Abstract
Introduction
Understanding the temporal dynamics of gene expression is vital for
interpreting biological responses, especially in drug treatment
studies. Conventional visualization techniques, such as heatmaps and
static clustering, often fail to effectively capture these temporal
dynamics, particularly when analyzing large-scale multidimensional
datasets. These traditional methods tend to obscure fine-grained
temporal transitions, resulting in overcrowded visualizations,
diminished clarity, and limited interpretability of biologically
significant patterns.
Methods
To address these visualization challenges, we introduce Temporal
GeneTerrain, an advanced method designed to represent dynamic changes
in gene expression over time. We applied Temporal GeneTerrain to
compare transcriptomic perturbations induced by mefloquine (M),
tamoxifen (T), and withaferin A (W), both individually and in
all-pairwise and triple combinations (TM, TW, MW, and TMW), in LNCaP
prostate cancer cells using the [36]GSE149428 dataset (0, 3, 6, 9, 12,
and 24 h). Expression values were first Z-score normalized, and the
1,000 most variably expressed genes were selected. To ensure
coordinated temporal dynamics, we calculated Pearson correlation
coefficients among these genes and retained those with r ≥ 0.5,
resulting in 999 strongly co-expressed candidates. We then constructed
a protein-protein interaction network for these genes and embedded it
in two dimensions using the Kamada-Kawai force-directed algorithm.
Finally, for each time point and treatment, we mapped the normalized
expression values of the corresponding genes onto the fixed
Kamada-Kawai layout as Gaussian density fields (σ = 0.03), generating a
distinct Temporal GeneTerrain map for each time-condition combination.
Results
The application of Temporal GeneTerrain revealed intricate temporal
shifts in gene expression, particularly unveiling delayed responses in
pathways such as NGF-stimulated transcription and the unfolded protein
response under combined drug treatments. Compared to traditional
heatmap visualizations, Temporal GeneTerrain significantly improved
both resolution and interpretability, effectively capturing gene
expression patterns’ multidimensional and transient nature. This
enhancement provides a solid foundation for further research and
analysis, assuring the scientific community of the method’s
reliability.
Discussion
Temporal GeneTerrain addresses the limitations of traditional
visualization methods by offering an intuitive and detailed
representation of gene expression dynamics. Compared to other
approaches, such as heatmaps and static clustering, Temporal
GeneTerrain uniquely captures the transient nature of gene expression
patterns. This method significantly enhances the interpretability of
complex biological datasets, thereby supporting informed
decision-making in biological research and therapeutic development.
Keywords: cancer cell lines, gene expression, drug screening, precision
medicine, bioinformatics, data visualization
1 Introduction
Precision medicine, at its core, centers on selecting optimal
treatments for individual patients. Given the heterogeneity of
patients’ genetic profiles, it is crucial to identify appropriate drugs
and to administer them at the most effective time or disease stage. In
translational science, extensive efforts have focused on characterizing
gene function and determining the drugs specifically targeting those
genes. Importantly, as a disease progresses, its molecular drivers--the
specific genes or gene sets involved--may change, influencing both
disease propagation and therapeutic outcomes. This dynamic nature
necessitates treatment strategies that adapt over time, a challenge
underscored by in vitro drug screening studies, where traditional
visualization techniques such as heat maps, clustering, bar charts, and
box plots, are employed to represent drug-induced gene expression
changes ([37]Friedman et al., 2015; [38]McGranahan and Swanton, 2015;
[39]Huang et al., 2017; [40]Konno et al., 2019; [41]Malone et al.,
2020; [42]Zeng et al., 2021).
Traditional gene expression visualization techniques have been
instrumental in depicting transcriptomic changes and discerning
patterns for various conditions within complex datasets. Heat maps
effectively capture expression gradients but cannot typically
illustrate gene interactions or integrate functional annotations.
Clustering methods, including self-organizing maps ([43]Kohonen, 1997;
[44]Tamayo et al., 1999), can handle large-scale data, yet often yield
results that are difficult to interpret or may not accurately reflect
the underlying biology. Bar charts and box plots provide useful
summaries but are limited in representing the complexity and
multidimensionality of gene expression profiles ([45]Tamayo et al.,
1999).
Moreover, these conventional techniques suffer from data overcrowding
and loss of resolution when applied to large datasets ([46]Dietzsch et
al., 2006; [47]Katz et al., 2010; [48]Deng et al., 2014; [49]Metsalu
and Vilo, 2015), and are inadequate for capturing dynamic interactions
and temporal transitions in gene activity. Clustering methods may not
always reflect the actual biological function of genes, leading to
potential misinterpretations ([50]Eisen et al., 1998; [51]Tamayo et
al., 1999). In addition, a general shortcoming of conventional
visualization techniques is their limited integration of different data
types that overlook potential correlations, trends, or causal
relationships ([52]Gehlenborg and Wong, 2012).
These limitations underscore the need for more advanced, flexible, and
comprehensive visualization methods. Newer techniques are required to
handle the complexity of gene expression data while providing
interactive and integrative capabilities for deeper insights and
interpretations. Such advanced methods could significantly enhance our
understanding of gene interactions and functions, particularly in
complex biological systems, and offer a more comprehensive view of drug
responses in cancer cell lines over time.
While traditional visualization approaches effectively capture gene
expression at discrete time points, they fail to convey the continuous
evolution of gene regulatory networks during disease progression. The
temporal aspect of gene involvement is critical; genes may be
differentially engaged as the disease evolves, influencing both
progression and treatment outcomes. Static representations do not
capture these transitions, limiting our understanding of the dynamic
interplay between gene regulation and therapeutic interventions.
