Finally, thanks to Katie Galloway and Christopher Johnstone for this thoughtful perspective. We are indeed excited about the "learning/training" aspect of the Perceptein network! www.science.org/doi/10.1126/...
As shown in our preprint a while ago, we took a small step further and created a protein circuit, made of de novo designed protein heterodimers and engineered split viral proteases, that carries out weights-tunable winner-take-all neural network computation in mammalian cells.
Since our last preprint, we have
1) redirected the classification outcome to cell death, demonstrating the interfacability of protein circuits (thanks to Shiyu Xia)
3) cleaned up the chemical reaction network diagram (from left to right). Here, each circle is a unique (left) or a group of (right) protein species, and each orange dot represents one (left) or a group of (right) chemical reactions
It was tremendous fun brainstorming with
@elowitzlab.bsky.social
in the early days of this project, and collaborating with all co-authors on this paper. A nice cover art made by the talented Ehmad Chehre:
This work was inspired by the seminal paper by Cherry and Qian, where they showed one could recognize handwritten digits using DNA molecules in test tubes: www.nature.com/articles/s41...
2) scaled up the neural network to be 2-input and 3-output, showcasing the scalability of the Perceptein architecture (left, simulation; right, experiments)
www.nature.com
A synthetic protein-based winner-take-all neural network controls cell fate decisions
DNA-strand-displacement reactions are used to implement a neural network that can distinguish complex and noisy molecular patterns from a set of nine possibilities—an improvement on previous demonstra...
Good point on the learning part. Backprop seems difficult to implement using molecules, currently contemplating other means, e.g. the exhaustive search strategy you mentioned, or Hebbian learning
I am so glad the algorithm has brought me to this gem
