Aim Reduce

AIM    REDUCE program:

stands for An Intelligent Machine-learning Model for Real-time Early Detection of Undesirable Cancer Events

  • Project 1: Natural language processing (NLP) methodologies to extract accurate and detailed information from radiology reports

  • Project 2: NLP methodologies to identify the presence of precancerous lesions in surgical pathology reports


Dr. Geoffrey Liu Lab–Lung Cancer Research Summary


Clinic-Translational Focus


The lung cancer laboratory of Geoffrey Liu at Princess Margaret Cancer Centre-University Health Network is focused on improving lung cancer screening and effective treatment strategies of all stages of lung cancer, using cell lines, mouse models, patient samples, population and epidemiological methods. Early Detection of Lung Cancer Using New Technologies:Lung-CALIBRE Lung Cancer And Liquid biopsy, Informatics, Breathomics, Radiomics forEarly detection. Lung cancer screening by low dose CT works well for detecting lung cancer early among traditionally high-risk individuals(older people with a significant smoking history).Unfortunately, more than half of lung cancer patients treated at Princess Margaret are lifetime light-or never-smokers, who do not qualify for CT screening. As a result, the majority of these patients do not discover their cancer until it has advanced to late stages.This includes the roughly3-7% of lung cancer patients with anALK molecular rearrangement(ALK+),which disproportionately impacts never-smokers and younger patients (i.e.,under 60 years old).


The overarching goal of this research program is to improve lung cancer survival rates by developing and improving early detection strategies. We are actively testing all innovative approaches we can get our hands on. Some of our current projectsi nvolve:

(A)Radiomics: Image analysis to predict lung cancer risk from low dose screening CT scans;

(B)Liquid biopsies:Blood sample analysis to detect cell-free tumour DNA, DNA methylation status, or protein signatures; and

(C)Breathomics: Analysis of volatile organic compounds found in breath to predict cancer risk.


All these approaches involve a number of ongoing research projects, various technologies, as well as internal and commercial partners.Liquid biopsy and Breathomics research are particularly promising because these technologies have potential to become rapid, effective, and affordable point-of-care tests that can predict lung cancer risk in a general population; using Breathomics technology, we have been able to discriminate never-smoker lung cancer patients from never-smoker healthy spouses. Among other exciting projects, we are currently using blood samples collected from patients over the last 15 years to develop a pan-cancer blood test.New Models of Lung Cancer:Patient-Derived Organoids Lung cancers with certain driver mutations (ALK, EGFR, KRAS, etc.)can be specifically targeted by “designer” drugs.Some of these tumors respond very well to these treatments, but eventually grow resistant. Determining which drug to offer next is a major challenge. We have developed a new model system to study drug resistance in these patients:“Patient-Derived Organoids” (PDOs), PDOs are innovative 3-dimensional cancer cell clusters that can be rapidly grown from patient biopsies, maintain important characteristics of patient tumours, and are efficient to study; we are among a small number of groups worldwide that have developed these models in lung cancer. We hypothesize that extensive drug testing and profiling of PDOs can help scientists identify new effective personalized treatments, and help doctors to make better treatment decisions for lung cancer patients following treatment failure.We have funding from the Canadian Institute for Health Research to optimize this technology and assess its ability to generate meaningful personalized data that can assist next-therapeutic decision-making on a clinically-relevant time frame.We have funding from AstraZeneca to use this technology to investigate mechanisms of resistance to osimertinib in EGFR+ lung cancer patients. We are currently seeking funding to pursue PDO models of ALK+and other targetable forms of lung cancer.


National Networks for Generating Real-World Evidence:

CARMA-BROSIn oncology, identification of molecular alterations has led to an improved understanding of the pathways involved in malignant transformation, and the development of drugs that target molecules in order to inactivate or stop the on cogenic process, thereby treating the cancer. More and more, tumours are being characterized and defined by these molecular alterations as they can have profound implications for how the cancer develops and how it can be managed. Non-small cell lung cancers (NSCLC) account for > 80% of lung cancers; however, this tumour is now classified as a myriad of small subtypes (including but not limited to ALK, ROS1, NTRK and NRG1 fusions; EGFR, BRAF, and MET mutations; and more).Some of these alterations, such as ALK, ROS1, and NTRK, occur across disease sites.Most of these mutations occur in a minority of lung cancer patients.The rarity of these molecular alterations and the lack of standardized treatment practices makes it very difficult to study these patient populations at a single institution.Our team has built a national network (CARMA-BROS; principal investigator, G. Liu)that is enabling the collection and sharing of rich clinical data between academic hospitals across Canada.Through this network, we are conducting collaborative research to understand real-world patterns in the management, treatment, and outcomes of patients with rare molecular alterations.This work is just starting but has already been successfully used to gain funding for the use of effective therapeutics in half a dozen these countries already. Growing and maintaining this network is incredibly time and resource intensive. In addition, most of the pressing questions we want to answer are valuable academic pursuits but are not perfectly aligned with the interests of pharmaceutical companies. This makes it difficult to attract funding to expand our objectives.

Empowering Medicine Through  Machine Learning :AIM2REDUCE

As cancer care improves and patients live longer, there has been an increased emphasis on quality-of-life factors, particularly for symptom and toxicity management. Recent research has shown that rapid reporting and management of symptoms and toxicities is associated with a 30% improvement in survival for patients with advanced-stage cancer, rivalling the most effective novel therapeutics.


At The Princess Margaret, we have vast amounts of complex clinical data that, historically, has been difficult to access, curate, and leverage for research. Harnessing real-world evidence, particularly relating to cancer symptoms and treatment toxicities, has the power to change how patients are treated at our institution and beyond.


2BLAST: Biostatistical and Bioinformatic Longitudinal Analysis of Symptoms and Toxicities

In collaboration withThe Princess Margaret Data Science Team, we have completed the 2BLAST Data Catalyst Project to build institutional infrastructure for accessing, collecting, and validating real-world clinical evidence for research.Through this work, our team developed approaches for curating clinical data from many different sources. We now have raw clinical data such as notes, radiology reports, and laboratory values for more than 60,000 patients with cancer over a period of 18 years, making this the richest dataset of this nature in Canada.


AIM2REDUCE:An Intelligent Machine-learning Model for Real-time Early Detection ofUndesirableCancerEventsLeveraging the vast clinical data now available, we are building a computational core facility, supported by other researchers and clinicians with relevant expertise, to develop tools that use artificial intelligence to accurately extract clinical data from free-text notes.So far, our computer science students have developed natural language processing(NLP) methodologies that can successfully extract accurate and detailed information from radiology reports, as well as identify the presence of precancerous lesions in surgical pathology reports. This enables the identification and retrieval of archival tissue for foundational research into molecular biomarkers of pre-cancer, with the goal of preventing cancer entirely.


Recognizing that large-scale data analysis is the future of improved healthcare, we are now starting two major initiatives that leverage the 2BLAST dataset for state-of-the-art artificial intelligence algorithms to predict and prevent, in real time, adverse cancer outcomes–such as emergency room visits and death–through early and appropriate intervention.We are in need of transitional funding to support our pilot efforts until we are able to obtain peer-review funding. Eventually we hope to leverage these methods across our CARMA-BROS network across Canada

Screen Shot 2022-09-12 at 4.29.10 PM.png
Screen Shot 2022-09-12 at 4.29.03 PM.png
Screen Shot 2022-09-12 at 4.33.36 PM.png