Design the binder. Don't just screen for it.
Sanghyun is building an agentic discovery and design engine for RNA and DNA aptamers and oligonucleotide binders. Scientific agents and generative sequence models design, optimize and rank candidates before anything is synthesized, then learn from every result that comes back from the bench.
possible sequences in a 40-nucleotide random region. A physical SELEX library samples roughly 1014–1015 of them.
of selection and amplification in a typical SELEX campaign, often taking weeks to months before a single lead is characterized.
for a typical aptamer, roughly a fifteenth the mass of an antibody. Smaller binders reach tissue that larger ones do not.
SELEX finds what survives the pool. Not what is best.
Systematic Evolution of Ligands by Exponential Enrichment is how nearly every known aptamer was found. It is also slow, low-throughput and blind to most of sequence space.
- Trapped in local optima. Each round enriches what is already in the pool, so campaigns converge on the first adequate family rather than the best possible one.
- Amplification bias. PCR favors sequences that copy well, and parasitic sequences can outcompete true binders.
- Developability comes last. Stability, nuclease resistance and synthesis cost are checked after selection, when changing the sequence often breaks binding.
- Little is learned from failure. Non-binders are discarded, so each new campaign starts almost from scratch.
An engine that reasons, generates and learns in one loop
We pair specialized generative models with agentic scientific reasoning, and close the loop with high-throughput binding data from our wet-lab collaborators.
Agentic hypothesis & target modeling
A reasoning agent reads the literature on a target, predicts binding pockets and sets design constraints before any sequence is generated.
How it works 02Hybrid ML design loop
Generative sequence models propose candidates. Structure prediction, folding thermodynamics and docking score them against the constraints.
How it works 03Closed-loop wet-lab validation
SPR, BLI and NGS-SELEX readouts flow back to the agents, which explain the misses, update structure–activity hypotheses and tune the next cycle.
How it worksEvery assay makes the next design better
Design, synthesis, measurement and analysis run as one instrumented cycle rather than four disconnected projects.
Target prospects, not a pipeline
We are an early-stage company. Nothing below is a result. These are the target classes we are building the engine against first, chosen because a small synthetic binder has a real structural argument and because the biology is measurable.
Secreted growth factors
Local delivery to a confined compartment such as the eye, where a small binder does not need to survive long systemic circulation and dose is measured in micrograms.
Coagulation proteases
The setting that uses the modality's distinguishing feature: a binder whose effect can be reversed on demand by a complementary strand designed alongside it.
Cell-surface receptors for delivery
Here the binder is the address label rather than the drug. Internalization efficiency and tissue penetration matter more than headline affinity.
Structured viral RNA elements
Binding a fold rather than a sequence, which is harder for a virus to escape by silent mutation, and the hardest test of our structural reasoning.
Three things changed at once
Nucleic acid models got good
Pretrained RNA language models and structure predictors now give useful representations of folding and function, not just sequence statistics.
Agents can run a scientific workflow
Reasoning agents can read a target's literature, call folding and docking tools, interpret the output and decide what to try next, with a scientist reviewing each decision.
Readouts became dense
High-throughput SPR and BLI plus deep sequencing turn one experiment into thousands of labeled examples instead of a handful.
Notes from the design loop
Learning from the misses: post-assay error analysis in a closed design loop
A candidate that fails to bind is a labeled example. Here is how we extract the signal from it.
Developability first: the constraints we set before generating a single aptamer
Nuclease resistance, synthesis cost and immune motifs belong in the objective, not in a later rescue campaign.
Why SELEX gets stuck: enrichment is not optimization
Rounds of selection reward replication fitness as much as affinity. That distinction explains a lot of failed campaigns.
Have a target that antibodies struggle with?
We take on a small number of collaborations each year, from single-target design campaigns to multi-program platform partnerships.