Why nucleic acid binders, and why they have underdelivered
Aptamers have most of what you want in a targeting molecule. We think the reason so few have reached patients is not the modality itself. It is the discovery process.
Chemically synthetic
Made on a solid-phase synthesizer, not in cells. Batch-to-batch identity is a chemistry problem rather than a biology problem, and there is no cell line to maintain.
Small and non-immunogenic
At roughly 8 to 15 kilodaltons an aptamer is an order of magnitude smaller than an antibody, penetrates tissue better and does not typically raise an antibody response of its own.
Reversible by design
A complementary oligonucleotide can act as a dedicated antidote, which is difficult to arrange for a protein therapeutic and valuable in acute settings.
What actually goes wrong in a selection campaign
Each of these is well documented in the aptamer literature. Together they explain why campaigns stall.
Coverage is vanishingly small
A 40-nucleotide random region spans about 1024 sequences. A generous starting library contains 1014 to 1015 distinct molecules, so the pool samples roughly one in a billion of the space. Selection can only improve on what was present at the start.
Amplification selects for replication, not affinity
Every round includes a polymerase step. Sequences that amplify efficiently gain representation regardless of how well they bind, and parasitic sequences with no affinity at all can dominate a late-round pool.
Enrichment is a noisy label
Read counts from sequencing a selection pool reflect binding, amplification efficiency, partitioning stringency and sampling depth at once. Treating enrichment as a proxy for affinity is a useful approximation that fails exactly where it matters, among the top candidates.
Specificity is bolted on
Counter-selection against related proteins is typically introduced after a binder appears. By then the lead family is fixed, and the structural features responsible for cross-reactivity may be the same ones responsible for binding.
Stability conflicts with the hit
Unmodified RNA is degraded within minutes in serum. Adding resistance afterwards, by substituting sugars or backbone chemistry, changes the fold that produced the affinity. Rescue campaigns are common and often unsuccessful.
Nothing carries forward
Once a campaign ends, the discarded pool, the failed modifications and the assay context usually leave with it. The next target starts with no accumulated model of what binds what.
Structure is the part worth getting right
An aptamer binds through a three-dimensional shape that its sequence only implies. Two sequences with eighty percent identity can fold into unrelated structures, and one of them can be a potent binder while the other is inert.
So we score the fold, not just the sequence. For every candidate we evaluate the minimum free energy structure, the suboptimal ensemble around it, and how dominant the intended motif is within that ensemble. A candidate whose target fold is merely the most likely of many is weaker than one whose fold is strongly preferred, even at equal predicted affinity.
We also check what else the sequence can do. Unintended G-quadruplexes, alternative stems that sequester the binding loop, and folds that only appear at low magnesium are all common failure modes that a sequence-only view misses entirely.
A candidate as the engine sees it
Illustrative example. Dot-bracket notation shows paired and unpaired positions.
Modification strategy is decided before generation
Each choice buys stability or half-life and costs something in synthesis, fold or affinity. The agent treats the modification budget as part of the design problem rather than a later fix.
| Strategy | What it addresses | Trade-off we model |
|---|---|---|
| 2′-fluoro / 2′-O-methyl pyrimidines | Nuclease resistance, with a large body of precedent | Alters sugar pucker and helix geometry, so the fold must be rescored, not assumed |
| Locked nucleic acid | Raises duplex stability at specific positions | Can over-rigidify a loop that needs conformational freedom to bind |
| Phosphorothioate backbone | Exonuclease resistance | Introduces stereochemistry and non-specific protein binding that confounds assays |
| 3′ inverted deoxythymidine cap | Blocks 3′ exonuclease attack cheaply | Minimal structural cost, so usually a default rather than a decision |
| Polyethylene glycol conjugation | Slows renal clearance of a molecule well below the filtration cutoff | Adds manufacturing complexity and can sterically block the binding face |
| Mirror-image L-nucleotides | Near-complete nuclease resistance | Incompatible with enzymatic selection, which is precisely why designed generation helps |
What we trust from an assay, and what we do not
A dissociation constant quoted without a method is close to meaningless. We record kinetics rather than endpoints, because association and dissociation rates separate candidates that an equilibrium number makes look identical. For most therapeutic uses a slow off-rate matters more than a headline affinity.
We also assume the first measurement is wrong in a specific, identifiable way. Mass transport limitation flattens apparent kinetics. Avidity from a multivalent surface inflates affinity. Nonspecific binding to the chip tracks with charge rather than structure. Refolding is sensitive to magnesium and to how the sample was heated and cooled. The agent checks for each of these signatures before it believes a result and updates a model on it.
Standing checks on every dataset
- Reference-subtracted and buffer-blanked traces, inspected rather than summarized
- Replicate agreement across independent surface preparations
- Immobilization density series to detect mass transport and avidity
- Counter-target measured in the same session and buffer
- Folding protocol recorded, including magnesium and cooling profile
- Scrambled and truncation controls included by design, not on request
Common questions
Does this replace SELEX entirely?
No, and we would be suspicious of anyone claiming it does. Selection is still the most direct way to interrogate an enormous pool against a real target. We use designed libraries to start from a far better place, and we run hybrid campaigns where a focused, designed library enters a short selection arm. The difference is that selection becomes a measurement step inside a design loop rather than the whole strategy.
How accurate is structure prediction for aptamers?
Secondary structure prediction for RNA is reliable enough to be useful and not reliable enough to be trusted alone, particularly for tertiary contacts, pseudoknots and metal-dependent folds. We treat predictions as rankings with uncertainty, keep structurally diverse candidates rather than only the top-scored one, and let the assay arbitrate. Where tools disagree, that disagreement is surfaced to a scientist instead of being resolved silently.
What targets suit this approach best?
Extracellular and secreted proteins, cell-surface receptors used for delivery, coagulation factors where reversibility matters, and structured RNA elements. Targets where a small synthetic binder has a genuine advantage over an antibody, and where a clean biochemical assay exists, are the strongest fit. Intracellular targets remain hard because delivery, not affinity, is the limiting step.
Who owns what comes out of a collaboration?
Partners own the sequences and data generated against their targets under the terms we agree. We retain the platform, the models and the methods. Improvements to the engine that come from a campaign are ours; the molecules are the partner's. We put this in writing before work starts.
How much of this is automated?
The data handling, scoring and analysis are automated. The decisions are not. A scientist approves the constraint specification before generation and the shortlist before synthesis. We built the audit trail precisely so those reviews can be substantive rather than a formality.
How far along are you?
Early. We are a research-stage company building the engine and testing it against our first target prospects. We have no clinical programmes, nothing has been evaluated in humans or animals, and we are not going to describe research directions as though they were results. The target prospects page sets out what we would need to demonstrate before calling any of it a programme.
Go deeper with the team
We are happy to walk through our folding and scoring approach in detail, including where it currently falls short.