Susan Cox, Reader in Cell Biophysics, King’s College London, UK

Revealing biological structure and heterogeneity with deep learning

The use of multiple views of difference instances of a biological structure to produce a single highly resolved image of the structure is well established in cryo-EM, and increasingly under development for fluorescence microscopy.
We take a deep learning approach to allow a free fit of the underlying biological structure and a parametrised fit of the heterogeneity of the structure.

By training the network we derive the optimal parameters for both the base model of the structure, and its variability.
We demonstrate that this approach can allow results to be derived about the way that biological structures vary, without the need to define input assumptions about the structures present in the sample.