Applying PySCMGroup to Breast Cancer Biomarkers Discovery

Mazid Abiodoun Osseni

1 a

, Prudencio Tossou

1,2 b

, Jacques Corbeil

1 c

and Franc¸ois Laviolette

1,3 d

Universit

e Laval, Qu

ebec, Canada

InVivo AI, Montr

eal, Canada

Mila, Montr

eal, Canada

Keywords:

Multi-omics, Breast Cancer, Interpretability, Penalty-weight, Pathway Guided Selection.

Abstract:

Background. The identiﬁcation of biomarkers associated with triple-negative breast cancer (TNBC) is still

an active area of research due to the complexity of ﬁnding robust biomarkers associated with the disease. Pre-

vious methods have attempted to tackle the problem from a mono-perspective view by analyzing each omics

individually in the search of biomarkers. The majority of these methods mainly focus on gene expression

analysis since their impact on the phenotype is easier to measure and possibly more direct. However, it is

common understanding that genes belong to pathways and tend to work together within various metabolic,

regulatory, and signalling pathways. Hence, in this work, we tackled the TNBC biomarker discovery prob-

lem as a multi-omic pathway-based problem by efﬁciently combining the biological knowledge from multiple

pathways using a novel machine learning algorithm. The proposed algorithm, called GroupSCM, is an exten-

sion of the Set Covering Machine (SCM) that incorporate the pathway features as priors.

Results. Although the GroupSCM performed similarly to the SCM, metric-wise, it helps identify new

biomarkers not previously found by the SCM. By leveraging the pathway priors, the GroupSCM was able

to uncover two miRNAs: hsa-mir-18a and hsa-mir-190b, already known to be associated with various cancers

including breast cancer and yet to be linked to the Triple-Negative Breast Cancer phenotype.

Conclusion. The addition of priors to the SCM leads to interpretable, complete and sparser models which

are easier to analyze in vivo settings. It also provides insight into the omics interaction by highlighting the

miRNAs and epigenome contribution to the prediction task.

Code Availability: The code is available at: https://github.com/dizam92/BRCA experiments and paper

1 INTRODUCTION

Based on their genes expression, breast cancer (BC)

cases can be sub-classiﬁed into ﬁve categories: lumi-

nal A, luminal B, HER2++, Basal like and Triple-

Negative Breast Cancer (TNBC) (Lehmann et al.,

2011). TNBC, characterized by the non-expression

of estrogen (ER), progesterone (PR), and HER2 re-

ceptors, represents 10-20% of all breast cancers and is

known to be the most aggressive form, i.e. most likely

to spread beyond the breast and recur post-treatment

(Weigelt et al., 2010). Unfortunately, TNBC still re-

mains poorly diagnosed, as clinical, microarray-based

studies and immunohistochemical proﬁling are often

inconclusive, due to its similarity with the basal-like

https://orcid.org/0000-0001-7358-7402

https://orcid.org/0000-0002-9841-9867

https://orcid.org/0000-0002-9973-2740

https://orcid.org/0000-0002-1937-2512

breast cancer type. Consequently, there is a need for

the discovery of additional biomarkers to improve the

clinical diagnostic and prediction prognosis between

TNBC and other breast cancer types.

Genetic studies of cancer diseases in general have

often focused on extracting information from as-

sociation analysis using only a few types of data.

For example, (Iorio et al., 2005) showed that, com-

pared with normal breast tissue, miRNAs (speciﬁ-

cally miR-125b, miR-145, miR-21 and miR-155) are

also aberrantly expressed in human Breast Cancer.

In the case of the Colorectal Cancer (CRC), (Lao

and Grady, 2011) put more emphasis on the fact that

the hyper-methylation of some CpGs site ahead of

certain genes (CXCL12) can promote the metastatic

behaviour of colon cancer cell lines. However, to

better understand the disease’s complex genetic ex-

planation and in order to provide robust biomark-

ers, it is becoming increasingly important to inte-

Osseni, M., Tossou, P., Corbeil, J. and Laviolette, F.

Applying PySCMGroup to Breast Cancer Biomarkers Discovery.

DOI: 10.5220/0010375500720082

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 3: BIOINFORMATICS, pages 72-82

ISBN: 978-989-758-490-9

grate and utilize the full scope of available omics

information, recorded from a wide range of exper-

imental modalities (Freedman et al., 2011), (Kar-

czewski and Snyder, 2018). Embracing this philos-

ophy, numerous studies combined omics to tackle

the biomarkers discovery problem. Using unbiased

multi-omics approaches, (Yuan et al., 2016) reported

that DNA topoisomerase 2 beta (TOP2B) shows a

male-biased difference in DNA methylation in Blad-

der Cancer. In the case of Calciﬁc Aortic Valve Dis-

ease (CAVD), (Heuschkel et al., 2020) demonstrated

that the multi-omics integration of overlapping pro-

teome/transcriptome molecules, with the miRNAs,

identiﬁed a CAVD protein–protein interaction net-

work. Meanwhile (Rappoport et al., 2020) introduced

MONET (Multi Omic clustering by Non-Exhaustive

Types) aimed at discovering modules of similar sam-

ples such that each module is allowed to have a

clustering structure for only a subset of the omics.