In response to these challenges, we propose the Temporal GeneTerrain, a
novel visualization technique designed to capture and represent dynamic
changes in gene expression over time. Unlike static snapshots, Temporal
GeneTerrains generate a continuous, integrated view of gene expression
trajectories that evolves during disease progression and treatment
response. By revealing temporal transitions in gene expression, this
method provides new insights into the molecular dynamics underlying
pathophysiological and therapeutic responses.
Temporal GeneTerrains build upon established visualization concepts
while introducing several key innovations. Rather than displaying a
series of disconnected time-point snapshots, our approach creates a
continuous representation that emphasizes the trajectories of
expression changes. This design facilitates the identification of
coordinated expression patterns, transient effects, and delayed
responses that might be overlooked using conventional methods.
Furthermore, the technique incorporates interactive elements that
enable users to explore the data at different levels of granularity,
focusing on specific genes, pathways, or time periods of interest to
gain multiscale insights ([53]Ahmed et al., 2021).
A significant strength of Temporal GeneTerrains is their capacity to
manage large-scale, multidimensional datasets without compromising
interpretability. Dimensionality reduction algorithms preserve
biologically meaningful relationships, allowing complex temporal
patterns to be visualized intuitively. This capability is especially
valuable in precision medicine, where integrating heterogeneous data
sources is essential for informing clinical decisions.
Another distinctive feature of Temporal GeneTerrains is its integration
of functional annotations and interaction networks within the
visualization. By overlaying known biological pathways, protein-protein
interactions, and gene-disease associations, the technique provides
context for interpreting observed expression changes. This integrative
approach bridges the gap between molecular data and clinical
implications, facilitating the translation of research findings into
actionable insights for patient care ([54]Gonzalez-Hernandez et al.,
2018).
Key innovations of Temporal GeneTerrain include:
* • Continuous Temporal Mapping: Rather than discrete snapshots, our
method interpolates expression changes to form a smooth trajectory,
exposing transient waves and sustained shifts in gene activity.
* • Integration of Functional Context: By overlaying pathway
annotations and PPI connections, each terrain conveys mechanistic
insights, linking molecular interactions to dynamic expression
patterns.
* • Invariant Network Topology: Re-optimizing layouts at each time
point introduces visual jitter, impeding clear trend tracking.
Freezing node coordinates on a single baseline layout eliminates
variability, enabling unambiguous comparison of gene trajectories
over time.
* • Adaptive Noise Smoothing: Fixed smoothing can either blur sharp
spikes or overemphasize noise. Dynamic modulation of the
GeneTerrain’s parameter according to expression-change magnitude
sharpens meaningful transients and highlights sustained patterns,
balancing sensitivity and clarity.
* • Scalability and Interactivity: Advanced dimensionality-reduction
techniques ensure that even large, multidimensional datasets remain
interpretable. Interactive controls allow users to explore
different temporal resolutions and focus on specific subnetworks
([55]Ahmed et al., 2021).
In [56]Table 1, we compared Temporal GeneTerrain against complementary
dynamic-network and trajectory-inference approaches. These include
TS-OCD for detecting overlapping temporal protein complexes
([57]Ou-Yang et al., 2014), DyNet/DyNetViewer for synchronized network
timelines ([58]Goenawan et al., 2016), TimeNexus for multilayer network
construction ([59]Pierrelée et al., 2020), TVNViewer for web-based
rewiring exploration ([60]Curtis et al., 2011), NACEP for module-aware
expression comparison ([61]Huang et al., 2010), TETRAMER for modeling
temporal transcriptional-regulation cascades ([62]Cholley et al.,
2018), and BioTapestry for hierarchical regulatory maps
([63]Longabaugh, 2011). TSEE excels at uncovering latent trajectories
in single-cell landscapes by embedding temporal information directly
into a dimensionality-reduction framework ([64]An et al., 2019).
TrendCatcher provides a robust statistical toolkit for pinpointing and
visualizing distinct gene- and pathway-level dynamics from longitudinal
data ([65]Wang et al., 2022). While each of these tools represents a
state-of-the-art solution within its specialized domain, Temporal
GeneTerrain complements them by fusing temporal dynamics with molecular
interaction networks into an interpretable spatial terrain, making it
uniquely suited for studies where understanding the co-regulation and
modular behavior of genes over time is critical. [66]Supplementary 1
provides additional details on the implementation and capabilities of
each method.
TABLE 1.
Feature matrix comparing Temporal GeneTerrain against leading dynamic
gene-expression visualization tools.
Method Continuous temporal map PPI integration (spatial) Statistical
DDEG modeling Regulatory GRN modeling Cytoscape Plugin
Temporal GeneTerrain ✔ ✔ ✖ ✖ ✖
TSEE ✔ ✖ ✖ ✖ ✖
TrendCatcher ✖ ✖ ✔ ✖ ✖
TS-OCD ✖ ✔ ✖ ✖ ✖
DyNet ✖ ✔ ✖ ✖ ✔
DyNetViewer ✖ ✖ ✖ ✖ ✔
TimeNexus ✖ ✖ ✖ ✖ ✔
TVNViewer ✖ ✖ ✖ ✖ ✔
NACEP ✖ ✖ ✔ ✖ ✖
TETRAMER ✖ ✖ ✖ ✔ ✖
BioTapestry ✖ ✖ ✖ ✔ ✖
[67]Open in a new tab
✔ indicates that the method provides the listed capability; ✖ denotes
lack thereof.
