Advancing Cancer Prognosis with Hierarchical Fusion of Genomic, Proteomic and Pathology Imaging Data from a Systems Biology Perspective

Junjie Zhou; Bao Xue; Meiling Wang; Wei Shao; Daoqiang Zhang

cs.CVMar 14, 2026

Advancing Cancer Prognosis with Hierarchical Fusion of Genomic, Proteomic and Pathology Imaging Data from a Systems Biology Perspective

Authors: Junjie Zhou, Bao Xue, Meiling Wang, Wei Shao, Daoqiang Zhang

Organizations: The College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics · School of Computer Science, Nanjing University of Posts and Telecommunications

Abstract

To enhance the precision of cancer prognosis, recent research has increasingly focused on multimodal survival methods by integrating genomic data and histology images. However, current approaches overlook the fact that the proteome serves as an intermediate layer bridging genomic alterations and histopathological features while providing complementary biological information essential for survival prediction. This biological reality exposes another architectural limitation: existing integrative analysis studies fuse these heterogeneous data sources in a flat manner that fails to capture their inherent biological hierarchy. To address these limitations, we propose HFGPI, a hierarchical fusion framework that models the biological progression from genes to proteins to histology images from a systems biology perspective. Specifically, we introduce Molecular Tokenizer, a molecular encoding strategy that integrates identity embeddings with expression profiles to construct biologically informed representations for genes and proteins. We then develop Gene-Regulated Protein Fusion (GRPF), which employs graph-aware cross-attention with structure-preserving alignment to explicitly model gene-protein regulatory relationships and generate gene-regulated protein representations. Additionally, we propose Protein-Guided Hypergraph Learning (PGHL), which establishes associations between proteins and image patches, leveraging hypergraph convolution to capture higher-order protein-morphology relationships. The final features are progressively fused across hierarchical layers to achieve precise survival outcome prediction. Extensive experiments on five benchmark datasets demonstrate the superiority of HFGPI over state-of-the-art methods.

Figures & tables

Figure 1: Overview of the HFGPI framework. The pipeline consists of two main components: (1) Feature Extraction : Whole slide histopathology images are partitioned into patches and encoded via a pre-trained vision backbone, while genomic and proteomic data are encoded by our Molecular Tokenizer that integrates identity embeddings with expression profiles. (2) Model Architecture : HFGPI models the biological hierarchy from genes to proteins to histology images. Specifically, Gene-Regulated Protein Fusion (GRPF) employs graph-aware cross-attention with structure-preserving alignment to obtain the gene-regulated protein embeddings. Then, Protein-Guided Hypergraph Learning (PGHL) associates proteins with semantically related patches through hypergraph convolution, capturing higher-order protein-morphology relationships. Finally, the hierarchically fused features are aggregated via Transformer encoder and attention pooling to produce patient-level embeddings for survival prediction.

Model	G.	P.	I.	BLCA	BRCA	GBMLGG	LUAD	UCEC	Overall
Gene exp. [ 29 ]	✓			0.672 $\pm$ 0.059	0.635 $\pm$ 0.072	0.805 $\pm$ 0.050	0.627 $\pm$ 0.043	0.645 $\pm$ 0.098	0.677
Protein exp. [ 29 ]		✓		0.648 $\pm$ 0.057	0.666 $\pm$ 0.112	0.754 $\pm$ 0.090	0.585 $\pm$ 0.056	0.580 $\pm$ 0.133	0.647
ABMIL [ 21 ]			✓	0.657 $\pm$ 0.044	0.698 $\pm$ 0.061	0.788 $\pm$ 0.065	0.602 $\pm$ 0.034	0.582 $\pm$ 0.112	0.665
CLAM-SB [ 35 ]			✓	0.672 $\pm$ 0.027	0.679 $\pm$ 0.061	0.790 $\pm$ 0.090	0.612 $\pm$ 0.036	0.585 $\pm$ 0.186	0.668
CLAM-MB [ 35 ]			✓	0.682 $\pm$ 0.038	0.675 $\pm$ 0.086	0.790 $\pm$ 0.087	0.632 $\pm$ 0.012	0.594 $\pm$ 0.123	0.675
TransMIL [ 47 ]			✓	0.688 $\pm$ 0.043	0.691 $\pm$ 0.049	0.814 $\pm$ 0.078	0.618 $\pm$ 0.052	0.586 $\pm$ 0.115	0.679

Table 1: Comparisons of C-index (mean ± std) with SOTA methods over five cancer datasets. G., P. and I. refer to genomic, proteomic and pathology imaging data, respectively. The best results and the second-best results are highlighted with bold and in underline , respectively. Methods marked with † are variants that replace the original textual data with proteomic data.

