arxiv: 2605.07838 · v1 · submitted 2026-05-08 · 🧬 q-bio.QM · cs.AI· cs.LG

Recognition: 2 theorem links

· Lean Theorem

PPI-Net connects molecular protein interactions to functional processes in disease

Dennis Veselkov, Guadalupe Gonzalez, Ivan Laponogov, Kirill Veselkov, Kyle Higgins

Pith reviewed 2026-05-11 02:09 UTC · model grok-4.3

classification 🧬 q-bio.QM cs.AIcs.LG

keywords graph neural networkprotein-protein interactionpathway hierarchycancer predictionhierarchical modelingRNA-seqmulti-omicspathway analysis

0 comments

The pith

A hierarchical graph neural network links molecular protein interactions to disease pathways, achieving over 90 percent accuracy in cancer type prediction.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents PPI-Net as a way to model how molecular alterations in proteins affect higher-level functional processes in disease. It embeds patient RNA-seq profiles into a protein-protein interaction network and propagates the information through a pathway hierarchy using graph attention layers. This structure supports both accurate classification of cancer types and identification of the involved biological programs. A reader would care because it bridges the gap between raw molecular data and interpretable disease mechanisms, potentially aiding both prediction and biological understanding. The approach shows measurable gains from using the full hierarchy and multi-level training.

Core claim

PPI-Net is a hierarchical graph neural network that integrates a protein-protein interaction network with a pathway hierarchy. Patient molecular profiles are embedded in the protein interaction graph and aggregated via graph attention across multiple levels of the pathway hierarchy to predict disease states and functional processes. On RNA-seq data from ten cancer types, the model attains balanced accuracy above 90% in multiple cohorts. Adding the pathway hierarchy improves accuracy by 6.7% over a protein-interaction-only version, and multi-level supervision provides an additional 12.3% gain. In a multi-omics setting with RNA-seq and methylation, it recovers known modules like TP53-AKT and p

What carries the argument

Hierarchical graph attention network that propagates signals from a protein interaction graph through a multi-layer pathway hierarchy to aggregate molecular data into functional representations.

If this is right

The pathway hierarchy integration yields a 6.7% accuracy improvement over protein-interaction-only models on breast cancer RNA-seq data.
Hierarchical multi-level supervision adds a 12.3% accuracy boost compared to single top-level prediction.
Multi-omics integration recovers canonical oncogenic modules and reveals coherent functional programs.
The model provides mechanistic insights into how molecular changes drive cancer biology.
Robust performance holds across ten cancer types from public genomics data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This hierarchical integration method could be tested on other complex diseases to see if it consistently improves over flat network models.
The attention mechanisms might highlight patient-specific pathway activations useful for personalized medicine approaches.
Applying the model to different omics layers or time-series data could reveal dynamic aspects of disease progression not captured here.
Generalization to non-cancer conditions would require checking if the same databases capture relevant biology adequately.

Load-bearing premise

The protein interaction network combined with the pathway hierarchy accurately captures the biological relationships needed to propagate molecular signals to disease processes without significant bias or omission.

What would settle it

Observing that the model achieves low accuracy or fails to recover known cancer pathways when evaluated on a fresh, independent set of patient RNA-seq samples from the same cancer types would indicate the claim does not hold.

read the original abstract

Understanding how molecular alterations propagate across biological systems to drive disease remains a central challenge. Although high-throughput profiling enables comprehensive characterization of tumor states, most models neglect structured biological relationships or lack interpretability across scales. Here we present PPI-Net, a hierarchical graph neural network that integrates protein-protein interaction (PPI) networks with pathway-level representations to model disease from molecular interactions to functional processes. Patient-specific molecular profiles are embedded within a shared interaction network from STRING and propagated through a multi-layer Reactome hierarchy using graph attention, enabling aggregation of gene-level signals into higher-order biological programs. Across RNA-seq data from ten cancer types from The Cancer Genome Atlas, PPI-Net achieves robust predictive performance, with balanced accuracy exceeding 90% in multiple cohorts. Comparative analysis on RNA-Seq data from breast cancer demonstrated that PPI-Net's integration of the Reactome hierarchy improved balanced accuracy by 6.7% relative to a PPI-only model, while hierarchical multi-level supervision improved balanced accuracy by 12.3% relative to using only a single top-level prediction head. Applying a multi-omics approach using RNA-seq and methylation data improves model interpretation, recovering canonical oncogenic modules, including TP53-AKT signaling and stress response pathways, while revealing convergence onto coherent programs such as ion signaling and cellular responses to stimuli. These results demonstrate that integrating interaction networks with pathway hierarchies enables accurate prediction while providing mechanistic insight into cancer biology.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PPI-Net adds Reactome hierarchy supervision to a PPI GNN and reports accuracy lifts on TCGA data, but the evaluation details are too thin to judge if the gains are real or scaffold artifacts.

