Toward generalizable phenotype prediction from single-cell morphology representations

Abstract

Background Functional cell processes (e.g., molecular signaling, response to stimuli, mitosis, etc.) impact cell phenotypes, which scientists can measure with cell morphology. However, linking these measurements with phenotypes remains challenging because it requires manually annotated labels. We propose that nuclear morphology can be a predictive marker for cell phenotypes that are generalizable across contexts. Methods We reanalyzed a pre-labeled, publicly-available nucleus microscopy dataset from the MitoCheck consortium. We extracted single-cell morphology features using CellProfiler and DeepProfiler, which provide robust processing pipelines. We trained multinomial, multi-class elastic-net logistic regression models to classify nuclei into one of 15 phenotypes such as ‘Anaphase,’ ‘Apoptosis’, and ‘Binuclear’. We rigorously assessed performance using F1 scores, precision-recall curves, and a leave-one-image-out (LOIO) cross-validation analysis. In LOIO, we retrained models using cells from every image except one and predicted phenotype in the held-out image, repeating this procedure for all images. We evaluated each morphology feature space, a concatenated feature space, and several feature space subsets (e.g., nuclei AreaShape features only). We applied models to the Joint Undertaking in Morphological Profiling (JUMP) data to assess performance using a different dataset. Results In a held-out test set, we observed an overall F1 score of 0.84. Individual phenotype scores ranged from 0.64 (moderate performance) to 0.99 (high performance). Phenotypes such as ‘Elongated’, ‘Metaphase’, and ‘Apoptosis’ showed high performance. While CellProfiler and DeepProfiler features were generally equally effective, concatenation yielded the best results for 9/15 phenotypes. LOIO showed a performance decline, indicating our model could not reliably predict phenotypes in new images. Poor performance was unrelated to illumination correction or model selection. Applied to the JUMP data, models trained using nuclear AreaShape features only increased alignment with the annotated MitoCheck data (based on UMAP space). This approach implicated many chemical and genetic perturbations known to be associated with specific phenotypes. Discussion Poor LOIO performance demonstrates challenges of single-cell phenotype prediction in new datasets. We propose several strategies that could pave the way for more generalizable methods in single-cell phenotype prediction, which is a step toward morphology representation ontologies that would aid in cross-dataset interpretability.