JP2007319016A - Method for specifying or classifying target bacterium or phage as specific genus, species or serum type - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a relative-entropy algorithm for genomic fingerprinting capturing host-phage similarity. <P>SOLUTION: The method comprises the classification of target bacterium or phage by the following steps; (a) a step to specify a target genome sequence; (b) a step to form a randomized background genome derived from the target genome; (c) a step to specify an oligonucleotide influencing the difference between the background genome and the target genome by executing the iterative argorithm; and (d) a step to compare the oligonucleotide sequence selected by the step (c) with known bacterium or phage sequence. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The relative entropy algorithm for genomic fingerprints captures host-phage similarity.

  Genome analysis does not cover many sequence differences in organisms. Both mononucleotide and dinucleotide content, as well as codon usage, vary widely within the genome (6). The size of an equally small bacterial genome is statistically sufficient to determine features based on a substantially abundant population of sequences that describe each organism.

  However, many of these features remained unintelligible, especially in the coding domain, due to complex constraints. Each (protein coding) gene encodes a specific protein that constrains its potential nucleotide sequence. Since the genetic code is degenerate, this constraint further addresses the vast number of promising DNA sequences for each gene. Furthermore, the overall codon usage in each gene is probably known to show a strong biological outcome determined by equally allowing tRNA abundance (5). These constraints must be excluded from consideration in order to isolate new features within the coding region.

  To solve this problem, the inventors create a background genome that shares exactly the above constraints with the actual genome, but is otherwise random (4). The background genome encodes all the same proteins, and the codon usage is closely matched for each gene. The hidden features we are looking for are included in the difference between the background genome and the actual genome. The problem is reduced until these differences are drawn.

  The inventors have incorporated information theory into an algorithm that systematically computes a string of nucleotides (words) that are over and under-represented in the actual genome compared to those in the background (for details see Materials And methods). The main difficulty in finding these words is that they are not independent. For example, when the word ACGT is under-displayed, the ACGTA is also under-displayed in the same manner as ACG. The assumption is that only one of these words is biologically significant while the other is "ride the horse".

  This problem extends to all words. Since the set of words of a given length is limited and so is the genome, the frequency of any one word affects the frequency of all others. The inventors have modified an iterative algorithm that uses information logic measures to select the words that most contribute to the difference between the actual and background genomes. At each stage, this word was added to the list, and then its effect was excluded from consideration by resizing the background genome. In this regard, we have a list of words, each of which is likely to exhibit biological significance that contributes independently to the difference between the actual and background genomes. The size of the genome determines the length of the word that has the statistical power to solve. For typical bacteria such as Escherichia coli, they can conservatively contain up to 7 nucleotides in length. The amino acid order by gene and codon usage are fixed, so that features not covered by the algorithm of the present invention are complementary to mononucleotide content and codon usage. For typical bacteria, the algorithm finds 100 to 200 sequences between 2 and 7 nucleotides in length (Table 1). These previously hidden signals contain the value of biological information.

Materials and Methods Relative entropy algorithm.

  In order to find the most significantly over- and under-represented words in the coding region of the genome, we first created a randomized background genome that we used to compare with the actual genome. This is accomplished by randomly substituting the codon corresponding to each amino acid within all genes (4).

New coding sequences were created that show the same amino acid content and codon usage per gene as the actual genome, but are otherwise random. We then counted the number of occurrences for each word, w, of length 2 to 7 in this randomized genome (our selection of 7 considered the longest word is Read by the total length of the coding sequence, the average number of occurrences of each word should be> 0 because our algorithm is robust. The means for generating a random genome is repeated 30 times, at which point the standard deviation in the number of occurrences converges to that word. We then computerized the average N _B (w) background total for each word w. For reasons explained below, we determine N _B (w) by considering only a 7-word word and generating value for a short-length word by measuring a subsequence. Chose to do. We made L (w) equal to the length of the word w and C (W ⁱ ₇ , w) equaled the number of times the symbol string w was contained in the symbol string W ⁱ ₇ of length 7. As an example, if w is AAC and W ²⁵⁷ ₇ is AACAAAAC, then L (w) is equal to 3 and C (W ²⁵⁷ ₇ , W) is equal to 2.