Ablation	Model	Avg.
Full Model	HFGPI	0.753
Modalities	$g,p,i$ $\Rightarrow$ $g,p$	$0.708$ ( $-4.5\%$ )
	$g,p,i$ $\Rightarrow$ $p,i$	$0.708$ ( $-4.5\%$ )
	$g,p,i$ $\Rightarrow$ $g,i$	$0.713$ ( $-4.0\%$ )
Tokenizer	Molecular. $\Rightarrow$ Families	$0.739$ ( $-1.4\%$ )
Tokenizer	Molecular. $\Rightarrow$ Pathways	$0.743$ ( $-1.0\%$ )

Table 2: Ablation study on key components, with C-index averaged across five datasets.

Table 3: Performance of using different (a) VLMs and (b) LLMs on average C-index across five datasets.

Figure 2: Visualization results on BRCA dataset using PCNA as an exemplar protein.

Figure 3: Kaplan-Meier Analysis of predicted high-risk (red) and low-risk (blue) groups on five cancer datasets.

$N_{g}$	BLCA	BRCA	GBMLGG	LUAD	UCEC	Overall
1000	0.711 $\pm$ 0.033	0.709 $\pm$ 0.049	0.862 $\pm$ 0.056	0.650 $\pm$ 0.042	0.779 $\pm$ 0.074	0.742
2000	0.717 $\pm$ 0.022	0.715 $\pm$ 0.043	0.873 $\pm$ 0.064	0.680 $\pm$ 0.039	0.782 $\pm$ 0.062	0.753
3000	0.710 $\pm$ 0.066	0.698 $\pm$ 0.038	0.847 $\pm$ 0.071	0.662 $\pm$ 0.042	0.751 $\pm$ 0.071	0.734

Table 4: Impact of the number of selected genes (

N_{g}

) on model performance across five TCGA datasets.

$k_{g}$	BLCA	BRCA	GBMLGG	LUAD	UCEC	Overall
50	0.703 $\pm$ 0.037	0.715 $\pm$ 0.027	0.864 $\pm$ 0.087	0.674 $\pm$ 0.041	0.778 $\pm$ 0.062	0.747
100	0.717 $\pm$ 0.022	0.715 $\pm$ 0.043	0.873 $\pm$ 0.064	0.680 $\pm$ 0.039	0.782 $\pm$ 0.062	0.753
150	0.711 $\pm$ 0.055	0.705 $\pm$ 0.069	0.865 $\pm$ 0.090	0.668 $\pm$ 0.051	0.763 $\pm$ 0.073	0.742

Table 5: Impact of the number of neighbors (

k_{g}

) in gene graph construction on model performance across five TCGA datasets.

$k_{p}$	BLCA	BRCA	GBMLGG	LUAD	UCEC	Overall
10	0.698 $\pm$ 0.047	0.694 $\pm$ 0.050	0.842 $\pm$ 0.131	0.670 $\pm$ 0.065	0.762 $\pm$ 0.081	0.733
20	0.717 $\pm$ 0.022	0.715 $\pm$ 0.043	0.873 $\pm$ 0.064	0.680 $\pm$ 0.039	0.782 $\pm$ 0.062	0.753
30	0.699 $\pm$ 0.042	0.717 $\pm$ 0.037	0.854 $\pm$ 0.103	0.673 $\pm$ 0.034	0.781 $\pm$ 0.031	0.745

Table 6: Impact of the number of neighbors (

k_{p}

) in protein graph construction on model performance across five TCGA datasets.

$k$	BLCA	BRCA	GBMLGG	LUAD	UCEC	Overall
8	0.694 $\pm$ 0.031	0.684 $\pm$ 0.039	0.871 $\pm$ 0.047	0.665 $\pm$ 0.038	0.730 $\pm$ 0.069	0.729
16	0.707 $\pm$ 0.061	0.697 $\pm$ 0.051	0.878 $\pm$ 0.049	0.672 $\pm$ 0.048	0.770 $\pm$ 0.024	0.745
32	0.717 $\pm$ 0.022	0.715 $\pm$ 0.043	0.873 $\pm$ 0.064	0.680 $\pm$ 0.039	0.782 $\pm$ 0.062	0.753
64	0.707 $\pm$ 0.053	0.708 $\pm$ 0.041	0.866 $\pm$ 0.080	0.678 $\pm$ 0.025	0.775 $\pm$ 0.122	0.747
128	0.703 $\pm$ 0.055	0.704 $\pm$ 0.051	0.864 $\pm$ 0.071	0.683 $\pm$ 0.024	0.760 $\pm$ 0.110	0.743

Table 7: Impact of the number of top-

k

patches selected per protein in PGHL on model performance across five TCGA datasets.

