Boosting the Translational Impact of Scientific Competitions by Ensemble Learning

点击翻译按钮获取中文摘要

 DESCRIPTION (provided by applicant): "Big data" such as those arising from sequencing, imaging, genomics and other emerging technologies are playing a critical role in modern biology and medicine. The generation of hypotheses about biological processes and disease mechanisms is now increasingly being driven by the production and analysis of large and complex datasets. Advanced computational methods have been developed for the robust analysis of these datasets, and the growth in number and sophistication of these methods has closely tracked the growth in volume and complexity of biomedical data. In such a crowded environment of diverse computational methods and data, it is difficult to judge how generalizable the performance of these methods is from one setting to another. Crowdsourcing-based scientific competitions, or challenges, have now become popular mechanisms for the rigorous, blinded and unbiased evaluation of the performance of these methods and the identification of best-performing methods for biomedical problems. However, despite the benefits of these challenges to the biomedical research enterprise, the impact of their findings has been remarkably limited in laboratory and clinical settings. This is likely due to two important aspects of current challenges: (i) their over-emphasis on identifying the "best" solutions rather than tryig to comprehensively assimilate the knowledge embedded in all the submitted solutions, and (ii) the absence of a stable channel of communication and collaboration between problem and solution providers due to a lack of sufficient incentives to do so. The aim of this project is to boost the translational impact of scientific challenges through a combination of novel machine learning methods, development of novel scalable software and unique collaborations with disease experts to ensure the effective translation of knowledge accrued in challenges to real clinical settings and practice. These novel methods and software are designed to effectively assimilate the knowledge embedded in all the submissions to challenges into "ensemble" solutions. In a first of its kind effort, the ensemble solutions derived from disease-focused challenges under the DREAM project will be brought directly to scientists and clinicians that are experts in these disease areas. Initial effort in this project will focus on active DREAM challenges aiming at the accurate prediction of drug response and clinical outcomes respectively in Rheumatoid Arthritis (RA) and Acute Myeloid Leukemia (AML). Both these diseases are difficult to treat and thus they pose major medical and public health concerns. In collaboration with RA and AML experts, the ensemble solutions learnt in these challenges will be validated in independent patient cohorts and carefully designed clinical studies. This second-level validation is essential to judge the clinical applicability of any method, but is rarely done As the methodology is general, similar efforts will be made for other diseases in later stages of the project. Overall, using a smart combination of crowdsourcing-based challenges and computational methods and software, we aim to demonstrate a unique pathway for studying and treating disease by truly leveraging the "wisdom of the crowds".