N _B (W ⁱ ₇ ) = 1/30 × (total number of measurements of W ⁱ ₇ in all 30 background genomes)

同様に、我々は、Ｎ_R（ｗ）を、実際のゲノムでのｗの計数と等しくさせた。続くものの中で、我々は、計数よりむしろ頻度（または等価に確率）で作業し、それにより、我々は、式Ｐ_B（ｗ）＝Ｎ_B（ｗ）／ＬおよびＰ_R（ｗ）＝Ｎ_R（ｗ）／Ｌ（式中、Ｌは、我々のコーディング配列の総長さである）を用いて各ワードについての頻度を形成した。

Similarly, we made N _R (w) equal to the count of w in the actual genome. In what follows, we work with frequency (or equivalently probabilities) rather than counts, so that we use the formulas P _B (w) = N _B (w) / L and P _R (w) = N _R (w) / L (where L is the total length of our coding sequence) was used to form the frequency for each word.

The two frequency distributions P _B and P _R were the starting point for the word search algorithm. This algorithm consists of two stages repeated in sequence. In the first stage, the word that most clearly separates the actual from the background distribution was selected based on the measure of significance to be described below. In the second stage, the background probability distribution was resized to an appropriate size to exclude the word differences seen in stage 1 from consideration. These two steps were repeated a fixed number of times or until the background distribution was close enough to the actual one.

  The Kullback-Leibler distance between actual and background probability distribution is

によって示される。

Indicated by.

Then we either word w of length from 2 to 7, wanted to obtain a number of entities that measures the range to give the D _KL. Natural measurements are

によって示される。

Indicated by.

  This can also be thought of as a Kullback-Leibler distance between two probability distributions, in particular between a rough actual and background distribution that we only know if a given word is w or not (12). In the first stage of the iteration, we selected a word w of length 2 to 7 that maximized the significant measurement S (w).

The next step was to rescale the background distribution to an appropriate scale with minimal means so that the contribution of w was consistent in both the measured and background distributions. For resizing to a minimum, the ratio of the frequency of the length 7 words W ⁱ ₇ containing w the same number should not be changed. That is, we wanted to resize all the words W ⁱ ₇ to the same C (W ⁱ ₇ , w) with equal factors. Therefore, we had to work with a detailed probability distribution with the appropriate texture. Our distribution for the background was defined as a set of ₇ words W ⁱ ₇ of length _{7 with} probability P _B (W ⁱ ₇ ). We divided W ⁱ _{7 of} this set into discrete subsets where each element of a given subset contains the word w an equal number of times. These sets are

である。
そしてＪ＝｛０、．．．、６｝および

It is.
And J = {0,. . . , 6} and

我々は、実際およびバックグランド分布で所定の部分集合中にある確率が等しいように、これらのばらばらの部分集合Ｋ_J（ｗ）を適切な規模に直したかった。

We wanted to rescale these disjoint subsets K _J (w) so that the probability of being in a given subset in the actual and background distributions is equal.

それらが、古い確率分布（およびそれらの確率を加えた）からグループ分けされる要素であるので、十分に定義された確率分布がある。確率を保存しつつ、ｗの寄与を除外する規模修正は、

Because they are elements that are grouped from the old probability distribution (and their probabilities added), there is a well-defined probability distribution. The scale correction that excludes the contribution of w while preserving the probability is

（式中、全てのｉに関して、Ｗⁱ ₇ ∈Ｋ_J）
によって示される。この規模修正で、ｗの実体の数字は、ここでＳ規模修正（ｗ）＝０で、それによりＤ_KLに対するｗの寄与は、除かれたことをに特に言及する。その後、我々は、次のワードｗ’などを得る段階１を繰返した。

(W ⁱ ₇ ∈K _{J for} all i)
Indicated by. In this scale modification, the numbers w entities, here the S-scale modification (w) = 0, whereby the contribution of the w for D _KL is particularly mentioned that it was removed. Then we repeated step 1 to get the next word w ′ and so on.

