https://orcid.org/0000-0002-0503-4176
Figures
Abstract
The huge amount of data acquired by high-throughput sequencing requires data reduction for effective analysis. Here we give a clustering algorithm for genome-wide open chromatin data using a new data reduction method. This method regards the genome as a string of 1s and 0s based on a set of peaks and calculates the Hamming distances between the strings. This algorithm with the systematically optimized set of peaks enables us to quantitatively evaluate differences between samples of hematopoietic cells and classify cell types, potentially leading to a better understanding of leukemia pathogenesis.
Author summary
High-throughput sequencing provides us huge amounts of data about gene regulation. In order to extract useful information from the data, data reduction is needed. Although RNA-seq data analysis has been extensively studied, where the focus is mainly on genetic loci, tools for epigenetic sequencing data, such as ATAC-seq data which represent chromatin accessibility, are comparatively lacking. Since the binding of transcription factors mainly occurs in open chromatin regions, it is presumably important to understand how chromatin accessibility landscape affects cell phenotype. In this context, we developed a systematic algorithm to select a set of peaks representing the open state of chromatin for a given sample of ATAC-seq data. This algorithm quantifies the difference between samples by regarding the genome as a string of 1s and 0s with Hamming distances and then performs hierarchical clustering. This algorithm has less computational cost and gives a reasonable cell type classification compared to a previous method. In this work, as an application of this algorithm, we present a comparative analysis of leukemia samples with healthy hematopoietic cells and provide new insights about the relationship between chromatin structures, cell surface proteins, and symptoms in leukemia.
Citation: Tanaka A, Ishitsuka Y, Ohta H, Fujimoto A, Yasunaga J-i, Matsuoka M (2020) Systematic clustering algorithm for chromatin accessibility data and its application to hematopoietic cells. PLoS Comput Biol 16(11): e1008422. https://doi.org/10.1371/journal.pcbi.1008422
Editor: Avner Schlessinger, Icahn School of Medicine at Mount Sinai, UNITED STATES
Received: May 7, 2020; Accepted: October 6, 2020; Published: November 30, 2020
Copyright: © 2020 Tanaka et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All ATAC-seq and RNA-seq data needed to reproduce this study have been deposited at the DNA Data Bank of Japan (DDBJ) under accession number DRA010939. The source code is available from https://github.com/tanakanishi/findclosest.
Funding: This research was supported by JSPS KAKENHI Grant Numbers JP19K16740 (AT), JP18J40119 (AT), JP19H03689 (MM), JP20H03514 (JiY), and by Japan Agency for Medical Research and Development (AMED) Grant Numbers JP20fk0108088h0002 (MM), JP17km0405207h0002 (AF), JP18km0405207S0103 (AF), and by a grant from the Naito Foundation (AT). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors declare that they have no conflict of interest.
Introduction
Cellular phenotypes are governed by epigenetic mechanisms. For example, information about how human DNA is packed and chemically modified in the nucleus plays an important role in understanding the differentiation and regulation of cells [1–4]. Methods such as chromatin immunoprecipitation sequencing (ChIP-seq) and assay for transposase accessible chromatin using sequencing (ATAC-seq) have proven useful for understanding the modification and detection of open chromatin on a genome-wide scale [5–9]. Those epigenetic data analysis methods usually start with data enrichment along the whole genome, also known as “peak calling” [10, 11].
Compared to RNA-seq data analysis, whose target regions are mainly in certain loci or genes across samples, the target regions on epigenetic sequencing data are undetermined. To determine the target regions, peak calling with an appropriate tool is often performed for the entire genome of every sample, and the target regions are defined as merged peaks among all samples. Then the total number of reads or fragments present in each region is counted for each sample, leading to a matrix, X = (xi,j), where xi,j represents the number of reads/fragments from sample i in region j. The matrix elements are normalized by quantile normalization to reduce the biases arising from variations in the data size over samples, followed by downstream processing [7–9].
However, this process raises two concerns. First, we do not fully understand the effect of merging all the peaks from different samples. For example, if two peaks from different samples slightly overlap, those two peaks are considered as one peak after the peak merging step. Therefore, the difference of the two peak positions, which may reflect cell identity, may be unintentionally ignored. The second concern is that we have no justification for applying quantile normalization over samples that are phenotypically different [12, 13].
Thus, the aim of the present study is to avoid these concerns by constructing an algorithm that systematically classifies epigenetic data obtained from high-throughput sequencing. In this analysis, toward cell type classification, we provide a systematic algorithm to select a set of peaks used for the downstream analysis, where the difference between samples are quantified by using the Hamming distance from information theory [14]. This algorithm has less computational cost while still producing reasonable classification compared to a previous method [7].
As an application of the developed algorithm, we use it to obtain new insights on samples of leukemia cells from chronic lymphocytic leukemia (CLL), acute myeloid leukemia (AML), and adult T-cell leukemia (ATL) at the chromatin level. In particular, using this algorithm, we infer the phenotype of a given leukemia sample as output by using only ATAC-seq data of that sample as input.
Results
ATAC-seq samples
In this paper, we mainly focused on 77 ATAC-seq datasets from 13 human primary blood cell types [7] as test data. The 13 cell types are comprised of hematopoietic stem cells (HSC), multipotent progenitor cells (MPP), lymphoid-primed multipotent progenitor cells (LMPP), common myeloid progenitor cells (CMP), megakaryocyte-erythroid progenitor cells (MEP), granulocyte-macrophage progenitor cells (GMP), common lymphoid progenitor cells (CLP), natural killer cells (NK), B cells, CD4+T cells (CD4+T), CD8+T cells (CD8+T), monocytes (Mono) and erythroids (Ery). These cell types are experimentally categorized by immunophenotypes described by the combination of cell surface markers shown in Table 1.
Types of hematopoietic cells and their corresponding cell surface markers in [7]. For example, CD34+ and CD38- for cell type ν means that a cell of type ν expresses CD34 but not CD38 at its surface.
For convenience, denotes a set of the thirteen cell types;
For all 77 samples, we assigned ATAC-seq reads to reference genome hg19 (http://hgdownload.cse.ucsc.edu/goldenPath/hg19/database/), and among them only those which had high mapping quality values (MQ ≥ 30) were used for the peak calling by MACS2 (see S1 Appendix for details of the preprocessing) [15]. The peak calling results consisted of the location with a peak width and the associated p-value. Concretely, the location of the k-th peak is expressed by gk = (γk, αk, βk), where γk is the chromosome number, αk is the start position, and βk is the end position. Note that we used MACS2 to call all ATAC-seq peaks with the following parameters (--nomodel --nolambda --keep-dup all -p pG), where the number of peaks is affected by the peak calling parameter ‘-p pG’. The parameter pG is larger than any p-values of the peak calling results. (See Materials and methods for details of the peak-calling).