This is a multi-omics clustering approach which dif-

fers from the other algorithms by not assuming a

common structure across all omics data. (Delogu

et al., 2020) proposed that upgrading multi-omics

toolkits with traditional absolute measurements un-

locks the scaling of core biological questions to dy-

namic and complex microbiomes, creating a deeper

insight into inter-organismal relationships that drive

the greater community function. (Bareche et al.,

2018) used copy-number aberrations, somatic muta-

tions and gene expression data to demonstrate the

substantial genomic heterogeneity that characterizes

TNBC molecular subtypes, allowing for a better un-

derstanding of the disease biology as well as the iden-

tiﬁcation of several candidate targets paving novel ap-

proaches for the development of anticancer therapeu-

tics for TNBC. (Singh et al., 2016) presented DIA-

BLO (Data Integration Analysis for Biomarker dis-

covery using a Latent component method for Omics

studies) a multivariate analysis method based on the

dimension reduction maximizing the correlation be-

tween each omics pairs. It underlines the importance

of the processing of the interaction between omics

data before using them for the prediction task.

Following this trend towards the integration of

multi-omics data this work presents a machine learn-

ing algorithm based on pathways interactions be-

tween gene expression, epigenome expression and

miRNAs expression to learn interpretable and sparse

models uncovering potentially new candidate targets

for the development of anticancer therapeutics for

TNBC. The use of the multi-omics data helps empha-

sis the interaction between the different omics which

leads to a better interpretation of the mechanism so-

lution. Here we focus on the interpretable model

because we strive to explain why these features are

selected in the model decision path. There are two

types of models in machine learning, generally speak-

ing: the black box model and the interpretable model.

An interpretable machine learning refers to methods

and models that make the behaviour and predictions

of machine learning systems understandable to hu-

mans. It differs from the black box model which

is a system that does not reveal its internal mecha-

nisms. Interpretability is more of a concept there-

fore there is no real mathematical deﬁnition of inter-

pretability. A (non-mathematical) deﬁnition reported

by (Miller, 2019) is: Interpretability is the degree to

which a human can understand the cause of a deci-

sion. The need for interpretability here lies in the fact

that for some tasks (especially the biological data), it

is not enough to get the prediction (the what) (Doshi-

Velez and Kim, 2017). The model must also explain

how it came to the prediction (the why), because a

correct prediction only partially solves your original

problem (Molnar, 2019). Moreover, interpretability

favours the acceptance of the model by health practi-

tioners. Before the learning phase, there is a need to

integrate the omics data. (Ritchie et al., 2015) deﬁne

the integration of omics as the combination of multi-

ple omics datasets to develop multivariate models to

predict complex phenotypes. Mainly, it is the con-

catenation method (Liu et al., 2013) and the ensem-

ble method (G

unther et al., 2012). The concatenation

method consists of putting all the omics together i.e.

side-by-side and treating them as only one mega-set.

The ensemble method consists of learning a classiﬁer

on each omics dataset and building a majority vote

with the classiﬁers, for the ﬁnal prediction. With time,

new methods were also explored including the clus-

tering (Rappoport et al., 2020) and the graph network

((Singh et al., 2016), (Heuschkel et al., 2020)). Here,

in this work, the goal is to let the algorithm infer po-

tential interaction between the features with the inte-

gration of new priors based on pathways of gene inter-

actions. Therefore the concatenation method which

does not imply any priors on the feature is more suit-

able for the task at hand. Depending on what a feature

refers to, pathway-based feature selection methods

can be classiﬁed into three categories: pathway-level

selection, bi-level selection and pathway-guided gene

selection (Tian et al., 2019). In the pathway-level se-

lection, the goal is to select the whole pathways as-

sociated with the phenotypes of interest ((Dinu et al.,

2007), (Wu et al., 2010)). The bi-level selection is a

process, which identiﬁes not only relevant pathways

but also important genes that contribute critically to

the signiﬁcance of identiﬁed relevant pathways. In the

pathway-guided gene selection, a feature corresponds

Applying PySCMGroup to Breast Cancer Biomarkers Discovery

to an individual gene.

The algorithm (GroupSCM) presented here is

an extension of the Set Covering Machine (SCM)

which was originally introduced by (Marchand and

Shawe-Taylor, 2002). The SCM is a learning algo-

rithm using a greedy approach to learn a conjunc-

tion set or a disjunction set of rules. The rules are

the decision stumps (omic value expression ≤ v or

omic value expression ≥ v). The SCM applicability

to the biomarker discovery task world is an effective

algorithm in the domain. Kover (Drouin et al., 2016),

a software used in the prediction of antibiotic resis-

tance in bacteria is based on the SCM. This a robust

algorithm yields great results in the antibiotic resis-

tance prediction problem in accordance with the pub-

lished literature. The GroupSCM extends the classi-

cal SCM to integrate the knowledge of variables in-

teracting in biological pathways. The intuition behind

this is to favour a rule selection based on their path-

ways association. The GroupSCM use the pathway-

guided gene selection as priors and modify the utility

function of the SCM into taking in account different

weights on each feature depending on the pathways to

which they belong.

2 METHODS

Let’s S be a set of data such as S

def

{

, y

)

}

i=0

∼ D

where x

∈ X is a training example, y

∈ Y its asso-

ciated label, D is a data generating distribution and

m the size of the dataset. In this work, we consider

the diagnosis of TNBC for a patient as a binary clas-

siﬁcation task where x

represents the ith patient data

(multi-omics features) and y

∈

{

−1, 1

}

whether the

patient is TNBC or non-TNBC. Below we ﬁrst de-

scribe the original SCM algorithm then we present

our extension, the GroupSCM.

