Revealing Developmental Logic with Deep Neural Networks

Embryonic pattern formation is a classic case of emergent phenomena, where macroscopic structures result from recursive interaction of simple identical units like cellular automata. Being fascinating these phenomena themselves, the reverse design problem is particularly challenging: given a pattern or function, what could the underlying microscopic logic be like?

Here we demonstrate deep neural network to be a powerful tool tackling this problem. With the cellular rule to be determined being represented by fully-connected network, the developing embryo become a recurrent-convolutional network, and is trained with back propagation.

Though DNNs are well-known black boxes themselves, we still wonder, particularly, whether this approach can help revealing real biological mechanism given phenomena or some measured data. This idea was tested on Turing pattern and Drosophila gap-gene system, and could result in 1) a similar-looking regulatory network compared with existing knowledge, and 2) predictions on mutant behavior which bare notable similarities with experiment data.

We believe in the near future, this technique can help settle various unresolved issues in embryonic development.