Robust Causal Methodology for Planning and Learning from Interventions in the Face of Uncertainty
Abstract
Advances in causal inference have provided a rigorous mathematical framework to reason about cause and effect. While early results in causal inference have focused on learning cause and effect from purely observational data, therecent development of the CRISPR-Cas system for genomic editing combined with high-throughput genomic sequencing technology have led to the collection of large-scale observational and interventional datasets. In many applications, the end goal of causal modeling is to predict how to manipulate a system to move it towards a desired global state. Current causal inference algorithms require strong assumptions and are therefore not robust to different forms of uncertainties. In particular, there is a shortage of causal inference methodology that is robust to the presence of latent variables, data corruption, or uncertainties in the interventions. But only if we are able to quantify the uncertainties in the system can we design and plan local interventions in the system to achieve a desired global outcome. With a focus on learning from interventional data, our goals are to develop a causal inference framework that is robust to the three most common forms of uncertainty, namely (1) in the data collection process (such as data corruption or the presence of latent variables), (2) in theinterventions (such as imperfect interventions or off-target effects), and (3) in the model itself. This new framework will then allow us to develop robust strategies for iteratively planning, performing, and learning from interventions to movethe underlying system towards a particular desirable global outcome. We will validate the developed methodology based on available large-scale observational and interventional single-cell RNA-seq data from embryonic stem cells. Due to the shortage of donor organs, the problem of how to induce directed differentiation from embryonic stem cells into particular cell types has been recognized as a major goal in stem cell biology. We will apply our methodology towards determining the best set of target genes for directing the differentiation of pluripotent cells towards a specific cell type and ultimately forward-engineering morphogenesis through iterative rounds of planning and learning from interventions. Taken together, this proposal aims to build a transformative theory of directed interventions by developing a causal inference framework that is robust to the different forms of uncertainties in real-world applications. Because of the unique opportunity to access high-throughput observational and interventional data in genomics, most of this proposal is illustrated in light of biological applications. However, the developed methodology will be readily applicable to other large, interconnected systems such as social networks or naval operations. Navy s global operations are becoming increasingly complex, making careful control and design of new strategies integral to its success. The intricacy of the naval system calls for the use of novel computational methods to predict the effects ofdifferent decisions and strategies. It is particularly critical for the Navy to understand how to plan new operations to achieve a desired outcome. Causal inference can be used to model naval operations and understand the mechanisms that are activated by various actions, thereby enabling data-driven decision-making. For robust decision-making it is crucial to take into account the uncertainties at all stages. Causal inference is still at its infa ncy and the need for robust methods is acute. Our proposed research will deliver a set of statistical tools and algorithms for planning interventions under uncertainty to robustly move the system towards a desirable state. This new set of tools will equip the Navy with better and more robust strategy design and has the potential to give the Navy a competitive advantage because of its unrivaled access to data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 27, 2018
- Source ID
- N000141812765
Entities
People
- Caroline Uhler
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy