Use of genetic variance to identify potential therapeutic targets
Dr John Quackenbush – Dana-Farber Cancer Institute, Boston, USA
Professor Quackenbush, we’re here at WIN 2012 and you’ve given a very interesting presentation today. Can you discuss this with us a little bit?
What I did today was I presented some of the work that we’ve been doing in which we’ve tried to really understand how we look at biological problems and question the basic assumptions that we’ve made over the years about how we analyse biological data. The genesis of this is that I was working with Christine Wells, who is now at the University of Queensland, and Jess Mar, who is now at Albert Einstein College of Medicine, and we were looking at data that was generated in neuronal stem cells from healthy patients, from patients who had schizophrenia, from patients who had Parkinson’s disease.
As we looked at the data, we realised that in most instances in biology when we make a measurement, what we do is we look at two groups, say, treated and controlled or tumour and normal or chemoresistant and chemosensitive and we ask, for the measurements we make, on average is there a difference between the measurement in group one and the measurement in group two. That’s a fundamental question that’s really summarised mathematically in the T-test because the average itself is an informant.
On average men are taller than women but height is not a good discriminator of gender because the variance, the spread, is large. So the question become, on average is there a significant difference given the spread, given the variance? What this allows us to do is to find very good markers in a number of different diseases, in a number of different models that tell us about what’s driving the differences between the phenotypes we observe. But in the study of neuronal stem cells we looked at that question and asked, on average, in key pathways associated with stem cell differentiation, is there a difference?
What we realised was when we looked at this data that there was another question that was much more informative and the question we started asking was, independent of the average in these populations, is there a difference in the variance? How variable is gene expression in these populations? And the amazing thing we found was that in our schizophrenics we saw a compression of the variance relative to the controls. In our Parkinson’s patients in the same neuronal stem cells we saw an increased variance in key pathways associated with differentiation.
Now, you might ask what does that tell you about the biology? If we look at this variation, the way we interpret it is that a compressed variance is actually associated with increased regulatory control or decreased plasticity; an increased variance is associated with the degradation of regulatory control – the system is letting things become much more variable. As we looked at this data what we realised was that there are models of schizophrenia, in fact, that suggest that what drives the disease may in fact be diminished plasticity in neuronal development, completely consistent with our observations. If we look at our Parkinson’s patients with the increased variance, one of the things that we observed was that this variance, in fact, is something which people commonly associate with other diseases of aging, that there’s a degradation of regulatory control. So it really made us go back and think about how we analyse biological systems.
So what does this have to do with cancer? Well, what’s very interesting about this is that as we started to look at cancer what we’ve seen is that for most genes in the genome there’s an increase in the variance. Cancer is a de-differentiation disease, the cells are becoming less well defined, much more variable. On the other hand, there is some subset of genes in each tumour type where the variance becomes compressed, they’re more tightly regulated. Now why is that if cancer is letting things go, why is it grabbing on to these? Our hypothesis is that these may in fact be key genes in key pathways that the cancer needs to regulate very tightly to survive. If we put this in the context of other methods we’ve been developing to tease out pathways from genomic data, we can look for pathway differences but we can also look for differences in the variability. Our hope is that when we start to find these genes where the variance is compressed that those may, in fact, represent the best drug targets. They may be fragile points in the system where the tumour is holding on to them very tightly and if we can hit them with a drug that will perturb them, we may have a better chance of killing the tumour and, if the variance is broader in the normal tissue, then not disrupting the function of the normal tissue.
Do any of the classical mutated genes fit into this?
Often the mutated genes don’t fall into these key genes where there’s this over-regulation. But what we’re hoping to see in a very large dataset that we’ve amassed is there’s an over-representation of drug targets, even among the drugs today that fall into these pathways.
For instance, would P53 be one of them maybe?
Maybe, it’s not one of the genes that we’ve seen early on in our analysis but mostly we’ve been working with small sample sizes. I think the other problem that we really face is that even though we’ve been able to amass large bodies of data, increasingly we’re coming to understand that tumours are defined molecularly. I think we’ve seen this very clearly in breast cancer where we recognise three or four major molecular subtypes. But even in those subtypes we realise there are additional alterations. So some of the analysis we’re doing may point to lesions in particular mutations if we could subset patients based on knowing which patients have particular genetic mutations. And that’s part of the challenge, that the datasets we have to date typically don’t have multiple different types of genomic data.