In 1993, a consortium led by James Gusella and Nancy Wexler identified the gene that causes Huntington's disease: a CAG trinucleotide repeat expansion in the HTT gene on chromosome 4. The mutation is dominant. It is fully penetrant above 40 repeats — meaning every person who carries it will develop the disease if they live long enough. The number of CAG repeats predicts age of onset with known statistical accuracy: higher repeat counts correlate with earlier onset. A person with 45 repeats will likely develop symptoms in their mid-40s. A person with 60 repeats may develop juvenile Huntington's in their teens.
No neurodegenerative disease is more genetically understood. The mutation is known. The pathological mechanism — mutant huntingtin protein aggregation, excitotoxicity, mitochondrial dysfunction — has been studied for three decades. The biomarkers are established: neurofilament light chain (NfL) in plasma as a progression marker, mutant huntingtin protein (mHTT) in cerebrospinal fluid as both a diagnostic and pharmacodynamic biomarker. The outcome measure is standardized: the Unified Huntington's Disease Rating Scale (UHDRS), a 124-point instrument that captures total motor composite, functional capacity, and behavioral symptoms.
Zero FDA-approved disease-modifying therapies exist for Huntington's disease as of 2026.
This is not a failure of biological knowledge. It is a failure of learning architecture.
What Thirty-Three Years of Trials Have Produced
The TRACK-HD study (2008–2012) established the natural history of HD progression in pre-manifest and early manifest patients across four international sites. The PREDICT-HD study (2001–2013) tracked 1,078 pre-manifest individuals across 32 sites to characterize the prodromal phase. ENROLL-HD, launched in 2012 and still ongoing, is the largest observational HD platform in history — more than 22,000 participants across hundreds of sites in North America, Europe, Latin America, and Australia.
Every one of these studies generates longitudinal outcome data: UHDRS trajectories, NfL plasma levels, mHTT CSF concentrations, cognitive assessments, imaging data, functional decline timelines. Every one of them feeds data into retrospective repositories and research databases. The European Huntington's Disease Network (EHDN) registry and ENROLL-HD together represent the most comprehensive natural history dataset for any hereditary neurodegenerative disease in human history.
And none of it routes in real time between active clinical trials.
When a gene silencing trial runs at 40 sites simultaneously, the outcome signals accumulating at Site 12 — a slightly different NfL trajectory response in patients with CAG repeat counts above 55 — do not route to Site 7, which is enrolling a similar patient population and making dosing decisions based on protocol assumptions established at trial launch. The signal is there. The architecture to move it across the network while the trial is running does not exist.
Tominersen: When the Signal Was in the Data Before the Trial Was Stopped
In March 2021, Roche and Genentech halted the GENERATION HD1 Phase III trial of tominersen — an antisense oligonucleotide designed to reduce total huntingtin protein expression. The trial enrolled more than 800 patients. Development investment exceeded $2 billion. The halt came from a data safety monitoring board review that found the drug was worsening outcomes in patients with more advanced disease, while showing potential benefit in lower-disease-burden patients.
The dose-stage interaction signal — that tominersen's effect varied with baseline disease burden in ways that could cause harm — was not manufactured at the halt. It was accumulated progressively across enrollment waves at dozens of sites. The question that matters architecturally is not whether the DSMB caught it. The question is: could a routing layer have surfaced that signal from early-enrollment sites to trial operations weeks or months before the point at which harm had accumulated enough to trigger a protocol halt?
Each early-enrollment participant's UHDRS trajectory and NfL response was an outcome observation. The clinical teams at those sites accumulated that signal. None of it routed to the sites enrolling the next cohort in real time. The architecture produced a trial that learned as a batch, not as a network.
A network that routes pre-distilled outcome signals from early-enrollment participants to later-enrollment sites — without sharing any patient data — could have surfaced the dose-disease-burden interaction earlier. The intervention would have been the same: adjust enrollment criteria or dosing protocol. The timing would have been different.
Why Federated Learning Cannot Solve This
Federated learning — training a shared model across institutions by sharing gradients rather than raw data — has been proposed as the architecture for cross-trial HD intelligence. It fails for the same structural reasons it fails in every rare neurodegenerative disease.
The CAG repeat stratification problem is severe. Huntington's clinical behavior differs meaningfully across repeat length ranges: patients with 36–39 repeats (reduced penetrance) follow a different disease course from those with 40–50 (typical adult onset) and radically different from those with 55–70+ (juvenile HD). A federated learning model that aggregates gradient updates across these subgroups learns a blended average that predicts none of them accurately.
