Summary
Ribonucleic Acids (RNAs) are crucial polymers in biology: RNAs might be the key to understanding the origins of life and are extremely promising therapeutic tools. Just like proteins, RNAs numerous properties stem from their 3-dimensional structure, which itself arises from their nucleotide sequence. However, RNA molecules are particularly dynamical, making their structure noticeably difficult to study, both experimentally and theoretically. Moreover, RNA dynamics are critical for biological function.
Recently, deep learning (DL) methods have revolutionised computational biophysics by mastering the task of protein native structure prediction from sequence. Comparable success can be hoped for RNA, but it is impaired by the less abundant data and the lack of direct transferability of current methods. More importantly, no attempt has been made at developing a DL approach to directly predict RNA dynamics.
In this proposal we aim to adapt the emergent framework of diffusion models to the task of one-shot sampling of RNA conformational ensembles. Diffusion models constitute a new paradigm in generative machine learning (ML) that has attained astounding successes in conditional image generation, audio, graph and geometric shape synthesis. Our goal is to produce a neural network-based software which, given an arbitrary input sequence, efficiently samples the 3D coordinates of RNA conformations according to their equilibrium probabilities. To do so we suggest an original approach combining (1) coarse-grained internal coordinates, (2) a diffusion-based generative framework, (3) an attention-based architecture inspired by state-of-the-art DL biomolecular approaches and (4) a mixed training procedure based on experimental fragments and molecular dynamics simulations. Our approach could be used as a replacement to extensive MD simulations which are voracious both in human time and energy expense, hence accelerating biophysical research and decreasing its carbon footprint.
Recently, deep learning (DL) methods have revolutionised computational biophysics by mastering the task of protein native structure prediction from sequence. Comparable success can be hoped for RNA, but it is impaired by the less abundant data and the lack of direct transferability of current methods. More importantly, no attempt has been made at developing a DL approach to directly predict RNA dynamics.
In this proposal we aim to adapt the emergent framework of diffusion models to the task of one-shot sampling of RNA conformational ensembles. Diffusion models constitute a new paradigm in generative machine learning (ML) that has attained astounding successes in conditional image generation, audio, graph and geometric shape synthesis. Our goal is to produce a neural network-based software which, given an arbitrary input sequence, efficiently samples the 3D coordinates of RNA conformations according to their equilibrium probabilities. To do so we suggest an original approach combining (1) coarse-grained internal coordinates, (2) a diffusion-based generative framework, (3) an attention-based architecture inspired by state-of-the-art DL biomolecular approaches and (4) a mixed training procedure based on experimental fragments and molecular dynamics simulations. Our approach could be used as a replacement to extensive MD simulations which are voracious both in human time and energy expense, hence accelerating biophysical research and decreasing its carbon footprint.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/101152924 |
| Start date: | 01-12-2024 |
| End date: | 30-11-2026 |
| Total budget - Public funding: | - 172 750,00 Euro |
Cordis data
Original description
Ribonucleic Acids (RNAs) are crucial polymers in biology: RNAs might be the key to understanding the origins of life and are extremely promising therapeutic tools. Just like proteins, RNAs numerous properties stem from their 3-dimensional structure, which itself arises from their nucleotide sequence. However, RNA molecules are particularly dynamical, making their structure noticeably difficult to study, both experimentally and theoretically. Moreover, RNA dynamics are critical for biological function.Recently, deep learning (DL) methods have revolutionised computational biophysics by mastering the task of protein native structure prediction from sequence. Comparable success can be hoped for RNA, but it is impaired by the less abundant data and the lack of direct transferability of current methods. More importantly, no attempt has been made at developing a DL approach to directly predict RNA dynamics.
In this proposal we aim to adapt the emergent framework of diffusion models to the task of one-shot sampling of RNA conformational ensembles. Diffusion models constitute a new paradigm in generative machine learning (ML) that has attained astounding successes in conditional image generation, audio, graph and geometric shape synthesis. Our goal is to produce a neural network-based software which, given an arbitrary input sequence, efficiently samples the 3D coordinates of RNA conformations according to their equilibrium probabilities. To do so we suggest an original approach combining (1) coarse-grained internal coordinates, (2) a diffusion-based generative framework, (3) an attention-based architecture inspired by state-of-the-art DL biomolecular approaches and (4) a mixed training procedure based on experimental fragments and molecular dynamics simulations. Our approach could be used as a replacement to extensive MD simulations which are voracious both in human time and energy expense, hence accelerating biophysical research and decreasing its carbon footprint.
Status
SIGNEDCall topic
HORIZON-MSCA-2023-PF-01-01Update Date
19-12-2024
Images
No images available.
Geographical location(s)
Search