Note that the peak position depends on parameter pG of the MACS2 algorithm as shown in Fig 1. For example, the start and end positions of a peak could change and one peak could split into two peaks depending on pG. Thus, we need to take into account the dependence of a set of peaks on different values of pG for careful analysis.
The plots show representative data of Mono obtained from SRA with accession number SRR2920475. (A) The number of reads Yx at each position x along chr 1 (γ = 1) and the peak region (αk, βk) as determined by the MACS2 algorithm with peak calling parameter pG = 10−2 (pink shaded regions) is shown. The peak region and its associated p-value ((αk, βk), pk) are (1092756, 1094068, 10−20.36428). (B) The obtained peak regions are ((1092817, 1093330), 10−20.36428) and ((1093480, 1094025), 10−8.19447) for pG = 10−4.
Parameterized binarization
First we ranked the peak results in the order of ascending p-values and then investigated the relationship between the peak width and the corresponding ranking. We found that as the p-value increased, the width of the ATAC-seq peaks became shorter statistically, which suggested the feasibility of robust data reduction against small noise in the data by selecting peaks with smaller p-values (Fig 2).
Distribution of peak width (βk − αk) and its corresponding ranking k obtained from the peak calling result of CD4+T cells with peak calling parameter pG = 10−2. The bin size is 400 × 400. The color code indicates the number of data in each bin.
Thus, we define Mcut as the threshold such that only peaks with rankings not greater than Mcut are used for the analysis hereafter. Then, for a given set of (Mcut, pG), we introduce B = {hγ,x}, where hγ,x = 1 when position x in chromosome γ is inside a peak and 0 otherwise (Fig 3). The process to obtain the binary sequence from the reads data is illustrated in Fig 4. Note that we do not perform any coarse-grained description for the genome position x but keep 1bp resolution. (See Materials and methods for details of the binarization).
Schema of the Hamming distance calculation from the peak locations with two samples . Each locus is converted to 1 or 0 based on the peak overlapping status.
(A) The number of reads Yx at each position x along chr 3 (γ = 3) and the peak region (αk, βk) as determined by the MACS2 algorithm with peak calling parameter pG = 10−2 (pink shaded regions). This figure shows representative data of NK cells obtained from SRA with accession number SRR2920495. The peak regions and the associated p-values ((αk, βk), pk) in the left and right peaks are ((188271079, 188271985), 10−422.5872) and ((188286401, 188287077), 10−329.52139), respectively. Thus, the width of the peaks (βk − αk) in the left- and right-hand sides are 906 and 676, respectively. (B) Binary sequence (hx) as determined by the peak regions seen in (A) when we chose Mcut satisfying .
Quantifying differences between two binary sequences by Hamming distance
Let us move onto the situation when one considers a set of samples to evaluate the difference between two binary sequences B. Here our strategy is to find the proper distance that can be measured from the normalized ATAC-seq data of two samples. Using that distance, we try to obtain hierarchical clustering of a set of hematopoietic cell samples to quantitatively characterize the relationship among those samples.
Let Ns be the number of samples. We then write the set of samples as
where Ns = 77 in this study. For sample
, we add index c to related objects as a superscript. For example, we write a binary sequence B associated to sample c as
.
There are many methods to evaluate the difference between a binary sequence Bc from sample and
from sample
. In this paper, we evaluated the difference between two samples (c, c′) by using the Hamming distance
between two binary sequences, Bc and
.
is calculated as the sum of the number of pairs with different values at every position x between Bc and
(Fig 5). We used the distance as an initial condition for the hierarchical clustering and then used Ward’s method to complete the hierarchical clustering [16]. Examples of hierarchical clustering with (Mcut, pG) = (2000, 10−2) and (80000, 10−2) are shown in Fig 6. (See Materials and methods for details of the Hamming distance and hierarchical clustering).
Matrix of Hamming distances dij between samples i and j. This matrix is used for the downstream analysis.
Hierarchical clustering obtained by Ward’s method with parameters (Mcut, pG) = (2000, 10-2) (A) and (80000, 10−2) (B).
Optimization of hierarchical clustering toward cell-type classification
By using the methods explained above, we can obtain a clustering dendrogram that depends on (Mcut, pG). We then need to systematically determine the best clustering, which is the clustering closest to the “perfectly classified dendrogram” where each set of all samples with type
coincides with an offspring set. This condition can be restated as an optimization problem by introducing a cost function “penalty” for the performance of clustering as follows.
Concretely, to quantitatively evaluate the obtained dendrogram for each combination of (Mcut, pG), we define type penalty λν for a given cell type . Type penalty λν corresponds to the number of samples from different cell types in cluster ν formed when all samples of cell type ν meet together from the bottom of the dendrogram (Fig 7). Additionally, we define global penalty
as the “cost function” of the optimization. Note that λ ≥ 0, and a “perfectly classified dendrogram” gives λ = 0. (See Materials and methods for details of the penalty).
Note that this dendrogram is constructed by artificial data to explain how to calculate the penalty, though we use the same labels such as HSC1. This dendrogram has six leaves, and three of them are classified to type HSC. To explain details of this dendrogram, we freely use the symbols and definitions in Materials and methods in this caption. We can see that τ(HSC) = 10. The corresponding node is n10 (displayed by the blue dot), and the corresponding cluster is the set {HSC1, HSC2, HSC3, MPP} (surrounded by the blue dashed line). Among the elements of
, one leaf, MPP, is not in type HSC, but the three others are. Hence, the type penalty of HSC in this figure is computed as λHSC = 4 − 3 = 1.
Determination of the best parameters for the optimization
As mentioned above, the optimization problem we have to solve is to find that minimizes the cost function λ(Mcut, pG). The schematic workflow in our algorithm is shown in Fig 8.
First we took into account all the peaks by setting Mcut = ∞ and checked how the dendrograms and λ(∞, pG) depended on pG, as shown in Fig 9. Considering the tendency of the parameter searching, we concluded that .
Global penalty λ(Mcut = ∞, pG) obtained by Ward’s method.
