Impact of regularization methods and outlier removal on unsupervised sample classification

Description

Carol Heckman studied why image-based cell assays often fail to repeat, testing whether data cleaning steps change how samples get classified. In five trials, matched samples received either the same chemical mixture or a plain solvent, and one low-dimensional measure summed up much of the image data. Batch effects from materials, staff, and culture conditions could not be fully controlled.

Scaling the data within each trial let the treated and control groups separate clearly on that measure. Standardising against large reference databases left the classification pattern unchanged, whichever suitable database was used. Removing outliers hurt instead, wiping out more than three percent of many samples and creating both false positives and false negatives. Heckman argues that some non-repeatability is built into cell assays, and that correct class assignment is a fairer quality check than matching repeat means exactly.

Catalog record

Source: https://europepmc.org/article/MED/PMC13505170