Abstract
Ovarian cancer is a malignant tumor with different clinicopathological
and molecular characteristics. Due to its nonspecific early symptoms,
the majority of patients are diagnosed with local or extensive
metastasis, severely affecting treatment and prognosis. The occurrence
of ovarian cancer is influenced by multiple complex mechanisms
including genomics, transcriptomics, and proteomics. Integrating
multiple types of omics data aids in predicting the survival rate of
ovarian cancer patients. However, existing methods only fuse
multi-omics data at the feature level, neglecting the shared and
complementary neighborhood information among samples of multi-omics
data, and failing to consider the potential interactions between
different omics data at the molecular level. In this paper, we propose
a prognostic model for ovarian cancer prediction named Dual Fusion
Channels and Stacked Graph Convolutional Neural Network (DFASGCNS). The
DFASGCNS utilizes dual fusion channels to learn feature representations
of different omics data and the associations between samples. Stacked
graph convolutional network is used to comprehensively learn the deep
and intricate correlation networks present in multi-omics data,
enhancing the model’s ability to represent multi-omics data. An
attention mechanism is introduced to allocate different weights to
important features of different omics data, optimizing the feature
representation of multi-omics data. Experimental results demonstrate
that compared to existing methods, the DFASGCNS model exhibits
significant advantages in ovarian cancer prognosis prediction and
survival analysis. Kaplan-Meier curve analysis results indicate
significant differences in the survival subgroups predicted by the
DFASGCNS model, contributing to a deeper understanding of the
pathogenesis of ovarian cancer and providing more reliable auxiliary
diagnostic information for the prognosis assessment of ovarian cancer
patients.
Introduction
Ovarian cancer is a tumor with a range of distinct clinicopathological
and molecular features [[34]1]. Due to its inconspicuous early
symptoms, it is often diagnosed at an advanced stage, leading to high
recurrence rates and low survival rates for patients [[35]2,[36]3]. The
occurrence of ovarian cancer is influenced by complex mechanisms at
multiple levels including genomics, transcriptomics, and proteomics
[[37]4–[38]6], and different types of omics analysis aid in predicting
the survival rate of ovarian cancer patients [[39]4,[40]6,[41]7].
Ovarian cancer is a heterogeneous disease characterized by molecular
and omics diversity. While single omics data focus on specific
molecular aspects of ovarian cancer, integrating different omics data
can provide complementary information from different molecular
perspectives. Comprehensive analysis of multi-omics data is crucial for
understanding the pathogenesis of ovarian cancer and predicting patient
prognosis. Boehm et al. [[42]8] integrated histopathological,
radiological, and clinical genomic data of high-grade serous ovarian
cancer to predict patient prognosis using a risk stratification model.
Zhang et al. [[43]9] utilized multi-omics data including gene
expression, somatic DNA alterations, miRNA expression, and DNA
methylation in serous ovarian cancer (SOC) to determine patient
prognosis and treatment efficacy, offering new insights for improved
treatment strategies. Therefore, ovarian cancer prognosis prediction
based on multi-omics data not only improves prediction accuracy but
also deepens researchers’ understanding of the biological
characteristics and molecular mechanisms of ovarian cancer. This
provides valuable insights for clinical practice and personalized
treatment strategies for patients.
Due to different omics data providing varying perspectives on ovarian
cancer, the integration of multi-omics data can comprehensively reveal
the molecular characteristics of ovarian cancer diversity from the
aspects of genomics, transcriptomics and epigenomics [[44]10],
contributing to a more comprehensive understanding of the potential
biological processes underlying ovarian cancer development
[[45]11,[46]12]. Existing researchers mainly fuse multi-omics data
through two methods: feature-based fusion and graph-based fusion.
Feature-based fusion methods integrate the feature representations of
different omics data into a unified feature space to reveal the
inherent correlations among multi-omics features. For example, Cheerla
et al. [[47]13] utilized unsupervised encoders to compress clinical
data, mRNA expression, microRNA expression, and histopathology whole
slide images (WSIs) into 512-dimensional feature vectors. They
aggregated the four feature vectors into a common feature space using
similarity loss for predicting the survival rate of pancreatic cancer
patients. The GPDBN method [[48]14] and HBSurv method [[49]15] employed
a bilinear feature encoding module to fuse the features of multi-omics
data and clinical data into a feature space, effectively utilizing
relationships between and within modalities to enhance the predictive
performance of the model. Graph-based multi-omics fusion methods
represent different types of omics data as graph structures, where
nodes represent cancer samples and edges represent relationships
between samples. These methods fuse the graph structures of different
omics data to achieve comprehensive analysis across omics data types,
fully considering the correlations between samples of different omics
data. The GCGCN method [[50]16] fused the sample-sample similarity
matrices of multi-omics data (including gene expression, copy number
alteration, DNA methylation, and exon expression) and clinical data
into a sample network graph structure using a similar network fusion
algorithm (SNF). It obtained shared and complementary information from
different data sources and achieved survival prediction for Breast
cancer (BRCA) and Lung squamous cell cancer (LUSC). Wang et al.
[[51]17] used cosine similarity to construct a weighted sample
similarity network for cancer multi-omics data (mRNA, DNA methylation,
and miRNA), and classified cancer subtypes by learning the correlation
between omics data and samples through graph convolutional network.
Although these methods outperform single-omics approaches,
feature-based fusion methods only focus on low-dimensional feature
representations of multi-omics data, while graph-based methods
prioritize the associations between cancer samples. Therefore, this
paper proposes a dual fusion channel strategy that combines feature
fusion and graph fusion. The strategy not only fully learns the feature
representations of different omics data at the feature level but also
considers the inherent correlations between ovarian cancer samples,
obtaining more reliable information for prognosis prediction of ovarian
cancer patients.
In recent years, researchers have used deep learning methods for cancer
subtype classification and patient prognosis prediction based on
multi-omics data by learning local features of each omics data type and
directly concatenating the local features of various omics data for
cancer classification or prediction. Liu et al. [[52]18] employed
convolutional autoencoders (CAE) to extract low-dimensional feature
representations of RNA and CNV separately, concatenated them, and used
a univariate Cox proportional hazards (Cox-PH) model to select features
associated with cancer survival for patient prognosis prediction.
