A binder is not
a sensor.

Continuous biosensors report a change in shape, not a binding event. We design the aptamer switches that produce one — and we tell you what the design rests on.

GGTGGCAGGAGGACTATTTATTTGCTTTTCTCCTCCTGCC

published anti‑IL‑6 parent, 31 nt · assumed binding core 1–13 · appended competing tail

−2.0ΔΔG, kcal/mol+2.0

Too stable and it never switches. Too loose and it never folds. The usable band is about 4 kcal/mol wide — narrower than the folding model's own error bar, which is why a plate tiles this axis instead of ranking along it.

What it does

Type a biomarker. Get 96 wells a lab can order.

An agent searches the published literature for a parent aptamer, folds it, enumerates every switch variant, filters them, and lays out a plate — with the citation, the thresholds and the assumptions attached.

It greps the full text of papers for sequences rather than ranking them by topic, and keeps the words printed either side of each hit, so a candidate that binds the receptor rather than the ligand can be caught before it reaches a plate.

Library

every tail length, register, mismatch and linker — enumerated, not sampled

8,577

In window

ΔΔG within ±2 kcal/mol of balance

3,366

Pass filters

off‑target tail binding, self‑dimer, G‑runs, homopolymer, GC

642

On the plate

tiled across 8 energy bands · 8 wells kept for controls

The part most tools skip

We test our own assumption, and it usually fails.

Every energy above is computed against an assumed binding core — which stretch of the aptamer actually touches the target. For most published aptamers, no paper maps it. So we re-design the whole plate under each plausible core and ask whether the answer survives.

0 of 88

designs survived every core hypothesis on our IL‑6 plate. Two of four cores produced a usable plate at all, and they shared no candidates. The plate was an artefact of an assumption nobody can check.

The tempting response is to advise against ordering. But a wet lab has a budget and a synthesis slot; refusing to choose just hands the problem back. So we spend the plate on the uncertainty instead.

Single-core plate

88 test + 8 control

All wells designed under one assumed core. If the guess is wrong, every well fails together and the round teaches nothing.

Hedged plate

44 + 44 test + 8 shared control

Split across both surviving hypotheses, every well labelled, positions randomised against hypothesis. One synthesis run answers which switches work and which core was right.

Design → build → test → learn

The bench answers, and the next round moves.

Upload the plate reader's CSV and it is read before anything is designed. A hypothesis with no responsive well is eliminated; the window recentres on the measured optimum rather than the predicted one.

Scatter of measured signal change against designed ΔΔG, one point per well, coloured by core hypothesis. One hypothesis clusters at high signal near ΔΔG 0; the other sits flat along the baseline. — Every correlation is tested against 2,000 random shuffles of the same numbers before it is allowed to move a design. Above p = 0.05 the window stays where it is — recentring on noise would spend a second synthesis run confirming an artefact of the first. The data shown is a simulated example, labelled as such on the figure; no wet-lab results have been collected yet.

Limits, stated up front

What this does not do.

It cannot invent an aptamer. De novo aptamer design has no validated computational method. Without a published parent there is nothing honest to build on, and the agent says so rather than generating something plausible.
It cannot predict affinity. Apparent K_d is reported only when a paper measured one for that exact sequence. Usually none has, and every affinity-derived figure is omitted rather than estimated.
It cannot reach healthy baseline. IL‑6 circulates at pg/mL while aptamer K_d values are nanomolar, and switching costs further affinity on top. This sensor class reaches sepsis and cytokine-release range, not routine monitoring.
The binding core remains an assumption. We tried an AlphaFold3-class complex predictor to resolve it; a scrambled aptamer of identical composition scored higher than the real one, so we do not use it.