Alpha-Tri

Alpha-Tri is a deep neural network to score the intensity similarity using all possible fragment ions, resulting in the improvement in peptide detections.

Hardware

A GPU with CUDA support

Main Packages

• PyTorch
• Pyteomics
• numba
• Prosit
• tensorflow

Example on a local PC (win10, NVIDIA GPU 1060)

1. Compile the modified DIA-NN:

   cd Alpha-Tri/DIA-NN_v1.7.12/mstoolkit
   make

   This will generate diann-alpha.exe in the same path.

2. Make a workspace folder containing:

   • diann-alpha.exe compiled by step 1
   • HeLa-1h.mzML, (test data could be downloaded from figshare or PXD005573)
   • lib.tsv (this spectral library could be downloaded from figshare or Pan-Human library, SAL00023)

3. Configure the operating environment by conda

   conda create -n alpha python=3.6 numpy=1.18 pandas=1.0 numba scikit-learn --yes
   conda activate alpha 
   conda install -c bioconda pyteomics python=3.6 --yes
   conda install tensorflow-gpu=1.11 keras=2.2.4 pytorch=1.1.0 cudatoolkit=9.0 -c pytorch --yes

   Of note, as Prosit is trained by tf1.11, advanced version of tf and CUDA may not be compatible.

4. Run Prosit to predict the MS2 for each precursor in lib

   cd Alpha-Tri/Prosit
   python prosit.py --lib workspace_dir/lib.tsv

   This will append the predicted MS2 to each precursor and store the result to lib.pkl.

5. Run DIA-NN:

   cd workspace
   ./diann-alpha.exe --f *.mzML --lib lib.tsv --out diann_out.tsv --threads 4 --qvalue 0.01
   or
   diann-alpha.exe --f *.mzML --lib lib.tsv --out diann_out.tsv --threads 4 --qvalue 0.01 

   Meanwhile, the modified DIA-NN will generate the scores file in workspace.

6. Run Alpha-Tri:

   cd Alpha-Tri/Alpha-Tri
   python main.py -ws workspace_dir --tri (post-scoring only by Alpha-Tri)
   python main.py -ws workspace_dir --xic (post-scoring only by Alpha-XIC)
   python main.py -ws workspace_dir --tri --xic (post-scoring by Alpha-Tri & Alpha-XIC)

   Finally, we get the identification and quantitative result, alpha_out.tsv, in the workspace folder.

Example on Google Colab Pro (linux, NVIDIA GPU P100)

see alpha_tri_on_Colab.ipynb

Troubleshooting

1. AttributeError: 'str' object has no attribute 'decode'

2. TypeError: add_weight() got multiple values for argument 'name'

These may be raised by the version incompatibility of Keras in Prosit (see the discussions in keras-team/keras#14265 and keras-team/keras#13540). You may have to install the same version of Keras:

conda install tensorflow-gpu=1.11 keras=2.2.4 pytorch=1.1.0 cudatoolkit=9.0 -c pytorch --yes