We then sought the best parameters to optimize the dendrograms and found that was close to (64000, 10−2), which gave the smallest penalty λ in our searching resolution, as shown in Figs 10 and 11. Note that 64000 is the midpoint of (60000, 62000, 64000, 66000, 68000) which give the same minimum penalty in our searching resolution. Hereafter, to investigate the property of the best clustering, we set
as (64000, 10−2). In our searching resolution, the increment in terms of Mcut was 2000 near Mcut = 64000. Note that more-refined resolutions might give better estimates of the optimized value
, but naturally the computational costs get higher. Even then, the following procedures are operationally unchanged.
The distribution of global penalty λ (A) and type penalty λν for each cell type ν (B) along with Mcut with parameter pG = 10−2 by Ward’s method.
Hierarchical clustering obtained by Ward’s method with (Mcut, pG) = (64000, 10−2).
The value of the minimum penalty achieved at was 18. This minimum was smaller than the penalty value of 27 for the clustering of the data from GSE74912_ATACseq_All_Counts.txt in [7]. The procedure of the latter clustering was as follows. First we performed a quantile normalization of the reads count in the distal elements (> 1000 bp away from a transcription start site (TSS)). Then we calculated the Pearson coefficients over all samples leading to a distance matrix where each entry is 1-(Pearson coefficient). By using Ward’s method, we finally obtained the clustering dendrogram. Note that for this case, Ward’s method gives penalty λ = 27 and UPGMA gives λ = 29.
Computational cost of the algorithm
As explained above, after obtaining data of the reads positions, we perform the MACS2 algorithm to get peak regions, and then finally we produce a hierarchical clustering. Here we consider the computational cost of our algorithm after acquiring the data of the reads positions and until acquiring a distance matrix to produce the hierarchical clustering. Note that the computational cost of the MACS2 algorithm is not more than O(Ns), where O() is the Landau notation and Ns is the total number of samples. We consider two situations. (i) One is the case where new samples to analyze are given. (ii) The other is the case where one new sample to analyze is added to the already analyzed samples, for which peak regions and the distance matrix are already calculated. For case (ii), we use the symbol Ns to write the total number of already analyzed samples. We claim that the computational cost of our algorithm is significantly lower than that of a previous method using target regions merged over samples [7] for large values of Ns for case (ii) and, in our case with Ns = 77, that the computational cost of our algorithm is practically lower for case (i).
Specifically, in case (i) for our algorithm, the corresponding computational cost is , which comes solely from the calculation of the Hamming distance. In case (ii), the corresponding computational cost is K2 Mcut Ns, which also comes solely from the calculation of the Hamming distance. Note that K1 and K2 are constants that do not depend on Mcut or Ns.
In the context of estimating the best optimization parameter , by using Mm different values for Mcut, the computational cost becomes
for case (i) and K2 Mcut MmNs for case (ii), where Mm does not depend on Ns or genome size L and can be adjusted according to the searching resolution of the optimization. Note that K1 and K2 do not depend on Mm. In addition, we optimize pG by Mp different values for pG. Since this optimization can be done for any algorithm, we do not take into account this cost for the comparison of different algorithms. Typically, we set (Mm, Mp) ≃ (30, 10) in our optimization corresponding to case (i). Note that in the section of “Application to leukemic cells” discussed later, corresponding to case (ii), we use the optimized parameters
, leading to (Mm, Mp) = (1, 1).
The previous method using targeted regions merged over samples in [7] includes (a) the merging of reads before peak calling and (b) calculating the distance matrix by the Pearson coefficients which automatically depend on Ns. Thus, for a given number Nnew of unanalyzed samples, the computational cost corresponding to the process of (a) and (b) is at least , where Nr is the minimum reads number over all samples, and L1 is the number of target regions merged over all samples. The first term comes from counting the reads and the second term comes from calculating the distance matrix. Note that Kr is a constant that does not depend on Nr or Nnew, and KL is a constant that does not depend on L1 or Ns. This form of the computational cost
is the same for case (i) with Nnew = Ns and case (ii) with Nnew = 1, leading to the conclusion that the computational cost of our algorithm is significantly lower than the previous method, especially for case (ii) with sufficiently large Ns. We do not have the exact estimate of the coefficients K1, K2, Kr, KL, but because
and
in our case, then
could be costly compared to
. In practice, even in case (i) with Ns = 77, we numerically found that the computational cost of our algorithm is lower due to our algorithm not using the process of merging reads unlike [7].
How to relate the best parameters to genomic context
In order to understand why ATAC-seq data under the condition of (Mcut, pG) = (64000, 10−2) was well classified, we analyzed the properties of the peaks with higher rankings.
The result of the previous section suggested that peaks of with
included key regions for characterizing cell types. Therefore, we investigated which functional genomic regions such as promoters, enhancers, etc. are dominantly related to these top 64000 peaks.
Functional annotation of peaks depending on rank.
In order to investigate functional annotations on the genome overlap with ATAC-seq peaks data, we applied the top 80000 peaks in three cell types (HSC, B cells, and Mono) to the 15-state ChromHMMmodel data. One can obtain data of the biological functions on the genome for HSC, B cells, and Mono from an integrative analysis of 111 reference human epigenome datasets, where we used the data of E032 for B cells, E035 for HSC, and E029 for Mono (https://egg2.wustl.edu/roadmap/data/byFileType/chromhmmSegmentations/ChmmModels/coreMarks/jointModel/final/) [17].
ATAC-seq peaks were ranked according to p-values and divided into groups consisting of 1000 peaks. Then we calculated the average ratio and the standard deviation for each of the 15 states over all samples in each cell type. For an explicit description, let us introduce a set of functional annotations, , where
is the set of regions on the genome, each of which corresponds to functional annotation y. We want to know how many peaks, k, of every 1000 peaks belong to each functional annotation y. For this purpose, we define
where gk = (γk, αk, βk) is the peak position. We computed
for
, as shown in Fig 12. Note that we used the position of the peak center, (αk + βk)/2, to annotate biological function.
Percentage () of functional annotations in every 1000 peaks for B cells (A), Mono (B), and HSC (C). Only the functional annotations that have maximum percentages ≥ 12%, y ∈ {FlankingActiveTSS, ActiveTSS, Enhancers, Quiescent_Low}, are shown.
