Practical challenges for biomedical modeling using HPC.

The concept underlying precision medicine is that prevention, diagnosis and treatment of pathologies such as cancer can be improved through an understanding of the influence of individual patient characteristics. Predictive medicine seeks to derive this understanding through mechanistic models of the causes and (potential) progression of diseases within a given individual. This represents a grand challenge for computational biomedicine as it requires the integration of highly varied (and potentially vast) quantitative experimental datasets into models of complex biological systems. It is becoming increasingly clear that this challenge can only be answered through the use of complex workflows that combine diverse analyses and whose design is informed by an understanding of how predictions must be accompanied by estimates of uncertainty. Each stage in such a workflow can, in general, have very different computational requirements. If funding bodies and the HPC community are serious about the desire to support such approaches, they must consider the need for portable, persistent and stable tools designed to promote extensive long term development and testing of these workflows. From the perspective of model developers (and with even greater relevance to potential clinical or experimental collaborators) the enormous diversity of interfaces and supercomputer policies, frequently designed with monolithic applications in mind, can represent a serious barrier to innovation. Here we use experiences from work on two very different biomedical modeling scenarios - brain bloodflow and small molecule drug selection - to highlight issues with the current programming and execution environments and suggest potential solutions.