Teach the lab
what to ask next.

01 — CAMPAIGN OPTIMIZATION

Automate the experiment.
Then improve the campaign.

Automating therapy discovery is one of the most powerful ways we can imagine tackling human disease. Making it work requires intelligence at two levels: running the experiment and choosing the research direction.

Self-driving labs excel at searching within a space scientists have defined. They can vary conditions, measure the result, and select the next experiment. Their success depends on the promise of that search space.

For complex problems, choosing the space is often the harder task. In Alzheimer’s research, a campaign targeting amyloid plaques calls for different biological systems, interventions, and measurements than a campaign studying astrocyte ageing. The choice shapes both the science and the lab you need to build.

Before a lab can optimize an experiment,
someone has to choose the campaign.

In 20 interviews with people working in therapy discovery, we heard two recurring frustrations with AI-generated research ideas: familiar suggestions and impractical designs. The opportunity is to identify new, high-value directions and turn them into campaigns a lab can execute.

Intelligence Engine is our approach to campaign optimization. It maps the scientific frontier, prioritizes promising research combinations, and helps define the experiments, measurements, and biological systems an autonomous lab should explore next.

02 — TWO LEVELS OF LEARNING

Better experiments.
Better questions to begin with.

The experiment loop improves conditions within a campaign. The campaign loop improves the campaign itself: the interventions we try, the measurements we make, and the biological substrates we work with.

THE CAMPAIGN LOOPExplore the two levels

Experiment loopCampaign loop

Optimize the campaign, not just its parameters.

The outer loop uses evidence to improve the research setup itself. Inside it, an experiment loop runs interventions and returns measurements.

TRY DIFFERENT CAMPAIGNS

Aβ42 FibrillationAPOE4 Senescence

BIOLOGICAL SYSTEM

Purified monomeric Aβ42 peptide

INTERVENTION

Small-molecule β-sheet destabilizers

MEASUREMENT

Real-time ThT fluorescence

03 — BENCHMARK RESULTS

Benchmark on a
quantifiable question.

Improving campaign recommendations requires a way to evaluate them. Testing every proposed experiment in a wet lab would be prohibitively slow and expensive. We framed the challenge around a testable question: which new research combinations will scientists explore next?

We rank previously unconnected research concepts using the literature available at a historical cutoff, then check which connections appear in later therapy-discovery publications. This provides a scalable test of the engine’s ability to anticipate new research directions. Practical feasibility and therapeutic value are the next stage of validation.

Solving the Unoriginality Problem

Intelligence Engine anticipates connections that researchers go on to investigate. In our full experimental-type benchmark, the engine achieves 0.896 ROC AUC across 14.8 million candidate pairs, with 6 of its top 20 connections appearing in later publications.

We then compared Intelligence Engine with Claude Sonnet 5.5 and GLM-5.3 on a uniform sample from that same candidate population. Intelligence Engine led both models in this head-to-head ranking test.

ONE BENCHMARK · TWO VIEWSDownload results ↗

FULL BENCHMARK

0.896ROC AUC

6/20top-ranked links found later

14.8Mcandidate research connections

Experimental-type candidates · Literature through 2024 → outcomes in 2025

SAMPLED MODEL COMPARISON

The same 10,000 pairs.
Three different models.

A uniform sample of the full candidate set, scored by every model.

ROC AUCROC curve

Higher AUC means better ranking of connections that appear in the held-out literature. Dashed line: random ranking.

Benchmark methodology +

The full benchmark contains 14,775,304 experimental-type candidate pairs, including 17,435 later links. Intelligence Engine uses the graph-ranking method selected on validation: common neighbors.

The head-to-head panel is a uniform 10,000-pair sample from the same population, with 10 later links. All three models score identical pairs. Full-set and sample scores are separate estimates; the top-20 shortlist changes with the candidate population.

Head-to-head performance on identical sampled pairs

Model	ROC AUC	Average precision
Intelligence Engine · graph ranker	0.8303	0.04117
Claude Sonnet 5.5	0.7446	0.00248
GLM-5.3	0.6948	0.00445

The outcome is new abstract co-occurrence in the held-out, topic-bounded corpus. Model knowledge may include outcome-year papers. Approximate paired AUC-difference intervals are −0.236 to +0.064 for Claude minus the engine, and −0.308 to +0.037 for GLM minus the engine; they include zero and assume independent pairs.

04 — THE GRIFFIN LABS EDGE

A focused intelligence layer
that can keep pace with science.

Scientific literature changes every day. Our approach makes the representation of that knowledge explicit: concepts, connections, and the frontier between established work and new combinations.

A focused domain graph can be updated and re-ranked as evidence changes, without retraining a frontier language model. LLMs can work with that structured representation to turn promising directions into research designs. We are building a specialized layer that strengthens the wider AI-for-science workflow.

An advantage for a biological data factory

Building a new automated lab capability takes time. Anticipating promising experimental setups creates an opportunity to invest in the right instruments, biological substrates, and assays earlier—and produce data the next generation of biological AI will need.

That is our proposed business wedge: use campaign intelligence to guide a focused biological data operation, then expand toward therapy discovery and pharma co-development as the platform delivers results.

05 — BUILDING THE BUSINESS

From campaign intelligence
to autonomous discovery.

Our next milestone is practical feasibility. The engine identifies new research directions; expert biologists will help turn them into campaigns that can be run.

1. 01 / VALIDATE THE CAMPAIGNS

   Bring wet-lab expertise into the engine.

   Hire biologists to evaluate recommendations, refine proposed designs, and define the operational blueprint for our first focused autonomous lab.

2. 02 / BUILD THE DATA CAPABILITY

   Generate high-value biological data.

   Build the first lab capability around the strongest reviewed campaigns. Develop a data-generation business serving AI and biotechnology partners, with brain research as an initial application.

3. 03 / ADVANCE THERAPEUTIC DISCOVERY

   Expand into asset discovery and co-development.

   Use the lab’s evidence and campaign intelligence to pursue in-house therapeutic research and partnerships with pharma. The long-term ambition is a discovery platform that can transfer across diseases.