Microsoft develops AI to help cancer doctors find the right treatments

Seattle: There are hundreds of new cancer drugs in development and new research published minute to minute, helping doctors treat patients with personalized combinations that target the specific building blocks of their disease. The problem is there’s too much to read and too many drug combinations for doctors to choose the best option every time.

Enter a Microsoft Research machine-learning project, dubbed Hanover, that aims to ingest all the papers and help predict which drugs and which combinations are most effective, according to the company.

Researchers at Oregon Health & Science University’s Knight Cancer Institute are working with Hanover’s architect, Hoifung Poon, to use the system to find drug combinations effective in fighting acute myeloid leukemia, an often-fatal cancer where treatment hasn’t improved much in decades. They include Jeff Tyner, and the institute’s director, Brian Druker, best known for pioneering Gleevec, a blockbuster drug for a different type of leukemia now owned by Novartis, that’s helped double those patients’ five-year survival rate since the 1990s.

Cancer is caused by genetic mutations that make cells grow and multiply out of control. Better ability to find those specific mutations has enabled new types of drugs that target the disease more precisely, raising the odds of survival. There are more than 800 medicines and vaccines in clinical trials to treat cancer, according to a 2015 report by Pharmaceutical Research and Manufacturers of America. At the same time, the rising speed and falling cost for sequencing genes has boosted research, the development of therapies and means more cancer patients can gain exact data on their case.

“It’s exciting, but it also provides us with the challenge of then what to do with all the data,” Knight Institute’s Tyner said. “That is where the idea of a biologist working with information scientists and computationalists is so important. The combination of all those resources is going to help make the ultimate breakthroughs for more effective, less-toxic therapies.”

In a recent interview, Poon showed a slide of a melanoma patient with tumours pocking nearly every inch of his chest. In a second photo, the lesions completely cleared up after a targeted therapy. But a third showed most of the lesions returned a few weeks later after a different mutation enabled the cancer to roar back to life. A combination of targeted drugs could work here, Poon said, but how to find it?

“There are already hundreds of these kinds of specifically targeted drugs, so even if you think let’s pair two drugs there are tens of thousands of options,” he said. “It’s very hard to wrestle with. You might need several drugs to lock down all of the tumour’s pathways.”

Hanover, a nod to a seminal 1956 AI workshop at that city’s Dartmouth College, is part of several projects announced Tuesday by Microsoft to develop computational approaches to better cancer care and study. Other projects involve a machine-learning and computer-vision system to help radiologists understand tumour progress and an effort that could one day allow scientists to program cells to fight disease.

Machine learning, where computers use data to glean insights without being explicitly programmed by humans, is increasingly aiding cancer research by parsing data in research papers, results from clinical trials, radiology reports and electronic medical records. International Business Machines Corp.’s Watson Oncology system is helping doctors interpret clinical data and develop individualized treatments. Google’s DeepMind Technologies Ltd has a medical unit and is working with the UK government health service to study whether computers can be trained to spot degenerative eye problems early enough to prevent blindness. Start-up Deep 6 Analytics in Pasadena, California, mines unstructured data like health records to find candidates for clinical trials of new drugs. Flatiron Health, backed by a Google VC unit, compiles research and patient information from cancer centers into a database and uses that to make clinical trials more efficient.

Poon’s Hanover wants to augment the work of so called tumour board reviews, where a number of doctors gather to discuss the best treatment option for patients.

“One of the bottleneck’s right now for the tumour board is to understand all this knowledge and how can you extrapolate,” he said. “This is what people have to deal with unless we can automate that process.”

Mark Craven, a professor of Biostatistics and Medical Informatics at the University of Wisconsin-Madison, has used the previous iteration of Poon’s work for research on genes potentially related to breast cancers that are resistant to the most common types of treatments, called triple-negative breast cancer.

While these kinds of machine-learning based approaches are important to new cancer care, challenges remain, according to Anil Goud, a medical director who works on things like clinical information at Cedars-Sinai hospital in Los Angeles. One is getting the information gleaned by machine-learning software into the hands of clinicians and into their regular workflow. And while some health insurers can be convinced to cover new and varied combinations of drugs if the research supports that, overall this is new territory for the insurance industry, Goud said. Researchers will also need to find enough patients who are appropriate candidates to make sure varied and new combined therapies work. He’s still hopeful though. “In oncology, it’s almost impossible for us to think this is doable without machine learning,” Goud said. “The amount of data to be able to sift through and understand what’s significant would have taken much, much, much longer if we hadn’t had this.”