It is not difficult to find with this iterative algorithm where the background distribution P _B (W ⁱ ₇ ) converges to the actual distribution P _R (W ⁱ ₇ ). This is because D _KL decreases monotonously (see below). It is well known that D _KL is not negative and is zero if the two distributions match or just match. The equation S (w) = 0 for all w also implies that the actual and background distributions match, so the algorithm does not stop before convergence is achieved. Eventually, D _KL will not be able to converge to a positive value since it would then be able to find a word that would reduce it below that value.

  For the application we had to conclude when the iteration no longer gave the list a statistically significant word. This interruption is when the likelihood variation is likely to create the most significant remaining word [the set of all words of length L (w)] appropriately modified for multiple hypotheses. . Interruption is selected word w

（式中、Δ（ｗ）は、ｗについてのバックグランド計数の標準偏差である）
を満足するときに起こる。本文献でのアプリケーションについては、我々は、１００回反復の後停止し、そしてそれは、実質的に中断より下である。

(Where Δ (w) is the standard deviation of the background count for w)
Happens when you are satisfied. For the application in this document, we stop after 100 iterations, which is substantially below the interruption.

Proof that D _KL decreases monotonically with scale modification. Considering the two probability distributions {p _j } and {q _j }, together with jεS and the set of promising results, the Kullback-Leibler distance is

である。これは、負でなく、そして、分布が一致する場合のみゼロである。

It is. This is not negative and is zero only if the distributions match.

Considering the disjoint division of S in the r set, S ₁ . . . S _r , ie
If k ≠ l and ∪ _i S _i = S, then S _k · S _i = ·

次に、きめが粗い確率を定義すると、

Next, if you define the coarse-grained probability,

Ｑ_iは、全てのｉについて＞０であると推定して、我々は、Ｐ_iおよびＱ_iの両方が、それら自体、確率分布であることに特に言及する。

Estimating that Q _i is> 0 for all i, we specifically mention that both P _i and Q _i are themselves probability distributions.

  If we define a distribution that is scaled appropriately,

そして、新たなＫｕｌｌｂａｃｋ−Ｌｅｉｂｌｅｒ距離は、Ｐ_iは、全てのｉについてＱ_iに等しい場合のみ対等である。

And the new Kullback-Leibler distance is comparable only if P _i is equal to Q _i for all i.

スコア・アルゴリズム。長さｇのゲノムＧに関して長さｓのコーディング配列を記録するために、我々は、最初に、以下の修飾を伴って、上に記述されるとおりＧについてのワードリストを作成した。それらが長さｓの配列に有為である場合のみに、ワードを、リストに加えた。この有為性は、各ワードについての総数および標準偏差を、規模ｓに直すことによって決定された。我々は、バックグランド・ゲノムおよび実際のゲノム中の各ワードの総数に、配列Ｓに関する予測される総数Ｎ_bおよびＮ_rを示すｓ／ｇを掛けた。標準偏差は、√ｓ／ｇにより適正な規模に直し、そしてΔ^sを示す。ワードが、式｜Ｎ_r−Ｎ_b｜＞３×Δ^sを満足する場合、それにより、それは、リストに含まれた；そうでなければ、それは飛ばされる。ｓは、ｇよりかなり少ないので、この基準は、上に記述される複数の仮説で修正された中断より実質的にいっそう厳密であった。バックグランド分布の規模を改めることを含めた残りの反復手段は、上に記述されるものと同じであった。

