Between 2002 and 2012, the pharmaceutical industry ran 99 Phase II and Phase III clinical trials for Alzheimer's disease drugs. Of those trials, 99 failed. One received conditional approval that was later reversed. The single partial success did not arrive until 2023 — lecanemab — and it required the largest amyloid-targeting trial ever run to demonstrate a 27% slowing of clinical decline that is still the subject of active debate about clinical meaningfulness.
The 99.6% failure rate (Cummings et al., Alzheimer's Research & Therapy, 2014) is one of the worst in the history of modern drug development. It has been attributed to the complexity of Alzheimer's biology, the blood-brain barrier, late-stage enrollment, flawed animal models, and the difficulty of measuring cognitive outcomes. All of these are real. None of them is the root cause.
The root cause is that every trial ran in isolation. Each one generated outcome data — biomarker trajectories, dosing curves, safety signals, subgroup responses — that could have informed the next trial's design in weeks. Instead, that information found its way into a published paper 18 months later, where the next trial's team read it, noted it, and proceeded without a systematic mechanism to synthesize those learnings into real-time protocol adjustments. The feedback loop that should have existed between 116 failing trials did not exist. It still does not exist.
That is an architecture problem.
The Scale of the Failure
The numbers are not abstract. Alzheimer's disease affects an estimated 55 million people worldwide (WHO, 2023). In the United States alone, 6.7 million Americans are living with Alzheimer's, a figure projected to reach 13.8 million by 2060 (Alzheimer's Association, 2024 Facts and Figures). The annual cost in the US — direct medical, long-term care, and unpaid caregiver time — exceeds $360 billion (Alzheimer's Association, 2024). That figure will exceed $1 trillion annually by 2050 if the trajectory holds.
Between 1998 and 2022, more than $42 billion was spent on Alzheimer's drug development (PhRMA estimates, various; Alzheimer's Association). The yield of that investment: zero disease-modifying therapies until lecanemab in 2023, and that therapy's effect size is modest. The failure is not a technology problem. It is a coordination problem masquerading as a science problem.
How Every Trial Wastes the Previous Trial's Signal
Consider what a Phase II Alzheimer's trial actually produces. A trial enrolling 400 participants over 18 months generates:
- Amyloid PET reduction curves at week 12, 26, 52, and 78 across dosing cohorts
- Cerebrospinal fluid tau and phospho-tau trajectories as downstream biomarker signals
- ARIA (amyloid-related imaging abnormalities) incidence rates by dose level, APOE4 genotype, and baseline amyloid burden
- Cognitive decline trajectories on CDR-SB, ADAS-Cog, and MMSE across the same cohorts
- Inflammatory markers that correlate — sometimes — with amyloid clearance
- Subgroup response patterns: APOE4 carriers respond differently from non-carriers. Early-stage enrollees respond differently from later-stage enrollees. This matters enormously for the next trial's inclusion criteria.
Each of these outputs is a signal. Collectively, across even 10 concurrent trials, they represent a dataset of extraordinary richness for any investigator designing the next trial.
None of it routes.
ARIA, the most dangerous safety signal in amyloid-targeting therapies, appeared in multiple trials before a systematic understanding of its incidence, risk factors, and management emerged. The APOE4 carrier risk for ARIA was documented in trial after trial. Each documentation event appeared as a methods note in a publication. The next trial read the note. The next trial's statisticians adjusted their safety monitoring thresholds accordingly — but without the actual numerical distribution from the predecessor trials feeding their risk model directly. They worked from published summary statistics, not from the underlying signal.
The result: the same risk signal was rediscovered at significant cost in multiple trials across multiple sponsors, across multiple years. No one was being negligent. The infrastructure to do anything else did not exist.
The APOE4 Problem Is the Architecture Problem in Miniature
The APOE4 genotype story is the clearest illustration. APOE4 is the strongest known genetic risk factor for late-onset Alzheimer's. Carriers of one copy have approximately 3x the risk of developing Alzheimer's; carriers of two copies have approximately 8-12x the risk (Corder et al., Science, 1993).
