Toward generalizable phenotype prediction from single-cell morphology representations

Functional cell processes (e.g., molecular signaling, response to environmental stimuli, mitosis, etc.) impact cell phenotypes, which scientists can easily and robustly measure with cell morphology. However, linking these morphology measurements with phenotypes remains challenging because biologically interpretable phenotypes require manually annotated labels. Automatic phenotype annotation from cell morphology would link biological processes with their phenotypic outcomes and deepen understanding of cell function. We propose that nuclear morphology can be a predictive marker for cell phenotypes that is generalizable across cell types. Nucleus morphology is commonly and easily accessible with microscopy, but annotating specific phenotypic information requires labels. Therefore, we reanalyzed a pre-labeled, publicly-available nucleus microscopy dataset from the MitoCheck consortium to predict single-cell phenotypes. We extracted single-cell morphology features using CellProfiler and DeepProfiler, which provide fast, robust, and generalizable data 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’. In a held-out test set, we observed an overall F1 score of 0.84, where individual phenotype scores ranged from 0.64 (indicating moderate performance) to 0.99 (indicating high performance). Notably, phenotypes such as ‘Elongated’, ‘Metaphase’, and ‘Apoptosis’ showed high performance. While CellProfiler and DeepProfiler morphology features were generally equally effective, combining feature spaces yielded the best results for 9 of the 15 phenotypes. However, leave-one-image-out (LOIO) cross-validation analysis showed a significant performance decline, indicating our model could not reliably predict phenotype in new single images. Poor performance, which we show was unrelated to factors like illumination correction or model selection, limits generalizability to new datasets and highlights the challenges of morphology to phenotype annotation. Nevertheless, we modified and applied our approach to the JUMP Cell Painting pilot data. Our modified approach improved dataset alignment and highlighted many perturbations that are known to be associated with specific phenotypes. 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.