Toward generalizable phenotype prediction from single-cell morphology representations

Deleted Journal, Год журнала: 2024, Номер 1(1)

Опубликована: Дек. 8, 2024

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

Язык: Английский

Cell Morphology accurately predicts the nuclear shape of adherent cells

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Дек. 28, 2024

Summary Cells are internally tensed, or prestressed, largely by actomyosin contractility. We hypothesized that nuclear shape is quantitatively predictable from cell since prestress couples them both. trained machine learning models on a publicly available image database of the WTC-11 line and predicted modes nucleus with high accuracy. develop U-Net architecture-based model, Cell2Nuc, voxels membrane accuracies between 74%-87%. To investigate prestress, we cultured imaged HeLa cells after inhibiting The Cell2Nuc model retrained slightly lower Statistical analysis revealed changes in size chromatin organization upon inhibition. Similar trends were seen images taken NIH3T3 cells. Thus, encodes features shape, their coupling partly due to contractility, whose abrogation leads mechanosensitive origin.

Язык: Английский