The Undiagnosed T2T Digital Twin for AI Pre-Diagnosis

1.Executive summary and the honest stance

This document is a contract with the future. It states, in advance and in falsifiable form, what we expect the Undiagnosed T2T Digital Twin to show, so that when the analyses complete we can score the results against the predictions rather than rationalize them afterward. The twin is a patient-owned, multi-omic, long-read reconstruction of one human genome, built to be read as a set of signals. The thesis of the whole program is a single corrected lesson, earned the hard way on protein data: a one-dimensional spectrum of a biological sequence is, by default, a null result, and the real signal lives in a representation the one-dimensional view cannot reach. For proteins that was three-dimensional structure. For the methylome it is the single molecule, and, looking forward in time, the dynamics of the regulatory system itself.

2.The subject and the datasets

All datasets are assumed complete for the purpose of this pre-registration. The twin is referenced only by its content-derived fingerprint and custodian biowallet; no personal name appears in this clinical-scope document.

3.The signal-domain lineage and its honest negative result

The signal-domain reading of biology has a real lineage. The Cosic Resonant Recognition Model (RRM) encodes each residue by its electron-ion interaction potential (EIIP), takes the discrete Fourier transform of the sequence, and reads a characteristic frequency from the cross-spectral consensus of a functional family. A companion periodic-charge model (PSM) looks for periodic charge architecture. These are principled instruments.

They were also tested honestly, at scale, and the result is the engine of this whitepaper. As a variant-effect predictor the 1-D Cosic-RRM spectrum was null: across a 223,987-variant benchmark the cohort ΔAUC was about −0.0065, and the spectral-only score fell below 0.5 (worse than chance) on several genes. The single metric that appeared to carry weak signal reduced, by Parseval's theorem, to a one-residue ΔEIIP difference (Spearman ρ ≈ 0.9998 to that trivial feature): there was no spectrum in it at all. PSM, by contrast, behaved like a real instrument, winning exactly where periodic charge exists (the SCN5A S4 arginine ladders) and correctly returning a clean null on the curvature-gated Piezo channels. And the genuine protein signal, when it appeared, lived in three-dimensional structure (elastic-network normal modes), not in the one-dimensional transform.

4.The corrected family for the methylome

Methylation is, unlike a protein sequence, a genuine one-dimensional spatial signal along the genome with real biological scales. That makes the transform-domain reading better motivated here than it ever was for protein sequence. It does not exempt it from the discipline. We split the family into a falsifier, a representational pivot, and a two-faced Laplace term.

4.1 Fourier / Cosic-RRM: demoted on arrival

The bulk spatial Fourier (or Cosic-RRM) spectrum of the per-CpG track is the exact same one-dimensional object that already came back null on proteins. We register it as the expected-null falsifier. There is even a mechanistic reason to expect the null here: CpG sites are spaced irregularly at roughly 100 bp, which aliases the 10.5 bp helical period out of existence past the Nyquist limit, and the ~180 bp nucleosome footprint is already captured by ordinary windowed-mean and nucleosome-depletion features. Reporting this null cleanly is the first honest contribution.

4.2 The structure-to-single-molecule pivot

For proteins, the escape from the null 1-D spectrum was three-dimensional structure. The methylome's analog is the single molecule. Each ultra-long read is a methylation pattern over hundreds of CpGs on one chromosome, and that pattern carries information bulk averaging and short-read sequencing mathematically cannot reconstruct: co-methylation, methylation haplotypes, and allele-split bimodality. This is the representation we bet on.

4.3 Laplace, two faces kept strictly apart

Face A (static). The co-methylation decay-length λ is the exponential constant of how methylation concordance falls off with genomic distance within a read. We bound this honestly: it is the s = 0 (DC) limit of the co-methylation kernel, a single operating point, not a transfer-function sweep, and we will not dress it in Hz it does not have.

Face B (dynamic), the frontier. The genuinely new contribution is not a spectrum at all. Disease onset and relapse have a rigorous dynamical-systems signature, critical slowing down: as a regulatory control loop approaches a tipping point, its dominant eigenvalue (pole) drifts toward instability and the system shows rising lag-1 autocorrelation, rising variance, and slower recovery from perturbation. That is the literal "pre" in pre-diagnosis. It is measured not in one snapshot but across time, which is why the single-timepoint twin is timepoint zero of a longitudinal early-warning system (see §7).