DeepMO [[53]19] learned local features of mRNA, DNA methylation, and
CNV through corresponding encoding subnetworks, concatenated these
features as inputs to a classification subnetwork, and achieved breast
cancer subtype classification. ConcatAE [[54]20] used autoencoders (AE)
to learn latent variable features of different types of omics data,
including Gene Expression, DNA Methylation, miRNA Expression, and CNV,
and concatenated them for survival prediction of breast cancer
patients. However, these methods learn local features of multi-omics
data and fuse the features of multi-omics data directly by
concatenation, neglecting the mutual influence of different omics data
at the molecular level and failing to consider the direct or indirect
relationships among samples, thus failing to comprehensively learn the
interactions and potential structures among multi-omics data, which
reduces the interpretability and generalization of the model. Recently,
graph convolutional network (GCN) have shown promising performance in
various fields such as cancer patient survival prediction and cancer
subtype classification by simultaneously leveraging both omics features
and the similarity networks describing correlations among samples
[[55]21]. Ling et al. [[56]22] proposed the survival model AGGSurv
based on GCN, constructing different sparse graphs using random subsets
of multi-omics high-dimensional features, which aids in survival
analysis of cancer patients. However, traditional GCN can only capture
the direct neighbor information of nodes in a single convolution
operation, neglecting the potential network structure among the nodes,
which limits the extraction of global information among different omics
data. Therefore, this paper proposes Stacked Graph Convolutional
Network (SGCN), which builds upon a single-layer GCN by stacking
multiple layers of graph convolutions. This progressively expands the
receptive field, allowing for the indirect capture of relationships
between more distant nodes, thereby extracting global features among
multi-omics data. By stacking multiple graph convolutional layers, this
paper integrates features of all multi-omics data and relationships
among samples, effectively learning the local and global complementary
information of ovarian cancer multi-omics data, thereby improving the
accuracy of the model for prognosis prediction.
Attention mechanisms regularize feature maps and word embeddings,
allowing deep learning models to focus more on specific regions
relevant to the model’s objectives [[57]23–[58]25]. Research has shown
that there exist important features highly correlated with prognosis
and patient survival in cancer multi-omics data [[59]26–[60]28]. By
utilizing attention mechanisms to focus on key features and optimize
the learning process of these critical features, the accuracy and
interpretability of models can be effectively improved. Choi et al.
[[61]29] developed the breast cancer subtype classification model
moBRCA-net, which uses attention modules to learn the importance of
features from different omics data, enhancing classification
performance. Sanghyuk et al. [[62]30] proposed the Multi-Prognosis
Estimation Network (Multi-PEN), which employs gene attention layers for
both mRNA expression and miRNA expression, identifying genes associated
with the prognosis of low-grade glioma (LGG). Prognostic genes play
crucial roles in cancer regulation, and the selection and attention to
prognostic genes significantly impact the survival status of patients.
Therefore, this paper introduces attention mechanisms to select
important features closely related to ovarian cancer prognosis from
multi-omics data, and assigns corresponding attention weights to the
features according to their importance, so as to optimize the feature
representations in order to improve the model’s prognosis prediction
performance.
In this study, we propose the deep learning model DFASGCNS, which uses
a dual fusion channel and stacked graph convolutional network for
predicting the prognosis of ovarian cancer patients. The DFASGCNS model
proposed a dual fusion channel strategy to select and effectively learn
the feature representations of various omics data, merging them at the
feature level while constructing graph structures corresponding to
different omics data. This approach enables an in-depth analysis of the
associations and potential network structures between samples of
different omics data, thereby revealing the complex interactions among
them. Stacked graph convolutional network (SGCN) is proposed, where a
single GCN layer learns the feature representations and direct
neighborhood information of multi-omics data, while multiple stacked
GCN layers explore the deeper connections and complex relational
networks among multi-omics samples. This comprehensive capture of
feature representations and sample associations enhances the overall
understanding of the data. Attention mechanism is introduced to select
important genes related to ovarian cancer prognosis, assigning
corresponding attention weights based on feature importance, focusing
the model more on features highly relevant to prognosis prediction. The
experimental results show that compared with the existing methods, the
model DFASGCNS in this paper fully considers the feature representation
of different omics data and the association between samples,
comprehensively acquires the complementary information and deep
association of multi-omics data, improves the performance of ovarian
cancer prognosis prediction, and contributes to the understanding of
the pathogenic mechanism of ovarian cancer and the identification of
personalized treatment plans for patients.
Materials and methods
Data collection
This paper downloaded multi-omics data and clinical data of ovarian
cancer from the TCGA database ([63]https://portal.gdc.cancer.gov/). The
omics data includes mRNA expression, DNA methylation, miRNA expression
and copy number variation (CNV)), the corresponding number of features
are 46610, 24923, 1874 and 24740 respectively. The details are shown in
[64]Table 1. The clinical data describes the clinical information of
587 ovarian cancer patients, including age, race, FIGO stage, survival
time, survival status and other characteristics.
Table 1. The summary of ovarian cancer data.
Omics type Number of samples Number of features Summary
mRNA 367 46610 HTSeq-FPKM
DNA methylation 363 24923 [65]Illumina Human Methylation 27k
miRNA 499 1874 BCGSC Illumina HiSeq
CNV 606 24740 Affymetrix SNP Array 6.0
[66]Open in a new tab
Data preprocessing
This paper preprocesses the downloaded data using a 4-step methods. The
first step is to intersect the samples of the mRNA expression, DNA
methylation, miRNA expression and CNV data sets to obtain 325 common
samples in the cross dataset. Secondly, to filter the features with
more than 20% missing values and combine the expression values genes
with ’0’ were converted to ’NA’, and the R package ‘ImputeMissings’
[[67]31] was used to fill missing values based on the median. Thirdly,
due to the large number of omics data features, the variance threshold
method was applied to select the features with variance, calculated
over all patients, higher than the given threshold [[68]32,[69]33], the
variance thresholds of mRNA, DNA methylation and CNV were initially
determined to be 7, 0.02 and 0.1 respectively. Finally, we use z-score
techniques to normalize data. After preprocessing, the number of
features for mRNA, DNA methylation, miRNA and CNV are 8492, 6125, 454,
and 2274 respectively, and the number of samples is all 325, as shown
in [70]Table 2.
Table 2. Summary information of the multi-omics data of ovarian cancer after
preprocessing.
Omics type Number of samples Number of features Summary
mRNA 325 8492 HTSeq-FPKM
DNA methylation 325 6125 [71]Illumina Human Methylation 27k
miRNA 325 454 BCGSC Illumina HiSeq
CNV 325 2274 Affymetrix SNP Array 6.0
[72]Open in a new tab
Model architecture
To leverage the complementary information of different omics data and
fully consider the shared and neighborhood information among different
omics samples, this paper proposes a novel ovarian cancer prognosis
prediction model, DFASGCNS. This model is based on stacked graph
convolution network (SGCN) and employs a dual fusion channel strategy,
aimed at simultaneously learning the omics feature data and the
intrinsic correlations between samples. Specifically, DFASGCNS
preprocesses ovarian cancer’s mRNA expression, DNA methylation, miRNA
expression, and CNV as inputs, selects genes closely related to ovarian
cancer prognosis through a feature selection algorithm, and employs an
attention mechanism to allocate different weights to important features
across different omics data, optimizing the representation of key
features in various omics data. For each type of omics data, a
relational graph between samples is constructed, and these graphs are
fused into a unified graph structure. The fused graph, together with
the fusion feature representation weighted by the attention mechanism,
is input into the SGCN to comprehensively learn the feature information
of the ovarian cancer multi-omics data and the sample network
structure. Finally, the prognosis of ovarian cancer patients is
predicted using a softmax classifier. The architecture of the DFASGCNS
model is depicted in [73]Fig 1.