2.1 SCM

The SCM is an iterative algorithm whose goal is to

produce either a conjunction of rules type model or

a disjunction of rules type. It does so, by selecting

the rule i with the maximum utility function at each

iteration until there is no more examples or the num-

ber of rules predeﬁned is reached. This is a maxi-

mization optimization algorithm type. At each iter-

ation, a utility score U

is computed for each of the

remaining rules. Then an optimal rule i is selected

∗

← argmax

) and so on to build a conjunction or

disjunction of rules as the ﬁnal model. Algorithm 1

presents the pseudo-code of the SCM. The SCM takes

as input a set of examples (the training set) S ; the type

of machine T which can be either a conjunction or a

disjunction; a trade-off parameter p which inﬂuence

how much we want to penalize the error on the pos-

itive examples and ﬁnally a parameter s which is the

maximum number of rules to consider in the ﬁnal re-

sults.

2.2 GroupSCM

The GroupSCM is based on the roots of the SCM. The

goal of the GroupSCM is to overweight by a previ-

ously computed term, every utility score U

of rules

i belonging to the same pathway(s) as the selected

rule(s) at the previous iteration. This overweight term

is denoted the prior of rule i p

. All of the p

belong

to the set PR = {p

}

i∈[1,n]

with n the number of fea-

tures. In addition to the SCM inputs i.e S, T , p and s,

the GroupSCM also takes as an input PR and GR, the

set of the groups for each rule already chosen. At the

beginning i.e iteration 0, no rule is chosen therefore

GR =

0. Let’s denote G the set of groups (pathway)

G = {g

} with k ∈ [1, |G|]. PG is the set of the prior

(preference) on the groups: PG = {p

}

g∈[1,G]

. For

each rule ri, the corresponding rule prior p

is corre-

lated to the sum of p

: p

= exp(c·(

∑

g∈g

)). The pri-

ors of the rules p

are heavily dependent on the priors

on the groups p

. Therefore, p

is where researchers

and domain users should put their knowledge. Here

we use a function that depends on the length of g

= f (−c · |g|) with f : f (x) = exp(−c · x). Algo-

rithm 2 presents the pseudo-code.

2.3 Pathways Building

The pathways are retrieved from two (2) databases:

BioGRID database (Stark et al., 2006) and MSigDB

(Liberzon et al., 2011). BioGRID is an open-access

database dedicated to the annotation and archival of

protein, genetic and chemical interactions for all ma-

jor model organism species including humans. The

Molecular Signatures Database (MSigDB) is a col-

lection of annotated gene sets for use with GSEA

software. 78384 pathways of gene interactions are

retrieved from BioGRID. In MSigDB, the targets

are the gene sets that were curated from pathway

databases. Usually, those genes sets are canonical

representations of a biological process compiled by

domain experts. There are 2232 canonical gene sets

in MSigDB. MSigDB also provides great information

on the miRNA. The miRNA gene sets extracted from

MSigDB represent potential targets of regulation by

transcription factors or microRNAs. The 1756 sets

consist of genes grouped by elements they share in

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

their non-protein-coding regions. In total there are

82372 gene sets (or pathways).

3 RESULTS

3.1 Datasets

The datasets were obtained from TCGA and all eth-

ical regulations were followed. The samples used

are those coming from the primary solid tumor (sam-

ple type code 01) and to the ﬁrst vial (vial code

A). The database contains numerous biological data

types including DNA methylation, mRNA expres-

sion, miRNA expression, SNPs, and clinical informa-

tion. The clinical view contains information regarding

cancer stage, menopause status, race, and tumor sta-

tus among others. The TNBC phenotype is retrieved

from that clinical ﬁle by verifying the PR, ER, and

HER2 phenotype i.e. positive or negative status. All

the patients with negative PR, negative ER, and neg-

ative HER2 phenotypes are classiﬁed as TNBC pa-

tients and all the other combinations are classiﬁed as

non-TNBC. There are 146 TNBC patients vs. 756

non-TNBC patients. It is worth noting that the dataset

is a bit imbalanced (1 for 5).

The epigenome data is obtained using Illumina

Methylation Assay technologies. There are 2 types

of Illumina Methylation technologies: the 27k probes

and the 450k probe technologies. Since TCGA is a

collection of datasets from multiple independent stud-

ies, there are many discrepancies in the global dataset.

For example, the information about the epigenome for

some patients is either only available in the 27k tech-

nology or in the 450k technology and is sometimes al-

together missing. In regards to this situation, we build

a view denoted the methyl fusion view. It essentially

consists of taking into account just the CpG available

in both of the technologies (27k or 450k) to build a

methyl fusion view for most of the patients. The new

feature size of this methyl fusion view is 19984 CpG

dinucleotides. The RNA expression data is available

in 2 forms: the gene expressions (20531 features) and

the genes isoforms expression, which includes splice

variants, expressions (73599 features). The miRNA

expression has 1046 features. Since the dimension

is enormous, a variable selection based on the me-

dian absolute deviation (MAD) as presented in (Singh

et al., 2016) was made. Two thousand (2000) features

were selected from the methyl fusion view. The same

amount was also selected from the RNA expression

view and ﬁnally, two hundred and ﬁfty (250) from the

miRNAs view. The ﬁnal dataset has 4250 features

of CpGs, RNA, and miRNA. Each of the remaining

CpGs sites and the RNA isoforms were linked to their

corresponding genes names. When it is impossible to

link the names of the gene to the features, the feature

is discarded from the dataset. Once the dataset is ﬁ-

nalized, the next logical step is the integration of the

groups pathways. After that preprocessing, each fea-

ture is assigned to its corresponding pathways groups.

