Sophisticated new algorithms offer opportunities when it comes to understanding why cancers develop and finding treatments that work. The idea is to look at a patient's genome and look for genes where cancer mutations cluster in some specific regions of the gene (standard algorithms look for genes that have more mutations than would be expected according to the average mutation rate for that gene)-and that might be linked to a specific type of cancer or cancer drug known to work for other patients with that mutation.
But not every algorithm shows the complete picture. It's only when insights are combined from several algorithms that it becomes clear whether or not the right mutations are being correctly linked to the cancers they lead to or drugs that work for patients with those cancers, explained Adam Godzik, PhD, Director and Professor of the Bioinformatics and Structural Biology Program at Sanford Burnham Prebys Medical Discovery Institute in La Jolla, Calif.
Those were the findings of a paper Godzik and his colleagues recently published in Nature Methods (2017;14:782-788).
"The point here is that when results from various cancer genomic studies are presented, people usually see the final result of two actually independent things: data from the patients and the algorithm used to analyze these data," Godzik said. "When results from different studies are compared-with both data and algorithms being different-at the end we don't know what contributed to the difference."
This new research was designed to compare the effectiveness of multiple algorithms on the same set of data to reveal the difference each algorithm was finding.
"The key finding is that each algorithm in the group we compared is adding something unique, missed by other algorithms," Godzik noted. "The picture we get from using all of them is more complete than that obtained using just one algorithm."
In an interview with Oncology Times, Godzik elaborated about why these findings are significant, and what they reveal about the future of using algorithms for cancer diagnoses.
1 Why is this research important-especially as more and more algorithms for cancer are being developed?
"When you see two papers [that evaluate an algorithm for pinpointing a gene mutation that could lead to cancer], each of them have a different dataset and a different algorithm. So you may not actually be so sure what is causing the differences. Is it the difference in the data or is it the difference in how people analyze it, the algorithm?
"So in this case, these algorithms that we looked at-each of them is giving us part of the truth. And the best insights we get is that, when we combine all these algorithms together, we see the global picture."
2 So you're not saying your data shows any one algorithm is necessarily wrong? What's the takeaway about each specific algorithm you looked at?
"Each of the observations might be true. If you have a group of people who are looking at people with breast cancer, somebody analyzes that and says there's a group that has mutations in [a specific] gene and this group has low survivability and should get more aggressive treatment. We're saying, well if you use another algorithm you might see another group [that has that gene but needs another type of treatment].
"It's not like one algorithm is correct and another algorithm is wrong. Each algorithm is showing you part of the truth and you would be missing something if you don't use multiple algorithms. Most observations they would make would be correct, but it wouldn't be all of the truth. Some algorithms would miss some of those subgroups. And we still don't know, in most cases, what it means. But in the future, we anticipate we'll know how the different subgroups could be treated differently.
"For instance, a big drug trial may fail because they take 1,000 people and give the drug to these people and it only helps 20 people. So, they say it's not good enough and the drug is not approved. But with tools like [these algorithms] we can say yes, it only helped 20 people, but it helped 100 percent of people who have a certain feature or a certain set of mutations. And this could be very useful because now we can say, well, the drug is not for everybody, but the drug can work for this one specific group.
"So this is the big picture here. This is the dream."
3 What is most important for oncology clinicians to know about your findings-and the future of using algorithms to diagnose and classify cancers?
"Doctors: Pay attention to how the data was analyzed, it could be as important as what is being analyzed.
"I think it would be too early to say these tools are usable now-or give a lot of insights for practicing clinicians. At this point in day-to-day clinical practice, genomic information is very rarely used or may only focus on one or two specific genes, like BRCA1/2. But this research is part of a bigger trend of more and more detailed analyses coming from cancer genomes."