Fig 1. The architecture of DFASGCNS.
[74]Fig 1
[75]Open in a new tab
Feature selection method: RLASSO
Ovarian cancer multi-omics data are characterized by low sample size
and high dimensionality. Feature selection methods can effectively
capture important features from the high-dimensional data, thereby
enhancing the predictive performance of the model [[76]34]. As shown in
[77]Table 2, the number of features after preprocessing for mRNA, DNA
methylation, miRNA, and CNV are 8492, 6125, 454, and 2274,
respectively. To identify features highly correlated with ovarian
cancer prognosis, this study introduces the use of a random forest and
the RLASSO feature selection method. Important features identified by
the random forest serve as supplements to the features lost during
LASSO regression. The final counts of features for each type of omics
data are shown in [78]Table 3.
Table 3. Summary information of the multi-omics data of ovarian cancer after
RLASSO.
Omics type Number of samples Number of features
mRNA 325 143
DNA methylation 325 142
miRNA 325 128
CNV 325 136
[79]Open in a new tab
In RLASSO, initially, LASSO regression employs L1 regularization by
adding a penalty term to the least squares error component in the
objective function. During the optimization process, this causes the
coefficients of some features to shrink towards zero, and ultimately,
reduces the coefficients of certain features entirely to zero, thereby
facilitating feature selection. The formula for feature selection using
LASSO regression is as follows:
[MATH: Min∑j=1N(
yji−∑k=1dixjkωk)2+
λ∑k=1di||ωk||1
subjectto:∑k=1di||ωk||1<
/mn><c
:MATH]
(1)
Where, i represents the i-th omics data type, N denotes the number of
samples, d^i indicates the total number of features for the i-th omics
data, y[j] represents the label of the j-th sample, and λ is the
regularization parameter.
Importance ranking of all features using random forest to construct a
decision tree, and the first K features with high importance are
selected based on the criteria of feature importance. In the feature
set of omics data
[MATH:
F={f1<
/mn>,f2,f3
,…,f(di<
/msup>) :MATH]
}, the feature importance collection
[MATH:
I={I1<
/mn>,I2,I3
,…,I(di<
/msup>) :MATH]
} is outputted based on feature importance. Here, d^i represents the
total number of features for the i-th omics data, and the formula for
I[x] is as follows:
[MATH:
Ix=1N∑k=1N(Rnoob
−Rnjoob
) :MATH]
(2)
Where,
[MATH:
Rpo
ob :MATH]
and
[MATH:
Rpj
oob :MATH]
represent data outside of the bag before and after the decision tree
disturbance (i.e. samples not sampled during decision tree resampling),
and the number of correctly classified samples is tallied. The top K
features are selected in descending order of importance for feature
selection. In this study, the optimal predictive performance is
achieved when K = 100. The important features selected by random forest
and the features retained by LASSO regression are combined as the total
features for specific omics data.
Construction and fusion of relationship diagrams between samples
While feature fusion can capture the complementary information at the
feature level from different omics data for each patient, it overlooks
the shared and complementary neighborhood information among samples
from different omics data. Addressing this issue, the DFASGCNS model
integrates multiple graphs constructed from ovarian cancer single omics
data and generates a fused graph, accurately revealing the intrinsic
correlations among samples. Given the data representation
[MATH: {xi}i=1n :MATH]
for n samples, where ρ(x[i],x[j]) represents the Euclidean distance
between samples x[i] and x[j], we construct the adjacency matrix A of
the k-nearest neighbor graph using the exponential similarity kernel.
Where, A∈R^n×n, and the calculation process is as follows:
[MATH: A(i,j)={exp(−ρ2(<
mrow>xi,xj)μδ2
),
j∈Ni0,otherwise<
/mi> :MATH]
(3)
Where N[i] represents the k nearest neighbors of sample i,δ^2 is used
to address scale issues and is experimentally set to the median of the
squared distances between all pairs of samples. μ is the
hyperparameter, which is set to 0.5.
According to the construction method of the k-nearest neighbor graph
described above, four single omics graphs were built using mRNA
expression data X^(1), DNA methylation data X^(2), miRNA expression
data X^(3), and CNV data X^(4), respectively denoted as
A^(1)、A^(2)、A^(3) and A^(4). These graphs were fused by taking the
average of the four adjacency matrices to obtain the final fusion graph
of adjacency matrix A.
Introducing attention mechanism for feature fusion
In ovarian cancer, different omics data have varying degrees of
importance in prognosis prediction [[80]37], such as mRNA expression,
DNA methylation, miRNA expression, and CNV. The self-attention
mechanism, through the training process, can autonomously learn the
feature weights of each omics data type and adaptively evaluate the
importance of each feature, enabling feature weighting. Compared to
traditional static feature selection methods, the self-attention
mechanism better captures the interdependencies and hidden
relationships between multi-omics data. Through feature weighting and
fusion, it enhances the model’s focus on key features, ultimately
improving the performance and accuracy of prognosis prediction.
Therefore, this study introduces the self-attention mechanism to
explore the contribution of different omics data to ovarian cancer
prognosis prediction. By assigning different weights to each type of
omics data, the model can better represent the important features
during feature fusion, thus improving the representation of crucial
multi-omics data. After preprocessing, mRNA expression, DNA
methylation, miRNA expression, and CNV data through RLASSO feature
selection, resulting in low-dimensional representations denoted as
[MATH:
Z(k)∈Rn×dk :MATH]
, where k = 1,2,3,4 represents the k-th omics data type, n denotes the
number of samples. d[k] denotes the number of features for the k-th
omics data. Each omics data type is treated as a feature matrix where
each row corresponds to a sample and each column corresponds to a
feature. To explore the contribution of different omics data to
prognosis prediction, this paper introduces a self-attention mechanism
that utilizes adaptive attention weights to assign importance to the
features of various omics data, thereby enhancing the model’s
prognostic performance. Specifically, an attention weight vector
[MATH:
α(k)∈Rd<
/mi>k :MATH]
is introduced for important features of each omics data, representing
the model’s attention to each feature and its importance in the fusion
process. The calculation process of attention weights α^(k) is as
follows:
[MATH:
α(k)=softmax(W(k)h(k)+b(k)<
/msup>) :MATH]
(4)
Where,
[MATH:
W(k)∈Rd<
/mi>k×d :MATH]
represents the learnable weight matrix, b^(k)∈R^d represents the bias
vector,
[MATH:
h(k)∈Rd<
/mi>k :MATH]
represents the feature representation of the k-th omics data, and d
denotes the number of hidden units in the attention mechanism. The
attention weight vector α^(k) is multiplied with the feature
representation Z^(k) to obtain the attention-weighted omics feature
representation. Consequently, the fused feature representation
Z^(fusion) is obtained. The calculation process is as follows:
[MATH:
Z(fus<
/mi>ion)=∑k=1
4 α(k)
Z(k) :MATH]
(5)
Where Z^(k) represents the feature representation of the k-th omics
data, and α^(k) is the attention weight vector for the corresponding
omics data.