As shown in Fig 12, most of the peaks with higher rankings belonged to “Active TSS”, which was related to the promoters of active genes, but as the rank went down, the ratio of peaks from enhancer regions started to increase. As the rank went down further, the ratio of peaks from “quiescent-low” regions started to increase. The ratio of peaks from promoters and enhancers crossed at around peak rank 10000 and the ratio of peaks from enhancers and “quiescent-low” regions crossed at around peak rank 60000. Therefore, we concluded that the number around the 64000th peak is strongly related to the point that the contribution of “quiescent-low” regions to the Hamming distances exceeds the contribution of enhancer regions to the Hamming distances.
Note that the type penalty of HSC under the condition was not as good as that of B cells or Mono, and the functional annotation result of HSC did not show clear behaviors compared with B cells and Mono (Fig 12C), which may partially explain the worse type penalty of HSC (Fig 10B).
Variations of hierarchical clustering methods
In general, when one performs data clustering, the effect of variations of the clustering algorithms and the effect of loss of data on the clustering output should be considered.
First we considered the dependence of the clustering results on the variations of the clustering algorithms. Besides Ward’ method which we used until here, there are several hierarchical clustering methods including UPGMA (Unweighted Pair Group Method with Arithmetic mean), WPGMA (Weighted Pair Group Method with Arithmetic Mean), UPGMC (Centroid Clustering or Unweighted Pair Group Method with Centroid Averaging), and WPGMC (Median Clustering or Weighted Pair Group Method with Centroid Averaging). We performed optimization also with UPGMA, as shown in Fig 13, and found that the minimum value of the penalty is 36 with Mcut = 12000. The other methods give worse results in general. Specifically, the minimum values of the penalty we found were 59 for WPGMA with Mcut = 20000, 127 for UPGMC with Mcut = 30000, and 149 for WPGMC with Mcut = 35000. These results suggested that Ward’s method giving 18 as the minimum value of the penalty was a better choice than that of the other methods for our purpose.
The distribution of global penalty λ (A) and type penalty λν for each cell type ν (B) along with Mcut with parameter pG = 10−2 by using UPGMA.
Robustness of our best clustering against the loss of data
Regarding the loss of data, let us consider making new reads data from original data R. Specifically, we set r with 0 ≤ r ≤ 1 as the probability of randomly removing ⌈rNr⌉ reads from R with the uniform distribution, where ⌈χ⌉ means the minimum integer larger than or equal to χ. Thus we can obtain
, where
is one read in R. Using this procedure, we computed λ for
. As shown in Fig 14B, when ratio r was increased, the value of λ was constant until r = 0.007 and gradually increased thereafter. In the region r ≥ 0.7, λ increased dramatically. Note that r = 0 gave λ = 18 and the highest possible value of λ for 77 samples is 924. Thus, we concluded that for small r, the average penalty tended to be stably close to that of r = 0.
The effect of the loss of reads on the global penalty λ. Reads were removed randomly from the uniform distribution with probability r. Then global penalty λ was calculated with parameter Mcut = 30000 (A), Mcut = 64000 (B) or Mcut = 80000 (C). Each circle indicates one sample and each square indicates the average over samples at the same r value.
Further, we investigated λ for different values of Mcut than 64000 to check the robustness of against random selections. Specifically, we investigated the behavior of λ by varying r for Mcut = 30000 and 80000 with
. The minimum value of λ as a function of r was 27 for Mcut = 30000 and located at r = 0 (Fig 14A) and was 38 for Mcut = 80000 and again located at r = 0 (Fig 14C). Note that in the region r ≥ 0.08, λ for Mcut = 30000 was smaller than λ for Mcut = 64000, which suggested that
becomes less than 64000 when the data size is decreased.
Thus, for the present data size, we concluded that our algorithm was stable against small losses of the data and it could also work well by adjusting Mcut for losses of data up to 50 percent. The obtained results imply that when the given data size is increased, our algorithm becomes more stable or potentially achieves better clustering with a smaller penalty than our current best clustering.
Application to leukemic cells
To evaluate the practicality of our algorithm with the optimized parameters on cancer research, we analyzed three types of leukemia: CLL, AML, and ATL, by calculating Ward’s distance function,
, between a given leukemia sample ζ and all samples
of cell type ν. (See Materials and methods for details of
).
To separate normal and leukemic cells effectively, information about the cell surface markers was used. CLL is a disease that is characterized by the clonal proliferation of malignant B lymphocytes. Leukemic cells from CLL patients were purified by using the cell surface markers CD5 and CD19, which are commonly used as markers for CLL (Table 2) [18].
Immunophenotype of CLL [8]: Note that B cells are CD19+, as shown in Table 1. Immunophenotype of AML [7]: SSC-high means that the intensity of side scatter in the flow cytometry is high. Note that HSC, MPP, and LMPP are Lin-, CD34+, CD38- as shown in Table 1. Immunophenotype of ATL [24, 25]: Note that CD4+T cells are CD4+, as shown in Table 1.
The AML samples analyzed in this study were divided into three stages, preleukemic HSC (pHSC), leukemia stem cells (LSC), and AML blasts by cell surface markers according to [7] (Table 2). Briefly summarizing these three types, HSC that acquired founder mutations become pHSC, which expand to generate preleukemic clones. The subsequent acquisition of progressor mutations creates LSC, which can self-renew and produce AML blasts [19]. It has been reported that mature LSC populations more closely resemble normal GMP, and immature LSC populations are functionally similar to LMPP [20]. A recent study has revealed that CD99-positive cells are almost entirely composed of LMPP-like cells in the sense of Ref. [21]. Thus, the LSC used in our study, which are CD99-positive, can be presumed to be LMPP-like LSC.
Human T-cell leukemia virus type 1 (HTLV-1) is a causative agent of ATL and HTLV-1-associated myelopathy/tropical spastic paraparesis (HAM/TSP) [22]. ATL has been subclassified into four clinical subtypes: acute, lymphoma, chronic, and smoldering. The chronic and smoldering subtypes are considered indolent, while patients with the acute or lymphoma subtype generally have a poor prognosis. HTLV-1 can infect a variety of cell types, but more than 90% of infected cells are CD4+ memory T cells in vivo [23]. In order to specifically separate HTLV-1-infected cells from other normal T-cells, Cell adhesion molecule 1 (CADM1/TSLC1) is used because of its sensitivity and specificity [24, 25]. Thus, in this study, to purify leukemic cells (HTLV-1 infected cells) from the peripheral blood mononuclear cells (PBMC) of ATL patients, we used the cell surface markers shown in Table 2.