If the feature does not belong to any pathways, it is

classiﬁed into a new pathway created just for those

types of features. For the GroupSCM experiments,

two (2) variants of the dataset are used: one with all

the features (902, 4250) and another without the fea-

tures not belonging to a pathway (902, 4110).

3.2 Metrics

In this section, we present the different metrics used

to evaluate the models learned. Since the problem is

a classiﬁcation task, four (4) metrics are reported: the

Accuracy

t p+tn

t p+ f p+tn+ f n

, the F1 score 2 ·

precision·recall

precision+recall

the Recall

t p

t p+ f n

and the Precision

t p

t p+ f p

. t p stands

for true positive; f p for false positive; tn for true neg-

ative and f n for false negative. Due to the data im-

balance, the most important metric to look for to as-

sess the performance of the models learned is the: F1

score. Being a mean between the precision and the

recall, it is well suited to understand the model per-

formance on both the positive and negative examples

simultaneously.

3.3 Experiments Results

All the experiments and their results are presented

here. Initially, two algorithms are evaluated: the

Decision Trees (DT) and the Set Covering Machine

(SCM) on the dataset. These evaluations serve as the

baseline results to compare the results of the Group-

SCM experimentation. As stated before, the multi-

omics data combination used here is CpG - RNA iso-

forms - miRNA. To eliminate the randomness due to

the creation of the train - test splits, the experiments

are run 15 times with different random train - test

splits at different random seeds. The dataset is split

using the 80/20 % split i.e. 80% for the training

set and 20% for the test set. In the learning phase,

a 5-fold cross-validation is applied to select the best

hyperparameters. If not stated otherwise the metrics

score presented here are only the metrics on the test

set.

At each run, i.e. for each split the metrics on

the test and train sets are reported and an average is

computed to see the mean performance of the algo-

rithms overall. Figure 2 shows the main results of

Applying PySCMGroup to Breast Cancer Biomarkers Discovery

Algorithm 1: Set Covering Machine.

1: procedure SCM(S, p, s, T, R )

2: INPUT : S: A set of m training examples, T: a type of machine conjunction or disjunction, p: The class

trade-of parameter, s: The maximum number of rules in h, R =

{

(x)

}

|R |

i=1

: a set of boolean-valued rules;

3: R

∗

← φ

4: P ← the set of examples in S with label 1

5: N ← the set of examples in S with label 0

6: stop ← False

7: while N 6= φ or |R

∗

| ≤ s or ¬stop do

8: ∀r

inR , N

← the subset of N correctly classiﬁed by r

9: ∀r

inR , P

← the subset of P correctly misclassiﬁed by r

10: ∀r

inR , U

= |N

| − p · |P

| if |N

| ≥ |P

| and −∞ otherwise

11: i

∗

← argmax

12: if U

∗

6= ∞ then and |N

∗

| > 0 or |P

∗

| > 0

13: R

∗

← R

∗

∪ r

∗

14: N ← N − N

∗

15: P ← P − P

∗

16: else

17: stop = True

18: return h, where h(x) =

∗

∈R

∗

(x) if T = conjunction or h(x) =

∗

∈R

∗

(x) if T = disjunction

Algorithm 2: Group Set Covering Machine (GroupSCM) (in bold the principal differences between SCM and GroupSCM.

1: procedure GROUPSCM(S, T, p, s, PR,GR, R )

2: INPUT : S: A set of training examples, T: a type of machine conjunction or disjunction, p: The class

trade-of parameter, s: The maximum number of rules in h, PR = {p

}

i∈[1,n]

, R =

{

(x)

}

|R |

i=1

: the set of

boolean rules; GR the subset of the groups chosen

3: OUT PU T : A conjunction or a disjunction f(x) of a subset R ⊆ H

4: R

∗

← φ

5: P ← the set of examples in S with label 1

6: N ← the set of examples in S with label 0

7: stop ← False

8: while N 6= φ or |R | ≤ s or ¬stop do

9: ∀r

inR , N

← the subset of N correctly classiﬁed by r

10: ∀r

inR , P

← the subset of P correctly misclassiﬁed by r

11: ∀r

inR , U



| − p · |P



· p

if |N

| ≥ |P

| and −∞ otherwise

12: i

∗

← argmax

13: Tie break choose i

∗

such as argmin

∗

14: Update PR: For all remaining rules,

15: . Case 1: favor rules from same group (inner group) p

= p

∗ exp(|g

∪ GR|)

16: . Case 2: favor rules from different groups (outer group) p

= p

∗ exp(−|g

GR|)

17: Update GR ← GR ∪ g

with g

the groups/pathways of rules choosen at 12

18: if U

∗

6= ∞ then and |N

∗

| > 0 or |P

∗

| > 0

19: R

∗

← R

∗

∪ r

∗

20: N ← N − N

∗

21: P ← P − P

∗

22: else

23: stop = True

24: return h, where h(x) =

∗

∈R

∗

(x) if T = conjunction or h(x) =

∗

∈R

∗

(x) if T = disjunction

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

Table 1: Mean F1 Scores from the best seed for each algorithm and the best values from the best seed. * is the GroupSCM

applied to the dataset without the features without a pathway. c is the value of the hyperparameter of the GroupSCM.

