In partnership with
Replace 16 supplements in one daily drink.
David Beckham. Giannis Antetokounmpo. Aryna Sabalenka. Inter Miami CF. The world's best choose IM8 because performance starts with what you do every day.
Daily Ultimate Essentials and Longevity deliver a complete, clinically formulated system for energy, digestion, sleep, focus and healthy aging. In a 12-week randomized controlled trial, 95% felt a boost in daily energy, 85% felt less bloated with better digestion, 80% reported better sleep and 75% reported sharper focus and mental clarity.
IM8 is NSF Certified for Sport, tested against 280+ banned substances and recommended by 300+ clinicians. With 50 million servings sold across 46 countries, discover the IM8 routine trusted at the highest levels of sport—and built for everyday life. See the clinical results.
LONGEVITY LATEST ISSUE 31 · 7 OCTOBER 2026
LONGEVITY LATEST · DEEP DIVE
The Phase 2 Wall
What AI really speeds up in the hunt for ageing drugs, where it still stalls, and how to read the next “AI drug reverses ageing” headline.
By Christian Thomsen · Companion to Issue 31 · 7 October 2026 · ~7-minute read
I expected the AI drugs to be failing early. Odd chemistry from a black box, the sort of molecule that falls over the first time it meets a human liver. That’s not what the data showed.
In a 2024 analysis, researchers at Boston Consulting Group tracked AI-discovered molecules that had completed a clinical phase. In phase 1, the first test in people and mostly a test of short-term safety and dosing, 80 to 90% succeeded, against a historical norm of roughly 40 to 65%. In phase 2, the first real test of whether a drug works in patients, about 40% succeeded, close to the historical norm. The phase 2 figure rests on just ten molecules.
On early evidence, AI is good at making molecules that clear first human testing. It hasn’t yet shown it is better at picking ones that work, and ten molecules are far too few to rule that out.
Treat that as an early reading, not a verdict. But it points at the right question: which parts of drug discovery AI demonstrably speeds up, and which it hasn’t yet been shown to touch.
Three questions, three different problems
Drug discovery asks three questions in sequence. Which protein should we hit? What molecule hits it? Does hitting it help a person? AI now touches all three, but its record differs sharply between them.
Structure: a protein’s 3D shape tells chemists where a drug could bind. AlphaFold, whose developers shared the 2024 Nobel Prize in Chemistry, predicts those shapes from sequence alone. This part is settled science.
Chemistry: generative models propose molecules, and simulation plus machine learning predicts how tightly they’ll bind and how the body will handle them. This is where the phase 1 numbers come from: fewer toxic surprises, faster design cycles.
Biology: choosing a target that actually drives disease in people. The inputs are messy (gene expression, tissue samples, animal models), and the answer only arrives in phase 2.
A simplification worth flagging: real programmes blur these steps, and many drugs labelled “AI-designed” used AI at only one of them. That’s where most headline confusion starts.
Two drugs, two kinds of AI
Zasocitinib is the closest thing to an approved computationally designed drug. Nimbus Therapeutics began work on TYK2 around 2016, using physics-based free-energy simulation and machine learning to assess more than 13,000 compounds on the computer before making them; the eventual molecule was identified in 2020. The target was already being pursued elsewhere: Bristol Myers Squibb’s TYK2 drug, Sotyktu, reached the US market first, in September 2022. So the hard part for Nimbus was chemistry, hitting TYK2 cleanly without its JAK relatives, not proving TYK2 mattered. Takeda bought the programme for $4 billion upfront. In the LATITUDE Atlas head-to-head trial, presented in full on 30 September, more than 2.5 times as many patients reached complete skin clearance at week 16 as on Sotyktu. The FDA decision is due in early 2027. Nimbus’s own R&D head declines the “first AI-approved drug” label.
Rentosertib is the opposite bet: AI picked the target, TNIK, as well as generating the molecule, and the programme reached the end of phase 1 in under 30 months. After a 71-patient, 12-week phase 2a (graded in Issue 31), its 320-patient phase 3 dosed its first patient on 9 September.
Put simply, zasocitinib shows that AI-assisted chemistry can produce a winner against a target someone else validated. Rentosertib is testing whether AI can also pick the target. Only the second question really matters for ageing, because most ageing targets have never been validated in humans at all.
Fierce Biotech on Nimbus and zasocitinib (2026) · Takeda NDA acceptance (2026) · Xu et al. GENESIS-IPF. Nat Med (2025) · Insilico first patient dosed (September 2026)
The senolytic hunt
Senolytics, drugs that clear senescent “zombie” cells, show the same pattern in miniature. Two 2023 studies used machine learning to find new ones.