The objective of our analysis using leukemic samples was to evaluate which type of hematopoietic cell is closest to a given leukemic sample at the chromatin level. Specifically, we added the ATAC-seq data of a leukemic sample to healthy hematopoietic ATAC-seq data and calculated the Hamming distances where is used. We computed
as the distance between cell type
and leukemic sample ζ; in this case, sample ζ was extracted from one patient.
We define the q-th closest cell type of sample ζ as type to provide the qth minimum of
in terms of ν. Using this quantity, we define the rank gap between a given reference type
and sample ζ as
such that
. In particular, we call
the closest type of sample ζ. Note that rank gap
holds when
. Thus, we not only revealed the closest cell type, but also identified the second, third, and so on closest cell type, and quantified the difference between the characterization results of our algorithm and a given type as the “rank gap”.
As shown in Table 3, by calculating the Hamming distance between each CLL sample and a set of hematopoietic cells, we found that the closest cell type for all CLL samples was B cells, which coincides well with the characteristics of CLL cell surface markers. This result led us to conjecture that our method could infer the cell type of a given leukemic cell characterized by immunophenotypes with using only its ATAC-seq data.
“Closest cell type” computed by our algorithm.
In order to assess the applicability of our method to leukemia whose cell of origin is not uniform and has high levels of heterogeneity between cases, we analyzed AML samples [7]. We found that the results of our analysis for pHSC and Blast had substantial overlap with those of a previous study [7], where 12 out of 16 samples for pHSC and 13 out of 18 samples for Blast are overlapped, as shown in Table 4. However, in the case of LSC, we found differences between the results of our analysis and those from [7]. Most of the LSC samples were closest to LMPP using our algorithm, but to GMP in [7]. As mentioned above, the LSC used in the present study were CD99-positive and are presumed to be composed of LMPP-like cells, which suggests that our characterization by using information of the Hamming distance infers the cell type with high accuracy, though further investigation is required.
Comparison between the “closest normal cell” in Fig 6i of [7] and “closest cell type” computed by our algorithm. The second, the third, and …-th “closest type” were also identified by our algorithm. The “rank gap” represents the difference of the result between the two analytical methods. For example, the “closest normal cell” of sample SU351-pHSC is MPP in [7], but is LMPP by our algorithm. “MPP” was the second “closest cell type”. Thus, the rank gap was calculated as 2-1 (= 1). If the results from the two analytical methods coincide with each other, the rank gap is 0.
Finally we analyzed ATL samples (See Materials and methods for details of sample preparation). When we calculated the Hamming distance between each ATL sample and a set of hematopoietic cells, we found that the closest cell type for two ATL samples was Mono (hereafter we term these samples “Mono-like ATL”), while that of the other samples was CD4+T, as shown in Table 5. Surprisingly, the two Mono-like ATL samples were categorized into chronic-type ATL. Since CD14 is the marker of Mono (Table 1), we investigated the CD14 gene expression pattern in CD4+T, Mono and ATL samples. Particularly, we calculated the ratio of the CD14 reads count to the CD4 reads count from RNA-seq data and found that the two Mono-like ATL samples exhibited higher values among all ATL samples (Fig 15). In this way, the obtained results led us to conjecture that our algorithm could infer the cell phenotype, potentially including clinical subtypes, only using ATAC-seq data. However, we need to analyze more samples to validate this conclusion.
The reads count of CD14 over the reads count of CD4 from RNA-seq data of CD4+T, Mono, and ATL samples.
Clinical subtypes of ATL samples and “closest cell type” computed by our algorithm.
Discussion
In this paper, we presented a new algorithm to systematically perform clustering of epigenomic data using the Hamming distance, which enabled us to find optimal parameters of the data reduction toward cell-type classification. This algorithm has one clear advantage in terms of computational cost compared to a previous method using targeted regions merged over samples [7]. Especially, when adding new samples to the analysis, we only have to calculate the distances between newly appearing pairs of samples and not between preexisting samples. The computational cost of the presented systematic algorithm is significantly lower for this situation compared to the previous method with merging targeted regions. Furthermore, this algorithm was found to effectively detect the closest cell type of a leukemic sample, with the results being broadly consistent with the characterization of leukemic samples by cell surface markers or RNA-seq. Thus, the developed algorithm potentially serves as a screening for the phenotype of a leukemia sample by using the ATAC-seq data of the sample as input.
As a next step, we need to investigate if our constructed algorithm is robust for other existing methods and data. For example, for the same data of hematopoietic cells, we replaced the Hamming distance with the Dice coefficient, which has been used in the CODEX project [26] to quantify the differences between two samples, but found the results with pG = 10−2 were not improved in terms of the penalty. We also compared our algorithm with DiffBind [27], which is commonly used as a ChIP-seq differential analysis tool, but again found that DiffBind with its default setting did not give a better clustering result. Note that there are other existing methods and data to be checked in the future.
A unique point of our constructed algorithm is that we only used ATAC-seq data without gene expression data. Our analysis suggests that ATAC-seq data itself contains enough information to determine cell types even in the absence of regional annotation data such as promoters or enhancers. This feature implies that our algorithm reveals elusive epigenomic properties that significantly affect the phenotype of cell types. Another advantage of our algorithm is that we do not assume a strong property for the statistics of the reads data, which is otherwise implicitly assumed when quantile normalization is performed. Instead of using the strong assumption, we took a data-driven approach for the normalization of the reads data, where we pre-analyzed the statistics of the reads data before performing any normalization.
Finally, our algorithm could extend its application to leukemic samples whose properties are uncertain. We also expect that our whole approach with slight modifications will be applicable to other epigenetic sequencing data such as ChIP-seq and bisulfite sequencing available, for example, from The International Human Epigenome Consortium (https://epigenomesportal.ca/ihec/), ROADMAP Epigenomics (http://www.roadmapepigenomics.org/) and many other resources, whose target regions for the analysis are not uniform between samples.
Materials and methods
Ethics statement
Experiments using clinical samples were conducted according to the principles expressed in the Declaration of Helsinki and approved by the Institutional Review Board of Kyoto University (permit numbers G310 and G204). ATL patients provided written informed consent for the collection of samples and subsequent analysis.
