Predicting With Confidence: Using Conformal Prediction in Drug Discovery.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Quantifying prediction reliability in machine learning is a significant challenge.
• Quantitative structure-activity relationship (QSAR) models require robust confidence estimates for new chemical entities.

Purpose of the Study:

• Introduce conformal prediction as a method to provide valid confidence estimates for predictive models.
• Demonstrate the application of conformal prediction in QSAR modeling for drug discovery.

Main Methods:

• Utilized conformal prediction, a framework applied on top of existing machine learning algorithms.
• Generated prediction intervals (upper/lower bounds for regression, sets for classification) specific to each predicted object.
• Defined prediction interval size based on confidence level and object nonconformity.

Main Results:

• Conformal prediction provides mathematically proven guarantees on error rates.
• The framework intrinsically handles the applicability domain of machine learning models.
• Successfully applied conformal prediction to model ABC transporters.

Conclusions:

• Conformal prediction offers a rigorous approach for reliable in silico modeling.
• It enhances the trustworthiness of QSAR predictions in drug discovery.
• The method ensures consistent handling of model applicability and error rates.