The minimum cohort problem applies at every non-major site. A dedicated HD clinic in Barquisimeto, Venezuela — which exists because Venezuela has the world's highest regional HD prevalence, concentrated around Lake Maracaibo through a founder effect traceable to a single ancestor in the early 19th century — may have 15 to 30 active patients at any time, stratified across repeat lengths and disease stages. Federated learning's mathematical foundation (McMahan et al., "Communication-Efficient Learning of Deep Networks from Decentralized Data," AISTATS 2017) requires sufficient local data to produce gradient updates that are statistically meaningful and do not destabilize the global model. A site with 15 patients, stratified further by CAG repeat range, does not meet that threshold.
The Venezuelan clinics are excluded from the federated learning network by architecture. The patients with the most concentrated HD experience on earth cannot contribute to a federated model because the math does not accommodate them.
The Architecture That Routes the Signal
The discovery made by Christopher Thomas Trevethan on June 16, 2025 — now covered by 39 provisional patents — is not a better federated learning algorithm and not a new trial design framework. It is an architectural discovery about how clinical outcome intelligence can scale across a distributed network without centralizing any patient data.
The Quadratic Intelligence Swarm protocol applied to Huntington's disease works as follows. When a gene silencing trial site completes a treatment interval with a patient, the clinical system distills the observation into an outcome packet — approximately 512 bytes. The packet encodes:
- A semantic fingerprint: CAG repeat range bracket, UHDRS total motor composite tier, disease stage class (prodromal, early manifest, mid-stage, advanced), intervention class (ASO, gene therapy, small molecule), baseline NfL plasma tier, baseline mHTT CSF tier if available
- An outcome delta: UHDRS TMC slope relative to CAG-matched natural history baseline, NfL trajectory direction, mHTT reduction percentage for gene silencing interventions, adverse event classification flag
- A confidence score and time-to-live
The packet contains no patient identity, no raw clinical data, no genetic sequence, no imaging data. Privacy is guaranteed not by policy or encryption but by data model: the only thing that leaves the clinical system is the distilled signal.
That packet is posted to a deterministic semantic address — a routing key computed from the fingerprint that places this outcome packet next to outcomes from patients with similar CAG repeat ranges at similar disease stages receiving similar interventions. Any site querying that address with a matching fingerprint pulls back the relevant packets from every contributing site.
The routing mechanism — distributed hash table achieving O(log N) lookup, vector similarity index, semantic search over a shared database — is an implementation choice, not the discovery. Christopher Thomas Trevethan's discovery is that any mechanism that maps fingerprints to deterministic addresses at O(log N) or better cost enables the complete loop: distill at edge, route by semantic similarity, synthesize locally. That loop is what scales intelligence quadratically without scaling compute.
import hashlib
import json
from datetime import datetime
class HDOutcomePacket:
"""
~512-byte distilled outcome from one HD patient treatment interval.
No patient identity. No raw clinical data. No PHI.
"""
def __init__(
self,
cag_repeat_bracket: str, # e.g. "40-45", "46-55", "56+"
uhdrs_tmc_tier: int, # 0-3: prodromal / early / mid / advanced
intervention_class: str, # "ASO" | "gene_therapy" | "small_molecule"
uhdrs_slope_vs_baseline: float, # negative = slower decline than CAG-matched baseline
nfl_trajectory: str, # "stable" | "rising_slow" | "rising_fast"
mhtt_reduction_pct: float, # 0.0 if not a gene-silencing trial
adverse_event_flag: bool,
confidence: float,
ttl_days: int = 90
):
self.fingerprint = self._compute_fingerprint(
cag_repeat_bracket, uhdrs_tmc_tier, intervention_class
)
self.payload = {
"uhdrs_slope_vs_baseline": round(uhdrs_slope_vs_baseline, 3),
"nfl_trajectory": nfl_trajectory,
"mhtt_reduction_pct": round(mhtt_reduction_pct, 1),
"adverse_event_flag": adverse_event_flag,
"confidence": round(confidence, 2),
"expires": ttl_days
}
self.routing_address = self.fingerprint[:16]
self.timestamp = datetime.utcnow().isoformat()
def _compute_fingerprint(self, cag, tier, intervention):
raw = f"HD:{cag}:tier{tier}:{intervention}"
return hashlib.sha256(raw.encode()).hexdigest()
def to_bytes(self) -> bytes:
data = {
"fp": self.routing_address,
"ts": self.timestamp,
**self.payload
}
encoded = json.dumps(data, separators=(',', ':')).encode()
assert len(encoded) < 512, f"Packet too large: {len(encoded)} bytes"
return encoded
# Example: early-manifest patient on ASO trial, faster-than-expected benefit
packet = HDOutcomePacket(
cag_repeat_bracket="46-55",
uhdrs_tmc_tier=1, # early manifest
intervention_class="ASO",
uhdrs_slope_vs_baseline=-0.18, # 18% slower decline than baseline
nfl_trajectory="stable",
mhtt_reduction_pct=42.0,
adverse_event_flag=False,
confidence=0.81
)
print(f"Routing address : {packet.routing_address}")
print(f"Packet size : {len(packet.to_bytes())} bytes")
print(f"Payload : {json.dumps(packet.payload, indent=2)}")
Any site treating an early-manifest patient with 46–55 CAG repeats on an ASO intervention queries that routing address and receives every relevant packet in the network. The site synthesizes locally. No patient data left any site.