Sequencing sample preparation
ATL patient PBMCs were thawed and washed with PBS containing 0.1% BSA. To discriminate dead cells, we used the LIVE/DEAD Fixable Dead Cell Stain Kit (Invitrogen). For cell surface staining, cells were stained with APC anti-human CD4 (clone: RPA-T4) (BioLegend) and anti-SynCAM (TSLC1/CADM1) mAb-FITC (MBL) antibodies for 30 minutes at 4°C followed by a wash with PBS. HTLV-1 infected cells (CADM1+ and CD4+) were sort-purified with FACS Canto (Beckman Coulter) to reach 98–99% purity. Data was analyzed by FlowJo software (Treestar). Soon after the sorting, 10000-50000 HTLV-1 infected cells were centrifuged and used for ATAC-seq as previously described [5]. Total RNA was isolated from the remaining cells using the RNeasy Mini Kit (Qiagen). Library preparation and high-throughput sequencing were performed by Macrogen Inc. (Seoul, Korea). The diagnostic criteria and classification of clinical subtypes of ATL were performed as previously described [28]. 77 ATAC-seq datasets from 13 human primary blood cell types and datasets from 42 AML patients were obtained from the Gene Expression Omnibus (GEO) with accession number GSE74912 [7]. ATAC-seq datasets from 7 CLL patients were obtained from GSE111015 [18] and RNA-seq datasets of CD4+T and Mono cells were obtained from GSE74246 [7].
Sequencing data analysis
ATAC-seq reads were aligned using BWA version 0.7.16a [29] with default parameters. SAMtools [30] was used to convert SAM files to compressed BAM files and sort the BAM files by chromosome coordinates. PICARD software (v1.119) (http://broadinstitute.github.io/picard/) was then used to remove PCR duplicates using the MarkDuplicates options. Reads with mapping quality scores less than 30 were removed from the BAM files. For peak calling, MACS2 (v2.1.2) software was used [15]. RNA-seq data were aligned to human reference genome hg19 using STAR 2.6.0c [31] with the --quantMode GeneCounts function. Normalization was not performed, and only raw reads count data of CD14 and CD4 were used in this study.
Principles of data reduction
When we analyze preprocessed ATAC-seq data with , we have to care for biases caused by the fact that the amount of reads, Nr, depends on the setting of the sample preparation and on the sequencers used. (See S1 Appendix for the explicit construction of
.) Normalization is done to remove such biases.
A conventional way to perform normalization is to use quantile normalization, where the distribution of the reads number on certain regions in the DNA is assumed to be the same for all samples [12, 13]. However, there is no strong reason to support this assumption, particularly for sample sets of different cell types. Furthermore, under this assumption, there is a risk that we overlook important differences between different cell types. Therefore, in this paper, we do not assume this property.
An alternative way to perform normalization is to reduce the data into a simple binary value hγ,x ∈{0, 1} on each genomic position (γ, x), where hγ,x depends on the data size Nr as little as possible. For example, one could determine the state of hγ,x = 1 and hγ,x = 0 as an “open” and “closed” chromatin status, respectively, on genomic position (γ, x).
In this direction, our ultimate purpose is to look for the “best” principle that determines two states for hγ,x, by which a set of samples including different cell types are completely classified into groups of the same cell type. We use no information about cell types when determining the value of hγ,x, because we would like to have an algorithm that can be applied without knowing the cell types.
Peak-calling with ranking
Currently we do not have the best solution to properly determine two effective states for hγ,x. As a candidate to approach the best solution, we use the MACS2 algorithm, which was originally invented to analyze ChIP-seq data [15] but is now widely used to estimate the location of open chromatin regions from ATAC-seq data [32, 33].
We would like to find the set of position (γ, x) where the number of reads overlapping with position (γ, x), , is relatively high in the neighborhood (γ, x). The MACS2 algorithm is likely to detect those positions from the data of the reads described by
. In our calculation, we use the MACS2 (v2.1.2) callpeak command with option “--nomodel --nolambda --keep-dup all -p pG”, where we need to set parameter pG as a parameter of peak inference (for details, see [15]).
By applying MACS2 to the input ATAC-seq data, we obtain the following output data structure:
- The label
of the chromosome to which the k-th peak has a start position 1 ≤ αk ≤ Lγ and end position 1 ≤ βk ≤ Lγ for 1 ≤ k ≤ M (here M is the number of peaks). We call gk = (γk, αk, βk) the k-th peak region.
- For each gk, p-value pk with pk ≤ pG is associated to the k-th peak. Note that MACS2 outputs log10(1/pk) = −log10 pk instead of pk.
and Lγ are the set of all chromosomes and the length of chromosome γ, respectively (see S1 Appendix for details of the notations). We define A as
By reordering the terms of k, we can set pk ≤ pk′ for any k < k′ without loss of information.
In Fig 2, we show the distribution of the peak width |βk − αk| versus ranking k. Note that gk with high pk could be affected significantly by the conditions of the experiments including sequencing, because the data above rank value 40000 unnaturally touches the value of the lower limit of width 200, which is predetermined by the MACS2 algorithm. Thus, there is a possibility that peaks with higher p-values could strongly depend on both the inference algorithm and the number of reads Nr. Those peaks would presumably not contribute to the detection of cell phenotypes. This observation suggests we should remove peaks with higher p-values as mentioned in Results.
Parameterized binarization by cutting off low-ranked peaks
Next we reconsidered how to alleviate biases in the data by introducing threshold number Mcut, such that
which leads to the removal of
as a candidate for the normalization of the ATAC-seq data. Note that
. Then, by using
, we may introduce a binary sequence
such that hγ,x = 1 if there is k satisfying αk ≤ x ≤ βk with
; otherwise hγ,x = 0 as shown in Fig 4.
pG and Mcut can be regarded as parameters for determining the value of hγ,x within the MACS2 algorithm and what part of the data is taken into account, respectively. Thus, our task under the principle above turns out to be how to determine a proper set of (Mcut,pG) for the cell-type classification.
Hamming distance
The Hamming distance is often used to compare two binary sequences in information theory (see Section 13 in [14]) and is equal to the number of positions on which two symbols have different values. See Fig 3 for an illustrative explanation.
The Hamming distance between two binary sequences and
with
is defined as
where we define
Algorithm of hierarchical clustering
In this and the next subsection, we recall algorithms for agglomerative hierarchical clusterings and drawing dendrograms. We use two methods, UPGMA and Ward’s. Though they are described in many textbooks (for example, see Chapter 4 in [34]), we need the description in order to define the global penalty and the type penalty. Our description of the algorithms follows [16].
