Name: Towia Libermann
Position: Associate professor of medicine at Beth Israel Deaconess Medical Center and Harvard Medical School; Director of BIDMC Genomics Center at the Dana Farber-Harvard Cancer Center Proteomics Core
Towia Libermann's group at Beth Israel Deaconess Medical Center has been working to overcome a difficulty in developing gene expression-based ovarian cancer prognostics stemming from the fact that patients lie along a continuum of survival chances. Prognostic profiles based long-surviving and short-surviving patients within this continuum run the risk of including gene data from patients who have signatures that do not necessarily add to the predictive value due to their individual, variable survival times.
Instead, Libermann and colleagues take the extreme cases in a survival continuum to create a profile. Pharmacogenomics Reporter spoke to Libermann at the Microarrays in Medicine meeting last week in Boston to learn how he and colleagues create prognostic profiles to avoid the continuum problem.
Does it always make sense to use only the extreme cases when building prognostic profiles?
It really depends on which type of samples were using, so it's a question we have about what is the best approach. And that's sort of primarily for all the studies we do as we get to survival — or other kinds of data where you look at a dynamic range of things that involve a time[point] that you don't want to make into an arbitrary cut-off.
In survival analysis, people typically do some sort of cut-off — 5 months, 20 months — whatever they felt was appropriate. Sometimes it is completely arbitrary. Sometimes it is based on some kind of a clinical parameter where they said, "OK — the majority of people have recurrences within that time period."
We felt it was very hard to make an arbitrary cutoff because, for example, if two years is the cutoff, why is someone who dies at one year and 11 months a short survivor, and a person surviving two years a long survivor? It's really one month's difference, and it doesn't make much sense biologically, so we decided that's not the way to do it, so we have to have a way to use the dynamic.
Even if some people who are longer survivors fall into the other category or vice versa, there may be many parameters that contribute to survival, but it's not only the biology of the cancer — there might be other things as well.
Because we don't know which labels to assign to patients, we thought the best way to do it was to take the extreme ones who have the best-defined biology. So, if someone's really dying very rapidly, and on the other end someone is living 5 years, 10 years — they should be in well-defined groups of people that are biologically significantly different.
So we use the extremes of the shortest survivor and the longest survivors as a starting point — a seeding ground — in order to derive the gene signatures. We derive the signatures from the extremes, and then use that to separate the patients who are anywhere in between, and then look on the Kaplan-Mayer curve [to help] separate them into two different curves.
And indeed, it did so nicely. Then we used that gene signature to go to a dataset [made up of the middle, non-extreme cases] and used this gene signature that was derived from the extremes on the dataset, and it fit very well and gave very nice discrimination with regard to survival.
We are doing this kind of thing in other cancers as well. What we haven't done is take this approach to published datasets — we should do that also because even with regard to ovarian cancer, there have been other studies published by other groups with regard to survival where they did arbitrary cutoffs.
Is there a way to say whether and by how much profiles created this way work better?
I think that's really only feasible if you do the side-by-side comparison and do some arbitrary cutoffs. Presumably, you will find different sets of genes in these types of analysis — it might still overlap because there might be some overlapping biology that will come out — but I assume there will be quite a lot of subtle differences between the kind of gene signatures.
But we were definitely very encouraged by the types of genes that came out in our survivor gene signature, because a lot of them made biological sense. A lot of the genes were involved in oncogenesis, proliferation, and those kinds of things. It wasn't just a set of random genes — lots of them had been associated with survival in other types of cancers. That gave us a lot of confidence that this might be actually a very appropriate way to do it.
What is the main advantage?
I think the main thing is that it allows you to really apply [it] to anything where you cannot make an arbitrary cutoff, or where you have no rationale for making a cutoff.