It is possible to classify images simply by analysing the way that trillions of DNA molecules in a test tube connect together into different shapes.
Conventional computers must be fully assembled before they can run any computations. Inspired by biological systems 鈥� which must operate while they are still building themselves 鈥� at Maynooth University in Ireland and his colleagues wanted to see whether a computer that starts as many disconnected components can solve problems by assembling itself in a specific way.
鈥淏iology with its genetic program is the premier example of molecular computation where the point of the computation is to create something. We wanted to explore that angle of computation,鈥� says at the California Institute of Technology who worked on the project.
Advertisement
The researchers used 鈥淒NA tiles鈥� to make their computer. These are DNA molecules folded into rectangular shapes that can attach to one another and produce larger 2D structures. In this case, there were about 900 different types of DNA tile, and they were designed so that they would form large numbers of three different structures 鈥� a tiny letter 鈥淗鈥�, a letter 鈥淎鈥� or a letter 鈥淢鈥�.
Next, the researchers used these DNA tiles to solve a pattern recognition problem: classify 18 greyscale images into one of three categories represented by the three letters H, A and M. For instance, the M category contained an image of the Nobel laureate May-Britt Moser, a magnolia flower and a mockingbird.
Crucially, each image contained 900 pixels, each of which could be linked to one of the 900 DNA tiles. Constantine says this let the team generate a specific test tube mix of DNA tiles for each image in such a way that an M image produced a final mixture rich in M letters, while an H image produced a final mix with plenty of H letters and an A image produced mostly A letters in the final mix. In this way, the self-assembling tiles converted a pattern of concentrations into a final answer. In essence, the system behaved like a neural network that might classify an image as a cat or a dog based on its pixel patterns.
Similar ideas have been theorised before, but the researchers showed that they really can work in the lab, says at the University of Arkansas. The new experiment could provide insight into what is happening when cells self-assemble in biology and open the door for building computational systems that can take unusual inputs like molecular concentrations, he says.
at the University of Waterloo in Canada says this is an 鈥渋ngenious way鈥� of using DNA tile self-assembly, and that it shows how computation does not have to be synonymous with what we are used to seeing in electronic computers. This work challenges us to understand how information processing can happen in more natural and physical systems, she says.
鈥淲e鈥檙e hoping that people who don鈥檛 care about DNA at all鈥� will be convinced that this is a new way to do computation,鈥� says at the University of Chicago who was part of the team.
Nature
