On September 16, 2026, Maynooth University in Ireland announced the "scaffolded DNA computer" (SDC), which performs operations such as addition through the binding of DNA molecules. The peer-reviewed paper, with Tristan Stérin and Abeer Eshra as co-first authors, was published online in Nature the same day. The research team designed the DNA so that the structure representing the correct answer is the most stable one, and tested more than 700 computations across 10 programs. However, the 30-second run time applies to a short computation, while the 100-bit scale comes from a separate, larger experiment. What was demonstrated, and under which conditions, changes the meaning of each result.
Designing DNA So the Correct Answer Is the Most Stable Structure
The SDC uses a long strand of DNA as a scaffold and lines up short "computation strands" along it. Each computation strand binds at a designated position on the scaffold and also links to neighboring strands through regions on its left and right. These neighbor-to-neighbor pairings carry the input digits and intermediate states of the computation.
In addition, for example, each position receives two input bits and a carry from the previous digit, and determines the information passed to the next digit. Choosing the DNA sequences is equivalent to building the rules of the computation into the system. In the paper, the team ran 4-bit additions such as "10 + 3 = 13" and read out each bit of the answer in separate experiments.
To make the computation proceed correctly, the researchers designed the system so that incorrect pairings between adjacent computation strands are energetically unfavorable in terms of binding energy. The idea is that, as strands bind and displace one another, the system moves toward the structure representing the correct answer. The target answer is mapped to a state of "thermodynamic equilibrium," in which the system settles under given conditions.
Professor Damien Woods, who led the research, explained how heating and cooling lead DNA to form structures in the university's announcement:
"Molecules interact to build a structure, and that structure is the answer."
This is not the first attempt to compute with DNA, nor the first to use equilibrium for computation. What the paper shows is that a scaffold-based design can be run experimentally as multiple programs, including addition, multiplication by three, and division by two. These include a process that checks whether the number of 1s in the input is odd or even, so the approach is not limited to building one specific structure.
The 30-Second Run and the 100-Bit Run Used Different Conditions
The experiments were conducted in 35 microliters of solution containing DNA. The buffer contained 12.5 millimoles per liter of magnesium ions, and the scaffold concentration in the short-scaffold system was 100 nanomoles per liter. The concentration of regular computation strands was 1,000 nanomoles per liter, ten times that of the scaffold, a formulation intended to make it easy for strands to bind at each scaffold position.
Outputs were distinguished as 0 or 1 by fluorescence intensity. Fluorescence was measured with a real-time PCR instrument capable of temperature control, and the measured curves were averaged over at least two replicates. In the series of experiments examining the 10 programs, seven plates were used including controls and replicates.
In the controls, only one type of strand was allowed to bind at each position, removing the competition that accompanies computation and simply assembling the target structure. By comparing against the fluorescence of these controls, the team evaluated how closely the computing samples approached the intended output. The yield in the paper therefore refers to the proportion of structures estimated from fluorescence, not a correct-answer rate indicating how many of the 700-plus computations were right.
Changing how the temperature is lowered makes short computations proceed faster. But short and large systems differ in both scaffold length and concentration, so run time alone cannot be lined up side by side to decide which is faster.
| Type of experiment | Scale and main conditions | Reported time |
|---|---|---|
| Standard short system | 4-position scaffold. Cooled from 80°C to 20°C | 3 hours of cooling, then 45 minutes held at 20°C |
| Short system with fast cooling | 4 positions or fewer. Cooled from 80°C to 55°C in under a minute | 0 and 1 distinguished in about 1 minute, in some cases 30 seconds |
| Large addition | Adding two 25-bit numbers on 25 positions. Scaffold concentration 10 nanomoles per liter | Good output within 14 hours for one example of a complex input |
The table organizes the addition, fast-run and scale-up sections of the paper by computation scale and temperature conditions. The university cited "10 + 3" as its 30-second example. This is not the total working time including DNA synthesis and sample preparation, nor a value showing computing speed that exceeds silicon computers.
The paper's "100 bits" is the sum of two 25-bit inputs, 25 carry bits and a 25-bit output. Adding up the breakdown in Figure 6a gives 25 + 25 + 25 + 25 = 100. This does not mean a single 100-bit integer was input, and there is no report that this largest scale was processed in 30 seconds.
In the large system, the team used a 624-base region of M13, a DNA of biological origin, for the computation. Such scaffolds are easy to obtain, but the binding strength varies from position to position, which is said to be a factor that slows the computation. Simply lengthening the scaffold did not allow the speed of the short system to be maintained as is.
Up to 25 Reuses Required Adding Input Strands
The experiment with up to 25 reuses was achieved with a bit-copy program that switches the input between 0 and 1. A different addition program was repeated 9 times, and a counter 24 times. Summarizing this as "25 different computations run in succession" obscures what was actually repeated.
For reuse, strands representing the new input and complementary strands that seal off the old input are added. The system is then heated and cooled again, moving it to the equilibrium corresponding to the new answer. In the bit-copy experiment, 0.2 microliters were added each time, and the system was cooled from 80°C to 48°C over 12 minutes. The time needed for the input-adding operation is not included in these 12 minutes.
The significance lies in being able to update the input and use the system repeatedly while keeping the molecular system used for computation. However, additional reagents and temperature operations are required. It does not mean a device has been completed that can run any number of times without adding new materials.
Energy Savings and Use Inside Cells Remain to Be Verified
The research team has in view uses such as DNA data storage and computationally controlling the formation of DNA structures. The university's announcement mentions the possibility that, in the future, this could lead to molecular systems that operate inside cells to detect signs of disease. All of these must be distinguished from uses actually realized in this experiment.
This study is not an observational study that infers causes from naturally found correlations; it is an experiment in which DNA composition, inputs and temperature were set and outputs measured. The subject is DNA in solution, and operation in humans, animals or cells was not tested. There is still a gap to be verified between being able to compute under specified conditions and the mechanism working inside the body.
Furthermore, thermodynamic calculations and simulations show a prediction that the correct-answer structure is favored under the designed conditions. Confirming the output experimentally does not mean every molecular reaction pathway leading to that answer was observed. The authors also consider that multiple pathways may be possible depending on temperature, and leave open the possibility that larger systems will require designing reaction rates.
On energy savings as well, the principle of mapping computation results to an equilibrium state must be considered separately from the power used by the whole apparatus. The experiments involve DNA preparation, heating and cooling, and fluorescence measurement. The paper does not give power consumption including these, or a reduction rate relative to silicon, so it is not at a stage where it can be called a "zero-power computer."
The paper's DOI is 10.1038/s41586-026-10996-5, and the authors have made the data and analysis code public. Replication was carried out within the same research team, but in the public materials we could check, no independent replication of the SDC by third parties was found. Stérin, Eshra and Woods have declared that they are inventors on a related pending patent.
Even as the computation scale grows, the challenge ahead is whether variation in binding can be suppressed and the target answer reached fast enough. To move toward the data processing on DNA that the authors envision, the advantages of this design will need to be confirmed across the full sequence of operations, including updating inputs and reading outputs, not just the computation itself.