One of the pressing challenges in cancer genomics is to distinguish driver mutations (i.e. mutations involved in tumorigenesis) and passenger mutations (i.e. functionally neutral mutations). Current approaches to identifying driver mutations in cancer look for recurrent events in multiple samples. However, it is now acknowledged that individual tumors of the same type are highly heterogeneous and have diverse genomic alterations. We developed a novel method, called DriveRank, to discover patient-specific molecular driver events. Our probabilistic model accounts for both gene interactions and somatic mutations, which allow us to quantitatively rank a list of mutations in any individual tumor based on mutations propensity to become advantageous alterations. DriveRank was applied to Glioblastoma, Ovarian Cancer, and Breast Cancer data from The Cancer Genome Atlas, and it was shown to rank proposed driver mutations from other predictive software accurately by displaying a median rank for candidate mutations to be over the 92 percentile for all three cancers. Our results showed that the usual suspects (TP53, EGFR, RB1, APC, etc..) appeared frequently along with rarer, yet functionally important, suspects not normally classified as drivers. The DriveRank method will potentially enable the determination the optimal treatment strategy for patients through individualized assessment.
All right, so I'd like to thank the entire cancer symposium for maybe making this event possible. I really enjoyed this event and
thank you for the opportunity to help me present this. So I'm here to talk about DRIVE-RANK,
which is a method that we developed in order to detect patient-specific driver mutations in cancer and
and I'd like to open by starting about describing the mystery of driver mutations in a computational sense. So over the last few years
recent advances in DNA sequencing technology have greatly enhanced our ability to
analyze cancer genomes because next-generation sequencing technology has allowed us to
have allowed us to detect a large spectrum of somatic mutations in cancer genomes and
from those and with the more complete genome set we are able to find more and more mutations in cancer which could
which could lead to a selective advantage over tumor cells and
knowing and knowing these mutations would greatly enhance our ability to
find the source and find possible treatment methods in cancer.
So using so these mutations called driver mutations are the mutations that we are looking for.
However, it is important to note that not every mutation in a cancer genome is going to be a driver mutation.
In fact, most mutations actually provide no selective advantage. These mutations are called
passenger mutations and thus we have the ability to sift out the
driver mutations from the passenger mutations has been a very big mystery in computational biology today. So
there have been many approaches in order to tackle this problem and there have been many
many previous methods have been developed. Most of these previous methods
operate under the idea of recurrence and recurrence is basically to say that if you have if you're given any patient any
any group of patients
whatever mutation occurs a lot occurs with a high frequency in those mutations is are going to be driver mutations.
There have been more sophisticated models using the using recurrence
by adding another term called exclusivity in which a subgroup of patients would include both a
subgroup of patients would have a driver mutation if both that driver mutation had
if that driver mutation was recurrent in that subgroup and exclusively recurrent in that subgroups. And this was a MIMO
MIMO
developed by Cyril et al and Vanden et al's dendrix path dendrix program both use both
recurrence and exclusivity and these software programs have been able to find many many types of common driver mutations.
However, these methods are limited because they are not necessarily patient specific
and the problem with not being patient specific is the fact that every patient would have a every patient has a different tumor.
No, there are no two tumors that are alike and it is very possible because not all tumors are homogenous
that there could be multiple driver mutations for multiple for different types of tumors. So thus
these methods might not provide the most direct answer in order to find a patient specific driver mutation.
So our goal is to address these
limitations in driver mutation software by directly assessing the impact of mutations in each individual patient.
And we decided to use this we decided to take a different approach than other methods have been looking for. Instead of looking for a patient
population or subpopulation what we decided to do is we decided to take a look at every individual patient and try to find out what
genes and that what mutations in that patient are most likely
most likely to be driver mutations.
And we do this by handling this like a mystery in a way like how Sherlock Holmes would solve a case.
And so we divide this into suspects and clues. Suspects we are going to limit to mutated genes.
So only driver mutations that actually affect the gene will be considered. And the clues or
characteristics of a driver mutation we decided to see if we decided to focus on two of them specifically.
One of them is interaction. How connected are
mutated genes? A suspect mutated gene in theory would connect to many and would have a high impact and thus connect to many other
genes in a network of genes. And also the other clue we would like to look at is expression, differential expression for it to be specific.
And expression is to say that if a
suspect mutation would be highly differentially expressed in cancer and also propagate other highly differentially expressed genes downstream.
So by looking at interaction in expression, we will be able to find a method to
method to rank individually what kind of what genes are most likely to be in cancer via a driver mutation.
So our method is a method of varying a Google PageRank and also a
developed and Morrison et al. a variation of Google PageRank called GeneRank which ranks
and Google PageRank the idea behind Google PageRank was it's it's the heart of Google's search engine
which was developed by Larry Page and Sergey Brin by using and used a stochastic random walk to determine the impact of websites. And
the stochastic random walk for those of you who are not familiar with that. It's a it's a relatively different. It's a it's a
it's a into it can be an intuitive process. My old statistics professor used to tell me that
just imagine a network for example a network of bars on Green Street for example over down the corner.
