#15 · Epigenomics

Benchmark validity fails across datasets, platforms and graph-based scoring models

The dominant thread is whether predictive performance in biological machine learning survives evaluation beyond its originating dataset, algorithm, or platform. The network-based index MNDI reaches AUC 0.941 on WDBC but 0.538 on BreaKHis, documenting cross-dataset fragility. A factorial benchmark of 63 scRNA-seq cell-type annotation tools across seven families reports that cell-type separability in embedding space predicts accuracy and outweighs algorithm choice. A reproducibility-based framework builds and transfers DNA methylation algorithms on the Illumina MSA, using 40,786 stable CpGs and EPIC-MSA paired data. A self-audit of a graph neural network metabolic reaction scorer on Recon3D detects label leakage, memorized noise, and input-invariant models, and withdraws a prior claim with code released. The studies differ in modality, so the shared thread is evaluative rather than methodological. What is new relative to single-dataset ranking is the emphasis on reproducibility and transferability of features and scores across platforms.