Design
1,000+
Candidates per run
Models like RFdiffusion, ESM, AlphaFold, and Boltz generate binders faster than traditional assays can test them.
Ground truth for agentic biology.
Agents design candidates. Models predict interactions. Measured binding at scale is still the missing step. Lagomics is building pooled, barcoded, sequencing-based validation to return a quantitative interaction matrix.
Design moved faster than measurement
The gap between generating candidates and proving which ones bind is where agentic biology still breaks down.
Prediction
Sparse
Interaction training data
Models that predict binding still lack dense, quantitative matrices to learn from or benchmark against.
Ground truth
Missing
The closed loop
Without measured hits and misses at library scale, design and prediction pipelines cannot validate or improve.
What we're building
One experiment. Every candidate. A real readout for each pair.
01
Candidate library in
Sequences and targets. The interaction question defined upfront.
02
One pooled experiment
Barcoded, cell-free, all pairs in one tube.
03
Interaction matrix out
Ranked binding, off-targets, structured data for teams and models.
Who this is for
Built for the teams generating more candidates than any bench can keep up with.
AI protein & binder design
Teams generating candidates with RFdiffusion, ESM, AlphaFold, Boltz, or in-house models. Validate the full output at the scale it is produced.
Peptide discovery
Teams screening peptide libraries against a target. Test the whole set at once, not just the top few that fit an assay budget.
Interaction-prediction models
Teams training or benchmarking models on binding. Generate the dense, quantitative ground truth that sparse public data cannot provide.
Activate Fellowship · MIT Engine · NVIDIA · Based at MBC Biolabs
Building this with the teams who need it
Lagomics would love to hear about the validation problems worth solving. A conversation is the best place to start.