NETWORK BASED DATA ANALYSIS REPORT

Andrea Policano 24/07/2024

“Analysis of clustering methodologies on
Colorectal Cancer samples”

1. ABSTRACT
When cancer is present usually other complications arise,
such an example is the presence of homeostasis problems in
the patients affected by it. One of the main problems of this
situation is that the pathogenesis related to these problems
is still poorly understood and further characterizations and
analysis are needed. Thanks to the study of Isella C. et al.
(2022) we now know that there is a correlation between the
MET oncogene, to which these homeostasis problems are
highly associated, and the coagulant factor XII(F12), in the
end, effectively finding a correlation between the two
biological components.
By starting off with the same data of Isella C.’s project, a
general analysis regarding the samples and the genes
associated to the data was executed, together with
comparisons between the different applicable methods that
can be used, in addition external softwares such as
STRINGdb were used for further data and information
retrieval and processing.

2. INTRODUCTION
2.1. COLORECTAL CANCER (CRC)
CRC is a type of cancer characterized by the abnormal
growth of cells in the first and longest part of the large
intestine, called “Colon”. Even
though it may become present at
any age, it is frequently present
in age-advanced individuals,
usually starting from the
formation of small cell lumps,
called “polyps” that can be easily
identified inside the colon, they
do not cause symptoms neither Figure 1, Colon section of the
are directly dangerous, although Large Intestine
they possess the potential to become cancerous elements,
their removal is also associated with better recovery against
the cancer itself. Some of the symptoms that may arise and
that can be related to CRC are:
  Change in bowel habits: frequent diarrhoea or
constipation
 Rectal bleeding
 Belly discomfort: characterized by cramps, gas or
general pain
 Weakness
 Loss of blood from the rectum
 Loss of weight
 Bowel un-emptiness: feeling that the bowel is still
empty after peristaltic movements
As stated, one of the main factors associated to CRC is the
old age of the patient, other risk factors that go hand in hand
with it are:
 RACE: the disease seems to have a link to the race of the
individual, especially for Black people in the US
 Family and personal history
 Inflammatory bowel diseases: such as Crohn’s disease
or ulcerative colitis
 Inherited disease
 Low-fiber, high-fat diet: typical Western diet, also
results were seen to be linked with processed and red
meat
 Absence of exercise
 Obesity
 Diabetes
 Smoking
 Alcohol consumption

2.2. GSE52060
The dataset here presented is the one used in the
beforementioned study, it is derived and produced by 46
samples coming from patients presenting Colorectal Cancer
(CRC), from them one sample coming from the neoplastic
tissue and one coming from the normal mucosal tissue were
taken, later on these two subdivisions will become the main
factors for methodology comparisons and analysis.
The GSE file present on GEO (Gene Expression Omnibus) was
fetched and loaded on the virtual environment of Rstudio,
from it pheno-related data was extracted and a metadata
dataframe was built, mainly used to understand additional
information regarding the samples, how they are grouped
and possibly find some potential targets to use as factors for
later stages in analysis. From that we can see that we have
23 samples of Normal mucosa and 23 of Neoplastic tissue,
for a grand total of 46 analysable samples.
Figure 2, Metadata dataframe obtained from the GSE52060 file

2.3. CLUSTERING TECHNIQUES INVOLVED

Throughout the study different clustering techniques will be
applied in order to understand how efficient they are in a
sample classification task, considering the given dataset and
how they will fare individually. Those which will be tested
are: PCA, K-means and Hierarchical clustering, Random
Forest, Linear Discriminant Analysis, LASSO and RSCUDO.
Each of these having their own methodology and way of
working, which will triumph over the others given the
situation?

3. METHODS
3.1. EXPRESSION MATRIX RETRIEVAL
After loading the GSE file and retrieving the metadata
dataframe, the next step in the project was to obtain the
expression matrix of the data present in the file. That was
done swiftly by few command lines. After the acquisition,
results of the expression matrix were box plotted, picturing a
before-and-after situation in the context of a log2

normalization in which an improved organization of the data

can be seen.