Algorithm Mean Metrics Values Best Metrics Values

DT 0.89 +/- 0.03 0.94

SCM 0.91 +/- 0.03 0.96

GroupSCM

Outer Inner Outer Inner

c=0.4; 0.88 +/- 0.04 c=0.1; 0.88 +/- 0.04 0.94 0.96

GroupSCM* c=0.1; 0.83 +/- 0.07 c=0.1; 0.83 +/- 0.07 0.90 0.90

the experiments. The F1-Score with the DT mod-

els is: 0.89 +/- 0.03 (ﬁg 2 a). The F1-Score with

the SCM models is : 0.91 +/- 0.03 (ﬁg 2 b). The

baseline scores can be judged as good scores since

they are around 90%. The SCM slightly outperforms

the DT. The models decision paths are similar be-

tween the SCM and the DT. Indeed on the 15 repeti-

tions, the DT outputs exclusively uc002vwt.2 MLPH

as the only tree 4 times and uc002vwt.2 MLPH as

the root of the trees 9 other times. Meanwhile, the

SCM outputs uc002vwt.2 MLPH as the only rule 9

times and in a conjunction of rules 3 other times.

Figure 1 presents the number of features retrieve

by each algorithm at each of the 15 seeds. The

SCM is mostly the sparser model at each seed fol-

lowed by the GroupSCM extension then the DT.

Furthermore, to better investigate the biomarkers re-

trieved, only the biomarkers selected by the model

with the best F1 score on the test set will be ana-

lyzed. The best model obtains by the DT is a tree

with only root uc002vwt.2 MLPH with 0.94 F1 score.

The best model obtains by the SCM is a conjunc-

tion of rules uc002vwt.2 MLPH, uc002hul.3 RARA,

uc009wsd.2 HDGF and uc001jpo.1 TSPAN15 with

0.96 F1 score.

The last two plots in the Figure 2 show the results

of the GroupSCM on the dataset with all the features

(ﬁg 2 c) and the dataset without the features not be-

longing to a pathway (ﬁg 2 d). With all the same

features (ﬁg 2 c), GroupSCM performs relatively like

the SCM metrics-wise. Indeed with the best hyper-

parameters, the average F1 score is 0.88 +/- 0.04.

The best model has an F1 score of 0.96. Mean-

while, the rules chosen are totally different from the

ones obtained from the baseline results. The rules se-

lected here are a conjunction of cg17095936 TBX19,

cg10305797 KRTDAP, and cg18267381 ZNF385D.

Their respective belonging pathways are: [G 82372],

[G 82372], [G 82372]. This is the same pathway.

The immediate conclusion from this is the effective-

ness of the prior given the fact that it helped guide the

decision paths of the algorithm. However, G 82372 is

the pathway of the features that did not belong to any

pathway within the databases data. Therefore even

though this is a great result there is a lack of biological

interpretation with this result. Why? Because the bi-

ological impact of the pathways cannot be explained

directly here.

The ﬁnal experimentation explores the perfor-

mance of the GroupSCM on the dataset without the

features that did not belong to any pathway within

the databases data retrieved. This experiment is more

suitable to correctly assess the performance of the

prior and the algorithm overall. With the best hyper-

parameters, the average F1 score is 0.82 +/- 0.05 on

the test set (ﬁg 2 d). Despite being roughly 9% lower

than the simple SCM and the GroupSCM on the com-

plete dataset, this is still a good score since it’s over

80% on average. The best model from this experi-

mentation has 0.90 F1 score. This model selects the

conjunction of these rules: hsa-mir-18a and hsa-mir-

190b. Table 1 presents a summary of all the results

of all the experimentation. One important observa-

tion to notice is the fact that the score values are not

different from the inner-group our outer-group update

(see algorithm 2). Another important observation is

despite the drop in performance, the algorithm is still

pretty good but is also more sparse on the pathways

levels. Most of the new rules selected belong to one

pathway. Therefore their interaction can be biologi-

cally interpreted and analyzed. In table 2 the features

selected by the best model for each experiment are

presented. Along with those features, the pathways to

which they belong are also presented.

4 DISCUSSION

In this paper, we tackle the breast cancer triple neg-

ative prediction problem with the purpose of provid-

ing an interpretable and sparse model. We elect to

do this task by learning a GroupSCM which is essen-

tially a SCM with a prior on the rules selection. The

new utility function is set to increase the weights of

the rules previously selected in the previous iteration.

There are two types of overweighting process used

in the algorithm: the inner-group-weighting and the

outer-group-weighting. In the ﬁrst case (the inner-

Applying PySCMGroup to Breast Cancer Biomarkers Discovery

Figure 1: Number of feature return per algorithm at each seed (iteration).

Table 2: Features and Pathways selected by the best model of each algorithm type.

