NECROGENESIS: A Pre-Registered Pilot Study Using Cross-Tissue Gene Regulatory Network Inference to Identify Candidate Age-Associated Regulators, with an Extension of the CYTOS Tree-Tensor-Network-vs-Graph-Neural-Network Framework to Single-Cell Aging

Using publicly available single-cell transcriptomic data from the Tabula Muris Senis (accessed via the CZ CELLxGENE Census), we built a pentest-inspired, falsification-driven pipeline to search for transcription factors whose regulatory centrality changes with age across independent mouse tissues. Gene regulatory networks were inferred with a GENIE3-style ensemble-regression baseline, validated against curated aging-gene databases (GenAge, CellAge), and cross-checked for cell-type-composition confounds. Across three independent tissue/cell-type pairs (skeletal muscle satellite cells, epidermal keratinocyte stem cells, intestinal crypt stem cells; two sexes), one transcription factor — Zfp212, an understudied KRAB zinc-finger protein — showed a consistent rise in regulatory centrality with age in all three tissues, while four other candidates generated in the discovery tissue did not replicate consistently. We separately extended the CYTOS framework (a pre-registered comparison of Tree Tensor Networks, TTN, against parameter-matched Graph Neural Networks, GNN, originally validated on the DREAM4 benchmark) to a classification task built from this real biological data, after diagnosing and correcting a numerical-collapse bug in the original TTN implementation for deep community hierarchies. In this extension, TTN outperformed GNN in 10 of 15 seeds (mean accuracy 59.4% vs. 57.3%, both above the 50% trivial baseline), a result directionally consistent with, but not statistically confirming, the original CYTOS finding (Wilcoxon p=0.124). Multiple follow-up analyses intended to explain the Zfp212 signal — functional (GO) enrichment of its inferred targets, a candidate DNA-binding-motif scan, and a cross-species check against public human aging transcriptomic studies — returned null or inconclusive results and are reported in full. We report every negative result and methodological correction made during this pilot, and explicitly do not claim causal or therapeutic relevance without experimental validation.