May 20, 2021
This spring, researchers from across Ontario gathered virtually at the 2021 OICR Translational Research Conference (TRC), including interdisciplinary experts in the use of machine learning (ML) and artificial intelligence for cancer research and eventual clinical implementation. During the Conference, a collaborative session was held to identify current challenges and opportunities in using ML within the clinic and identify what a path forward would look like, to accelerate the use of this technology in Ontario.
The discussion addressed various steps in realizing this vision, including improving data collection, developing and validating machine learning models, and facilitating the transition of these systems into clinical use. Panelists examined opportunities to share data, create common standards for ML model development and cited a need to foster cross-disciplinary collaborations to advance ML research in Ontario. The session also reflected on tempering expectations around the use of ML in cancer care, and how the field could deliver tangible results while cutting through the hype surrounding the technology and addressing data bias.
“The opportunities for machine learning in precision cancer medicine are immense, but we face hurdles at various points in the workflow, which limits our ability to transition discoveries into clinical use,” says Dr. Shraddha Pai, Principal Investigator, OICR and co-moderator of the session. “This session allowed those involved in the space to connect and reflect on what’s needed to map out a path forward and help make routine use of machine learning for cancer care a reality in Ontario.”
Among key challenges identified by the group were the lack of well-curated datasets that accurately represent relevant patient populations, a problem fueled by data silos and limited cross-disciplinary conversations to define problems in a manner that facilitates collecting the right data for model-building. A list of key takeaways was generated from the session.
“There are real logistic hurdles around disconnected data and disciplines, which faces researchers who want to develop machine learning models for pre-clinical biomarker discovery and eventual clinical implementation. However, if we manage to fix these issues, Ontario’s system offers features to our advantage,” says Pai. “For example, compared to parts of the U.S., Ontario has the significant advantage of having a universal healthcare system which links patient data through various clinical settings. When set up correctly, data from the OHIP (healthcare system) program, when integrated with biological patient profiles like genomic and imaging data, could accelerate the development of more accurate models representative of the patient population. To realize the value of these data to create impactful clinical tools, we need to break down barriers to data sharing and interdisciplinary collaboration, and create partnerships to identify existing biases in data collection and ensure they don’t lead to the development of misleading models.”
The session included some of Ontario’s foremost experts on machine learning in the context of cancer care:
- Dr. Anne Martel, Sunnybrook Research Institute, Senior Scientist (Machine learning in clinical imaging).
- Dr. Michael Hoffman, Princess Margaret Cancer Centre, Senior Scientist (Machine learning in genomics; liquid biopsy; member of Temerty Centre for AI Research and Education in Medicine)
- Dr. Harriet Feilotter, Queen’s University, Professor, Department of Pathology and Molecular Medicine (Clinical genomics; Ontario Health Data Platform).
- Dr. Amber Simpson, Associate Professor, Department of Biomedical and Molecular Sciences and School of Computing, Queen’s University (Machine learning in clinical imaging; Ontario Health Data Platform).
- Dr. Michelle Brazas (co-moderator), OICR, Senior Program Manager.
- Dr. Shraddha Pai (co-moderator), OICR, Principal Investigator (Machine learning and genomics; member of Temerty Centre for AI Research and Education in Medicine).
Key takeaways from the session can be found here: http://pailab.oicr.on.ca/assets/docs/2021TRC_AIinML_KeyTakeaways.pdf
January 8, 2021
Q&A with new OICR Investigator Dr. Shraddha Pai on uncovering the hidden differences between cancers
OICR is proud to welcome Dr. Shraddha Pai to its Computational Biology Program as a Principal Investigator. Here, Pai discusses current challenges in understanding diseases and what motivates her to tackle some of the biggest challenges in biomedical research.
What are some of the research questions you’re interested in?
