/analogue-split

pip install analoguesplit

Primary LanguageJupyter NotebookMIT LicenseMIT

Analogue Split

A Chemically Biased Parametric Data Splitting Method

Overview

The Analogue Split method is designed to analyze and improve the robustness of machine learning models by considering activity cliffs in molecular datasets. Activity cliffs are pairs of similar molecules with significantly different biological activities, which can challenge the performance of predictive models.

This package provides tools to:

  1. Ensure a specified fraction of the test set molecules are involved in activity cliffs.
  2. Analyze model performance as a function of the proportion of activity cliffs in the test set.
  3. Visualize these analyses through gamma plots.

Installation

You can install the package from PyPI using:

pip install analoguesplit

Usage

Parameters

  • gamma: Fraction of the test set comprising of activity cliffs.
  • omega: Similarity threshold to create edges between molecules.
  • test_size: Fraction of the dataset to be used as the test set.
  • X: Feature vector (molecular fingerprints).
  • y: Label vector (biological activities).

API

func set_random_seed

Sets a random seed for reproducibility.

def set_random_seed(seed: int) -> None:

func calculate_fp

Calculates molecular fingerprints for a list of molecules.

def calculate_fp(mols: list[Chem.rdchem.Mol], fp: str = "ecfp4") -> np.ndarray:

func convert_smiles_to_mol

Converts a list of SMILES strings to RDKit molecule objects.

def convert_smiles_to_mol(smis: list[str]) -> list[Chem.rdchem.Mol]:

func calculate_simmat

Calculates the similarity matrix for molecular fingerprints using a specified similarity function.

def calculate_simmat(fps: np.ndarray, similarity_function) -> np.ndarray:

func tanimoto_similarity

Calculates the Tanimoto similarity coefficient between two binary vectors.

def tanimoto_similarity(fp1: np.ndarray, fp2: np.ndarray) -> float:

func find_activity_cliffs

Identifies activity cliffs in the dataset.

def find_activity_cliffs(fps: np.ndarray, labels: np.ndarray, threshold: float) -> list[tuple[int, int]]:

func analogue_split

Splits the dataset into training and test sets, ensuring a specified fraction of the test set molecules are activity cliffs.

def analogue_split(fps: np.ndarray, labels: np.ndarray, test_size: float, gamma: float, omega: float) -> tuple[np.ndarray, np.ndarray]:

func train_and_evaluate_models

Trains and evaluates models using the analogue split and returns evaluation results.

def train_and_evaluate_models(gammas: list[float], fps: np.ndarray, labels: np.ndarray, models: dict, test_size: float, omega: float) -> dict:

func plot_evaluation_results

Plots evaluation results for different gamma values.

def plot_evaluation_results(results: dict, gammas: list[float], title: str) -> None:

How to use analoguesplit ?

  1. Identify Activity Cliff Molecules: Determine which molecules are part of activity cliffs based on their similarity and class labels.
  2. Generate Test Sets: For each gamma value, create test sets with the desired proportion of activity cliff molecules.
  3. Evaluate Model Performance: Train models on the training set and evaluate them on the test sets, calculating metrics such as accuracy, precision, recall, and F1 score.
  4. Create Gamma Plot: Visualize the model performance metrics against gamma values to understand the impact of activity cliffs on model robustness.

Example

Please check Notebook to learn how to use analoguesplit.

License

This project is licensed under the MIT License.

Acknowledgments

This package relies on several excellent Python libraries including RDKit, scikit-learn, NumPy, and Matplotlib.