Stacked graph convolutional network (SGCN)
To extract deeper and complex network of relationships for survival
prediction from ovarian cancer multi-omics data, this study proposes a
dual fusion channel strategy to merge feature representations between
omics data and relationships among samples to obtain richer features.
Stacked Graph Convolutional Network (SGCN) is proposed to learn the
fused graph A and the global feature representation Z^(fusion) after
the fusion of multi-omics data. In SGCN, graph convolutional network is
a deep learning model for processing graph-structured data. When
calculating the convolution matrix, the graph structure needs to be
modeled first. Specifically, the procedure of calculating the
convolution matrix
[MATH: A^
:MATH]
in SGCN is shown below:
[MATH: A^=
D˜−12A˜D˜−12 :MATH]
(6)
Where
[MATH: A˜=A<
/mi>+In,In
:MATH]
is the n-order identity matrix.
[MATH: D˜=d<
/mi>iag(d˜1,d˜2,…,d˜n) :MATH]
represents the degree matrix derived from
[MATH: A˜
:MATH]
, where
[MATH: d˜i=∑j <
/mrow>A˜ij :MATH]
. For the node representation matrix Z^(fusion) = [z[1],z[2],…,z[n]]
with n samples and the convolution matrix
[MATH: A^
:MATH]
, in the stacked graph convolution network, the calculation of each
graph convolutional layer’s output is as follows:
[MATH:
Z(l+1<
/mn>)=f2(f
1(A^Z(fusion
)W1(l))W2(l)<
mo>) :MATH]
(7)
Where
[MATH:
W1(
l) :MATH]
and
[MATH:
W2(
l) :MATH]
are the weight matrices of the l-th graph convolutional layer, f[1] and
f[2] are tanh and sigmoid activation functions, respectively. Repeated
the calculation process of the graph convolutional layers, experimental
validation shows that the optimal predictive performance of the model
is achieved when the number of convolutional layers is set to 5.
Through SGCN, the fused omics data features and the relationship graph
among samples are learned, with the output Z^h of the last layer of the
graph convolutional network serving as the final high-level feature
input to the softmax classifier. For a given layer i, this study
employs Relu as the activation function between the input data Z^h and
the output layer y, with the calculation process as follows:
[MATH:
y=fi(Zh<
/msup>)=relu(
Wi(Zh)+bi<
mo>) :MATH]
(8)
Where Z^h and y are two vectors of length l and q, respectively. W[i]
represents the weight matrix of size l×q, and b[i] is a bias vector of
length q. The cross-entropy function is utilized as the loss function,
with the calculation process as follows:
[MATH:
Cross−e<
mi>ntropy=−1n∑i=1
n [yi⋅log(pi)+(1−<
/mo>yi)⋅log(1−pi)]
:MATH]
(9)
Where y[i] represents the label of sample i, with high risk being 1 and
low risk being 0, and p[i] represents the predicted probability of
sample i.
Model training
In this study, DFASGCNS was implemented based on Torch 1.10.0 and
Python 3.6.11. During the training process, the learning rate, number
of epochs, and batch size were set to 0.001, 1000, and 32,
respectively. The Adam algorithm was utilized to optimize the objective
function. To prevent overfitting, dropout and weight decay (L2
regularization) were implemented to ensure the effectiveness of the
model, with dropout rate and weight decay rate set to 0.2 and 0.001,
respectively.
Evaluation metrics
In this paper, the performance of DFASGCNS in predicting ovarian cancer
prognosis was evaluated using the following evaluation metrics:
accuracy (ACC), F1-score, and the area under the receiver operating
characteristic curve (AUC). Accuracy (ACC) is defined as:
[MATH:
ACC=TP+TNTP+<
/mo>TN+FP+FN :MATH]
(10)
where TP, TN, FP, and FN represent true positives, true negatives,
false positives and false negatives, respectively. The F1-score is a
weighted average of precision and recall, and it is defined as:
[MATH:
F1−scor<
mi>e=2×precision
×recall<
/mrow>precision+rec
all :MATH]
(11)
where precision represents the percentage of accurately predicted
positive samples out of all positive samples, and recall represents the
rate of accurately predicted positive samples out of all accurate
positive samples. They are defined as follows:
[MATH:
precisi<
mi>on=TPTP+FP
:MATH]
(12)
[MATH:
recall=<
mfrac>TPTP+<
/mo>FN :MATH]
(13)
The AUC is the area under the ROC curve. The larger the area, the
better the prediction effect of the model.
Results
Selection of k value in k-neighbor graph
The hyperparameter k in the model DFASGCNS represents the size of the
neighborhood in the k-nearest graph. Its selection is determined
through cross-validation on the training data. To assess its
robustness, we trained DFASGCNS using different k values within a
relatively large range. We set the range of k values from 2 to 16, as
depicted in [81]Fig 2. Despite variations in the F1-score with changing
k values, DFASGCNS demonstrated strong robustness, particularly within
the range of k values from 10 to 16, where it achieved the highest
F1-score, indicating optimal model performance. It is noteworthy that
having different k values within a relatively large range suggests the
robustness of the DFASGCNS model in ovarian cancer prognosis
prediction.
Fig 2. The results of ovarian cancer prognosis prediction with different
values of k.