Figure 3,
Boxplot of
pre-
normalized
GSE52060
expression
data
 Figure 4,
Boxplot of
pre-
normalized
GSE52060
expression
data
3.2. PRINCIPAL COMPONENT ANALYSIS (PCA)
PCA is a machine learning method focused on the
dimensionality reduction in order to simplify a large dataset
into a smaller one, maintaining all the important information,
pattern and trends. This methodology of course still pays the
price when applied, the reduction itself will cost us accuracy
but in the project’s context it is highly profitable for the
identification of homogeneous subgroup of genes with
similar expression profiles or samples which present akin
trends.
PCA was performed through the use the prcomp command
together with the execution of a t-test on the dataframe used
in order to maximize gene prioritization and feature
selection, following these processes both a summary and a
screeplot were produced together with other graphs
produced using the autoplot function, the colours were
assigned by taking in consideration the column 8 of the
metadata_df variable “source_name_ch1” which showed the
subdivision of the two tissues recovered.

Figure 5, PCA of metadata dataframe using as reference

the two tissue types, in this case the “skeleton” of the
PCA can be seen, together with the absence of obvious
clustering or grouping of the samples due to the lack of
characterization using colours
3.3. HIERARCHICAL AND K-MEANS CUSTERING
Following the creation of the PCA plot, the next step of the
analysis revolved around the use of clustering methods for
further information retrieval and processing.

Figure 6, Hierarchical clustering raw plot, showing the respective distance

measures for each sample and hints on their relationships

3.3.1. CLUSTERING
It is an unsupervised machine learning methodology, in
which no class values hinting at a priori grouping of the data
are given. It focuses on finding similarity groups in a data set
and it is a daily-life technique that is commonly put in
practice, some of its day-to-day applications are for example:
grouping clothes, objects, videogames, tools, documents or
even target possible marketing subjects.
It is characterized by an algorithm in which a distance is
contained, the quality of a clustering process is measured by
taking in consideration some the data, the distance function
and in general the algorithm itself, as in this case we ought
to consider a trade-off between accuracy/quality and
computational feasibility.

3.3.2. K-MEANS CLUSTERING

K-means is a partitional clustering methodology used to
assign data points to one of the K clusters depending on the
distance of the points themselves. The algorithm starts by
randomly assigning cluster centroids in the space and
grouping the datapoints to them, while taking in
consideration the distance from the centre of the cluster.
After this point new centroids are selected and the process is
repeated iteratively, until the change in the Sum of Squared
Errors (SSE) goes below a certain threshold, or the iterations
are completed.
It is highly effective and time-convenient, but it also presents
its downsides, such as the need of an optimal K to choose in
much more complicated case, when few clusters are not
suitable for the study, or the need to define the mean and its
sensitivity to outliers but it is highly effective in many
situations.
In the project context the “K-means” algorithm was used in
order to distinguish the samples using the tissue types to
which they belonged, the K-value used was 2 as the
maximum number of tissues. This step produced a plot
similar to that of the PCA one, with the presence of the label
of the samples and the same colouring used in the phase
before.

3.3.3. HIERARCHICAL CLUSTERING

As for the one before is a clustering algorithm, with the only
difference being the use, in the Hierarchical, of a distance
matrix, in which the distance measure can be defined by
different methods such as: the Euclidian, Manhattan or the
Minkowski distance, with the last being used only in specific
cases. Other than this we ought to take in consideration also
the production of a different end result with respect to the k-
mean one, by applying this methodology we obtain a tree
called dendrogram with the samples taking the part of the
leaves.
Hierarchical clustering can function both as a bottom-up or a
top-down approach, using either a single or complete-linkage
approach. In the project this technique was used to produce
different dendrograms to see how the samples are related,
taking into account, as in the steps before, their tissue
subdivisions. In this case the main command used in this
phase are: hclust for the tree production with the
methodology “ave”; cutree for squads/groups production
using the metadata; rect.hclust for dendrogram production
with a border identifying the two groups and finally the
production of the dendrogram of the samples with the
branches colours related to the tissue types.
Figure 7, One of the dendrogram produced by applying Hierarchical
Clustering, here is one the three dendrogram produced and represent the
starting point for the others production

