Using metabolic networks as artificial neural network architectures

Type de poste
Niveau d'étude minimal
Durée du poste
Contrat renouvelable
Contrat non renouvelable
Date de prise de fonction
Date de fin de validité de l'annonce
Nom de la structure d'accueil

Campus INRA de Jouy
Domaine de Vilvert
78352 Jouy-en-Josas

Jean-Loup Faulon
Email du/des contacts

We have recently demonstrated that it was possible to engineer a perceptron (one-layer neural network) in an E. coli cell extract and to use it for sample classification. Having shown that metabolism can be used to process information in engineered biological systems, we are seeking to which extent this is the case in natural systems, in particular with bacteria where signaling pathways are not as developed as in higher organisms. Answering this question is the main purpose of the internship.

To this end, we propose to use the metabolic networks of several bacterial species as artificial neural network architectures. The networks will be trained and cross-validated for classical machine learning datasets and results will be compared with those obtained using state-of-the-art architecture. Particular attention will be paid on defining the inputs and outputs of the networks. Among the various possibilities, inputs could be restricted to metabolites that can cross cell membrane and outputs could be metabolites that directly trigger transcriptional and translational response. Various learning strategies will be tested including reservoir computing to deal with recurrent loops.

To perform the work the recruited student will benefit from preliminary work we have carried out exhibiting examples where metabolism is transducing signals along with a database of molecules acting on gene expression and the RetroPath and SensiPath software tools. To build the network architecture the student will be using a Python script we have been developing to encode metabolite and enzyme layers. Libraries to be used include Keras, PyTorch (for artificial neural networks) and CobraPy (for metabolic networks).

The recruiting team (15 members from INRA, U. Manchester and U. Evry) develop in silico (retrosynthesis and machine learning), in vitro (cell-free) and in vivo (E. coli) methods to design and engineer metabolic pathways. Our pathways are used in the context of metabolic engineering, biosensing, and biocomputing (see The recruited student will work closely with a research engineer, a research associate, and a PhD student who are currently designing and engineering signal transduction, integration and processing using metabolic pathways in the context of whole-cell and cell-free biosensors.

Selected references
• Pandit A. et al. Metabolic Perceptrons for Neural Computing in Biological Systems. Nature Communications, 2019, 10(1): 3880.
• Voyvodic PL, Pandi A, et al., Plug-and-play metabolic transducers expand the chemical detection space of cell-free biosensors. Nature Communications, 2019, 10(1):1697.
• Delepine, B., et al., RetroPath2.0: A retrosynthesis workflow for metabolic engineers. Metab Eng, 2018, 45: 158-170.
• Koch, M., et al., A dataset of small molecules triggering transcriptional and translational cellular responses. Data Brief, 2018, 17: 1374-1378.
• Delepine, B., et al., SensiPath: computer-aided design of sensing-enabling metabolic pathways. Nucleic Acids Res, 2016, 44(W1): W226-31.

Equipe adhérente personne morale SFBI
Equipe Non adhérente