Score algorithm. To record a coding sequence of length s for a length G of genome G, we first created a word list for G as described above with the following modifications. Words were added to the list only if they were significant for an array of length s. This significance was determined by converting the total number and standard deviation for each word to scale s. We multiplied the total number of each word in the background and actual genomes by s / g which indicates the expected total number N _b and N _r for sequence S. Standard deviation, fix the proper scale by √s / g, and shows the Δ ^s. If the word satisfies the expression | N _r −N _b |> 3 × Δ ^s , then it was included in the list; otherwise it is skipped. Since s is significantly less than g, this criterion was substantially more rigorous than the interruptions corrected with the multiple hypotheses described above. The remaining iterations, including changing the size of the background distribution, were the same as described above.

  This new list L formed a scoring template using the number of words X. To obtain the score, we formed a background B of sequence S by the same Monte Carlo shuffle procedure as described above. We then implemented the following iterative algorithm: At each stage we pulled the word W from the indicated list L. We then compare the total number of that word in sequence S and background B, and the direction of bias for W between S and B is for W between genome G and its background. Our score only if it is the same as that of, i.e. if W is over-displayed in both G and S or under-displayed in both compared to their individual background. 1 was added. We then resized B with the means described above, excluding the effects of W, and proceeded to the next stage. Looking back at the full list L, we got the number Y of promising words X with a match between the genome and the sequence. The final score was C × (X−Y / 2) / √Y and C was constant. For each short sequence, scoring is performed on all 164 bacterial species in the NCBI database (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=Genome) and it is 253 chromosomes were included.

  Metrics for phylogenetic trees. The metric utilized a 50 kb slice and the scoring method described above. The distance between the two genomes A and B was calculated in three steps. First, all 50 kb slices of genome A were scored against whole genome B and then the scores were averaged. The same process was repeated for a 50 kb slice of genome B scored against genome A. Second, the two averages were made symmetric. Finally, the symmetric score was subtracted from the largest promising score. Although not following the triangle inequality, this distance exhibits most of the metric-symmetric positive positive zero property only when A is equal to B. We used the nearest cluster group and used the PHYLIP software package that outputs the tree (3).

Results The word list arises from the entire bacterial genome, but they correspond to DNA sequence features that are homogeneous throughout the genome. We confirmed this in two distinct ways. First, e.g. Using E. coli, we divided the genome in half and independently operated the algorithm on the two halves. The resulting list was the same until statistical variation. For the 100 word list, the top 80 words were in both lists. The process was repeated multiple times with various distributions and the results were similar.

For our secondary investigation that under- and over-representation of words is a local feature of the genome, we have created a basic algorithm that records the sequence of coding DNA based on the word list obtained from each genome. It was. This algorithm takes as its input a coding DNA sequence and a list of words and assigns a score to that sequence based on the under- and over-representation of the words in the sequence (see Materials and Methods). 253 bacterial chromosomes larger than 100 kb in length in the NCBI database were divided into 50 kb and 100 kb slices. These sequences were scored separately against all 164 species. 92 percent of the 100 kb slices scored highest in their own species. Using the 50 kb sequence, 86% scored highest in their own species. This confirms that the word corresponds to a feature that is homogeneous throughout each bacterial genome. Neither GC content nor codon usage show this property of homogeneity; both vary substantially within a single genome. Furthermore, this result suggests that the scoring means based on these hidden words is a useful sorter of sequences. For example, sequences obtained from microorganisms in the Sargasso Sea described by Venter et al. (9) can be compared to known bacteria without the need for similar genes. (The best known bacterial genome sorter is the oligonucleotide approach developed by Karlin and Cardon [6]. Even with our innocent first version of the scoring algorithm, The results are slightly better than those using the most comprehensive oligonucleotide approach involved in comparing the frequency of oligonucleotides showing lengths up to 4. Our scoring is based on Venter et al. [9 ] Substantially better than that of the dinucleotide approach used by
Our approach is also well suited to study the relationship between viruses and their hosts. Since the viral DNA is copied and expressed inside the host, it may be expected that the viruses and their hosts will share some evolutionary pressure. However, mononucleotide content and codon usage vary dramatically between the host and the phage. Some information was obtained from oligonucleotide comparisons, but our score system described in “Materials and Methods” is better than 60%. Of the population of DNA phage sequenced on the NCBI website, 185 phage has a known primary host. Many of the phage are known or assumed to have multiple host species within the same genus. For this reason we considered host targets at the genus level. The 164 species fall into 108 different genera. For our algorithm, the correct host genus gave the highest score for 93 out of 185 phage, and 131 phage had the correct host with the top three scores (Table 2). For comparison, the best oligonucleotide score system showed 58 out of 185 correct host genera. Both codon usage and mononucleotide content are few predictors of the phage host.

