Health & Medicine Dr Benjamin Haibe-Kains ︱ ARTIFICIAL INTELLIGENCE AND to improve prediction recognised that, due to the high costs CANCER RESEARCH of patient survival and response to of clinical trials, the existing clinical data Artiicial Intelligence and their underlying therapies. During his graduate training related to given treatment and cancer Computational machine learning algorithms could be at the Université Libre de Bruxelles in subtypes were extremely scarce. Given of particular value for research exploring Belgium, he worked on developing that machine learning usually requires complex diseases, when trying to identify predictors of survival in breast cancer a large sample size to avoid artefactual effective pharmacological treatments for patients based on high dimensional gene discoveries, it was time to investigate methods of researching them. expression (messenger RNA) data. He “preclinical models”, which are cancer continued his research as a postdoctoral cells derived from patient tumours that Cancer, one of the leading causes of fellow at the Dana-Farber Cancer one can replicate ininitely. These models cancer treatments death worldwide, is a perfect example of Institute/Harvard School of Public Health, therefore provide a fantastic advantage this. Scientists have not yet been able to leveraging a large collection of data to compared to clinical trials as the same identify a systematic treatment for cancer develop “gene expression signatures” to cancer cells can be tested with multiple Artiicial Intelligence (AI) and machine learning algorithms have the he computational analysis of big that is successful in curing the disease in robustly identify molecular subtypes of therapies to assess which one is the most potential to bring substantial advances in the ields of research exploring data is set to bring huge advances many of its most aggressive subtypes. breast and ovarian cancers. eficient, something impossible to do complex diseases and trying to identify effective treatments. Dr Benjamin Tin terms of the methods used to Cancer derives from an uncontrolled with patients. Haibe-Kains, working at The Princess Margaret Cancer Centre in Toronto, carry out medical and basic research. A division of abnormal cells in a given Using large collections of cancer has spent over a decade developing machine learning tools and databases number of researchers worldwide have part of the body, which can invade and molecular data and machine learning Dr Haibe-Kains’ laboratory invested most that could help scientists gain a better understanding of different sub-types already started developing algorithms destroy surrounding healthy tissue algorithms, Dr Haibe-Kains and of its resources in compiling and curating of cancer and their response to pharmacological treatments, in order to that can analyse large amounts of and organs. It is an extremely complex his collaborators identiied several the largest anticancer drug screens identify more effective drugs for individual patients. biological and medical data, identifying disease, with more than 200 subtypes, “prognostic biomarkers”, that are in preclinical models. Such screens new hypotheses for future investigation. each of which is contain not only the A particular ield of medical research often diagnosed and Dr Benjamin Haibe-Kains explores genomic make-up that could highly beneit from the use treated differently. of the cancer cells, of machine learning algorithms is that of As cancer arises from the potential of machine learning but also the way exploring cancer genomics and trying to aberrations in the these cells react to determine the best treatments for different genomic materials algorithms, and chemical treatments; subtypes of the disease. of the cells, scientists these complex data have developed computational genomics to improve are referred to as THE VALUE AND CHALLENGES OF sophisticated prediction of cancer patients’ survival . BIG DATA proiling These efforts Throughout the years, scientists have technologies to and response to therapies resulted in the collected vast amounts of data from measure these development experiments aiming to achieve a aberrations and use them to personalise computational models using speciic of PharmacoDB1, a web-application better understanding of the molecular therapies. Still, the most common molecular features to predict the allowing researchers to quickly access the dynamics behind complex diseases. treatments for cancer are chemotherapy, probability of survival of cancer patients pharmacogenomic data to investigate However, the complexity of such data using drugs to kill the most proliferative treated with standard-of-care treatments. the possible associations between makes it dificult for human researchers cells, and radiotherapy, using high energy He says: “Amongst many discoveries, genomic aberrations and drug response. to run in-depth analyses and extract the X-rays. These treatments can sometimes my research unravelled the landscape of relevant information. This is where the be successful in reducing or eradicating cancer pathway activities associated with As part of his current research, Dr use of high performance computers, cancerous cells, yet they can be highly patients’ survival in each of the molecular Haibe-Kains is testing machine learning coupled with the right programs could toxic and are not tailored toward the subtypes, allowing me to further improve algorithms on the database, to try be of great help. In the last decades, speciic set of genomic aberrations that my molecular prognostic models.” and pinpoint predictors of treatment computer scientists have worked hard make each tumour unique. reaction. Paired with the right machine to develop machine learning algorithms EFFECTIVE TREATMENTS FOR learning algorithms, PharmacoDB could – programs allowing machines to Artiicial Intelligence tools could assist INDIVIDUAL PATIENTS be used to develop an AI tool assisting quickly analyse large amounts of data researchers by analysing the complex In 2012, Dr Haibe-Kains started his own in selection of the most effective and “learn” models useful to identify genomic make-up of each individual independent laboratory, broadening treatments for each individual cancer the relevant pieces of information and tumour to develop accurate predictors his ield of research to explore ways in patient. Dr Haibe-Kains’ laboratory also make predictions. These algorithms can of treatment response. This would in turn which machine learning algorithms could leveraged these valuable data to address then be leveraged to develop artiicial help to identify more effective treatments be used to predict therapy response another important issue in cancer intelligence (AI) tools that can assist for individual patients, a major step in patients. With big players like IBM research: how to classify drugs based humans to analyse data that are beyond towards personalised medicine. announcing its Watson initiative to on their mechanism of action? Although our reach. As they become increasingly develop AI to assist oncologists in their biologists, chemists and pharmacologists advanced and sophisticated, AI tools are COMPUTATIONAL METHODS FOR treatment decision process, it was an teamed up to develop a large portfolio of opening up a new world of possibilities CANCER RESEARCH exciting opportunity for Dr Haibe-Kains drugs with high anticancer potential, it is for big data analysis, by transforming the Throughout his career, Dr Benjamin to extend the biomarker discovery unclear for many of these drugs how they way in which studies and investigations Haibe-Kains has explored the potential beyond patients’ survival, and to make actually inhibit the growth of cancer cells. are conducted, leading to important of machine learning algorithms, personalised medicine a reality. However, Dr Haibe-Kains and his team developed discoveries in a much shorter time. bioinformatics and computational Dr Haibe-Kains and his team quickly the Drug Network Fusion (DNF), a

