pi-HelixNovo

pi-HelixNovo is a de novo sequencing model based on the Transformer architecture, using a MS2 spectrum and its complementary spectrum as inputs and generating the corresponding peptides. The model weights we have trained are avaliable at https://zenodo.org/records/10405582. If you use pi-HelixNovo in your work, please cite the following publication:

Tingpeng Yang, Tianze Ling, Boyan Sun, Zhendong Liang, Fan Xu, Xiansong Huang, Linhai Xie, Yonghong He, Leyuan Li, Fuchu He, Yu Wang, Cheng Chang, Introducing π-HelixNovo for practical large-scale de novo peptide sequencing, Briefings in Bioinformatics, Volume 25, Issue 2, March 2024, bbae021, https://doi.org/10.1093/bib/bbae021

Hardware requirements

GPU

• Train a model from scratch: An NVIDIA GPU with enough computing power and memory (e.g., Tesla V100 with 32GB of memory).
• Evaluate a pretrained model or de novo sequencing: CPU (slow) or an NVIDIA GPU (e.g., GeForce RTX 3050 with 4GB of memory, ..., Tesla V100 with 32GB of memory). If the error “CUDA OUT OF MEMORY” occurs, please decrease the “predict_batch_size” in the config.yaml.

Storage space

• Windows system: C drive with enough storage space.
• Linux system: A hard disk with enough storage space.

The usage of our code

Preparation:

For linux users

Enter the code folder

conda env create -f main_env.yaml
conda activate main_env 

For windows users

Refer to Run pi-HelixNovo in Docker

Specify the device

--gpu=-1 # run pi-HelixNovo on CPU
--gpu=0 # run pi-HelixNovo on GPU 0
--gpu=0,1 # run pi-HelixNovo on GPU 0,1; Distributed deep learning

Train a model from scratch:

python main.py --mode=train --gpu=0 --config=./config.yaml --output=train.log --peak_path=./sample_data/training_set/*.mgf --peak_path_val=./sample_data/validation_set/*.mgf

Evaluate a pretrained model

python main.py --mode=eval --gpu=0 --config=./config.yaml --output=evaluate.log --peak_path=./sample_data/validation_set/*.mgf --model=the_path_of_your_model

De novo sequencing

python main.py --mode=denovo --config=./config.yaml --gpu=0 --output=denovo.log --peak_path=./sample_data/denovo_sample/*.mgf --model=the_path_of_your_model

The results will be shown in the current folder as denovo_denovo.txt because --output=denovo.log

TITLE	Peptide	p
27	VLEGHAEK	0.95
28	LQHEAATATQK	0.93
29	KEAAPPPK	0.96

"TITLE" is the TITLE information of the MS spectrum in the corresponding mgf file, and "p" is the confidence score of the peptide sequence.

The config.yaml used in pi-HelixNovo

To train models on the nine-species benchmark dataset, please use config.yaml

To train models on the merged dataset of PXD008808, PXD011246, PXD012645 and PXD012979, please use merge-config.yaml.

To train models on the MSV000081142 dataset, please use config.yaml

Recommendation

For practical large-scale de novo peptide sequencing, we highly recommend utilizing the model weight "MSV000081142-epoch-5-step-800000.ckpt", which was trained on the MSV000081142 dataset, while employing the "config.yaml" configuration file.