patmjen / maldi-learn

Software library for Maldi-Tof preprocessing and machine learning analysis.

maldi-learn

maldi-learn is a Python library for MALDI-TOF preprocessing and machine learning analysis, developed by members of the Machine Learning and Computational Biology Lab of Prof. Dr. Karsten Borgwardt.

This library is a work in progress. Below, we provide some information for loading “DRIAMS”, a large-scale data set of mass spectra and performing some simple machine learning tasks with it. Stay tuned for more information!

Installation

The installation of this package requires poetry, a Python package handler. This is only required if you want to try out the code for yourself or even contribute to maldi-learn. In cases where maldi-learn is used as a dependency, there is no need for you to do anything.

To set up the development environment, run poetry install in the project root. To run commands in the associated virtual environment of this package, run poetry shell to spawn a shell.

Handling multiple Python versions

This project requires at least python version 3.7. If you want to contribute to maldi-learn, we would recommend a development setup in which an appropriate Python version is installed using pyenv, and then marking this folder for usage with this version:

$ pyenv install 3.7.4  # Install python 3.7.4 using pyenv
$ pyenv local 3.7.4    # Mark python version 3.7.4 for usage in this folder
$ poetry install       # Setup the virtual environment for development

Getting the data

Download the data from DOI: 10.5061/dryad.bzkh1899q and extract all files into one shared directory, which we will refer to as the root data directory.

Examples

The folder examples contains scripts that demonstrate a part of the API. We recommend to start with classification.py. This examples demonstrates a full, albeit simplified example, which we will subsequently discuss here.

We first import all required packages. Notice that we use dotenv, a package for handling environment variables in .env files. This makes it very easy to specify the root folder of the data set. If you downloaded the data set, you can add the following line to a .env file in the root folder of this repository:

DRIAMS_ROOT=/path/to/your/download

The code below automatically looks for this file and stores its value in the variable DRIAMS_ROOT. Alternatively, you can also set an environment variable using export DRIAMS_ROOT=/path/to/your/download.

import dotenv
import os

from maldi_learn.driams import DRIAMSDatasetExplorer
from maldi_learn.driams import DRIAMSLabelEncoder
from maldi_learn.driams import load_driams_dataset

from maldi_learn.utilities import stratify_by_species_and_label
from maldi_learn.vectorization import BinningVectorizer

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score

dotenv.load_dotenv()
DRIAMS_ROOT = os.getenv('DRIAMS_ROOT')

Moving on, we perform some book-keeping and loading in the code:

explorer = DRIAMSDatasetExplorer(DRIAMS_ROOT)

driams_dataset = load_driams_dataset(
            explorer.root,
            'DRIAMS-A',
            ['2015', '2017'],
            'Staphylococcus aureus',
            ['Ciprofloxacin', 'Penicillin'],
            handle_missing_resistance_measurements='remove_if_all_missing',
)

As you can see, the data set will contain spectra from 2015 and 2017, from the DRIAMS-A site, with a species of Staphylococcus aureus, and the two antibiotics ciprofloxacin and penicillin. The most interesting part of this call is the parameter for handle_missing_resistance_measurements. With its current value, a sample will be removed if there are no resistance measurements available. Other values are remove_if_any_missing, meaning that a single missing value already triggers removal, or keep, meaning that no removal will be performed.

Next we continue with binning the spectra. For this example, we only use 100 bins. Afterwards, we prepare a stratified split based on ciprofloxacin (thus ensuring that the prevalence of the resistance values is the same; strictly speaking this is not necessary here as we only loaded one species, but this example is supposed to be somewhat instructive).

# bin spectra
bv = BinningVectorizer(100, min_bin=2000, max_bin=20000)
X = bv.fit_transform(driams_dataset.X)

# train-test split
index_train, index_test = stratify_by_species_and_label(
    driams_dataset.y, antibiotic='Ciprofloxacin'
)

Next, we create (binary) labels and train a standard logistic regression classifier. In a real application, we would also be performing a hyperparameter search, but for this example, a single demonstration based on our train–test split is sufficient.

y = driams_dataset.to_numpy('Ciprofloxacin')

lr = LogisticRegression()
lr.fit(X[index_train], y[index_train])
y_pred = lr.predict(X[index_test])

print(accuracy_score(y_pred, y[index_test]))