5.Methods: feature construction and discipline

From the aligned, mod-tagged reads we extract, per locus: the bulk windowed methylation (mean and variance), CpG density, the standard co-methylation scalar (PDR / epipolymorphism), and the candidate single-molecule features (λ, read-level Shannon entropy, mixture/cluster count, allele-split state). The transform-domain features (the Fourier falsifier, the wavelet multi-scale descriptor) are computed over true genomic coordinates, not CpG index, to avoid spectral leakage.

Every positive claim must clear four gates, each ported directly from the proteome null result:

• Four baselines. A candidate must beat windowed mean, windowed variance (beat the second moment, not just the first), CpG density (not re-encode the sampling grid), and the standard PDR/epipolymorphism scalar.
• Bulk-collapsed control. Every single-molecule feature is recomputed after averaging reads into a pseudo-bulk track; the single-molecule version must beat its own collapsed self, the only proof the long reads were necessary.
• The ρ > 0.97 algebraic-identity kill-switch. The literal port of the Parseval catch: before any positive claim, if a feature correlates above ρ = 0.97 with a trivial feature (variance, CpG density), it is declared the trivial feature in costume and reported null.
• GroupKFold-by-region add-gate. A feature must add held-out separation over the baselines under cross-validation held out by genomic region, not merely correlate with the label.

6.Diagnostic inferences: a five-tier hypothesis engine

For an undiagnosed case the twin produces ranked hypotheses across five tiers. Each output is a lead for a clinician, never a verdict.

7.The dynamic-Laplace early-warning frontier

The most distinctive prediction of this program cannot fire on a single timepoint, and we say so plainly. Critical slowing down is a property of a trajectory, not a snapshot. With three or more consented, time-separated samples, a region or regulatory module approaching a tipping point should show the early-warning triad (rising autocorrelation, rising variance, slowing recovery) before any overt methylation or clinical change. The single-timepoint twin establishes the baseline against which that drift is later measured.

The system also accepts an input. In control terms a psychological or physiological input u(t) moves the operating point of the regulatory transfer function H_regulatory(s). Psychoneuroimmunology supplies a real, falsifiable example: sustained stress shifts the Conserved Transcriptional Response to Adversity (the CTRA inflammatory program), and connection and meaning shift it back. This is a measurable input to the dynamic side, and it is exactly the longitudinal, patient-owned sampling that the governance layer (§9) is built to enable.

8.The pre-registered hypotheses

Seven hypotheses, registered now. Each lists its prediction, its pre-defined null, the baseline it must beat, and how it is falsified.

9.Expected processing time and computational cost

Per-stage estimates for the methylome and single-molecule track, using the protocol's measured throughput. The HiFi variant calling and de-novo assembly are already complete and are treated as sunk cost. The v2 pipeline aligns two ONT libraries concurrently, roughly halving the alignment wall-clock and cost versus the sequential v1.

The per-twin marginal compute cost for the full epigenetic and single-molecule analysis is on the order of one hundred and fifty dollars, dominated by the irreducible ultra-long alignment. The foundation-model training is a cohort-scale, one-time investment amortized across all twins, not a per-patient cost.

10.Governance: BioNFTs and patient ownership

Ownership is a first-class component of the twin, not metadata. It is also load-bearing for the science: the dynamic-Laplace early-warning frontier can only fire on consented, longitudinal, repeated sampling, and that is exactly what the governance rails enable.

This is the operational form of "the patient as CEO of their own health": the owner decides each use, the chain of custody is auditable, and the longitudinal participation that the predictive frontier requires is sustained by ownership, dividends, and hope rather than coercion.

11.Honest limitations and comparison plan

This is a methods-only pre-registration on a single genome. The strongest single-molecule and episignature results are testable at n = 1; the primary λ contest is an honest coin-flip whose value is its falsifiability; the bulk Fourier null is near-certain and is a contribution precisely because it is reported cleanly. The dynamic-Laplace early-warning frontier cannot be evaluated here at all and waits on at least three longitudinal timepoints. The cohort foundation model is gated behind a population of consented twins and a beat-the-simple-baseline requirement.