Use RNA-FM Online
Commercially Available RNA-FM No-Code Web Server
RNA-FM: RNA Foundation Model for Structure and Function Prediction
RNA-FM is an interpretable RNA foundation model pretrained on unannotated RNA sequences, introduced in the paper "Interpretable RNA Foundation Model from Unannotated Data for Highly Accurate RNA Structure and Function Predictions" by Jiayang Chen, Zhihang Hu, Siqi Sun, Qingxiong Tan, Yu Li and colleagues, with affiliations including The Chinese University of Hong Kong, Fudan University, Shanghai AI Laboratory, MIT and the Broad Institute. Code and weights are published on GitHub.
Many RNA tasks, from structure prediction to function annotation, are limited by the small amount of labeled data. RNA-FM instead learns from sequence alone, so its embeddings can be reused across tasks.
The model is a BERT-style transformer trained with masked nucleotide prediction on about 23 million non-coding RNA sequences from RNAcentral. The authors report that the learned embeddings capture sequential and evolutionary information and improve results on structure and function tasks.
Property | Detail |
|---|---|
Architecture | 12 transformer encoder blocks, 640 hidden size, 20 attention heads |
Maximum input length | 1024 |
Pre-training data | RNAcentral, 23.7 million sequences after 100 percent identity deduplication |
Objective | Masked language modeling (15 percent of tokens) |
Training | Eight A100 80 GB GPUs for about one month |
Why a Foundation Model for ncRNA Matters
Because RNA-FM needs no labels for pre-training, it can serve as a general feature extractor. The authors apply it to secondary and 3D structure prediction, SARS-CoV-2 genome analysis, protein-RNA binding prediction and 5' UTR mean ribosome load prediction.
RNA-FM Benchmark Results for Structure and Function
RNA 3D closeness (long-range contacts)
On the RNAcontact benchmark with an 8 angstrom contact threshold, top precision values (L/10, L/5, L/2, L/1) were:
Method | L/10 | L/5 | L/2 | L/1 |
|---|---|---|---|---|
RNAcontact (100-model ensemble) | 0.48 | 0.45 | 0.40 | 0.33 |
ResNet32 with RNA-FM | 0.68 | 0.66 | 0.62 | 0.53 |
ResNet32 with RNA-FM and transfer learning | 0.88 | 0.85 | 0.79 | 0.66 |
Other tasks
Protein-RNA binding: mean AUPRC across 17 RBP subsets was 0.815 for sequence alone and 0.824 with RNA-FM.
5' UTR mean ribosome load: on a random test set R2 was 0.853 with RNA-FM versus 0.826 for sequence alone.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with RNA-FM on Tamarind Bio
Structure prediction support: use embeddings for secondary structure and 3D contact or distance prediction.
Viral genome analysis: the authors apply the model to SARS-CoV-2 genome structure and evolution.
Protein-RNA binding: improve prediction of RNA-binding protein interactions.
UTR engineering: predict mean ribosome load from 5' UTR sequences.
How to Use RNA-FM on Tamarind Bio
Log in: sign in at tamarind.bio and open the RNA-FM tool.
Provide RNA sequences: enter or upload your RNA sequences (the model was trained with a maximum input length of 1024).
Select the output: choose embeddings or the prediction task, if exposed.
Set parameters: adjust any options the tool exposes.
Run the job: submit and wait for completion.
Download outputs: save embeddings or predictions.
Use downstream: feed embeddings into your own models or compare predictions with experiments.
Things to Keep in Mind
The authors note that gains in functional tasks are smaller than in structural tasks, partly because of sequence distribution differences from the pre-training data.
RNA-FM struggles to separate short from long ncRNAs and to distinguish lncRNAs by species, though it captures a rough evolutionary trend.
For small datasets the authors suggest transfer learning to avoid overfitting; fine-tuning may suit large, complex tasks.
Source: Chen J, Hu Z, Sun S, Tan Q, Wang Y, Yu Q, Zong L, Hong L, Xiao J, Shen T, King I, Li Y. Interpretable RNA Foundation Model from Unannotated Data for Highly Accurate RNA Structure and Function Predictions. arXiv:2204.00300. Code: https://github.com/ml4bio/RNA-FM. Web server: https://proj.cse.cuhk.edu.hk/rnafm/