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

Abstract:

Graph Neural Networks (GNNs) have emerged as the leading approach for inductive graph-level learning. However, existing benchmarks predominantly address scenarios where the number of graphs ($n$) significantly surpasses the number of nodes per graph ($p$), a condition known as the $n \gg p$ regime. This focus neglects biological fields like omics, which function under the contrasting $n \ll p$ paradigm. In this setting, datasets feature large-scale networks of genes, transcripts, or proteins but are limited to a small number of patient samples. This disparity prompts a critical inquiry: \textit{what is the performance of GNNs in such low-sample, high-node omics environments?}

To address this, we present \texttt{OgBench} (Omics-Graph Bench), the inaugural benchmarking platform designed for graph-level prediction within the $n \ll p$ framework typical of omics research. The platform offers a standardized, end-to-end modular infrastructure that transforms raw omics data into families of featured graphs with diverse structural characteristics. We evaluate classical GNNs, alongside models specifically engineered for large graphs and omics tasks, comparing them against Multi-Layer Perceptrons (MLPs) and other machine learning baselines to set performance references. Our analysis reveals that commonly utilized GNNs frequently fail to surpass the capabilities of simple MLPs and traditional baselines. These outcomes contest the widely held belief that graph structure inherently provides added value in this context, urging a thorough reevaluation of current learning methodologies. By highlighting these constraints, OgBench establishes an open-source ecosystem that enables the community to create and test new architectures specifically adapted for biological graphs. The source code is publicly accessible at https://github.com/geometric-intelligence/ogbench.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC