Research map / Graph learning / GNN architectures

OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data

NeurIPS 2026 · Graph learning › GNN architectures · approach: GNN generalization theory · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces OgBench, a benchmark and modular pipeline for graph-level classification on omics expression data in the n ≪ p regime, finding that standard GNNs often fail to beat MLPs and classical ML baselines.

Problem

Existing GNN graph-level benchmarks operate in the n ≫ p regime, whereas omics data have few samples and many features per sample; it is unclear whether GNNs help in this low-sample, high-node setting.

Approach

A standardized end-to-end omics-to-graph benchmark where each patient sample becomes a graph over genes/proteins with expression features, with configurable node selection (variance, correlation, distance correlation, random) and edge construction (co-expression vs. STRING PPI), used to compare GNNs against MLPs, SVM and elastic net.

Key results

• On Heritage, SAGE reached the best macro F1 of 0.576 ± 0.062 versus 0.495 ± 0.017 for MLP.
• On BRCA, MLP achieved the best macro F1 of 0.840 ± 0.048 and AUROC 0.962, with GNNs such as GPS at 0.824 and MLA-GNN at 0.837 close behind.
• On Parkinsons, MLP (0.747) and GIN (0.748) were roughly tied, while GCN (0.654) and GATv2 (0.646) lagged; no architecture consistently beat simple baselines across datasets.

Surprising findings

• Widely used GNNs often do not outperform simple MLPs and classical linear baselines on omics graphs.
• Random node selection, rather than correlation-based selection, gave the strongest graph-based performance on Heritage.

Limitations

• Only four reference datasets across three omics modalities, limiting generalization.
• Fixed shared adjacency matrix across samples may underrepresent individual biological heterogeneity.
• Results are based on 3 seeds with high variance and noisy validation selection.

Experimental setup

Datasets: HERITAGE Parkinson's transcriptomics (GENEPARK) AddNeuroMed BRCA (TCGA via MLOmics) STRING

Benchmarks: HERITAGE Parkinson's AddNeuroMed BRCA

Metrics: Accuracy Macro F1 Weighted F1 Macro AUROC parameter count GPU memory training time

Baselines: MLP SVM Elastic Net GCN GraphSAGE GATv2 GIN GPS SAGN ChebNet MLA-GNN

Builds on: TopoBench STRING WGCNA ChebNet MLA-GNN

Closest papers by meaning

• On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.87)
• GLBench: A Comprehensive Benchmark for Graph with Large Language Models (NeurIPS 2024 · similarity 0.86)
• PRAGA: Prototype-aware Graph Adaptive Aggregation for Spatial Multi-modal Omics Analysis (AAAI 2025 · similarity 0.86)
• Core Knowledge Learning Framework for Graph (AAAI 2025 · similarity 0.85)
• Rethinking Cancer Gene Identification Through Graph Anomaly Analysis (AAAI 2025 · similarity 0.85)
• Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2025 · similarity 0.85)