Heterogeneous Graph Attention Network With Multimodal Feature Selection for Cancer Multiomics Integration

Sina Tabakhi; Charlotte Vandermeulen; Ian Sudbery; Haiping Lu

cs.LGAug 5, 2024

Heterogeneous Graph Attention Network With Multimodal Feature Selection for Cancer Multiomics Integration

Authors: Sina Tabakhi, Charlotte Vandermeulen, Ian Sudbery, Haiping Lu

Organizations: School of Computer Science, The University of Sheffield, Sheffield, UK · Centre for Machine Intelligence, The University of Sheffield, Sheffield, UK · School of Biosciences, The University of Sheffield, Sheffield, UK

Abstract

The increase in high-dimensional multiomics data demands advanced integration models to capture the complexity of human diseases. Graph-based deep learning integration models, despite their promise, struggle with small patient cohorts and high-dimensional features, often applying independent feature selection without modeling relationships among omics. Furthermore, conventional graph-based omics models focus on homogeneous graphs, lacking multiple types of nodes and edges to capture diverse structures. We introduce a Heterogeneous Graph ATtention network for omics integration (HeteroGATomics) to improve cancer diagnosis. HeteroGATomics performs multimodal feature selection through a multi-agent system, creating dedicated networks of feature and patient similarity for each omics modality. These networks are then combined into one heterogeneous graph for learning holistic omics-specific representations and integrating predictions across modalities. Experiments on three cancer multiomics datasets show that HeteroGATomics outperforms baseline methods across most evaluation metrics in cancer diagnosis. External validation on an independent breast cancer cohort further demonstrates its generalizability, where HeteroGATomics outperforms graph-based multiomics integration methods across evaluation metrics. Moreover, HeteroGATomics enhances interpretability by identifying important biomarkers contributing to the diagnosis outcomes.

Figures & tables

Figure 1: HeteroGATomics stage I: multimodal feature selection. First, HeteroGATomics represents the preprocessed multiomics data as feature similarity networks, where each network represents a specific omics modality with nodes corresponding to features and edges denoting their correlations. Then, a multi-agent system (MAS) selects informative features, considering both intra- and cross-modality interactions. The MAS iteratively constructs and evaluates candidate feature subsets while updating dynamic node scores, dynamic edge weights, and omics importance. After the predefined number of iterations, the MAS selects the top $B$ features based on the weighted node scores and constructs the multiomics data with the selected features for heterogeneous graph learning in stage II.

Figure 2: HeteroGATomics stage II: heterogeneous graph learning . First, HeteroGATomics builds a patient similarity network for each omics modality and combines it with the feature similarity network to construct a heterogeneous graph ( $\mathcal{G}_{H}^{m}$ ). Then, stacked GAT layers ( L ) learn heterogeneous graph representations, where each layer uses three GATs to model the three relation types in the graph. Afterward, a single-layer neural network predicts patient labels for each omics modality ( $\hat{\mathbf{Y}}^{m}$ ). Finally, late fusion combines the modality-specific predictions into a cross-omics discovery tensor and feeds it into VCDN to generate the final prediction ( $\hat{\mathbf{Y}}$ ) for downstream tasks and interpretation.

Dataset	Categories	#Patients	#Original features			#Preprocessed features
			DNA	mRNA	miRNA	DNA	mRNA	miRNA
BLCA	High-grade: 397, Low-grade: 21	418	485,577	20,530	2,210	7,999	2,373	249
LGG	Grade II: 254, Grade III: 268	522	485,577	20,530	2,157	8,277	1,166	287
RCC	KICH: 65, KIRC: 201, KIRP: 294	560	485,577	20,530	1,847	4,107	2,456	238

Table 1: A Summary of multiomics data characteristics.