3.4. HEATMAP
A heatmap is a 2-dimensional data visualization technique in
which the data are represented through the use of a gradient
of colours, usually using a darker one for greater quantity of
that value, the main aim of this procedure is that of finding
how the samples are correlated between each other and to
exploit these relationships in future steps.
As for other common methodologies it can be used in many
different fields, going from criminology to gene-expression
analysis, with lots of different type of it depending on the
study type in which the methodology is applied.
The first thing done before the heatmap production was
creating a distance matrix with the data coming from the
previous step, along with the setting of a colour palette that
will be needed for the final plot. The pheatmap function was
used for the creation of the heatmap using as main data the
distance matrix, while incorporating the already available
distance_mat variable obtained in the Hierarchical Clustering
step, and the palette of colours.

3.5. RANDOM FORESTS

Even though they are not directly correlated to clustering
algorithms, they still play a role in classification and
regression operations in machine learning models. Making
use of the ensemble mechanism “bagging” they are based
on the production of many decision trees which will later be
merged in order to obtain a final model able to have a much
more accurate output value, increasing the quality of the
learning model. An additional feature that separates this
type of algorithm from the rest is the use of labelled data,
thus showing how its origin is related to the supervised
learning typology.
Its application in our project starts with the creation of a new
dataset that will be specifically used only for this phase of
the project, together with it the conditions for the dataset’s
related samples were obtained and set from the metadata_df
obtained in the first part of the analysis, the conditions are
the same that we used for the other cases: Neoplastic and
Normal mucosal tissues. Following along the model was
produced using the randomForest function, later on, graph
related to the model were also produced, focusing on the
most variable relevance(fig.6) and the main probes useful for
the model quality(fig.7).

Figure 8, Plot showing the relevance of the first samples of the used
dataframe in the quality of the model, meaning that the only the starting
samples are relevant for the model

Figure 9, Plot showing which are the probe having the highest
relevance/importance in the model created, with the addition of the
visualization of how much the loss of quality will affect the model if that
probe were to be modified
3.6. LINEAR DISCRIMINANT ANALYSIS (LDA)
LDA is a supervised machine learning approach that aims to
solve multi-class classification problems by separating them
through data dimensionality reduction. It is especially useful
to optimize machine learning models.
This technique is able to make predictions by using Bayes
theorem and calculate the probability of a particular data set
to belong to an output, while also modelling the data
distribution. It focuses on maximizing the between-class
distance and minimizing the within-class one, in addition LDA
works by identifying a linear combination of features that can
be used to separate two or more classes of objects. It is done
by projecting the data into a one-dimensional graph for
easier classification, proving a high level of versatility also
for multi-class data.
In order to evaluate the classifier obtained through the LDA
the ROC and CARET methodologies were used. At the
beginning of this step the creation of factors "tissue type:
adjacent normal mucosa (N)" and "tissue type: neoplastic
tissue (T)" was completed, after that t-tests were conducted
on the expression dataframe using the function rowttests
while also accounting for the new factors, extraction of
probes having a p-value lower than 0.05 followed and a new
dataframe was built, additionally a column was inserted in
the df, called “AFFECTED” hinting whether or not the sample
is derived from the neoplastic tissue.

Figure 10, LDA derived plot showing

the separation between the samples
clusters, here are visible some
elements that may result in outliers

Following this path the model went under training using the
function lda, taking as input parameters:
 AFFECTED (as Function): needed for understanding
which are the variables/samples that we want to guess
as having neoplastic tissue against the normal
mucosal one
 The dataframe created before that has built-in the
factors of interest
 The prior probability of each class
 The subset of elements that we want to train
After these steps were completed, the mod.values were
predicted using the function predict coming from the , taking
as input the model built and the dataframe’s training subset.

