AMPL: A Data-Driven Modeling Pipeline for Drug Discovery.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• cheminformatics

Background:

• Reproducibility and traceability are critical for machine learning (ML) in drug discovery.
• Existing tools may lack the modularity and extensibility needed for comprehensive model building and evaluation.

Purpose of the Study:

• To develop an end-to-end, modular, and extensible software pipeline for building and sharing ML models.
• To predict key pharmaceutical-relevant parameters using ML.

Main Methods:

• Developed the ATOM Modeling PipeLine (AMPL), an open-source software extending DeepChem.
• Utilized an array of ML and molecular featurization tools.
• Benchmarked AMPL on diverse pharmaceutical datasets.

Main Results:

• Traditional molecular fingerprints showed suboptimal performance compared to other feature representations.
• Larger dataset sizes directly correlated with improved prediction performance.
• Uncertainty quantification's correlation with model error varied across datasets and model types.

Conclusions:

• An extensible and shareable pipeline is essential for accessible and reproducible ML model building in drug discovery.
• Expanding public pharmaceutical datasets is crucial for enhancing prediction performance.
• AMPL facilitates reproducible ML model development and deployment in the pharmaceutical industry.