How Owkin narrows the search space in aging research

The hardest problem is choosing what to test

In any therapeutic area, the number of plausible hypotheses a research team could pursue is effectively unlimited. Combinations of biology, patient subgroup, mechanism, and dataset multiply quickly, and a small team can only investigate a handful of them in a year. Deciding which ones are worth the lab time is, more than any single step, what determines whether a program produces useful science.

In our aging research program, internally, we call it the Aging POD, we have built a process to turn that decision into a structured weekly exercise. The program works like an AI-driven discovery factory: each week it surveys a large space of possible hypotheses, scores them on the same criteria, and surfaces a small number that are credible, feasible, and defensible enough to take further. The goal is a tighter shortlist — a small set of directions worth real lab time.

Angles and strategies: organising the search space

Every drug discovery program is asking two questions at once. Where to look — which biological context — and how to look — which research approach.

We refer to these as Angles and Strategies.

A Research Angle is a specific biological question. For example, what drives sarcopenia in a particular tissue and patient population. A Research Strategy is the method we use to investigate it — searching for a single novel target, exploring combination therapies, or mining clinical datasets for unexpected drug–disease links.

A traditional process moves through one Angle × one Strategy at a time. The matrix we have built runs many Angles against many Strategies in parallel, in the same week. Today that is around 100 aging angles across 7 strategies, with plans to grow both.

This way of working is operationalized through campaign mode in K Pro, a capability our biomedical team is already using in a dedicated instance of the product, with a broader rollout planned. Campaign mode lets a researcher run a biological angle against multiple research strategies in parallel, with the same scoring framework applied across every combination.

Three components

Three pieces of infrastructure make the matrix workable end-to-end.

The first is a Virtual KOL Board: a multi-agent system that simulates expert debate across disciplines — molecular and cellular biology, clinical longevity, pharmacology, regulatory — and stress-tests each angle before it consumes compute and human power. Weak hypotheses are filtered out early.

The second is the Aging Data Hunt, a deterministic pipeline that sources, filters, and quality-scores multi-omic datasets from GEO, CellXGene, and our internal cohorts. The same query run twice returns the same scored shortlist; reproducibility is a hard requirement.

The third is Lab Feasibility Scoring, a module calibrated against Owkin's lab catalog that grades each candidate on whether we can actually test it: reachable cell lines, available readouts, and realistic timelines. Ideas that no available lab can validate are flagged before they reach the selection committee.

The weekly cadence

The process runs over five days.

On Monday and Tuesday, scientists refine the week's biological questions and tighten each hypothesis.

On Wednesday, the matrix runs: computational pipelines execute every Angle × Strategy combination, with in-silico QC on each output.

On Thursday, scientists shift from data search to evaluation, examining the strongest leads in detail.

On Friday, a committee selects the candidates that will go forward to IP review and lab validation.

The following Monday, the process begins again with a new set of angles.

What is changing as the program scales

As the matrix grows, the limiting factor stops being whether we can generate novel hypotheses and starts being whether the ones we surface are defensible and deliverable. Two changes are underway.

An IP and regulatory layer scores patentability and freedom-to-operate during the discovery week, rather than after candidate selection. To clear Friday's committee, a target now needs to be both novel and defensible.

Closer integration with the lab ensures that each selected candidate has a clear, near-term path into validation, either internally or with an external partner.

A different unit of output

Both modes still rely on expert scientists at the centre. What changes is how those scientists spend their week: less time pulling and cleaning data, more time choosing between options the matrix has already scored.

Aging research is still mapping its own territory. Many of the central biological questions remain open — multiple plausible drivers of cellular aging, multiple candidate intervention points, limited consensus on which combinations are most worth pursuing.

In an open field like this, coverage matters as much as speed. The system can autonomously explore combinations that no single research team would have the bandwidth to walk through manually, and surface them in a scored, comparable form for scientists to evaluate.

The set of directions our researchers get to choose from grows. The people choosing stay the same.

Where this goes next

The aging program is the first place we have run this process at scale, and the same architecture is being applied to other therapeutic areas across Owkin. For partners, the practical implication is that the unit of work we can offer is no longer a single hypothesis after a long project, but a structured shortlist of viable directions, with feasibility and IP context already considered.

That is the change we think is most useful to talk about at this stage, not the individual results yet, but the way the search itself is organised.

Once campaign mode rolls out more broadly in K Pro, the same workflow will be available to research teams running their own programs.