To describe the algorithms, we define two distance functions between two subsets, as follows (for inductive definitions and other distance functions, see Section 4.2 in [34]). One distance function,
comes from the UPGMA method and is defined as the average of all the distances between samples in C1 and C2. Equivalently, we define
If C1 or C2 is empty, we set
.
Another choice of the distance function, , comes from Ward’s method and is defined as
where we define
Again, if C1 or C2 is empty, we set
.
In the following, we fix as
or
. We sometimes identify sample
and subset {c} of single element c. For example, we write
for
. Note that
where K = 1 for
and K = 2−1/2 for
by definition.
We define a cluster as subset of
with a specified order of elements. Hierarchical clustering is an algorithm that can construct set
of clusters and order the elements in
to draw dendrograms.
- We set
for 1 ≤ τ ≤ Ns. We do not consider the order of the elements in
because they are sets of a single element.
- We define the list of uncombined clusters as
and set the historical list of clusters as
.
- At the t-th step (1 ≤ t ≤ Ns − 1), we define
and
inductively.
- (a). We look up the pair
and
with τ′ < τ″ in
such that their distance is a minimum; that is,
Note that 1 ≤ τ′ < τ″ < t + Ns by construction. We consider only the case when the pair is uniquely determined.
- (b). We define a new cluster
. If the elements of
are ordered as c1, c2, …, cz and the elements of
are
, then the elements of
are ordered as
- (c). We define
If t < Ns − 1, go to the (t + 1)-th step.
- (a). We look up the pair
We can easily see that if we do not consider the ordering, then we have as a set. Thus we finally obtain a list of 2Ns − 1 clusters
and an ordering of all elements of
from
.
How to draw dendrograms
The (rooted) dendrogram displays how our clustering combines pairs of clusters and the distance of the pairs. In the following, we explain an algorithm that introduces new symbols. For details, see [16].
- If sample
appears in the ordering of
as the aτ-th element, then we associate point nτ = (aτ, 0) in two-dimensional coordinate space to cluster
. We call point nτ the leaf, which corresponds to
.
- For 1 ≤ t ≤ Ns − 1, we inductively associate point
to cluster
. If
is constructed as the union of
and
with 1 ≤ τ′ < τ″ < t + Ns, we associate to
the node
Note that
and
are uniquely determined. We call
the node associated to the (t + Ns)-th cluster
.
- We connect
with nτ′ and nτ″.
Since each node or leaf n corresponds to cluster , we can define the offspring set
of n as set
without ordering. Graphically, the offspring set of node n is the set of samples corresponding to leaves branched from node n, as displayed in Fig 7. This intuitional explanation is justified, since the y-coordinate of the “mother node”
is larger than or equal to those of the “child nodes” nτ′, nτ″ if we use Ward’s method or UPGMA. Note that there are many choices to draw dendrograms; for example, at any branching node, we can exchange two branches without any essential change in the data structure.
Global penalty as a cost function
In this section, we discuss the global penalty, a quantity that measures how the obtained hierarchical clustering differs from our knowledge of cell type classifications. We also give examples displaying the computation of the penalties and extreme situations that represent the theoretical bounds of the penalties. Note that these examples are just for explanation and not obtained from actual data.
In our settings, each sample is previously classified by types. Explicitly, set consists of thirteen types:
For each type
, we denote the set of samples classified to type ν as
. This set could be empty, though it is not in our case. For every pair ν, ν′ of distinct types, there are no common elements in
and
, and the union of
among all types
coincides with
. Equivalently,
For a given hierarchical clustering constructed in the manner of the previous section, the type penalty for type ν is the quantity λν defined as follows. If is empty, we set λν = 0. Otherwise, since the cluster grows step by step, there is the minimum τ for 1 ≤ τ ≤ 2Ns − 1 such that
. We denote the minimum τ by τ(ν). Then we define λν as the number of elements in
that are not of type ν. In other words, we set
Since
includes all elements of type ν, we find λν ≥ 0. Also since
is a subset of
, we find
. Thus we have
(See Fig 7 for an explanation of type penalties).
For a given hierarchical clustering, the global penalty λ is defined to be the total sum of type penalties,
λ is bounded as
(1)
In our case, since
and
, we have 0 ≤ λ ≤ (13 − 1) ⋅ 77 = 924. Note that for a certain class of trees, these upper and lower bounds are not achieved. Fig 16 displays examples of the upper and lower bounds.
Note that this dendrogram is constructed using artificial data to explain how to calculate the penalty, though we use the same labels such as Mono1. Both of these dendrograms have six leaves () that are classified into three types (in these examples,
). (A) This example gives the lowest global penalty 0. (B) In this example, we have
. Since the corresponding cluster
is the whole set
, the local penalty is 6 − 2 = 4 for each type, and the global penalty is 4 × 3 = 12. This result gives the upper bound
in Eq (1).
Further, we write λ(Mcut, pG) as λ to point out that λ depends on (Mcut, pG). Note that is equal to
, which was defined in the previous section.
Supporting information
S1 Appendix. Additional details of sequencing analysis.
https://doi.org/10.1371/journal.pcbi.1008422.s001
(PDF)
Acknowledgments
The authors thank P. Karagiannis for valuable comments and proofreading of this manuscript. They also thank MACS Program at Graduate School of Science Kyoto University which allowed this collaboration to be carried out.
References
- 1. Klemm SL, Shipony Z, Greenleaf WJ. Chromatin accessibility and the regulatory epigenome. Nature Reviews Genetics. 2019;20:207–220.
- 2. Gaspar-Maia A, Alajem A, Meshorer E, Ramalho-Santos M. Open chromatin in pluripotency and reprogramming. Nature Reviews Molecular Cell Biology. 2011;12:36–47.
- 3. John S, Sabo PJ, Thurman RE, Sung MH, Biddie SC, Johnson TA, et al. Chromatin accessibility pre-determines glucocorticoid receptor binding patterns. Nature Genetics. 2011;43:264–268. pmid:21258342
- 4. Thurman RE, Rynes E, Humbert R, Vierstra J, Maurano MT, Haugen E, et al. The accessible chromatin landscape of the human genome. Nature. 2012;489:75–82. pmid:22955617
- 5. Buenrostro JD, Giresi PG, Zaba LC, Chang HY, Greenleaf WJ. Transposition of native chromatin for fast and sensitive epigenomic profiling of open chromatin, DNA-binding proteins and nucleosome position. Nature Methods. 2013;10:1213–1218.
