Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning Framework for RNA Inverse Folding with Geometric Structure Potentials

Annabelle Yao

TL;DR

This work tackles RNA inverse folding by predicting sequences that realize a predefined 3D conformation. It introduces an end-to-end GVP-augmented Transformer that encodes geometric RNA features and autoregressively designs sequences, trained on ~2500 experimentally solved single-chain RNAs. Across standard benchmarks and RNA-Puzzles, the method achieves state-of-the-art recovery (0.481 on RNA-Puzzles) and TM-scores (0.332 on RNA-Puzzles), with strong cross-family generalization validated by a leave-one-family-out setup and AlphaFold3 folding confirming structural plausibility. The framework accelerates RNA design, enabling rapid, geometry-informed sequence generation with broad implications for RNA therapeutics and synthetic biology; the code is open-source for community use.

Abstract

RNA's diverse biological functions stem from its structural versatility, yet accurately predicting and designing RNA sequences given a 3D conformation (inverse folding) remains a challenge. Here, I introduce a deep learning framework that integrates Geometric Vector Perceptron (GVP) layers with a Transformer architecture to enable end-to-end RNA design. I construct a dataset consisting of experimentally solved RNA 3D structures, filtered and deduplicated from the BGSU RNA list, and evaluate performance using both sequence recovery rate and TM-score to assess sequence and structural fidelity, respectively. On standard benchmarks and RNA-Puzzles, my model achieves state-of-the-art performance, with recovery and TM-scores of 0.481 and 0.332, surpassing existing methods across diverse RNA families and length scales. Masked family-level validation using Rfam annotations confirms strong generalization beyond seen families. Furthermore, inverse-folded sequences, when refolded using AlphaFold3, closely resemble native structures, highlighting the critical role of geometric features captured by GVP layers in enhancing Transformer-based RNA design.

Deep Learning Framework for RNA Inverse Folding with Geometric Structure Potentials

TL;DR

This work tackles RNA inverse folding by predicting sequences that realize a predefined 3D conformation. It introduces an end-to-end GVP-augmented Transformer that encodes geometric RNA features and autoregressively designs sequences, trained on ~2500 experimentally solved single-chain RNAs. Across standard benchmarks and RNA-Puzzles, the method achieves state-of-the-art recovery (0.481 on RNA-Puzzles) and TM-scores (0.332 on RNA-Puzzles), with strong cross-family generalization validated by a leave-one-family-out setup and AlphaFold3 folding confirming structural plausibility. The framework accelerates RNA design, enabling rapid, geometry-informed sequence generation with broad implications for RNA therapeutics and synthetic biology; the code is open-source for community use.

Abstract

RNA's diverse biological functions stem from its structural versatility, yet accurately predicting and designing RNA sequences given a 3D conformation (inverse folding) remains a challenge. Here, I introduce a deep learning framework that integrates Geometric Vector Perceptron (GVP) layers with a Transformer architecture to enable end-to-end RNA design. I construct a dataset consisting of experimentally solved RNA 3D structures, filtered and deduplicated from the BGSU RNA list, and evaluate performance using both sequence recovery rate and TM-score to assess sequence and structural fidelity, respectively. On standard benchmarks and RNA-Puzzles, my model achieves state-of-the-art performance, with recovery and TM-scores of 0.481 and 0.332, surpassing existing methods across diverse RNA families and length scales. Masked family-level validation using Rfam annotations confirms strong generalization beyond seen families. Furthermore, inverse-folded sequences, when refolded using AlphaFold3, closely resemble native structures, highlighting the critical role of geometric features captured by GVP layers in enhancing Transformer-based RNA design.
Paper Structure (13 sections, 6 equations, 5 figures, 1 table)

This paper contains 13 sections, 6 equations, 5 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. Data Preprocessing
  4. Structural Features
  5. Feature Encoding
  6. Transformer
  7. Model Training and Evaluation
  8. Results
  9. Model Performance and Benchmarking
  10. Mask-Family Validation
  11. RNA-puzzle Evaluation
  12. Discussion
  13. Conclusion

Figures (5)

  • Figure 1: The overall method used to preprocess my data. (Graph created by the student researcher using Canva and the structures downloaded from RCSB PDB)
  • Figure 2: The workflow of my framework. First, key atom coordinates are extracted from the PDB file. Then, the scalar and vector features of the key atoms are computed along with the rotation matrix to ensure rotational invariance. These features will be used to make a graph where the nodes are each key atom and the weights are the features. The graph is then passed through the GVP and the Transformer to generate an RNA sequence. (Graph created by the student researcher using Canva and the structures downloaded from RCSB PDB)
  • Figure 3: GVP structure. (Graph recreated by the student researcher using Canva based on jing2020learning)
  • Figure 4: Transformer structure. (Graph created by the student researcher using Canva based on prior research yao2024novel)
  • Figure 5: Results on RNA-Puzzles and Mask-Family. (A) Bar chart of the recovery rate for each RNA-Puzzle ID using my method. (B) Bar chart of the validation results for the Mask-Family task, reporting both recovery rate and fold-back TM-score for each family. (C) Comparison of RNA design outcomes for PZ23, contrasting LEARNA's predictions (blue structure) with ours (green), both folded using AlphaFold3 and overlaid with the ground-truth structure (red). (D) The same comparison for the RF01151 family. (Figures A and B are created by the student researcher using GraphPad Prism and Figures C and D are created by the student researcher using PyMOL)