Science
The pillars of CMMS
The research of CMMS is built on three pillars, each one is described here in more detail.
Pillar 1 – Development of integrated theoretical and experimental approaches
A current challenge is the equal dovetailing of theory and experiment to jointly formulate approaches to describe biological processes. One difficulty to date has been convincing experimenters that the different levels of hypothesis generation are equivalent and complementary. Therefore, in CMMS, experiments are jointly planned by experimenters and theorists so that i) a sufficiently large number of different usable data are incorporated into modeling and simulation, ii) model predictions are experimentally tested, and iii) theoretical methods and experimental approaches are optimized. This circular approach leads to data usable in models, theoretical descriptions and algorithms, new predictions and new information as a basis for further optimization steps. The efficient use, further development and new development of data analysis methods (e.g. in image processing, omics analysis, dimensionality reduction, fitting, correlation analysis, statistics, machine learning) necessarily takes place. Also, through joint planning, relevant biological and medical experiments are related as models.
Pillar 2 – Multiscale-Modelling and Analysis
In multiscale analysis, data obtained from independently conducted experiments, which may run at different contrasts and at multiple spatiotemporal scales, are combined. While data analysis links the individual levels, modeling and simulation must lead to hypotheses that can be performed at the scale of a specific experiment. Examples include using atomic coordinates with a coarse grain method to simulate the dynamics of biomolecules in large complexes and membranes, or incorporating data from image analysis of high-resolution microscopic techniques into mathematical models to describe the dynamics and structure of these systems. A fundamental problem is that one does not initially know what accuracy of data is required. Development of multiscale analysis, modeling, and simulation methods is ongoing.
Multiscale approaches are based on a combination of methods, e.g., the numerical solution of partial differential equations in complex domains or the coupling of agent-based models to describe the function, structure, or dynamics of an object with networks or systems of differential equations to describe internal regulatory mechanisms. The models created by coupling different approaches are numerically challenging, and often studies on the stability and accuracy of such coupled numerical methods and on their implementation on high-performance computers do not yet exist. Progress in this area requires deep mathematical analyses of the models and methods, as well as the development of efficient implementation strategies and the optimal choice of computer architecture. Furthermore, existing methods for multiscale integration have to be further developed and tested and improved by concrete application in pillar 1 for modeling and simulation of specific systems.
Pillar 3 – High performance computing
A template for integrated data strings from microscope to computer (model) is being developed, networking high-resolution microscopes with high-performance computers. Data and metadata from various light and electron microscopes will be stored in standardized format on HPC mass storage devices, allowing efficient access. Neural networks will be involved at every level in the long term.
Data security is ensured especially with regard to possible medical data. Databases for efficient and standardized access to these data are being developed. The first stage of the data path is automatic pattern recognition and selection of irrelevant areas for physical data reduction. Further stages, such as segmentation, incorporate experience from particle physics.
Another aspect is the visualization of complex data. For this purpose, standardized tools are to be developed and integrated into an analysis platform that efficiently displays different formats. A generic simulation and analysis platform should integrate data import and export with appropriate conversion routines, as well as incorporate data analysis and modeling, statistical packages, visualization, and a standardized scripting language. Since cross-scale models require a lot of computation time, the efficiency of the algorithms is of great importance. Various computer science techniques are used here, such as optimizing data structures, vectorization, generating high parallelism, and using GPGPUs. The different approaches are further developed and combined in libraries.