Machine learning / artificial intelligence in biosimulations

KIT/SCC, FZJ, DKFZ and University of Duisburg/Essen develop a self-adapting variant of dynamic particle swarm optimization of biomolecule simulations - Paper published in nature machine intelligence.

Biomolecules are the molecular Swiss Army knives in our bodies and perform a wide variety of functions. The protein shown is an enzyme that regulates the energy carrier ATP. In the background, the HPC cluster of the SCC used for the project.

Life functions on the molecular scale through a complex interplay of biomolecules, for which the structure of the biomolecules involved plays a central role. Experimental methods can determine such structures and are based on the evaluation of primary data, but their interpretation is not always unambiguous. Molecular simulations are a powerful tool to evaluate such ambiguous experimental data.

An essential challenge is to weight the experimental interformation against the underlying physical simulation model. In a collaboration of FZJ, KIT, DKFZ and the University of Duisburg/Essen, a self-adaptive variant of the dynamic particle swarm optimization was developed to solve this weighting problem. Each individual parameter is learned at runtime, resulting in a dynamically evolving and iteratively refined search space topology. It has been shown for several biomolecular systems that the method makes very efficient use of computation time while yielding highly accurate structures. Since such parameter problems are common in molecular simulations, applications such as material simulations are conceivable in addition to biomolecular simulations.

To the paper: Dynamic particle swarm optimization of biomolecular simulation parameters with flexible objective functions


Contact person at SCC: Prof. Dr. Alexander Schug


(translated wirh DeepL.com)