As a proof of concept, we applied Temporal GeneTerrains to study the
effects of single drug perturbations and their combinations in prostate
cancer cell lines. Prostate cancer was chosen as an ideal model due to
its well-characterized progression patterns and variability in
patients’ treatment responses. Using our generated temporally
integrated GeneTerrain visualizations, we observed distinct gene
expression transitions throughout treatment, revealing both immediate
and delayed responses that static methods failed to capture.
Subsequent gene set enrichment analysis further enhanced the biological
interpretation of our results by associating coordinated gene
expression changes with key pathways implicated in disease progression
and drug sensitivity or resistance. These insights underscore the
potential of Temporal GeneTerrains to elucidate molecular processes
that drive disease dynamics, and to identify novel targets for
therapeutic intervention ([68]Zhang et al., 2022).
The remainder of the paper is structured as follows: [69]Section 2
details the datasets involved and the algorithmic design of the
Temporal GeneTerrain method. [70]Section 3 presents a comprehensive
evaluation of results from a prostate cancer cell line case study under
various drug perturbations. [71]Section 4 discusses the biological
significance, clinical implications, and potential applications of our
findings, concluding with a summary of contributions and suggestions
for future research.
2 Materials and methods
2.1 Dataset
The [72]GSE149428 dataset ([73]Diaz et al., 2020) was retrieved from
the Gene Expression Omnibus (GEO) Database
([74]https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE149428).
This dataset comprises transcriptomic profiles of the LNCaP prostate
cancer cell line treated with DMSO (vehicle control), three distinct
drugs, and all possible combinations of these drugs, with each
condition performed in triplicate. Samples were collected at six time
intervals (0, 3, 6, 9, 12, and 24 h), including a triple drug
combination explicitly evaluated in LNCaP cells.
2.2 Temporal GeneTerrain: Conceptual framework and implementation
The Temporal GeneTerrain consists of an innovative visualization
methodology that extends conventional gene expression analysis by
incorporating temporal dynamics into a network-spatial framework that
enables researchers to observe the evolution of gene expression
patterns over time within a biologically meaningful spatial context.
The method builds upon the GeneTerrain visualization framework
initially proposed by [75]You et al. (2010) by incorporating an
additional temporal dimension that reveals dynamic changes in gene
expression patterns. The implementation follows a two-phase process:
(1) generating individual GeneTerrain visualizations at discrete time
points and (2) temporally integrating these visualizations to reveal
continuous dynamic expression patterns.
2.3 Data preparation and preprocessing
Before visualization, gene expression data were normalized using
Z-score normalization to standardize measurements across time points
and conditions. Differential expression analysis was performed at each
time point relative to baseline or control samples. Protein-protein
interaction (PPI) network data were obtained from HAPPI ([76]Chen et
al., 2017) and BEERE ([77]Yue et al., 2019) to provide the underlying
network structure for network-based visualization.
2.4 GeneTerrain generation
2.4.1 Network layout construction
Each GeneTerrain visualization is founded on a two-dimensional spatial
layout, where each node represents a gene, and the inter-node distances
reflect functional relationships derived from the PPI network. To
construct this layout, we employed the Kamada-Kawai force-directed
graph drawing algorithm ([78]Kamada and Kawai, 1989) to generate this
layout based on protein-protein interaction data. This algorithm
positions genes in two-dimensional space such that genes with
functional relationships (as defined by the PPI network) are placed in
closer proximity.
The Kamada-Kawai algorithm ([79]Kamada and Kawai, 1989) determines
optimal vertex (gene) positioning by minimizing the system’s energy,
which is computed using a combination of attractive forces between
connected vertices and repulsive forces between all pairs of vertices.
The system energy (E) is formulated as:
[MATH: E=∑i,j∈Ek2dij<
/msub>−lij2l
ij+∑i≠jk2dij :MATH]
Where:
* •
[MATH: E :MATH]
represents the set of edges in the PPI network
* •
[MATH:
dij :MATH]
is the Euclidean distance between nodes i and j in the
visualization
* •
[MATH:
lij :MATH]
represents the desired distance between nodes i and j, computed as:
[MATH:
lij=kArea
V
Lij :MATH]
* • |V| is the total number of vertices (genes) in the graph
* • Area is the display area of the visualization
* •
[MATH:
Lij :MATH]
is the shortest path length between nodes i and j in the PPI
network
* •
[MATH: k :MATH]
is a constant scaling factor
The algorithm iteratively adjusts vertex positions to minimize this
energy function until convergence criteria are met.
2.5 Gene expression signal visualization
After constructing the network layout, gene expression data are mapped
onto this spatial framework by generating a continuous signal field.
Each gene’s expression level is modeled as a three-dimensional Gaussian
distribution centered at its corresponding coordinates in the layout.
The amplitude of the Gaussian reflects the gene’s expression value,
where positive amplitudes denote upregulation and negative amplitudes
denote downregulation.