read the letter

PPI-Net embeds patient RNA-seq and methylation profiles into a fixed STRING PPI graph and propagates them upward through a multi-layer Reactome hierarchy using graph attention, with supervision at multiple levels. On ten TCGA cancer types it claims balanced accuracy above 90% in several cohorts, plus a 6.7% gain over a PPI-only baseline on breast cancer and a further 12.3% from the multi-level heads. The multi-omics version also surfaces some expected modules such as TP53-AKT signaling. That combination of explicit hierarchy, multi-level loss, and reported ablations is not in the cited prior work, so the architecture itself is the clearest new piece. The ablation on supervision levels is useful and the recovery of known pathways gives a basic sanity check on interpretability. The soft spots sit in the validation. The abstract states the numbers but gives no information on cross-validation folds, cohort stratification, statistical testing of the gains, or controls for data leakage across TCGA projects. Without those, the high accuracies and the hierarchy lift are difficult to trust. The fixed STRING/Reactome scaffold is another open question: if the structure itself encodes well-studied hubs or general pathway over-representation, both the absolute performance and the ablation improvements could reflect database regularities rather than patient-specific signal propagation. No scrambling or randomization test of the network is described. This is a paper for computational biologists who build or compare graph models on multi-omics cancer data. A reader looking for a concrete hierarchical GNN example with pathway supervision will find something to try or benchmark against. It deserves a serious referee because the method is described clearly enough to reproduce once the missing protocol details are supplied, and the central claim can be tested directly. I would send it to review with requests for code, full CV description, and at least one check on whether the scaffold biases are driving the results.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces PPI-Net, a hierarchical graph attention network that embeds patient-specific RNA-seq profiles into a shared STRING PPI network and propagates signals through a multi-layer Reactome hierarchy to predict cancer type or subtype while recovering functional modules. It reports balanced accuracy exceeding 90% across ten TCGA cohorts, a 6.7% gain from adding the Reactome hierarchy versus a PPI-only model on breast cancer data, a further 12.3% gain from multi-level supervision, and improved interpretability when methylation data are added, recovering known oncogenic programs such as TP53-AKT signaling.

Significance. If the reported performance and ablation gains are reproducible under proper controls, the work would demonstrate a concrete route for injecting curated biological hierarchy into GNNs to improve both accuracy and mechanistic interpretability in cancer genomics. The multi-omics interpretation results and explicit comparison to a PPI-only baseline are positive features that could be built upon.

major comments (3)

[Abstract / Results] Abstract and Results: The central claim of balanced accuracy >90% across ten TCGA cohorts and the 6.7% improvement from Reactome integration are presented without any description of the cross-validation strategy, statistical testing of the reported deltas, handling of class imbalance, or explicit controls for data leakage between training and test splits. These omissions are load-bearing for the predictive claims.
[Methods] Methods / Model description: The manuscript applies a single fixed STRING PPI graph and Reactome hierarchy to all patients and cohorts. No analysis is provided to quantify potential database biases (e.g., degree distribution favoring well-studied hubs, pathway over-representation) or to test whether the observed performance and ablation gains could arise from these structural artifacts rather than patient-specific signal propagation.
[Results] Comparative analysis: The 6.7% balanced-accuracy lift attributed to the Reactome hierarchy (breast-cancer cohort) and the 12.3% lift from multi-level supervision are reported without the corresponding baseline model definitions, hyperparameter search protocols, or significance tests, making it impossible to determine whether the gains are robust or merely reflect differences in model capacity.

minor comments (2)

[Abstract] The abstract states that multi-omics integration 'improves model interpretation' but does not quantify this improvement or provide a direct comparison of recovered modules with and without methylation data.
[Methods] Notation for the graph-attention layers and the hierarchical aggregation operator should be defined more explicitly, ideally with a small diagram or pseudocode, to allow readers to reproduce the multi-scale message passing.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We agree that greater transparency in experimental protocols, statistical controls, and analysis of potential database artifacts is necessary to support the reported claims. We have revised the manuscript to address each point and provide the following point-by-point responses.

read point-by-point responses

Referee: [Abstract / Results] Abstract and Results: The central claim of balanced accuracy >90% across ten TCGA cohorts and the 6.7% improvement from Reactome integration are presented without any description of the cross-validation strategy, statistical testing of the reported deltas, handling of class imbalance, or explicit controls for data leakage between training and test splits. These omissions are load-bearing for the predictive claims.

Authors: We agree that these details are essential. In the revised manuscript we have expanded the Methods and Results sections to describe the patient-stratified 5-fold cross-validation procedure, which ensures no patient overlap between folds and uses balanced sampling to address class imbalance. We now report mean balanced accuracy plus standard deviation across folds and include Wilcoxon signed-rank tests confirming statistical significance of the 6.7% and 12.3% gains (p < 0.01). These additions directly support the central claims. revision: yes
Referee: [Methods] Methods / Model description: The manuscript applies a single fixed STRING PPI graph and Reactome hierarchy to all patients and cohorts. No analysis is provided to quantify potential database biases (e.g., degree distribution favoring well-studied hubs, pathway over-representation) or to test whether the observed performance and ablation gains could arise from these structural artifacts rather than patient-specific signal propagation.