Metric	0	0.1	0.3	0.5	0.7	0.9	1.0
Average C-index	0.738	0.741	0.753	0.747	0.741	0.735	0.730

Table 8: Impact of the regularization parameter (

\lambda

) on overall model performance (average C-index) across five TCGA datasets.

Figure 4: Visualization results on BLCA dataset using EGFR as an exemplar protein.

Figure 5: Visualization results on GBMLGG dataset using CD31 as an exemplar protein.

Methods	#Params	Inference Time	Training Time
Methods	(M)	(s / slide)	(s)
PS3	1.08	0.013	1876
ICFNet	6.18	0.104	11232
Ours	2.03	0.017	2228

Table 9: Comparisons of model complexity and efficiency. We report the number of parameters (MB), inference time per slide (ms), and training time (s) on the BRCA dataset.

Cancer Survival Prediction

Cancer survival prediction estimates patient outcomes and time to events from clinical, molecular, and medical-imaging data.

See all 92 papers

HGP-Mamba: Integrating Histology and Generated Protein Features for Mamba-based Multimodal Survival Risk Prediction
Mar 17, 2026Jing Dai, Chen Wu, Ming Wu +4Cancer Survival Prediction Multimodal Survival Analysis
Cancer Hospital of Dalian University of Technology, Shenyang, China · School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology, Dalian, China · Key Laboratory of Integrated Circuit and Biomedical Electronic System, Dalian University of Technology, Dalian, China
HGASurv: Hierarchical Granularity-Aware Multimodal Integration for Survival Prediction
Oct 19, 2026Peike Yu, Wuchao Liu, Jihong Guan +1Multimodal Survival Analysis Cancer Survival Prediction
School of Computer Science and Technology, Tongji University, 1239 Siping Road, Shanghai 200092, China · College of Computer Science and Artificial Intelligence, Fudan University, 2005 Songhu Road, Shanghai 200438, China
ProtoPathway: Biologically Structured Prototype-Pathway Fusion for Multimodal Cancer Survival Prediction
May 20, 2026Amaya Gallagher-Syed, Costantino Pitzalis, Myles J. Lewis +2Cancer Survival Prediction Oncology
Queen Mary University of London

Mar 17, 2026J/K move · Enter open · S save

HGP-Mamba: Integrating Histology and Generated Protein Features for Mamba-based Multimodal Survival Risk Prediction

Jing Dai, Chen Wu, Ming Wu, Qibin Zhang+3

Cancer Hospital of Dalian University of Technology, Shenyang, China · School of Biomedical Engineering, Faculty of Medicine, Dalian University of Technology, Dalian, China · Key Laboratory of Integrated Circuit and Biomedical Electronic System, Dalian University of Technology, Dalian, China

Cancer Survival Prediction Multimodal Survival Analysis Computational Pathology Multimodal Fusion Whole-Slide Image Analysis State Space Models

Open paper AskPDF arXiv

Recent advances in multimodal learning have significantly improved cancer survival risk prediction. However, the joint prognostic potential of protein markers and histopathology images remains underexplored, largely due to the high cost and limited availability of protein expression profiling. To address this challenge, we propose HGP-Mamba, a Mamba-based multimodal framework that efficiently integrates histological with generated protein features for survival risk prediction. Specifically, we introduce a protein feature extractor (PFE) that leverages pretrained foundation models to derive high-throughput protein embeddings directly from Whole Slide Images (WSIs), enabling data-efficient incorporation of molecular information. Together with histology embeddings that capture morphological patterns, we further introduce the Local Interaction-aware Mamba (LiAM) for fine-grained feature interaction and the Global Interaction-enhanced Mamba (GiEM) to promote holistic modality fusion at the slide level, thus capture complex cross-modal dependencies. Experiments on four public cancer datasets demonstrate that HGP-Mamba achieves state-of-the-art performance while maintaining superior computational efficiency compared with existing methods. Our source code is publicly available at https://github.com/Daijing-ai/HGP-Mamba.git.

Abstract

Cancer Survival Prediction

Explore similar work

HGP-Mamba: Integrating Histology and Generated Protein Features for Mamba-based Multimodal Survival Risk Prediction

HGASurv: Hierarchical Granularity-Aware Multimodal Integration for Survival Prediction

ProtoPathway: Biologically Structured Prototype-Pathway Fusion for Multimodal Cancer Survival Prediction