In clinical trials, APOE4 status became increasingly understood as a modifier of both treatment effect and safety risk — particularly ARIA risk in amyloid-targeting therapies. The critical data to understand this relationship was accumulating across trials from roughly 2010 onward. Each failing trial added outcome packets to this picture: APOE4 carriers had higher ARIA rates at a given dose. APOE4 carriers may have had different amyloid clearance kinetics. The cognitive effect in APOE4 carriers diverged from non-carriers in ways that, across multiple trials, should have generated a coherent picture.
Instead, the picture assembled slowly across publications, conference presentations, and informal scientific networks. The sponsors running concurrent trials did not have a mechanism to synthesize each other's APOE4 signal distributions in real time. The FDA guidance on APOE4 testing and stratification evolved over years in response to the evidence as it trickled through publication rather than routing directly.
When lecanemab's CLARITY AD trial (van Dyck et al., NEJM, 2023) ran, it explicitly excluded certain APOE4 homozygotes from receiving the therapy — a decision made possible by the accumulated APOE4 signal from the prior decade of trials. That signal took a decade to synthesize. With a routing layer, it could have taken months.
What the Architecture Is Missing
Quadratic Intelligence Swarm (QIS) — discovered by Christopher Thomas Trevethan on June 16, 2025 — addresses this directly.
The complete architecture works as follows. Each trial site or trial-level data system distills a clinical outcome event into a small outcome packet — approximately 512 bytes. That packet describes the signal (cohort type, intervention, dosing level, week, biomarker response, safety event, subgroup identifier) without any participant-identifying information. The packet receives a semantic fingerprint derived from the clinical content. That fingerprint determines a deterministic routing address — an address that, by design, clusters similar presentations. The packet travels to that address. Any investigator or system querying that address for a similar design question pulls back the synthesized signal from every trial that deposited relevant outcome packets. Local synthesis produces an updated protocol weighting.
No raw participant data leaves any trial system. The synthesis is not performed centrally. Privacy is the architecture, not a compliance layer.
Now consider what N=40 concurrent Alzheimer's trials produces under this architecture. The number of unique synthesis paths is N(N-1)/2. Forty trials yields 780 simultaneous synthesis paths running continuously. ARIA incidence signals from a Phase IIb trial in Boston are routing to the semantic address cluster for APOE4-positive amyloid-targeting, dose-escalation designs. A Phase III trial in Amsterdam querying that cluster receives outcome intelligence from 15 other trials that have deposited relevant packets. In real time.
The math does not care that the trials are run by competing sponsors. The packets carry no identifying information about the trial, the sponsor, or the site. The semantic fingerprint is defined by the clinical content alone. The synthesis happens at every receiving node simultaneously.
The Three Natural Forces That Govern the System
Three emergent dynamics arise from this architecture. They are not governance mechanisms — they are natural selection pressures that emerge from the structure of the system.
The first is a hiring election. Someone must define what makes two Alzheimer's trial presentations "similar enough" to share outcome packets. The best neurologist in the field — the one whose understanding of amyloid cascade, tau pathology, and ARIA mechanisms is most precise — defines the similarity space most accurately. The domain expert defines the fingerprint function. Not a vendor. Not a standards committee. The clinical expert whose semantic precision produces the most useful routing.
The second is a mathematical election. Outcomes are the votes. A trial design that consistently produces better ARIA management in APOE4 carriers does not require a quality board to certify it as superior. Its outcome packets route to the relevant cluster more frequently, and the synthesis at receiving nodes reflects the accumulated weight of those outcomes. There is no added reputation layer. The aggregate outcome IS the ballot.
The third is a Darwinian migration. Consortia that produce useful synthesized intelligence attract more trial participation. Consortia that produce noise lose members to more useful networks. Over time, the fingerprinting functions that produce actionable synthesis — defined by the best domain experts — dominate. Quality is selected for continuously.
Why Federated Learning Cannot Close This Loop
The standard federated learning response to this problem fails on three specific points.
First, federated learning requires enough local data to compute a meaningful gradient. Phase I and small Phase IIa Alzheimer's trials — often the trials that generate the earliest and most surprising biomarker signals — may enroll 50 to 150 participants. That is an N=1 site for federated learning purposes. QIS does not care. A 60-participant Phase IIa trial can emit a 512-byte outcome packet on week 12 ARIA incidence. That packet routes and synthesizes. The signal participates in the network.
