Platform
The Loop
Five stages. One feedback cycle. The model improves at every step — from the first cell assay to the last animal study.
Generative Design
Proprietary generative stack produces binder candidates conditioned on antigen and proprietary developability levers. New sequences generated end-to-end. Thousands of candidates per cycle.
In Silico Pre-screen
An ML-guided ranking step prioritizes generative outputs for downstream evaluation. Lower-confidence sequences are deprioritized before any physical experiment.
Cells-on-Array Validation
Proprietary functional display platform. Each candidate receives an independent functional binding label against live receptor-expressing cells.
PK Scaffold Conjugation
Peptide hits are conjugated to modular PK scaffolds using site-specific chemistry. Half-life and exposure tuned independently of the binding domain.
In Vivo PK/PD
Peptide-scaffold conjugates and antibody-format hits enter rodent PK/PD studies. Plasma concentration, tissue distribution, and pharmacodynamic endpoints are measured and recorded.
All data retrains the model
Cell assay labels · conjugate characterization · in vivo PK/PD
What makes it different
The integration is the moat.
Generative Model — What it produces
Sequence and structure for a binder in a single forward pass — not a ranking of existing sequences, but generation of new ones. The model satisfies antigen complementarity and proprietary developability constraints simultaneously. Developable by construction.
Cells-on-Array — What it measures
A functional binding label per candidate per receptor. The specific format is proprietary. The output: per-sequence signal that discriminates binders from non-binders at throughput compatible with closed-loop retraining. Not bulk selection. Not NGS. Individual functional measurements.
PK Scaffold Conjugation — Why it matters
Cell binding and in vivo exposure are different problems. A peptide that binds its receptor in a well does not automatically have acceptable plasma half-life, tissue distribution, or target engagement in vivo. Peptide hits are conjugated to modular PK scaffolds to bridge the two. Antibody-format hits, with native half-life properties, route directly to in vivo studies. PK/PD data from both paths feeds back into the model alongside cell data.
The Model Retraining Loop
Cell assay labels, conjugate characterization data, and in vivo PK/PD profiles are structured into training examples and used to fine-tune the generative model adapter and the in silico ranking layer. The model accumulates target-specific knowledge. Hit rate on designed candidates improves measurably round over round.
Throughput
The display platform scales.
The biology does not change.
Increasing throughput is an instrumentation problem, not a chemistry problem. The same display format and live-cell readout runs at every scale tier. The model accumulates more training data per unit time as throughput increases.
Questions about the platform?
We are selective about early partnerships. If you have a target and want to understand how the loop applies to your program, reach out.
Get in Touch