Bioinformatics, 36(12), 2020, 3927–3929

doi: 10.1093/bioinformatics/btaa205
Advance Access Publication Date: 27 March 2020
Applications Note

Downloaded from https://academic.oup.com/bioinformatics/article-abstract/36/12/3927/5812779 by University Libraries | Virginia Tech user on 27 June 2020
Data and text mining
debCAM: a bioconductor R package for fully
unsupervised deconvolution of complex tissues
Lulu Chen1, Chiung-Ting Wu1, Niya Wang2, David M. Herrington3, Robert Clarke4 and
Yue Wang1,*
1
Department of Electrical and Computer Engineering, Virginia Polytechnic Institute and State University, Arlington, VA 22203, USA,
2
Search Ranking Unit, Google LLC, Mountain View, CA 94043, USA, 3Department of Internal Medicine, Wake Forest University,
Winston-Salem, NC 27157, USA and 4Department of Oncology, Lombardi Comprehensive Cancer Center, Georgetown University,
Washington, DC 20057, USA
*To whom correspondence should be addressed.
Associate Editor: Jonathan Wren
Received on October 18, 2019; revised on March 5, 2020; editorial decision on March 19, 2020; accepted on March 23, 2020

Abstract
Summary: We develop a fully unsupervised deconvolution method to dissect complex tissues into molecularly dis-
tinctive tissue or cell subtypes based on bulk expression profiles. We implement an R package, deconvolution by
Convex Analysis of Mixtures (debCAM) that can automatically detect tissue/cell-specific markers, determine the
number of constituent subtypes, calculate subtype proportions in individual samples and estimate tissue/cell-
specific expression profiles. We demonstrate the performance and biomedical utility of debCAM on gene expres-
sion, methylation, proteomics and imaging data. With enhanced data preprocessing and prior knowledge incorpor-
ation, debCAM software tool will allow biologists to perform a more comprehensive and unbiased characterization
of tissue remodeling in many biomedical contexts.
Availability and implementation: http://bioconductor.org/packages/debCAM.
Contact: yuewang@vt.edu
Supplementary information: Supplementary data are available at Bioinformatics online.

1 Introduction We have demonstrated real biomedical utilities of debCAM tool on

Tissue heterogeneity serves as both an underexploited information
source in characterizing complex tissue phenotypes and a major con-
founding factor in studying individual tissue or cell subtypes.
Having analytic tools to define the molecular landscape of tissue
heterogeneity, and to determine how subtypes are remodeled with
phenotypic transitions, will be essential for the next step in systems
biology research. We have developed a fully unsupervised deconvo-
lution method, namely, deconvolution by Convex Analysis of
Mixtures (debCAM), to dissect complex tissues into molecularly dis-
tinctive tissue or cell subtypes based on bulk expression profiles
(Avila Cobos et al., 2018; Chan et al., 2008; Wang et al., 2016).
Importantly, debCAM requires no a priori information on the num-
ber, signatures or compositions of the subtypes present in heteroge-
neous samples, and does not require pure subtype references.
Supported by a well-grounded mathematical framework (Chan
et al., 2008; Wang et al., 2016), debCAM automatically detects
tissue/cell-specific markers, determines the number of constituent
subtypes, calculates subtype proportions in individual samples and
estimates tissue/cell-specific expression profiles.
The debCAM R package implements and tests the latest func-
tionalities of the debCAM algorithm pipeline in the literature.
gene expression (Wang et al., 2016), proteomics (Herrington et al.,
2018), imaging (Chen et al., 2011) and methylation data. These
applications have led to novel findings and hypotheses. The present
software also enhances data preprocessing, accelerates robust scatter
simplex identification and integrates supervising information, such
as known markers.
2 Description
2.1 Methodology and software
Ideally, a molecularly distinctive subtype would be composed of in-
dividual molecular features that are exclusively expressed in the cog-
nate cell or tissue subtype of interest while in no others—so-called
molecular markers. Fundamental to the success of debCAM is the
newly proven mathematical theorems (Chan et al., 2008; Chen
et al., 2011; Wang et al., 2016), showing that the scatter simplex of
mixed expressions is a rotated and compressed version of the scatter
simplex of pure subtype expressions, where the molecular markers
are located at each of the vertices (Fig. 1A–E). debCAM works by
geometrically identifying the vertices (and their resident markers) of
the scatter simplex of globally measured expressions. Tissue samples

C The Author(s) 2020. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com
V 3927
3928 L.Chen et al.