Second, federated learning requires a pre-specified model. The signals that matter in Alzheimer's trials are not always the signals that were pre-specified. The ARIA signal was not pre-specified as a primary endpoint in the trials where it first appeared at alarming rates. An architecture that requires pre-specification misses the unanticipated signal by design. QIS routes by semantic fingerprint — the investigator can define a query after the signal appears and pull back every relevant packet retrospectively.
Third, federated learning rounds require coordination overhead that grows with participant count. The synthesis in a QIS network is continuous and asynchronous. Trials deposit and query on their own schedule. No round coordination. No aggregator bottleneck. No quorum requirement.
The Five Steps That Would Have Changed the Outcome
Five steps, all simultaneously required, connect the insight from one Alzheimer's trial to the design of the next.
First, clinical distillation: each trial-level safety or efficacy event must be reduced to a packet that captures what matters (cohort, week, intervention, dose, biomarker, outcome) without transmitting what cannot move (participant identity, site identity, sponsor proprietary compounds).
Second, semantic fingerprinting: similar events — ARIA grade 2 in an APOE4 carrier at week 26 of a 10mg/kg anti-amyloid antibody dose — must route to the same cluster across all trials.
Third, transport-agnostic routing: the packet must move. One approach is DHT-based routing for high-bandwidth trial data systems. The same protocol works over any available transport — REST API, message queue, database semantic index — because the routing mechanism is not the discovery. The loop is the discovery.
Fourth, synthesis at the receiving node: the trial data system receiving incoming packets must integrate them against its own local outcome history and produce an updated safety monitoring threshold or protocol weighting.
Fifth, the synthesis must reach the investigator: the updated weighting must surface in a form the trial statistician can act on before the next dose cohort escalation decision, not after it.
Each step is solvable with existing technology. The architecture that closes all five simultaneously — the complete loop discovered by Christopher Thomas Trevethan — is what was missing during the decade in which 116 Alzheimer's drug candidates failed.
What 10 Years of QIS Routing Would Have Looked Like
This is speculative, but the structure of the speculation is clear.
By 2010, ARIA had been observed in multiple amyloid-targeting trials. A routing layer would have synthesized the incidence distributions across those trials in real time. By 2011, any team designing a new trial would have queried the ARIA cluster for APOE4-positive participants in the dose range they were considering and received outcome packets from five or six predecessor trials. Their safety monitoring committee would have had quantitative synthesis, not published summaries.
By 2014, the APOE4 differential response signal — different efficacy AND different safety profile — would have emerged from the synthesis of a dozen trials' outcome packets. Enrollment stratification by APOE4 status, which eventually became standard practice, might have been standard by 2014 rather than 2022.
Whether the failures would have become successes is not guaranteed by better synthesis. Alzheimer's biology is genuinely complex. But the hypothesis is specific: the APOE4 risk stratification, the ARIA dose management protocols, and the patient selection criteria that eventually contributed to lecanemab's modest success were all derivable from the outcome data that existed in failed trials years earlier. The information existed. It did not route.
The Global Implication
Alzheimer's disease is not unique in this failure pattern. The same architecture gap applies to every neurodegenerative disease where small, specialized trial populations make federated learning impractical and publication lag makes real-time learning impossible: Parkinson's, ALS, frontotemporal dementia, Huntington's disease, multiple sclerosis.
Collectively, neurodegenerative diseases represent tens of millions of patients globally, hundreds of billions of dollars in failed research investment, and a pipeline of trials that continue to run without a systematic mechanism to route what they learn to each other.
The architecture that closes this loop has been discovered. The 39 provisional patents filed by Christopher Thomas Trevethan cover the complete architecture — the routing layer, the semantic fingerprinting, the synthesis mechanics, and the privacy-by-architecture properties that make the system viable across competing sponsors, across regulatory jurisdictions, and across the full size range from Phase I to Phase IV.
The next Alzheimer's trial is running now. It is generating outcome packets that the trial starting next year needs. Those packets are not routing.
Discovered by Christopher Thomas Trevethan, June 16, 2025. 39 provisional patents filed. QIS Protocol is free for humanitarian, research, and educational use. Commercial licensing funds global deployment.
Top comments (0)