Skip to content

Latest commit

 

History

History
595 lines (436 loc) · 14.6 KB

README.md

File metadata and controls

595 lines (436 loc) · 14.6 KB
Raw

Table of content

Documentation

What in this project ?

This project predict basic molecule's properties, from its fingerprint features, using machine learning

  • Setup python environment using miniconda
  • Parse argument from the CLI
  • Load a dataset and perform feature extraction
  • Split the dataset
  • Train a classification machine learning model
  • Save the pretrained model
  • Predict the molecule properties using pretrained model
  • Evaluate the model
  • Package the app inside a docker image
  • Access the predict method via a Flask API
  • Documentation

Data Exploration

dataset_single.csv

    P1	mol_id	    smiles
0	1	CID2999678	Cc1cccc(N2CCN(C(=O)C34CC5CC(CC(C5)C3)C4)CC2)c1C
1	0	CID2999679	Cn1ccnc1SCC(=O)Nc1ccc(Oc2ccccc2)cc1
2	1	CID2999672	COc1cc2c(cc1NC(=O)CN1C(=O)NC3(CCc4ccccc43)C1=O...
3	0	CID5390002	O=C1/C(=C/NC2CCS(=O)(=O)C2)c2ccccc2C(=O)N1c1cc...
4	1	CID2999670	NC(=O)NC(Cc1ccccc1)C(=O)O

  1. The target we want to predict is binary, 0 or 1. A classification model (such as RandomForest) is suitable for this problem P1

  2. The target (P1) had 82.17% (4108) of 1's and 17.82% (891) of 0's

    1 - 82.17 % - 4108
0 - 17.82 % - 891

  3. A vector (of size 2048) is obtained for each molecule after extracting its fingerprint features (using fingerprint_features method from features_extration module). Therefore, we concatenate this vector to the dataset for each molecule.

  4. Load dataset and create features X and target y vectors for a dataset file using pandas

    def LoadAndProcessData(path=config.path_single):

    print('[INFO] Load the data and extract features')
    isinstance(path, str)

    # load the csv file
    df = pd.read_csv(path)

    # extract features
    df['smiles_features'] = df.smiles.apply(
        lambda x: np.array(fe.fingerprint_features(x)))

    # merge the extracted features to the dataframe
    df_feat = pd.DataFrame()
    for i in range(len(df.smiles_features)):
        df_feat['fe_'+str(i)] = df.smiles_features[i]
    df_feat = df_feat.T.reset_index(drop=True)

    # [fe_0, fe_2, ... fe_2047]
    df_feat.columns = ['fe_'+str(i) for i in range(len(df.smiles_features[0]))]

    X_all = pd.concat([df, df_feat], axis=1)
    X_all.drop(columns=['mol_id', 'smiles_features', 'smiles'], inplace=True)

    y = X_all['P1']
    X = X_all.drop(columns=['P1'], axis=1)
    return X, y

Machine Learning Model

  • Split the data to train and validation 
    def SplitDataset(X, y):
"""
Randomly Split the dataset into training and validation sets: 80%-20%
"""
print('[INFO] Split the dataset: train - validation sets')
X_train, X_valid, y_train, y_valid = train_test_split(
    X, y, test_size=.2, random_state=0)
return X_train, X_valid, y_train, y_valid
  • Classification model: Tree based such as Random Forest using scikit-learn package 
    model = RandomForestClassifier(
        n_estimators=config.n_est,
        max_depth=config.max_depth,
        n_jobs=-1,
        random_state=0)
  • Train the model 
    model.fit(X_train, y_train)
  • Perform prediction 
    model.predict(X_test)
  • Evaluate the model by computing accuracy 
    accuracy = metrics.accuracy_score(y_true, y_pred)

Project Structure

This project is organized as follow

.
├── Dockerfile
├── environment.yaml
├── README.md
├── run_flask_api.sh
├── app
│   ├── cli.py
│   ├── data
│   │   ├── dataset_multi.csv
│   │   ├── dataset_single.csv
│   │   └── dataset_single_folds.csv
│   ├── __init_.py
│   ├── media
│   │   ├── P1.png
│   │   ├── test_api.png
│   │   ├── test_flask.png
│   │   └── test_terminal.png
│   ├── models
│   │   ├── model_pred_multi.bin
│   │   └── model_pred_single.bin
│   └── src
│       ├── config.py
│       ├── feature_extractor.py
│       ├── flask_api.py
│       └── main.py
└── setup.py