Method	AUROC	Accuracy	BACC	MCC	NPV	PPV	Sensitivity	Specificity
KNN	0.779 $\pm$ 0.191	0.955 $\pm$ 0.027	0.779 $\pm$ 0.191	0.517 $\pm$ 0.331	0.978 $\pm$ 0.020	0.517 $\pm$ 0.329	0.583 $\pm$ 0.382	0.975 $\pm$ 0.020
MLP	0.744 $\pm$ 0.194	0.962 $\pm$ 0.026	0.744 $\pm$ 0.194	0.513 $\pm$ 0.368	0.973 $\pm$ 0.023	0.583 $\pm$ 0.417	0.500 $\pm$ 0.387	0.987 $\pm$ 0.017
RF	0.684 $\pm$ 0.149	0.955 $\pm$ 0.022	0.684 $\pm$ 0.149	0.444 $\pm$ 0.317	0.968 $\pm$ 0.016	0.600 $\pm$ 0.436	0.383 $\pm$ 0.299	0.985 $\pm$ 0.020
Ridge	0.720 $\pm$ 0.129	0.962 $\pm$ 0.016	0.720 $\pm$ 0.129	0.513 $\pm$ 0.271	0.971 $\pm$ 0.018	0.650 $\pm$ 0.391	0.450 $\pm$ 0.269	0.990 $\pm$ 0.017
XGBoost	0.760 $\pm$ 0.175	0.964 $\pm$ 0.016	0.760 $\pm$ 0.175	0.552 $\pm$ 0.302	0.976 $\pm$ 0.019	0.650 $\pm$ 0.369	0.533 $\pm$ 0.356	0.987 $\pm$ 0.013
mRMR-mv	0.683 $\pm$ 0.146	0.952 $\pm$ 0.018	0.683 $\pm$ 0.146	0.402 $\pm$ 0.288	0.968 $\pm$ 0.015	0.500 $\pm$ 0.387	0.383 $\pm$ 0.299	0.982 $\pm$ 0.020

Table 2: Classification performance comparison (mean

\pm

standard deviation over 10-fold cross-validation) for the BLCA dataset. The best performance for each metric is in bold; second-best results are underlined.

Method	AUROC	Accuracy	BACC	MCC	NPV	PPV	Sensitivity	Specificity
KNN	0.670 $\pm$ 0.052	0.667 $\pm$ 0.053	0.670 $\pm$ 0.052	0.353 $\pm$ 0.102	0.630 $\pm$ 0.056	0.737 $\pm$ 0.066	0.552 $\pm$ 0.116	0.788 $\pm$ 0.077
MLP	0.697 $\pm$ 0.043	0.695 $\pm$ 0.044	0.697 $\pm$ 0.043	0.396 $\pm$ 0.085	0.673 $\pm$ 0.050	0.726 $\pm$ 0.043	0.653 $\pm$ 0.084	0.740 $\pm$ 0.050
RF	0.704 $\pm$ 0.055	0.703 $\pm$ 0.056	0.704 $\pm$ 0.055	0.415 $\pm$ 0.109	0.686 $\pm$ 0.071	0.735 $\pm$ 0.058	0.664 $\pm$ 0.118	0.745 $\pm$ 0.083
Ridge	0.650 $\pm$ 0.049	0.649 $\pm$ 0.049	0.650 $\pm$ 0.049	0.303 $\pm$ 0.098	0.634 $\pm$ 0.048	0.671 $\pm$ 0.059	0.631 $\pm$ 0.073	0.670 $\pm$ 0.078
XGBoost	0.679 $\pm$ 0.078	0.678 $\pm$ 0.078	0.679 $\pm$ 0.078	0.365 $\pm$ 0.161	0.657 $\pm$ 0.076	0.715 $\pm$ 0.108	0.638 $\pm$ 0.107	0.721 $\pm$ 0.114
mRMR-mv	0.687 $\pm$ 0.051	0.686 $\pm$ 0.053	0.687 $\pm$ 0.051	0.380 $\pm$ 0.099	0.670 $\pm$ 0.069	0.716 $\pm$ 0.047	0.645 $\pm$ 0.126	0.729 $\pm$ 0.072

Table 3: Classification performance comparison (mean

\pm

standard deviation over 10-fold cross-validation) for the LGG dataset. The best performance for each metric is in bold; second-best results are underlined.

Method	Accuracy	Macro F1	Micro F1	Weighted F1	Precision	Recall
KNN	0.946 $\pm$ 0.028	0.944 $\pm$ 0.025	0.946 $\pm$ 0.028	0.947 $\pm$ 0.028	0.949 $\pm$ 0.027	0.946 $\pm$ 0.028
MLP	0.954 $\pm$ 0.024	0.953 $\pm$ 0.024	0.954 $\pm$ 0.024	0.953 $\pm$ 0.025	0.958 $\pm$ 0.021	0.954 $\pm$ 0.024
RF	0.950 $\pm$ 0.026	0.950 $\pm$ 0.020	0.950 $\pm$ 0.026	0.950 $\pm$ 0.026	0.953 $\pm$ 0.025	0.950 $\pm$ 0.026
Ridge	0.954 $\pm$ 0.024	0.949 $\pm$ 0.025	0.954 $\pm$ 0.024	0.954 $\pm$ 0.024	0.956 $\pm$ 0.024	0.953 $\pm$ 0.024
XGBoost	0.955 $\pm$ 0.022	0.955 $\pm$ 0.019	0.955 $\pm$ 0.022	0.955 $\pm$ 0.022	0.959 $\pm$ 0.020	0.955 $\pm$ 0.021
mRMR-mv	0.954 $\pm$ 0.021	0.953 $\pm$ 0.018	0.954 $\pm$ 0.021	0.954 $\pm$ 0.022	0.956 $\pm$ 0.021	0.954 $\pm$ 0.021

Table 4: Classification performance comparison (mean

\pm

standard deviation over 10-fold cross-validation) for the RCC dataset. The best performance for each metric is in bold; second-best results are underlined.

