Computer scientists at Harvard University have come up with a way to convert algorithms that teach machines to learn into a form that would allow artificial intelligence to be programmed into complex chemical reactions.
The ultimate result could be “smart” drugs “programmed” to react differently depending on which of several probable situations they might encounter – without the need to use nano-scale electronics to carry the instructions.
“This kind of chemical-based AI will be necessary for constructing therapies that sense and adapt to their environment,” according to Ryan P. Adams, assistant professor of computer science at Harvard’s School of Engineering and Applied Sciences (SEAS), who co-wrote the paper explaining the technique. “The hope is to eventually have drugs that can specialize themselves to your personal chemistry and can diagnose or treat a range of pathologies.”
The full paper is available here as a PDF.
The techniques are part of a larger effort to program the behavior of molecules in manufacturing, decision-making and diagnostics, using both nano-scale electronics and the still-relatively-new study of bionanotechnology.
Adams and co-author Nils Napp, a fellow at Harvard’s Wyss Institute for Biologically inspired Engineering, created a tool that could convert algorithms that represent unknown quantities as a matrix of probable outcomes into a set of chemical reactions. Those reactions can’t be observed directly, but do follow the same steps an artificial-intelligence program would follow in specific situations.
Biological organisms are able to operate in chaotic situations by setting their own priorities and picking their way through the “noise” – aberrant or irrelevant feedback – created by competition or interference from other organisms, according to Napp. Computers can sift signal from noise faster and more completely, but only with instructions from humans on how to identify which is noise and which is not, and a process to follow that will help make those decisions simpler over time.
Among the state-of-the-art techniques proven to teach machines to learn are a set of algorithms that create graphical models of a set of probabilities, in order to make it easier to choose which is the most probable among all the available possibilities. The technique, which is common in search engines, fraud detection, and error correction for mobile phones, is referred to as “message passing inference on factor graphs.”
It is those algorithms, and their ability to decide which of many possibilities is most probable, that Adams and Napp have begun to convert into chemical reactions that can estimate how frequently each of many variables can appear without having seen many or made a count of them all.
While the mechanism is biochemical rather than electronic, molecular programming relies on computational abstractions using biomolecules and functions more common in computer science than bioscience, including logic gates, neural networks and linear systems.
Adding that capability to a drug or other chemical would allow it to gauge the biochemical composition of its new environment and tailor itself to operate within the conditions it finds, according to Adams and Napp.
“These algorithms allow today’s robots to make complex decisions and reliably use noisy sensors. It is really exciting to think about what these tools might be able to do for building better molecular machines,” Napp said.
“Probabilistic graphical models are particularly efficient tools for computing estimates of unobserved phenomena,” according to Adams. “It’s very exciting to find that these tools map so well to the world of cell biology.”
Image: Shutterstock.com/ Nadezda Razvodovska