New method analyzes whole-genome gene expression to predict whether tumors respond to chemotherapy
From oracle bones to astrology to psychic hotlines, countless generations of humans have aspired to predict the future. While still mostly impossible, there are certain areas where the right people with the right tools can make a really good guess. How well a patient responds to chemotherapy might soon be one such area, if R. Stephanie Huang, PhD, and her colleagues have their way.
Chemotherapeutics can be as devastating on healthy cells as cancerous ones. Unfortunately, how patients respond to these drugs can be variable, and ineffective courses of treatment are often physically, emotionally and financially taxing.
The ability to predict a tumor’s drug response would have profound implications in patient care and drug development. It would aid clinicians in deciding what chemotherapeutics would work best for individual patients, and help reduce the enormous time and financial costs needed to run clinical trials. It could also be used to discover new uses and targets for existing drugs.
Scientists have worked hard to identify biomarkers – some measurable biological indicator of whether a tumor will shrink under chemotherapy. However, the effects of biomarkers on drug response are generally small, as a host of complex genetic and environmental factors are involved. For Huang, assistant professor in the Department of Medicine, the solution to this challenge is a matter of scale. Rather than using single or small groups of biomarkers to predict whether a tumor responds to chemotherapy, she is using the entire genome.
“Everything in the genome may play a role in drug response,” Huang said. “Some genes might contribute a tiny amount, some might be large. Our thinking is, let’s not pick and choose. Let’s let the data speak for itself.”
To accomplish this, Huang, along with postdoctoral researcher Paul Geeleher, PhD, and Nancy Cox, PhD, professor and section chief of genetic medicine, utilized data from a massive library of almost 700 cell lines representing dozens of tumor types from the Wellcome Trust and Sanger Institute’s Cancer Genome Project. Each of these tumor lines had previously been tested for their sensitivity to almost 140 different drugs, and their whole-genome gene expression – a measure of which genes are being actively turned into proteins, and at what levels – profiled.
With this information in hand, the researchers built a computational tool that estimated the contribution of each gene to a tumor’s sensitivity toward a drug. They ran this tool to generate estimates on every drug and cell line combination, and then used it to analyze tumors from actual patients.
In three independent clinical trials for which patient data was previously published, Huang and her colleagues were able to predict whether a tumor would be sensitive or non-responsive to the drug observed in that trial with between roughly 70 to 90 percent accuracy. They reported these results earlier this year in the journal Genome Biology.
To generate their prediction, the researchers assigned the estimated contribution values for genes from the cell line data to the same genes in tumors from the clinical trials. When combined, this gave a likelihood for whether that tumor would be sensitive to a particular drug. Accuracy was verified by how the tumor actually fared under treatment.
“Despite different drugs and different disease settings, our model was able to clearly separate sensitive from resistant,” Huang said. “What we can do with this tool is better than almost everything else out there, whether a single biomarker or groups of markers.”
While happy with these initial results, Huang acknowledges there are obstacles. How each individual tumor sample is collected and treated can affect gene expression variability, muddling the predictive power of the model. To improve its accuracy, additional studies need to be conducted to gather gene expression and sensitivity data from as many different combinations of tumor cell lines and drugs as possible – a feat no single laboratory can perform on its own. And lastly, the model in its current form can only predict the effects of one drug on a tumor, not a multi-drug course of treatment.
None of these challenges are insurmountable, however. “Our model is built off data from the worst case scenario, and it’s still remarkably effective,” Huang said. “We can only improve its accuracy as we develop better collection and testing protocols, and incorporate additional data.”
Huang and her colleagues are now working to improve their model, and are collaborating with oncologists to test and prove its utility it in a small-scale clinical trial of neuroblastoma, a pediatric cancer. In the meantime, they have published the source code for their model in the open access journal PLOS ONE in order to share it as widely as possible.