www.researchoutreach.org www.researchoutreach.org Behind the Bench Dr Benjamin Haibe-Kains

E: [email protected] T: +1 416 581 8626 W: http://www.pmgenomics.ca/bhklab/ https://www.linkedin.com/in/benhaibekains/ https://twitter.com/bhaibeka

Research Objectives Research Society, Ministry of Economic postdoctoral fellowship at the Dana-Farber Dr Haibe-Kains’ research focuses Development/Ministry of Research and Cancer Institute and Harvard School of on the computational integration of Innovation of Ontario, Ontario Institute for Public Health (USA). high dimensional molecular data to Cancer Research simultaneously analyse multiple facets of Contact carcinogenesis. Bio Benjamin Haibe-Kains, PhD Dr Benjamin Haibe-Kains is Scientist at Princess Margaret Cancer Research Tower, Funding the Princess Margaret Cancer Centre Floor 11, Room 310 Princess Margaret Cancer Foundation, and Assistant Professor in the Medical 101 College Street Canadian Institutes of Health Research, Biophysics and Computer Science Toronto Terry Fox Research Institute, Stand Up To departments of the University of ON M5G 1L7 Cancer Canada, Cancer Research Society, Toronto (Canada). He earned his PhD in Canada Natural Sciences and Engineering Research Bioinformatics at the Université Libre de Council of Canada, Canadian Cancer Bruxelles (Belgium), before conducting a

