Question 1

What are the three main approaches to RNA structure prediction?

Accepted Answer

The three principal approaches are: (1) thermodynamic free energy minimization using nearest-neighbor parameters to compute the MFE secondary structure; (2) data-directed chemical probing integration, incorporating SHAPE, DMS-seq, and icSHAPE signals as pseudo-energy constraints; and (3) deep learning and foundation models, including CNNs, bidirectional LSTMs, transformers, and graph neural networks trained on sequence or probing data.

Question 2

How accurate are deep learning RNA structure prediction methods compared to thermodynamic approaches?

Accepted Answer

Deep learning methods have substantially outperformed thermodynamic baselines on benchmark datasets. UFold achieved F1 = 0.91 — approximately 10–30% improvement over thermodynamic models. DeepFoldRNA achieved an average RMSD of 2.69 Å and is approximately 350–4,000× faster than Monte Carlo methods. DRfold achieved greater than 75.6% improvement in TM-score over prior approaches. trRosettaRNA achieved mean RMSD 5.5 Å on RNA-Puzzles versus 10.5 Å for the best human groups.

Question 3

Who are the key patent assignees in RNA structure prediction?

Accepted Answer

Within this dataset of 9 identified patents, Chinese academic institutions dominate filing volume: Huazhong University of Science and Technology (2 patents, protein-RNA complex prediction), Shandong Province Computing Center (2 patents, supercomputing infrastructure), Hunan University (2 patents, RNA-binding residue prediction via graph neural networks), and Fudan University (1 patent, RNA language model). Atomic AI, Inc. holds 2 US and WO patents covering ML with chemical mapping data for RNA tertiary structure prediction.

Question 4

How is RNA structure prediction applied to drug discovery?

Accepted Answer

RNA structure prediction has been applied directly to SARS-CoV-2 genomic RNA to identify unusually stable RNA structures as potential small-molecule drug targets. The Atomic AI patents explicitly target therapeutic RNA applications including human transcriptome mapping. Chemical probing data integration (SHAPE/DMS) is a key differentiator for therapeutic RNA applications, with RASP covering 161 datasets across 18 species and 18 methods.

Question 5

What is the role of RNA foundation models in structure prediction?

Accepted Answer

RNA foundation models such as RNA-FM and UNI-RNA use self-supervised pre-training on tens of millions of unannotated RNA sequences using masked language modeling, then fine-tune for structure and function prediction. RNA-FM was pre-trained on 23 million non-coding RNA sequences and infers sequential and evolutionary information without labels. Fudan University's 2025 patent explicitly claims an RNA language model pipeline for secondary structure prediction.

Question 6

What emerging directions are shaping RNA structure prediction research in 2025–2026?

Accepted Answer

Five emerging directions are identified: (1) RNA foundation models and self-supervised pre-training on tens of millions of sequences; (2) ML integration with chemical mapping data (SHAPE/DMS), as covered by Atomic AI's 2024 patents; (3) end-to-end differentiable 3D folding architectures that jointly learn geometry prediction and structure construction; (4) equivariant graph neural networks for RNA-protein interface modeling; and (5) high-performance computing infrastructure for distributed RNA structure prediction pipelines.

AI AGENTS

AI APPLICATIONS

OTHERS

INDUSTRIES

EXPLORE

ENGAGE

SUPPORT & SERVICES

Your Agentic AI Partner
for Smarter Innovation

Great, Please verify your email.

RNA Structure Prediction 2026 — PatSnap Eureka

RNA Structure Prediction: Deep Learning Displaces Thermodynamics