The parameter search engine for the nervous system.
Averic is building a self-driving lab for neuromodulation. A researcher specifies the neural outcome they want and the constraints they must respect. Averic chooses the most informative stimulation experiment, runs it in a living model, measures the response, learns, and repeats, until it hands back a protocol validated in biology, not simulation.
The bottleneck is parameter search.
An estimated 3.4 billion people live with a neurological condition, the leading cause of ill health and disability worldwide (Lancet Neurology, 2024). Neuromodulation (steering neural activity with ultrasound, electricity, or light) is used or studied across epilepsy, Parkinson’s disease, depression, chronic pain, and paralysis. For many of these conditions it is one of the only paths that does not run through a new molecule.
Whether any of it works comes down to parameters. Target, frequency, intensity, pulse duration, duty cycle, timing: every modality opens a search space too large to test by hand, and every point in that space costs a living experiment. Interventional datasets are tiny, so there is no shortcut through published data. Labs tune by intuition and grid search, effective settings take months to find, and protocols are delayed, missed, or never discovered at all.
| Parameter | Coarse range | Levels |
|---|---|---|
| Carrier frequency | 0.2 to 3 MHz | 6 |
| Intensity | device-safe span | 5 |
| Pulse duration | 1 ms to 1 s | 5 |
| Duty cycle | 5 to 50% | 4 |
| Inter-pulse timing | 0.1 to 10 s | 5 |
| Anatomical target | candidate sites | 4 |
| = 12,000 CONFIGURATIONS · AT 20 MIN PER MANUAL TRIAL: ROUGHLY 5.5 MONTHS OF CONTINUOUS BENCH TIME | ||
A search engine with a lab in the loop.
Averic treats protocol discovery as a sequential learning problem. The machine learning is not a model trained on papers: it sits inside the experimental decision loop, chooses what the lab tests next, and learns from what the biology answers.
- 01
Define.
A lab specifies the outcome it wants, the model system, the stimulation modality, parameter bounds, and hard safety constraints.
- 02
Select.
An active-learning optimizer chooses the single most informative protocol to test next, trading expected performance against uncertainty.
- 03
Run and measure.
The protocol executes in a living model. Computer vision and recordings score the biological response automatically.
- 04
Learn.
The result updates the map of the intervention-response landscape, and the loop repeats until the map is good enough to act on.
What comes back is not a dashboard. It is a validated protocol, the mapped response surface with uncertainty estimates, and the complete experimental record behind both. Every campaign the system runs is designed to make the next one cheaper.
Built on a system that already closed the loop.
Averic is not starting from a pitch deck. Over three years at the University of Minnesota, founder Evan Morris built a closed-loop platform, SILENCE, that watched CRISPR-engineered C. elegans for seizure-like activity and fired ultrasound, in real time, to interrupt it. An engineered whole-organism seizure model, live computer-vision detection, RF-amplified ultrasound stimulation, and automated intervention, integrated end to end by one person.
It is the run-and-measure core of the Averic loop, closed end to end. What Averic adds on top is the layer that decides which experiment to run next, and learns from every answer.
| Measured characteristic | Result |
|---|---|
| Seizure recurrence after intervention | 70.1% reduction |
| Overall seizure-like activity | ≈30% reduction |
| Time per experiment | ≈5 minutes |
| Marginal cost per experiment | under $2 |
- 2026
Regeneron ISEF: First Place, Cellular and Molecular Biology.
- 2026
Mary Sue Coleman Award for Life Science Innovation and Impact: a $10,000 Top Award.
- 2026
Dudley R. Herschbach SIYSS Award: representing ISEF during Nobel Week in Stockholm.
- 2026
First ISEF finalist to receive two Top Awards in the same year.
- 2025
Regeneron ISEF: Second Place, for earlier work on a low-cost optogenetic neural-plasticity tool.
The worm is the proving ground, not the ceiling.
C. elegans models a surprising range of neurological disorders, and a five-minute, sub-$2 experiment is one of the cheapest questions anyone can ask a living nervous system. That is where the loop learns to search. It is not where it stops.
- STAGE 01
Sonogenetic seizure control in C. elegans.
The first optimization campaigns, on the assay the founder already built. This is the stage under construction today.
- STAGE 02
More phenotypes, more modalities.
Optogenetic and electrical stimulation across additional C. elegans disease models, on the same optimization layer.
- STAGE 03
Higher-order models.
Zebrafish, organoids, and rodents, with partner labs connecting their own rigs to the loop.
- STAGE 04
Devices, then the clinic.
Preclinical protocol development for neural-device companies, and, much later and under full regulatory controls, patient-specific programming.
Every campaign feeds a dataset that does not exist anywhere today: controlled neural interventions linked to measured outcomes, across parameters, models, and modalities. Optimizers are replaceable. That dataset is not.
Stage 01 is what exists. Worm protocols do not transfer directly to humans, and Averic does not claim they do. What transfers is the search: knowing which mechanisms and protocol regions deserve the expensive experiments before anyone spends a year finding out.
Founder
Evan MorrisFounder
Evan is an 18-year-old studying neuroscience and finance at the University of Minnesota. He has spent three years building optogenetic and sonogenetic systems across UMN laboratories, and his current research spans rodent traumatic brain injury and a swine model of temporal-lobe epilepsy. He was the first Regeneron ISEF finalist to receive two of the fair’s Top Awards in a single year.