2026-09-17
A scaffolded DNA computer runs 10 equilibrium programs, including 25-bit addition; N=4 typical-anneal yield is ~95%, short runs finish in a minute, BitCopy is reused 25 times.
Digital computers, like cells, stay out of equilibrium and spend energy on proofreading, redundancy, and tight timing. Theory has long said a computation can be the relaxation of a physical system to a thermodynamically favoured state, with the output sitting at the energy minimum. Energy-based models and simulated annealing use that idea, then run it on silicon that is itself far from equilibrium, at large energy cost. The missing piece is a programmable medium whose energy landscape can be engineered.
Most DNA computers to date have been non-equilibrium machines: algorithmic self-assembly, strand-displacement circuits, molecular robots. They work, and they leak, nucleate off-pathway, demand multi-step handling, and hate temperature drift. The Maynooth University group set out to show that the correct output can be the favoured equilibrium, with a protocol close to a single-pot anneal.
A Scaffolded DNA Computer is a one-dimensional scaffold with N unique binding domains. Compute tiles carry program and data bits on left and right, here ℓ=3 bits per side. A program plus an input is a subset of tiles mixed with the scaffold. A tile binds its bottom domain to a colour-matched scaffold site, then matches neighbours on the compute domains. A colour mismatch is an algorithmic error with an enthalpic penalty, later corrected by replacement. Position 1 has a single anchor tile, so the target configuration is unique.
Any finite-state machine compiles to a 1D SDC. Tiles are typically at 10× the scaffold concentration, so every site gets a tile. Domain strengths are kept weak enough for reversible binding and to limit off-scaffold junk. Readout is quenched fluorescence: low for 0, high for 1. The standard protocol is a 3-hour single-pot anneal from 80 °C to 20 °C at 100 nM scaffold, then a 45-minute hold. Short scaffolds were also run with super-fast anneals on the order of a minute. To reuse a sample, new input strands plus a blocker of the old input are added and the mix is annealed again.
Ten programs, more than 700 runs, binary inputs up to 50 bits, 100 bits of compute on the largest addition, plus base-3 strings.
On N=4 typical anneals, completion levels were linearly separable from the off-target. 4-bit Addition averaged 96.7% yield (s.d. 0.027), with the worst output bit at 91.6%. Across ten N=4 programs the mean was 95.3% (s.d. 0.035). Sites B and C averaged 98.1%; site D dropped to 94.4%. Raising competitive complexity from 2 to 8 did not hurt the 3-hour yield much (95.1% at complexity 8). A 20-hour post-anneal hold showed neither leak nor slow completion. One plate, dried out after 1.5 years, still ran MultiplyBy3 and Parity after adding water.
Super-fast anneals from 80 °C to 55 °C in under a minute still separated 0 from 1. Addition averaged 82.4% yield at 1 minute; all programs averaged 81.2%. The paper calls this the fastest non-trivial DNA programs in the literature, and also the speed floor of their lab gear.
Reuse: BitCopy 25 times, Addition 9 times, Counter 24 times, about 12 minutes per cycle. Counter kept a clear output across 24 rounds despite eight competing tiles per non-anchor site, with normalized fluorescence falling about 0.094 every eight repeats.
Scale-up to 25 sites used a biologically derived M13 fragment, scaff-624, at 10 nM, with almost all strands unpurified. 25-bit Addition uses sink / non-sink logic to swallow carries and cut error paths. Random 25-bit pairs have mean longest non-sink run M=4.0; M=4 finishes in about an hour, an M=5 example in Fig. 6a completed well within 14 hours, and M≤5 covers more than 84% of 25-bit pairs. IsoenergeticBitCopy rearranges compute domains so configurations with a fixed number of mismatches sit on a flat energy plateau: 71% yield at 20 sites, 59% at 25, both on a 14-hour anneal.
| Setting | Scale | Time | Yield |
| N=4 typical anneal | 10 programs | 3 h | 95.3% |
| N=4 Addition | 4 output bits | 3 h | 96.7% (worst bit 91.6%) |
| N=4 super-fast | all programs | 1 min | 81.2% |
| IsoenergeticBitCopy | 20 / 25 sites | 14 h | 71% / 59% |
For people who work on energy-based models, simulated annealing, or reversible computing, this is a wet-lab demonstration that "output equals ground state" can be more than a slogan: a single-pot anneal, minute-scale runs on short instances, dozens of input swaps in the same tube, no purification between rounds. It is not a substrate for training a network. The computing model is the point. In DNA nanostructures and archival DNA storage, designing the free-energy landscape may be cheaper than bolting on another proofreading layer.
N=4 is already solid. N=25 runs, at about 60% yield and tens of hours.
Larger scaffolds need slower anneals. M13 domains are uneven in binding energy, about 6.6 kb of leftover single strand comes along, and the tenfold drop in concentration hurts signal. 59% yield at 25-site BitCopy is not a scaling story yet. Super-fast yield falls to the low eighties; linear separability is not a zero bit-error rate.
25-bit Addition does use sink logic to suppress errors, which sits slightly against the claim that no explicit error correction is needed. The ternary digit-2 channel was a serendipitous temperature-dependent quench, characterized after the fact. The N² time bound from the bind/replace model was never enforced in the lab; the anneal path is still a black box. Heating and cooling costs were not compared seriously with silicon.