Interpretable single-cell perturbations from decoder gradients

Generative models implicitly learn underlying dynamics of data and can do more than just reconstruction. By leveraging output gradients with respect to the latent dimensions, we explore a simple approach to infer arbitrary perturbation effects which generates interpretive flow maps within high-dimensional biological datasets. By applying this method to several cases in single-cell RNA-sequencing, we demonstrate its use in inferring effects from knockdown, overexpression, toxin response and embryonic development. This approach can further add global structure to dimensionality reductions which normally only preserve local patterns. Needing only a decoder, our method simplifies analyses, is applicable to already trained models, and offers clearer insights into cellular dynamics without complex setups. In turn, this gives a more straightforward interpretation of results, making it easier to discern underlying biological pathways with easily understandable visual representations. A temporarily anonymized codebase is found on https://github.com/trackpointer/perturbations.