By: Jon Farrow
26 Mar, 2020
Cancer is the world’s second leading cause of death, responsible for approximately one in six deaths. It’s a challenging disease to diagnose and treat because every tumour is different. Cancer can take many routes as it evolves and spreads. New research uses AI to reconstruct that path in order to determine a cancer’s causes and most likely effective treatments.
Quaid Morris is a Canada CIFAR AI Chair at the Vector Institute and a Professor at the Donnelly Centre at the University of Toronto whose lab applies machine learning methods to biological problems like genetics and cancer. Morris is a lead author, along with Lincoln Stein at the Ontario Institute for Cancer Research, on two recent papers in Nature Communications which analyzed the genomes of over 2,600 tumours to identify common patterns as part of the Pan Cancer Analysis of Whole Genomes (PCAWG) consortium.
Cancer occurs when mutations in cells cause them to grow and divide unchecked. Often, the mutations deactivate a cell’s normal DNA-repair mechanisms, which Morris believes may be the key to tracing a cancer’s source. Because the cancer cells are unable to repair their DNA, mutations accumulate, like a record of everywhere the cell has been.
“Cancer cells evolve from normal, healthy cells to cancer cells through accumulation of driver mutations that give them a selective advantage,” says Morris. “And while they are gaining these driver mutations, the cells are also picking up thousands of passenger mutations.”
Morris and his collaborators believe these passenger mutations, so called because they aren’t actively involved in making the cancer grow and divide rapidly, reflect the environment of the cell. Exposure to carcinogens like tobacco smoke, and the conditions of the tissue where the cell lives, will leave distinct signatures of mutations. Researchers like Morris can use these signatures to find out where the cancer started in the body, and why.
The mutational signatures are complex. With billions of DNA base pairs in the human genome and many different ways each one could change, there is a lot of data to analyze. Machine learning thrives on data.
In the first paper, Morris and his team used a deep learning system to identify the source tissue of cancer.
Metastasis, when a cancer spreads from one tissue to another, is common, but a cancer’s susceptibility to treatment depends on its tissue of origin. As Morris and his team report in the paper, three per cent of patients present with metastatic cancers without knowing where the cancer started. Even to many trained pathologists, metastatic tumours are difficult to distinguish from primary ones.
Morris’s team, knowing that a cancer’s journey through the body will be reflected in its passenger mutations, trained a deep learning algorithm to classify tumours based on their pattern of mutations. First they showed the system a labelled set of data from the thousands of tumour genomes available through PCAWG. Once the system was trained, they presented the genomes of tumours the system hadn’t seen before and asked it to classify the tumour as one of 24 types.
The algorithm passed the test with flying colours, correctly identifying 83 per cent of the metastatic tumour origins. This is significantly better than the average 50 per cent human pathologists are able to accomplish.
Funding and capacity-building through the Canada CIFAR AI Chairs program allowed Morris to expand the range of AI tools he could use. “We used deep learning techniques that we normally wouldn’t have used in my lab,” he says. “Through the CIFAR funding and through our connection with the Vector Institute, we became interested in these [deep learning] techniques and I think it led to some excellent work.”
The second paper, which demonstrates a piece of software called Tracksig, traced the evolution of cancer based on the mutations it accumulates.
“These passenger mutations don’t have any functional impact,” says Morris. “But they’re not fully random, because the type of mutation your cells acquire depends on what caused the mutation.” Morris and others in this field have found that generally the carcinogen – the substance that caused the cancer – will leave its signature early on. By identifying the timing of the mutations, they can get closer to finding the cause.
TrackSig uses machine learning to find patterns in the mutations and identify the likely timing of mutations in a tumour, essentially tracing its evolutionary history. Not only does this provide a clue to the cancer’s cause, but also informs treatments. “A number of cancer therapies target elements of the DNA damage repair pathway,” says Morris. “If we can see the impact of the absence or deficit of the DNA damage repair pathway [using TrackSig], that gives us a secondary signal that can help us interpret whether or not someone might be more responsive to the treatment.”
Closer ties to the machine learning community through the Canada CIFAR AI Chairs program has helped Morris learn about the latest techniques and recruit talented students to help apply them to biological problems. “[The Canada CIFAR AI Chairs program] has certainly increased the transfer of technology into clinical settings,” he says.
Morris believes that combining advances in genomics and biomedical science with the latest computer science techniques will transform cancer from an acute and deadly disease to a chronic, long term one. For him, it comes down to understanding the patterns in DNA. “The biggest problem of the century is figuring out what our DNA means and how our cells work,” he says. “And I think that problem can actually be solved in our lifetime.”
Quaid Morris is a Canada CIFAR AI Chair at the Vector Institute.
The Canada CIFAR AI Chairs program is the cornerstone program of the CIFAR Pan-Canadian AI Strategy. A total of $86.5 million has been earmarked for this program.
The goal of the Canada CIFAR AI Chairs Program is to recruit and retain in Canada some of the world’s leading researchers in AI and provide them with long-term, dedicated research funding to support their research programs, and help them train the next generation of AI leaders.