Filtered circular fingerprints improve either prediction or runtime performance while retaining interpretability.

Area of Science:

• * Cheminformatics
• * Computational Chemistry
• * Drug Discovery

Background:

• * Circular fingerprints, developed over 50 years ago, remain crucial for building predictive quantitative structure-activity relationship (QSAR) models.
• * Historically used for database searching, circular fingerprints are often folded into short bit-strings for compactness.
• * Folding introduces bit collisions, compromising interpretability and adding noise to structural information.

Purpose of the Study:

• * To evaluate the efficacy of using filtered circular fingerprints for QSAR modeling.
• * To compare the performance of filtered fingerprints against folded and unprocessed fingerprints.
• * To introduce a web service for applying filtered fingerprints and providing prediction rationales.

Main Methods:

• * Application of supervised feature selection to filter circular fingerprint fragments.
• * Development and utilization of the Collision-free Filtered Circular Fingerprints (COFFER) web service.
• * Validation of QSAR models using filtered, folded, and unprocessed fingerprints.

Main Results:

• * Filtered circular fingerprints significantly enhance predictive performance compared to folded fingerprints, maintaining interpretability.
• * Filtered fingerprints offer reduced computational effort, improved compactness, and less redundancy than unprocessed fingerprints.
• * The area under the precision-recall curve is proposed as a more suitable metric for QSAR model validation in virtual screening.

Conclusions:

• * Unfolded circular fingerprints should be used for QSAR modeling to preserve interpretability.
• * Filtering circular fingerprints is an effective strategy to manage computational costs associated with using all fragments.
• * Filtered fingerprints provide a balance of predictive power, interpretability, and computational efficiency for QSAR applications.