- 6. Buenrostro JD, Wu B, Litzenburger UM, Ruff D, Gonzales ML, Snyder MP, et al. Single-cell chromatin accessibility reveals principles of regulatory variation. Nature. 2015;523:486–490. pmid:26083756
- 7. Corces MR, Buenrostro JD, Wu B, Greenside PG, Chan SM, Koenig JL, et al. Lineage-specific and single-cell chromatin accessibility charts human hematopoiesis and leukemia evolution. Natature Genetics. 2016;48:1193–1203. pmid:27526324
- 8. Rendeiro AF, Schmidl C, Strefford JC, Walewska R, Davis Z, Farlik M, et al. Chromatin accessibility maps of chronic lymphocytic leukaemia identify subtype-specific epigenome signatures and transcription regulatory networks. Nature Communications. 2016; 7:11938. pmid:27346425
- 9. Qu K, Zaba LC, Satpathy AT, Giresi PG, Li R, Jin Y, et al. Chromatin Accessibility Landscape of Cutaneous T Cell Lymphoma and Dynamic Response to HDAC Inhibitors. Cancer Cell. 2017; 32:27–41.e4. pmid:28625481
- 10. Tu S and Shao Z. An introduction to computational tools for differential binding analysis with ChIP-seq data. Quantitative Biology. 2017; 5(3):226–235.
- 11. Yan F, Powell DR, Curtis DJ, Wong NC. From reads to insight: a hitchhiker’s guide to ATAC-seq data analysis. Genome Biology. 2020; 21(1):22.
- 12. Meyer SU, Pfaffl MW, Ulbrich SE. Normalization strategies for microRNA profiling experiments: A’normal’ way to a hidden layer of complexity? Biotechnology Letters. 2010; 32:1777–1788.
- 13. Hicks SC and Irizarry RA. quantro: A data-driven approach to guide the choice of an appropriate normalization method. Genome Biology. 2015; 16:1–8.
- 14.
MacKay DJC. Information theory, inference and learning algorithms. Cambridge University Press; 2003
- 15. Zhang Y, Liu T, Meyer CA, Eeckhoute J, Johnson DS, Bernstein BE, et al. Model-based Analysis of ChIP-Seq (MACS). Genome Biology. 2008; 9:R137. pmid:18798982
- 16.
Müllner D. (2011). Modern hierarchical, agglomerative clustering algorithms. arXiv:1109.2378. [Preprint]. 2011 Available from: https://arxiv.org/abs/1109.2378
- 17. Ernst J. and Kellis M. Chromhmm: automating chromatin-state discovery and characterization. Nature Methods. 2012; 9:215–216.
- 18. Rendeiro AF, Krausgruber T, Fortelny N, Zhao F, Penz T, Farlik M, et al. Chromatin mapping and single-cell immune profiling define the temporal dynamics of ibrutinib response in CLL. Nature Communications. 2020; 11:1–14. pmid:31996669
- 19. Döhner H, Weisdorf DJ, Bloomfield CD. Acute myeloid leukemia. The New England Journal of Medicine. 2015; 373:1136–1152.
- 20. Goardon N, Marchi E, Atzberger A, Quek L, Schuh A, Soneji S, et al. Coexistence of LMPP-like and GMP-like leukemia stem cells in acute myeloid leukemia. Cancer Cell. 2011; 19:138–152. pmid:21251617
- 21. Chung SS, Eng WS, Hu W, Khalaj M, Garrett-Bakelman FE, Tavakkoli M, et al. Cd99 is a therapeutic target on disease stem cells in myeloid malignancies. Science Translational Medicine. 2017; 9(374). pmid:28123069
- 22. Matsuoka M and Jeang KT. Human T-cell leukaemia virus type 1 (HTLV-1) infectivity and cellular transformation. Nature Reviews Cancer. 2007; 7:270–280.
- 23. Richardson JH, Edwards AJ, Cruickshank JK, Rudge P, Dalgleish AG. In vivo cellular tropism of human T-cell leukemia virus type 1. Journal of Virology. 1990; 64:5682–5687.
- 24. Manivannan K, Rowan AG, Tanaka Y, Taylor GP, Bangham CR. CADM1/TSLC1 Identifies HTLV-1-Infected Cells and Determines Their Susceptibility to CTL-Mediated Lysis. PLoS Pathogens. 2016; 12:1–18.
- 25. Nakahata S, Saito Y, Marutsuka K, Hidaka T, Maeda K, Hatakeyama K, et al. Clinical significance of CADM1/TSLC1/IgSF4 expression in adult T-cell leukemia/lymphoma. Leukemia. 2012; 26:1238–1246. pmid:22289924
- 26. Sánchez-Castillo M, Ruau D, Wilkinson AC, Ng FS, Hannah R, Diamanti E, et al. CODEX: A next-generation sequencing experiment database for the haematopoietic and embryonic stem cell communities. Nucleic Acids Research. 2015; 43(D1):D1117–D1123. pmid:25270877
- 27. Ross-Innes CS, Stark R, Teschendorff AE, Holmes KA, Ali HR, Dunning MJ, et al. Differential oestrogen receptor binding is associated with clinical outcome in breast cancer. Nature. 2012; 481(7381):389–393. pmid:22217937
- 28. Shimoyama M. Diagnostic criteria and classification of clinical subtypes of adult t-cell leukaemia-lymphoma. a report from the lymphoma study group (1984-87). British Journal of Hematology. 1991; 3:428–437.
- 29.
Li H. Aligning sequence reads, clone sequences and assembly contigs with bwa-mem. arXiv:1303.3997. [Preprint]. 2013 Available from: https://arxiv.org/abs/1303.3997?upload=1
- 30. Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics. 2009; 25:2078–207. pmid:19505943
- 31. Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, et al. STAR: Ultrafast universal RNA-seq aligner. Bioinformatics. 2013; 29:15–21. pmid:23104886
- 32. Corces MR, Granja JM, Shams S, Louie BH, Seoane JA, Zhou W, et al. The chromatin accessibility landscape of primary human cancers. Science. 2018; 362:1–58. pmid:30361341
- 33. Denny SK, Yang D, Chuang CH, Brady JJ, Lim JS, Grüner BM, et al. Nfib Promotes Metastasis through a Widespread Increase in Chromatin Accessibility. Cell. 2016; 166:328–342. pmid:27374332
- 34.
Everitt BS, Landau S, Leese M, Stahl D. Cluster Analysis, 5th Edition. Wiley Series in Probability and Statistics. John Wiley & Sons, Ltd.,; 2011