3.6.1. CARET
After completing the first part of the LDA phase, the Caret
methodology was implemented. It provides a wide range of
functions focused on data preparation, modelling and
evaluation. With the objective to streamline the evaluation
and building process of predictive models, the package used
also includes functions related to data preparation, modelling
and evaluation.
CARET mainly works by working on Hold-out specific samples
and fitting the model on the remaining ones, in the end the
average performance is calculated across all the hold-out
predictions and the determinations of the optimal
parameters set is given. When starting with the application
of CARET, a control group and a metric needed to be defined:
1. Control  built by the trainControl function using as
input:
a. Method: “cv”, setting the resampling method to
cross-validation
b. Number: “10”, indicating the 10-fold cross-
validation
c. Repeats: “NA”, iteration of the methods
2. Metric  set as “Accuracy”
Later training of the two models was conducted by using the
train function using the lda and rf methodologies and in both
cases considering also the metric established. Results were
obtained using also the resamples function and ggplot for
the representations.

3.6.2. ROC CURVE

The ROC Curve is a probability graph able to show the
classification model performance considering all the
classification thresholds, potting two parameters: TPR (True
Positive Rate) vs FPR (False Positive Rate), at different
thresholds meanwhile it also tries to separate the “signal”
from the potential “noise” detected.
It is used mainly to evaluate the performance of a binary
classifier considering the two metrics already cited:
 TPR  Sensitivity, proportion of correctly identified
positives that are actual ones
 FPR  Measuring the proportion of actual negatives
incorrectly identified by the model
Results in this part of the experiment by simply redoing a
prediction of the model, using in this case the test subset,
and plotting thir results using the plot.roc function. In
addition to that the value of AUC (Area Under the Curve) was
calculated.

3.7. LASSO/RIDGE REGRESSION

These two are very famous algorithms heavily used in the
field of machine learning, both of them are used to reduce
errors deriving from a linear regression model, with each one
presenting a different key feature that characterize it.
 RIDGE REGRESSION  introduces a regularized
(penalty) term (λ) in the cost function in order to
prevent overfitting, it can reduce all the coefficients
by a small amount
 LASSO REGRESSION  also in this case the penalty
term is present but in this case is regularization L1
instead of L2 like in the RIDGE case. L1 represent the
sum of the absolute values of the coefficients then
multiplied by a constant λ, in this case the main
objective is to reduce the features.

Figure 11, Plot showing the trend of the quality of the model depending on
the chosen regularization parameter

During the study LASSO was the used algorithm instead of

the RIDGE due to problems and anomalies related to the
data sets. In this case we reutilize the function train already
implemented in the CARET steps changing some of the input
parameters:
1. Function  AFFECTED, recalling the LDA methods
2. Data  the model matrix deriving from LDA
3. Method  glmnet
 4. Family binomial
5. Alpha  1, indicating the penalty for LASSO
6. Control  control, recalling the CARET section
7. Metric  the metric already produced in the previous
part of the project
In the end, plots regarding the relationship between
coefficient, binomial deviance and log lambda were produced
and the model was compared to the ones obtained from RF
and LDA.

3.8. RSCUDO
After the implementation of LASSO, the next algorithm taken
into consideration and used in the project was RSCUDO. It
implements decision trees in order to identify robust
subgroups through iterative clustering, it is mainly used in
the biological context, focusing specifically on
transcriptomics and genomic data.
In this project the whole phase starts with the factorization
(all over again) of the conditions of the tissue and creation of
both a train and test group, the following step led to the
training through the specific scudoTrain function, signatures
check was the successive event in the phase, subsequently
the creation of different networks took place. The criteria for
the differentiation of the cluster created was the type of
subdata used in the conditions and a classification procedure
carried out only in the final graph, based on the factorized
in_training data.

Figure 12, Example of network built while using the RSCUDO package,
here can be seen the main two clusters and some elements "on the line"
between them

