Researchers seek to identify biological markers which accurately di erentiate cancer subtypes and their severity from normal controls. One such biomarker, DNA methylation, has recently become more prevalent in genetic research studies in oncology. This project seeks to apply the innovative and adaptive machine learning methodology in case-based reasoning (CBR) to examine DNA methylation levels in breast cancer. Instead of relying on a generalized knowledge-base, CBR uses highly speci c information extracted from similar cases which can also greatly expedite the process of nding 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.
The term epigenetics was rst introduced into modern biology by Conrad Waddington as a means of de ning interactions between genes and their products that result in phenotypic variations. Waddington's landscape presents a cell becoming more di erentiated as time goes on. One of the events that can cause this di erentiation 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 methylated 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 references 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 e ects 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 stable 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 decreased (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
applications using microarray data. Anaissi, Goyal, Catchpoole, Braytee, and Kennedy
[
A bit unorthodox, Yao and Li, [
Ramos-Gonzalez et al., [
More recently, Lamy, Sekar, Guezennec, Bouaud and Seroussi [
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 speci c subgroups of samples.
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 speci c 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 classi cation 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 epigenetic 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 ability (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 classi cation, prediction becomes a natural and swift process. Here, sample similarities will be retrieved and used to determine patient outcomes with modi cations 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 classi cation, prediction and survival analysis. 3
Work was just completed using DNA methylation to classify breast cancer samples from normal tissue samples. The rst stage was to investigate the most diverse of these cases, stage 4 cancer versus normal tissue. Classi cation was performed using naive bayes (NB), random forest (RF), and k-nearest neighbor with 3 iterations of k at a stage after surrogate variable analyses, after di erentially-methylated position analyses, and after di erentially-methylated region analyses. Finally, methylation probes at each genomic region within a particular gene were averaged and features were selected to nd the highest performing 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 samples 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 classi cation process are in Table 1.
While this methodology held strong results, all iterations of the dataset suffered from a class-imbalance and whether or not over tting 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 classi cation 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 heterogeneous group in regards to the epigenetic state, but the small sample size removed the possibility of separating by clinical factors and still locating meaningful information. It is believed that a case-based reasoning approach would mitigate these issues and produce stronger results.