Berlin 2024 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

BP: Fachverband Biologische Physik

BP 10: Computational Biophysics II

BP 10.7: Hauptvortrag

Dienstag, 19. März 2024, 11:15–11:45, H 0112

RNA Contact Prediction by Data Efficient Deep Learning — Oskar Taubert¹, Fabrice von der Lehr², Alina Bazarova^3,4, Christian Faber³, Philipp Knechtges², Marie Weiel^1,4, Charlotte Debus^1,4, Daniel Coquelin^1,4, Achim Basermann², Achim Streit¹, Stefan Kesselheim^3,4, Markus Götz^1,4, and •Alexander Schug^3,5 — ¹Scientific Center of Computing, Karlsruhe Institute of Technology — ²Institute for Software Technology, German Aerospace Centre — ³Jülich Supercomputing Centre, Forschungszentrum Jülich — ⁴Helmholtz AI — ⁵Faculty of Biology, University of Duisburg-Essen

On the molecular level, life is orchestrated via many biomolecules. To gain detailed understanding of biomolecular function, one needs to know their structure. Yet the structural characterization of many important biomolecules and their complexes remains experimentally challenging. For proteins, the richness of labeled training data enables highly successful deep-learning approaches. Deep learning on RNA, however, is hampered by the lack of such data. The limited data, however, can still be used to predict spatial adjacencies (*contact maps*) as proxy for 3D structure. Statistical physics based approaches such as direct coupling analysis can provide such contact maps. Going beyond such approaches, our recent model BARNACLE combines using unlabeled data through self-supervised pre-training and efficient use of the sparse labeled data. We observe a considerable improvement over both the established classical baseline DCA and other neural networks.

Keywords: RNA Structure Prediction; Data Inference; Machine Learning; Molecular Dynamics