3.9. PATHFINDR
With the implementation on the starting dataset of all the
wanted clustering/classification algorithms, pathfindR was
applied in order to study the enrichment results of the used
genes in the dataset. Before doing any type of operation in
this scenario a filtered list (p-value < 0.01) of genes was
extracted from the df, which will later be used in the manual
enrichment carried out using the software enrichR available
online (see 3.10.)
Once done with the extraction, the process for pathfinder
application begins. The starting point is represented by the
developing of a model matrix based on the conditions of the
tissue, followed by a “fit model” operation and production of
a contrast matrix, later on an analysis on said matrix was
carried out using limma, from which we could obtain the list
of probes of interest, further filtered by p-value (<0.01).
Until now all the operations that were enacted had a role in
the formation of the final dataframe that would later be used
for the effective run of pathfinder, still before that a merge
between the results obtained until now and another df,
created through the use of both probeID and their respective
gene symbols, associated through the use of ensemble and
biomart.
Subsequently to the merging, different databases were opted
to run the algorithm, in the end 3 of them were chosen:
KEGG, GO-BP, Reactome. With everything done and ready to
go pathfinder was applied through the run_pathfindR
function for each of the selected databases, together with
the upcoming creation of a plot for each case.

3.10. ENRICHMENT
After obtaining the df having the ILLUMINA probes, a new
empty dataframe was built. From previous analysis the logFC
and the p-value of each probe were extracted and
implemented in the new structure, in addition a column
having the gene symbol of the probes was created through
the use of the “illuminaHumanv3.db” and “Annotation.dbi”
libraries and the mapIDs function, able map the probe id to
the respective gene symbol using the given database, the
inputs were as such:
 Database  “illuminaHumanv3.db”
 Keys  “probe_IDs” identify the list of elements that
you want to convert, in this case the column having
the probes ID
 Columns  “SYMBOL”, which type of column to take
in consideration for the retrieval from the db
 Keytype  “PROBID”, indicating what is the object
that you want to be converted
Concluding the mapping phase, a df having as columns both
the probe and gene symbol IDs were obtained in the end,
leading to the merging of the newly obtained element with a
modified, t-tested df retrieved from the starting data and
metadata. Concluding the merging between the two we will
have a new structure in which the IDs will be present
together with statistics such as p-value, logFC and dm.
From that point the objective is to retrieve the gene symbols
ID having p-value < 0.05, therefore further filtering is
needed. Completed this procedure, a list of the gene symbol
is extracted from the dataframe and uploaded on the online
software of EnrichR, available at the maayanlab.cloud
website. After the insertion of the list of gene symbol IDs, the
main categories taken in consideration for the enrichment
were: Transcription, Pathways, Ontologies and
Diseases/Drugs.

3.11. STRING DB
As the final step of the project, the list of gene symbols
previously used in the enrichment analysis was, first of
all, sorted according to the p-value in an ascending
way, and later the top 150 rows were selected and
translated to their respective UNIPROT ID with the aim
of subjecting them to STRINGdb analysis, in order to
find possible relationships and insights that can be used
for future studies. The software itself is able to produce
highly informative networks using the provided list of
IDs and the information coming from many different
databases, with the possibility of choosing the organism
of preference.
In the end a network of 133 elements was built,
showing how each node (protein) interacts with its
neighbours, their relationship and if there are any
solitary nodes.
4. RESULTS
4.1. PCA
Following the procedures applied in the Principal Component
Analysis step (see Methods) plots related to the sample
clustering and distribution were produced. Regarding this
last one, everything seems to be fine, no errors nor unusual
elements or modification were found on the other end the
PCA plot, when produced, presented one only issue.
Figure 13, Plot deriving from the PCA phase, here almost two perfectly
divided groups can be seen except for two datapoints in the bottom part
of the graph

In Figure 13 we are able to see how the samples used can be

separated in an almost perfect manner taking into account
the tissue type to which they belong, the only exception is
the presence of two datapoints, later confirmed to be
GSM1258081 and GSM1258082, which seem to have been
mixed up during the process, this hypothesis was proposed
due to them being the only anomalies in the procedures. A
better visualization of the “overlapping” between the two
groups can be seen through Figure 14:

Figure 14, PCA plot showing a better

clustering divisions of the two
groups, in it the overlapping
between the two of them can be
easily seen

4.2. K-MEANS AND HIERARCHICAL CLUSTERING

After completing the PCA procedure, different clustering

algorithms were applied on the sample study, leading to the
correct grouping of each element to its respective class. The
results can be seen in Figure 15 and Figure 16, where the
labelled samples do not present overlapping.
Figure 15, K-means plot obtained in step 3.2., it presents the two classes
and the labelled samples belonging to each one, together with the GSM ID
of each sample