Algorithm Features Pathways

uc002vwt.2 MLPH [G 455, G 1029, G 10245, G 11700, G 15952, G 25302,

G 30827, G 33710, G 44607, G 51709, G 52337]

SCM

uc002vwt.2 MLPH, uc002hul.3 RARA,

uc009wsd.2 HDGF, uc001jpo.1 TSPAN15 Not Shown (too much values)

GroupSCM

cg17095936 TBX19, cg10305797 KRTDAP, [G 82372], [G 82372],

cg18267381 ZNF385D [G 82372]

GroupSCM* hsa-mir-18a, hsa-mir-190b [G 80747], [G 81521]

group case), if the model selects some rules outside

the same pathways the conclusion is that the correct

explanation (from the experiment) is inter-groups re-

lated (since the rules belong to different pathways). In

the second case (the outer-group case), if the model

selects rules from the same pathways (i.e. it uses the

priors) meaning that the correct explanation is intra-

groups related. But overall, looking at the results in

table 1, the metric scores are not different from in-

ner versus outer update for the GroupSCM. Even if

it is for different c values. Despite being a little bit

less performing than the SCM, the GroupSCM pro-

duce sparser rules: three rules selected versus ﬁve

rules selected for the SCM. This comparison is on

the dataset with all the features. The results are even

sparser when considering the experiment with just the

features belonging to a known pathway. This situation

selects two rules. The sparser model implies easier in

vivo experimentation and validation.

Interestingly, the SCM selects almost all rules in

the RNA omics components and the GroupSCM se-

lect the rules either in the CpGs omics part or the

miRNA part (table 2). The main observation here

is while both the DT and the SCM seem to select

only RNA isoforms rules therefore the RNA omics

views only, the GroupSCM overlooks this view to se-

lect rules from either the CpGs (methylome view) or

the miRNA (miRNA views). This is useful ﬁnding

since CpGs and miRNA impact the translation there-

fore the gene expressions levels. Biologically speak-

ing, the algorithm is putting greater emphasis on the

upstream of the biology principal theory. Let’s recall

here one of the hidden goals is to see how well in-

vestigating all the omics together would perform on

the prediction task of determining the TNBC vs. non-

TNBC patients. In that case, the algorithm still con-

centrates on just one component of the omics view

but in a more integrative fashion. Indeed knowing if

the CpGs sites are hyper or under methylated is also

informative on the genes expression level, since we

have the pathways of the genes, those sites regulate.

It is a similar process for the miRNA. Their levels of

expression (over or under) also affect the gene expres-

sions level, and knowing their pathways, we can see

the downstream impact of the miRNAs on gene ex-

pressions. The GroupSCM enables a bigger picture

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

(a) Decision Tree (b) SCM

Figure 2: DT-SCM-GroupSCM Mean Metrics Results (Accuracy, F1-score, Precision and Recall)(Train (Gray) Test (Red)).

and observation of the biomarker discovery process

in a more global way.

We elect to analyze the features returned by

the GroupSCM in the experiment without the fea-

tures not belonging to a pathway. The model ob-

tains from this experimentation is a two-rule con-

junction: hsa-mir-18a and hsa-mir-190b. Accord-

ing to (Weizmann, 2020a) diseases associated with

mir18A include thyroid gland anaplastic carcinoma

and medulloblastoma. Among its related pathways

are Parkinson’s disease pathway and DNA damage

response. In our database: mir18A belongs to path-

way G 80747 in our experimentation (http://mirdb.

org/cgi-bin/mature mir.cgi?name=hsa-miR-18a-3p).

This biomarker is known to heavily impact numer-

ous pathways or interactions in many cancers. (Ko-

matsu et al., 2014) described mir18A as an impor-

tant biomarker in cancer since mir18A, which is lo-

cated in the potentially oncogenic miR-17-92 clus-

ter, is a highly expressed microRNAs in several

types of cancers. (Li et al., 2016) demonstrated

that tamoxifen resistance in breast cancer cells is en-

hanced through a miR-18a-HIF1 feedback regulatory

loop. Recently, (Zhang et al., 2019) also showed

that SREBP1, targeted by miR-18a-5p, modulates

epithelial-mesenchymal transition in breast cancer

via forming a co-repressor complex with Snail and

HDAC1/2. This literature review conﬁrms that our

algorithm targets an already known biomarker in can-

cer studies. But there is no link speciﬁcally to the

TNBC phenotype. According to (Weizmann, 2020b),

there are no diseases linked to mir190B. Nevertheless,

(Cizeron-Clairac et al., 2015) proved that mir190B is

the highest up-regulated miRNA in ER-positive com-

pared to ER-negative breast tumors. Making it a po-

tential new biomarker for the triple-negative breast

cancer. (Zhao et al., 2020) recently demonstrated that

long non-coding RNA TUSC8 inhibits breast cancer

growth and metastasis via miR190b-5p/MYLIP axis.

These evidences show the ability of our model to dis-

cover biomarkers for the TNBC prediction problem.

As we can see for both of these miRNAs there is no

direct connection between them and the TNBC phe-

notype but our study suggests otherwise. The next

step will be to analyze those biomarkers in in vivo set-

tings to have solid conﬁrmation of the discovery. We

proceeded with a statistical analysis of the biomark-

ers retrieved and analyzed the expression level in the

population with the classical t-test and the p-values.

