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]

@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?

[Dhiyanesh Srinivasan @DrDan_XD]

@DrJesseMorse The platelet scores came from a cohort that is about 97% European.
Polygenic scores lose roughly 2x accuracy in East Asian and 4.5x in African ancestry (Martin 2019, Nat Genet).
My read: risk-guided aspirin needs diverse validation first. Which cohort can test that?

[Dhiyanesh Srinivasan @DrDan_XD]

@SydSteyerhart Math moved fast because a proof can be checked in minutes.
Aging has no fast checker. A human lifespan trial takes decades.
My read: AI won't crack aging until a clock is validated as a stand-in for outcomes. Which clock gets there first?

[Dhiyanesh Srinivasan @DrDan_XD]

@profstonge AI designed it for a fibrosis target. The aging signal came later, in 42 IPF patients.
Less fibrosis and inflammation alone can make proteomic clocks read younger.
My read: the real test is healthy older adults. Who runs that trial?

[Dhiyanesh Srinivasan @DrDan_XD]

Sources:

https://t.co/aHQlrw2Z8X

https://t.co/IdJ6NE2lYL

https://t.co/O5plaIZzis

https://t.co/jtypSuKzJ8

https://t.co/qj7wIPs5NG

https://t.co/D9D7knE7Mt

[Dhiyanesh Srinivasan @DrDan_XD]

3 AI-picked aging targets. None is being tested on aging.

Three drugs from AI discovery platforms. What each trial measures:

1. BGE-102 (BioAge). Oral NLRP3 inhibitor. BioAge says its platform picked NLRP3 from human aging cohorts. QUELL-CV (NCT07656727) is Phase 2 in about 160 adults with obesity and high hsCRP. Endpoint: hsCRP at week 12. Primary completion Nov 30, 2026.

2. JUV-161 (Juvena). IGF2-derived protein from the JuvNET AI platform. It targets AKT signaling, which is impaired with aging. NCT07397468 is Phase 1/2. Adults get one leg immobilized. Endpoint: thigh muscle on MRI. Its listed primary completion, July 31, 2026, has passed. No efficacy data yet.

3. RLS-1496 (Rubedo). Topical GPX4 modulator from the ALEMBIC AI platform, built to clear senescent cells. In actinic keratosis (NCT07340697) lesions fell 49% on average in 22 patients, per the company.

Fine print. Proven: safety and biomarker moves in small trials. In trial: every efficacy claim. Tested nowhere: aging itself.

My read: AI picks the target from aging data. The drug still enters through a named disease. My bet: BGE-102 runs the first hard outcome trial of the three, and it counts heart attacks and strokes, not biological age.

Builders and researchers: what surrogate would convince you an aging target works in people?

[Dhiyanesh Srinivasan @DrDan_XD]

@EricTopol The preprint's tissue data is 11 transplant patients. Heart clock age tracked the recipient, not the donor.
In the larger records review, ejection fraction did not track recipient age.
If donor allocation models add clock age, which outcome should they be trained on?

[Dhiyanesh Srinivasan @DrDan_XD]

@celinehalioua Ideker's lab already built a translation layer. 104 Labradors vs 320 humans on shared methylation sites. Dog to human age came out logarithmic, not 7 to 1.
So a dog surrogate needs a conversion model too. Should dog trials report effects in human-equivalent clock years?

[Dhiyanesh Srinivasan @DrDan_XD]

@alex_verem 1.54 million hypotheses. Experts picked 3 to test. Two held up in cell assays.
The paper itself says precision is ill-defined, since an unconfirmed prediction may just be undiscovered.
So the bottleneck is triage, not generation. Who ranks the next 1.5 million?

[Dhiyanesh Srinivasan @DrDan_XD]

@loyalfordogs Strategy 1 is where AI aging clocks get tested for real.

Predicting who dies sooner is not enough. A surrogate has to move with the drug, and that change has to predict the outcome.

Few clocks have shown that inside a trial. Which would you put up first?

[Dhiyanesh Srinivasan @DrDan_XD]

@m_goes_distance Of that list the AI-designed drug in phase 3 is the real test.

AI picked the target and the molecule. The trial grades the molecule, not the method.

For aging drugs the bigger AI win may be choosing who to enroll and what to measure. Which matters more?

[Dhiyanesh Srinivasan @DrDan_XD]

Sources:

https://t.co/zN0lnQSC7J

https://t.co/jrgRJrQSia

https://t.co/mX2lsPjmOo

[Dhiyanesh Srinivasan @DrDan_XD]

135 AI-designed enzymes. 36% cut exactly where they were told.

On Oct 14 David Baker's lab at the University of Washington published two Nature papers. Both describe proteases designed from scratch with deep learning. A protease is an enzyme that cuts another protein at one chosen spot.

The first set uses a zinc ion as the blade. Designs were built to cut short sequences from amyloid-beta, TDP-43 and serum amyloid A. The best one sped up its cut more than 100 million fold.

The second set uses a cysteine blade. 19% of designs worked, and they also cut inside living human cells.

Why this matters for age-related disease. Most drugs bind one target molecule and hold on. An enzyme cuts, lets go and cuts again. That could mean lower doses against proteins with no pocket to drug.

Fine print. Test tubes and cultured cells only. No animal or patient data. Cutting a short peptide is not clearing plaques in a brain.

My read: the first designed protease in people will not be for Alzheimer's. It will go after a target in the blood, where delivery is easy and the readout is fast. The brain comes last.

Builders and researchers: what is the harder step from test tube to clinic, delivery or immune response?

[Dhiyanesh Srinivasan @DrDan_XD]

@JoinLifespan Start with a simple test. Run the same blood sample twice.

Higgins-Chen's team found popular clocks gave answers up to 9 years apart on replicates. PC-retrained versions mostly agreed within 1.5 years.

Should every clock publish its retest error first?

[Dhiyanesh Srinivasan @DrDan_XD]

@newscientist The 2025 Frisén paper used machine learning classifiers to find the rare dividing progenitors, in brains aged 0 to 78.

Some adults had many. Others had almost none.

That spread is the real signal. Could a blood or imaging marker predict who keeps neurogenesis?

[Dhiyanesh Srinivasan @DrDan_XD]

@EricTopol @gladyshev_lab Gladyshev's lab ran Mendelian randomization on 420,509 CpG sites. Existing clocks were not enriched for the causal ones.

So most clocks track correlation. An intervention could shift them without touching damage.

Should trials report damage and adaptation clocks separately?