And if us if a random college student took got into a random bar and
spent some time there had a little bit too much to drink and couldn't
lost a sense of direction and just stumbled randomly into a near into a bar next door.
And then repeated this process over time until he can know he could no longer do that.
We could be able to find out the popularity of each bar by
simulating this over thousands and thousands of college students for example like on a Friday night we could simulate this and thus we would be
able to determine the impact of
every bar. And in a general theory is that a bar's popularity will be based off by the
popularity of the bars that link to it. So we by using this idea
we developed a per patient ranking system that could that use that use this concept as well as
factoring and differential expression genes a differential expression for the patient. So we took so by using a
differential expression for by exome
from exome or an RNA-seq data coupled with the network pathway we are able to detect patient specific driver mutations.
So this is the iterative the top on the top is the iterative form of our equation the bottom is the matrix form of our equation.
And it's a little bit complicated to describe but in the most abstract form
I can say is like it requires the several terms one is differential expression.
So this is a differential expression weight here and a connectivity form which is the which is a matrix or
iterative representation of how well gene connects to other to other downstream genes. For example
I would have three connections here, which would be positive a damping factor, which is
which is a parameter that we trained. It's basically the weight between expression and connectivity that we have.
This is this was trained using a previous data sets and we also because it's a stochastic random walk and
it is and we use the rank of the outlinks.
So basically downstream genes the rank of the downstream gene will affect the will affect the rank of the gene itself. So thus
so it's a little bit different than Google PageRank, which was ranked the in links.
However, the ideas are relatively the same.
So it is important to note from this equation that every gene in a possible network will be included.
However, we will only look at genes that were mutated to as a comparison tool.
So to validate our data what we decided to do is we decided to take a look or compare our data with
with MIMO, which was which has been a very popular standard these days. And
from MIMO's data so for so MIMO detected a 12
drivers potential driver mutated genes and
in
glioblastoma and we and we did this over glioblastoma ovarian cancer and breast cancer for TCGA and
from from this from this chart over here
we ranked all these all the genes that MIMO had MIMO detected and we found that and we
for each patient and we took the median of that to just so we could get a nice statistic over here and we found that
the majority of the genes in blue are ranked in the 90th percentile or higher in
in drive rank where and also like all but two are actually ranked in the 80th percentile or higher.
So we have a relatively high concordancy with MIMO.
So after validating our data, we decided to take a look at a frequency analysis.
So we decided to look up how many times a gene was mutated how frequently was significantly mutated with our results.
So we took 92 glioblastoma patients and every time a gene was mutated
we counted that we added that as a total hit and every time it was significantly mutated
we would add that or mutated it within the 90th percentile. It would be considered a significant mutation.
So by by looking at this graph over here
we can tell we can we can roughly classify these driver mutations at three types of they're in three categories.
One is common driver mutation mutations that are relatively commonly known throughout and
there's a consensus in literature about that. Other commonly cancer related genes
these are oncogenes that have a very high potential to become driver mutations as well.
And the third group is the most interesting actually. These are less common.
These are less commonly heard in literature, but that does not necessarily mean they're not driver mutations yet.
So we wanted to take an in-depth look at both the common and less common driver mutations.
So our sanity check over here was TP53 and it's the probably most well-known driver mutation as most people here would have had some
some course that did talked about TP53 in cancer. And TP53 was mutated in our data set about 51% of the time.
But every time it was mutated we detected it as significant, which means that which which is concordant
how other data sets found you by using recurrence.
They also found the same that TP53 was mutated almost every time.
And the reason why it's mutated as you can see in the graph in the graph on the right the center TP53
connects to about 200 different genes.
So it's extremely highly connective and in the genes in green are have
drive ranks of at least in the top quartiles, the top 25%.
And you can see that there's a disproportionate amount of green of green nodes here.
So thus we can see that like because it's both highly connected and highly
and other genes are highly ranked that connect from TP53, TP53 has a very high rank itself.
