A MOGONET-Style Multi-Omics Biomarker Pipeline: Why a Near-Random Graph Net Still Earns Its Place

TL;DR (Quick Answer)

This is an honest engineering write-up of a MOGONET-style multi-omics consensus biomarker pipeline built as an internal R&D project at sysofti.

The headline — on a small synthetic cohort (n=30), the graph network alone scores near-random in leak-free 5-fold cross-validation (AUC 0.53 ± 0.16). Yet as one voter in a 5-evidence consensus, the top-10 ranking is 90% real markers (9 of 10 are known periodontitis genes).

The lesson — a single model that looks weak in honest evaluation can still be a useful voter. That contrast is the whole point of the consensus design, and we show it with data.

What it is — per-omics Graph Convolutional Networks (GCN) over a sample-similarity graph, attention-fused, contributing to a consensus score alongside differential-expression hubs, Random Forest, a DNN, and co-expression modules.