bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024.

The copyright holder for this preprint

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

gen3DNet: An R Package for Generating 3D Network Models

Paul Morrison2,, Tina Tang2, Charles Lu1,2, Shamim Mollah1,2
1
Department of Genetics, Washington University School of Medicine St. Louis, MO 63110, USA.
2
Institute for Informatics and Data Science, Washington University School of Medicine St. Louis, MO 63110, USA.

*
Corresponding author
*
[email protected]

Abstract
Motivation: Networks are ubiquitous to organize and represent data in the real world such as the
biomedical networks, social networks, etc. Due to the attractive property of networks which can
explicitly capture the rich relationships among data samples, there is an increasing demand to
transform linkage-free data to graph-structure data for efficient downstream analyses. Existing
works typically focus on representing a single data object like gene expression profile in a
homogeneous 2D network, which fail to deal with situations where two different data objects are
involved to create a 3D heterogeneous network.
Results: In this paper, we introduce an R package, gen3DNet (a generic version of the iPhDNet), for
generating 3D network models from two correlated objects with shared common factors.
Specifically, gen3DNet builds the relationships between samples and shared factors using the non-
negative matrix factorization, where three clustering techniques are evaluated to determine the
number of functional modules. In addition, it builds the relationships between samples from the two
distinct data objects based on a partial least squares regression model. Usage of the package is
illustrated through a real-world application.
GitHub URL Source code is available at https://github.com/mollahlab/gen3DNet

Introduction
Many real-world data and complex systems are in the form of networks, such as protein-protein
interaction networks , gene regulatory networks , document networks , etc The structure and
function of complex networks. The structure of scientific collaboration networks [1–4]. A typical
network model can be represented as 𝐺 = (𝑉, 𝐸), where 𝑉 ∈ 𝑅* is a set of 𝑛 unique vertices and
𝐸 ∈ 𝑅*×* is a set of edges capturing the pairwise relationships among vertices. Compared with the
plain linkage-free data, data organized in networks have many attractive merits. First, the complex
relevance and dependency relationships between data instances can be explicitly reflected by the
edge connections. In addition to the explicit relations, networks are able to capture rich intermediate
bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

or implicit correlations between instances through the relationship propagation over the entire
topology structure, i.e., two vertices with similar structure tend to be relevant. Another appealing
factor is that data represented in networks are beneficial for efficient downstream applications, e.g.,
clustering , classification , prediction and visualization tasks, with the support of many off-the-shelf
network-based learning algorithms [5, 6] . As a result, the past few years have seen a plethora of
works which seek to transform linkage-free data to the equivalent network-structured data (a.k.a
network construction). Specially, network construction has attracted tremendous attention in the
biomedical domain. For examples, gene regulatory network construction aims at modeling the
complex regulatory activities that may occur among genes by exploring the gene expression data [4,
7–9]. Gene co-expression network construction aims to build the pairwise transcript–transcript
associations among genes by calculating their co-expression values [10]. Protein-protein interaction
network construction aims to build the high-confident correlation relationships among proteins in
various organisms using the high-throughput techniques [11, 12]. One common point for these
network constructions lie in that they try to derive a 2D and homogeneous network model from the
single data object such as gene expression and protein-protein interaction observations, which fails
to consider situations in which two distinct correlated data objects are involved to create a 3D
heterogeneous network. For instance, in the complex biological system of breast cancer
microenvironment , protein regulations and histone regulations in the extracellular matrix can be
viewed as two different data objects which are causally mediated by the shared growth promoting
factors. Analogically, in the service computing domain, Mashups and APIs can be viewed as two
different objects which are associated with the shared functional category factors. In general, the
challenge of constructing a 3D network model from two different but correlated data objects is
twofold: 1) the relationship characterization between the data object and shared factor; and 2) the
relationship characterization between the two distinct data objects in a unified network model.

Figure 1: Implementation framework of the proposed gen3DNet.

Here, we extends a recently proposed network model iPhDNet [4] (e.g., capturing the interactions
among peptides, histons and drugs) to introduce gen3DNet encapsulated in an R package for
generating 3D network models from the two input data objects with shared relevant factors. Figure
1 shows the high-level implementation scheme of the proposed gen3DNet which takes two data
objects as inputs. First, it performs the clustering on the first data object, e.g., using Non-negative
Matrix Factorization (NMF) [13], to obtain the linkage coefficients between samples from the first
object and shared factors, where 𝐾 clusters (e.g., 𝐶/ , 𝐶0 , ⋯ , 𝐶2 ) are generated for all samples and
bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

shared factors respectively. We implemented three methods to determine the optimal selection of
𝐾, including NbClust [14] , Cophenetic [15], Silhouette [16] based techniques and evaluation metrics.
Meanwhile, the linkage coefficients between samples from the two different objects are obtained
based on the Partial Least Squares Regression (PLSR) model [17]. Finally, the two interaction
networks generated in above two steps are integrated to a single 3D network.
The rest of paper is organized as follows. In the second section, we briefly describe the network
modeling based on NMF and PLSR techniques. In the third section, we present the gen3DNet package
with examples. In the forth section, we discuss the analysis results of a real-world application in the
bioinformatics area based on gen3DNet. Finally, the last section concludes the paper.

Materials and Methods

Let the two input data objects be represented as 𝑀/ ∈ 𝑅4×5 and 𝑀0 ∈ 𝑅*×5 , where 𝑚 and 𝑛 are
respectively the numbers of samples from the first and second objects, and 𝑓 is the number of shared
factors between the two objects. First, the relationships between the samples and shared factors are
calculated from the first data object 𝑀/ , which involves a non-negative matrix factorization is
performed as:
𝑀/ ≈ 𝑊𝐻; (1)

where samples and shared factors are mapped in a 𝑘-dimensional latent space represented in 𝑊 ∈
𝑅4×2 and 𝐻 ∈ 𝑅*×2 , respectively. 𝑘 is a hyper-parameter which can be determined by the three
evaluation methods including NbClust, Cophenetic and Silhouette shown in Figure 1. It is necessary
to note that in order to guarantee the non-negativity of the input matrix in NMF, 𝑀/ has been
previously log-transformed which is a prevalent practice in the biomedical domain . Then, all samples
are grouped into 𝑘 clusters based on 𝑊 in which each sample is assigned with a cluster with the
largest membership value across the 𝑘 latent dimensions, i.e., each of the 𝑘 latent space dimensions
corresponds to a cluster. The same clustering process applies to all shared factors based on 𝐻. Finally,
samples and shared factors from the identical clusters build the pairwise edge connections, where
the weights on the edges are given by the corresponding coefficient values in 𝐻, i.e., assume sample
𝑠@ and factor 𝑝B are in the 𝑘 CD cluster, then the connection weight between 𝑠@ and 𝑝B equals to 𝐻B,2 .

Then, the relationships between samples from the two different objects are calculated based on the
PLSR model:
𝑀0 = 𝑊EFG5 𝑀/ + 𝐸 (2)

where 𝑊EFG5 ∈ 𝑅*×4 is the coefficient matrix revealing the contributions of the first object 𝑀/ to
the formulation of the second object 𝑀0 , and 𝐸 is the residual or error matrix for the PLSR model.
We then build the connection between samples from object 1 and samples from objects 2, where
the edge weights are given by 𝑊EFG5 . By integrating the output networks from the NMF and PLSR
modelings, a 3D network model can be finally produced shown in Figure 1.
bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Default k selection algorithm

An important component of iPhDNet [4] is the number of clusters present within the data, but there
is no single domain-based method for choosing this parameter. For this reason, we undertook an
evaluation of some common k-selection algorithms to determine which can be expected to
generalize best. To do this, we began by generating hundreds of synthetic datasets, each created to
have a specific number of clusters, ran the candidate algorithms on these datasets, and observed the
correlation between the true and predicted k values under varying conditions. Based on these
results, we chose a combination of the Ward clustering algorithm and KL index (implemented in the
NbClust library) as the default algorithm for k selection in gen3DNet.
Each dataset was generated by starting with k cluster centers, and then adding other data points in
a Gaussian distribution around each center, with the standard deviation of this distribution being an
important parameter. Another parameter was the number of dimensions (columns) used in the
dataset. Both of these were varied using the following values:
• Options for stdev. within / stdev. between: 1, 3, 5, 7, 9, 11.

• Options for the true number of clusters: 2, 3, 4, 5, 6, 7, 8, 9, 10.

• Options for the number of dimensions: 2, 4, 8, 16, 32.

The number of samples (rows) in each artificial dataset was left constant (at 1000). Examples of
sample datasets are shown in Figure 2a, 2b. Both have the same number of clusters, but this harder
to tell in Figure 2b because the variation within clusters is similar to the variation between clusters.

(a) Ground truth: 3 clusters (b) Ground truth: 3 clusters

Figure 2: Example of simulated ground truth data.

The detailed dependency relations between modules in gen3DNet is depicted in Figure 3 .

bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

Figure 3: The dependency relations between modules in the gen3DNet package.