Setup the development environment

  1. First download miniconda :

    wget --quiet https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
Miniconda3-latest-Linux-x86_64.sh -b -p ~/miniconda

    Export an environment variable for miniconda

    export PATH=~/miniconda/bin:$PATH

  2. Create an isolated environment to develop the App

    Update conda packages

    conda update -n base conda

    Create and activate a virtual environment app

    conda create -y --name app python=3.6
conda activate app

  3. Install required packages

    conda env update -n app --file environment.yaml

    with environment.yaml

    name: app
channels:
  - conda-forge
  - anaconda
dependencies:
  - rdkit
  - flask
  - gunicorn
  - pip
  - numpy
  - pandas
  - matplotlib
  - scikit-learn
  - click

  4. Check if the environment is activated:

    If the setup is correctly done, the name of the environment (app) should appear at the beginning, like that:

    (app) amine@amine-HP-ZBook-15-G2:~/$

Running the App

You can run the App from a terminal by executing the main.py file with python3

python3 app/src/main.py --help

output

usage: main.py [-h] [--data DATA]

optional arguments:
-h, --help   show this help message and exit
--data DATA  Specify the path of the dataset

Example:

python3 app/src/main.py --data app/data/dataset_single.csv

Packaging the App

Setup

To do that we need to create two files, setup.py and cli.py

  1. setup.py: configure the application

    from setuptools import setup

setup(
    name="predmol",
    version="1.0.0",
    license='MIT',
    author='Amine Hadj-Youcef',
    author_email='amine.hadjyoucef@gmail.com',
    description='This application uses machine learning to predict basic properties of a molecule...',
    packages=["app"],
    entry_points='''
        [console_scripts]
        predmol = app.cli:main
        '''
)

  2. cli.py contains four commands: train, predict, evaluate

     import sys
 import click
 from app.src.main import Train, Predict, Evaluate, LoadAndProcessData, SplitDataset


 @click.group()
 @click.version_option("1.0.0")
 def main():
     """Molecule's basic properties prediction App"""
     pass


 @main.command()
 @click.option('--data_path', '-d', type=str, required=True, default='app/data/dataset_single.csv', help="Please enter the path of data in order to train the model")
 def train(data_path):
     """Train a machine learning model for prediction and save the pretrained model to disk"""
     X, y = LoadAndProcessData(data_path)
     X_train, X_valid, y_train, y_valid = SplitDataset(X, y)
     click.echo(Train(X_train, y_train))


 @main.command()
 @click.option('--path_x_test', '-p', type=str, required=True, default='data/dataset_single.csv', help="Please enter the path of data in order to perform prediction")
 def predict(path_x_test):
     """Perform prediction using a pretrained Machine Learning prediction model"""
     click.echo(Predict(path_x_test))


 @main.command()
 @click.option('--y_pred', '-p', type=float, required=True, help='Enter the predicted array')
 @click.option('--y_true', '-t', type=float, required=True, help='Enter the true array')
 def evaluate(y_pred, y_true):
     """Evaluate the prediction model by measuring accuracy"""
     click.echo(Evaluate(y_pred, y_true))


 if __name__ == "__main__":
     main()

Install the packaged App

Install the predmol package by entering in the terminal (from the root of the project)

pip install -e .

Check if the app is correctly installed

which predmol
/home/amine/miniconda/envs/app/bin/predmol

the application is correctly installed under the app environment.

How to use it?

Issue the following command to know how to use the App

predmol --help

output:

Usage: predmol [OPTIONS] COMMAND [ARGS]...

Molecule's basic properties prediction App

Options:
--version Show the version and exit.
--help Show this message and exit.

