NEW YORK – Large-scale genomic studies like The Cancer Genome Atlas have yielded detailed maps of the somatic mutations that exist in many types of cancers, even some rare ones, allowing clinicians to treat patients more precisely by picking therapies that target specific alterations in the cancer genome.
However, according to William Hahn, a medical oncologist at the Dana Farber Cancer Institute and the Broad Institute, knowing a patient's mutations doesn't always lead to a therapy. For many tumors, sequencing the genome is possible, but interpreting the data is too hard, or drugs don't yet exist to target the mutations that are found.
At the virtual annual meeting of the Association for Molecular Pathology on Friday, Hahn described his and his colleagues' efforts to comprehensively map genes required for the fitness of human cancers using genome-scale approaches. He also gave an update on the Cancer Dependency Map (DepMap) project, an initiative that was launched in 2019 by the Broad, the Wellcome Sanger Institute, and several industry partners to create a comprehensive preclinical reference map that connects tumor features with tumor dependencies.
"If you think about TCGA data as a parts list [for a car], we now know how many parts there are, what kind of parts they are, and even how often those parts are broken," Hahn said in his presentation. "But if we really want to use this information to reverse engineer our understanding of cancer, we have to have other knowledge of what the parts do, so that we can put this back together and understand, like an engine in a car, how a tumor is really working in a patient."
His lab has taken two widely used approaches and performed them at genome scale. One approach involves performing gain-of-function assays, to see if they produce a certain cancer phenotype, and then performing the opposite loss-of-function experiments to suppress or delete genes.
The other approach is to systematically study the cancer genome without focusing on any particular genes identified by genome sequencing studies. In that way, Hahn and his colleagues are aiming to create the cancer dependency map.
As an example of what could be accomplished with the first approach, Hahn showed an analysis that was done in his lab to analyze p53, and answer the question of why hotspot mutations occur more frequently in the gene than at other alleles. Hahn and his team conducted detailed saturation mutagenesis screens to assess p53 function in isogenic cancer cell lines, produced a library comprising all p53 missense and nonsense mutants, and performed selection screens in the presence of p53-activating agents. But while the screens produced a lot of data about p53, they didn't answer the question about the hotspot mutations, indicating that those mutated positions likely don't have a special function that can be measured using the assays the researchers were employing.
Instead, they turned to work that was being done defining specific cancer signatures in sequence data, and applied the technique to p53. They saw the hotspot mutations that had a strong functional impact in their original assays, and noted that it was very likely that those mutated positions were very common mutational processes found in many tumors.
What was interesting, however, was that the team also saw mutations in these analyses that were in p53 alleles that had profound functional impact in the original assays, but that were not the targets of know mutational processes, and in fact are rather scarcely found in most tumors. Conversely, the researchers also found some alleles that had no discernible activity in their assays, but which seemed to be the targets of very common mutational processes and are frequently found in tumors, likely representing the fact that they're being mutated but don't have any functional impact.
"So, if you take this into consideration, you can see that our phenotypic selection model of our screen doesn't match up very well, and if you just take the mutational signatures model that doesn't predict the frequency of mutations," Hahn said. "But if you put those two things together, you come up with a model that is very, very close…. So, I think what this tells us is that, in fact, we now have an answer for the question [of] why are there hotspot mutations in p53. It's not that those alleles have a special function. It's just that they're also the target of common mutational processes. And so, this saturation mutagenesis approach allowed us to get to a definitive answer."
The researchers also decided to do a similar set of experiments with the oncogene KRAS. Here, they took a wild-type KRAS allele and introduced mutations at every position in a saturated way, and looked to see which alleles would transform cells. Opposite to p53, they found that there were about a half dozen positions that were very clearly mutated and had a function in transforming cells. These positions were the ones that were also the most frequently mutated in KRAS, Hahn said.
In looking at the whole set of alleles, he added, one can ask what the correlation is between the strength of the transformation and the frequency of alleles. In the case of KRAS, the analyses showed that the more transforming allele was, the more commonly it was to be found in a tumor.