The following candidate k-selection methods were used:

1. Cophenetic correlation
Given a hierarchical classification, the cophenetic correlation is the correlation between the true
distances and dendrogrammatic distances of each pair of data points.
As the number of clusters increases, the cophenetic correlation continues to improve, so choosing
the maximum will favor the maximum number of clusters. For that reason, a separate algorithm is
needed to find the "knee", where increasing k any further has no benefit. For this purpose, we used
the Kneedle algorithm [18]. Since the knee often appears in an "L" shape, the Kneedle algorithm
works by drawing a line from the upper left to the lower right and locating the point that is furthest
from this line. This leads the algorithm to the bend of the "L".

2. Silhouette
defines the silhouette for an individual point as a number from -1 to +1 comparing the closeness of
other points within the same cluster to the rest of the points in the dataset. For a point i, let a(i) be
bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

the average distance between i and the other points in its cluster, and b(i) be the average distance
between i and the closest cluster to it. The silhouette for this point is
𝑏(𝑖) − 𝑎(𝑖)
(3)
𝑚𝑎𝑥{𝑎(𝑖), 𝑏(𝑖)}

For an entire dataset, the silhouette is the average over all points. This method was also combined
with the Kneedle algorithm for knee detection.

3. Ward KL
This k-selection method uses the Krzanowski-Lai index [19] to choose the number of clusters, with
the cluster assignments being made by Ward hierarchical clustering algorithm [20].
The Ward algorithm, described in , is a hierarchical clustering algorithm that minimizes a measure of
within-cluster variance (in this case, the squared Euclidean distance between points). Initially, each
point begins as a separate cluster. In each iteration, a pair of clusters must be merged - and the pair
with the lowest within-cluster distance is always the one chosen. The "ward.D2" version of the Ward
algorithm, from the NbClust library , was the implementation used.
To turn this clustering method into a k-selection algorithm, the NbClust library finds clusters for each
number from 1 to 32 and picks the one with the highest score. In this case, the score used for the
decision was the index proposed in .

The Krzanowski-Lai index is based on the quantity 𝑘 0/S 𝑆U2 , where p is the number of columns and 𝑆U2
is the trace of the pooled within-group covariance matrix. The difference between a given value of k
and the previous value is then referred to as 𝐷𝐼𝐹𝐹2 .
Finally, the Krzanowski-Lai index itself is defined as
𝐷𝐼𝐹𝐹2
𝐾𝐿(𝑘) = Z Z (4)
𝐷𝐼𝐹𝐹2[/

Therefore, the KL index will be maximized when the change from adding another cluster (|𝐷𝐼𝐹𝐹2[/ |)
is the smallest compared with the change from adding the last cluster (|𝐷𝐼𝐹𝐹2 |).

The gen3DNet package

We encapsulate above modeling process in the R package gen3DNet which is available from the
CRAN at https://CRAN.R-project.org/package=gen3DNet.
Installation
The package can be installed in two ways.
1) The release version from CRAN:
install.packages("gen3DNet")
bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

or 2) Directly from GitHub:

install.packages("gen3DNet")
remotes::install_github(“MollahLab/gen3DNet”)

Experiments and Results Analysis

Test cases
In this section, we demonstrate an application of gen3DNet to build a 3D network model linking
phosphoproteins, histones and drugs data obtained from Mollah SA and Subramaniam S study [4]
and compared the results with iPhDNet. Gen3DNet results are in perfect agreement with iPhDNet
results.

Evaluation
To evaluate Gen3DNet, 20 datasets were generated/simulated for each combination of (n_rows,
stdev. within / stdev. between). Each of the three k-selection methods was applied to the datasets.
The results for one pair of settings are shown in Figure 4, which shows the true k on the x axis and
the predicted k on the y axis. As the figure shows, the correlation between the true number of
clusters and the Ward prediction is perfect (45 degree angle line), while it is weaker for the other
methods: Kneedle Silhouette and Kneedle Cophenetic (scattered throughout).

Figure 4: Comparison of the three adopted methods for identifying the optimal number (K) of clusters.
bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

To compare the performance of these clustering methods across different settings, each of these
plots was represented as a correlation - and the results were gathered into heatmaps in Figure 5.
From these figures, we see that the Ward method improves as the number of dimensions increases
and the clusters become sharper. This makes sense, as either change will increase the distance
between clusters. This pattern is not seen for the other k-selection methods.

Figure 5: Average correlation in random datasets with different cluster sharpness and basis.

Discussion
The past few years have seen an increasing demand to transform and represent linkage-free data in
networks, which will deliver higher efficiency for many downstream applications. In this paper, we
introduced a simple and generic tool for generating 3D network models from two input matrices with
shared column properties. We have demonstrated the usage of the proposed gen3DNet R package
and its application to construct a 3D network involving phosphoproteins, histones and drugs in the
breast cancer study. Beside studying the histone signatures and precise mechanism of action in
cellular reprogramming of histone modifications in breast cancer, gen3DNet can be widely used to
reveal 3D connectivity and shared associations within any complex systems.

