Fig. 3: cpDistiller can effectively correct well-position effects while preserving biological variation for ORF profiles in 12 batches.
From: Triple-effect correction for Cell Painting data with contrastive and domain-adversarial learning
a, b UMAP visualizations of embeddings obtained by different methods in Batch_1 colored by row (a) and column (b). Suffixes “-row” and “-column” on scDML and scVI indicate the corrected effects. c Dendrograms illustrate the clustering of ORF perturbations induced by 12 reagents in treatment, graphically rendered based on low-dimensional representations generated by different methods. Baseline represents the preprocessed but uncorrected CellProfiler-based features. d Quantitative evaluation of the performance of different methods on ORF profiles across 12 batches, where each batch is analyzed independently to assess the correction of well-position effects while preserving biological variation. The ASW and tASW metrics yield three results based on row labels, column labels, and plate labels, respectively. In the boxplots, center lines indicate the medians, box limits show upper and lower quartiles, whiskers represent the 1.5 × interquartile range, and notches reflect 95% confidence intervals via Gaussian-based asymptotic approximation. Each data point represents one biologically independent batch (n = 12), with batches differing in their experimental context and typically containing distinct gene perturbations. e Heatmap shows p-values from one-sided paired Wilcoxon signed-rank tests. Each cell in the heatmap reflects the statistical significance of one method’s (row) superiority over another (column), derived from 132 evaluations across 12 batches using 11 metrics (n = 132).