Figure 16, K-means algorithm plot showing the frame of the clustering
groups, no overlapping is in sight this time

Using the algorithms applied in the previous step we were

able to effectively produce a perfect division of the two
groups without the presence of any errors or anomalies,
solving the initial problem presented in the PCA algorithm. In
Figure 15,16 the red cluster is characterized by the
Neoplastic Tissue, meanwhile the blue is representative of
the Normal Mucosa tissue type.
 Subsequently to the application of the K-means algorithm,
the same test/trial was carried out using the Hierarchical
clustering methodology, which resulted also in this case in
the perfect subdivision of the two sample groups (check

Figure 17 and Figure 18)

4.3. HEATMAP
After obtaining the modified distance matrix using the
already available data from the clustering phase, an

Figure 18, A better visualization of the

Figure 17, Dendrogram plot produced previous dendrogram with the
with the hierarchical clustering algorithm, branches coloured according to the
the two marking in red are used for group: B for Neoplastic and R for
visualization purposes in order to Normal Mucosa, and the leaves having
distinguish between the two groups their GSM ID
heatmap was produced:

Figure 19, Heatmap presenting the comparison between each sample

From the plot we can see that with an inner-group

comparison, the values obtained are quite low, on the other
end doing an inter-group comparison leads to high value of
correlation between the samples compared, maybe
underlying possible relationships between different sample’s
interactions that may be targeted for further and more
specific analyses, while possibly focusing at the gene-
expression level to denote key differences or common
points/grounds that can be exploited for future studies.

4.4. LDA
During part 3.6. procedures related to the Linear
Discriminant Analysis were applied, one of the first plot
produced was that related to the quality evaluation of the
LDA algorithm in both the training and test group.
Figure 20, LDA quality evaluation in both [UP] Normal Mucosa group and
[DOWN] Neoplastic group, the X axis represents the division point
between the two groups.

A fairly good classification was done by the LDA with the

presence of few outliers mainly present in the control group.
A possible explanation, also in order to reconnect to what
was already noted from previous results, is that the outliers
represent the nodes that were later fixed using the k-means
clustering algorithm in step 3.3., therefore possibly hinting
again to a certain relationship between these samples,
considering also the possibility of erroneous data collection
and elaboration. Further analysis and consideration shall be
taken if the aim is that of solve this doubt.

Additional evaluating operations regarding the

clustering/grouping of the sample while using LDA were
executed:
Figure 21, LDA clustering plot, in [blue] we have the samples related to
the control group (Normal Mucosa), in [green] we can see the samples
belonging to the Neoplastic tissue group

As we can see also in this case an almost perfect grouping of

the samples is done except for the one that are already
known to cause some anomalies, referring, for example, to
data point n.32 in the bottom right part of the graph, and
using the information we already know we have the
possibility to identify the outlier as the sample having the ID
GSM1258090.
Additionally, in order to understand how the LDA faired in
this test, the trade-off between specificity and sensitivity was
calculated, with a final AUC value of 0.88, demonstrating
that the model itself does a good job when it comes to
classify examples in this particular situation, not perfect but
good enough.

Figure 22, ROC curve and representations of the AUC of the LDA
methodology
Following the AUC calculation, the next phase of evaluation
concerns the comparison of the models used (RF and LDA) in
terms of Accuracy. Using model prediction functions, the
following results related to their efficacy were obtained
(check Figure 23):

Figure 23, comparison between the application of LDA against RF, as we

can see the accuracy of LDA is much higher than expected, with both its
mean and IQR being grater than the RF model

By looking at the graph we are able to see that the accuracy

of the LDA model is still high, but it does not seem to be due
to the presence of overfitting of the training dataset, for the
RF, instead, the accuracy seems to be much more in
standard ranges, the causes of this event are still no known
in our case.
As a supplementary operation, the application of the CARET
methodology was applied so as to se how both the models
would act in presence of more trials. The results showed that
in an opposite respect to what we have seen before, with the
implementation of more trials the best model to apply is
Ranndom Forest instead of LDA (see Figure 24):
Figure 24, Plot showing the comparison of the Accuracy of LDA vs RF, in
the case of CARET application

