Our Technology
Our cutting-edge hardware captures hidden behavioral features with unprecedented detail, while our AI platform classifies and models behavioral phenotypes at scale.
The Platform
The Virtual Labs platform is three layers, and each product is one of them. The Bridge Layer — the PolarEye camera — carries the assays labs already run across the bridge from legacy, hand-scored protocols into one standardized, poolable data format. The Discovery Layer — a new instrument class we unveil at SfN 2026 — adds the signals those assays never measured. The Substrate — MouseOS — is the analysis platform and unified schema everything streams into: the data foundation on which virtual control groups, and eventually virtual animals, will be built.
Inside the platform, every video moves through three processing stages — capture, classification, and pharmacological modeling — to produce a complete, reproducible picture of how a compound affects the CNS. Two vocabularies, on purpose: the products are the platform's layers — Bridge, Discovery, Substrate — while every video moves through the pipeline's stages — capture, classification, modeling.
Stage 01 — Capture
Our capture layer uses multi-camera arrays and infrared illumination to track subjects with millimeter precision. Full-body pose estimation — 85 anatomical keypoints spanning body, head, tail, and face, at 95 frames per second — extracts kinematic data that is invisible to human observers.
Every joint angle, velocity vector, and postural transition is encoded as a time-series feature — and dedicated facial keypoints support grimace-scale expression analysis and facial recognition for individual-subject identification alongside the kinematics. In practice: no ear notching, no tail marks, no RFID implants — and minimal handling stress. Observer bias is eliminated by design.
The Environment on the Record
Rodent behavior shifts with room brightness, background noise, ultrasonic sound, temperature, pressure, and humidity — variables that are almost never measured, and become unexplained variance when they change. Every session on our platform ships with a synchronized environmental record that converts uncontrolled variables into measured covariates.
This is what makes the rest of the platform trustworthy. Virtual Control Groups are only valid when baseline context is matched — the metadata is how they're matched. And when a cohort behaves unexpectedly, the environmental record is where the answer usually lives: reproducibility, explained.
Stage 02 — Classification
A transformer-based classifier segments the continuous pose stream into discrete behavioral states — grooming, rearing, locomotion, freezing, and over 40 additional ethological categories — with frame-level precision.
We are training the classifier on a growing corpus of annotated behavioral data spanning the field's most widely used assay paradigms — built to generalize across strains, ages, and experimental conditions.
Stage 03 — Modeling
The behavioral output will feed a pharmacodynamic model calibrated against validated reference compounds — benzodiazepines, antidepressants, antipsychotics, and cognitive enhancers — across the major assay paradigms. We are building that calibration now, reference compound by reference compound.1
Novel compounds will be screened against this model to generate predicted dose–response signatures, identify high-probability candidates early, and sharply reduce the number of animals needed for subsequent in vivo confirmation studies. The model will grow more predictive with every experiment run on the platform.
Performance
Validation
Our validation standard is simple: no assay ships until it is benchmarked against published pharmacology — peer-reviewed literature spanning more than 30 years of rodent behavioral neuroscience. Deep learning-based behavioral classification has been shown to reach human expert accuracy across standard assay paradigms.2
Where This Leads
For decades, Structure-Activity Relationship (SAR) analysis has been the foundation of medicinal chemistry — relating the structural features of a molecule to its potency and selectivity at a defined target. The next frontier extends SAR into behavioral space.
The long-term scientific vision is direct: a molecule's structural features should predict its rodent behavioral phenotype, and that phenotype should predict human clinical outcomes. Drug discovery experiments conducted computationally — without synthesizing every candidate, without running every animal study — for the compound classes where the models are mature enough to support it.3
That future is not here. But it is precisely what this platform is building toward. Every experiment run on the Virtual Labs system — every compound profiled, every behavioral phenotype characterized — contributes to the training dataset that makes predictive behavioral SAR possible. The dataset that does not yet exist is the one we are creating.
The Roadmap
References