In ﬁgure 3 we plotted the expression levels of both

Applying PySCMGroup to Breast Cancer Biomarkers Discovery

(a) hsa-mir-18a

(b) hsa-mir-190b

Figure 3: Expression levels of the hsa-mir-18a (a), hsa-mir-

190b (b) and both (c) in the TNBC vs non-TNBC.

features independently alone (ﬁg 3 a & b) and to-

gether. The conclusion is both of them are statisti-

cally signiﬁcant with hsa-mir-18a more expressed in

the TNBC and hsa-mir-190b in the non-TNBC. To-

gether they are substantially expressed in the TNBC

examples improving the signiﬁcance.

5 CONCLUSION

We present and apply an extension of the SCM to

an algorithm using a prior on the pathways of appar-

tenance of the features. In this particular case, the

algorithm yields interesting results biomarkers while

maintaining a good statistical scores overall. This al-

gorithm is a good addition to the precision medicine

ﬁeld using the pathway interaction to ﬁnd the appro-

priate biomarkers related to a speciﬁc disease. It is

sparse and interpretable which suits clinician expec-

tations. Despite the ﬁndings, additional in vivo exper-

imentations should be completed continuing the im-

provements of the model performance of data-driven

predictions and to provide validated evidence linking

speciﬁc biomarkers to a disease phenotype.

COMPETING INTERESTS

The authors declare that they have no competing in-

terests.

AUTHOR’S CONTRIBUTIONS

OMA designed the algorithm and the experimentation

and wrote this article. PT co-designed the algorithm.

JC and FL supervised the work and contributed to the

redaction of the manuscript.

ACKNOWLEDGEMENTS

A special thanks to Emmanuel Noutahi for his inputs

in this work. The work was supported by the Canada

Research Chair in Medical genomics (JC), Compute

Canada, Institute Intelligence and Data (IID) and

Centre de recherche en donn

ees massives (CRDM).

REFERENCES

Bareche, Y., Venet, D., Ignatiadis, M., Aftimos, P., Piccart,

M., Rothe, F., and Sotiriou, C. (2018). Unravelling

triple-negative breast cancer molecular heterogeneity

using an integrative multiomic analysis. Annals of On-

cology, 29(4):895–902.

Cizeron-Clairac, G., Lallemand, F., Vacher, S., Lidereau,

R., Bieche, I., and Callens, C. (2015). Mir-190b,

the highest up-regulated mirna in erα-positive com-

pared to erα-negative breast tumors, a new biomarker

in breast cancers? BMC cancer, 15(1):499.

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms

Delogu, F., Kunath, B., Evans, P., Arntzen, M., Hvid-

sten, T., and Pope, P. (2020). Integration of abso-

lute multi-omics reveals dynamic protein-to-rna ratios

and metabolic interplay within mixed-domain micro-

biomes. Nature Communications, 11(1):1–12.

Dinu, I., Potter, J. D., Mueller, T., Liu, Q., Adewale, A. J.,

Jhangri, G. S., Einecke, G., Famulski, K. S., Halloran,

P., and Yasui, Y. (2007). Improving gene set analysis

of microarray data by sam-gs. BMC bioinformatics,

8(1):242.

Doshi-Velez, F. and Kim, B. (2017). Towards a rigorous sci-

ence of interpretable machine learning. arXiv preprint

arXiv:1702.08608.

Drouin, A., Gigu

ere, S., D

eraspe, M., Marchand, M., Tyers,

M., Loo, V. G., Bourgault, A.-M., Laviolette, F., and

Corbeil, J. (2016). Predictive computational pheno-

typing and biomarker discovery using reference-free

genome comparisons. BMC genomics, 17(1):1–15.

Freedman, M. L., Monteiro, A. N., Gayther, S. A., Coet-

zee, G. A., Risch, A., Plass, C., Casey, G., De Biasi,

M., Carlson, C., Duggan, D., et al. (2011). Principles

for the post-gwas functional characterization of cancer

risk loci. Nature genetics, 43(6):513–518.

unther, O. P., Chen, V., Freue, G. C., Balshaw, R. F.,

Tebbutt, S. J., Hollander, Z., Takhar, M., McMaster,

W. R., McManus, B. M., Keown, P. A., et al. (2012). A

computational pipeline for the development of multi-

marker bio-signature panels and ensemble classiﬁers.

BMC bioinformatics, 13(1):326.

Heuschkel, M. A., Skenteris, N. T., Hutcheson, J. D.,

van der Valk, D. D., Bremer, J., Goody, P., Hjortnaes,

J., Jansen, F., Bouten, C. V., van den Bogaerdt, A.,

et al. (2020). Integrative multi-omics analysis in cal-

ciﬁc aortic valve disease reveals a link to the formation

of amyloid-like deposits. Cells, 9(10):2164.

Iorio, M. V., Ferracin, M., Liu, C.-G., Veronese, A., Spizzo,

R., Sabbioni, S., Magri, E., Pedriali, M., Fabbri, M.,

Campiglio, M., et al. (2005). Microrna gene expres-

sion deregulation in human breast cancer. Cancer re-

search, 65(16):7065–7070.

Karczewski, K. J. and Snyder, M. P. (2018). Integrative

omics for health and disease. Nature Reviews Genet-

ics, 19(5):299.

