Redesigning StemRIM's R&D around AI — not bolting it on top. How the bench, the Dry team, and discovery itself change.
Three mutually-reinforcing moves deliver it — and one precondition makes all three possible. We ask for three decisions to start this quarter.
Generative AI proposes peptides; a model ranks them before synthesis; the bench validates only the winners; every result trains the next round.
One Discovery-Loop pod measured in validated candidates per bench-week — not experiments run. The bench proves ideas; it no longer originates them all.
Turn the ~400-patient rare-disease ceiling into an enriched, approvable population — a path to approval the old playbook can't reach.
Precondition: a proprietary data lake that captures every experiment by default. Moves 1, 2 and 3 all sit on it — so we build it first, this quarter.
Paul David's lesson (1890s–1920s): factories that swapped a steam engine for an electric one and changed nothing got almost no gain. The payoff came only when each machine got its own motor — and the floor was rebuilt around the flow of work.
New power, old layout. Every machine chained to one central shaft. Output barely moved.
Machines freed from the shaft. The factory rearranged around the work. Then productivity jumped.
Adding AI on top of our bench-first process is a faster steam engine. The payoff comes when we redesign the work itself.
It speeds up a step that already exists.
Useful, cheap, worth doing — but it is not the transformation. Budget ~20% of effort here.
It changes the workflow, who decides, or what flows into the data.
This is where the gains live. Budget ~80% of effort here.
Every slide that follows is labeled with this test — that's the horizontal logic of this deck.
How many high-quality, well-validated shots on goal can we generate before the cash runs? That is the question every move below answers.
Our out-licensing model (upfront + milestones + royalties; redasemtide→Shionogi) does not change. AI multiplies the assets that feed it.
We have a Dry (bioinformatics) team beside the Wet (experimental) team — but Dry is a service downstream of Wet, analyzing data and building apps after the bench decides what to make. That is the electric motor wired onto the old shaft.
scientist proposes
synthesize & assay
disease models
sequencing, after the fact
out-license
Serial and human-paced. Each run informs that run, then sits in folders. Faster instruments make each box quicker — the shape never changes.
In one sentence: the wet lab is the origin of every idea, and computation is tacked on at the end to report.
The steam-engine column isn't wrong — we ship it in Phase 0 to free up your hours. We just don't call it transformation.
Computation proposes. The bench validates. Every result trains the next proposal.
Why we can do this: TRIM3/4/5 are designed peptides with a measurable activity readout — an ideal design–assay loop. Most biotechs would kill for a target this clean.
Bench decides what to make → runs it → Dry analyzes and reports. Two departments, one direction.
Your day: run my experiment, then analyze it.
Wet + Dry fuse into a single team with shared OKRs, measured in validated candidates per bench-week — not experiments run.
Your day: review the model's ranked proposals, design the experiment that validates them, feed results back.
The bench stops being where ideas are born and becomes the high-value place they are proven.
The fear is real: "the machine replaces the founder's intuition." It does the opposite. The loop's objective function is Tamai-sensei's biology — the mechanism defines what "good" means. The model searches a space no human team can; the CSO adjudicates a ranked queue instead of originating every idea by hand.
Every hypothesis starts from one scientist's insight, one at a time.
The mechanism guides AI to propose thousands; the CSO judges the best.
Blind test: the model proposes peptides; the CSO ranks AI vs human designs without knowing which is which.
An amplifier for founder intuition — not a replacement.
Dystrophic EB has only ~400 patients nationwide. A large Ph III is impossible. Small-n is treated as a statistical ceiling.
We already see it: in EB, 7 of 9 improved, 4 markedly — but we don't model who responds and why.
A multi-omics + clinical responder-signature model predicts who responds — turning n≈400 into an enriched, approvable population.
Org change: a small translational data squad owns the biomarker as a first-class, milestone-bearing deliverable — itself licensable.
Precision medicine as the path to approval where the old playbook has none.
Every assay, sequencing run, animal study and key decision flows into one structured, versioned lake — not a sprawl of shared folders. "Log it to the lake" becomes part of every protocol, enforced like GLP.
Index the shared drives, patents and R&D-meeting recordings into a system you can ask questions of — plus always-on IP / freedom-to-operate scans on the peptide platform.
Our 20 years of screening data is the moat a generic model can't have — but only once it's unified. This is what we start this quarter.
Moves 1, 2 and 3 all sit on the data lake. Build it first, or everything stalls. This is the single most important line in the plan.
Thin labeled data hobbles generative design → start with active learning, graduate to generation as the lake fills.
Without real workflow authority, the loop slides back to Dry-as-service → the org change must lead the tooling.
AI-designed trials depend on PMDA & academic PIs → apply Stage-4 redesign to assets we still own (TRIM3/4/5), not Shionogi's.
Irreversible foundation first; the exciting parts second. A 69-person factory can be redesigned in a year.
| Risk | Cheap test — run it first |
|---|---|
| Cold start — not enough labeled data | ~2 weeks. Backtest: can a model trained on past data re-rank known hits to the top? |
| Org resistance to the inversion | One quarter. Run one program as a fused pod; measure candidates/bench-week vs a control. |
| "AI can't capture the founder's intuition" | A blind test. CSO ranks AI vs human peptide designs, unlabeled. |
| Data lake becomes an IT boondoggle | 30-day mandate. One assay captured end-to-end — no 2-year platform. |
| PMDA / PIs reject AI-designed trials | One meeting. Informal PMDA pre-consult before any build. |
The data lake + knowledge brain. The precondition for everything.
Fuse Wet+Dry on a single TRIM3/4 program. Run the blind test.
Workflow authority, reporting to the CEO, peer to the CSO.
We keep exactly what StemRIM is. We change how the factory is laid out — for more, better-validated assets per yen burned.