MIT scholars have established a novel theoretical framework for examining the mechanisms of treatment interactions. Their methodology enables researchers to effectively gauge how combinations of treatments will influence a set of units, such as cells, allowing investigators to conduct fewer expensive experiments while obtaining more precise data.
For instance, in exploring how interconnected genes impact cancer cell proliferation, a biologist might need to utilize a blend of treatments to target multiple genes concurrently. However, due to the potential for billions of combinations for each experimental iteration, selecting a limited set of combinations for testing might skew the results generated by their research.
Conversely, the new framework addresses a situation in which the investigator can systematically design an unbiased experiment by administering all treatments concurrently, while also adjusting each treatment’s dosage to manage the outcomes.
The MIT researchers theoretically demonstrated a nearly optimal strategy within this framework and conducted a variety of simulations to evaluate it in a multiround trial. Their approach minimized error rates in every instance.
This methodology may eventually aid scientists in deepening their understanding of disease mechanisms and creating novel therapies for cancer or genetic conditions.
“We’ve introduced a concept that people can contemplate further as they investigate the most effective method for selecting combinatorial treatments at every phase of an experiment. Our aspiration is that this can eventually be applied to address biologically significant inquiries,” expresses graduate student Jiaqi Zhang, an Eric and Wendy Schmidt Center Fellow and co-lead author of a paper discussing this experimental design framework.
Along with her on the paper is co-lead author Divya Shyamal, an MIT undergraduate; and senior author Caroline Uhler, the Andrew and Erna Viterbi Professor of Engineering in EECS and the MIT Institute for Data, Systems, and Society (IDSS), who also directs the Eric and Wendy Schmidt Center and is a researcher at MIT’s Laboratory for Information and Decision Systems (LIDS). This research was recently showcased at the International Conference on Machine Learning.
Simultaneous treatments
Treatments can interact with one another in intricate ways. For example, a researcher aiming to ascertain whether a particular gene contributes to a specific disease symptom may need to target several genes at the same time to analyze the effects.
To achieve this, scientists utilize what are referred to as combinatorial perturbations, where they administer multiple treatments concurrently to the same group of cells.
“Combinatorial perturbations will provide you with a high-level overview of how various genes interact, offering insights into how a cell functions,” Zhang clarifies.
Given the high costs and time demands of genetic experiments, scientists strive to select the optimal subset of treatment combinations to examine, which proves to be a formidable challenge due to the vast number of possibilities.
Choosing a suboptimal subset may produce biased results by concentrating solely on combinations pre-selected by the researcher.
The MIT scholars tackled this issue with a different approach by adopting a probabilistic framework. Instead of concentrating on a chosen subset, each unit randomly receives combinations of treatments based on user-defined dosage levels for each treatment.
The user establishes dosage levels based on the objective of their experiment—perhaps this researcher aims to investigate the effects of four distinct drugs on cellular growth. The probabilistic method yields less biased data as it doesn’t confine the experiment to a pre-established subset of treatments.
The dosage levels function like probabilities, and each cell acquires a random combination of treatments. If the user sets a high dosage, it is more probable that most cells will receive that treatment. Conversely, a smaller fraction of cells will acquire that treatment if the dosage is lower.
“From there, the question is how do we design the dosages to estimate the outcomes as accurately as possible? This is where our theory plays a role,” Shyamal reflects.
Their theoretical framework illustrates the most effective method for structuring these dosages to optimize learning about the characteristic or trait being examined.
Following each experiment round, the user gathers the outcomes and incorporates those into the experimental framework. This will yield the optimal dosage strategy for the subsequent round, and this process continues, gradually adapting the strategy over multiple iterations.
Optimizing dosages, minimizing error
The researchers demonstrated that their theoretical approach produces optimal dosages, even when dosage levels are hindered by a limited supply of treatments or when noise within the experimental results fluctuates at each round.
In simulations, this innovative approach exhibited the lowest error rates when assessing estimated versus actual outcomes of multiround experiments, surpassing two baseline approaches.
Moving forward, the researchers aim to refine their experimental framework to account for interactions between units and recognize that certain treatments could result in selection bias. They also aspire to implement this technique in real experimental contexts.
“This represents a new method to tackle a particularly intriguing problem that is challenging to resolve. With this innovative framework at our disposal, we can conceptualize more effective experimental designs for numerous applications,” Zhang comments.
This research is partially supported by the Advanced Undergraduate Research Opportunities Program at MIT, Apple, the National Institutes of Health, the Office of Naval Research, the Department of Energy, the Eric and Wendy Schmidt Center at the Broad Institute, and a Simons Investigator Award.