Deep generative models for the prediction of single cell gene expression from its image and vice versa

Robotic microscopy coupled with image analysis of millions of cells and single-cell sequencing, each making the cellular heterogeneity of tissues and organisms accessible, have been crucial advances for a fine understanding of life and pathologies.

In this context, Minos Biosciences has developed a technology allowing, for the first time at very high throughput, to explore the relationship between the gene expression profile of individual cells and their visual phenotype. The project will be carried out by the Computational Bio-Imaging and Bioinformatics teams of the ENS Institute of Biology and the ESPCI Paris Bio-Chemistry Laboratory, in partnership with the startup. For each individual cell, the relationship between the two types of available data (image and transcriptome) will be studied along three axes. First, the candidate will evaluate and compare the predictive capacity of transcriptome and single cell images for a task of cell classification into subpopulations. In a second, he will evaluate if the concatenation of the two types of data in input of a deep network allows to increase its predictive capacity. Finally, he will develop a deep-learning approach to evaluate with what accuracy it is possible to predict the visual phenotype of a cell from transcriptomic data and vice versa.

The proofs of concept developed during this project will have numerous applications for research and/or, in a personalized medicine perspective, for the identification and validation of “biomarkers” (cell profiles) for prognostic or diagnostic purposes.