Downloaded from https://academic.oup.com/bioinformatics/article-abstract/36/12/3927/5812779 by University Libraries | Virginia Tech user on 27 June 2020
Fig. 1. debCAM workflow and case studies. (A) The implemented debCAM pipeline with four major functional modules: data preprocessing, simplex and marker detection,
deconvolution of mixed expression profiles and model selection (Supplementary Material). Matrix ‘A’ is the mixing proportions whose column vectors correspond to the verti-
ces of the scatter simplex, matrix ‘S’ contains subtype-specific expression profiles and K is the number of subtypes present; all are unknown and need to be estimated. (B)
Scatter simplex of 12 in vitro mixtures shows six vertices, hosting the subtype-specific markers (detected by debCAM). (C) A priori known markers (five immune cell types)
superimposed onto the scatter simplex are closely resided around the corresponding vertices. (D and E) Scatter simplex of purified neuron and glia subtypes (astrocytes and ma-
ture oligodendrocytes) shows clearly three distinctive vertices. As expectedly, with varying mixing proportions, the scatter simplex of bulk tissue samples shows a rotated and
compressed version of the original simplex. (F) Neuron-specific expression profiles obtained from purified samples and from bulk tissue samples are highly correlated over all
CpG and marker sites

to be deconvolved by debCAM contain unknown number and vary-
ing proportions of molecularly distinctive subtypes. Expression of a
given molecular feature (e.g. mRNA and protein) in a specific sub-
type is modeled as being linearly proportional to the abundance of
that subtype. The minimum description length (MDL) information
criterion determines the number of subtypes present (Chen et al.,
2011; Wang et al., 2016).
The debCAM software tool performs the following major ana-
lytics pipelines (Fig. 1A), in additional to normalization and antilog-
arithm: (i) data preprocessing. Molecule features whose expression
levels are lower or higher than pre-fixed threshold are removed as
noise or outliers. The number of mixture profiles is reduced by prin-
cipal component analysis. On the scatter simplex constructed by
perspective projection and local outlier factoring, molecule features
are aggregated into representative cluster centers using K-means or
Affinity Propagation Clustering. The interior cluster centers are
removed by QuickHull and a greedy/floating research algorithm. (ii)
Simplex and Marker detection. An exhaustive combinatorial search
is conducted to identify the optimal scatter simplex for a given num-
ber of vertices among the peripheral cluster centers, guided by a
convex-hull-to-data fitting criterion. The members of the identified
simplex vertex clusters are considered subtype-specific markers. (iii)
Deconvolution of mixed expression profiles. The mixing propor-
tions are first estimated by the collective expression levels of
subtype-specific markers, followed by non-negative least squares
estimation of subtype-specific expression profiles. (iv) Model selec-
tion. The optimal number of subtypes present in samples is deter-
mined by the MDL information criterion among the competing
models.
2.2 Case study with validation
We use three datasets of two omics types (mRNA and methylation)
and three validation schemes (gold standard, ground truth and
cross-validation) to assess the performance of debCAM R package.
We first tested debCAM on biologically mixed gene expression
profiles (GSE64385). The 12 samples represent the mixtures of five
immune cell subtypes and one cancer cell-line in various known pro-
portions. The optimal scatter simplex blindly identified by debCAM
is given in Figure 1B, showing six distinctive vertices. To validate
the accuracy of subtype-specific markers blindly detected by
debCAM, we superimpose the color-coded known markers in the
scatter simplex showing close proximity to the corresponding verti-
ces (Fig. 1C, the cancer cell-line-specific markers are unavailable for
comparison). More convincingly, the sample-wise subtype propor-