[82]Fig 2
[83]Open in a new tab
Sensitivity analysis of the attention mechanism
To evaluate the effectiveness of the attention mechanism in integrating
multi-omics data for predicting the prognosis of ovarian cancer
patients, we conducted a sensitivity analysis on the number of hidden
units d in the attention mechanism. The aim was to reveal how attention
mechanisms of different scales affect the model’s performance. In this
paper, we used preprocessed mRNA expression, miRNA expression, DNA
methylation, and CNV datasets as inputs for DFASGCNS. After RLASSO
feature selection, we performed five-fold cross-validation experiments
on ovarian cancer prognosis prediction using different numbers of
hidden units d in the attention mechanism. The number of hidden units d
in the attention mechanism of DFASGCNS was set to 50, 80, 110, 140,
170, and 200, respectively, and validated. The experimental results are
shown in [84]Fig 3. In [85]Fig 3, when the number of hidden units in
the attention mechanism of DFASGCNS was set to 140, the Acc value
reached 71.16%, and the model’s performance was optimal. Thereafter, as
the number of hidden units d increased, the Acc showed a downward
trend, indicating that an increase in the number of hidden units in the
attention mechanism of DFASGCNS may lead to gradient vanishing,
reducing the model’s ability to learn important features, thereby
affecting the prediction performance.
Fig 3. Prognostic prediction of OV by different number of hidden units in
attention mechanism.
[86]Fig 3
[87]Open in a new tab
Ablation experiment
Ablation experiments with different omics data
To investigate the impact of different types of omics data on ovarian
cancer prognosis prediction, this study conducted experiments using
single-omics data, arbitrary combinations of two types of omics data,
arbitrary combinations of three types of omics data, and integration of
four omics data sets. The dataset was randomly divided into a 70%
training set and a 30% validation set, and the process was repeated 10
times. The results are presented in Tables [88]4 and [89]5.
Table 4. The prognostic prediction results of ovarian cancer using single
omics data and arbitrary integration of two omics data (%).
mRNA √ √ √ √
DNA methylation √ √ √ √
miRNA √ √ √ √
CNV √ √ √ √
ACC 62.26 59.81 56.28 59.45 64.32 62.47 62.89 61.55 63.36 60.92
F1-score 63.92 61.17 59.76 60.24 69.70 66.19 65.32 61.93 64.70 62.57
AUC 60.59 58.70 56.52 58.99 62.19 60.73 60.68 60.28 59.51 58.99
[90]Open in a new tab
Table 5. The prognostic prediction results of ovarian cancer using arbitrary
integration of three omics data and integration of four omics data (%).
mRNA √ √ √ √
DNA methylation √ √ √ √
miRNA √ √ √ √
CNV √ √ √ √
ACC 67.59 68.38 65.71 64.96 71.16
F1-score 73.72 75.34 72.98 71.23 79.25
AUC 63.13 63.29 62.76 60.60 64.77
[91]Open in a new tab
From [92]Table 4, it is evident that different types of omics data
exhibit distinct performance in ovarian cancer prognosis prediction.
Specifically, when using a single type of omics data, mRNA expression
demonstrates the best performance, followed by DNA methylation, while
miRNA expression and CNV perform the poorest. This difference reflects
the varying roles of different omics molecular information in ovarian
cancer development and prognosis [[93]35,[94]36]. When integrating any
two types of omics data, combining mRNA expression and DNA methylation
yields the best results, while integrating miRNA expression and CNV
data yields the worst results. Furthermore, the prognostic predictions
based on mRNA expression combined with other omics data are superior to
those of integrating two types of omics data without mRNA expression.
This indicates that mRNA expression contributes most prominently to
ovarian cancer prognosis prediction, followed by DNA methylation, while
the contributions of mRNA expression and CNV are relatively minor,
consistent with previous research findings [[95]37].
The results in [96]Table 5 demonstrate that integrating four types of
omics data yields the best performance in ovarian cancer prognosis
prediction, with ACC, F1-score, and AUC values of 71.16%, 79.25%, and
64.77%, respectively. Among the combinations of any three types of
omics data, the combination of mRNA expression, DNA methylation, and
CNV follows the integration of four types of omics data, with
performance indicators of ACC, F1-score, and AUC decreasing by 2.78%,
3.91%, and 1.48%, respectively. Comparatively, integrating DNA
methylation, miRNA expression, and CNV exhibits the largest decrease in
each evaluation metric compared to integrating four types of omics
data, with ACC, F1-score, and AUC decreases of 6.20%, 8.02%, and 4.17%,
respectively. This indicates that the removal of mRNA expression data
has the most significant impact on model performance, further
validating the importance of mRNA expression data. Combining the
analyses from Tabled 4 and 5 reveals that integrating multiple types of
omics data significantly outperforms single or partial omics data,
underscoring the importance of complementary nature and comprehensive
use of multi-omics data to enhance ovarian cancer prognosis prediction.
Consequently, the comprehensive consideration of multiple types of
omics data would provide stronger support for personalized treatment in
clinical practice and the formulation of ovarian cancer management
strategies [[97]2].
Ablation experiments of model structures
To validate the contributions of each module in the proposed model
DFASGCNS to ovarian cancer prognosis prediction, we conducted ablation
experiments by altering different parts of the model configuration. We
evaluated the importance of each module in the DFASGCNS model using the
following five different configurations:
* SGCN: Constructing feature graphs using single omics data that
underwent RLASSO feature selection and applying SGCN on each single
omics graph.
* NoAttention-SGCN: Predicting using fused features after RLASSO and
fusion graph A, concatenating the low-dimensional features
Z^(1)、Z^(2)、Z^(3) and Z^(4) of the four types of omics data without
including an attention layer.
* NoAttention-FF: Employing only fused features after RLASSO and
fully connected layers for prediction, concatenating the
low-dimensional features Z^(1)、Z^(2)、Z^(3) and Z^(4) of the four
types of omics data without attention layers and SGCN module.
* Attention-FF: Utilizing fused multi-omics features with added
attention mechanism and fully connected layers for prediction
without the SGCN module.
* Without-Graph: Removing the graph from the entire model structure
and replacing it with an identity matrix in SGCN.
This paper adjusted different parts of the ovarian cancer dataset to
implement the five variants of DFASGCNS mentioned above. Among them,
SGCN is a single-omics method, NoAttention-SGCN uses a feature fusion
method without attention mechanism, removing the attention mechanism
from DFASGCNS. NoAttention-FF is a feature fusion method without
attention mechanism, and a fully connected network is introduced after
advanced feature fusion, removing the attention mechanism and SGCN from
DFASGCNS. Attention-FF is based on NoAttention-FF, using a feature
fusion method with attention mechanism, removing only SGCN from
DFASGCNS. Without-Graph is a DFASGCNS without graph structure,
replacing the fusion graph A in DFASGCNS with an identity matrix. For
fair comparison, the same neural network is used to extract features of
omics data in the above five configurations, as shown in [98]Table 6.
Table 6. The predictive results using different variant model architectures
(%).
