:sparkles: SMRTnet :sparkles:

This is a PyTorch implementation of our study:

:dart: SMRTnet: Predicting <ins>s</ins>mall <ins>m</ins>olecule and <ins>R</ins>NA <ins>t</ins>arget interactions using deep neural <ins>net</ins>works

Small molecules can bind RNAs to regulate their fate and functions, providing promising opportunities for treating human diseases. However, current tools for predicting small molecule-RNA interactions (SRIs) require prior knowledge of RNA tertiary structures, limiting their utility in drug discovery. Here, we present SMRTnet, a deep learning method to predict SRIs based on RNA secondary structure. By integrating two large language models, convolutional neural networks, graph attention networks, and an attention-based multimodal data fusion model, SMRTnet achieves high performance across multiple experimental benchmarks, substantially outperforming existing state-of-the-art tools. For wet-lab validation, we conducted a large-scale experimental assessment on SMRTnet predictions for 10 disease-associated RNA targets (e.g. mRNA of undruggable proteins, onco-miRNAs, viral RNAs, and RNA repeat expansions), identifying 40 hits of RNA-targeting small molecules with nanomolar-to-micromolar dissociation constants using microscale thermophoresis (MST). Focusing on the MYC internal ribosome entry site (IRES) as a target, SMRTnet-predicted small molecules showed binding scores correlated closely with observed validation rates. Notably, one predicted compound downregulated MYC expression, inhibited proliferation, and promoted apoptosis in three cancer cell lines. Taken together, SMRTnet expands the scope of feasible RNA targets and accelerates the discovery and development of RNA-targeting therapeutics. <img src="figs/workflow.png" width=100% /> Overview of SMRTnet

:round_pushpin: Cite us

If you use this tool in your research, we kindly ask that you cite our paper:

Title: Predicting small molecule–RNA interactions without RNA tertiary structures

Author: Fei Y, Wang P, Zhang J, Shan X, Cai Z, Ma J, Wang Y, and Zhang QC

Journal: Nature Biotechnology, 2026 (5-year Journal Impact Factor: 59.5)

Paper ink: https://www.nature.com/articles/s41587-025-02942-z

:telephone: Contact us

Please contact us if you are interested in our work or potential academic collaborations.

Dr. Yuhan Fei, School of Life Sciences, Tsinghua University, Posdoc, yuhan_fei@outlook.com
Jiasheng Zhang, School of Life Sciences, Tsinghua University, PhD student, zjs21@mails.tsinghua.edu.cn

:book: Table of contents

:one: Getting Started
- Install via PyPI
- Run SMRTnet via Google Colab
:two: Download pre-trained models
:three: Repo Structure
:four: Datasets
:five: Usage
- How to train your own model
- How to test model performance
- How to perform interence for novel interactions
- How to benchmark on known interactions
- How to identify potential binding sites
:six: Referenced Repos
:seven: Copyright and License
:eight: Patent Declaration
:nine: Disclaimer

Getting started

Please run the following command to check your CUDA version before installing SMRTnet:

nvidia-smi

nvcc --version

:heavy_exclamation_mark: Note: To install Torch and DGL versions compatible with your CUDA setup, please refer to the following URLs:

Torch: https://pytorch.org/get-started/previous-versions/
DGL: https://www.dgl.ai/pages/start.html

:pushpin: Install via PyPI

1) The Stable version for installation (Recommend)

conda create -n smrtnet python=3.8.10
conda activate smrtnet
pip install torch==2.4.1+cu118 torchvision==0.19.1+cu118 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu118
pip install smrtnet
conda install dglteam/label/th24_cu118::dgl

The stable version of SMRTnet environment is also available on Zenodo (https://zenodo.org/records/14970392) for offline installation.

