Will you live to 120? · The Cells That Won’t Die

Where the machines come in

And now the honest, balanced status of the whole field, because the lesson is “be disciplined,” not “give up.” Senolytics are not dead. Small human trials have shown that drugs like dasatinib-and-quercetin really can reduce the burden of senescent cells in people — the first direct evidence that human senolysis is even possible — in conditions ranging from a lung-scarring disease to diabetic kidney disease, and trials in other diseases including Alzheimer’s are underway. The eye program has shown promise. But as of now, the truth is this: no senolytic has been approved for anything; no large trial has clearly demonstrated the anti-aging — or even firm disease-modifying — promise at scale; and the first marquee test of the whole idea failed. Senescence remains one of the most compelling targets in aging biology and one of its least proven therapies. Both halves are true at once. A reader who hears only “mice lived longer when we cleared zombie cells” and not “the first big human trial failed” has, once again, learned the wrong half.

Where the machines come in

Two places in this story are genuinely machine-shaped, and one of them traces directly back to a problem we flagged early in the chapter.

The first is the marker problem. Because there is no single test for senescence — because it is a heterogeneous state recognized only by a combination of features — identifying senescent cells reliably is exactly the kind of pattern-recognition task that machine learning is built for. Researchers now train deep-learning image classifiers to spot senescent cells from their morphology and from multiplexed molecular imaging, learning the combination of subtle features that no single stain captures. A major coordinated effort (the NIH’s Cellular Senescence Network, SenNet) is building tissue-by-tissue atlases of senescent cells and the classifiers to find them. This matters because better detection is the precondition for everything else: you cannot study, target, or confirm the clearance of cells you cannot reliably identify. Note, though, the now-familiar caution carried over from Chapter 2 — a classifier that predicts “this cell is senescent” is still a predictor, and it is only as good as the labeled examples it learned from, in a field where even the experts’ labels are uncertain.

The second is drug discovery itself: using computational screening to find new senolytics and senomorphics — predicting which compounds will selectively pull a senescent cell’s survival crutch, and, more valuably, designing ones with wider therapeutic windows than navitoclax, targeting survival pathways that senescent cells depend on but vital healthy cells don’t. This is where artificial intelligence could earn real ground, because the selectivity problem — kill the zombie, spare the platelet — is partly a molecular-design problem, and molecular design is exactly what the newest AI tools are getting good at.

But hold onto the chapter’s hardest lesson when you weigh these. The Unity trial did not fail for want of a better algorithm or a cleverer molecule. It failed because a piece of mouse biology did not hold in a human being. AI can sharpen our detection of senescent cells and accelerate the search for drugs that kill them. It cannot, on its own, tell you whether clearing those cells will cure a human disease — that question is answered only in the clinic, and so far the clinic’s answer has been “not yet.” The binding constraint in this field is not computation. It is the stubborn, expensive gap between a mouse and a person.

✦ ✦ ✦

Carry these three things

1. A senescent cell is not exhausted — it is locked into an active, harmful state on purpose. Triggered by danger (a short telomere, an oncogene, irreparable damage), it bolts itself out of the cell cycle with p21 and p16, reinforces its resistance to dying, and pours out the SASP — an inflammatory, tissue-corroding secretion that spreads senescence to its neighbors and, summed across the body, is much of what we call inflammaging. The chronic inflammation of aging has a cellular address.

2. Senescence is the brake-and-price trade-off again, now as an evolutionary principle. It suppresses cancer and aids healing when you are young, and poisons your tissues when you are old — a textbook case of antagonistic pleiotropy, a trait selected for its early benefit despite its late cost. So “kill every senescent cell” is, like “turn on telomerase,” removing a brake that was there for a reason; the real target is the chronic harmful cells, not the acute useful ones, and we cannot yet cleanly tell them apart. They accumulate because aging raises their production and lowers their clearance at the same time — $N^* = P/c$, with the numerator climbing and the denominator shrinking — which is why the burden doesn’t creep but accelerates.

3. The most beautiful idea in the field met the most expensive disappointment. Clearing senescent cells rejuvenated mice; the first major human trial — Unity’s, run by the field’s own founders with serious money behind it — failed for knee osteoarthritis. Not a failure of chemistry or AI, but the same mouse-to-human gap from the last two chapters, in a third costume. Senolytics remain genuinely promising and genuinely unproven: real evidence that we can clear these cells in people, no proof yet that doing so cures anything at scale. So far we have been trying to remove the damaged cells, or measure their damage. The next chapter asks a more ambitious question — whether we can slow the rate of damage itself, by turning the ancient dials a cell uses to sense food. It is where the field’s hardest fights live, and where its single most oversold molecule is waiting.

All 10 chapters of Will you live to 120? · Foundations · All courses