  By limiting our analysis to double-stranded DNA (dsDNA) phage that encompass most of the known phage, our host prediction has clearly improved. Removing 35 single-stranded DNA phage improved scoring to 87/150 or 58% for the top score and 123/150 or 82% for the top three scores. Phages can be further stratified into temperature and lysine (1). Note that for the temperature dsDNA phages that make up the majority of the sequenced phages, our prediction for the host was very good (93% for the top 3 and 70% for the top score). Lysine phage are still better than 50% in the top three, but similar scores are not obtained, indicating that their DNA is not under the same evolutionary pressure as that of the host.

我々は、我々のスコア・アルゴリズムを、ゲノム間の距離を形成するのに適合させた（「材料および方法」を参照）。階層クラスター群を、１６４の細菌種の集合の距離の行列に使用して、我々は、系統樹を作成した（図１ａ）。この樹は、標準細菌分類学の大半を捕捉する。例えば、図１ｂを参照すると、エンテロバクテリア（Enterobacteria）を、同じ分岐群にグループ分けされることを示している。これは、ワードリストによりコーディングされた特性が、進化的に保存されることを示唆する。距離は、全ゲノム特性に基づいているので、我々は、遺伝子水平移入のような系統樹木を作成する上で共通の落とし穴のいくつかを避けた。さらに、この方法は、均質な遺伝子または多量の配列決定されたゲノムを必要せずに、この樹での新たな種の追加が出来た。

We adapted our scoring algorithm to form a distance between genomes (see “Materials and Methods”). Using the hierarchical cluster group in the distance matrix of a set of 164 bacterial species, we created a phylogenetic tree (FIG. 1a). This tree captures most of the standard bacterial taxonomy. For example, referring to FIG. 1b, it shows that Enterobacteria are grouped into the same branch group. This suggests that the characteristics coded by the word list are evolutionarily preserved. Since distance is based on whole-genome traits, we avoided some of the common pitfalls in creating phylogenetic trees like horizontal gene transfer. Furthermore, this method allowed the addition of new species in this tree without the need for homogeneous genes or large amounts of sequenced genomes.

We have introduced an algorithm that finds over 100 new signals in the coding region of each bacterial genome. The set of applications represented includes the use of these signals (words) as a sorter, and the genomic linkage between phage and their hosts, as well as the creation of statistical trees. These are just a subset of the algorithm potential uses. Certain promising uses for eukaryotes include slice site detection, mRNA degradation or stabilization signals, tissue specificity, and host-virus relationships. The actual exon has an over-representation signal like the exon split enhancer (2). Our algorithm can measure an aggregated list of excess and under-represented sequences in actual exons, and it can be used to dissociate actual exons from confusing intronic sequences. is there. For mRNA stability, a few groups measured decay constants for a number of mRNAs in diverse organisms including humans (8, 11). The half-life range is in the order of two orders of magnitude, but the signal or structure that determines this difference in stability is unknown. If our algorithm is used for the 1,000 most rapidly decaying mRNAs and the 1,000 most stable set of mRNAs, the difference between the two lists should provide an important set of signals. Absent. For tissue specificity, genes first expressed in various tissues have been shown to exhibit unique properties over the last few years; their codon usage and GC content are different (7, 10). We should be able to find another signal that separates the tissues. These signals indicate the possibility of providing information about the host tissue regarding the virus. Unlike codon usage and mononucleotide content not shared by phage and bacterial hosts (or by human viruses and their host tissues), our algorithm is a very good predictor of viral hosts.