The continuous signal S (x,y) at any point (x,y) in the visualization
is given by:
[MATH: Sx,y=∑i=1nei·exp−x−xi2+y−yi22σ2 :MATH]
Where:
* •
[MATH: ei :MATH]
is the expression value of gene
[MATH: i :MATH]
* • (
[MATH: xi :MATH]
,
[MATH: yi :MATH]
) are the spatial coordinates of gene
[MATH: i :MATH]
in the layout
* •
[MATH: σ :MATH]
controls the dispersion (width) of the Gaussian distribution
The resulting continuous signal field is rendered using a divergent
color scheme, with intense red indicating strong upregulation, intense
blue indicating strong downregulation, and intermediate values
represented by shades of green. [80]Figure 1 of [81]Supplementary 2
presents the complete GeneTerrain workflow and demonstrates how varying
the Gaussian smoothing width (σ) systematically modulates the spatial
resolution and feature sharpness of the resulting terrain.
FIGURE 1.
[82]FIGURE 1
[83]Open in a new tab
A framework for constructing Temporal GeneTerrains involves Step-1,
Generating the GeneTerrain for visualizing using the network biology of
genes and gene signals; Step-2, Stitching the GeneTerrain generated in
Step-1 based on their chronological order.
2.6 Temporal integration
The second phase of the Temporal GeneTerrain method involves
integrating individual visualizations to capture temporal dynamics. For
each cell line, a series of GeneTerrain visualizations was generated
for discrete time points (e.g., 0, 6, 12, 24, and 48 h). To facilitate
direct comparison across time, the network layout coordinates were held
constant while updating the gene expression values for each time point.
This strategy enables the clear visualization of temporal trajectories
in gene expression, thereby revealing coordinated regulatory changes
and identifying genes with significant temporal modulation. [84]Figure
1 schematically illustrates the framework for constructing and
temporally integrating GeneTerrains, and [85]Algorithm 1 presents the
pseudocode for the proposed method.
Algorithm 1
Pseudocode of Temporal GeneTerrain Method.
* Require: GeneExpressionData, PPINetwork, TimePoints, σ, k, area, ϵ,
max- Iterations, η
* Ensure: TemporalVisualization
* 1: Data Preprocessing:
* 2: for each sample in GeneExpressionData do
* 3: Normalize (e.g., via Z-score)
* 4: end for
* 5: Network Layout Construction:
* 6: Let
[MATH: V=set of genes
:MATH]
in PPINetwork
* 7: for each pair
[MATH: i,j∈V,i≠j :MATH]
do
* 8:
[MATH:
Lij ← ShortestPathPPINetwork,i<
mo>,j :MATH]
* 9:
[MATH:
lij ← kar
eaVLij
:MATH]
* 10: end for
* 11: for each gene
[MATH: i∈V :MATH]
do
* 12: Initialize position
[MATH: pi :MATH]
* 13: end for
* 14:
[MATH:
iterat<
mi>ion ← 0,Eprev<
mtext> ←∞ :MATH]
* 15: while
[MATH:
iterat<
mi>ion < maxI
terations :MATH]
do
* 16:
[MATH:
Ecur←0
:MATH]
* 17: for all
[MATH:
i,j∈V,<
mi>i≠j :MATH]
do
* 18:
[MATH:
dij←
pi−pj
:MATH]
* 19:
[MATH:
Eij←
k2dij−lij2
lij+k2d
ij
:MATH]
* 20:
[MATH:
Ecur←Ecur
+Eij<
/mrow> :MATH]
* 21: end for
* 22: if
[MATH: Eprev
−Ecur< ϵ :MATH]
then
* 23: break
* 24: end if
* 25: for each
[MATH: i∈V :MATH]
do
* 26:
[MATH: Gi←∑j≠i∇piEij :MATH]
* 27:
[MATH:
pi←pi−η G
i :MATH]
* 28: end for
* 29:
[MATH:
Eprev ← Ecur,iterat
ion ← iteration+1 :MATH]
* 30: end while
* 31: Signal Field Generation:
* 32: for each
[MATH: t∈TimePoints
:MATH]
do
* 33: for each grid point (x, y) do
* 34:
[MATH: Stx,y←∑i∈Veitexp−x−p
i,x2+y−p
i,y22σ2
:MATH]
* 35: end for
* 36: Map
[MATH: Stx,y :MATH]
to a color map
[MATH: Vt :MATH]
* 37: end for
* 38: Temporal Integration:
* 39: Assemble
[MATH: Vt :MATH]
into TemporalVisualization
* 40: return TemporalVisualization
3 Results
In this article, we employed the Temporal GeneTerrain methodology to
perform a comparative analysis of transcriptomic perturbations induced
by three drugs, mefloquine (M), tamoxifen (T), and withaferin A (W),
administered individually and in combination (TM, TW, MW, and TMW) on
the LNCaP prostate cancer cell line. This case study demonstrates how
Temporal GeneTerrain can elucidate both single-drug and
combination-based perturbation at the transcriptomic level. We used the
[86]GSE149428 dataset ([87]Diaz et al., 2020) for our case study, which
consists of the drug perturbation data from cancer cell lines cultured
for six distinct time intervals (0, 3, 6, 9, 12, and 24 h).
We initially selected 1,000 genes exhibiting high expression variance
for data preprocessing, a key indicator of dynamic regulation and
potential interaction with drug response. To refine our gene set, we
conducted a thorough correlation analysis to identify genes with a
correlation coefficient threshold above 0.5. This stringent selection
criterion allowed us to distill our initial group to a core subset of
999 genes, demonstrating a strong inter-correlation, suggesting a
significant role in drug sensitivity and resistance mechanisms.
