International crowd-sourced data-mining competition resulted in improved prediction of prostate cancer patient survival
Results of the Prostate Cancer DREAM Challenge published in the top clinical oncology journal Lancet Oncology today beautifully demonstrate the benefits of sharing clinical trial data. The winning prediction model was developed by a team headed by FIMM Group Leader Tero Aittokallio. Overall, this collaborative scientific effort led to novel insights and improvements in cancer treatment and management.
In 2015, professor Aittokallio’s team excelled in Sage Bionetworks’ Prostate Cancer DREAM Challenge, a crowd-sourcing competition for highly demanding scientific problems. The participating teams were asked to develop new outcome prediction models for a subtype of prostate cancer, called metastatic castrate resistant prostate cancer (mCRPC), accounting for one third of all prostate cancer patients with metastatic disease.
A total of 50 competing teams comprised of diverse group of international experts created models using data hosted on the Project Data Sphere ® Online Service – a broad-access research platform that collects and curates patient-level data from completed, phase III cancer clinical trials. Data from more than 150 clinical variables from more than 2000 patients were made available to the competing teams.
"My group has a long-term expertise in developing multivariate machine learning models for various biomedical applications, but this DREAM Challenge provided us with the unique opportunity to work on clinical trial data, with the eventual aim to help patients with metastatic castration-resistant prostate cancer," Aittokallio says.
In addition to outperforming all other teams, the winning team’s model was also more accurate than a recently published state-of-the-art prognostic model.
In the report published in Lancet Oncology 15 November, both the winning model and the novel discoveries made by this machine learning approach are described. These results demonstrate how data-mining can identify combinations of clinical features most predictive of individual medical outcomes as an effective way to reveal new insights about disease from patterns in patients’ clinical data.
“Our model was based on not only single clinical measurements, but actually certain biomarker combinations were shown most accurate to predict how a patient’s disease will progress. For instance, in addition to immune system biomarkers and renal and hepatic function, our algorithm identified an under-reported cancer biomarker, aspartate aminotransferase, as an important factor in making prognoses”, Aittokallio continued.
Image: Teijo Pellinen. Multiplexed fluorescent staining of human prostate cancer tissue, showing cancer glands in red and green, and one normal epithelial gland with pink cell lining
"To me, the Challenge proved that clinical trial data can be utilized in prognostic modelling. More open sharing of trial data with the cancer research community would be important”, says member of the winning team, MD Tuomas Mirtti, from HUSLAB.
Organizers and Prostate Cancer DREAM Challenge winners include participants from: Sage Bionetworks, US; the University of Texas Southwestern Medical Center, US; the University of Turku, Finland; University of Helsinki, Finland; Helsinki University Hospital, Finland; the University of Colorado, US; AstraZeneca, US; IBM T.J. Watson Research Center, US; Prostate Cancer Foundation, US; Harvard Medical School, Boston, MA, US; the University of California, San Francisco, US; Memorial Sloan-Kettering Cancer Center and Weill Cornell Medical College, US; Tulane Cancer Center, US; University of Texas Southwestern Medical Center, US; and the University of Colorado Comprehensive Cancer Center, US.
Guinney et al. Prediction of overall survival for patients with metastatic castration-resistant prostate cancer: development of a prognostic model through a crowdsourced challenge with open clinical trial data. Lancet Oncology Online first publication, 15 November 2016.
Tero Aittokallio, PhD, Professor
Institute for Molecular Medicine Finland (FIMM)
University of Helsinki