Prosit

Prosit is a deep neural network to predict iRT values and MS2 spectra for given peptide sequences. You can use it at proteomicsdb.org/prosit/ without installation.

Hardware

Prosit requires

• a GPU with CUDA support

Installation

Prosit requires

• Docker 17.05.0-ce
• nvidia-docker 2.0.3 with CUDA 8.0 and CUDNN 6 or later installed
• make 4.1

Prosit was tested on Ubuntu 16.04, CUDA 8.0, CUDNN 6 with Nvidia Tesla K40c and Titan Xp graphic cards with the dependencies above.

The time installation takes is dependent on your download speed (Prosit downloads a 3GB docker container). In our tests installation time is ~5 minutes.

Model

Prosit assumes your model to be in a directory that includes:

• model.yml - a saved keras model
• config.yml - a model specifying names of inputs and outputs of the model
• weights file(s) - that follow the template weights_{epoch}_{loss}.hdf5

You can download a pre-trained model for HCD fragmentation prediction on https://figshare.com/projects/Prosit/35582.

Usage

The following command will load your model from /path/to/model/. In the example GPU device 0 is used for computation. The default PORT is 5000.

make server MODEL=/path/to/model/

Example

Please find an example input file at example/peptidelist.csv. After starting the server you can run:

curl -F "peptides=@examples/peptidelist.csv" http://127.0.0.1:5000/predict/

The example takes about 4s to run. An expected output file can be found at `examples/output_msms.txt`.

Using Prosit on your data

You can adjust the example above to your own needs. Send any list of (Peptide, Precursor charge, Collision energy) in the format of /example/peptidelist.csv to a running instance of the Prosit server.

Please note: Sequences with amino acid U, O, or X are not supported. Modifications except "M(ox)" are not supported. Each C is treated as Cysteine with carbamidomethylation (fixed modification in MaxQuant).

Pseudo-code

1. Load the model given as MODEL environment variable
2. Start a server and wait for inputs
3. On incomming request
   • transform peptide list to model input format (numpy arrays)
   • predict fragment intensity with loaded model for given peptides
   • transform prediction to msms.txt output format and return response