Sturgeon: Open-Source Deep Learning for Simultaneous Base Calling and Methylation Detection in Nanopore Sequencing

This project introduces Sturgeon, a deep learning framework enabling accurate base calling and DNA methylation quantification using nanopore sequencing data. We address limitations in protein accessibility and model training by engineering commercial-grade Pif1 helicase and CsgG/HfaB proteins via random amino acid substitution and providing an open-source platform. Sturgeon combines convolutional neural networks (CNNs) and reverse LSTMs to process raw signals, achieving 88.76%-91.88% base calling accuracy and high methylation correlation with BS-seq (Pearson ≥0.96) across species. Designed for extensibility, Sturgeon serves as a foundation for advancing nanopore sequencing of genetic and epigenetic modifications.

Key innovations: Novel protein designs, open-source framework, and simultaneous sequence/modification analysis.

Building and Using Sturgeon

Preparing the Python Environment

Create a virtual environment using Conda. The Python scripts require numpy (version 20.0 or higher) and pytorch (version 2.0 or higher) with CUDA above 12.0

conda create -n Sturgeon python=3.10
conda activate Sturgeon
pip install numpy torch==2.6.0+cu124

Building the cpp_components

Sturgeon utilize a C++ code to accelerate CTC decoding process. It should be built before running basecalling.

git clone https://github.com/xiaochuanle/Sturgeon.git
cd Sturgeon/cpp_components
mkdir build && cd build
cmake .. -DCMAKE_CUDA_COMPLIER="your nvcc path" -DCMAKE_PREFIX_PATH="your libtorch path"

You should change the cuda complier and libtorch path in CMakeLists.txt or add the correct path while running cmake.

Running Sturgeon Basecalling and Methylation detection

Suturgeon combines basecalling and methylation detection together in one run. And our trained model with Qitan Sequencing data is in ./Saved_Models

Here is an example.

python ./basecall.py 
./Saved_Models/cnn_reverselstm_ctc.pt
/h5_dir/
-o 
/output_file_path
--batch_size 1024 
--num_get_chunk 16 
--num_decode 8 
--num_stitch 8 
--output_moves 
--call_mods // optional
--ref_genome 
/your_ref_genome_path
--mod_module_path 
./Saved_Models/mod_bilstm_attn.pt
--mod_result_path 
/path_to_save_mod_res/mod_result.txt