Using Neural Networks to Extend Experimental Trajectories in Genetic Circuits

There are many processes in nature that exhibit noisy time series data and for which data collection is difficult. Examples in biology include protein expression in genetic circuits, cell shape, and molecular simulation dynamics. In all these examples, there are practical limitations to data collection including cost and technological or computational difficulties. We present a procedure to extend time series data with machine learning. A gated recurrent unit (GRU) neural network is trained on a simple gene network as a proof of concept. The GRU network then extends a protein trajectory from the gene network. The principle of maximum caliber (MaxCal) is used to assess whether the GRU network has accurately and faithfully extended this protein expression trajectory. Our results indicate that the machine learning model can accurately extend a time trajectory of protein number. This procedure for data extension, combined with the quality control provided by MaxCal, can be extrapolated to more challenging gene circuits and other biological problems where extension of stochastic time series data would be useful.