Using Jupyter Notebooks for re-training machine learning models.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning applications in drug discovery
• Pharmacokinetics and drug transport modeling

Background:

• Machine learning (ML) models necessitate extensive molecular descriptor selection for accurate prediction of active and inactive compounds.
• Data privacy concerns and limited access restrict the chemical space, impacting ML model generalization and reliability.
• Existing ML workflows often require significant user expertise for descriptor selection and model implementation.

Purpose of the Study:

• To develop a framework for re-trainable and transferable ML models to overcome data limitations and privacy issues.
• To reduce the extensive user-driven selection of molecular descriptors required for ML model training.
• To enable decentralized, facile, and rapid updating of ML models across different local instances.

Main Methods:

• Proposed a framework for re-trainable ML models adaptable to various local instances.
• Shared ML models via a Jupyter Notebook to facilitate broader chemical space evaluation.
• Kept most tunable parameters pre-defined to simplify model implementation and reduce descriptor selection burden.

Main Results:

• The proposed framework demonstrated general applicability across six diverse transporter datasets (BCRP, BSEP, OATP1B1, OATP1B3, MRP3, P-gp).
• The method allows for less extensive descriptor selection compared to traditional ML approaches.
• Decentralized model updating proved to be facile and fast, enhancing model adaptability.

Conclusions:

• The developed framework offers a robust solution for building reliable ML models with reduced data requirements and enhanced privacy.
• This approach significantly simplifies the implementation and updating of ML models in cheminformatics.
• The generalizability across multiple transporter datasets highlights the potential of this framework for broader applications in drug discovery.