Edinburgh: models trained on just 58 known senolytics, set among 2,523 compounds from the published literature, flagged ginkgetin, periplocin and oleandrin. All three cleared senescent human cells in the dish, at a several-hundred-fold lower screening cost.
MIT and the Broad Institute: graph neural networks trained on a 2,352-compound screen scored more than 800,000 molecules. One hit, BRD-K56819078, reduced senescent-cell burden in the kidneys of aged mice.
Here’s what the hype misses. Oleandrin comes from oleander, one of the more poisonous plants in the garden centre. It’s a cardiac glycoside, from the same family as the heart drug digoxin, with a narrow gap between effect and harm. A senolytic in a dish is a molecule that kills stressed cells. Whether it kills the right ones, safely, in a 70-year-old is a separate question nobody has answered.
The model did what it was built to do. One of its best leads is still a known poison.
None of these compounds has published human ageing data. That isn’t a criticism of the software. It’s a reminder of where the software’s job ends.
Why the comparison itself can mislead
Before reading too much into 80 to 90% against 40 to 65%, it’s worth knowing why comparisons with historical trial data are slippery.
Different targets: AI programmes have often chosen well-understood targets, where phase 1 is easier to pass. Some of the advantage may be target choice, not the algorithm.
Different definitions: “AI-discovered” has no agreed meaning, and “success” in phase 1 can mean anything from clean safety data to a decision to keep going.
Different eras: historical benchmarks cover drugs from earlier decades, with different regulators, trial designs and disease mixes.
Different visibility: companies announce successes faster than quiet terminations, so small tallies can lean optimistic.
None of this means the phase 1 advantage is false. It means the numbers describe a young field’s first few dozen molecules, not a settled law. The comparison to watch is AI-assisted versus conventional programmes on the same kinds of target, followed for long enough to see phase 3.
Why phase 2 is the wall
Phase 2 is biology’s first honest test, and three things make it harder for ageing than for a skin disease.
No agreed target list: psoriasis has pathways proven in people. Ageing has hallmarks with mostly animal evidence behind them.
No quick endpoint: a psoriasis score moves in 16 weeks. Fewer heart attacks, less frailty or longer life take years to show, and regulators don’t recognise ageing itself as something to treat.
Surrogates still under construction: clocks are the obvious shortcut. Yale’s TranslAGE work, 51 intervention studies run through 16 clocks, shows they respond unevenly. The rentosertib clock analysis can’t separate slower ageing from treated disease, a point its peer reviewers pressed hard.
There’s a deeper problem underneath. A model learns to predict what works in humans from examples of things that worked in humans. For ageing, there are very few such examples. AI can’t shortcut a missing answer key.
What would change my view: these are two separate bars. A phase 3 that slows lung decline over a full year would make rentosertib a credible IPF drug; that’s evidence about a disease. Evidence about ageing would need more: an independent group finding a clock shift in people without the disease, then fewer age-related events. The second bar is much higher, and nobody has cleared it for any drug.
Five checks on an AI-drug headline
Which step: did AI find the target, design the molecule, or both? “AI-designed” often means chemistry help on a known target.
Which stage: cells, mice or people? Phase 1 success is mostly about safety.
Which endpoint: symptoms, events and function, or a clock or blood marker?
Who wrote it: company-authored analyses aren’t wrong by default, but they need independent replication.
How long, how many: 12 weeks and 71 people is a signal. A year and 320 people is a test.
Run those five before letting a result change what you do, or what you buy.
What this means for you
The two drugs that matter here are investigational and will reach patients, if at all, through prescriptions for specific diseases. The laboratory hits are not supplements either, oleandrin above all, whatever an online shop claims.
If I were designing a protocol from this evidence, it would be boringly simple: attend the screening programmes AI is already improving, keep the risk factors with decades of outcome data under control (blood pressure, lipids, glucose, smoking, activity), and watch two milestones. Zasocitinib’s FDA decision, expected in early 2027, will show what AI-assisted chemistry can deliver. Rentosertib’s phase 3 can’t read out before its last patient finishes 52 weeks of treatment; when it does, it will show whether an AI-picked target holds up, and its blood samples at weeks 12, 26 and 52 will give the clock question a second, larger test.
The software is fast. Biology still sets the pace.
Sources and further reading
Evidence reviewed through 3 October 2026. Phase-success figures come from one analysis with small numbers at phase 2 (ten molecules). Rentosertib human data come from a single 12-week phase 2a trial, and the clock analysis was led by the developer; the senolytic findings are cell and mouse studies only. Educational content only; rentosertib and zasocitinib are not approved, and no compound named here should be taken outside a clinical trial or prescription.
© 2026 FrontWave Media Ltd · Longevity Latest