Commands:
evaluate Evaluate model
predict Perform prediction using a pretrained Machine Learning...
train Train a machine learning model for prediction

For instance, in order to invoke the train command, you can do

premol train --help
Usage: predmol train [OPTIONS]

Train a machine learning model for prediction and save the pretrained
model to disk

Options:
-d, --data_path TEXT  Please enter the path of data in order to train the
                        model  [required]

--help                Show this message and exit.

so you need to mention the path to the dataset you want to train your model on

predmol train --data_path app/data/dataset_single.csv

Similarly, you can use predict and evaluate commands:

predmol predict --help

Usage: predmol predict [OPTIONS]

  Perform prediction using a pretrained Machine Learning prediction model

Options:
  -p, --path_x_test TEXT  Please enter the path of data in order to perform
                          prediction  [required]

  --help                  Show this message and exit.

and

predmol evaluate --help

Usage: predmol evaluate [OPTIONS]

  Evaluate the prediction model

Options:
  -p, --y_pred FLOAT  Enter the predicted array  [required]
  -t, --y_true FLOAT  Enter the true array  [required]
  --help              Show this message and exit.

Note: These commands could be personalized depending on the way we want to interact with them.

Dockerization

Another method for packaging the App and ensure scalability, we use Docker by creating a Dockerfile.

Setup

For simplicity and stability we start from the base image of miniconda3, then we install required python packages and create a virtual conda environment.

Dockerfile

FROM  continuumio/miniconda3
LABEL Author, Amine HadjYoucef

ENV APP_HOME /app
WORKDIR $APP_HOME
COPY . $APP_HOME

#---------------- Prepare the envirennment
RUN conda update --name base conda &&\
    conda env create --file environment.yaml
SHELL ["conda", "run", "--name", "app", "/bin/bash", "-c"]

ENTRYPOINT ["conda", "run", "--name", "app", "python", "src/main.py"]

Build a docker image

We build a docker image, named app, with the following commands

docker build --tag app .

Note:

  1. The Dockerfile must be located in the root directory of the application (in this case)

  2. Make sure Docker is installed on your machine (https://www.docker.com).

Run the docker image

To run the App from the docker image we created (app)

docker run --rm -ti app

Flask API

Another way to distribute and use the App is by packaging it as an API, for instance using the Flask API framework

Setup

Install Flask inside conda virtual environment

conda install -n app -c anaconda flask

Create a python file for the api (app/src/flask_api.py) and configure Flask

export FLASK_APP=api
export FLASK_ENV=development

with flask_api.py

from flask import Flask
from app.src.main import Predict, LoadAndProcessData

app = Flask(__name__)

@app.route('/')
def hello():
    return 'Hello, World!'

if __name__ == '__main__':
    app.run('0.0.0.0', 5000)

Run the API from the terminal

flask run

Open the browser on the port 5000 to check if the API is running: http://127.0.0.1:5000/

test_flask

Simplify the execution with a script

For simplifying the setup I created a script (run_flask_api.sh) to gather Flask setup from one file.

run_flask_api.sh :

#!/bin/bash

export FLASK_APP=app/src/flask_api.py
export FLASK_ENV=development

flask run --port 5000

and it can be executed from the terminal

sh run_flask_api.sh

Create a Prediction Endpoint

To create a prediction endpoint, we import the Prediction method from the main file (app/src/main.py) and create an endpoint for it in the flask API.

The endpoint takes two arguments parameters: the model path and the dataset path, path_model, path_X_test, respectively.

Source code:

from flask import Flask
from app.src.main import Predict, LoadAndProcessData

app = Flask(__name__)

@app.route('/')
def hello():
    return 'Hello, World!'

@app.route('/predict/<path:path_X_test>')
def predict(path_X_test):
    """
    Example:
    path_X_test=app/data/dataset_single.csv
    Link:
        http://127.0.0.1:5000/predict/app/data/dataset_single.csv
    """
    isinstance(path_X_test, str)
    y_pred = Predict(path_X_test)
    return {'status': 'OK', 'y_pred': y_pred.tolist()}

if __name__ == '__main__':
    app.run('0.0.0.0', 5000)

Access the predict endpoint of the API with this URL: http://127.0.0.1:5000/predict/app/data/dataset_single.csv

If everything works fine, on the browser you should see something like that: test_api

and on the terminal you should expect something like that: test_terminal

Futur work

  • Tune hyperparameter of RandomForest model
  • Implement Stratified Cross Validation
  • Implement others Machine Learning model
  • Improve the interaction with the app

Found an issue ? or You want to contribute?

Please create an issue and thank you for your contribution