Skin in the Game: How AI Could Improve Odds of Immunotherapy Treatment

by Gary Rainville

Queensland, Australia, is blessed with an abundance of sun, sea and sand along its seemingly endless miles of golden beaches. Perhaps not surprisingly, the Australian state is cursed with one of the world’s highest rates of melanoma, a relatively infrequent but serious form of skin cancer, as well.

Max Kelsen, a Brisbane-based startup and NVIDIA Inception program member, is employing AI to boost the chances of melanoma patient success using immunotherapy, a treatment that works with the body’s immune system to promote an enhanced response to disease.

Immunotherapy can have life-changing results, but its effectiveness varies greatly. One cancer patient may go into remission and have no evidence of disease, another may find their immune system attacking healthy organs and cells.

Due in part to its high cost and generally low response rates, immunotherapy in many cases has been limited to the treatment of late-stage cancer. Max Kelsen researchers are using AI and genomics to identify reliable markers and develop a test to better predict which patients will benefit from immunotherapy prior to treatment.

Building a Test to Better Predict Treatment

Backed by funding from the Australian federal government, Max Kelsen has teamed with genomiQa, BGI Australia, Metro North Health and Hospital Service, and QIMR Berghofer Medical Research Institute.

Each brings a vital area of expertise to the project — Max Kelsen uses AI to mine big data and explore novel insights, genomiQa specializes in genomic data analysis, BGI is a major supplier of genomic sequencing, and QIMR Berghofer is a leader in immunology and cancer genomic research.

“Our project is looking at genome sequencing, not just single markers,” said Nicholas Therkelsen-Terry, CEO of Max Kelsen. “This project will firmly establish the role of AI and whole genome analysis in the future of precision medicine.”

The application of AI in genomics has been slow due to the small amount of high-quality data with associated health outcomes and the vast compute power needed.

The entire sequence of 3.2 billion genomic base pairs in about 3,000 persons amounts to an insane 30 petabytes of compressed data. Another option is to look at a processed file called a variant call format (VCF) file, which stores about 50MB per person for a total of 150GB of data for the same 3,000 persons.

To compute this data, Max Kelsen is relying on two IBM Power9 systems packed with six NVIDIA Tesla V100 Tensor Core GPUs on each. “With NVIDIA, we can process the size of data necessary to make a real difference in cancer treatment today,” said Therkelsen-Terry.

Using sophisticated AI approaches, Max Kelsen plans to integrate genomic, transcriptomic and patient clinical information to identify a classifier and develop a test of treatment response. The classifier will be developed using genomic data from several melanoma projects. This will then be validated and refined in a second cohort of 400 lung cancer patients collected using routine practice within the Australian health system.

The initial approach is semi-supervised learning involving building general models to interpret the signature of a genome. By framing genomic information as a language, the entire structure of the genome can be analogized to a structured corpus of chapters, paragraphs, sentences and phrases.

“We’ve done a lot of work around a technique where you have an agent that dreams new scenarios and trains on its dreams,” said Therkelsen-Terry. “It’s learning a representation of the real world in dreams or synthesizes false similar scenarios where it is able to create a greater search base than what the existing training data provides.”

Reducing Morbidity, Mortality Rates

Max Kelsen and its partners have three years to work through these problems. They aim to have the first paper on VCF modeling and melanomas out by July 2019.

Ultimately, the goal is reduced morbidity and mortality associated with melanoma and lung cancer by ensuring clinicians are best informed on how to treat patients and who will be most suited to receive immunotherapy. It will then be applied to derive classifiers for other cancer types.

“Far beyond whether a single tumor is going to respond to immunotherapy, I want to set deep learning as a key tool in the toolbox of the future of medicine and the professionals working as a way to solve some of the very big problems we have today,” said Therkelsen-Terry.