Most previous neural networks consisted out of a physically connected network of neural cells. But can a soup of interacting molecules also show brain-like behavior? Apparently, it can.
Researchers at Caltech (the same team that, a little over a month ago, made a DNA computer that can calculate square roots) created a neural network of four neurons, each made out of 112 distinct DNA strands. This network was subsequently ‘trained’ to recognize four scientists, whose identities were represented by a set of specific answers to four yes/no questions. The network, then, tried to identify one of these scientists based on an incomplete set of answers. So, a human player thought of one of the four characters, fed the network some clues in the form of DNA strands in the test tube and awaited the answer, which was communicated through fluorescent signals. The network could even ‘say’ that it didn’t have enough information to unequivocally pick the correct ‘mystery scientist’. The game was played with 27 different answer combinations, and the network got it right each time.
The network is thus able to ‘consider’ an incomplete pattern and ‘figure out’ what it might mean. Such brain-like behavior in biochemical systems has been hypothesized for some time, but, until now, it hadn’t been realized. The network is based on a simplified model of a neuron, which receives input signals, assigns a certain weight to them, and, when the weighted sum of the inputs exceeds a certain threshold, produces an output.
The technique used is known as a strand-displacement cascade. Basically, single and partially double-stranded DNA molecules (double helices, with one strand that sticks out) are used. Single stranded DNA molecules float around and might run into a partially double stranded colleague. If their bases match, the single strand binds and removes the other strand of the helix. Since the researcher can sequence DNA strands with whatever base sequence they want, they can program the interactions.
For more details on how it was done, be sure to check the following videos (by the way, big kudos for the researchers for putting these videos on youtube):
Video, Part I introduces the motivation, the mathematical model of a neuron, and the design of a DNA-based neural network that is capable of recalling a memory when presented with partial information. (Source: luluqian)
Video, Part II introduces how to make such a DNA-based neural network in the laboratory, how it remembers four scientists and identifies one of the given an incomplete answer to a set of yes-or-no questions, and how this work could be related to the evolutionary history of intelligence. (Source: luluqian)
This remarkable study shows that DNA-based networks can exhibit brain-like behavior. Of course, four neurons isn’t much (in no way does this diminish the authors' achievement), and making larger networks would be challenging. Furthermore, the system is slow. It took the network roughly eight hours to identify the ‘mystery scientist'. Nevertheless, this is a wonderfully well-crafted study and this technique certainly merits further inquiry.
In the words of the authors:
Our results suggest that DNA strand displacement cascades could be used to endow autonomous chemical systems with the capability of recognizing patterns of molecular events, making decisions and responding to the environment.
Qian, L.; Winfree, E. and Bruck, J. (2011). Neural network computation with DNA strand displacement cascades. Nature. 475, pp. 368 – 372. doi: 10.1038/nature.10262.