for the scientiic community. With these better schedule the series of medical Q&A large amounts of pharmacogenomic data appointments involved in the treatment in hand, we applied machine learning of complex diseases such as cancer. It When and how did you irst start techniques to better classify drugs (DNF4) will help hospitals better monitor their being interested in computational and discover new biomarkers predictive performance, and ensure the highest models to be used in cancer genomics of drug response in cancer cell lines5. standards of safety, diagnosis accuracy research? Even though these discoveries will have and treatment eficacy. The mining of When I did my bachelor in Computer to undergo further validation before data from wearable devices will allow Science, I was interested in the their translation into clinic, they show that continuous monitoring and patients development of Artiicial Intelligence in machine learning can yield promising to be more proactive when symptoms robotics. In the early 2000s, AI robotics results with potential clinical relevance. arise. And of course, AI will enable new new technique integrating multiple basic and translational cancer research Dr Haibe-Kains’ approach to research is was still in its infancy and the applications discoveries, further expanding our pharmacogenomic data to design a could mark the beginning of a new era highly collaborative and multidisciplinary, were still limited. This is when my former How long do you believe it might take to knowledge of cancer and other diseases. start witnessing a major introduction of comprehensive drug similarity map for personalised medicine, characterised merging the expertise of scientists from supervisor at the Université Libre de Limitless. Bruxelles, Prof Gianluca Bontempi AI technology within medical settings? (or taxonomy). DNF allows researchers by quick and advanced data analysis, a number of different ields. In future, advised me to consider bioinformatics, Not long, probably a few years from now. What are your plans for future to assess the similarity between a a booming ield in need of researchers As more hospitals have the patients’ research? drug with unknown mechanism of with expertise in machine learning. I electronic health records connected to First, I want to validate our irst action with well-characterised drugs, Application of artiicial intelligence in followed his advice and started a PhD the experimental data derived from their discoveries – predictors of drug therefore providing an eficient tool basic and translational cancer research under a co-supervision with Dr Christos tumour materials, we will inally have access response and new drugs predicted to identify potential new indications Sotiriou, a breast cancer oncologist at the to the large, high quality data required for to be eficacious in aggressive cancer Institut Jules Bordet. This is how I started AI to show its full potential for biomedical types – in animal studies to get them for approved or experimental drugs could mark the beginning of a new era using computational models in cancer applications. Recognising that large close to clinical applications. The Princess (a process called “drug repurposing”). genomics research. cohorts of patients and derived materials Margaret Cancer Centre enjoys a strong As more pharmacogenomic data for personalised medicine are necessary to make major discoveries drug development group, which will be becomes available and machine learning So far, how effective have the and build the new generation of AI-based key in this endeavour. Second, I want to algorithms are further improved, which was previously unattainable. Dr the computational methods developed machine learning tools developed tools in the clinic, hospitals are joining make our pharmacogenomic platforms databases such as PharmacoDB could Haibe-Kains’ work is a perfect example of by him might lead to ground-breaking by you been in improving biomarker forces, and even pharmaceutical companies tools of choice for hospitals and become extremely valuable resources. this, as he has introduced computational discoveries, which could inform discovery and drug selection from have started to share more and more pharmaceutical companies, to further pharmacogenomic data? data related to clinical trials. These are the facilitate data sharing and large-scale methods that could speed up cancer oncologists on how to select the most My lab was deinitively not the irst to necessary steps we must take to unleash the computational analysis. Finally, I want to A NEW ERA FOR RESEARCH research signiicantly, by analysing large effective treatments for individual cancer tackle these important challenges. power of AI for personalised medicine. integrate molecular data with imaging If developed and used correctly, the datasets and identifying predictors of patients in clinical settings. However, we built on our previous data, both pathological and radiological use of machine learning and AI tools in treatment response. experience in meta-analysis of large In the years to come, what role do you images, to predict the best course compendium of gene expression data feel AI will have in terms of research and of treatments and better follow the to develop computational platforms patients over time. Our irst application 1 http://pharmacodb.pmgenomics.ca/ 2 PharmacoGx: An R package for analysis of large pharmacogenomic datasets. Smirnov P, Saikhani Z, El-Hachem N, Wang innovation within the medical ield? D, She A, Olsen C, Freeman M, Selby H, Gendoo DM, Grossman P, Beck AH, Aerts HJ, Lupien M, Goldenberg A, Haibe-Kains B. Bioinformatics. 2015 Dec 9. 3 for pharmacogenomic data analysis. It is hard to predict the roles of AI in the of deep learning on radiological images PharmacoDB: an integrative database for mining in vitro anticancer drug screening studies. Smirnov P, Koia V, Maru A, Freeman M, Ho C, El-Hachem N, Adam GA, Ba- PharmacoGx2 and PharmacoDB3 future of medicine; the applications are is promising, supporting this line of alawi W, Saikhani Z, Haibe-Kains B. Nucleic Acids Res. 2017 Oct 9, gkx911. 4 Integrative cancer pharmacogenomics to infer large-scale drug taxonomy. El-Hachem N, are open-source and freely available close to limitless. It will help patients research for future clinical applications. Gendoo DM, Soltan Ghoraie L, Saikhani Z, Smirnov P, Chung C, D eng K, Fang A, Birkwood E, Ho C, Isserlin R, Bader G, Goldenber g A, Haibe-Kains B. Cancer Res. 2017 Mar 17 5 Gene isoforms as expression-based biomarkers predictive of drug response in vitro. Saikhani Z, Smirnov P, Thu KL, Silvester J, Lupien M, Mak TW, Cescon D, Haibe-Kains B. Nat Commun 2017 oct, in press.

www.researchoutreach.org www.researchoutreach.org