🧠 An Algorithm Walked Into a Memory Clinic and Predicted Your Future

This week belongs to Alzheimer's research. Three separate papers dropped within weeks of each other, collectively suggesting we're getting uncomfortably good at predicting the disease and simultaneously learning that our favorite prediction tools have a specificity problem. We've also got a placenta company selling its wound care business to chase immortality, and a longevity app that costs more than most people's rent. Here's what we've got.

RESEARCH

Here's a question nobody enjoys at a memory clinic: "How fast will I decline?" For most Alzheimer's patients, the honest answer has been a vague gesture at population averages, which is about as useful as telling someone the average weather in their country.

Researchers at Imperial College London and the UK Dementia Research Institute decided to fix that. Their machine learning model, published in Communications Medicine, predicts individualized 12-month cognitive and functional decline using nothing but routine clinic data. No MRI. No PET scan. No spinal tap. Just sub-item scores from standard cognitive assessments, demographics, and comorbidity codes your doctor already collects.

The best-performing model (an ElasticNet regression, for the stats crowd) achieved R² of 0.74 for cognitive decline and 0.77 for functional decline, with prediction errors below one standard deviation of average annual decline. Cognitive predictions validated externally on the US-based ADNI cohort with comparable accuracy, though no equivalent functional data existed for external testing. The team even built a clinician-facing tool called Theia that generates predicted scores with explainability baked in.

The catch? Training relied on 79 eligible cognitive trajectories and 74 functional ones from a pool of 153 total, which is modest by ML standards. Multi-center validation is needed before anyone deploys this in practice. But the real appeal is accessibility: while everyone else is chasing expensive biomarkers and imaging, this approach works with the data clinics already have. To be fair, even an imperfect crystal ball beats the current standard of educated shrugging.

RESEARCH

If you've been reading this newsletter for a while, you'll remember we covered Alzheimer's blood tests back when Roche got FDA clearance for its pTau181 test. Those tests answer a yes-or-no question: is amyloid pathology present? A team at Washington University just asked a much scarier one: when will you develop symptoms?

Published in Nature Medicine, the study introduces "clock models" built from plasma %p-tau217, the ratio of phosphorylated to non-phosphorylated tau at position 217. The concept is surprisingly intuitive. First author Kellen Petersen compared it to tree rings: amyloid and tau accumulate in consistent patterns, and once you know when they first became elevated, you can estimate when symptoms will arrive.

The timing turns out to be age-dependent. If your p-tau217 goes positive at 60, you've got roughly 20.5 years before symptoms. At 80, it's closer to 11.4. Prediction accuracy hit a median absolute error of 3.0-3.7 years across validation cohorts, and the team open-sourced all the code, because apparently, some still believe in open sharing.

Important caveat: the NIH summary explicitly states these models aren't accurate enough for individual use yet, though they could help select participants for trials. The study cohorts were also predominantly non-Hispanic White, so generalizability needs work. Still, going from "you might get Alzheimer's someday" to "probably around age 78, give or take three years" is a meaningful shift.

RESEARCH

Just when we got excited about p-tau blood tests (see above), a European team drawing on samples from four centers published what amounts to a cold shower. In Nature Medicine, researchers led by Mathias Jucker at the University of Tübingen found that p-tau181 and p-tau217 are also significantly elevated in systemic amyloidosis, a group of protein-misfolding disorders that damage the heart, kidneys, and nerves through amyloid deposition. The elevation? Two- to threefold, essentially the same magnitude as in Alzheimer's.

This matters because ATTR amyloidosis (the transthyretin variant) affects an estimated 10-25% of adults over 80. That's the exact population being screened for Alzheimer's with these blood tests. The risk of diagnostic false positives is not theoretical.

The study analyzed 280 serum samples across cohorts in Tübingen, Heidelberg, Groningen, and Pavia, finding that elevations held even after adjusting for age, sex, and kidney function. Polyneuropathy from non-amyloid causes showed no elevation, which at least means p-tau can distinguish between types of nerve damage.