Parameter optimization involved manually setting the sigma parameter
for the GeneTerrain algorithm to 0.03. Subsequently, GeneTerrain
visualizations were generated for all samples. [88]Figure 2 illustrates
GeneTerrain visualizations that capture the dynamics of gene expression
alterations across successive time intervals (0→3, 3→6, 6→9, 9→12, and
12→24 h). At each interval, the GeneTerrains identify sets of
significantly upregulated or downregulated genes; these gene sets are
subsequently subjected to enrichment analysis to elucidate critical
pathways underpinning the observed temporal dynamics. The top row
corresponds to the DMSO (vehicle control), while the following three
rows display results from single-drug treatments, and the final four
rows depict drug combination treatments. Notably, treatments M and T
exhibit minimal perturbations, maintaining relatively consistent
expression patterns over the 24-h period. In contrast, other remaining
treatments display pronounced gene expression alterations. The
color-coded patterns and trajectories depicted in the GeneTerrain maps
effectively capture these dynamic perturbations over time, with a
colormap range from −10 to +10.
FIGURE 2.
[89]FIGURE 2
[90]Open in a new tab
GeneTerrains at different time intervals forming the Temporal
GeneTerrain for DMSO as the control vehicle, three treatments, and
their combination for the LNCaP cancer cell line.
[91]Figure 3 provides an in-depth portrayal of gene expression
variability in prostate cancer cells subjected to different drug
treatments across time, organized into two key sections. [92]Figure 3a
features a pie chart detailing gene variation distributions across six
comparison groups: T_0 vs. T_24, W_0 vs. W_24, TM_0 vs. TM_24, TW_0 vs.
TW_24, MW_0 vs. MW_24, and TMW_0 vs. TMW_24. Each chart segment
corresponds to one comparison category, with segment sizes reflecting
the proportion of highly variable genes in each. This format clarifies
how each treatment comparison uniquely influences gene expression
variability, accompanied by percentage values for quantitative insight.
The TMW treatment displays the most significant variability between its
initial and 24-h time points. Panel 3b highlights temporal patterns of
gene expression changes by presenting a pie chart that illustrates the
percentages of gene variability across discrete time intervals (0–3,
3–6, 6–9, 9–12, and 12–24 h) for individual drug treatments (TMW, MW,
TW, TM, W). The segments, representing distinct time intervals, offer
insights into the dynamic nature of cellular responses, precisely
pinpointing substantial changes during the TM treatment at 12–24 h, the
MW treatment at 0–3 h, and the TMW treatment at 12–24 h. Notably, the
minimal gene expression changes observed in the M and T treatments,
encompassing both upregulated and downregulated genes, were omitted
from [93]Figure 3.
FIGURE 3.
[94]FIGURE 3
[95]Open in a new tab
(a) a pie chart representing the distribution of gene variation across
specified categories: T_0 vs. T_24, W_0 vs. W_24, TM_0 vs. TM_24, TW_0
vs. TW_24, MW_0 vs. MW_24, and TMW_0 vs. TMW_24. Each segment’s size
corresponds to the proportion of genes with high variation in that
category, and the percentages are shown for each segment. (b) The pie
chart shows the percentage of each portion, representing the
distribution of counts over the various time intervals for each drug
(MWT, MW, TW, TM, W) in the time intervals (0–3, 3–6, 6–9, 9–12, 12–24
h).
[96]Figure 4a presents the Temporal GeneTerrain visualization for the
TM treatment, depicting drug perturbations on a cancer cell line sample
cultured over intervals of 0, 3, 6, 9, 12, and 24 h. [97]Figures 4b, c
present heatmaps of synthetic pathway enrichment scores over time in
the TM treatment study, utilizing the PAGER tool ([98]Yue et al., 2018;
[99]Yue et al., 2015) for both upregulated and downregulated gene sets.
Each heatmap displays the enrichment level for various pathways at
different time points, with rows representing pathways and columns
denoting time points. Each cell’s color intensity and values reflect
the pathway’s enrichment score. The PAGER tool parameters were set to a
PAG range of 2–5,000, a similarity score threshold of 0.05, a minimum
overlap of one gene, a cohesion value of 1, and a p-value threshold of
0.05, utilizing data from the WikiPathways_2021 dataset for Homo
sapiens. The heatmaps represent p-values on a–log10 scale to enhance
the visualization of statistical significance, with more intense red
shading indicating higher pathway significance.
FIGURE 4.
[100]FIGURE 4
[101]Open in a new tab
Analysis of Pathway Dynamics in LNCaP Cancer Cell Line Under TM
Treatment. (a) illustrating the Temporal GeneTerrain of a LNCaP cancer
cell line sample treated with a combination TM treatment, cultured over
intervals of 0, 3, 6, 9, 12, and 24 h. (b, c) Two heatmaps of
visualizing the synthetic pathway enrichment scores over time for TM
treatment for genes that are upregulated and downregulated using the
PAGER tool.
Examination of the pathway heatmap for upregulated genes in [102]Figure
4b provides a nuanced view of the cellular response dynamics in the
LNCaP cell line over the 24-h period ([103]Horoszewicz et al., 1983).