So furin is the gene that we wanted to look for which is less commonly a concordant with literature and
the reason why we wanted to look at furin because it has a biological significance in cancer because it
has a lot of
leads to protein convertases such as the growth factor that could
lead to a selective advantage as you can see here PCSK and NOTCH in this in this
subnetwork right here as
links to furin quite well and furin appears relatively high in in glial blastoma patients 28% of the time
but it's only significant about 69% of the time and this and this is probably the reason why it wasn't detected by
recurrency methods because it wasn't significant enough times. However, we
however, however
we could because we only use individual patients to do this analysis. This could be an example of a
patient-specific driver mutation and previous literature have from Coppola and Neoplasia actually stated that can't driver mutations actually
that furin has furin inhibiting furin actually decreases the selective advantage of tumor
cells and so in summary drive rank is a patient-specific method to discover drive new driver mutations
however, and it can it can both find the common driver mutations from a per patient sample with very high frequency
but it can also as an advantage to other previous methods find identify rare mutations that are potential drivers on a per patient basis
so if you had if you were a cancer patient with a very rare cancer that no one else has actually
done you recurrency methods would fail in order to detect detector driver mutations
but we would be able to find at least a search engine like tool to figure it out
So we would like to improve the model using it by improving the damping factor using a Bayesian analysis actually
Instead of having it just a constant term and also we wanted to modify not just mutated genes
But also like non-coding regions and other changes as well other genomic alterations
So I would like to conclude by acknowledging Pratik Lahiri who is my partner in this project as well as Ryan Cunningham who helped me
Who helped us develop the drive rank algorithm as well as my advisor Dr. Jam Ma and the entire Ma lab
and also our funding sources the Illinois Distinguished Fellowship the National Science Foundation and the NIH, so
Thank you very much
Do we have any questions?
I guess I just have one practical question for treating patients on a per patient basis
Because it's nice what you've done in correlated expression with the disease
but how expensive do you think that would be for determining per patient treatment and at the end of the day
how do you determine, even if you determine some genetic alteration is that going to be determined?
Is it going to be a driver? How do you determine it?
Well, okay, the thing is that if right now we're limited by the cost of being able to sequence at each patient's genome
and we need exome or RNA-seq data for any patient that we want to categorize
So thus, I guess if we were a business and we wanted to charge for this
I guess our prices would be dependent on how much a next generation sequencing technology would be
But once we get the data and we will be able to have at least a certain
It would be an interesting tool for physicians I think because once we have the data
we have a list of driver mutations that we suspect
So thus, if a physician looks at this and a physician has the biological knowledge to interpret the data
he or she would be able to provide the best treatment
because by pinpointing the specific genes that were mutated that we found actually
Do you think that the driver mutations could change over time as the tumor or the body sort of co-evolves?
So require maybe monitoring in specific intervals?
That is actually an interesting question because a lot of
One interesting discussion I've had before was the fact that
if, for example, one gene was mutated and that gene was, for example, involved in a repair pathway
especially a DNA repair pathway, that gene could actually propagate other driver mutations
So yes, in theory, if we had a time factor, we currently are only using TCGA data
which is not time dependent at the moment
But if we did monitor this over time, we could actually find more information
and we could actually make a stochastic process within a stochastic process
our original stochastic process
and thus actually be able to model it that way
However, unless we have the time
But as the earlier question has said, it would probably require a lot more
That would actually increase the cost a lot more
Currently, next generation sequencing technology, although it is relatively cheap to
relatively cheap to sequence a genome
it's still not in that $1,000 range every day
So you can't just go to Walmart and get your genome sequenced
So if I understand your method correctly
you started off with your seed genes in your network
and then you use your driver genes
and after a number of hits, there's probably a cut-off in significance
that you choose as driver genes
My question is, if you use differentially expressed genes as your driver genes
what is your justification for using those genes as your drivers?
Because those may be differentially expressed
Those actually may be passenger genes to something else at a higher level
which could be control
So I'm not sure how you're
I don't know what you're considering
if you're a false accelerator or a p-value cut-off
but you probably have many differentially expressed genes
Among the many, how are you tuning down what to choose as driver genes
from a non-statistical standpoint as you use
but a mechanistic standpoint inside the network
My second question is, what is your PPI database that you use?
The second question first, we use TCGA data
We use three types of cancer cells
cancers, glioblastoma, ovarian cancer, and breast cancer
So as to actually what our significant cut-offs were
was actually the fact that, first of all
since we are modeling driver mutations, not differentially expressed genes
what we did is actually the first cut-off
before we even used differentially expressed genes
we only looked at genes that were mutated first
A driver mutation must be a mutation, if anything
and with mutations we are able to significantly cut off our gene network
to only about 35 or so genes that we can find
For our network, we rank every gene independently
whether or not it's mutated or not
However, we only care about the genes that are actually mutated
For any particular patient
the datasets that we found from TCGA
I guess the TCGA data came from UNC as well as Broad
and Washington University and Baylor
and a combination of those
The mutation data of all the mutated genes
only a couple of them were actually significant
We used about 10% cut-off
That actually made a very manageable number of 5 or 6 per patient
Going through that, we would actually have a very manageable number of genes
that we suspect to be driver mutations
A lot of them, like p53 and EGFR
are all very commonly documented as driver mutations anyway
The problem with the analysis itself
is the fact that we don't really have a good false discovery rate
but no one can because there's no gold standard in defining a driver mutation
There's no consensus about what is a driver mutation
and what is not a driver mutation
That's actually a very big problem
That's why we develop algorithms like this
in order to find out which mutations are the driver mutations
Thanks a lot