Wilby Labs designs DNA nanostructures and is building a simulator that predicts their shape, motion and assembly yield before a single strand is ordered.
Today you find out at the end. A DNA origami folds when dozens to a few hundred short staple strands pin a long scaffold strand into shape, over an anneal that can run for hours. If the design misfolds, the gel tells you afterwards. A good prediction moves that answer to the start.
The simulator is built in four layers. The fast ones answer most design questions in minutes. The slower ones check them, with oxDNA as the reference at the end.
Nearest-neighbor thermodynamics and crosstalk screens
Domain-level stochastic folding with loop-closure cooperativity
Base-pair elastic network and Brownian dynamics
oxDNA2 molecular dynamics, on a GPU where possible
It works at the level of binding domains1,2. Staples bind and unbind one domain at a time, and each bound domain makes the next loop easier to close. The report gives the assembly yield, the share of simulated anneals that end with every staple bound, along with where in the anneal the structure folds and which staples are the weak points.
rectangle_12h
Model output
From 80 °C to 65 °C at 1 minute per degree, then to 25 °C at 10 minutes per degree, in 12.5 mM magnesium.
s2s6s8s14A base-pair elastic network, built on the physics of CanDo3, relaxes the design in 3D and reports its twist, bend and flexibility. This sheet is laid out at 10.67 base pairs per turn, looser than the 10.5 that DNA prefers, so it winds up along its length. Deleting one base pair in every 64 roughly halves the twist.
FlatThe layout in the design file, before any mechanics.
179° per 100 nmPredicted for the sheet as drawn: 115° from one end to the other.
87° per 100 nmPredicted with one base pair deleted every 64.
Brownian dynamics of the elastic model shows how much each part of a structure moves. When a question needs single-nucleotide detail, the simulator writes a pre-relaxed configuration for oxDNA24 and drives the run. Starting relaxed avoids the large initial strains of a raw export: the largest bond is 1.9 oxDNA units, against 5.4 in scadnano’s own export.
Off-target matches on the scaffold and staple dimers are scored with the same nearest-neighbor energies5 as the intended binding, so the two can be compared directly. Hairpins are folded with the DNA parameters in ViennaRNA, and runs of four or more G are flagged from the sequence. The report names the strands involved and what to change.
Excerpt from the report for the test sheet. Free energies at 50 °C.
s2 · s6 · s8 · s14s49 ACGACGGCC → scaffold 713s24 .((((((((.........))))))))......s6 GGGG · s15 GGGGG · s27 GGGGGThe simulator is in active development. Each layer is checked against a reference: an exact result, a second implementation, a published measurement or oxDNA. The folding model has not yet been compared with a lab measurement. The open rows are the next work, starting with a measured annealing curve.
| Check | Compared with | Result |
|---|---|---|
| Duplex melting temperatures, four sequences in three buffers | An independent implementation of the same nearest-neighbor5 and salt-correction6 parameters (Biopython) | Agrees within 0.05 °C |
| Folding model, detailed balance | The exact Boltzmann distribution of each staple’s bound states | Agrees to one part in a million |
| Strand-displacement rate against toehold length | Measurements by Zhang and Winfree8, whose three-step model the simulator uses | Reproduces the steep rise for short toeholds and the plateau near 3 × 106 M−1 s−1 |
| Relaxed size of a test design | oxDNA2, same design, same starting structure | 61.8 × 26.9 × 6.6 nm, against 62.4 × 29.7 × 6.3 nm |
| oxDNA export | scadnano’s own oxDNA export | Same conventions, with the largest bond at 1.9 oxDNA units against 5.4 |
| Absolute folding temperature | A measured annealing curve | Open |
| Design-dependent yield changes | Published folding experiments by Dunn and colleagues2 | Open |
| Predicted yield | Gel yields of our own designs | Open |
| Duplex melting temperatures | Measured melting tables, rather than another implementation | Open |
We keep an annotated library of more than 60 papers, from DNA thermodynamics to molecular machines. We surveyed the tools the field already relies on, and the simulator’s architecture follows from that survey.
Design tools such as cadnano9, scadnano10, DAEDALUS, PERDIX, ATHENA and MagicDNA turn a target shape into strand routings and sequences. They don’t say whether the design will fold. Shape predictors such as CanDo3 and mrDNA11 give fast shape and flexibility, checked against cryo-EM, but not whether the structure assembles at all.
Nearest-neighbor tables, NUPACK12 and ViennaRNA handle duplexes and small complexes, without the 3D context that makes origami folding cooperative. oxDNA4 is the best-validated model at the level of single nucleotides, but folding a whole origami with it is orders of magnitude too slow. Domain-level research code1,2 predicted design-dependent yield changes that experiments then confirmed, but it was written for specific designs.
We haven’t found a tool that connects these, from a design file to a yield that is checked against the lab. That is the gap the simulator is built for. oxDNA stays the reference: we export to it rather than rebuild it.
DNA is a material you can program. Sequence decides what binds to what, so a long scaffold strand and a couple of hundred short staples can fold themselves into a chosen shape13. The same rules have already produced walkers14, a molecular assembly line15 and a rotary motor16.
Wilby Labs is working toward molecular assembly in the spirit of Drexler’s assemblers17, built from DNA parts in the tradition Seeman started18 rather than from diamond. The simulator comes first, so that every part is designed with a prediction in hand.
We’d like to hear from labs that fold DNA origami, from people who build design tools, and from anyone working toward molecular assembly. If you have folding data we could test our predictions against, we’d especially like to talk.