“Cancers are very complex,” explained Greene. “Our goal is to measure which genes are being expressed, and to what extent they’re being expressed, and then automatically summarize what the cancer is doing and how we might control it.”
Normally, it is difficult to apply computational models across different studies because the gene expression data is “noisy,” meaning that there are many factors that differ in the way gene expression is measured. To begin their analysis, Greene’s team added more noise to the data and then trained a computer to remove the noise. To remove the noise, the computer had to learn about key underlying features of breast cancer. “This approach of removing noise makes the models we constructed more generally applicable,” Greene said.
Greene and the Dartmouth team studied DAs, which train computers directly on the data without requiring researchers to provide known biological principles to the computer, as a method to identify and extract complex patterns from genomic data. The model that the computer constructs can then be compared to previous discoveries to understand where data supports those discoveries and where the data raises new questions. The performance of DAs was evaluated by applying them to a large collection of breast cancer gene expression data. Results show that DAs were able to recognize changes in gene expression that corresponded to the cancers’ molecular and clinical information.
“These techniques and findings will enable others to use the DAs to evaluate gene expression data in a variety of disease sites,” reported Greene. “While noise in data is usually viewed as a problem, adding noise to data can actually be a good thing because it can help reveal the underlying signal. When we did this to analyze data from breast cancers, we found gene expression features that generalize across studies and represent important clinical factors.”
Next for Greene’s research team are more complex models that take multiple levels of regulation into account. Their goal is to develop methods that not only model data but that can automatically explain to researchers what the models have learned.
The report can be found online at: http://psb.stanford.edu/psb-online/proceedings/psb15/tan.pdf