Omics data type Method ACC F1-score AUC
mRNA SGCN 63.28±1.31 64.16±1.87 60.41±1.34
DNA methylation 59.92±1.62 61.32±1.26 57.98±0.97
miRNA 56.78±1.75 59.11±1.60 57.12±1.54
CNV 59.53±1.84 60.59±1.54 58.81±1.04
Integrate four types of omics data NoAttention-SGCN 69.95±1.57
78.62±1.68 64.08±1.25
NoAttention-FF 65.46±1.77 73.94±1.42 62.53±1.84
Attention-FF 66.83±1.59 75.03±1.39 62.97±1.75
Without-Graph 65.38±1.24 74.65±1.80 62.17±1.63
DFASGCNS 71.16±1.52 79.25±1.32 64.77±1.26
[99]Open in a new tab
Comparative experiments with existing methods
To evaluate the effectiveness of the DFASGCNS model in predicting the
prognosis of ovarian cancer, this paper compared the DFASGCNS model
with existing traditional and deep learning-based prognostic prediction
methods. Among them, SVM [[100]38], RF [[101]39], and XGBoost [[102]40]
utilize machine learning methods to integrate multi-omics data for
prognostic prediction of ovarian cancer, while MDNNMD [[103]41], GCGCN
[[104]16], MOLI [[105]42], DeepMO [[106]19], MOGONET [[107]17], MOCSC
[[108]43], and MDCADON [[109]44] are all deep learning methods. Label
predictions were made by a majority vote of KNN in the training data,
with K set to 37 to minimize errors. In RF, multiple decision trees are
combined through ensemble learning to obtain the final prediction
results. XGBoost employs gradient boosting technique for early and
late-stage cancer classification. MDNNMD integrates patient prognosis
prediction results through score fusion. GCGCN adopts graph fusion
method, while MOLI, DeepMO, MOGONET, MOCSC, and MDCADON adopt late
fusion methods for multi-omics data. Among them, DeepMO, MOGONET,
MOCSC, and MDCADON utilize chi-square test, GCN, DAE, and RLASSO
feature selection algorithms for gene screening. To comprehensively
evaluate the proposed DFASGCNS model, this study employed repeated
hold-out cross-validation following the method proposed by Lee et al.
[[110]45] The dataset was randomly divided into 70% training set and
30% testing set. Different methods were trained on the training set to
build prediction models, and evaluation metrics were calculated by
predicting survival on the test patients. The experiments were repeated
10 times, and the results are shown in [111]Table 7.
Table 7. The results of prognosis prediction for ovarian cancer using
DFASGCNS compared to existing methods (%).
Method ACC F1-score AUC
SVM 54.28±1.87 53.59±1.66 54.43±2.25
RF 55.16±3.49 54.17±2.36 53.12±2.11
XGBoost 56.04±2.12 54.66±1.60 54.92±0.87
MDNNMD 60.85±1.88 68.42±2.03 57.22±1.89
GCGCN 64.52±1.75 71.19±1.64 61.71±1.53
MOLI 63.08±1.21 70.94±1.80 59.48±1.62
DeepMO 63.98±1.72 70.25±1.69 60.17±1.23
MOGONET 64.25±2.36 73.16±2.20 57.94±1.98
MOCSC 65.48±1.83 73.45±2.31 59.25±2.17
MDCADON 69.47±2.10 77.91±1.82 63.45±2.15
DFASGCNS 71.16±1.52 79.25±1.32 64.77±1.26
[112]Open in a new tab
In [113]Table 7, it can be observed that the proposed model DFASGCNS
achieves the highest values in evaluation metrics Acc, F1-score, and
AUC, with values of 71.16%, 79.25%, and 64.77% respectively. Compared
to other methods, it demonstrates better accuracy in predicting the
prognosis of ovarian cancer patients. This is attributed to the
consideration of inter-sample correlations in ovarian cancer within the
multi-omics feature fusion of the DFASGCNS model. By utilizing SGCN to
learn the deep network structure of the dual fusion channel and
introducing attention mechanism during feature fusion of different
omics data, DFASGCNS adequately addresses the importance of different
omics features, thus enhancing the model’s performance in predicting
ovarian cancer prognosis. The relatively high F1-score value of the
DFASGCNS model is due to its weighted combination of precision and
recall, which provides relatively accurate results even in the presence
of imbalanced datasets [[114]46]. Overall, the evaluation metrics of
deep learning methods are higher than those of machine learning methods
(SVM, RF, and XGBoost), indicating that deep learning methods can more
fully learn the high-dimensional features of multi-omics data related
to ovarian cancer and extract important information. Among the deep
learning-based methods for integrating multi-omics data, DeepMO
improves upon MOLI by incorporating a feature selection method, leading
to improvements in evaluation metrics Acc and AUC, underscoring the
importance of feature selection in ovarian cancer prognosis prediction.
Both MOGONET and DFASGCNS utilize GCN to learn advanced features of
omics data. However, DFASGCNS not only considers features of different
omics data but also incorporates the graph structure of relationships
between ovarian cancer samples, facilitating a comprehensive
exploration of inter-sample correlations and thereby enhancing ovarian
cancer prognosis prediction performance. Furthermore, compared to the
MDCADON model, the DFASGCNS shows improvements in all evaluation
metrics. In summary, compared to existing methods, the DFASGCNS model
presented in this study can more effectively learn feature
representations of multi-omics data and correlations between samples,
thereby improving the accuracy and reliability of predicting ovarian
cancer prognosis.
External validation of GEO datasets
To investigate the generalization performance of the DFASGCNS model
across different datasets, this study utilized four GEO datasets of
ovarian cancer, including [115]GSE26712 [[116]47], [117]GSE32062
[[118]48], [119]GSE17260 [[120]49], and [121]GSE140082 [[122]50].
Detailed information is provided in [123]Table 8. To ensure the
objectivity and comparability of the experimental results, the GEO
datasets were randomly divided into 70% training set and 30% testing
set, and this process was repeated 10 times to ensure the independence
and randomness of training and testing data. The DFASGCNS model was
compared with existing methods, and the results are illustrated in
[124]Fig 4.
Table 8. The properties of the GEO datasets.
Datasets Sample numbers Data category Gene annotation platform
[125]GSE26712 185 RNA-seq [126]GPL96 Affymetrix
[127]GSE32062 260 Gene expression [128]GPL6480 Agilent
[129]GSE17260 110 Gene expression [130]GPL6480 Agilent
[131]GSE140082 380 Gene expression [132]GPL14951 Illumina
[133]Open in a new tab
Fig 4. The results of DFASGCNS compared to other existing methods on the GEO
datasets.
[134]Fig 4
[135]Open in a new tab
(a) [136]GSE26712 dataset. (b) [137]GSE32062 dataset. (c) [138]GSE17260
dataset. (d) [139]GSE140082 dataset.
