Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

3D-based RNA function prediction tools in rnaglib

Carlos Oliver, Vincent Mallet, Jérôme Waldispühl

TL;DR

RNA 3D structural data are expanding, enabling data-driven discovery of structure–function relationships, but standardized representations and datasets remain challenging. The chapter presents rnaglib, a Python toolkit that encodes RNA 3D structures as expressive graphs using Leontis-Westhof base-pair geometry, provides dataset construction utilities, and supports self-supervised and supervised learning workflows. It introduces RNADataset and multiple Representations (graph, point cloud, voxel) with a training loop to predict functional attributes such as binding residues, enabling end-to-end ML pipelines. The work lowers barriers to geometric deep learning on RNA and supports design and discovery by linking 3D structure to function in a scalable, extensible framework.

Abstract

Understanding the connection between complex structural features of RNA and biological function is a fundamental challenge in evolutionary studies and in RNA design. However, building datasets of RNA 3D structures and making appropriate modeling choices remains time-consuming and lacks standardization. In this chapter, we describe the use of rnaglib, to train supervised and unsupervised machine learning-based function prediction models on datasets of RNA 3D structures.

3D-based RNA function prediction tools in rnaglib

TL;DR

RNA 3D structural data are expanding, enabling data-driven discovery of structure–function relationships, but standardized representations and datasets remain challenging. The chapter presents rnaglib, a Python toolkit that encodes RNA 3D structures as expressive graphs using Leontis-Westhof base-pair geometry, provides dataset construction utilities, and supports self-supervised and supervised learning workflows. It introduces RNADataset and multiple Representations (graph, point cloud, voxel) with a training loop to predict functional attributes such as binding residues, enabling end-to-end ML pipelines. The work lowers barriers to geometric deep learning on RNA and supports design and discovery by linking 3D structure to function in a scalable, extensible framework.

Abstract

Understanding the connection between complex structural features of RNA and biological function is a fundamental challenge in evolutionary studies and in RNA design. However, building datasets of RNA 3D structures and making appropriate modeling choices remains time-consuming and lacks standardization. In this chapter, we describe the use of rnaglib, to train supervised and unsupervised machine learning-based function prediction models on datasets of RNA 3D structures.
Paper Structure (16 sections, 2 figures)

This paper contains 16 sections, 2 figures.

Table of Contents

  1. Introduction
  2. RNA 3D Data in rnaglib
  3. Representing RNA 3D structures
  4. Objectives
  5. Materials
  6. Dataset Utilities
  7. Dataset preparation
  8. Obtaining annotated structures
  9. Annotating your own structure
  10. Inspecting the RNA 3D object
  11. The RNA 3D Dataset object
  12. Machine Learning Tutorial
  13. RNA Representations
  14. Training Loop
  15. Visualizing Structures
  16. ...and 1 more sections

Figures (2)

  • Figure 1: RNA 3D representation learning paradigm. We use the Leontis-Westhof base pair geometry classification to build expressive graphs of RNA 3D models. The graphs are then embedded into a learned space with an encoder network $\mathrm{ENC}_\theta$ to emit embedding space $\mathbf{Z}$ and property space $\mathbf{P}$ with the decoder $\mathrm{DEC}_\psi$.
  • Figure 2: Sample graph drawing of PDBID: 1NLF