Komatsu, S., Ichikawa, D., Takeshita, H., Morimura, R.,

Hirajima, S., Tsujiura, M., Kawaguchi, T., Miyamae,

M., Nagata, H., Konishi, H., et al. (2014). Circulat-

ing mir-18a: a sensitive cancer screening biomarker

in human cancer. in vivo, 28(3):293–297.

Lao, V. V. and Grady, W. M. (2011). Epigenetics and col-

orectal cancer. Nature reviews Gastroenterology &

hepatology, 8(12):686.

Lehmann, B. D., Bauer, J. A., Chen, X., Sanders, M. E.,

Chakravarthy, A. B., Shyr, Y., and Pietenpol, J. A.

(2011). Identiﬁcation of human triple-negative breast

cancer subtypes and preclinical models for selection

of targeted therapies. The Journal of clinical investi-

gation, 121(7):2750–2767.

Li, X., Wu, Y., Liu, A., and Tang, X. (2016). Long

non-coding rna uca1 enhances tamoxifen resistance

in breast cancer cells through a mir-18a-hif1α feed-

back regulatory loop. Tumor Biology, 37(11):14733–

14743.

Liberzon, A., Subramanian, A., Pinchback, R., Thor-

valdsd

ottir, H., Tamayo, P., and Mesirov, J. P. (2011).

Molecular signatures database (msigdb) 3.0. Bioinfor-

matics, 27(12):1739–1740.

Liu, Y., Devescovi, V., Chen, S., and Nardini, C. (2013).

Multilevel omic data integration in cancer cell lines:

advanced annotation and emergent properties. BMC

systems biology, 7(1):14.

Marchand, M. and Shawe-Taylor, J. (2002). The set cover-

ing machine. Journal of Machine Learning Research,

3(Dec):723–746.

Miller, T. (2019). Explanation in artiﬁcial intelligence: In-

sights from the social sciences. Artiﬁcial Intelligence,

267:1–38.

Molnar, C. (2019). Interpretable Machine Learning. https:

//christophm.github.io/interpretable-ml-book/.

Rappoport, N., Safra, R., and Shamir, R. (2020). Monet:

Multi-omic module discovery by omic selection.

PLOS Computational Biology, 16(9):e1008182.

Ritchie, M. D., Holzinger, E. R., Li, R., Pendergrass, S. A.,

and Kim, D. (2015). Methods of integrating data to

uncover genotype-phenotype interactions. Nature re-

views. Genetics, 16(2):85.

Singh, A., Gautier, B., Shannon, C. P., Vacher, M., Rohart,

F., Tebutt, S. J., and Le Cao, K.-A. (2016). Diablo-

an integrative, multi-omics, multivariate method for

multi-group classiﬁcation. bioRxiv, page 067611.

Stark, C., Breitkreutz, B.-J., Reguly, T., Boucher, L., Bre-

itkreutz, A., and Tyers, M. (2006). Biogrid: a general

repository for interaction datasets. Nucleic acids re-

search, 34(suppl 1):D535–D539.

Tian, S., Wang, C., and Wang, B. (2019). Incorporating

pathway information into feature selection towards

better performed gene signatures. BioMed research

international, 2019.

Weigelt, B., Baehner, F. L., and Reis-Filho, J. S. (2010).

The contribution of gene expression proﬁling to breast

cancer classiﬁcation, prognostication and prediction:

a retrospective of the last decade. The Journal of

Pathology: A Journal of the Pathological Society of

Great Britain and Ireland, 220(2):263–280.

Weizmann, I. o. S. (2020a). Mir18a. https://www.

genecards.org/cgi-bin/carddisp.pl?gene=MIR18A.

Accessed: 2020-10-21.

Weizmann, I. o. S. (2020b). Mir190b. https://www.

genecards.org/cgi-bin/carddisp.pl?gene=MIR190B.

Accessed: 2020-10-21.

Wu, D., Lim, E., Vaillant, F., Asselin-Labat, M.-L., Vis-

vader, J. E., and Smyth, G. K. (2010). Roast: ro-

tation gene set tests for complex microarray experi-

ments. Bioinformatics, 26(17):2176–2182.

Yuan, Y., Liu, L., Chen, H., Wang, Y., Xu, Y., Mao, H., Li,

J., Mills, G. B., Shu, Y., Li, L., et al. (2016). Com-

prehensive characterization of molecular differences

in cancer between male and female patients. Cancer

cell, 29(5):711–722.

Zhang, N., Zhang, H., Liu, Y., Su, P., Zhang, J., Wang, X.,

Sun, M., Chen, B., Zhao, W., Wang, L., et al. (2019).

Applying PySCMGroup to Breast Cancer Biomarkers Discovery

Srebp1, targeted by mir-18a-5p, modulates epithelial-

mesenchymal transition in breast cancer via forming

a co-repressor complex with snail and hdac1/2. Cell

Death & Differentiation, 26(5):843–859.

Zhao, L., Zhou, Y., Zhao, Y., Li, Q., Zhou, J., and Mao,

Y. (2020). Long non-coding rna tusc8 inhibits breast

cancer growth and metastasis via mir-190b-5p/mylip

axis. Aging (Albany NY), 12(3):2974.

BIOINFORMATICS 2021 - 12th International Conference on Bioinformatics Models, Methods and Algorithms