At 0 h, the metallothionein-binding pathway is notably active, with an
expression level of 5.1, suggesting an immediate cellular response to
the drug. This pathway remains consistently upregulated throughout the
24-h period, although with a gradual decrease in expression, indicating
a sustained but diminishing response. The choline catabolism pathway
shows a decrease in activity over time, starting at a moderate level at
0 h and progressively diminishing, reflecting a possible adaptation or
downregulation of this pathway in response to the treatment. Notably,
several pathways show no initial activity at 0 h but become
significantly upregulated later. For instance, the NGF-stimulated
transcription pathway ([104]Chen et al., 2021) remains inactive at 0 h
but shows a dramatic increase from 3 h onwards, peaking at 14.7 at 12
h. This suggests a delayed but robust response to the drug. Similarly,
the photodynamic therapy-induced unfolded protein response
([105]Firczuk et al., 2013) and the response of EIF2AK1 (HRI)
([106]Cordova et al., 2022) to heme deficiency pathways are activated
from 3 h onwards, indicating their roles in later stages of the drug
response. Interestingly, pathways such as the orexin receptor pathway
([107]Graybill and Weissig, 2017) and zinc homeostasis ([108]Li et al.,
2020) remain inactive until later stages (12 and 24 h, respectively)
before exhibiting markedly high expression levels, suggesting
involvement in long-term cellular adaptations. The heatmap in
[109]Figure 4c reveals distinct temporal response patterns among
various pathways in the LNCaP cancer cell line across the time points
of 0, 3, 6, 9, 12, and 24 h. At 0 h, pathways, including the nuclear
receptors’ meta pathway, the orexin receptor pathway, and spinal cord
injury-related pathways, exhibit significant activation, suggesting an
immediate cellular response to the drug perturbation. However, these
pathways show no activity at subsequent time points, suggesting a
transient or initial phase response. Subsequently, other pathways
become active as time progresses. Notably, the prostaglandin signaling
pathway shows a consistent increase in activity from 3 to 12 h,
indicating a sustained response to the drug over this period.
Similarly, the LTF danger signal response pathway exhibits a gradual
increase from 3 to 24 h, with a slight dip at the 12-h mark, suggesting
a prolonged involvement in the drug response. In contrast, some
pathways, such as the FGF23 signaling and unfolded protein response,
show activity only in the initial hours (up to 6 h) and then cease to
respond, implying a role in early-stage drug response mechanisms.
Interestingly, pathways associated with selective serotonin reuptake
inhibitors and activation of the NLRP3 inflammasome by SARS-CoV-2
exhibit delayed activation, with no initial activity but subsequent
triggering at 3 and 6 h, further highlighting temporal specificity in
drug response.
These pathways’ temporal specificity and high significance suggest
distinct roles in mediating the cellular response to drug treatment.
These findings are instrumental in identifying potential targets for
cancer therapy, elucidating the timing of cellular responses, and
informing the design of time-dependent therapeutic interventions.
Overall, the temporal pathway activation patterns in LNCaP cells reveal
phase-specific biological responses to treatment. Nuclear receptor
meta-pathway activation at baseline (0 h) aligns with studies showing
elevated LRH-1 expression in castration-resistant prostate cancer
(CRPC), where this nuclear receptor drives de novo androgen
biosynthesis via steroidogenic enzymes (CYP17A1, HSD3B1) ([110]Xiao et
al., 2018). The transient orexin receptor pathway activity at 0 h
corresponds to OX1R’s dual role in prostate cancer - baseline receptor
presence in LNCaP cells ([111]Graybill and Weissig, 2017) with
therapeutic potential through androgen receptor translocation
inhibition ([112]Graybill and Weissig, 2017; [113]Couvineau et al.,
2021). However, its rapid deactivation likely reflects receptor
internalization or downstream signaling adaptation. The sustained
activation of prostaglandin signaling from 3 to 12 h may intersect with
NF-κB-mediated survival pathways, as neuropeptides such as bombesin
have been shown to activate NF-κB in LNCaP cells to preserve androgen
receptor (AR) stability under castration conditions ([114]Jin et al.,
2008). Similarly, the prolonged activation of the LTF danger signal
mirrors mechanisms observed in CRPC progression, where chronic NF-κB
activation prevents tumor regression through AR/cyclin D1 maintenance
([115]Jin et al., 2008). The delayed NLRP3 inflammasome activation
(from 6 h) may reflect secondary stress responses, potentially linked
to LRH-1-mediated steroidogenesis creating pro-inflammatory
microenvironments ([116]Xiao et al., 2018). Although FGF23 signaling
was not directly examined in these studies, its early-phase activity
may be attributable to growth factor crosstalk during the initial
response to treatment.
4 Discussion
This study presents Temporal GeneTerrain, an innovative visualization
methodology that significantly enhances our ability to interpret the
dynamic landscape of gene expression in response to pharmacological
interventions. Moving beyond conventional static approaches such as
heat maps and standard clustering techniques, this method effectively
captures the temporal evolution of gene expression with enhanced
resolution and biological relevance. Similar advancements have been
observed in tools like TrendCatcher, which identifies dynamic
transcriptional signatures and biological processes over time, further
validating the importance of temporal analysis in transcriptomics
([117]Wang et al., 2022).
Our analysis revealed specific dynamic transcriptional events in LNCaP
cells, namely, the transient NGF-stimulated transcription surge peaking
at 12 h, the short-lived stress-response module activation between 6
and 9 h, the delayed zinc homeostasis upregulation at 24 h, and the
early but transient orexin receptor pathway engagement at baseline.
Previous studies examining NGF effects in LNCaP cells focused primarily
on proliferation assays conducted over multi-day intervals, without
accompanying genome-wide expression profiling at intermediate time
points ([118]Sortino et al., 2000). Earlier transcriptomic analyses
characterizing LNCaP progression toward castration resistance sampled
only broad intervals (e.g., 0, 3, 6, 12, and 24 h), without
highlighting transient activation spikes ([119]Vaarala et al., 2000).