2) The Latest version for installation

conda create -n smrtnet_latest python=3.8.10
conda activate smrtnet_latest
pip install torch torchvision
pip install smrtnet-latest
conda install dglteam/label/th24_cu121::dgl

:heavy_exclamation_mark: Note: The table now explicitly details these two installation pathways and provides complete, version-specific dependency lists for reference:

| Package | Stable version | Latest version | Remarks | |---------|----------------|----------------|---------| | babel | 2.17.0 | 2.17.0 | Up to date | |charset-normalizer| 3.3.2 | 3.3.2 | Required | |dgllife| 0.3.2 | 0.3.2 | Up to date | |dgl| 2.4.0.th24.cu118 | 2.4.0.th24.cu121 | Up to date | |matplotlib| 3.7.5 | 3.7.5 | Constrained by dependencies | |networkx| 2.8.8 | 3.1 | Constrained by dependencies | |huggingface-hub| 0.29.1 | 0.34.4 | Up to date | |notebook| 7.3.2 | 7.3.3 | Constrained by dependencies | |numpy| 1.20.3 | 1.24.4 | Constrained by dependencies | |pandas| 1.2.4 | 2.0.3 | Constrained by dependencies | |prefetch_generator| 1.0.3 | 1.0.3 | Up to date | |prettytable| 3.11.0 | 3.11.0 | Constrained by dependencies | |pytorch-lightning| 1.1.5 | 2.4.0 | Constrained by dependencies | |python| 3.8.10 | 3.8.10 | Required | |rdkit| 2022.3.5 | 2022.3.5 | Required | |scikit-learn| 0.24.2 | 1.3.2 | Constrained by dependencies | |scipy| 1.10.1 | 1.10.1 | Constrained by dependencies | |seaborn| 0.13.2 | 0.13.2 | Up to date | |tensorboard| 2.14.0 | 2.14.0 | Constrained by dependencies | |tensorboardX| 2.6.2.2 | 2.6.2.2 | Constrained by dependencies | |torch| 2.4.1+cu118 | 2.4.1 | Constrained by dependencies | |tqdm| 4.67.1 | 4.67.1 | Up to date | |transformers| 4.28.1 | 4.28.1 | Required | |xsmiles| 0.2.2 | 0.2.2 | Up to date |

‘Required’ denotes that SMRTnet requires this specific version of the indicated package for proper operation;
‘Up to date’ indicates that the dependency is at the latest version of the indicated package;
‘Constrained by dependencies’ explains that, although a newer version is available, compatibility with other dependencies limits the update.

:heavy_exclamation_mark: Note: We conducted usability tests of both installation methods with a diverse group of users to validate the setup process. The hardware and software details are listed below:

| GPUs | Driver version | CUDA version| Stable version | Latest version | |-------------|----------------|---------------|--------------|-------------| | H20 (96G) | 570.158.01 | 12.8 | :white_check_mark: | :white_check_mark: | | RTX 4090 (24G) | 570.124.06 | 12.8 | :white_check_mark: | :white_check_mark: | | RTX 4090 (24G) | 550.135 | 12.4 | :white_check_mark: | :white_check_mark: | | RTX 2080 (11G) | 535.216.03 | 12.2 | :white_check_mark: | :white_check_mark: | | A100 40G (40G) | 560.35.03 | 12.6 | :white_check_mark: | :white_check_mark: | | A800 80G (80G) | 450.248.02 | 11.0 | :white_check_mark: | :white_check_mark: |

:pushpin: Run SMRTnet via Google Colab

We have developed an online jupyter-notebook that allows installation-free execution of SMRTnet directly in a web browser via Google Colab. This solution supports both inference and interpretability functionalities while eliminating system-specific installation issues with limited GPU resources.

Step 1: Please click the followting link: https://drive.google.com/drive/folders/1HQo3o2saY5U9vPqebz4ZdpCVVQXqw0q_?usp=sharing, and copy the shared folder to your own Google Drive by dragging it into your Drive interface:

Step 2: Please follow the step-by-step instructions provided in the SMRTnet.ipynb notebook to run SMRTnet directly: https://colab.research.google.com/drive/1pm5ZCD8cFRvPA9RPvtEaCHoU1p5X5v4Y?usp=sharing

<img src="figs/colab_run.png" width=100% /> :bangbang: If you encounter any issues during the installation process, please feel free to report the problem in the 'Issues module' or contact us directly via email at yuhan_fei@outlook.com or zjs21@mails.tsinghua.edu.cn.

Download pre-trained models from Zenodo

<img src="figs/architecture.png" width=100% /> The architecture of SMRTnet  Since the pre-trained models used in SMRTnet are large, we have uploaded them to Zenodo for direct download. Users are required to download the pre-trained models, including the RNA language model (RNASwan-seq), the chemical language model (MoLFormer), and the SMRTnet model, from the link below and place them in the SMRTnet folder (see the “Repo Structure” section below for details).

Pre-trained models used in SMRTn

SMRTnet

Install / Use

README