Optimal Design of Interventions for Causal Discovery in Genomics
Details
Title : Optimal Design of Interventions for Causal Discovery in Genomics Author(s): Caroline Uhler Link(s) : https://icml.cc/virtual/2022/workshop/13456
Rough Notes
Invited talk at the Adaptive Experimental Design and Active Learning in the Real World @ ICML 2022.
Motivation from single-cell biology:
- We now have access to high-throughput observational and interventional single-cell gene expression data and imaging data.
- Given this, we want to discover e.g. wiring diagram of a cell, or do drug discovery, identify best pertubations for cellular reprogramming (, ) (i.e. move cell from one state like disease state, to another state) etc.
From a causal perspective there are 2 problems of interest, that is, given a fixed budget we may wish to identify optimal interventions to:
- Learn the most about the underlying causal model.
- Move the system to a desired state (causal matching problem).
When deciding optimal interventions, increasing the number of variables to intervene on does not necessarily give better and better results.