Publications
All entries are authored by Theodor Spiro (ORCID 0009-0004-5382-9346).
They are grouped by what stage each one is actually at, so the list cannot read as stronger than it is: a self-archived Zenodo preprint is not a peer-reviewed paper, and is not presented as one. Where a DOI resolves to Zenodo, that is the archived manuscript and its code — not a journal record.
The canonical BibTeX file is served at /papers.bib
and is the single source of truth for citation export; everything on this page is
generated from it.
Under review at a journal
-
Waveform asymmetry as a biomarker of neural aging: spatial degradation of oscillatory cycle shape across two independent cohortsPeak-trough asymmetry of EEG oscillations across five frequency bands in 215 (LEMON) and 608 (Dortmund Vital Study) adults, with 208-subject 5-year longitudinal follow-up. Beta-band asymmetry decreases with age (r = −0.326 LEMON, r = −0.314 Dortmund); effect size exceeds classical alpha slowing. Theta spatial entropy predicts memory independent of age.
Preprints
Publicly posted, not peer reviewed. arXiv entries are on arXiv; the rest are
self-archived on Zenodo with a permanent DOI.
-
Cancers of all lineages converge on the placental cytotrophoblast, a normal invasive cell state that malignancy deepensA coordinate map built from 44 signaling modules over 154 normal cell types — no cancer data — onto which the malignant cells of 25 cancers converge. The anchor is the placental cytotrophoblast, a normal cell state that naturally invades, proliferates, and evades immunity; within patients, malignancy deepens the state ~6× over matched normal tissue (colorectal 43/46, paired P < 10⁻¹²). Not a uniquely cancerous program but a normal placental/developmental one, pushed deeper.
-
Universal statistical signatures of evolution in artificial intelligence architectures935 ablation experiments compiled from 161 publications. The DFE of architectural modifications matches biological DFEs (heavy-tailed Student's t; 68% deleterious / 19% neutral / 13% beneficial for major ablations); architectural diversification follows logistic dynamics with punctuated equilibria; 14 architectural traits were independently invented 3–5 times in parallel with biological convergences.
-
Spectral exponents of the twelve-lead ECG reveal the anatomy of cardiac conduction disorders and a bifurcation between aging and diseaseSpectral exponent β extracted from 412,730 twelve-lead recordings across three continents (PTB-XL, Chapman-Shaoxing, CODE-15%). β is diagnostic of cardiac conduction anatomy: CLBBB vs CRBBB AUC = 0.982 (Germany), 0.982 (China), 0.979 (Brazil). Aging flattens the spectrum; disease steepens it. Honest null on independent mortality prediction (HR = 1.02, p = 0.83).
-
Transcriptomic noise accumulates within tissue identity across human aging: a systemic signature distinct from cell-composition driftThree-level variance decomposition on bulk transcriptomes from 263 GTEx v8 donors plus single-cell data from Tabula Muris Senis, Calico rat caloric-restriction atlas, and a rhesus macaque cross-species atlas. Tissue identity is preserved across forty years of aging; the signature is systemic noise, not selective accumulation. Caloric restriction acts as a noise filter, not a structure restorer. Cross-species lifespan scaling (α = −1.02, R² = 0.90).
-
Temporal architecture of signaling oscillations predicts cancer gene function across pathways157 genes across 14 oscillatory signaling pathways classified into rise-phase and recovery-phase by temporal role only. Rise genes enriched for oncogenes, recovery for tumor suppressors (OR = 27.5, p = 3.6 × 10⁻⁹); 12/12 pathways match the prediction. Predicted inversions in growth-inhibitory pathways (p53, TGF-β) confirmed. On 22 divergent cases, the temporal framework is correct in 19/22 (86%) vs 3/22 (14%) for naive biochemical classification.
Manuscripts in progress
Complete or near-complete write-ups that are not yet posted.
-
Cellular proteostasis modules causally enable memory consolidation through region-specific architecturesMemory consolidation, read through 44 cellular sensing modules, resolves into a three-layer architecture — perception (senses the trigger), sentinel (senses proteostatic load), and construction (builds the machinery) — that separates cleanly in time and mechanism. TLR9 knockout selectively removes the ER-chaperone arm (−60–75%) and impairs memory: direct causal evidence. 26 pre-registered predictions across 7 datasets; 28/32 published perturbations predicted correctly, 0 contradictions.
-
Functional differentiation generates universal fitness-effect distributions in neural networksAblation studies of 144 identity-tracked attention heads, 24 transformer layers, and 30 distributed-noise perturbations at each of 8 checkpoints (step 512 → 143,000) of Pythia 410M-deduped — 1,584 ablations. The DFE shape evolves from delta-peak-with-outliers to heavy-tailed Student's t (df ≈ 2.3, β ≈ 0.6 within biological range). 39% of eventually-critical heads emerge from noise; only 4% are born critical. L8H9 undergoes phase-transition-like specialization between steps 4k–8k.
-
21,000 attempts to think differently: a large-scale Russian adaptation of the Divergent Association Task reveals practice resistance and lexical predictors of semantic creativityRussian-language adaptation of the Divergent Association Task (Olson et al., 2021). 21,159 submissions; Cronbach's α = 0.899; split-half reliability 0.696; no measurable practice effect. Strongest predictor: semantic category diversity (r = 0.47). Theoretical ceiling 110.5 (best human attempt: 104.8). Live instrument runs at mool32.github.io/dat-ru.
