AMPL

AMPL extends DeepChem and integrates machine learning and molecular featurization to build, evaluate, and share predictive models for pharmacokinetic and safety parameters in in silico drug discovery.

Key Features:

• Integration with DeepChem and ML algorithms: Extends the DeepChem library and incorporates a variety of machine learning algorithms for predictive modeling.
• Molecular featurization: Supports multiple molecular featurization techniques, including comparisons of molecular fingerprints against alternative representations.
• Modularity and extensibility: Provides a modular architecture that allows integration of additional algorithms and featurization methods.
• End-to-end pipeline: Implements workflows from data preprocessing through model building and model evaluation to enable traceability of modeling steps.
• Benchmarking on pharmaceutical datasets: Performs benchmarking across extensive pharmaceutical datasets to evaluate model and feature performance.
• Uncertainty quantification and data-driven analysis: Includes uncertainty quantification and analysis showing correlations between dataset size and prediction performance, with variability across datasets and models.

Scientific Applications:

• Prediction of pharmacokinetic properties: Building predictive models for pharmacokinetic parameters relevant to drug disposition.
• Safety assessment of compounds: Modeling safety-related endpoints to assess potential toxicological liabilities of candidate molecules.
• Early compound triage: Identifying promising compounds through predictive models to prioritize candidates in early drug discovery.

Methodology:

Extends DeepChem; integrates multiple machine learning algorithms and molecular featurization tools; performs data preprocessing, model building, model evaluation, benchmarking on pharmaceutical datasets, feature representation comparisons (including molecular fingerprints), and uncertainty quantification.