Confidence in Inactive and Active Predictions from Structural Alerts.

Area of Science:

• Computational chemistry
• Toxicology
• Drug discovery

Background:

• Assessing biological activity of chemicals computationally is vital for risk assessment.
• Assigning confidence to inactive predictions from in silico tools is challenging.
• False-negative predictions are a major concern in hazard identification.

Purpose of the Study:

• To develop methods for assigning confidence to both active and inactive predictions from structural alerts for protein-binding molecular initiating events (MIEs).
• To enable responsible use of computational predictions in chemical safety evaluations.

Main Methods:

• Structural alerts were derived using an iterative statistical method.
• Confidence was assigned by measuring Tanimoto similarity between Morgan fingerprints of test and training set chemicals.
• A new, independent validation dataset was created with 648 experimental measurements for 27 compounds against 24 proteins, supplemented with ChEMBL25 data.

Main Results:

• Defined suitable cutoff values for Tanimoto similarity to establish different confidence categories.
• Successfully applied confidence categories to computational predictions on the new validation dataset.
• Enabled identification of chemicals with confident active or inactive predictions.

Conclusions:

• The developed methods allow for confident assignment of activity predictions, both active and inactive.
• This enhances the reliability of in silico tools for chemical risk assessment.
• Facilitates responsible application of computational predictions in regulatory science.