Figure 3: External validation on METABRIC. Models are trained on TCGA-BRCA (n = 875) and tested on METABRIC (n = 1,272) for five-class breast cancer subtype classification using mRNA expression and DNA methylation. Numbers in parentheses indicate the total number of selected features across both modalities.

Figure 4: Performance of HeteroGATomics MAS with different numbers of agents for random forest classification on the LGG dataset. The averaged values from 10-fold cross-validation are reported for each metric. Results are reported for 1, 5, and 10 agents per omics modality.

Figure 5: Performance comparison of feature selection methods for random forest classification on LGG. The averaged values from 10-fold cross-validation are reported for each metric.

Figure 6: Performance comparison of feature selection methods for random forest and Ridge classification for (a) the BLCA, (b) the LGG, and (c) the RCC datasets. The vertical bars show the mean, the black lines represent error bars indicating plus/minus one standard deviation, and each dot is a model’s performance on each fold.

Figure 7: Ablation studies to evaluate the feature selection module across five classifiers. HeteroGATomics MAS + [classifier] denotes the results of the feature selection module within HeteroGATomics for a classifier, while HeteroGATomics represents the results derived from the entire HeteroGATomics architecture.

Figure 8: Ablation studies to evaluate heterogeneous graph elements using LGG.

Figure 9: Ablation studies to evaluate the impact of the fusion strategy on HeteroGATomics using LGG.

Figure 10: Ablation studies to evaluate the impact of different omics combinations on HeteroGATomics performance using LGG.

ID	Design question	Ablation study	Key performance change
A1	Why use multiple agents per omics modality instead of a single agent?	1 vs 10 agents per omics modality (Fig. 4 ) RF classifier on LGG dataset 20–100 selected features	+2.2–5.3% AUROC +2.4–5.1% Accuracy
A2	Why use HeteroGATomics MAS rather than existing unimodal and multimodal feature selection methods?	HeteroGATomics MAS vs best feature selection baseline (Fig. 6 ) 100 features selected per omics modality	BLCA: +4.0% AUROC (RF) +5.6% AUROC (Ridge) LGG: +0.7% AUROC (RF) $-$ 0.7% AUROC (Ridge) RCC: $-$ 0.2% Weighted F1 (RF) $-$ 1.3% Weighted F1 (Ridge)
A3	Why add graph learning and multimodal fusion on top of feature selection?	Full HeteroGATomics vs HeteroGATomics MAS with the best classifier (Fig. 7 )	+20.9% AUROC (BLCA) +8.1% AUROC (LGG) +0.4% Weighted F1 (RCC)
A4	Why use heterogeneous rather than homogeneous graphs?	Heterogeneous (patient and feature similarity networks) vs homogeneous (patient similarity network) graphs (Fig. 8 ) LGG dataset	+2.1% AUROC +3.3% Accuracy +3.9% Sensitivity +3.1% Specificity
A5	Why retain node and edge information from the HeteroGATomics MAS module?	Both node and edge information vs only one of them (best variant) (Fig. 8 ) LGG dataset	+0.4% AUROC +2.0% Accuracy +2.5% Sensitivity +1.5% Specificity
A6	Why use VCDN for multimodal fusion?	VCDN vs intermediate fusion (concatenation of embeddings) (Fig. 9 ) LGG dataset	+0.5% AUROC +4.6% Accuracy +2.0% Sensitivity +7.0% Specificity

Table 5: Selected ablation results and quantified contributions of key HeteroGATomics design choices. Values are relative changes in mean performance over 10-fold cross-validation. Positive values indicate improvements of the proposed design over the corresponding ablated or simplified alternative.

