Ai+ Training

Modeling a cell's state e.g. during differentiation or in response to perturbations is a central goal of computational biology. Single-cell technologies now give us easy and large-scale access to state observations on the transcriptomic and more recently also epigenomic level. In particular, they allow resolving potential heterogeneities due to asynchronicity of differentiating or responding cells, and profiles across multiple conditions such as time points, space and replicates are being generated. This makes this an ideal application area for machine learning method development to understand cellular variation, contribution of particular transcripts as well as impact of perturbations.

In this talk, I will shortly review approaches from deep representation learning we and others have been using to identify the manifold of cellular state (=gene expression manifold). I will then introduce our recent perturbation model 'compositional perturbation autoencoder' (CPA), a deep autoencoder we developed to describe the impact of perturbations such as drug or genetic modification on this manifold. With CPA we can learn an interpretable model of perturbations and predict novel and/or optimal perturbations. I show examples of CPA predicting dosage-specific drug effects as well as combinatorial genetic interactions, and how CPA allows in-silico generation of putative interaction effects.