[Paper Review] Accurate RNA 3D structure prediction using a language model-based deep learning approach

Tao Shen, Zhigang Hu+15 more|arXiv (Cornell University)|Jul 4, 2022

RNA and protein synthesis mechanisms57 citations

TL;DR

E2Efold-3D is an end-to-end deep learning method for de novo RNA 3D structure prediction that achieves sub-4 Å RMSD on RNA puzzles and is substantially faster than FARFAR2.

ABSTRACT

Accurate prediction of RNA three-dimensional (3D) structure remains an unsolved challenge. Determining RNA 3D structures is crucial for understanding their functions and informing RNA-targeting drug development and synthetic biology design. The structural flexibility of RNA, which leads to scarcity of experimentally determined data, complicates computational prediction efforts. Here, we present RhoFold+, an RNA language model-based deep learning method that accurately predicts 3D structures of single-chain RNAs from sequences. By integrating an RNA language model pre-trained on ~23.7 million RNA sequences and leveraging techniques to address data scarcity, RhoFold+ offers a fully automated end-to-end pipeline for RNA 3D structure prediction. Retrospective evaluations on RNA-Puzzles and CASP15 natural RNA targets demonstrate RhoFold+'s superiority over existing methods, including human expert groups. Its efficacy and generalizability are further validated through cross-family and cross-type assessments, as well as time-censored benchmarks. Additionally, RhoFold+ predicts RNA secondary structures and inter-helical angles, providing empirically verifiable features that broaden its applicability to RNA structure and function studies.

Motivation & Objective

• Motivate accurate de novo RNA 3D structure prediction despite scarce experimental data.
• Propose an end-to-end differentiable DL pipeline that outputs 3D coordinates directly.
• Leverage RNA foundation model representations and secondary structure information to augment data.
• Incorporate self-distillation and recycling to enhance prediction accuracy and data efficiency.

Proposed method

• Use a fully differentiable end-to-end model to predict atom coordinates from sequence data.
• Initialize sequence representations with an RNA foundation model trained on 23M RNA sequences.
• Process features through a 4-layer E2Eformer to capture nucleotide interactions.
• Use an 8-layer structure module to generate final 3D RNA coordinates with structure-aware loss constraints.
• Augment training with secondary structure information and self-distillation to expand effective data size.
• Incorporate recycling to refine predictions similarly to AlphaFold-style iterative refinement.

Experimental results

Research questions

• RQ1 Can a fully differentiable end-to-end model predict de novo RNA 3D structures from sequence alone?
• RQ2 Does combining RNA foundation model representations with secondary structure constraints improve 3D predictions?
• RQ3 Do self-distillation and recycling strategies meaningfully boost prediction accuracy for RNA 3D structures?
• RQ4 How does E2Efold-3D perform on RNA puzzles and RNA complexes compared with state-of-the-art methods?

Key findings

• On an independent non-redundant test set of 50 RNAs, E2Efold-3D achieves an average RMSD of 3.486 Å, TM-score of 0.518, and lDDT of 0.739.
• E2Efold-3D obtains an average TM-score of 0.5175, outperforming FARFAR2-avg (0.249) and FARFAR2-best (0.294 approximate) across the reported comparison, indicating better global fold matching.
• Average inference time is 0.12 seconds per structure, versus FARFAR2 at about 4,777 seconds per structure, highlighting substantial speed advantages.
• Ablation studies show RNA-FM embeddings and recycling both contribute to performance gains, with RNA-FM boosting TM-score and recycling aiding RMSD refinement.
• On RNA-Puzzles benchmarks including monomeric and RNA-ligand complexes, E2Efold-3D achieved best scores across RMSD and TM-score compared with community results.