A Test Tube as a Processor: How DNA Molecules Learned to Solve Problems

The article traces the development of DNA computing. It begins with the restriction enzyme EcoRI, which cuts DNA at a specific six-nucleotide sequence, demonstrating molecular specificity. Leonard Adleman, a Turing Award winner and professor at the University of Southern California, realized in 1993 that enzymatic recognition and action constitute computation. In 1994, he published an experiment in Science solving a Hamiltonian path problem with a seven-city graph using DNA molecules: each city was assigned a random 20-nucleotide sequence, and road molecules were built to bridge them. Ligation and PCR filtered the correct paths, yielding the answer in about a week. Further work by Qi Ouyang (maximal clique, 1997) and Ravinderjit Braich (satisfiability with 20 variables, 2002) proved the approach but hit a wall: the number of possible paths grows factorially, so for 200 cities the required DNA mass would exceed Earth's mass by 10^328 times. Milan Stojanovic's MAYA automaton (2003) played tic-tac-toe using DNAzymes, but noise limited circuits to about ten gates. DNA storage, however, excelled: Yaniv Erlich and Dina Zielinski's 'DNA fountain' stored 2.14 megabytes in 2017, and Lee Organick's 2018 work achieved random access to files from a 200-megabyte pool. In biomedicine, strand displacement circuits enabled molecular neural networks; Lulu Qian's group in 2018 classified MNIST digits with a DNA-based network taking eight hours per computation. Despite not competing with silicon for speed, DNA offers extreme storage density and the ability to compute inside living cells.