From [140]Fig 4, it is evident that across the four GEO datasets, the
evaluation metrics of the DFASGCNS model are higher than those of other
methods. Specifically, in the [141]GSE26712 dataset, the predictive
performance of deep learning methods surpasses that of machine learning
methods, the DFASGCNS model achieving good predictive results in terms
of evaluation metrics ACC and F1-score. Similarly, when training the
DFASGCNS model on the [142]GSE32062, [143]GSE17260, and [144]GSE140082
datasets and evaluating it on the test sets, the DFASGCNS model
exhibits superior predictive capabilities compared to other methods.
Compared to machine learning methods, the performance of the DFASGCNS
model is significantly improved, while the improvement is less
pronounced compared to deep learning methods. This phenomenon can be
attributed to the fact that these four GEO datasets only contain single
types of omics data, while the DFASGCNS model proposed in this study is
capable of predicting on multiple omics data. Therefore, its
performance improvement on GEO datasets is relatively modest.
Nevertheless, the DFASGCNS model still achieves good performance even
on single-omics data, comparable to methods like MOLI and DeepMO,
indicating its good generalization ability and stable and excellent
predictive performance across different datasets. This provides greater
reliability and feasibility for clinical practice in ovarian cancer
prognosis prediction. Thus, the DFASGCNS model proposed in this study
not only performs well on multi-omics data but also demonstrates good
performance on single-omics data, further validating its effectiveness
in ovarian cancer prognosis prediction.
Survival analysis
To further evaluate the performance of DFASGCNS, logrank tests were
conducted to predict whether patients in high-risk and low-risk
subgroups exhibit significantly different survival curves.
Specifically, using the threshold of the median risk ratio, patients in
the ovarian cancer dataset were divided into high-risk and low-risk
subgroups based on the predicted risk ratios. Subsequently, a log-rank
test was performed to evaluate whether there is a significant
difference in the actual survival times between the two sample groups.
The p-value serves as a crucial indicator for assessing the statistical
significance of differences in survival data between groups. A p-value
of less than 0.05 indicates statistical significance, with smaller
p-values suggesting more pronounced differences between the two
subgroups, thereby reflecting a more effective prediction method. In
this study, Kaplan-Meier survival curves were plotted based on the
ovarian cancer prognosis predictions from all deep learning methods
used in the comparison. The log-rank p-values corresponding to each
method were obtained, as shown in [145]Fig 5. The x-axis represents
time, and the y-axis represents the survival rate, as depicted in
[146]Fig 5.
Fig 5. The Kaplan-Meier survival curves of ovarian cancer patients generated
by different methods.
[147]Fig 5
[148]Open in a new tab
From [149]Fig 5, it can be observed that all deep learning methods
applied to ovarian cancer patient survival prognosis yield p-values
less than 0.05, indicating significant effectiveness in prognostic
prediction. By comparing with seven existing methods, it was found that
the model DFASGCNS proposed in this study generated the most
significant p-value (p = 0.0027). This result indicates that the
DFASGCNS model achieves the best performance in survival prediction,
with the survival differences between the high-risk and low-risk
subgroups being more pronounced and highly significant. This highlights
the superiority of the model in predicting ovarian cancer survival
outcomes.This finding further emphasizes the importance and
effectiveness of the DFASGCNS model in ovarian cancer survival
prediction. By introducing a dual-channel fusion strategy in deep
learning methods, which learns feature representations of multiple
omics data and shares and complements neighborhood information among
samples, and by employing attention mechanism to assign corresponding
attention weights to important features of different omics data,
thereby optimizing the feature representation of multiple omics data.
The DFASGCNS can more accurately capture survival differences among
patients and provide more discriminative predictive results. This is
significant for clinical practice and treatment decision-making, as it
helps doctors better understand the prognosis of patients and thus
develop more effective treatment plans.
Enrichment of selected genes
To further investigate the impact of multi-omics data features on the
pathogenesis of ovarian cancer, we employed the RLASSO feature
selection algorithm to identify omics data features. These features
were ranked by their importance using random forests, and the top 20
important gene features were subjected to GO (Gene Ontology) and KEGG
(Kyoto Encyclopedia of Genes and Genomes) pathway analysis via the
online tool Metascape, as shown in [150]Fig 6. The top 20 important
gene features selected in this study include: ADH1B, USP25, VDR,
STK17B, TPST2, CXCL9, PTPRC, CXCL14, TP53, CXCL11, KRAS, PBK, NRGN,
SCNN1A, POLD2, POLR1C, SAR1A, PTEN, NRAS, ZNF826P.
Fig 6. Pathway enrichment analysis of identified genes.
[151]Fig 6
[152]Open in a new tab
The x-axis represents the -log10 p-value for each term, and the y-axis
represents the KEGG pathway terms. (a) GO pathway enrichment analysis.
(b) KEGG pathway enrichment analysis.
Among them, ADH1B encodes a protein that is a member of the alcohol
dehydrogenase family, and it has been proven to promote mesothelial
clearance and ovarian cancer infiltration [[153]51]. CXCL9 has been
confirmed to be associated with ovarian cancer prognosis; it is related
to the survival rate of patients with high-grade serous ovarian cancer
(HGSC) and is an independent marker of good prognosis in HGSC patients
[[154]52]. In ovarian cancer patients, abnormal overexpression of
CXCL14 in serum and ovarian tissues is associated with poor prognosis,
making it a novel auxiliary marker for early diagnosis of ovarian
cancer [[155]53]. TP53 mutations are closely associated with high
recurrence rates, chemotherapy resistance, and shorter survival in
ovarian cancer [[156]54]. CXCL11 is highly expressed in the
immunoreactive subtype, and its ligand CXCL11 and receptor CXCR3
characterize this subtype, with the CXCL11-CXCR3 signaling pathway
being a therapeutic target for ovarian cancer [[157]55]. KRAS mutations
are associated with low-grade serous ovarian cancer, often predicting
poor prognosis, particularly in patients with poor chemotherapy
response [[158]56]. PTEN is a tumor suppressor gene that negatively
regulates the PI3K/AKT signaling pathway. Studies have shown that loss
or mutation of PTEN is associated with increased invasiveness, enhanced
anti-apoptosis, and poorer prognosis in ovarian cancer [[159]57]. From
[160]Fig 6, GO/KEGG enrichment analysis revealed several important
pathways associated with ovarian cancer, providing valuable insights
into the mechanisms of ovarian cancer occurrence, metastasis, and
treatment. Among these, pathways such as vesicle-mediated transport and
human papillomavirus (HPV) infection are related to the development and
metastasis of ovarian cancer. Extracellular vesicles play a significant
role in cell-to-cell communication and have been implicated in tumor
formation and metastatic disease [[161]58]. Additionally, the
identified pathway related to HPV infection has been shown to be highly
associated with ovarian cancer [[162]59]. The Hippo signaling pathway
is a highly conserved pathway that regulates organ size and plays a key
role in ovarian physiology. Dysregulation of the Hippo pathway
contributes to loss of follicular homeostasis and reproductive
disorders such as polycystic ovary syndrome (PCOS), premature ovarian
insufficiency, and ovarian cancer [[163]60]. Furthermore, we also
identified pathways associated with other cancers or diseases,
including hepatocellular carcinoma and Parkinson’s disease [[164]37].
