Interpretation of Biomolecular Data Driven by Swarm Intelligence
The detailed structure of biomolecules is of key interest in understanding their crucial roles and interactions in the human body. In order to develop new drugs, for example, researchers must first determine these structures in complex experiments. However, the data obtained from such experiments is sometimes ambiguous and cannot be uniquely assigned to a specific structure. Scientists from JSC, KIT, DKFZ, and the University of Duisburg-Essen have jointly developed an AI-based method to evaluate such ambiguous data with the help of data-driven molecular simulations. This method is based on the concept of swarm learning in AI research. A supercomputer simulates many swarm members at the same time on over 1000 processors. Each member tests different parameter combinations and weights of the experimental data with a complementary physics-based computer model. According to Marie Weiel, a doctoral researcher and leading author of the study, a crucial ingredient is the communication that takes place between the swarm members to cooperatively find an optimal solution, which is essential for the best possible interpretation of the data as molecular structures. The method thus delivers very accurate structures and at the same time uses the available computing resources – in this case JUWELS and the HPC cluster at KIT – very efficiently. The results were published in the journal Nature Machine Intelligence (DOI:10.1038/s42256-021-00366-3).
Contact: Prof. Alexander Schug, al.schug@fz-juelich.de
from JSC News No. 283, 24 September 2021