ESPCI web site
In vitro and in vivo neural computations with metabolism
Engineering information processing devices in living systems is a long-standing venture of synthetic biology. Yet, the problem of engineering devices that perform basic operations found in machine learning remains largely unexplored. I will first present the in vitro (cell-free) engineering of enzyme catalyzed perceptrons. The performances of the perceptrons will be exemplified with biological samples classification based on their metabolic, protein and RNA compositions. With the goal of engineering in vivo learning devices, I will next show how hybrid mechanistic-neural models can perfectly surrogate classical constraint-based modelling, and make metabolic networks suitable for backpropagation and consequently be used as an architecture for machine learning. The performance of the hybrid models, trained with experimental datasets of E. coli growth rates in different media composition and different gene knockouts, will be showcased on cross-validation sets.