AI in Longevity Research: Where Machine Intelligence Helps, and Where It Cannot

The 2025 Scripps Research study on AI-driven longevity compound discovery — in which over seventy percent of candidates selected by a machine learning system extended lifespan in model organisms — is one of the more substantive recent demonstrations of what AI can contribute to aging research. The study is worth reading carefully, both for what it shows and for what it does not.

What it shows is that pattern recognition across very large multi-omic datasets can identify compounds whose lifespan-extending activity would have taken human researchers substantially longer to find. The AI did not invent biology. It surfaced relationships in published data that human researchers had not yet integrated. That is a specific contribution, and it is a real one.

What it does not show is that AI can replace the judgement layer of life-sciences work. The candidates the AI surfaced still required human researchers to evaluate, prioritise, formulate, and test. The lifespan extension was observed in model organisms; translation to human biology remains the same hard problem it has always been. And the regulatory, safety, formulation, and commercial layers that determine whether a candidate becomes a usable product are the layers AI is least able to navigate on its own.

The broader AI-in-aging research wave

Insilico Medicine, Recursion Pharmaceuticals, BioAge Labs, Isomorphic Labs, and a long tail of smaller computational biology firms have built businesses around the proposition that AI applied to aging biology produces faster, better, or more numerous drug candidates than conventional pipelines. The proposition is broadly correct at the candidate-generation layer. It is unevenly correct at the candidate-validation layer, where the harder biology lives. It is not yet correct at the regulatory, formulation, and commercial layers, where the work that actually produces a product is done.

The pattern that emerges from observing the AI-in-life-sciences sector for several years is consistent. AI is a force multiplier at certain stages of the pipeline. It is not a substitute for the stages where human judgement is what produces the work. The companies that perform well over time are organised around this distinction. The companies that have struggled have tended to be organised as though AI could carry layers it is not yet able to carry.

What HI-Centric-AI is

HI-Centric-AI — Human-Intelligence-Centric AI — is the institutional name for the AI architecture Atumnus Life Sciences operates. The name describes both the technical design and the working philosophy. The infrastructure is used to synthesise the bioregulator research literature at scale, to model peptide-class compound behaviour and delivery characteristics, to support institutional knowledge management across the IP portfolio, and to accelerate the work of human researchers and writers across the broader Atumnus programme.

The architecture is built around a specific design commitment. AI components produce drafts, surface relationships, and accelerate iteration. Human judgement decides what to publish, what to prosecute as IP, what to qualify as Opticeutical, and what to claim about the underlying biology. The decision authority is human. The acceleration is machine. Reversing those, in our considered view, produces work whose epistemic status the reader cannot evaluate — which is the failure mode that several AI-first life-sciences efforts have demonstrated in public.

“AI components produce drafts. Human judgement decides what to publish. The decision authority is human. The acceleration is machine. Reversing those produces work whose epistemic status the reader cannot evaluate.”

Why we surface this publicly

Three reasons. First, the AI-in-life-sciences sector is at a stage of its development where the design commitments of individual programmes matter. Programmes that have made the HI-Centric commitment behave differently from programmes that have made the AI-Centric commitment, and the difference will become visible over the next several years as the two architectural choices produce different outcomes. Readers, partners, and observers benefit from knowing which posture a given programme is operating under.

Second, the published Opticeutical Standard and the IP portfolio it supports are human-authored work. The standard reflects human judgement about what discipline matters in the category. The IP claims reflect human invention. The institutional reference at endogenicpharmacology.com is human scholarship. These are not AI outputs that humans audited. They are human work that AI accelerated. The distinction matters for how the reader should treat the material.

Third, the broader AI-in-aging wave is going to produce, over the next several years, a substantial volume of content claiming AI-driven longevity insights of one kind or another. Most of it will be derivative. Some of it will be genuinely valuable. The reader's job in distinguishing the two becomes easier when the programmes producing the material have stated their design commitments publicly. The Atumnus design commitment is HI-Centric. It is stated here. Programmes operating under different commitments are welcome to state theirs.

What this is not an argument against

This is not an argument against AI in life sciences. We use AI extensively. It is, when correctly positioned, one of the most productive tools available to a small institutional research programme operating in a category that requires synthesis across very large literatures. The argument is about the positioning, not the tool. The full institutional treatment of HI-Centric-AI as a research discipline is documented at hi-centric-ai.com. The treatment of how this discipline informs the bioregulator IP work specifically is documented at endogenicpharmacology.com.