Key result
Machine learning high-content imaging identifies failing cardiac fibroblasts with an F1 score of ~0.88.
Effect estimate: Weighted F1 score 0.88
A machine learning model using high-content imaging can accurately distinguish between healthy and failing human cardiac fibroblasts, offering a potential tool for discovering novel anti-fibrotic therapies.
Introduction: Cardiovascular disease remains the leading cause of mortality worldwide, accounting for nearly 18 million deaths each year. Many disease etiologies, including heart failure (HF) and atrial fibrillation, involve the development of cardiac fibrosis, or pathological expansion of the extracellular matrix, mediated by the proliferation and activation of resident cardiac fibroblasts into a myofibroblast state. Hypothesis: Given the impact of cardiac fibrosis on cardiovascular disease worldwide and the lack of effective therapies, more innovative strategies to discover new treatments are needed. Goals: By applying a Cell Painting assay, a tool for extracting unbiased cell morphology readouts, to primary cardiac fibroblasts isolated from myocardial biopsies from human heart failure patients and nonfailing donor controls, we aim to gain insights into the biological mechanisms regulating cardiac fibroblast activation in failing hearts. Methods: We modified the assay by substituting an F-actin stain, a key component upregulated in activated myofibroblasts. We applied the assay to cardiac fibroblasts isolated from four heart failure patients with idiopathic dilated cardiomyopathy and two healthy organ donor hearts that could not be placed for transplantation. We acquired microscopy images, segmented 16,887 cells, and processed high-content morphology features using image-based profiling. We optimized and trained a binary logistic regression model to predict single cells from failing or healthy hearts. Results: Our model demonstrates high performance with weighted F1 scores of 0.91 in the training and 0.88 in the testing sets. We saw high accuracy across each individual heart in the test set, averaging approximately 0.88. We further applied the model to cardiac fibroblasts treated with DMSO or a TGF-β inhibitor, which are negative and positive controls. Applying our model to failing cells treated with the positive control showed higher healthy probabilities, indicating that our model can accurately distinguish between cell conditions. Conclusion: Our results demonstrate the potential of using this approach to identify new therapeutic strategies that make failing single-cell cardiac fibroblasts healthy. We plan to evaluate our model in a large drug screen to discover novel therapeutic strategies to reverse cardiac fibrosis.
No takes yet. Share an insight, caveat, or question.
Tomkinson et al. (2025) studied Heart failure (idiopathic dilated cardiomyopathy) (n=6). High-content imaging and machine learning (Cell Painting assay) vs. Healthy donor cells was evaluated on Prediction of single cells from failing or healthy hearts (Weighted F1 score 0.88). A machine learning model using high-content imaging accurately distinguished cardiac fibroblasts from failing versus healthy hearts, achieving a weighted F1 score of 0.88 in the testing set.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: