=Paper=
{{Paper
|id=Vol-2567/paper22
|storemode=property
|title=Constructing prototypes for classification using epigenetic and genetic analysis
|pdfUrl=https://ceur-ws.org/Vol-2567/paper22.pdf
|volume=Vol-2567
|authors=Christopher L. Bartlett
|dblpUrl=https://dblp.org/rec/conf/iccbr/Bartlett19
}}
==Constructing prototypes for classification using epigenetic and genetic analysis==
https://ceur-ws.org/Vol-2567/paper22.pdf
Constructing prototypes for classification using
epigenetic and genetic analysis
Christopher L. Bartlett
Intelligent Bio Systems Laboratory, Biomedical and Health Informatics
State University of New York at Oswego, 7060 NY-104, Oswego, NY 13126
cbartle3@oswego.edu
1 Abstract
Researchers seek to identify biological markers which accurately differentiate
cancer subtypes and their severity from normal controls. One such biomarker,
DNA methylation, has recently become more prevalent in genetic research stud-
ies in oncology. This project seeks to apply the innovative and adaptive machine
learning methodology in case-based reasoning (CBR) to examine DNA methyla-
tion levels in breast cancer. Instead of relying on a generalized knowledge-base,
CBR uses highly specific information extracted from similar cases which can
also greatly expedite the process of finding a solution. Further, this can locate
targeted biomarkers by reusing homogenous factors, or revising to locate novel
biomarkers in highly heterogeneous samples. While locating these biomarkers,
this project proposes to use CBR to classify samples, predict prognoses and
determine survival factors.
1 Introduction
The term epigenetics was first introduced into modern biology by Conrad Wadding-
ton as a means of defining interactions between genes and their products that
result in phenotypic variations. Waddington’s landscape presents a cell becom-
ing more differentiated as time goes on. One of the events that can cause this
differentiation is methylation. Methylation is a covalent attachment of a methyl
group to cytosine. Cytosine (C) is one of the four bases that construct DNA
and one of only two bases that can be methylated. While adenine can be methy-
lated as well, cytosine is typically the only base that’s methylated in mammals.
Once this methyl group is added, it forms 5-methylcytosine where the 5 refer-
ences the position on the 6-atom ring where the methyl group is added. Under
the majority of circumstances, a methyl group is added to a cytosine followed
by a guanine (G) which is known as CpG. While the methyl group is added
onto the DNA, it doesn’t alter the underlying sequence but it still has profound
effects on the expression of genes and the functionality of cellular and bodily
functions. Methylation at these CpG sites has been known to be a fairly sta-
ble epigenetic biomarker that usually results in silencing the gene. Further, the
Copyright © 2019 for this paper by its authors. Use permitted
under Creative Commons License Attribution 4.0 International (CC BY 4.0).
amount of methylation can be increased (known as hypermethylation) or de-
creased (known as hypomethylation) and improper maintenance of epigenetic
information can lead to a variety of human diseases.
Within the domain of case-based reasoning (CBR), there exist several appli-
cations using microarray data. Anaissi, Goyal, Catchpoole, Braytee, and Kennedy
[1], for example, attempted to navigate the complexity of the highly-dimensional
and imbalanced datasets often found in microarray analysis by focusing on case
retrieval. Their framework uses a k-nearest neighbor (kNN) classifier with a
weighted feature-based similarity measure to retrieve similar patients from a
case base of acute lymphblastic leukemia. Gene expression data is employed to
determine this similarity, and the treatment and outcome is used to propose solu-
tions. Feature selection, dimensionality reduction, and feature weighting is used
to handle the high-dimensionality of the data and removal of irrelevant features.
They utilize oversampling to deal with the imbalanced classes. More specifically,
they use the synthetic minority oversampling technique (SMOTE) methodol-
ogy which artificially creates minority samples based on interpolation between
members of the original minority class. After these pre-processing stages, a new
sample is given to the kNN classifier to retrieve similar cases.
A bit unorthodox, Yao and Li, [4], considered microarray samples in each
class as one case-base. Then, given a sample, they retrieve several similar cases
from each of the case-bases. Testing on leukemia, colon, and cancer data, Yao
and Li retrieved results that outperformed several classic algorithms, including
a few which used case-based reasoning.
Ramos-Gonzalez et al., [3] used a two-level feature selection process for gene
expression data in squamous cell carcinoma and adenocarcinoma. Their method-
ology has a preliminary feature selection which uses a non-parametric Mann-
Whitney test to locate genes whose expression levels variation are statistically
differentiated between subtypes. Following is a feature selection stage with Gra-
dient Boosted Regression Trees that further refines the feature list into a greatly
reduced subset that still maintains a high classification accuracy. A distance-
based approach is used to retrieve similar cases, while additional diagnostic in-
formation may be requested that assists in correcting the prediction.
More recently, Lamy, Sekar, Guezennec, Bouaud and Seroussi [2] proposed
a CBR method that visualizes results. The CBR system was rather straight-
forward, retrieving cases through a distance measure, though their specialization
was in the explainability. Qualitative attributes between cases were shown us-
ing rainbow boxes, where labeled and colored rectangles extend through columns
that represent the cases, clearly showing what was similar or dissimilar between
cases. Quantitative attributes are provided in scatter plots that center on the
query case and accurately displays the similar cases.
