SpecEncoder

Source code for SpecEncoder: Deep Metric Learning for Accurate Peptide Identification in Proteomics. Link to paper: https://academic.oup.com/bioinformatics/article/40/Supplement_1/i257/7700866

Error in Code and model weights have been FIXED

Free for academic uses. Licensed under LGPL.

Visit https://predfull.com/ to check related works

Update History

• 2024.08.09: Error in training codes have been FIXED.
• 2024.07.14: Error in Code and model weights have been FIXED.
• 2024.03.07: Second version.
• 2023.10.28: First version.

Method

Based on the structure of the residual convolutional networks.

Different workflows:

Required Packages

Recommend to install dependency via Anaconda

• Python >= 3.7
• Tensorflow >= 2.5.0
• Pandas >= 0.20
• pyteomics
• numba
• Tensorflow-addons

Usage

After clone this project, you should download the pre-trained model (encoder.h5) from zenodo.org and place it into SpecEncoder's folder.

You can find model files for both charge 2+ and 3+. Note that charge 3+ model files is larger as we used 2x layers on charge 3+ for better performance.

Convert Spectra into vectors

First we convert query into vectors:

python encode.py --query query.mgf --model encoder.h5 --output query.pkl

Then we convert target library and decoy library into vectors:

python encode.py --query library.mgf --model encoder.h5 --output library.pkl

python encode.py --query decoy.mgf --model encoder.h5 --output decoy.pkl

Typical running speed: convert around 700 spectra in 1 second on a NVIDIA A6000 GPU.

Generate theoretical spectra

If we have sequences that don't have experimental spectra, we can predict theoretical spectra using predfull (you can find pm.h5 here, also, note that the predfull.py here is modified to suit this project):

python predfull.py --input example_db.tsv --model pm.h5 --output db_predicted.mgf --decoy decoy_predicted.mgf

Note that this script will also generate DECOY database using reversed peptides.

Searching

We can do final search in 3 approaches:

1. Spectral library search
2. Database search
3. Mixed search

In mixed search spectral library will overwrite database spectra when share a same peptide.

Then we can perform searching:

python search.py --query query.pkl --mode 1 --lib_vec library.pkl --lib_decoy decoy.pkl --output result.tsv

Train this model

See train.py for sample training codes. Note that the input MGF should have multiple instance for each peptide, i.e., only peptide have replicates used for training.