Owkin HCC offering: Predicting prognosis and identifying biomarkers for liver cancer
Owkin HCC offering: Predicting prognosis and identifying biomarkers for liver cancer
What is HCC (a type of liver cancer)?
Hepatocellular carcinoma, also known as ‘HCC,’ is the most common liver cancer and was also the fourth leading cause of cancer-related deaths worldwide in 20181. Early-stage HCC is resectable and has a favorable prognosis compared to late-stage HCC; however, people who develop unresectable tumors in later stages are not curable. Furthermore, even those who undergo procedural treatments to remove a tumor at earlier stages face high recurrence rates (up to 70-80%), and existing systemic therapies, i.e., subsequent post-operative treatment for patients, only incrementally increase patient survival (up to 1.5 years).
Currently, most potential therapies fail in phase 3 of the clinical trial. It is in this phase at which researchers test whether the new treatment is better than the standard care. Machine Learning (‘ML’) models are useful tools to enhance the late-stage development of drugs by optimizing the phase 3 trials. For example, models can be applied to a patient population to select high-value subgroups of patients that are likely to respond to systemic therapy. This reduces the sample size needed to prove the efficacy of the drug. It also maximizes the chance of success in shorter periods, which, in turn, significantly accelerates time-to-market.
Owkin’s AI Research
At Owkin, we believe that Artificial Intelligence (‘AI’) can be a crucial tool when tackling the most challenging scientific questions. What’s more, AI can accelerate vital collaborations in clinical research to solve unmet medical needs. Our challenge: How can we effectively learn from the millions of siloed patient data points without compromising data privacy and protection? Our solution? Firstly, to bring academic and pharmaceutical industry researchers together in a Federated Research Data Platform. Here, using proprietary infrastructure and technologies, researchers can train ML models on distributed data at scale across multiple medical institutions without centralizing the data.
When tackling a particular therapeutic area or research question, we organize the following key elements into specific AI focus areas: Access to expert-curated clinical datasets, AI predictive models, expertise from key opinion leaders across medical institutions worldwide, and Owkin’s internal biomedical AI experts.
In this blog post, we introduce and explain the Owkin HCC capabilities.
Owkin HCC capabilities
The HCC capabilities offers access to expertly-curated clinical datasets from multiple academic centers in the USA and Europe. Owkin partners with these leading medical centers to develop ML models. These have the potential to accelerate drug development and clinical care in the field of HCC. Current projects include; (i) the prediction of prognosis from whole slide images1, which we will discuss in more detail in this article; (ii) liver segmentation and lesion detection from ultrasounds2; and (iii) the prediction of response to treatment, leveraging radiomics on Magnetic Resonance Imaging (‘MRI’) and Computed Tomography (‘CT’) Scan data.
Pr. Valerie Vilgrain, Chief of Radiology at Beaujon Hospital, AP-HP/Greater Paris University Hospitals and a key partner for Owkin in the fight against this type of liver cancer said:
Prognostic markers of response to trans-arterial chemoembolization (TACE) in liver cancer are not well known. We are collaborating with Owkin to apply machine learning techniques to multimodal data from a large curated research cohort. We hope that these novel approaches will improve our understanding of the disease and we expect to use this knowledge to personalize and improve TACE treatments.
In addition to ongoing projects in the HCC capabilities, Owkin has published the results of two outcome prediction ML models (HCCInterpret and HCCExpert). These models improve the prediction of overall survival, compared to the existing staging approaches, and identify novel biomarkers of prognosis in patients with resected HCC. They were trained and validated on datasets from Henri Mondor Hospital, AP-HP/Greater Paris University Hospitals in Paris, and The Cancer Genome Atlas (‘TCGA’). This was achieved with the help of international academic experts such as pathologist Dr. Julien Calderaro and Owkin’s data scientists. These elements form an optimal collaborative environment to help tackle one of the world’s deadliest diseases: Hepatocellular carcinoma (‘HCC’).