Advantages of CBR are its ability to generalize, and explainability. These
factors will lend to an informative view of the epigenetic state of a cancer sample,
and will hopefully assist in determining the heterogeneity of specific subgroups
of samples.
2 Research Plan
The proposed research project seeks to employ CBR in an investigation of the
epigenetic factors of breast cancer. Feature selection methods will be tested and
evaluated to hone in on highly specific areas of the epigenome that have been
impacted. A CBR framework to classify cancer samples, predict cancer prognoses
and calculate survival is planned, with the underlying pathophysiological impacts
of the cancer being investigated along the way. Prototypical representations of
the the cancer and the clinical subgroups will also be researched.
2.1 Research Aims
1. To construct a case-based reasoning framework for classification of epigenetic
data in breast cancer which takes covariate factors into account. Primary
work here will focus on retrieving similar cases based on clinical and epige-
netic similarity and using previously located labels to classify novel cases.
In areas of dissimilarity, prior cases will be adapted to conform to the novel
case. Integrating clinical factors has been shown to increase prediction abil-
ity (van Vliet et al., 2012) and prognostic performance (Zhu et al., 2017).
It is hypothesized that the inclusion of these factors will lead to greater
heterogeneity of found biomarkers as well as greater biological relevance.
2. To extend the established framework to predicting cancer prognoses. After
the construction of a CBR framework for classification, prediction becomes
a natural and swift process. Here, sample similarities will be retrieved and
used to determine patient outcomes with modifications occurring where its
necessary.
3. To further extend the established framework for survival analyses. Similar to
Aim 1 and 2, similar samples will be retrieved though the goal at this phase
is to locate the epigenetic signatures relevant to prolonged patient survival.
4. To locate deep pathophysiological pathways that have been impacted by
cancer.
5. To establish a prototypical representation of cancer and clinical subgroups.
6. Extend the model for the reuse of prototypes for classification, prediction
and survival analysis.
3 Progress-To-Date
Work was just completed using DNA methylation to classify breast cancer sam-
ples from normal tissue samples. The first stage was to investigate the most
diverse of these cases, stage 4 cancer versus normal tissue. Classification was
performed using naive bayes (NB), random forest (RF), and k-nearest neigh-
bor with 3 iterations of k at a stage after surrogate variable analyses, after
differentially-methylated position analyses, and after differentially-methylated
region analyses. Finally, methylation probes at each genomic region within a
particular gene were averaged and features were selected to find the highest per-
forming genomic regions. The genes with the highest performing genomic regions
were then mapped to KEGG functional pathways and for the top 4 functional
pathways, the associated genes were used to classify a larger set of cancer sam-
ples from a variety of stages to normal tissue. The four pathways were olfaction
transduction, neuroactive ligand-receptor interaction, nicotine addiction, and
GABAergic synapse. Results of this classification process are in Table 1.
Functional Pathway NB RF K1 K2 K3
1. Olfaction Transduction 95.4 95 94.975 95.4 96.45
2. Neuroactive-Ligand 95.92 93.45 96.8 96.27 98.05
3. Nicotine Addiction 94 87.25 97.05 95.95 97.8
4. GABAergic 93.5 88.9 97.05 96.225 97.95
Table 1: Classification results from Naive Bayes, Random Forest and K-Nearest Neigh-
bor with 3 instances of K for genes located in the most associated functional pathways
While this methodology held strong results, all iterations of the dataset suf-
fered from a class-imbalance and whether or not overfitting occurred cannot
yet be deduced. With these issues in mind, it is hopeful that the generation of
a strong prototype through which to compare samples will allow a one-to-one
correspondence that eliminates class-imbalance and strengthens classification
results. If the prototype is able to be visualized, it would expand its strength
and allow for downstream views into which biological mechanisms lend to the
prototype’s accuracy. Further, stage 4 samples were selected to represent a het-
erogeneous group in regards to the epigenetic state, but the small sample size
removed the possibility of separating by clinical factors and still locating mean-
ingful information. It is believed that a case-based reasoning approach would
mitigate these issues and produce stronger results.
References
1. Anaissi, A., Goyal, M., Catchpoole, D.R., Braytee, A., Kennedy, P.J.: Case-
based retrieval framework for gene expression data. Cancer Informatics 14 (2015).
https://doi.org/10.4137/cin.s22371
2. Lamy, J.B., Sekar, B., Guezennec, G., Bouaud, J., Sroussi, B.: Explain-
able artificial intelligence for breast cancer: A visual case-based rea-
soning approach. Artificial Intelligence in Medicine 94, 4253 (2019).
https://doi.org/10.1016/j.artmed.2019.01.001
3. Ramos-Gonzlez, J., Lpez-Snchez, D., Castellanos-Garzn, J.A., Paz, J.F.D., Cor-
chado, J.M.: A cbr framework with gradient boosting based feature selection for
lung cancer subtype classification. Computers in Biology and Medicine 86, 98106
(2017). https://doi.org/10.1016/j.compbiomed.2017.05.010
4. Yao, B., Li, S.: Anmm4cbr: a case-based reasoning method for gene ex-
pression data classification. Algorithms for Molecular Biology 5(1) (2010).
https://doi.org/10.1186/1748-7188-5-14