For the first time, humans have played tic-tac-toe – also known as noughts
and crosses – with bacteria. These were no ordinary bacteria, but E. coli
extensively genetically modified and set up to act as a simple neural
network, a form of artificial intelligence.
This approach could have all kinds of applications, from creating living
materials capable of learning to making “smart” microbiomes, says Alfonso
Jaramillo at the Spanish National Research Council.
He and his team started with an E. coli strain genetically modified to sense
12 different chemicals and respond by altering the activity of any genes the
researchers chose. This strain, called Marionette, was created in 2019 by
another group.
Jaramillo and his colleagues further modified the Marionette strain so that
it had numerous copies of two bits of circular DNA, called plasmids, each
coding for a different fluorescent protein: one red and one green.
The ratio of the number of these two plasmids – and hence the colour of the
bacteria’s fluorescence – isn’t predetermined and can be altered by the 12
chemicals and by certain antibiotics. In the absence of any further input,
this ratio remains constant and is thus a form of memory.
What’s more, when the bacteria do get another input, the output – the colour
resulting from the ratio of fluorescent proteins – depends on the previous
ratio. This means that the bacteria behave in the same way as an electronic
component called a memristor that is being used to create computer chips
that mimic how the synapses in a brain work. Jaramillo calls these creations
“memregulons”.
The team decided to teach these memregulons to play tic-tac-toe, as this is
a benchmark often used to demonstrate new approaches in computing. The
bacteria were grown in eight wells corresponding with the outer squares of a
tic-tac-toe grid.
For simplicity’s sake, the team assumed that the human player always starts
and puts a cross in the centre square. The first bacterial nought is then
placed on the square corresponding to the well with the reddest colour.
The human plays next and the bacteria are “told” of the move by one of the
chemicals they can sense being added to each well – each chemical
corresponds to one square. That changes the protein ratio in each well,
indicating the next move. Each game takes several days as time is needed for
the bacteria to respond.
“In the beginning, the bacteria play randomly,” says Jaramillo. But they can
be trained by “punishing” wells that play a wrong move with a dose of
antibiotics.
After eight training games, the bacteria became expert players, says
Jaramillo. The team simulated how the trained sets of bacteria play games,
and these simulations show they could beat unskilled humans. But the
researchers didn’t play any further games after the training stage in which
the bacteria lost every time, so E. coli have yet to actually beat humans at
tic-tac-toe. “We did not bother to play those winning games,” says
Jaramillo.
“[It] is a powerful demonstration of adapting a complex biological system to
perform an entirely artificial task,” says Joanne Macdonald at the
University of the Sunshine Coast in Australia. In 2006, Macdonald created a
DNA-based computer that was unbeatable at tic-tac-toe.
“The tic-tac-toe game playing with bacteria is an excellent demonstration of
their innovative work,” says Sangram Bagh at the Saha Institute of Nuclear
Physics in India, who leads one of the two groups that have previously
created bacteria-based artificial neural networks.
He isn’t convinced that Jaramillo and his team’s set-up meets the definition
of an artificial neural network. “But still, it is a good strategy,” says
Bagh.
Jaramillo says his system is a simple form of neural network known as a
one-layer linear artificial neural network. His team is already creating
more complex neural networks with the bacteria that can do tasks such as
handwriting recognition, he says. “They can do very sophisticated things.”
Reference:
bioRxiv,
DOI: 10.1101/2022.04.22.489191