M3NetFlow: Multi-scale, Multi-hop Graph AI for Integrative Multi-omic Analysis

Overview

Multi-omic data-driven studies are at the forefront of precision medicine by characterizing complex disease signaling systems across multiple views and levels. The integration and interpretation of multi-omic data are critical for identifying disease targets and deciphering disease signaling pathways, yet remain challenging due to dense, multi-protein interactions.
M3NetFlow addresses this gap with a multi-scale, multi-hop, multi-omic network flow model that supports both hypothesis-guided and generic analyses. We validate M3NetFlow on two case studies:

1. uncovering mechanisms behind synergistic drug combinations (anchor-target guided analysis), and
2. identifying biomarkers of Alzheimer’s disease (AD) (generic analysis).
   Across evaluations, M3NetFlow achieves state-of-the-art predictive performance and highlights drug-synergy and disease-associated targets, and can be readily applied to other multi-omic tasks.

---

Contents

• Overview
• 1. Parsing the Raw Data Files
  • 1.1 Cancer datasets preprocessing
  • 1.2 ROSMAP AD dataset preprocessing
• 2. Run the Model
  • 2.1 Run M3NetFlow
  • 2.2 Run baselines
  • 2.3 Model performance and comparisons
• 3. Interpreting Cancer Results
  • 3.1 Extract subnetwork attention
  • 3.2 Average edge attention
  • 3.3 Filter attention network (R)
  • 3.4 Cell-gene degree maps (TF-IDF and non-TF-IDF)
• 4. ROSMAP AD Downstream Analysis
  • 4.1 Extract AD subnetworks
  • 4.2 Average attention and p-values
  • 4.3 Pathway enrichment and NetFlowVis-AD
  • 4.4 NetFlowVis-AD controls
• Citation

---

1. Parsing the Raw Data Files

1.1 Cancer datasets preprocessing

python ./parse/init_parse.py
python ./parse/parse_circle.py
python ./parse/parse_network.py

Afterwards, parse datasets into the graph-AI-ready format. In post_parse.py, you may set the dataset parameter to datainfo-nci or datainfo-oneil. You can also modify the number of folds via n_fold = [1-5].

python post_parse.py

Further processing details are available in data_processing_details.pdf.

1.2 ROSMAP AD dataset preprocessing

Change directory to M3NetFlow_AD and use the notebook ROSMAP_union_raw_data_process_AD.ipynb to parse raw datasets. Then run:

python load_data.py

---

2. Run the Model

2.1 Run M3NetFlow

From the project root (or ./M3NetFlow_AD), configure dataset and fold_n in geo_tmain_m3netflow.py, then run:

python geo_tmain_m3netflow.py

2.2 Run baselines

python geo_tmain_gcn.py
python geo_tmain_gat.py
python geo_tmain_gformer.py
python geo_tmain_mixhop.py
python geo_tmain_pna.py
python geo_tmain_gin.py

Tip: For ROSMAP AD model performance, change directory into M3NetFlow_AD.

2.3 Model performance and comparisons

---

3. Interpreting Cancer Results

3.1 Extract subnetwork attention

python geo_analysis_m3netflow.py

3.2 Average edge attention

python analysis_path_edge.py

3.3 Filter attention network (R)

setwd('/Users/muhaha/Files/VS-Files/M3NetFlow')
cell_map_dict_df = read.csv('./datainfo-nci/filtered_data/cell_line_map_dict.csv')
fold_n = 0
num_cell = length(cell_map_dict_df[[1]])
for (x in 1:num_cell){
  each_cell_search_gene(fold_n, x, edge_threshold = 0.2, giant_comp_threshold = 20, whole_net = F, reweight=F)
}

3.4 Cell-gene degree maps (TF-IDF and non-TF-IDF)

Add TF-IDF weights to node importance to obtain cell-line-specific gene importance:

python analysis_cell_node.py

kegg_path.ipynb produces Figure 4c.

---

4. ROSMAP AD Downstream Analysis

Change directory to M3NetFlow_AD for AD analyses.

4.1 Extract AD subnetworks

python ROSMAP_analysis_path_edge.py

4.2 Average attention and p-values

python ROSMAP_analysis_avg_and_pvalue.py

4.3 Pathway enrichment and NetFlowVis-AD

Use the R app ./M3NetFlow_AD/shinyapp_graph_neo_drug-omics-pvalues.R to visualize AD results. The figure below highlights AD-relevant genes (by p-values) and suggests mechanisms and potential targets.

Visit NetFlowVis-AD: https://m3netflow.shinyapps.io/NetFlowVis-AD/

4.4 NetFlowVis-AD controls

• Patient Type: AD vs Non-AD.
• Signaling pathway selection: 48 pathways; selected ones are emphasized.
• Edge-weight threshold: [0, 0.25] (default 0.106).
• Important gene threshold by p-value: enlarge/mark significant nodes (tomato).
• Component size threshold: filter small components.

Visualization panel options:

• Gene node size (common nodes)
• Important gene node size
• Label size for common nodes
• Label size for important nodes

---

Citation

If you use M3NetFlow in your research, please cite:

@article{zhang2025m3netflow,
  title={M3NetFlow: A multi-scale multi-hop graph AI model for integrative multi-omic data analysis},
  author={Zhang, Heming and Goedegebuure, S Peter and Ding, Li and DeNardo, David and Fields, Ryan C and Province, Michael and Chen, Yixin and Payne, Philip and Li, Fuhai},
  journal={iScience},
  volume={28},
  number={3},
  year={2025},
  publisher={Elsevier}
}

License

This repository is released for research use.