Dhiyanesh Srinivasan

Where AI meets aging science, anti-aging drugs and GLP-1s. New models and trials, explained for builders and researchers. Johns Hopkins. Ex longevity clinic.

[Dhiyanesh Srinivasan @DrDan_XD]

AI is changing how we study aging.

Most of what you see about it is hype.

I'm here to sort the signal from the noise.

[Dhiyanesh Srinivasan @DrDan_XD]

@MartinBJensen @impetusgrants Most of that tacit knowledge is failed experiments that never reach a journal.
An LLM can catch a stats error. It cannot know a rival lab already ran the mouse study and saw nothing.
My read: funders who collect grantee null results will train the better reviewer. Who starts?

[Dhiyanesh Srinivasan @DrDan_XD]

Source (Zhang et al., Age and Ageing 2026):
https://t.co/khk65iuxaL

[Dhiyanesh Srinivasan @DrDan_XD]

An MRI aging clock and a blood protein clock barely agree.

A new study in Age and Ageing used UK Biobank data from 46,734 people. An AI model (a Mixture of Experts) read MRI scans and body measurements. It estimated the age of four organs: brain, heart, liver and pancreas.

Predicted age tracked real age closely. Average correlation: 0.861.

The organ age gap is predicted age minus real age. It was linked to 20 age-related diseases. It helped predict disease in the same organ and in other organs.

The twist. Next to a blood protein age gap, the correlations were only weak. Combining the two predicted calendar age better. It did not consistently predict disease better than the organ scores alone.

Blood work found 54 to 154 proteins tied to each organ's gap. Nine showed up in all four organs. Those nine were also linked to death from any cause.

Fine print: one cohort and associations only. No one has shown that shrinking an organ age gap changes outcomes.

My read: the single "biological age" number is on its way out. Clocks will be reported organ by organ and by data type, because they measure different things.

If your MRI clock and your protein clock disagree, which one do you trust?

[Dhiyanesh Srinivasan @DrDan_XD]

@MartinBJensen One way to buy back incompressible time: shorter lives in real homes.
The Dog Aging Project follows tens of thousands of pet dogs. They age several times faster than us in our environments.
My read: that is the fastest longitudinal proxy we have. Should AI budgets fund it?

[Dhiyanesh Srinivasan @DrDan_XD]

@agingdoc1 A clock can move without risk moving.
No biological age test has shown in a trial that lowering the score leads to fewer heart attacks or deaths.
My read: AI diagnostics sell the readout. The missing piece is the outcome trial. Who runs it first?

[Dhiyanesh Srinivasan @DrDan_XD]

Sources:
https://t.co/zBgciyttWM
https://t.co/BZ0PnfiMfK
https://t.co/Qy50oNMNR2
https://t.co/fTqWWsKHms
https://t.co/Ut1sDkW6oZ

[Dhiyanesh Srinivasan @DrDan_XD]

Reprogramming is in humans. The AI-designed versions are not.

Partial reprogramming tries to make old cells act younger. Three programs, ranked by distance to the clinic:

1. ER-100 (Life Biosciences). In humans now. Gene therapy into the eye with the original Yamanaka factors OSK. NCT07290244 is Phase 1 in about 18 people with glaucoma or NAION. Primary endpoint: safety. Oct 8 interim: 3 glaucoma participants, no serious adverse events through day 56, visual fields better in 2 of 3. Primary completion May 2027.

2. NewLimit. Next in line. An AI system called Ambrosia predicts which transcription factor sets make old cells look young. Over 5,000 payloads tested. First product is RNA in lipid nanoparticles aimed at liver cells. Oct 21 update: first large animal studies run, guidance from regulators, trials planned for next year.

3. Retro Biosciences. Furthest out. With OpenAI, a protein model (GPT-4b micro) redesigned SOX2 and KLF4. Reprogramming marker expression rose more than 50-fold, in cells in a dish. Retro's one human trial, RTR242, is an autophagy drug, not reprogramming.

Fine print. Proven: short-term safety in 3 people. In trial: any benefit. Tested nowhere: whether reprogramming slows aging in people.

My read: the therapy in people uses factors found in 2006. AI's job is picking what comes after them.

My bet: NewLimit doses the first AI-picked reprogramming payload in a person in 2027, and its first readout is liver safety and function, not an aging clock.

Builders and researchers: what readout would convince you a reprogramming payload works in people?

[Dhiyanesh Srinivasan @DrDan_XD]

@kelway74 @Dr_Singularity Proof in humans is the part that can't be skipped.
Most drugs that look good in animals still fail in human trials.
My read: AI can pick better candidates. It can't replace the trial. What would convince you otherwise?

[Dhiyanesh Srinivasan @DrDan_XD]

@agingdoc1 Ovarian aging has a clean endpoint that arrives early: age at menopause.
Most aging clocks wait decades for mortality data to validate.
My read: a multi-omic panel matters only if it beats AMH alone at predicting it. Did they test that?

[Dhiyanesh Srinivasan @DrDan_XD]

@BowTiedPhys SELECT cut major heart events 20% over about 3 years. But only in people who already had heart disease (Lincoff 2023, NEJM).
Decade forecasts extend that to primary prevention.
My read: those models need real-world data to confirm the effect. Which cohort will test it first?

[Dhiyanesh Srinivasan @DrDan_XD]

Sources for the claim, the registries and the preprint:

https://t.co/248pHBD0gO
https://t.co/gHsIKMkfzf
https://t.co/1cKahSx3ue
https://t.co/7q3GEMh1mS

[Dhiyanesh Srinivasan @DrDan_XD]

Up to 10 years younger in 8 weeks? I checked the registry.

An AI-guided aging trial ran at Sheba Medical Center in Israel. 31 adults aged 50 to 80. In June the company's study director posted on LinkedIn: 80% responders, some with up to a 10-year drop on a glycan aging test.

The registry tells a smaller story.

Open label. No placebo arm. Each person got one of two drug combos plus two investigational supplements. One combo included semaglutide.
The primary endpoint is a muscle, immune and cognition composite. Not a glycan clock.

It was registered on ISRCTN in September 2026. The last visit was in February. The preprint itself says retrospectively registered.

The only results out so far are safety and feasibility. 30 of 31 finished. Primary endpoint results are still to come.

The AI part is real. An app scored daily face, tongue, eye and voice scans and sorted people into keep, augment, switch or stop. Doctors reviewed the calls. In this pilot they did not switch anyone's treatment.

My read: AI in aging trials will be judged on whether its early calls match the week 8 endpoint, not on clock drops.

Would you trust a responder label from a trial with no placebo?

[Dhiyanesh Srinivasan @DrDan_XD]

Preprint (Oct 19), plus the 2017 paper:

https://t.co/JruTIgG97F

https://t.co/6SXtlpb4Zx

[Dhiyanesh Srinivasan @DrDan_XD]

80% accurate. Saying "no" every time scores 77%.

A popular post this week said AI will solve aging. It cited a new bioRxiv preprint from the University of Kent.

The models predict which compounds extend lifespan in C. elegans worms. Best accuracy: 80.36%.

The fine print. 867 of the 1,120 compounds do not extend lifespan. That is 77.4%. A model that always guesses "no" scores almost the same.

The authors say accuracy misleads here. Their better metric, AUPRC, peaks at 0.53. Random ranking would score about 0.23. Real signal. Not a solved problem.

Also: worms only. Checked by cross validation, not on an outside dataset. No predicted compound has been tested in live worms yet.

A 2017 paper with the same senior author also reported about 80%.

My read: the field will stop citing accuracy and start citing hit rates. Of the top 20 picks, how many extend lifespan in the lab? The authors' own next test is compounds that hit the ser-5 receptor.

What baseline do you put next to accuracy in your models?

[Dhiyanesh Srinivasan @DrDan_XD]

@HaoYin20 @DogAgingProject A dog clock has a test human clocks lack: breed size. Big breeds live shorter lives.
If the residual tracks body size, it is mostly relearning a known risk factor.
My read: the useful number is mortality signal after adjusting for size. Did it hold?

[Dhiyanesh Srinivasan @DrDan_XD]

@m_goes_distance The aging line is the one that unlocks the rest.
A trial needs an endpoint. No aging clock is FDA-qualified as a biomarker, so TAME still counts age-related diseases and death.
My read: the first clock validated as a surrogate shortens every aging trial. Which is closest?

[Dhiyanesh Srinivasan @DrDan_XD]

@DrJesseMorse Short-lived controls are why the NIA's ITP tests every drug at 3 sites.
It matters for AI too. Lifespan prediction models train on DrugAge, which logs single-lab results like this one.
My read: unreplicated wins are noisy labels. Should models weight them by replication?

[Dhiyanesh Srinivasan @DrDan_XD]

@Dr_Singularity 80% cross-validation on known worm compounds mostly tests chemical similarity to the training set.
Worm lifespan effects also shift by strain and lab (Lucanic 2017, Nat Commun).
My read: the real test is prospective hits in new strains. Did they run any?

[Dhiyanesh Srinivasan @DrDan_XD]

Sources:

https://t.co/UxRLZqdU2b

https://t.co/hUwY3s6qn2

[Dhiyanesh Srinivasan @DrDan_XD]

FDA wants Alzheimer's drugs tested before symptoms start.

On Oct 22 it launched NAIN. It is a new neuroscience program that joins its drug, biologics and device centers.

Its first project is about trials that start before memory slips. FDA names three open questions. Who to enroll. Which endpoints fit each stage. What evidence shows decline was delayed.

The same week ARPA-H funded WashU to build an open, AI-enabled multi-omic platform. It also funded a biomarker-guided platform trial at Mass General Brigham.

My read: the bottleneck is not drug discovery. It is agreeing on what an early signal means. Whoever builds models that link biomarker shifts to later decline will help write the rules. Aging clocks will face the same test.

Builders and researchers: what would make an AI-predicted trajectory count as trial evidence?