- Notifications
You must be signed in to change notification settings - Fork11
The state of the art Deep CNN neural network for de novo sequencing of tandem mass spectra
lkytal/PepNet
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
Link to publication:AccurateDe Novo Peptide Sequencing Using Fully Convolutional Neural Networks
The state of the art Deep CNN neural network forde novo sequencing of tandem mass spectra, currently works on unmodified HCD spectra of charges 1+ to 4+.
Free for academic uses. Licensed under LGPL.
Visithttps://denovo.predfull.com/ to try online prediction
- 2023.04.27: 2nd Revised version.
- 2022.11.28: Revised version.
- 2021.12.28: First version.
Based on the structure of the residual convolutional networks. Current precision (bin size): 0.1 Th.
After clone this project, you should download the pre-trained model (model.h5) fromzenodo.org and place it into PepNet's folder.
- Will only output unmodification sequences.
- This model assumes aFIXED carbamidomethyl on C
- The length of output peptides are limited to =< 30
Recommend to install dependency viaAnaconda
- Python >= 3.7
- Tensorflow >= 2.5.0
- Pandas >= 0.20
- pyteomics
- numba
Packages Required for traning:
- Tensorflow-addons
Sample output looks like:
| TITLE | DENOVO | Score | PPM Difference | Positional Score |
|---|---|---|---|---|
| spectra 1 | LALYCHQLNLCSK | 1.0000 | -3.8919184 | [1.0, 0.9999956, 1.0, 1.0, 1.0, 1.0, 0.99999976, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0] |
| spectra 2 | HEELMLGDPCLK | 1.0000 | 4.207922 | [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99999976, 1.0] |
| spectra 3 | AGLVGPEFHEK | 1.0000 | 0.54602236 | [1.0, 1.0, 1.0, 1.0, 1.0, 0.99999917, 1.0, 1.0, 1.0, 1.0, 1.0] |
Simply run:
python denovo.py --input example.mgf --model model.h5 --output example_prediction.tsv
The output file is in MGF format
- --input: the input mgf file
- --output: the output file path
- --model: the pretrained model
Typical running speed: sequencing 10,000 spectra in ~59 seconds on a NVIDIA A6000 GPU.
We provide sample data on
example.mgf file contains ground truth spectra (randomly sampled fromNIST Human Synthetic Peptide Spectral Library), while theexample.tsv file contains pre-run predictions.
Also, you can runpython evaluation.py --novorst example_prediction.tsv to generate figures presenting the de novo performance.
Seetrain.py for sample training codes
As we demonstrated in the manuscript, we follow the DeepNovo-DIA's method to generate a pseudo-spectrum of each precursor, so we can perform De novo like it's a DDA spectrum. These steps are describe in DeepNovo-DIA's Method section 'Precursor feature detection' and 'In-house database searching'. We actually reused the pseudo-spectrum MGF generated by DeepNovo-DIA.
Also, Visithttps://www.predfull.com/ to check our previous project on full spectrum prediction
About
The state of the art Deep CNN neural network for de novo sequencing of tandem mass spectra