Figure 11: Known partners of selected top biomarkers for (a) the BLCA dataset and (b) the LGG dataset. Direct protein-protein interactions are recovered for DNA and mRNA omics. For the miRNA omics, known mRNA targets are recovered from starBase.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Notation	Description
${\mathbf{a}}_{r,s}^{(l)},{\mathbf{a}}_{r,d}^{(l)},{\mathbf{a}}_{r,e}^{(l)}$	Attention vectors to weigh source, destination, and edge attributes for relation $r$ in $l$ th layer
$\mathcal{D}^{m}=<\mathbf{X}^{m},\mathbf{Y}>$	Omics $m$ with patients $\mathbf{X}^{m}\in\mathbb{R}^{N\times d_{m}}$ and labels $\mathbf{Y}\in\mathbb{R}^{{N\times C}}$
${\mathbf{e}}_{uv}$	Edge attributes between nodes $u$ and $v$
$\eta_{\mathcal{V}}(\cdot),\eta_{\mathcal{E}}(\cdot)$	Static node score and edge weight
$\mathcal{G}^{m}=(\mathcal{V}^{m},\mathcal{E}^{m},w_{m})$	Undirected weighted graph with nodes $\mathcal{V}^{m}$ , edges $\mathcal{E}^{m}$ , and omics importance $w_{m}$
$\mathcal{G}_{H}^{m}=(\mathcal{V}^{m},\mathcal{E}^{m},\mathcal{O},\mathcal{R})$	Heterogeneous graph with nodes $\mathcal{V}^{m}$ , edges $\mathcal{E}^{m}$ , and corresponding types $\mathcal{O}$ , $\mathcal{R}$

Appendix

Table 6: Notations and descriptions.

Figure 12: HeteroGATomics classification performance across datasets with different hyperparameter configurations. (a) Results for the BLCA dataset with AUROC. (b) Results for the LGG dataset with AUROC. (c) Results for the RCC dataset with Weighted F1. Averaged values from 10-fold cross-validation are presented. Each fold divides the training set into training and validation sets, with the respective evaluation metric reported for the validation set.

Hyperparameter	Range ∗	BLCA	LGG	RCC
Number of heads	{ 2 , 3, 4}	3	2	2
Sparsity rate	{ 0.80 , 0.85, 0.90}	0.90	0.85	0.80
Dropout rate	{ 0.0 , 0.1, 0.2, 0.3}	0.0	0.3	0.0
Pre-training lr	{0.050, 0.010 , 0.001}	0.010	0.001	0.001
Training lr	{0.050, 0.010, 0.001 }	0.001	0.001	0.001
VCDN lr	{0.050, 0.010 , 0.001}	0.05	0.05	0.05

Appendix

Table 7: Dataset-specific hyperparameter configurations for HeteroGATomics.

Module	Hyperparameter	Value
HeteroGATomics MAS	Number of iterations	50
	Number of agents per omics	10
	Initial dynamic node score $(c_{\mathcal{V}})$	0.2
	Initial dynamic edge weight $(c_{\mathcal{E}})$	0.2
	Node decay coefficient $(\rho_{\mathcal{V}})$	0.1
	Edge decay coefficient $(\rho_{\mathcal{E}})$	0.1

Appendix

Table 8: Common hyperparameter configurations for HeteroGATomics across datasets.

Figure 13: Average execution times per fold over 10-fold cross-validation. (a) Execution time of HeteroGATomics MAS for different numbers of selected features across three datasets. (b) Execution time of HeteroGATomics GAT for different combinations of omics modalities on LGG.

Figure 14: Comparison of classification performance and execution time across methods, averaged over 10 cross-validation folds. Execution time is shown on a base-10 logarithmic scale.

Dataset	#Parameters	Peak GPU memory (MB)
BLCA	1.50 M	868
LGG	1.13 M	952
RCC	1.17 M	1233

Appendix

Table 9: Number of trainable parameters and peak GPU memory of HeteroGATomics GAT , averaged over 10-fold cross-validation.

Figure 15: Statistical comparison of HeteroGATomics with each baseline based on the results in Tables 2–4. Color shows the mean difference in performance over the 10 cross-validation folds (green: HeteroGATomics is better; pink: the baseline is better). Stars denote the significance of a two-sided Wilcoxon signed-rank test across folds (***: $p<0.01$ , **: $p<0.05$ , *: $p<0.1$ ).

Rank	Biomarker ID	Biomarker name ∗	Omics
1	MIMAT0000416	hsa-mir-1-3p	miRNA
2	MIMAT0004926	hsa-mir-708-5p	miRNA
3	MIMAT0004518	hsa-mir-16-2-3p	miRNA
4	MIMAT0009451	hsa-mir-1976	miRNA
5	MIMAT0000080	hsa-mir-24-3p	miRNA
6	DBC1	BRINP1	mRNA

Appendix

Table 10: Top 30 ranked biomarkers identified by HeteroGATomics in the BLCA dataset.

