DNA-based neural network reads ‘molecular handwriting’
Image credit: Dreamstime
Bioengineers based at California Institute of Technology (Caltech) have created an artificial neural network capable of identifying the molecular equivalent of handwritten numbers.
The work is part of an ambitious project to build artificial neural networks – machine learning systems which loosely mimic the behaviour of the biological brain – using DNA.
“Though scientists have only just begun to explore creating artificial intelligence in molecular machines, its potential is already undeniable,” said Professor Lulu Qian, the bioengineer heading up the Caltech laboratory engaged in this field.
“Similar to how electronic computers and smart phones have made humans more capable than a hundred years ago, artificial molecular machines could make all things made of molecules, perhaps including even paint and bandages, more capable and more responsive to the environment in the hundred years to come.”
Now, Qian and her colleagues have created DNA-based networks which function like small networks of neurons in order to process molecular information.
They tested their neural network with a standard test for machine learning; recognising handwritten figures (in this case, using “molecular handwriting”). This is challenging – often for human as well as computers – due to the enormous range of approaches people take to drawing letters and numbers, and requires a neural network not just to recognise figures, but also to account for variation when it compares a scribbled figure with its “memory” of figures.
In this case, Qian’s graduate student Kevin Cherry was able to demonstrate that a neural network constructed from custom-designed DNA sequences was capable of identifying molecular numbers encoded in 20 unique DNA strands representing pixels, which were mixed together in a test tube.
“The lack of geometry is not uncommon in natural molecular signatures yet still requires sophisticated biological neural networks to identify them; for example, a mixture of unique odour molecules comprises a smell,” Qian explained.
This DNA-based network was capable of classifying molecular handwriting representing numbers from one to nine. The learning process involved using an “annihilator” DNA molecule to select a winner when identifying a number; in this process this molecule forms a complex molecule between different competitors, eventually “[eating] up” competitor molecules until just one competitor species remains, which indicates the network’s decision.
Qian and her laboratory plan to develop more complex artificial neural networks which are capable of forming “memories” and therefore learning to perform a range of tasks.
“Common medical diagnostics detect the presence of a few biomolecules, for example cholesterol or blood glucose,” said Cherry. “Using more sophisticated biomolecular circuits like ours, diagnostic testing could one day include hundreds of biomolecules, with the analysis and response conducted directly in the molecular environment.”