New model reveals cancer-causing mutations

In this interview, News-Medical speaks with Maxwell Sherman, an MIT graduate student and one of the lead authors of a study that used a new method to examine cancer genomes.

Can you introduce yourself, tell us about your scientific background and what inspired your latest research?

We are a multidisciplinary team of computer scientists, mathematicians and biologists fortunate enough to work in the ecosystems of MIT and Harvard. Most previous work to identify mutations that cause cancer emergence and progression has focused on the 2% of the genome that codes for proteins. We wanted to enable the cancer research community to search 100% of the genome for mutations that can cause cancer.

Cancer cells can have thousands of mutations in their DNA. What is the difference between a mutation that promotes cancer progression and a relatively neutral mutation?

Cancer can be understood through the lens of Darwinian evolution. Driver mutations allow a cell to grow and divide faster, producing more cells for progeny. Cancer is a result of this cellular race: once a cell has accumulated enough of these driver mutations, it can divide indefinitely, escape the immune system and eventually spread to other tissues, all hallmarks of cancer. On the other hand, neutral “passenger” mutations are mutations that do not affect a cell’s ability to grow or reproduce and thus play no role in the Darwinian evolution of cells. The vast majority of somatic mutations in our cells appear to be neutral.

What do we currently know and do not know about mutations that cause cancer?

This is a big question that is difficult to answer both succinctly and accurately. Suffice it to say, decades of research have uncovered major causes of numerous cancers, leading to many breakthroughs in medicine’s ability to treat patients in the clinic. Yet there is still a huge amount that we do not know. We don’t know the full spectrum of driver mutations in the non-coding genome, unraveling all the complexities of copy number variation (shout out to the recent Nature papers making huge strides on this), or the role of repeated extensions . But there’s undoubtedly so much we don’t know waiting to be discovered.

With a new model you could scan the genome of cancer cells. Can you describe the model and what new insights it has yielded?

Our model uses a deep learning procedure to map the genome-wide somatic mutation rates for a cancer of interest. It then uses a custom probabilistic model to search those maps almost instantly to estimate the number of passenger mutations that should be in a particular region of the genome.

Our approach has several key features: 1) a mutation rate map only needs to be trained once for a particular cancer type (and we already have trained and publicly available maps for 37 cancer types). It can then be applied to any cohort of patients of that tumor type; 2) users have the flexibility to specify regions anywhere in the genome down to the resolution of a single base pair; 3) our model is fast and efficient enough that users can complete a genome-wide analysis on a PC in minutes.

One type of non-coding mutation you focused on was cryptic splice mutations. What are cryptic splice mutations and how do they cause cancer?

Cryptic splice mutations are mutations that occur far from the boundaries of a gene’s exons, but nevertheless confuse the cellular machinery responsible for splicing introns and rejoining the exons. These mutations thus lead to an incorrect splicing of the gene. This usually results in nonsense mRNA transcripts that the cell simply recycles or a non-functional protein. Either way, the right protein product of the gene is not made. Tumor suppressor genes generally inhibit cell division, preventing a cell from dividing uncontrollably. Cryptic splicing mutations can render these genes non-functional, taking away the cell’s own defenses against cancer.

With this new model, you could also look at known cancer-causing mutations. What have you learned about these mutations within the 37 different cancer types you’ve studied?

We found that genes that often cause one type of cancer can occasionally cause other types of cancer. Dig’s construction was key to this understanding. Because our model can be trained across one set of patients and applied to another set, we were able to pool thousands of patient samples from heterogeneous sequencing studies, providing the statistical power needed to investigate these rare events.

Considering that your model used a deep neural network, a kind of deep learning, how do you see kinds of machine learning that will influence cancer research in the future?

As the field generates greater data in sheer size and complexity, the need for tools that can automatically parse and extract meaning from these data sets will only increase. Machine learning algorithms can provide a powerful approach to this challenge. It can be particularly powerful for generating and prioritizing data-driven hypotheses about molecular mechanisms, which can then be explored experimentally, an approach that the field (including our lab) is increasingly embracing.

How might the findings of your research and the model itself influence the future development of cancer therapies?

We hope that the cancer community will make important discoveries about the biology of cancer by examining the non-coding genome. Each discovery has the potential to open new avenues for therapies. Our model is a tool to help the cancer community do that.

What is the future for you and your research?

We’re working on some exciting new stuff that we’re excited to share soon.

Where can readers find more information?

About Maxwell Sherman

I am currently a fourth year Ph.D. candidate jointly supervised by Professor Bonnie Berger and Professor Po-Ru Loh. My research focuses on developing algorithms to uncover the role of somatic mutations in human health and disease.