Rank	Biomarker ID	Biomarker name ∗	Omics
1	cg15024277	BEX3	DNA
2	cg20371266	CHRND	DNA
3	cg22373770	N/A	DNA
4	cg05165025	MIDEAS,RP5-1021I20.1	DNA
5	cg16503259	N/A	DNA
6	cg20253855	CUX1	DNA

Appendix

Table 11: Top 30 ranked biomarkers identified by HeteroGATomics in the LGG dataset.

Figure 16: Interaction network of top 30 biomarkers with known partners for BLCA. Direct protein-protein interactions are recovered for DNA and mRNA omics. For the miRNA omics, known mRNA targets are recovered from starBase [ 13 ] .

Figure 17: Interaction network of top 30 biomarkers with known partners for LGG. Direct protein-protein interactions are recovered for DNA and mRNA omics. For the miRNA omics, known mRNA targets are recovered from starBase [ 13 ] .

Figure 18: GO enrichment analysis of the top 30 biomarkers from the DNA and mRNA omics modalities for (a) LGG and (b) BLCA. The y-axis shows the top 10 most significant GO category terms, while the x-axis represents the percentage of biomarkers belonging to each GO category.

Multi-Omics Data Integration

See all 23 papers

CMGL: Confidence-guided Multi-omics Graph Learning for Cancer Subtype Classification
Apr 27, 2026Boyang Fan, Hengchuang Yin, Siyu Yi +5Oncology Cancer Subtype Classification
College of Computer Science, Sichuan University, Chengdu, China · Xinjiang Technical Institute of Physics and Chemistry, Chinese Academy of Sciences, Urumqi, China · College of Mathematics, Sichuan University, Chengdu, China +1
PathMoG: A Pathway-Centric Modular Graph Neural Network for Multi-Omics Survival Prediction
Apr 27, 2026Di Wang, Chupei Tang, Junxiao Kong +3Survival Analysis Graph Neural Networks
1Cuiying Honors College, Lanzhou University, China · School of Mathematics and Statistics, Lanzhou University, China · School of Mathematics and Statistics, Lanzhou University, China and Shanghai Innovation Institute +1
Biologically Informed Deep Neural Networks for Multi-Omic Integration, Pathway Activity Inference and Risk Stratification in Cancer
Jul 6, 2026Pedro Henrique da Costa Avelar, Le Ou-Yang, Min Wu +1Cancer Survival Prediction Autoencoders
Department of Informatics, King’s College London, London WC2B 4BG, United Kingdom · Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR), Singapore 138632, Singapore · Division of Informatics, Imaging and Data Science, Faculty of Biology, Medicine and Health, University of Manchester +1

Apr 27, 2026J/K move · Enter open · S save

CMGL: Confidence-guided Multi-omics Graph Learning for Cancer Subtype Classification

Boyang Fan, Hengchuang Yin, Siyu Yi, Yifan Wang+4

College of Computer Science, Sichuan University, Chengdu, China · Xinjiang Technical Institute of Physics and Chemistry, Chinese Academy of Sciences, Urumqi, China · College of Mathematics, Sichuan University, Chengdu, China+1

Oncology Cancer Subtype Classification Multimodal Graph Learning Multimodal Fusion Multi-Omics Data Integration

Open paper AskPDF arXiv

Motivation: Multi-omics integration can improve cancer subtyping, but modality informativeness and noise vary across cancer types and patients. Most graph methods for multi-omics data learn modality contributions within the downstream classification objective, leaving predictive reliability for each patient implicit. As a result, uninformative modalities can weaken the fused representation, while unreliable omics can introduce noisy patient relationships into graph propagation. To address these two problems, we propose CMGL, which produces a separate reliability estimate before fusion and uses consensus patient neighborhoods for graph classification. Results: CMGL estimates modality confidence for each patient through evidential deep learning, fixes these values during fusion across omics, and performs classification on an independently specified consistency graph. On four MLOmics cancer-subtype tasks and the 32-class pan-cancer task, CMGL consistently improves over the strongest baseline, surpassing it by 4.03% in average accuracy on the four single-cancer tasks. Its representations recover the PAM50 intrinsic subtypes of breast invasive carcinoma (BRCA), and the model trained on BRCA transfers without fine tuning to kidney renal clear cell carcinoma (KIRC), stratifying patients into prognostically distinct groups.

Abstract

Multi-Omics Data Integration

Explore similar work

CMGL: Confidence-guided Multi-omics Graph Learning for Cancer Subtype Classification

PathMoG: A Pathway-Centric Modular Graph Neural Network for Multi-Omics Survival Prediction

Biologically Informed Deep Neural Networks for Multi-Omic Integration, Pathway Activity Inference and Risk Stratification in Cancer