Data Analytics Model Accurately Determines Impact of Cancer Drugs

Using a new data analytics approach, researchers could enhance cancer drug response predictions by accounting for overlooked variation across and within cancer types, a study published in PLOS Computational Biology revealed.

With new technologies and large genomic datasets, researchers can examine shared features not just within a single type of cancer, but can look for patterns across many different types of cancer.

This information can provide hints that an approach that worked well in one type of cancer may also be successful against a different type of cancer based on common underlying features. However, going through all this data is challenging. A model could predict that a drug works against multiple cancer types, but may miss critical variation in within individual cancer types and fail to recognize that the drug isn’t helpful for certain subsets of patients.

Researchers from the University of Michigan Rogel Cancer Center aimed to improve anti-cancer drug response predictions. The team set out to develop a data analytics approach that could allow for simultaneous examinations of differences across multiple cancer types, as well as within individual types.

“It’s like the old argument about nature versus nurture,” said study co-senior author Jun Li, PhD, a professor of human genetics, and associate chair for research of computational medicine and bioinformatics.

“Obviously both contribute. The questions we set out to answer were: How much does each contribute? And can we use that information to make predictions that would be useful into the clinic?”

The study used MEK inhibitor response as a proof-of-concept example, and drew on two public datasets that each contained several hundred patient-derived cancer lines.

The model analyzed mRNA expression, point mutations, and copy number variations. The results showed that while predictions of drug response were highly accurate when comparing one cancer type as a whole to another cancer type, the predictions only held up for about five of ten cancer types when looking at that cancer type on its own.

“What this means is that to be useful in the clinic for helping individual patients, we have to be able to incorporate both between-cancer and within-cancer data,” said co-senior author Matthew Soellner, PhD, an assistant professor of chemistry in the U-M College of Literature, Science, and the Arts.

“Otherwise, you may capture the average response to a drug across cancer types, but completely lose where an individual patient falls within their cancer type.”

For example, most colorectal cancer cell lines are sensitive to MEK inhibition, but liver cancer cell lines show a much more mixed response. Therefore, additional biomarkers and information beyond just the cancer type would be important for determining the likelihood that an individual patient with liver cancer will respond well to an MEK inhibitor.

To achieve this goal, the team developed a visualization strategy called a cigar plot, where differences in response between cancer types can be viewed simultaneously with responses within each type of cancer.

The more elongated, or cigar-shaped, the distribution of results within a cancer type, the better it can be used to predict response across different individuals affected by that type of cancer – whether brain cancer or lung cancer.

Researchers noted that this approach of balancing considerations of both cancer type and individual variation within a type when making a treatment decision could be applied to other types of diseases as well. The approach could work with other conditions that affect multiple tissues or where predictions are drawn across broad, diverse populations.

“We hope that this approach can serve as a general tool for the field,” said Merajver, a professor of epidemiology and of internal medicine. “More patients than ever are participating in basket clinical trials, which select based on a cancer’s molecular features rather than where in the body the cancer originated, so prediction models will be increasingly important for matching patients with effective treatments.”