tions estimated by debCAM match the ground truth almost perfect-
ly, with correlation coefficient of 0.975–0.996.
We further assessed debCAM on DNA methylation data
(GSE41826) profiled from both purified (flow-sorted) neuron and
glia subtypes as well as bulk tissue samples from prefrontal cortex.
debCAM
Application of debCAM to both purified and bulk tissue samples
detects consistently three molecularly distinctive tissue subtypes,
whose cellular phenotypes are inferred as neuron, astrocytes and
mature oligodendrocytes according to their temporal cellular dy-
namics (Supplementary Fig. S6). The subtype-specific markers resid-
ing near the simplex vertices of bulk tissue samples are firstly
detected by debCAM and then cross-validated by the largely com-
mon markers residing around the simplex vertices of purified sub-
type samples (Fig. 1D and E, Supplementary Table S1).
Furthermore, the subtype-specific expression profiles estimated by
debCAM are highly correlated between purified and bulk samples,
achieving almost perfect correlation coefficient of 0.982–0.992 over-
all CpG sites and correlation coefficient of 0.884–0.962 over CpG
markers (Fig. 1F, Supplementary Fig. S4).
3 Discussion
The debCAM provides a completely unsupervised deconvolution
tool, complementary to numerous existing deconvolution methods
(Avila Cobos et al., 2018; Hart et al., 2015; Lobanova and
Lobanov, 2019; Newman et al., 2015). Moreover, the debCAM
software can readily perform semi-supervised deconvolution by
incorporating relevant a priori information such as known or
reference-derived subtype-specific markers. We expect that
debCAM method, with a Bioconductor R package, to be a very use-
ful software tool for unbiased molecular analysis of complex tissues
in their native environment. Though the case studies are illustrated
only in scatter space, debCAM is principally applicable to sample
space when significantly dense sampling is available, such as single-
cell profiling. Historically this technique was motivated by the need
to identify expressed marker genes for tissue/cell subtypes within a
mixed sample; however, the approach described here is applicable
to any molecular profiling technique for identifying the unique mol-
ecule marker or other features that are affiliated with distinct sub-
3929
types within the studied samples. As a fully unsupervised machine-
learning method, debCAM is expectedly not immune to hidden con-
founding factors, e.g. batch effect and collinearity (Supplementary
Figs S7 and S8).
Downloaded from https://academic.oup.com/bioinformatics/article-abstract/36/12/3927/5812779 by University Libraries | Virginia Tech user on 27 June 2020
Funding
This work was supported by the National Institutes of Health [HL111362-
05A1, HL133932, NS115658]; and the Department of Defence [W81XWH-
18-1-0723, BC171885P1].
Conflict of Interest: none declared.
References
Avila Cobos,F. et al. (2018) Computational deconvolution of transcriptomics
data from mixed cell populations. Bioinformatics, 34, 1969–1979.
Chan,T.-H. et al. (2008) A convex analysis framework for blind separation of
non-negative sources. IEEE Trans. Signal Process., 56, 5120–5134.
Chen,L. et al. (2011) Tissue-specific compartmental analysis for dynamic
contrast-enhanced MR imaging of complex tumors. IEEE Trans. Med.
Imaging, 30, 2044–2058.
Hart,Y. et al. (2015) Inferring biological tasks using Pareto analysis of
high-dimensional data. Nat. Methods, 12, 233–235.
Herrington,D.M. et al. (2018) Proteomic architecture of human coronary and
aortic atherosclerosis. Circulation, 137, 2741–2756.
Lobanova,E. and Lobanov,S. (2019) Efficient quantitative hyperspectral
image unmixing method for large-scale Raman micro-spectroscopy data
analysis. Anal. Chim. Acta, 1050, 32–43.
Newman,A.M. et al. (2015) Robust enumeration of cell subsets from tissue ex-
pression profiles. Nat. Methods, 12, 453–457.
Wang,N. et al. (2016) Mathematical modelling of transcriptional heterogen-
eity identifies novel markers and subpopulations in complex tissues. Sci.
Rep., 6, 18909.