-
Clonal crystallization as a shared signature of bone-marrow aging and neural-network trainingA two-metric framework (Gini + effective N) applied jointly to mouse cell-type proportions and to Pythia-410M head-importance distributions. Pythia moves through (ΔGini, Δeff_N) by (+0.145, −45.5) over 143k steps; bone marrow on FACS and Droplet platforms moves in the same quadrant. Caloric restriction in rat bone marrow rescues 64% of the Gini drift. A specific substrate-independent signature, narrower than full DFE universality.
-
The Oracle's Fingerprint: correlated AI forecasting errors and the limits of bias transmissionThree studies on whether the wisdom of crowds survives consultation of a small set of LLMs. GPT-4o, Claude, and Gemini show pairwise error correlation r = 0.78 on 568 resolved binary Metaculus forecasts (Study 1). Community forecasts move in the LLM-predicted direction (r = 0.20, p = 0.007), but the shift is fully explained by rational updating toward ground truth (Study 2). Pre-ChatGPT human biases already strongly resembled the LLM pattern (r = 0.87); post-ChatGPT the resemblance weakened (r = −0.28) — LLMs inherited human biases rather than introducing novel ones (Study 3).
-
Negative feedback loop architecture as a modular predictor of cancer vulnerability across signaling pathwaysAlgorithmic extraction of 128 unique negative feedback loops from 159 KEGG signaling networks, organized into 14 pathway-level modules. NFL genes are 59-fold enriched for Cancer Gene Census membership over non-NFL genes in the same pathways (Fisher exact p = 9 × 10⁻⁴⁴). Irreversible Authority metric predicts cancer gene fraction with Spearman ρ = 0.83 across modules. Eigenspace projection identifies 20 novel cancer candidates, 13 (65%) IntOGen-validated.
-
Epistasis mapping in transformer attention heads: cross-model multi-checkpoint ablation interactionsPairwise head-ablation interaction maps (ε_AB = Δ_AB − (Δ_A + Δ_B)) on Pythia 410M and OLMo-2 1B across the full training trajectory. 78% of significant top-30 pairs in Pythia 410M show ε_loss > 0 (synthetic-lethal-like in Costanzo's terminology). The epistasis transition co-locates with the Paper 2 DFE crystallization (training steps 512–1,000), rejecting lottery-ticket emergence.
-
Neural-network fitting as a structure sensor: a pre-registered negative on immune agingHonest negative result. A pre-registered test (A∧B∧C∧D, leave-donors-out) of whether an autoencoder's unsupervised fit can recover donor age from immune scRNA-seq (GSE233321, 167 donors). The latent never beats a matched linear PCA at recovering age (0/2 across two seeds × latent dims) and training erodes the signal; a synthetic positive control localises the regime where the method would win — high linear rank, low intrinsic dimension — which single-cell data does not occupy. The unsupervised latent independently re-derives the perceptome eigenspace (cross-validated CCA, all p = 0.001) but not its aging axis (orthogonal to the reference inflammaging direction).
Software and instruments
-
perceptome: a Python toolkit for cellular perception analysisA Python toolkit treating 44 cellular signaling pathways (NF-κB, mTOR, UPR, p53, nuclear receptors, …) as a cell's perceptual repertoire. Three capabilities other tools don't directly provide: (1) a 9-PC eigenspace derived from 154 normal human cell types (Human Protein Atlas) where any cell — yours, a tumor, a drug-treated — can be located on the same map; (2) a capacity-floor predictor — given a cell type and a pathway, will it ramp under stimulus or is it already saturated; (3) a validity scorecard with three null controls (random-gene panel, housekeeping, cell cycle) that catches common perturbation artifacts before interpretation. It is the toolkit behind the cancer-convergence finding — 25 cancers converge on the placental cytotrophoblast (Zenodo 10.5281/zenodo.20542130) — and implements only what that multi-year pre-registered program validated, deliberately excluding operations that were tested and falsified.
-
metric-autopsy: a metric-agnostic gate system for separating biological signal from QC and technical artifacts in single-cell metricsRed-team a computed single-cell metric before you believe it. Eight gates — dropout independence, factorial QC parity, n_genes matching, stratified controls, cross-dataset replication — probe a black-box metric and stop at the first blocking failure. Built from three of my own failures that each survived weeks before a 45-second QC check killed them. Ships as a pip package, a Claude Code skill, and an MCP server; 35 tests.
Reports and pre-registered negatives
Results that did not go the way the pre-registration predicted, written up in full rather than dropped.
-
Connectivity Contrast Index across five EEG datasets: an honest non-replication and a methods note on universal comprehension biomarkersHonest negative result. 18 metrics tested across 5 independent EEG datasets totaling 126 subjects, after a promising single-dataset finding (CCI window-level d = +0.564). No metric replicates across datasets; the strongest follow-up signal (wPLI alpha CV on DERCo, p = 0.015) reverses direction on ZuCo and shows null effects on STEW and Speech-in-Noise. Universal zero-calibration EEG comprehension detection is not supported.