  This algorithm can also be used to help find transcription factor binding sites. From the DPI interact database (http://arep.med.harvard.edu/dpinteract/), we A set of known binding sites for 13 transcription factors with 15 or more binding sites listed for E. coli was extracted. Binding sites were determined by the collection of weight matrices that score the binding motif. Actual e. The E. coli genome is developed into a weight matrix and they are added to the background e. By comparing with the coli genome we found that 12 of the 13 motifs are clearly under-represented (4 standard deviations) in the coding region. This measure can be used as a filter to determine if a motif that is readily available is real when a commonly used motif finder pops out an excess signal that is not an actual transcription factor binding motif.

図１（ａ）ＮＣＢＩデータベースで見られる１６４の細菌種についての系統学上の樹。三角形は、エンテロバクテリアの分岐群を含む。（ｂ）その樹のエンテロバクテリアの分岐群の引き伸ばし。アシネトバクター株ＡＤＰ１、ニトロソモナス・ヨーロピアエ、エルビニア・カロトボラ、エッシェリキア・コリ、サルモネラ・エンテリカ、サルモネラ・エンテリカ・セロバル・ティフィ、シゲラ・フレクスネリ、ホトロアブダス・ルミネッセンス、エルシニア・ペスチス、エルシニア・シュードツバキュロシス、イジオマリナ・ロイヒエンシス、シゲラ・オネイデンシス、ビブリオ・コレラエ、ビブリオ・パラハエモリチクス、およびビブリオ・ブルニフィカスについての結果が示される。この群から消えている唯一のエンテロバクテリアは、ブキネラ・アフィジコラである。

FIG. 1 (a) Phylogenetic tree for 164 bacterial species found in the NCBI database. The triangle contains a branch group of Enterobacteria. (B) Stretching the enterobacterial branch of the tree. Acinetobacter strain ADP1, Nitrosomonas europeiae, Ervinia carotobola, Escherichia coli, Salmonella enterica, Salmonella enterica ceroba tifi, Shigella flexneri, Hotroabdas luminescence, Yersinia pestisio Results are shown for Leuhiensis, Shigella Oneidensis, Vibrio cholerae, Vibrio parahaemoriticus, and Vibrio Brunificus. The only enterobacteria disappearing from this group is Bukinera aphidikola.

  For the wide use of these signals, we expect the development of a background model that clearly represents the true coding region. Many bioinformatics issues require looking for long motifs or sequences by comparing it with arbitrary criteria. These problems indicate the difficulty that there is no way to generate a background model that includes all of the basic principles in the actual genome. Our algorithm determines all of the short global principles. The creation of standards and models that anticipate these principles will make it possible to track a variety of difficult bioinformatics issues.

Claims (1)

A method for identifying or classifying a target bacterium or phage as a specific genus, species or serotype,
a) providing a genomic sequence of the target bacterium or phage;
b) providing a randomized background genome with the same amino acid content and codon usage per gene as the target genome;
c) Run an iterative algorithm to identify an oligonucleotide sequence of about 2 to about 7 nucleotides in the target genomic sequence that represents the sequence that most affects the difference between the background and target genomes in length The iterative algorithm comprises:
i) selecting the oligonucleotide sequence that most largely separates the target genome from the background genome;
ii) rescaling the background establishment distribution to factor out differences for the oligonucleotides selected in step (i);
iii) repeating the step (i) until the background distribution approaches the target distribution;
A step including:
d) comparing the oligonucleotide sequence selected in step (c) with a known bacterial or phage sequence;
A method characterized by comprising.