4.5. LASSO/RIDGE REGRESSION

During the 3.7 phase of the project, different results and
plots were obtained when the LASSO methodology was
applied, starting off with the plot needed in order for us to
understand which was the best λ value to use for producing
the lowest MSE value, with λ being equal to -2.46

Figure 25, plot showing the lambda value such that it is able to minimize
the binomial deviation, obtained in the LASSO analysis

Afterwards a plot representing the relationship between

regularization parameter specific for LASSO and the model
accuracy was drawn. From that graph we can gain some
insights regarding how the reg. parameter can be inserted in
the model while also preserving the quality of the
classification, showing also that with the increase of its value
the model itself probably becomes too much simplistic and is
not able to work as well as before (look up Figure 26).
Figure 26, Plot representing the relationship between the accuracy of the
model and the value of the regularization parameter to insert in it,
increasing the parameter means lowering the accuracy of the model

As the final step in this stage of the experiment, the model

obtained through the application of the LASSO methodology
was compared to those of RF and LDA, leading to the LASSO
model resulting as far worse to apply to the initial dataframe
respect to its two colleagues.

Figure 27, Comparison plot between the three methodologies used for
clustering/classification procedures, when implementing the LASSO model,
the accuracy of the classification greatly decreases in comparison to RF or
LDA

4.6. RSCUDO
Concluding the LASSO classification procedure, the RSCUDO
methodology was applied. The first results obtained showed
the clustering of the tissues groups in two main clusters with
the presence of sparse datapoints, with some of them still
connected, even if on their own. In Figure 28 we are able to
see what was mentioned above, in addition to that, it is clear
the existence of some nodes that seem to “trespass”
between the domains of the two key clusters, with a higher
probability of them being outliers that were considered
related to a group to which they are not actually associated
with. The following classification was focused on looking up
for the Signatures of the tissue’s groups:

Figure 28, Clustering plot obtained through the application of the

RSCUDO methodology, in this representation the two core clusters
can be seen together with a few of other sparse elements which
seem to intrude between each other’s domains.

A secondary classification procedure was carried out, with

respect to the first one, with the objective of further
decomposing the dataset, in order to create much more
subclusters. The aim here is to see how the algorithm
considers the “solitary” points already seen in Figure 28, on
top of the presence of possible changes in the network due
to these new applied requirements:

Figure 29, Clustering plot showing how the RSCUDO algorithm decompose
and recognise the different solitary nodes already seen in the previous
graph, here we are able to see the subgroups perceived by it
4.7. ENRICHMENT
After obtaining the gene list the result underwent enrichment
through the use of the online platform of EnrichR, as
specified in the Methods section 3.10., out of the ~13300
genes, only a fraction of them were found to have a
description in the online software, with a number around the
~9000 genes. In the end the enriched genes were around
~4000.The results coming from this step of the analysis
presented some associations with:
 Translation
 Influenza infection
 RNA and Ribosome modifications, formation and
binding
 Diamond-Blackfan anaemia
 Belong to the top 500 genes that are downregulated in
COVID-19
 Glycosylation-related disorders
 Bladder, Ovarian, Testicular carcinoma
 Other different types of aneamia
 Presence of bulk tissue in the kidney
Quite not what we were expecting, but as stated before
these results may be due to the high amount of filtering that
was imposed on the genes and the presence of duplicates
among them. eventually, with further research, some of the
pathway in which with elements were enriched were related
to the CRC context, for example:
 Colorectal carcinoma
 Colorectal neoplasm
 Acute myeloblastic leukemia
 Carcinoembryonic antigen
 Muir-Torre syndrome
 High association to the tissue of intestine and uterus
 Gut microbiota beta diversity
 Colon cancer association and intestine epithelial cells
 CL34, HT115, SNU16 and SNUC1 cell lineages of the
Large Intestine
 Presence of bulk tissue in the colon
 Association with HCT116 cell line

