NRSC 4132 Blog

When I was a kid I spent a lot of time trying to get my mom and dad to play board games with me. Unfortunately, most of these games required three or more players, and while Mom was usually a willing participant, Dad never relished the idea of spending an hour playing Clue or Candlyland with me. I never really got the hint, and to this day my closet at home is stuffed with board games that were taken out of the box once, set up neatly for three players, and then packed up and stuffed back in the closet without a single card drawn or the dice rolled once. It seems like some researcher may have shared my difficulty and was unable to find a single person to take part in that classic 1980s board game Guess Who?. But, rather than admit defeat and break out the cards for a lonely game of solitaire this individual decided a better option would be to create a bio-molecular system to approximate a computational neural network that could be the guessing game partner. Well, I sincerely wish that scenario was the inspiration for the research that resulted in the recent Nature paper Neural network computation with DNA strand displacement cascades, but even if it wasn?t this is still one of the most creative and interesting takes on artificial intelligence I?ve seen.
Taking inspiration from the biomolecular circuits that guided the ?intelligent? behavior of individual cells, researchers at Caltech designed a system of four artificial neurons trained to recall four single stranded DNA sequences and can identify the most similar one when presented with incomplete patterns. Using a linear threshold function a single neuron was modeled. This function allows the ?neuron? to receive an imput that is then multiplied by a positive or negative weight and then based on meeting or exceeding a predetermined threshold the ?neuron? fires producing an output. The the DNA neurons made from 112 discrete strands were created and connected in a Hopfield associative memory. This network of 4 neurons makes a tiny biomolecular brain capable of some interesting tasks.
The CalTech researchers ?trained? the tiny brain to identify four scientists based on answers to four yes or no questions such as ?Is the scientist a mathematician?? The human player in this modified game of Guess Who? answers these questions with corresponding strands of DNA that go into the test tube with the artificial neurons. The network then communicates its guess via fluorescent signals or it indicates that not enough information was given or that there was a contradiction. The network came up with the correct answer in every trial.
This biomolecule based brain may prove to have many practical applications in several fields. Medically, it would be very beneficial to have cells capable of identifying and signaling the presence of certain molecules or processes involved in disease, and in chemical engineering complex compounds could be produced one molecule at a time using these networks. This is truly a novel approach to modeling the brain and one which may prove very useful in the future.