Owkin HCC capabilities in Action
At Owkin, we value the impact of our collaboration with world-leading academic research centers. To build our HCC capabilities, we first accessed data from French hospital Henri-Mondor, to train proprietary algorithms: HCCInterpret and HCCExpert. These are two ML models built to accurately predict, from histology slides, the survival of patients with HCC after surgical resection.
HCCInterpret is an interpretable model that does not require pathologist annotations. It can therefore be used to identify biomarkers of survival.
HCCExpert includes an ‘attention mechanism’ allowing it to recognize pathologist annotations to further improve the predictive performance of overall survival.
Dr. Julien Calderaro is a member of the department of pathology at Henri Mondor Hospital, AP-HP/Greater Paris University Hospitals, and the corresponding author of the paper. He recognizes the need to develop adjuvant therapies for liver cancer. Dr. Calderaro was pleased to announce, “these models performed better than any known clinical, biological or pathological prognostic markers and provided novel insights on the biological features related to the aggressiveness of the tumor.”
In order to train both HCCInterpret and HCCExpert, we used a dataset from the Henri-Mondor Hospital of 390 histology slides from 194 patients with resected HCC. An expert liver pathologist annotated the tumor and non-tumoral areas to train HCCExpert. Figure 1 below shows an example of experts’ annotations next to a histological slide. By incorporating the expertise of pathologists and clinicians in the annotations of histology slides, the HCCExpert predictive model became more robust and yielded high-impact results.
Figure 1: A sketch of an expert-annotated histological slide used to train Owkin’s predictive models.
An independent dataset of 328 histology slides from The Cancer Genome Atlas (TCGA) validated both HCC models. Additionally, we compared the model performance to all existing clinical, biological and pathological prognostic markers.
The models significantly outperformed existing survival scores that incorporate all relevant clinical, biological, and pathological variables.
HCCInterpret analyzed cellular and immunological characteristics and identified previously unknown areas of tumors most predictive of poor survival. HCCExpert improved predictive performance by 14% as a result of its ability to recognize expert annotation. These results were outlined in Hepatology. They are one of many proofs-of-concept for Owkin’s Data Platform enhancing oncology state-of-the-art research methodologies through innovative AI predictive models.
Owkin HCC capabilities impact on liver cancer
Owkin’s HCC offering aims to power medical research, accelerate drug development and improve clinical care.
With Owkin Studio, medical researchers can access expert-curated models via our collaborative software platform featuring state-of-the-art AI models. (To learn more, explore the HCCInterpret predictive model on the TCGA validation cohort via Owkin Studio demo.) Researchers will also have access to a hands-on application of our model. They can use it to find new biomarkers of HCC and gain invaluable research insights into the disease.
Owkin’s data access, models, and know-how built within our HCC capabilities apply to Pharma at all stages of the drug development pipeline. Such elements are accessible via one of our pharma solutions. You can see the tangible impact in the application of both of our HCC models.
HCCInterpret leads to the identification of biomarkers of survival. This can be a great lead for pharmaceutical companies to narrow down druggable targets for HCC and enhance the HCC drug development process.
HCCExpert is an impactful resource in terms of clinical trial efficiency. We simulated the impact of the model applied during the patient selection process of phase 3 clinical trials2. Adjusting for the model’s prognostic scores in the trial design, the number of patients required to meet the trial endpoints and determine drug efficacy is significantly reduced. These lead to cost savings of ~15% and a reduced trial duration of ~1 year.
In conclusion, Owkin’s HCC capabilities highlights both the impactful academic research and the invaluable pharmaceutical impact which exists within the field of Oncology. This is just one of the many research collaborations and predictive models that we’ve developed with academic medical experts on their curated datasets. In future projects and collaboration, we plan to broaden our approach further. We will do this by including new data modalities such as genomics and radiology into our HCC models.
Llovet JM., et al., Hepatocellular carcinoma, Nat Rev Dis Primers 2016
Schmauch B., et al., Diagnosis of focal liver lesions from ultrasound using deep learning, Diagnostic and Interventional Imaging, 2019
Owkin simulated patient recruitment efficiencies for two major HCC trials: KEYNOTE-937 and CheckMate 9DX.