Of particular interest was the finding related to the Muir-Torre

syndrome, an autosomal dominant phenotypic variant of
hereditary non-polyposis colorectal cancer, characterized by
the presence of sebaceous tumours of the epithelial cells of
the intestine, usually accompanied by also colonic
carcinoma. The genetic modifications related to the
syndrome upbringing are known but still needs to be further
refined and could be an interesting point of start for a study.
Figure 30, Orphanet Augmented 2021 section of the enrichment approach
applied on the filtered genes, here are the argument presented by the
already cited section in which the protagonist is the Muir-Torre syndrome,
a form of colorectal cancer, non polypolis, characterized by the presence
of sebaceous tumours

In the end both expected and unexpected data were

recovered, with some of them with a higher possibility of
being used for the treatment of the disease, even though
more information are needed for a better characterization of
the condition and for better assessment of the weak point
that can be used against CRC, even though low relevance to
the disease was found from the obtained genes.

4.8. PATHFINDR
With the application of the PATHFINDR algorithm,
supplementary information regarding insights on the
enrichment of the genes used before, it showed much similar
results to those obtained using the EnrichR instrument. Such
an example is present in Figure 31, where the plot introduced
showed high level of relationship between the gene list
inserted and their respective relevance in ribosomal
functions, modification and structural features, which were
also present in the previous phase.
Figure 31, Enrichment plot that presents the results coming from the base
KEGG database used in PATHFINDR, here the results are highly correlated
to the one found in the previous phase using the EnrichR online
instrument

The results obtained from this phase mirror those coming

from the EnrichR online tool, with the only exception that
with the produced plots we are not able to see any relevance
to the Colorectal region from which the tissues were
obtained. Additional plots were produced:

Figure 32, Enrichment ploit produced by PATHFINDR in which the genes

are enriched with respect to Biological Process information
Figure 33, Enrichment plot related to the Reactome database, obtained
through the PATHFINDR package

4.9. STRING DB
Once the gene list from the merged dataframe was obtained,
the first result obtained from this step was a list of the
respectively translated genes into a suitable format for
STRING DB, that being the respective protein name for each
respective gene. The final completed list was then submitted
into the online software, with also the specification of the
organism of interest (Homo Sapiens). In the end a network
was generated showing how the elements are connected:

Figure 8, STRINGDB- derived plot presenting the main connected cluster

with few sparse solitary nodes
The list of the main connected components present in the
central cluster was obtained manually and was later put also
under enrichment for further investigation, those being:
CEP55, UBE2T, MND1, CKS2, ASPM, CCNA2, EXO1, CCNB1,
NCAPG, TPX2, MCM6, UBE2C, BUB1, GINS2, RFC3, MCM10,
KIF2C, KIF4A, TRIP13, RACGAP1, CENPW, SKA3, FAM83D,
CDK4, GMPS.
Results showed enrichment in gut microbiota association,
together with melanoma, gallbladder and renal cortical
glands diseases, connection to Huntington disease,
relationship with neurodegenerative disease, liver carcinoma
and Goldberg-Shprintzen megacolon syndrome.

5.0. FINAL THOUGTHS

Since the beginning this study aimed to focus on the
comparison of different methodologies to classify the most
correctly as possible samples coming from tissues known to
be related to the colorectal cancer affliction, while
considering their effectiveness and efficiency, trying in the
meantime to avoid errors and biases introduction as much as
possible, considering also each methodology behaviour with
the “perceived” presence of outliers. We now know that their
quality depends on many factors, such as the number of
trials introduced the number of samples, in example when
faced with multiple tests, the Random Forest methodology
showed the best accuracy mean and range, in the case of
the single run, instead, LDA reign over the others, in the
meanwhile LASSO was the single methodology which
managed to score the lowest accuracy. In addition, RSCUDO
clustering showed how the samples behaved and their
relationship between each other, clearing the path for the
further analysis that took place throughout the final sections
of the project
Even though the data in the end was heavily filtered, the
information obtained from the enrichment approach were
still significant and further attempts may be ideal for better
assessment of possible study-target for CRC treatment.
Unfortunately, some information are still missing, such has
been seen when using the STRING database with the
presence of some stand-alone nodes, demonstrating the
need for more advanced characterization is required.