The MOSES service for SingularityNET

The purpose of this service is to use MOSES for supervised classification of high dimensional data sets with many more features than samples, such as whole genome sequencing data or gene expression data. See the OpenCog wiki page about MOSES or this Quick Guide for more detailed information on MOSES.

The user supplies a csv file of sample data with samples/observations as rows and binary valued features (1 for "true" and 0 for "false") as columns with binary sample labels (1 for "case" and 0 for "control") in the first column, along with a yaml file of MOSES program options, cross-validation parameters, and score thresholds for filtering the evolved boolean models.

The service provides a URL link for downloading the set of output files, including copies of the input files, the MOSES log file, tables of output models from each cross-validation fold with their out-of-training-sample scores, a table of the filtered models with their scores on the complete input dataset and the scores for their majority-vote ensemble, and a table of feature counts from the ensemble model.

You can find a detailed description of using the service here.

Building and Running the Service

1. Clone the project:

   $ git clone --recursive https://github.com/MOZI-AI/moses-service.git

2. Go to the project folder to start the docker containers to run the gRPC server and its dependencies (redis, mongo, etc)

   2a. Define the APP_PORT and SERVICE_ADDR variables. Change <PORT_NUM> to the port number you would like to run the result_ui app and <SERVICE_ADDR> to the address of the host where you are going to run the app. If you are running this locally, set SERVICE_ADDR to localhost.

    $ export APP_PORT=<PORT_NUM>
    $ export SERVICE_ADDR=<ADDR>
    $ export FLASK_SERVER=http://<ADDR>:<FLASK_PORT> # e.g http://192.168.1.3:5000
   

   You can also set these values in the docker-compose.yml/docker-compose-dev.yml file in the environments section

   2b. Start the docker containers:

    $ docker-compose -f docker-compose-dev.yml up
   

   N.B If you make any changes to the code make sure you rerun the containers with --build flask. That is, docker-compose -f docker-compose-dev.yml up --build

3. Open a new terminal and install the python dependencies for running the service client on your local system. Run:

   $ pip install grpcio grpcio-tools pyyaml

4. Generate the gRPC code from the protobufs. Run the following:

   $ ./build.sh

   Note: Make sure you have set execute permission for build.sh. If not, just run chmod +x build.sh

5. On the new terminal, while still in the project directory, call the service client. Replace with a .yaml file containing the moses and cross-validation with the path to file you want to run analysis on Inputs:

• options: yaml file with MOSES algorithm and cross-validation parameters. See below for examples.

• data: csv file with observations or samples in rows and binary features in columns labeled 1 for TRUE and 0 for FALSE for each sample. The first column should indicate the category label of the sample (1 for case and 0 for control). See the doc here for a discussion of prepairing specific experimental data types.

  $ python -m service.moses_service_client <options-file> <dataset-file>

  This will output a link where you can poll the status and download the result files once the analysis is finished.

NOTE: You can find a sample options.yaml file in the tests/data directory of the project

Options

moses_opts: "-j8 --balance 1 \
  -m 10000 -W1 \
  --output-cscore 1 --result-count 100 \
# feature selection parameters
  --enable-fs 1 --fs-algo simple --fs-target-size 4 \
  --fs-focus all --fs-seed init \
# hill climbing parameters
  --hc-widen-search 1 --hc-crossover-min-neighbors 5000 \
  --hc-fraction-of-nn .3 --hc-crossover-pop-size 1000 \
  --reduct-knob-building-effort 1 --complexity-ratio 3"

cross_val_opts:
    folds: 3
    random_seed: 2
    test_size: 0.3

target_feature: "case"

see here for a complete description of MOSES options.