The Quadratic Scaling That Matters Here
With N active HD trial and observational sites in a network:
- N = 40 GENERATION HD1 sites: 780 synthesis paths
- N = 200 ENROLL-HD contributing sites: 19,900 synthesis paths
- N = 500 (including EHDN, observational registries, and HD clinics in Venezuela, Bolivia, Japan): 124,750 synthesis paths
Currently, the number of real-time synthesis paths between those sites is approximately zero. The data exists. The routing architecture does not.
The N=1 Sites That Cannot Be Excluded
The LMIC inclusion argument is unusually specific for Huntington's disease. Venezuela's Lake Maracaibo region has the world's highest known HD prevalence — an estimated 700 symptomatic patients per 100,000 in some villages, compared to 10–12 per 100,000 in typical US or European populations. This concentration exists because of a founder effect: thousands of cases trace to a single ancestor, and the disease has been documented in the region for over a century. The genetic homogeneity of that population — high CAG repeat counts, specific HTT haplotype — makes it one of the most scientifically valuable HD cohorts on earth for understanding CAG repeat-specific disease progression.
Federated learning cannot include these sites. The population is genetically unique, the clinical infrastructure is resource-constrained, and the sample size per intervention class is too small for gradient-based learning to function.
QIS has no minimum cohort. An outcome packet from a neurologist in Maracaibo observing a patient with 65 CAG repeats on a neuroprotective intervention is a valid network contribution. The routing address is deterministic from the fingerprint. Any research institution or trial site querying for high-repeat-count juvenile-onset observations receives that packet alongside observations from MGH, UCSF, and University College London. The Maracaibo clinic has identical architectural standing to any major academic medical center.
What Thirty-Three Years Could Have Built
The HD field has accumulated something genuinely rare in medicine: three decades of longitudinal natural history data in a genetically characterized population, with standardized outcome measures, validated biomarkers, and an active global registry. ENROLL-HD and EHDN represent infrastructure that most diseases would require a generation to build.
That infrastructure produces retrospective insight on timescales of months to years. A routing layer built on top of it — routing pre-distilled outcome packets by CAG-repeat-stratified fingerprints across active trial sites in real time — would compound the intelligence that infrastructure generates without creating a single new data collection burden.
The tominersen failure cost more than two billion dollars and closed a therapy window for thousands of patients. The question is not whether the signal was in the data. It was. The question is whether the architecture to route that signal across 40 sites in real time could have changed the timeline.
The architecture exists now. Christopher Thomas Trevethan discovered it on June 16, 2025. The 39 provisional patents cover the protocol, not the transport layer. Any implementation that closes the loop — distill outcome at edge, route by semantic similarity to a deterministic address, synthesize locally — realizes the discovery.
For the HD field, that means ENROLL-HD and EHDN continuing to do exactly what they are doing, with a routing layer added that lets the signal from a trial site in Caracas reach a trial site in Boston in real time, without any patient data crossing any border.
The Quadratic Intelligence Swarm protocol was discovered by Christopher Thomas Trevethan on June 16, 2025. 39 provisional patents filed. For the architecture specification: QIS is an open protocol. For the complete loop in 60 lines of Python: QIS in 60 lines.
Previous in this series: Alzheimer's drugs: 99.6% failure rate, every trial in isolation · ALS and Parkinson's: 150 centers, zero routing
Top comments (0)