Supervised learning in DNA neural networks

Learning enables biological organisms to begin life simple but develop immensely diverse and complex behaviors. Understanding learning principles in engineered molecular systems could endow non-living physical systems with similar capabilities. For 30 years, it has been hypothesized that neural computation occurs not only at the multicellular level in the brain but also within individual cells at the molecular level. This hypothesis suggests that brain-like learning rules could be implemented by molecular circuits. Recent experiments with enzymatic and enzyme-free DNA circuits have demonstrated rudimentary brain-like behavior in test tube chemistry. However, in these systems, learning was performed on electronic computers, with the results then implemented in molecular systems, limiting their functionality to predetermined tasks. In this study, we demonstrate that a test tube of DNA molecules can be programmed to carry out supervised learning. By exposing the system to molecular examples of inputs and desired responses, the information directly obtained from the molecular environment enables the system to develop its own pattern classification capabilities. We show that the same set of molecules can be trained to classify different sets of 100-bit patterns, suggesting the potential for molecular machines to learn autonomously from their environments.