Deep learning has become a popular tool in the life sciences. It is based on the architecture of neural networks, which consist of computing units connected in a layered fashion. Usually embodied as a computer program, they excel at mapping intricate relationships between input and output. Besides silicon-based computers, networks of interacting biomolecules also have the potential to perform complex computation. However, more challenges exist when one manipulates DNA than executes a line of code.
Anthony Genot from the University of Tokyo and colleagues present designs of networks of biomolecules that mimic neural networks’ structure and function. These networks are built with different sequences of nucleic acids and enzymes catalyzing their production, cutting and degradation. Output molecules from one layer are fed into the next, enabling a programmable and nonlinear transformation from the input signals of microRNA concentrations to output fluorescence signals. Compared to non-enzymatic systems, the advantages of this strategy include higher speed and sensitivity, as well as the ability to correct errors, notes Genot. Furthermore, “Enzymes bring high nonlinearity, which we exploit to improve the sharpness of computation. In practice this allows us to perform computations like majority voting on 10 bits, whereas non-enzymatic methods are limited to 3 bits.” Much effort was spent to fine-tune activity of the enzymes. “We built a microfluidic platform to scan experimental conditions and quickly pinpoint the optimal cocktail of reagents and temperature.”
This is a preview of subscription content, access via your institution