ULCYP: A Multitask Model for Predicting P450 Inducers Based on Positive-Unlabeled Learning.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Predicting cytochrome P450 (CYP) induction is crucial for mitigating drug-drug interactions and toxicity in early drug discovery.
• Limited availability of comprehensive inducer data hinders the development of accurate predictive models for CYP induction.

Purpose of the Study:

• To develop ULCYP, a novel multitask deep learning framework utilizing positive-unlabeled (PU) learning for enhanced CYP induction prediction.
• To leverage large unlabeled datasets to overcome the scarcity of reliable negative samples in CYP induction prediction.

Main Methods:

• Implementation of a multitask deep learning framework (ULCYP) based on positive-unlabeled (PU) learning.
• Utilizing integrated gradients for model interpretability and identifying key molecular substructures.
• Defining a rigorous applicability domain to ensure prediction reliability.

Main Results:

• ULCYP demonstrated superior performance compared to baseline models across multiple metrics.
• Achieved an average area under the curve (AUC) greater than 0.81 on a test set of CYP inducers and nonagonists.
• Successfully predicted induction mediated by key receptors: pregnane X receptor (PXR), constitutive androstane receptor (CAR), and aryl hydrocarbon receptor (AhR).

Conclusions:

• ULCYP effectively addresses data scarcity challenges in CYP induction prediction through PU learning.
• The model provides reliable and interpretable predictions, aiding in the early assessment of drug candidates.
• ULCYP offers a valuable tool for drug discovery, enhancing the prediction of potential drug-drug interactions and toxicity.