The mechanistic explanation is interesting: tau protein isn't exclusive to neurons. It's expressed in cardiomyocytes, peripheral nerves, skeletal muscle, and kidney podocytes. When amyloid deposits stress these tissues, they release phosphorylated tau into the blood, the same as brain cells do in Alzheimer's. Jucker is blunt about the implications: p-tau "should not serve as a standalone diagnostic criterion" for AD. On the flip side, it could become a simple blood marker for systemic amyloidosis, which currently requires tissue biopsy or bone scintigraphy. Every cloud, apparently.

NEWS

Celularity, the placenta-derived cell therapy company founded by Robert Hariri, just licensed its commercial biomaterials portfolio to NEXGEL in a deal worth up to $35 million. Translation: selling the stuff that currently makes money to focus on the stuff that doesn't make money yet. Because of course they are.

The deal includes $15 million upfront plus up to $20 million in milestone payments, with royalties on top. The licensed products (Biovance, Interfyl, Acelagraft, and others) are established regenerative biomaterials for wound care and orthopedics, most with existing insurance reimbursement. Celularity keeps exclusive manufacturing rights at its Florham Park facility, which is a nice hedge.

What remains is the IMPACT platform and NK cell therapy pipeline, including CYNK-001 (placental natural killer cells in Phase 1 for glioblastoma). The longevity pivot centers on using placental NK cells to clear senescent cells, with expansion planned via state-level investigational use pathways in Florida, Texas, and Arizona rather than traditional FDA trials.

Some context on the financials: Celularity reported just $738,000 in cash at year-end and a $38.4 million operating loss. NEXGEL still needs to secure roughly $14.9 million in financing to close this deal. So this reads less as "strategic pivot" and more as "survival strategy wrapped in longevity branding." The NK cell science is legitimate, though. The runway just needs to catch up with the ambition.

NEWS

Human Longevity, Inc., the company Craig Venter co-founded in 2013 before departing in 2018 to do other Craig Venter things, has launched an AI-powered longevity app. It gives HLI members continuous access to their health data, filtered through AI-generated insights and an interactive chat feature. The underlying Executive Health platform integrates whole genome sequencing, whole-body MRI, metabolomics, proteomics, microbiome sequencing, and 100+ biomarker panels. Executive Chairman Dr. Wei-Wu He describes the app as shifting health management from episodic testing to continuous, AI-driven monitoring.

The scientific foundation isn't vapor. A 2020 PNAS paper from HLI's early work showed that integrating genomics with imaging and metabolomics caught 1.7% of participants with early-stage cancer on MRI and identified significant rates of insulin resistance, elevated liver fat, and calcified coronary plaque that participants didn't know about. The app is a consumer interface built on that clinical infrastructure, six years later.

Whether this qualifies as "democratizing longevity" depends on your wallet. HLI generates over 150 GB of data per assessment, runs liquid biopsy for cancer screening, and recently added Nobel laureate Michael Levitt (Chemistry, 2013) to its scientific advisory board. But at roughly $8,000 per visit with no insurance coverage, this remains firmly in executive retreat territory. Meanwhile, Venter launched a potentially competing genomics firm (Diploid Genomics) in January 2026, which is either ironic or just very Craig Venter.

If this edition had a theme, it would be "prediction is messy." We can forecast cognitive decline from a clipboard, estimate Alzheimer's onset from a blood draw, and generate 150 GB of health data per person (for a price). But tau, the biomarker everyone was betting on, apparently moonlights in cardiac and renal amyloidosis. And the company pivoting hardest toward longevity reported $738,000 in cash, which is roughly what a single senior scientist costs for a year.

The real progress this week isn't any single paper. It's the emerging picture: routine clinical tools getting smarter, blood tests getting more predictive, and our understanding of what those tests actually measure getting more nuanced. Which, honestly, is how science usually works.

If any of this made you think (or argue with your screen), forward it to someone who'd appreciate the ride. And hit reply, we read everything.

Keep questioning everything (especially standalone biomarker criteria),

Prateek & Jere

P.S. Robert Hariri co-founded both Celularity and Human Longevity Inc. The man collects longevity companies like some people collect sourdough starters. At some point, you have to wonder if he just really, really hates aging.