1 Campaign optimization
Before building an autonomous lab you must decide on a campaign.
Self-driving labs excel at accelerating biological discovery—but only if human scientists give them the right search space. When the answer lies within narrow, well-defined constraints, AI can rapidly iterate through experiments to find it (the experiment loop).
For complex disorders like Alzheimer’s, there is no obvious, narrow search space. Choosing a direction—whether targeting amyloid plaques, tau vaccines, or astrocyte aging—requires completely different biological substrates, instruments, and interventions. Deciding which physical lab setup and search space to build in the first place is the campaign loop.
“If I want an innovative research program, even from frontier AI models I can only get AI slop or at best things that have already been tried before.”
Figure 1
The campaign loop
Optimize the campaign, not just the experiments inside it.
The campaign loop is used to improve the research setup itself.
Example campaigns
Intervention
Anti-amyloid antibody
Measurement
Plasma p-tau217
Substrate
iPSC cortical neurons
Impact index
82
2 Results
Results
For the campaign loop problem, our Campaign Engine—trained for 2 hours on a consumer GPU—performs better than frontier LLMs.
Figure 2
Performance in designing impactful research campaigns
| System | Score |
|---|---|
| Campaign Engine | 12 of 20 |
| Claude Fable 5.2 | 10 of 20 |
| Claude Opus 5.5 | 7 of 20 |
| Gemini Argon | 7 of 20 |
| GPT 6 Astra | 6 of 20 |
3 Previous innovation
Previous innovation
Before developing Campaign Engine, which we want to scale as a startup venture, we started as a nonprofit research organization, seizing the best opportunities we saw for using AI to automate scientific discovery in biomedicine. That work produced the projects below.
Applied AI
SUMMIT
The world’s first self-updating, multiscale platform for biological simulations, unifying thousands of causal connections from the scientific literature in minutes rather than weeks.
Fundamental AI
NDEs for metabolic simulations
Neural differential equations that overcome the limits of traditional metabolic simulations, with over 90% lower error and 1000× faster inference than existing methods.
Contact
Let’s build the campaign loop.
For investors, scientific collaborators, and prospective partners who want to get involved in using the Campaign Engine to automate therapy discovery.