Gast ::
| Hauptseite > Publikationsdatenbank > Contact Maps in RNA Structure Prediction |
| Hauptseite > Publikationsdatenbank > Contact Maps in RNA Structure Prediction |
| Poster (After Call) | FZJ-2024-07274 |
; ; ;
2024
Abstract: Predicting the spatial structure of non-coding RNA (ncRNA) is an important task for understanding fundamental processes in living nature. Physical force fields are used to infer the structure from a sequence using simulations on high-performance computers. However, the best results are obtained by incorporating evolutionary data via a binary mapping of contacts. The same phenomenon can be seen in protein structure prediction, where the groundbreaking AlphaFold2 also incorporates this step. Much work has been done in the past to optimise the algorithms for simulations, but what are good contacts and why are these contacts important in the first place is an unsolved puzzle. To find answers, we tried different contact map topologies on a well-defined test set of ncRNAs. We also looked at using fewer, but wisely chosen contacts and how this can improve prediction. To obtain our results, we ran many simulations for comparison on the high performance cluster JUWELS with the RNA folding software SimRNA and used convolutional neural networks (CNN) to select contacts. Our results suggest that it is important to pay more attention to the selection of contacts, especially when developing machine learning algorithms. Furthermore, good contacts not only ensure faster folding in the simulation, they are actually essential for correct folding. It seems that it is the additional constraints that bring the physical force field into the more correct form.
|
The record appears in these collections: |