This helps to elucidate the connections between ovarian cancer and
other diseases, providing important clues for deeper exploration of
ovarian cancer pathophysiology.
Generalizability across different cancer types
To validate the generalization capability of the DFASGCNS model, this
study conducted comparative experiments using single-omics and
multi-omics data on the Lower Grade Glioma (LGG) and Lung Squamous Cell
Carcinoma (LUSC) datasets. In the experiments, four types of omics data
(mRNA, DNA methylation, miRNA, and CNV) were used for LGG and LUSC. The
DFASGCNS model first employed the RLASSO feature selection method to
extract key features, which were then integrated using an attention
mechanism. For each type of omics data, the model constructed
corresponding graph structures to capture the relationships between
samples and underlying network structures. By stacking multiple layers
of graph convolutional networks, the model further learned deep
correlations and complex relationship networks among multi-omics
samples. Finally, a Softmax classifier was used for prognosis
prediction of LGG and LUSC. Ten-fold cross-validation was performed,
using Acc, F1-score, and AUC as evaluation metrics. The results of Acc,
F1-score, and AUC for single-omics and multi-omics data on LGG and LUSC
are shown in [165]Table 9. As seen from [166]Table 9, compared to
single-omics prognosis prediction, the AUC, Acc, and F1-score for
multi-omics data on LGG and LUSC all improved, demonstrating the
effectiveness of the DFASGCNS model in integrating multi-omics data.
Table 9. Prognostic results of different cancer datasets (%).
Datasets mRNA √ √ √ √
meth √ √ √ √
miRNA √ √ √ √
Acc 62.12 61.41 60.93 63.36 62.43 62.80 70.44
LGG F1-score 65.31 62.80 61.35 68.54 65.91 63.52 73.62
AUC 60.10 59.44 57.14 61.24 60.78 59.91 62.76
LUSC Acc 63.43 62.52 62.02 66.91 65.04 64.73 71.11
F1-score 66.21 63.11 62.93 70.12 69.90 68.56 74.80
AUC 61.22 58.96 58.31 62.25 61.72 60.42 63.03
[167]Open in a new tab
Additionally, further comparative experiments were conducted between
the DFASGCNS model and other methods on the LGG and LUSC datasets, as
shown in [168]Table 10. The results indicated that the DFASGCNS model
outperformed the other methods in terms of AUC, Acc, and F1-score on
both datasets. Among the comparison methods, SVM, RF, and XGBoost are
machine learning methods, while the others are deep learning methods.
As shown in [169]Table 10, deep learning methods achieved better
performance than machine learning methods on the LGG and LUSC datasets,
demonstrating the superiority of deep learning for cancer prognosis
prediction. Specifically, for LGG prognosis prediction, the DFASGCNS
model achieved AUC, Acc, and F1-score values of 70.44%, 73.62%, and
62.76%, respectively. For LUSC prognosis prediction, the AUC, Acc, and
F1-score values were 71.11%, 74.80%, and 63.03%, respectively. These
results indicate that the DFASGCNS model effectively integrates
multi-omics data and achieves favorable cancer prognosis prediction
performance, further validating the effectiveness and generalizability
of DFASGCNS in predicting the prognosis of different cancer types.
Table 10. Comparison of prognostic results of different methods on different
cancer datasets (%).
Method LGG LUSC
Acc F1-score AUC Acc F1-score AUC
SVM 60.11 63.54 57.90 62.88 65.52 59.25
RF 59.85 64.75 57.13 63.59 64.16 58.19
XGBoost 62.42 64.11 58.22 63.13 65.95 59.36
MDNNMD 65.12 67.21 59.10 64.92 67.21 60.93
GCGCN 64.93 66.96 59.37 65.75 66.49 61.14
MOLI 68.26 68.30 58.68 65.24 69.23 60.74
DeepMO 68.70 67.52 60.33 68.46 68.91 61.62
MOGONET 69.41 69.11 61.46 67.43 70.32 62.05
MOCSC 68.72 70.43 60.86 69.15 71.14 61.93
MDCADON 69.15 72.82 61.95 70.27 72.93 62.40
DFASGCNS 70.44 73.62 62.76 71.11 74.80 63.03
[170]Open in a new tab
Conclusion
This paper proposes an ovarian cancer prognosis prediction model,
DFASGCNS, which uses a dual fusion channel and stacked graph
convolutional network (SGCN). The introduction of attention mechanism
explores the importance of key features from different omics data for
ovarian cancer prognosis prediction. To capture the correlation between
different omics data and fully consider the inter-sample relationships
in ovarian cancer, a dual fusion channel strategy is proposed, enabling
comprehensive learning of feature representations of multiple omics
data and the sharing and complementing of neighborhood information
among samples. The use of SGCN learns the fused features of multiple
omics data and the latent network structure between samples,
facilitating a comprehensive understanding of the complex relationship
network in ovarian cancer multi-omics data. Additionally, external
validation of the DFASGCNS model was conducted using four GEO datasets
of ovarian cancer. The results demonstrate that learning feature
representations of multiple omics data and the relationship graph
between samples using a dual fusion channel can effectively learn
high-dimensional feature representations of multiple omics data from
different perspectives. The utilization of SGCN to capture the latent
network structure of multiple omics data significantly improves the
accuracy of ovarian cancer prognosis prediction. Furthermore,
Kaplan-Meier survival curves show significant differences in survival
subgroups predicted by DFASGCNS, contributing to an in-depth
understanding of the pathogenesis of ovarian cancer and to the search
for new therapeutic strategies.
Although the proposed model achieves promising performance in ovarian
cancer prognosis prediction, there are still some areas for
improvement. In future work, we plan to incorporate histopathological
images of ovarian cancer into the multi-omics data, leveraging image
information for comprehensive research on ovarian cancer and providing
a new research approach for ovarian cancer prognosis prediction.
Data Availability
[171]https://portal.gdc.cancer.gov/.
Funding Statement
This research was funded by the National Natural Science Foundation of
China, grant number 62176177; and the Natural Science Foundation of
Shanxi Province, grant number 202203021211121.
References