Comparison of Approaches for Determining Bioactivity Hits from High-Dimensional Profiling Data

Area of Science:

• * Computational Biology
• * Cell Biology
• * Drug Discovery

Background:

• * Phenotypic profiling assays measure thousands of cellular features to identify treatment effects.
• * Untargeted screening yields complex data, challenging the distinction between active and inactive compounds.
• * Robust hit identification is crucial for effective drug discovery and toxicological studies.

Purpose of the Study:

• * To compare diverse hit identification strategies for imaging-based phenotypic profiling.
• * To evaluate methods based on single- and multi-concentration analyses.
• * To optimize modeling parameters for a 10% false-positive rate while detecting subtle effects.

Main Methods:

• * Analysis of a Cell Painting dataset using various hit identification strategies.
• * Multiconcentration analysis: curve fitting at feature, category, and global levels; distance metrics (Euclidean, Mahalanobis); eigenfeatures.
• * Single-concentration analysis: signal strength (total effect magnitude); profile correlation among replicates.

Main Results:

• * Feature-level and category-based approaches identified the highest percentage of active chemicals.
• * Signal strength and profile correlation methods detected the fewest active hits at the fixed false-positive rate.
• * Distance metric fitting approaches showed the lowest likelihood of false positives from assay noise.

Conclusions:

• * Hit identification performance varies significantly across different computational strategies.
• * Most methods achieved high concordance, indicating robust hit calls across approaches.
• * Feature-level, category-based, and distance metric fitting methods offer distinct advantages for phenotypic profiling analysis.