Unraveling the Web of Genetics: A Causal Approach to Gene Function and Interaction

With approximately 20,000 genes in the human genome that can influence each other in intricate ways, identifying which gene groups to target poses a significant challenge. Furthermore, these genes operate in modules that mutually regulate one another.

Researchers from MIT have devised theoretical foundations for methods that can determine the most effective way to cluster genes into related groups, enabling a better understanding of the underlying cause-and-effect relationships among numerous genes.

Crucially, this new method relies solely on observational data, meaning researchers do not need to conduct costly or impractical interventional experiments to gather the data necessary for inferring causal relationships.

Ultimately, this technique could assist scientists in pinpointing specific gene targets to influence specific behaviors, paving the way for more precise treatments for patients.

“In genomics, understanding the mechanisms that underlie cell states is vital. However, cells have a complex multiscale structure, making the level of summarization equally important. By finding the right way to aggregate the observed data, the insights gained about the system become more interpretable and useful,” explains graduate student Jiaqi Zhang, an Eric and Wendy Schmidt Center Fellow and co-lead author of a paper detailing this technique.

Zhang is joined by co-lead author Ryan Welch, who is currently a master’s student in engineering, and senior author Caroline Uhler, a professor in the Department of Electrical Engineering and Computer Science (EECS) and the Institute for Data, Systems, and Society (IDSS). Uhler is also the director of the Eric and Wendy Schmidt Center at the Broad Institute of MIT and Harvard, and a researcher at MIT’s Laboratory for Information and Decision Systems (LIDS). This research will be presented at the Conference on Neural Information Processing Systems.

Learning from observational data

The challenge the researchers aimed to address was learning the programs of genes. These programs illustrate how certain genes collaborate to regulate other genes within biological processes such as cell development or differentiation.

Since it is not feasible for scientists to study the interactions of all 20,000 genes, they employ a technique called causal disentanglement to learn how to effectively combine related groups of genes into a representation that allows for an efficient examination of cause-and-effect relationships.

In previous efforts, the researchers demonstrated this could be effectively achieved using interventional data, which involves data obtained by altering variables within the network.

However, conducting interventional experiments is often costly, and in some cases, such experiments may be unethical or not technologically feasible.

Using only observational data presents a challenge, as researchers cannot compare gene behavior before and after an intervention to understand how groups of genes interact.

“Most research in causal disentanglement presupposes access to interventional data, making it unclear how much information can be extracted using only observational data,” Zhang remarks.

The MIT team created a more generalized approach utilizing a machine-learning algorithm to accurately identify and group observed variables, like genes, based solely on observational data.

This technique allows them to recognize causal modules and reconstruct a precise representation of the underlying mechanisms of cause and effect. “While our research was driven by the need to clarify cellular programs, we had to devise new causal theories to understand what insights could be derived from observational data. With this theoretical foundation established, we can apply our findings to genetic data in future studies to identify gene modules and their regulatory interactions,” Uhler states.

A layerwise representation

Using statistical methods, the researchers calculate a mathematical function known as the variance for the Jacobian of each variable’s score. Causal variables that do not influence any subsequent variables should have a variance of zero.

The team reconstructs the representation in a structured, layer-by-layer manner, starting with the removal of the bottom layer variables that exhibit a variance of zero. They then reverse this process, layer-by-layer, eliminating variables with zero variance to uncover which variables or groups of genes are interconnected.

“Pinpointing variances that are zero quickly becomes a complex combinatorial problem, so developing an efficient algorithm to solve it was a significant challenge,” Zhang shares.

Ultimately, their method produces an abstracted representation of the observed data, featuring layers of interconnected variables that accurately depict the underlying causal structure.

Each variable signifies an aggregated group of genes that function in unison, and the connection between two variables illustrates how one group regulates another. Their method successfully encapsulates all the information necessary for determining each layer of variables.

After verifying the theoretical validity of their approach, the researchers conducted simulations to demonstrate that the algorithm can effectively extract meaningful causal representations using only observational data.

In the future, the researchers aspire to apply this methodology to real-world genetics scenarios. They also intend to investigate how their method can yield additional insights when some interventional data is available or aid scientists in designing effective genetic interventions. Eventually, this approach could enable researchers to more efficiently identify genes that work together within the same program, which may assist in discovering drugs that target these genes to treat various diseases.