Investigations of time-course zinc treatments in prostate cancer cells
predominantly profiled gene expression at earlier stages (e.g., 3 and 6
h) and did not capture the late-phase activation at 24 h ([120]Lin et
al., 2009; [121]Kolenko et al., 2013). Similarly, studies of orexin
receptors reported negligible OX1R expression in standard LNCaP cells
and lacked temporally resolved gene expression data that could reveal
an immediate baseline engagement ([122]Saghapour et al., 2024).
Furthermore, while methods such as TrendCatcher have demonstrated the
utility of fine-grained temporal analyses in other contexts, such as
COVID-19 PBMC transcriptomes ([123]Wang et al., 2022), and tools like
TSEE have underscored the importance of high-resolution temporal
embedding in single-cell RNA-seq data ([124]An et al., 2019), analogous
approaches have not previously been applied to the [125]GSE149428 LNCaP
dataset.
This echoes findings from studies like the GeneTerrain Knowledge Map
(GTKM), which uses protein-protein interaction networks to graphically
represent differentially expressed genes, offering nuanced insights
into gene interactions and expression patterns ([126]Saghapour et al.,
2024). Notably, Temporal GeneTerrain adeptly delineates immediate and
delayed transcriptional responses to single and combination treatments
with mefloquine, tamoxifen, and withaferin A. This temporal dissection
reveals that pathway activations, such as NGF-stimulated transcription
and the unfolded protein response, occur in a staggered manner,
aligning with studies that emphasize transcriptional adaptation as a
key driver of tumor progression rather than a reflection of
pre-existing cellular states ([127]Bolis et al., 2021).
A principal advantage of the Temporal GeneTerrain framework is its
integration of protein-protein interaction networks via force-directed
layout algorithms. This approach provides a spatial representation that
mirrors the functional interrelationships among genes, facilitating a
more intuitive interpretation of underlying biological processes. By
superimposing gene expression values on these networks using a
continuous signal field, nuanced but significant shifts in cellular
behavior over time were visualized. Similar approaches, such as TSEE
(Time Series Elastic Embedding), have demonstrated the potential of
integrating temporal information into visualization frameworks to
enhance the resolution of dynamic transitions ([128]An et al., 2019).
The observed temporal dynamics suggest that while some genes respond
promptly to treatment, others display delayed regulation, indicating a
complex interaction between immediate drug effects and subsequent
adaptive responses.
Furthermore, the integration of gene set enrichment analysis reinforces
our visualization strategy. By mapping expression changes to known
biological pathways, we confirm that the temporal shifts captured by
Temporal GeneTerrain correspond to critical regulatory mechanisms. For
instance, the delayed yet pronounced activation of pathways such as
NGF-stimulated transcription underscores the prospective relevance of
temporal regulation in influencing treatment outcomes. These findings
are consistent with studies showing gradual transcriptional progression
during prostate cancer adaptation to androgen deprivation therapy
([129]Vaarala et al., 2000; [130]Romanuik et al., 2010). This dual
strategy—merging spatial network visualization with pathway enrichment
analysis provides a comprehensive framework for interpreting
high-dimensional transcriptomic data in clinically relevant contexts.
Unlike generating independent GeneTerrain snapshots for each time
point, which requires realignment of hundreds of gene nodes, our
temporal stitching strategy constructs a single force-directed layout
at baseline and preserves those exact coordinates across all subsequent
time points. By mapping time-specific expression fields onto an
invariant topology and then interpolating between them, users can
directly track the ‘flow’ of each gene’s activity without the visual
jitter introduced by re-optimizing the layout at every interval.
To support this, we modified the underlying algorithm in two ways.
First, we performed Kamada–Kawai embedding just once on the PPI network
and froze the node positions for all time points, which eliminates
layout variability. Second, we introduced an adaptive Gaussian
smoothing scheme, in which σ is automatically adjusted based on the
magnitude of expression change between consecutive intervals such that
smaller σ values sharpen transient spikes, while larger σ values
emphasize sustained trends. Together, these enhancements sharpen the
contrast between immediate versus delayed responses and reduce noise.
Despite these promising findings, several limitations merit
consideration. First, the accuracy of the spatial layouts produced by
our methodology is inherently dependent on the quality and completeness
of the underlying protein-protein interaction data. Incomplete or
inaccurate interaction networks could obscure subtle gene
relationships. In addition, the parameter selection process, such as
determining the optimal sigma value for Gaussian smoothing, requires
careful calibration. Although our settings were informed by preliminary
analyses, further optimization across diverse datasets is necessary to
establish standardized protocols, a challenge similarly encountered in
other time-series visualization frameworks like TSEE ([131]An et al.,
2019).
Despite the advances afforded by Temporal GeneTerrain, our approach is
inherently dependent on the quality and completeness of the underlying
protein–protein interaction (PPI) network. Gaps or inaccuracies in PPI
data can obscure true functional relationships and lead to misplacement
of genes in the terrain. In principle, our framework can accommodate
alternative network priors such as regulatory interactions derived from
ChIP-seq, RIP-seq, or Hi-C chromatin-contact maps by simply
substituting the PPI adjacency matrix used in the force-directed
layout. Employing a ChIP-seq–based network, for example, would
emphasize direct transcription factor–target relationships, whereas
Hi-C–derived contacts could reveal three-dimensional co-regulation
modules. Similarly, our high-variance gene filter was chosen to focus
on the most dynamically regulated genes, but more stringent criteria
(e.g., top 500 most variable) would produce a sparser terrain with
reduced visual clutter at the expense of potentially missing subtler
but biologically important signals, while looser thresholds (e.g., top
2,000) would increase coverage but could overwhelm the visualization
with noise. Systematic evaluation of these tradeoffs across multiple
datasets would be essential for establishing best practices for gene
selection.
Scalability to larger cohorts and more heterogeneous clinical samples
also presents challenges. While we demonstrated Temporal GeneTerrain on
∼1,000 highly variable genes in a controlled cell-line experiment,
extending this method to complete transcriptomes or patient-derived
multi-omics (e.g., scRNA-seq, proteomics, metabolomics) would require
strategies to manage both computational load and visual
interpretability. Possible solutions include hierarchical terrain
generation first mapping modules or pathway-level aggregates before
“drilling down” to individual genes and GPU-accelerated force-directed
layouts to handle millions of interactions. For clinical datasets,
batch effects and sample heterogeneity can distort both the network and
the temporal signal; incorporation of robust normalization,
batch-effect correction, and adaptive σ-smoothing schemes will be
necessary to ensure that inter-patient variability does not confound
the temporal trajectories. Future work will focus on integrating these
strategies to bring Temporal GeneTerrain toward large-scale, clinically
actionable analyses.
As a result, Temporal GeneTerrains reveal biologically meaningful
dynamics that discrete analyses would miss, such as the 3 h activation
of NGF-stimulated transcription preceding any visible change in the
unfolded protein response, and the transient peak in stress-response
genes between 6 and 9 h that disappears by 12 h. These patterns emerge
organically from the continuous terrain rather than as isolated signals
across separate plots, significantly improving interpretability in
time-series transcriptomic studies. When we used a traditional
clustered heatmap to visualize the 0–24 h expression data, transient
spikes were blurred and gene trajectories became misaligned.
[132]Figure 2 in [133]Supplementary 2 shows how static heatmaps obscure
fine-grained temporal transitions, with staggered clusters and shifted
patterns. In contrast, our dynamic Temporal GeneTerrain, when using
σ-modulated smoothing uncovers interaction patterns and transient
activation waves, fully resolved underlying temporal dynamics.
Moreover, while Temporal GeneTerrains prove effective in visualizing
dynamic gene expression in controlled in vitro environments, their
application to more heterogeneous clinical samples may pose challenges.
The variability inherent in patient-derived data encompassing
differences in sample quality, treatment regimens, and disease stages
could complicate the visualization and interpretation of temporal
patterns. Future research should focus on adapting this methodology for
clinical contexts, ideally integrating additional omics layers (e.g.,
proteomics or metabolomics) to yield a more holistic view of disease
dynamics ([134]Romanuik et al., 2010; [135]Aviñó-Esteban et al., 2025).
Looking ahead, several promising avenues for future research emerge.
Expanding the Temporal GeneTerrain framework to incorporate multi-omics
data could provide a more enriched, systems-level perspective on
cellular responses. Furthermore, coupling real-time data acquisition
with interactive visualization tools may facilitate dynamic decision
support in clinical settings, ultimately refining therapeutic
precision. Such advancements would broaden our approach’s applicability
and help bridge the gap between fundamental research and clinical
practice.
In conclusion, Temporal GeneTerrains represent a significant
methodological advancement in bioinformatics and precision medicine. By
capturing the dynamic nature of gene expression and linking these
alterations to functional biological pathways, this methodology
provides researchers and clinicians with a powerful tool for
understanding the complex temporal responses that underlie drug
efficacy and disease progression. Continued enhancements and validation
across diverse biological contexts will establish Temporal GeneTerrains
as an indispensable resource for developing time-sensitive,
personalized therapeutic strategies.
Funding Statement
The author(s) declare that financial support was received for the
research and/or publication of this article. This work was partially
supported by internal research grants from the University of Alabama at
Birmingham awarded to JC, as well as by the National Institutes of
Health under grant number 1UM1TR004771-01.
Data availability statement
The dataset utilized in this study can be found at:
[136]www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE149428, and the code
is available at:
[137]https://github.com/aimed-lab/Temporal_GeneTerrain.
Author contributions
ES: Conceptualization, Data curation, Formal Analysis, Methodology,
Validation, Visualization, Writing – original draft, Writing – review
and editing. RS: Writing – original draft, Writing – review and
editing. DH: Writing – review & editing. KS: Writing – original draft,
Writing – review and editing. ZS: Writing – review and editing. JC:
Conceptualization, Funding acquisition, Investigation, Methodology,
Project administration, Supervision, Validation, Writing – review and
editing.
Conflict of interest
The authors declare that the research was conducted in the absence of
any commercial or financial relationships that could be construed as a
potential conflict of interest.
Generative AI statement
The author(s) declare that no Generative AI was used in the creation of
this manuscript.
Publisher’s note
All claims expressed in this article are solely those of the authors
and do not necessarily represent those of their affiliated
organizations, or those of the publisher, the editors and the
reviewers. Any product that may be evaluated in this article, or claim
that may be made by its manufacturer, is not guaranteed or endorsed by
the publisher.
Supplementary material
The Supplementary Material for this article can be found online at:
[138]https://www.frontiersin.org/articles/10.3389/fbinf.2025.1602850/fu
ll#supplementary-material
[139]Table2.xlsx^ (27.7KB, xlsx)
[140]Presentation1.pptx^ (1.9MB, pptx)
[141]Table1.xlsx^ (16.2KB, xlsx)
References