNPV|
|10||588||100||(C149#1#16)(C248#4#298)(C305#2#41)(C442#2#208) (C541#1#25)||(TAACCC)n||1, 5,10,12||Cyprinid herpesvirus 1, Gallid herpesvirus 2, Falconid herpesvirus 1 Human herpesvirus 6A, Human herpesvirus 6B, Rquid herpesvirus 3|
|5||76||100||(C248#4#35)(C5#1#41)||(GT)n||2,10, 16, 19||Orgyia pseudotsugata MNPV|
|9||186||100||(C248#7#150)(C305#2#36)||(AGGGTT)n||2,12,13,18,22, X, Y||Falconid herpesvirus 1, Gallid herpesvirus 2|
4.6. Phenotypes: “Group I(dsDNS)” in Baltimore virus classification
It is observed that all of these viruses in Table 5 belong to the “Group I(dsDNS)” of Baltimore classification , as shown in Table 6, and most of them are from the family “Herpesviridae” and order “Herpesvirales.” Indeed, it is very interesting and attractive to have all the viruses compared with human whole genome and to inspect these co-occurrences of repeats for virus prevention from the genomic point of view in the future.
|Viruses||Class ID||The International Committee on Taxonomy of Viruses (ICTV)||Baltimore classification|
|Orgvia pseudotsugata MNPV||C5||Alphabaculovirus||Baculoviridae||N||Group I(dsDNA)|
|Gryllus bimaculatus nudiviras||C14||Alphanudivirus||Nudiviridae||N||Group I(dsDNA)|
|Cyprinid herpesvirus 1||C149||Cyprinivirus||Alloherpesviridae||Herpesvirales||Group I (dsDNA)|
|Rabbit fibroma virys||C284||Leporipoxvirus||Poxviridae||N||Group I (dsDNA)|
|Falconid herpesvirus 1||C305||Mardivirus||Herpesviridae||Herpesvirales||Group I(dsDNA)|
|Gallid herpesvirus 2||C305||Mardivirus||Herpesviridae||Herpesvirales||Group I(dsDNA)|
|Meleagrid herpesvirus 1||C305||Mardivirus||Herpesviridae||Herpesvirales||Group I(dsDNA)|
|Taterapox virus||C357||Orthopoxvirus||Poxviridae||N||Group I(dsDNA)|
|Human herpesvirus 6A||C442||Roseolovirus||Herpesviridae||Herpesvirales||Group I(dsDNA)|
|Human herpesvirus 6B||C442||Roseolovirus||Herpesviridae||Herpesvirales||Group I(dsDNA)|
|Human herpesvirus 7||C442||Roseolovirus||Herpesviridae||Herpesvirales||Group I(dsDNA)|
|Equid herpesvirus 3||C541||Varicellovirus||Herpesviridae||Herpesvirales||Group I(dsDNA)|
5. Conclusions and future works
Except considering the phenotypes that result from the epigenetics , it is believed that some of the phenotypes of creatures (or organisms) are determined by their genotypes as they are born in the beginning. This chapter proposes a novel approach to mine for genetic markers via comparing class frequency distributions of maximal repeats extracted from tagged genomic sequences of creatures, where the classes are derived from the tags given by domain experts. Once domain experts can divide the creatures into disjoint classes as precisely as possible according to their features (phenotypes), then they can adopt the scalable approach developed in  to extract the maximal repeats and compute class frequency distributions of these repeats via comparing the whole genomic sequences of these creatures. The repeats or the combination of some repeats that are with extremely biased class frequency distribution can be seen as class markers (genotypes) and can provide clues to biologists to analyze the relationship among these class markers (genotypes) and their corresponding features (phenotypes).
Due to the availability of cloud computing with flexible infrastructure, nowadays, it becomes possible to compute class frequency distributions of maximal repeats from a huge amount of tagged whole genomic sequences of many creatures across species via the scalable maximal repeat extraction approach  with Hadoop MapReduce programming model. The function mentioned in this chapter is somewhat like “Archimedes’ Law of the Lever,” as shown in Figure 11, the Archimedes, an outstanding ancient Greek scientist, said that “Give me a place to stand on, and I will move the Earth.” With scalable computing power and enough tagged genomic sequences, in other words, a domain expert can figure out the relationship among phenotypes and genotypes if the classes are properly and precisely defined. It is desired to have further cooperation with domain experts, especially who have collected the whole genomes of diverse organisms and desire to find or identify the relationship between genomic signatures and the features they concern in the future.
From a practical point of view, it is inconvenient for general users to have experiments of maximal repeat extraction by themselves in the beginning because there are a lot of preprocessing works and need considerable hardware infrastructure to support such a big-data computing. Furthermore, it might be a bottleneck or nightmare for general users, for example, biologists, to implement Hadoop MapReduce programming as described in . Therefore, it is highly desired if maximal repeat extraction can be provided in public cloud services, such as Amazon Elastic Container Service (AWS ECS) , Google Cloud Platform , and Azure Container Service (AKS) . It is highly expected that one will develop novel comparative genome with tagged genomic sequences and bring users with novel cloud services of computing class frequency distribution of maximal repeats in the future.
This study is supported by Ministry of Science and Technology, Taiwan, under project MOST 106-2632-E-468-002. I thank Prof. Jeffrey J.P. Tsai who provided computational environments and finical support. I also thank Prof. Yi-Chun Wang for collecting the human chromosomes and inspecting experimental results about viruses and humans. I also thank Prof. Rouh-Mei Hu who helped in explaining the SNPs and the relationship between genotypes and phenotypes; Prof. Jan-Gowth Chang and Charles C.N. Wang had valuable discussions; I thank Jazz Wang for providing valuable Hadoop programming discussions. Finally, I thank Ling-Yu Ji for offering her drawing to show the conceptual diagram of “Archimedes’ Law of the Lever.”