New study signals major advance in the future of precision cancer care

A new paper in Biology Methods and Protocols, published by Oxford University Press, indicates that a new computational method may help researchers identify effective precision treatments for cancer more quickly and efficiently.

Precision oncology is the promising, recently developed, approach to cancer treatment in which providers shape therapies to the unique molecular profile of a patient’s tumor. Current cancer therapy depends increasingly on matching the right drug to the right patient. Large-scale studies evaluate thousands of drugs on hundreds of cancer cell lines to find genetic biological markers to predict a drug’s effectiveness. In practice, however, this data is incredibly noisy. Hidden distracting details—unmeasured biological differences between cell lines—can create false leads and cause researchers to miss important signals.

Implementing precision oncology treatment effectively relies on preclinical discovery tools, most notably large-scale medical screenings. These screens test drugs on hundreds of cancer cell lines—each representing a different tumor type—to discover the genomic features that predict a drug’s effectiveness. However, a model based solely on observed features is insufficient because drug response depends on a multitude of unobserved variables reflecting hidden properties of both cancer cell lines and drugs. For instance, a cell line’s response is often determined by the position of a tumor in a patient’s body- something rarely measured directly in screenings. The presence of unobserved variables like this creates data problems that can mask true drug-gene associations or create false ones.

Researchers here generated a proposed framework, which they call the Structured Orthogonal Latent Variable Estimation (SOLVE), to overcome the challenges presented above by jointly and explicitly modeling the contributions of unobserved hidden factors alongside known specific predictors of cancer drugs within a single, unified model. This is a new statistical framework designed to separate significant gene–drug relationships from this hidden background noise. SOLVE analyzes three pieces of information at once: the genetic features of each cell line, the chemical properties of each drug, and a third latent component that represents unmeasured biology. A key feature of the method, according to the authors, is that it forces this latent component to capture only what cannot be explained by the measured data, making the different contributions identifiable and interpretable within the model.

The investigators here believe that SOLVE provides both a one-step, closed-form solution for standard regression and a repeatable algorithm that also works for classification problems such as predicting sensitive versus resistant responses to drugs. Applied to two major cancer data resources, the Cancer Cell Line Encyclopedia and the Genentech Cell Line Screening Initiative, researchers found that SOLVE uncovered well established relationships, like the critical link between the EGFR gene—the gene that provides instructions for a protein that helps control cell growth and division—and the drugs that inhibit it, which competing methods missed.

By more cleanly disentangling measured predictors from hidden variables, the investigators believe that SOLVE offers a robust computational tool for biomarker discovery and for investigating why some tumors respond to treatment while others do not. This has direct implications for precision oncology and could be adapted to other complex biomedical studies.

The paper, “SOLVE: A Structured Orthogonal Latent Variable Framework for Disentangling Confounding in Matrix Data,” is available (at midnight on January 28^th) at https://doi.org/10.1093/biomethods/bpaf094.

Direct correspondence to: 
Jialai She
Phillips Academy 
Andover, MA
shejialai@gmail.com

To request a copy of the study, please contact:
Daniel Luzer 
daniel.luzer@oup.com