Authors: We acknowledge the concern about fixed graph biases. Patient-specific RNA-seq values serve as node features, so attention weights adapt to individual profiles. To quantify this, the revised supplementary material now includes an ablation comparing the real STRING/Reactome structure against degree-preserved random graphs; the biological graphs yield significantly higher accuracy (p < 0.05), indicating that performance is not driven solely by generic structural properties. We have also added a Discussion paragraph on potential hub and annotation biases. revision: yes
Referee: [Results] Comparative analysis: The 6.7% balanced-accuracy lift attributed to the Reactome hierarchy (breast-cancer cohort) and the 12.3% lift from multi-level supervision are reported without the corresponding baseline model definitions, hyperparameter search protocols, or significance tests, making it impossible to determine whether the gains are robust or merely reflect differences in model capacity.

Authors: We agree the baselines required clearer specification. The PPI-only baseline is a graph-attention network on the STRING graph alone, using identical embedding dimensions and a single prediction head. The multi-level supervision variant adds auxiliary losses at each Reactome hierarchy level. In the revision we have added a dedicated subsection describing the shared grid-search protocol (learning rate, attention heads, hidden size) applied identically to all models, plus a table of hyperparameter values and paired statistical tests (p < 0.05) for both reported lifts. Model-capacity differences are now explicitly controlled. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical ML performance on external TCGA cohorts

full rationale

The paper describes a standard hierarchical GNN trained on patient RNA-seq profiles embedded in a fixed STRING/Reactome scaffold. Reported balanced accuracies (>90% on ten TCGA cohorts) and ablation lifts (6.7% from hierarchy, 12.3% from multi-level supervision) are direct empirical outcomes of supervised training and evaluation on held-out data, not reductions by construction. No self-definitional equations, no parameters fitted to a subset then renamed as predictions, and no load-bearing self-citations appear in the provided text. The derivation chain consists of conventional graph attention propagation plus cross-entropy loss, which remains independent of the final accuracy numbers.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claims rest on pre-existing curated databases and standard graph neural network components rather than new postulates; limited information is available from the abstract alone.

free parameters (1)

GNN hyperparameters
Number of layers, attention heads, embedding dimensions, and training hyperparameters are fitted to the TCGA data but not enumerated.

axioms (2)

domain assumption STRING database interactions accurately reflect biologically relevant protein relationships for cancer propagation.
The shared interaction network is taken directly from STRING without additional validation in the abstract.
domain assumption Reactome hierarchy correctly organizes functional processes at multiple scales.
Used as the propagation and supervision structure.

pith-pipeline@v0.9.0 · 5564 in / 1360 out tokens · 39107 ms · 2026-05-11T02:09:11.948299+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
hierarchical graph neural network that integrates protein-protein interaction (PPI) networks with pathway-level representations... propagated through a multi-layer Reactome hierarchy using graph attention
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
PPI-only model... Reactome hierarchy improved balanced accuracy by 6.7%

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

[1]

conveyor-belt

In contrast to prior approaches that operate directly on gene-level features, PPI-Net ﬁrst models patient-speciﬁc molecular states within a protein interaction network, enabling the capture of local molecular context before propagating signals through a structured pathway hierarchy using a sequential series of bipartite graph attention layers, allowing th...

work page 2026
[2]

conveyor belt

At the highest levels of the hierarchy, pathways related to cellular responses to external stimuli emerged as dominant contributors, integrating signals from stress response, DNA damage signaling and protein folding pathways. Key contributors included heat shock protein 90 components HSP90AA1 and HSP90AB1, which are associated with decreased breast cancer...

work page 2011
[3]

22 Vella, D

Nucleic acids research 52, D672–D678 (2024). 22 Vella, D. et al. MTGO: PPI network analysis via topological and functional module identiﬁcation. Scientiﬁc reports 8, 5499 (2018). 23 Zhou, G., Wang, J., Zhang, X. & Yu, G. in 2019 IEEE international conference on bioinformatics and biomedicine (BIBM). 1836–1841 (IEEE). 24 Hu, L., Wang, X., Huang, Y .-A., Hu...

work page 2024
[4]

Biochemical Pharmacology 230, 116573 (2024)

in cancer: emerging roles and therapeutic potentials. Biochemical Pharmacology 230, 116573 (2024). 57 Li, M., Tian, P ., Zhao, Q., Ma, X. & Zhang, Y . Potassium channels: Novel targets for tumor diagnosis and chemoresistance. Frontiers in oncology 12, 1074469 (2023). 58 Han, D., Wang, L., Jiang, S. & Yang, Q. The ubiquitin–proteasome system in breast canc...

work page 2024