"I think having this knowledge allows us to think about anticipating resistance as we make drugs, and [to] understand the biology of these oncogenes and tumor suppressor genes," Hahn said.
Importantly, he added, it raises the question of whether this kind of analysis could be done on every cancer gene. While the answer is yes, it would take a massive effort, and would depend on developing very robust assays to try to assess gene function.
According to Hahn, he and his colleagues have thought about whether gene expression signatures could act as proxies for phenotypes, whether they could capture meaningful phenotypes with general assays, and which specific phenotypes would be amenable to this type of approach. There's also the question of whether this approach would be cell-type dependent.
"One thing that occurred to us is that we could use the power of single-cell sequencing to try to identify the patterns of phenotypes," he said. "We've figured out how to do this type of experiment for hundreds of alleles. And now is the question of whether we can scale this to do saturation mutagenesis for any particular gene."
Updating the DepMap
Hahn also talked about the efforts he and his collaborators are making with the DepMap, to take genetic and molecular information on cell lines and tumors, and then perturb them at scale in order to build a comprehensive map of the genes and pathways required for the genesis and existence of particular subsets of tumors.
"[W]e take all that molecular information, and we take the preferential dependencies and we build predictive models that allow us to say not only which cell lines are dependent on particular genes, but what other features line up with that," Hahn said. "Now, when we've done this over many hundreds of cell lines, what we've learned is that, much to my surprise, mutation is not the best predictor of dependency. In fact, gene expression tends to be the best predictor of whatever feature you're looking at. And my guess is that it has to do with the fact that gene expression is an integrator of the phenotypic change in any given cell."
The researchers are using both RNAi and CRISPR screens, and then combining the two kinds of data. They've already analyzed more than 500 cell lines using RNAi and more than 1,000 cell lines using CRISPR.
The work is proving fruitful. For example, one experiment that was done in Hahn's lab found that the gene BIRC6 had a strong selective dependency in a model of breast cancer, that its depletion induced apoptosis in cells, and its suppression caused tumor regression in primary and metastatic tumors in an ER+ breast cancer model. Further analysis showed that three other genes formed a unique ubiquitin ligase with BIRC6, and that this complex is involved in regulating the integrated stress response, which likely explains the differential dependency of certain cell lines and tumors on this complex.
"I think that's a pretty good example that shows the power of how you can get to not only a gene, but a protein complex," Hahn said, adding that the DepMap site contains all the data the collaborators have collected from the RNAi and CRISPR screens, the molecular profiling data, and the small molecule screening data, as well as a raft of data-visualization tools.
Importantly, Hahn believes that clinical labs can benefit from this data as well.
"My vision would be that labs like mine, banding together with others in the world, would come up with ways to prospectively analyze all these variants for a large number of genes," he said. "We have started some discussions with investigators around the world in an alliance of groups that are interested in doing these kinds of studies. And I think that we would ultimately have databases for clinical labs to be able to just look up what's known about every variant as they need, to interpret them."
DepMap has also shown that gene expression data has a place in the clinic, he said.
"When people do the analysis to say what are the most predictive markers, it almost doesn't matter which question is being asked, it almost always comes out that gene expression is the best way to get to those answers," Hahn said. "The power that we've already started to unlock using single-cell RNA sequencing to understand cell types and identify new cell types, identify functions, even identify the receptor ligand pairs within tissues tells us that this is a very powerful way of determining biology by measuring a small number of molecules in each cell."
DepMap data itself will eventually be used in the clinic, he added. Gene dependency data itself may or may not be useful right away as a clinical tool, but one of the things that the researchers are actively engaged in is whether they can delineate complexes or pathways of significant clinical importance through the dependency map, such as the BIRC6 example.
"I think if we can do that, then it may well help clinical geneticists look at mutations, variants of unknown significance, and try to get a sense of whether or not they might contribute [to disease] in some way," Hahn said.
In the short term, he added, DepMap has spawned dozens of drug-discovery efforts, both in academia and industry. "I know of several that are into the clinical trials phase," Hahn said. "So, we'll see how the next few years go."