Acknowledgments
This work was supported by Washington University in St. Louis School of Medicine Genetics and
Institute for Informatics and Data Sciences departmental funds.

Author Contributions
Original concept: S.M.; experimental design: SM; implementation: P.M., S.M.; Evaluation: P.M., S.M.;
manuscript writing: P.M., S.M. R package submission: T.T., C.L.

References
1. Newman MEJ, Park J. Why social networks are different from other types of networks. Phys Rev E
Stat Nonlin Soft Matter Phys. 2003;68 3 Pt 2:036122.
bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

2. Franceschini A, Szklarczyk D, Frankild S, Kuhn M, Simonovic M, Roth A, et al. STRING v9.1: protein-
protein interaction networks, with increased coverage and integration. Nucleic Acids Res. 2013;41
Database issue:D808–15.

3. Hartwell LH, Hopfield JJ, Leibler S, Murray AW. From molecular to modular cell biology. Nature.
1999;402 6761 Suppl:C47–52.

4. Mollah SA, Subramaniam S. Histone signatures predict therapeutic efficacy in breast cancer. IEEE
Open J Eng Med Biol. 2020;1:74–82.

5. Sannigrahi M, Thandeeswaran R. Predictive analysis of network-based attacks by hybrid machine

learning algorithms utilizing Bayesian optimization, logistic regression, and random forest algorithm.
IEEE Access. 2024;12:142721–32.

6. Jahagirdar V, Bapaye J, Chandan S, Ponnada S, Kochhar GS, Navaneethan U, et al. Diagnostic

accuracy of convolutional neural network-based machine learning algorithms in endoscopic severity
prediction of ulcerative colitis: a systematic review and meta-analysis. Gastrointest Endosc.
2023;98:145–54.e8.

7. Khan MM, Mohsen MT, Malik MZ, Bagabir SA, Alkhanani MF, Haque S, et al. Identification of
potential key genes in prostate cancer with gene expression, pivotal pathways and regulatory
networks analysis using integrated bioinformatics methods. Genes (Basel). 2022;13:655.

8. de Campos LM, Cano A, Castellano JG, Moral S. Combining gene expression data and prior
knowledge for inferring gene regulatory networks via Bayesian networks using structural restrictions.
Stat Appl Genet Mol Biol. 2019;18.

9. Gao Y, Zhang S, Zhang Y, Qian J. Identification of MicroRNA-target gene-transcription factor

regulatory networks in colorectal adenoma using microarray expression data. Front Genet.
2020;11:463.

10. Stuart JM, Segal E, Koller D, Kim SK. A gene-coexpression network for global discovery of
conserved genetic modules. Science. 2003;302:249–55.

11. Filipov A, Mitova R, Slavchev G, Djambazov S, Vekov T. PCN230 A systematic literature review on
cost-effectiveness of next-generation sequencing (NGS): NGS compared to other high-throughput
sequencing methods in the context of personalized cancer therapy. Value Health. 2020;23:S464.

12. Chang KCN, Marton MJ. Next generation sequencing and its clinical applications: The growing
pains. Next Gener Seq Appl. 2016;3.

13. Lee DD, Seung HS. Learning the parts of objects by non-negative matrix factorization. Nature.
1999;401:788–91.

14. Charrad M, Ghazzali N, Boiteau V, Niknafs A. NbClust: AnRPackage for determining the relevant
number of clusters in a data set. J Stat Softw. 2014;61.
bioRxiv preprint doi: https://doi.org/10.1101/2024.11.11.623060; this version posted November 12, 2024. The copyright holder for this preprint
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
available under aCC-BY-NC-ND 4.0 International license.

15. Saraçli S, Doğan N, Doğan İ. Comparison of hierarchical cluster analysis methods by cophenetic
correlation. J Inequal Appl. 2013;2013.

16. Shahapure KR, Nicholas C. Cluster quality analysis using silhouette score. In: 2020 IEEE 7th
International Conference on Data Science and Advanced Analytics (DSAA). IEEE; 2020.

17. Krämer N, Sugiyama M. The degrees of freedom of partial least squares regression. J Am Stat
Assoc. 2011;106:697–705.

18. Satopaa V, Albrecht J, Irwin D, Raghavan B. Finding a “kneedle” in a haystack: Detecting knee
points in system behavior. In: 2011 31st International Conference on Distributed Computing Systems
Workshops. IEEE; 2011.

19. Krzanowski WJ, Lai YT. A criterion for determining the number of groups in a data set using sum-
of-squares clustering. Biometrics. 1988;44:23.

20. Murtagh F, Legendre P. Ward’s hierarchical clustering method: Clustering criterion and
agglomerative algorithm. arXiv [stat.ML]. 2011.