Cellular perception, aging biomarkers, and the comparative biology of learning systems. Vaika Inc. (aging-research nonprofit, Gudkov lab) · Tel Aviv · mool32.github.io ORCID 0009-0004-5382-9346 · LinkedIn · tspiro@vaika.org
Current result — cancers of all lineages converge on the placental cytotrophoblast. A 9-dimensional coordinate system for cellular state, built from 44 transcriptional modules across 154 normal human cell types and no cancer data at all. Project 25 quality-controlled single-cell cancer datasets into it and the malignant compartment of every major lineage lands in one small region of normal physiology — anchored on the placental cytotrophoblast, the normal cell that invades, proliferates and evades immunity. Within matched tumour–normal pairs from the same patient, malignancy deepens that state ~6× (colorectal 43/46, paired P < 10⁻¹²). Preprint (Zenodo, not peer reviewed): 10.5281/zenodo.20542130
| perceptome | Cellular perception analysis in Python. 44 signaling modules, a 9-PC eigenspace from 154 Human Protein Atlas cell types, a capacity-floor predictor, and a validity scorecard with three null controls. v0.2.2 · Zenodo DOI · 74/74 tests |
| metric-autopsy | Red-team a computed single-cell metric before you believe it. Eight gates catch dropout, library-size, batch and factorial-QC confounds; the run stops at the first blocking failure. Built from three of my own results that survived weeks before a short QC check killed them. v0.1.1 · Zenodo DOI · 35 tests |
- Universal statistical signatures of evolution in AI architectures — arXiv:2604.10571. 935 ablations from 161 ML papers; the distribution of fitness effects of architectural changes matches biological DFEs.
- Waveform asymmetry as a biomarker of neural aging — under review, Frontiers in Aging Neuroscience. 823 subjects across two cohorts plus a 208-subject 5-year longitudinal.
- Spectral exponents of the twelve-lead ECG — preprint. 412,730 recordings, three continents; CLBBB vs CRBBB AUC ≈ 0.98 cross-population.
- Proteostasis modules causally enable memory consolidation — summary. 26 predictions registered before analysis; TLR9 knockout removes one arm and impairs memory.
- EEG comprehension detection — 18 metrics, 5 datasets, 126 subjects. Nothing replicates. Zero-calibration EEG comprehension detection is not supported by the evidence.
- Neural-network fitting as a structure sensor — a pre-registered negative: an autoencoder does not beat matched linear PCA at recovering donor age, and training erodes the signal.
- Oscillation as a substrate for perception — the founding hypothesis was refuted mid-project; the coupling that holds the cycle is what suppresses memory capacity.
Every inferential project is pre-registered, code and data are released, and the methodology that enforces it is public.
2026 Emergent Ventures grant recipient.