I’m very driven to understand why different people with the same cancer type, have different outcomes and respond differently to the same treatment. As genomic assays get cheaper, we learn more about molecular interplay in different cells, and our population datasets become larger and mature, we are able to integrate different layers of the genome and cell types, to try to get at this question. For example, we now believe there are four main types of medulloblastoma with different underlying molecular networks and outcomes. This field of research is called ‘precision medicine’: using patient profiles to match them with the most effective treatment. But really this is just a new phrase to describe what doctors have been doing since the dawn of medicine; it just means that now we’re using powerful computers and algorithms to find patterns in much larger and complex genomic datasets. The principle is the same.
As a trainee in Dr. Gary Bader’s group, I led the development of an algorithm that integrates several types of patient data to classify patients by outcome. Our method – called netDx – adapts the idea of recommender systems, used by Netflix and Amazon, to precision medicine. Just as one would ask Netflix to “find movies like this one”, netDx helps identify patients “with a treatment profile like this”. In a benchmark, netDx out-performed most other methods in predicting binary survival in four different types of cancer. Importantly, netDx is interpretable, and recognizes biological concepts like pathways. This makes it a useful tool to get mechanistic insight into why a predictor is doing well, and provides a way to understand the underlying biology and perhaps drive rational drug design.
I also have a special interest in understanding the link between epigenetics and disease, particularly as this pertains to the brain. Epigenetics refer to molecular changes that change how the genome behaves – for example, turning a gene on or off in a given cell type. My own previous research in mental illness has found epigenetic biomarkers related to psychosis, which explain the distinctive features of this condition. The same may be the case in certain types of cancers, particularly those of developmental origin.
How do you plan to unravel these complex layers of biology?
My research program has two main goals. The first is to build models for precision medicine – predicting disease risk, treatment response – starting with population-scale datasets that have several types of patient data. I’m hoping to use existing and emerging data such as UK BioBank, CanPath, ICGC-ARGO and the Terry Fox Research Institutes’ datasets, and ongoing clinical trials, to identify which clinical outcomes are easily amenable to our approaches. The models my group builds will incorporate prior knowledge about genome organization and regulation, so that these are interpretable. For example, we will use epigenomic maps of specific tissue types, or data from single-cell resolution maps, pathway information, to find and organize relevant needles in the genomic haystack. This feature will give us interpretability, which is key to increasing confidence in a model, as well as to improving the understanding of cellular pathways that affect disease and eventual drug development.
My second goal is to understand the epigenomic contributions – particularly developmental changes – to cancer risk, using a combination of molecular biological, genomic and analytic techniques.
As I work toward these goals, I hope to collaborate on complementary projects, such as identifying DNA methylation changes in circulating tumour DNA and improving how we subtype adult tumours. These projects will hopefully lead to new biomarkers, and ultimately improvements to how we diagnose and treat cancer.
Importantly, the software that my team builds will also be openly available to the research community, so others can apply my methods to different types of diseases. I’m excited to get started.
Your work applies beyond cancer. How do you traverse these different disease areas?
The reclassification of disease based on molecular or other biomarkers, and how disease subtype affects risk and treatment response, isn’t unique to cancer – the same research questions extend to other types of disease such as metabolic diseases, autoimmune diseases and mental illness. At the end of the day, we are looking at the same system organized at the molecular, cellular and organ-level, with similar principles of genomic regulation and perhaps similar considerations for drug discovery. Our algorithms are based on these general principles and can therefore be used to answer similar questions for different disease applications, or very different types of cancer. Of course, it’s important to collaborate with teams that have domain expertise to make sure the algorithms are “fine-tuned” for a particular application, and I look forward to benefitting from those partnerships.
What excites you about this type of work?
I’m excited to join a community where basic research is so strongly connected to clinical purpose. Personally, I am very motivated by the prospect of a positive impact on patients within my lifetime and feel that my group’s work is more likely to have a valuable impact in an environment that combines basic and translational research. That said, we’re only just beginning to see the benefits of precision medicine and many challenges remain to bring genomic knowledge into practice. I hope that I can create more useful methods and models for precision medicine and improved clinical decision-making in the coming decade.
I’m especially excited to be at OICR because of the Institute’s access to clinical trials, strong genomics and computational biology program, and pharmacology team. If my group can find promising biomarkers and leads, we can work with OICR collaborators in the Genomics and Drug Discovery groups to move from basic research to application.