Single-cell analysis has advanced significantly at the genomic, transcriptomic, and proteomic levels, creating comprehensive cell atlases. However, tools for high-dimensional morphological analysis at single-cell resolution have lagged. Cell morphology, crucial for disease diagnosis and prognosis, reflects genomic and functional states. While fluorescence-activated cell sorting (FACS) provides some morphological data, it's limited. Existing image-based cell sorting methods often rely on biomarker staining, compromising cell viability, or use low-resolution images limiting morphological information capture. This study introduces COSMOS, a cloud-enabled platform for real-time cell imaging, analysis, and sorting, using deep learning to interpret high-resolution brightfield images for label-free cell classification and sorting. This addresses the challenge of real-time deep learning analysis and sorting of cells based on high-resolution brightfield images, a gap in current technology.

Recent advancements in image-based cell sorting utilize high-speed image capture, but limitations include reliance on biomarker staining (compromising cell viability for downstream assays) or deformability assays. Some approaches using feature engineering of low-resolution or reconstructed images may not capture the full complexity of morphological information. Deep learning and machine intelligence offer the potential to overcome these limitations by enabling simultaneous and generalized characterization of image traits. Prior studies have applied deep learning to cell classification on pathology slides and even to recapitulate immunohistochemistry signals. However, real-time deep learning classification and sorting decisions based on high-resolution images have been challenging due to computational constraints and data size limitations.

COSMOS is a cloud-enabled platform combining hardware and software for real-time cell imaging, analysis, and sorting. The hardware includes a microfluidic system for single-cell focusing and flow, high-speed imaging optics, and pneumatic valves for cell sorting. The software incorporates a deep learning model (Inception V3 architecture) trained on a large dataset of annotated single-cell images. This model generates high-dimensional morphological embeddings, used for cell classification. Dimensionality reduction techniques (UMAP) are employed for visualization and interpretation of these embeddings. The system is designed for real-time operation, distributing computational load across multiple processors and GPUs. A key aspect of the design is the microfluidic system that maintains cells in a tight focus range, crucial for high-resolution image acquisition. The workflow begins with single-cell image capture followed by real-time deep learning-based classification and sorting, driven by a user-selected model. Data is stored in a cloud-enabled database for future analysis. The methodology includes cell culture and preparation procedures for both cell lines and primary cells derived from patient samples. SNP assays are employed to assess sorting purity and enrichment. Cell viability and the impact on transcriptomic profiles are assessed through scRNA-Seq analysis, comparing COSMOS-processed cells to unprocessed controls. Two models were trained: a 'Circulating Cell Classifier' for blood cells and a 'Lung Tumor Classifier' for dissociated tumor cells.

COSMOS successfully distinguished various cell types (PBMCs, fnRBCs, NSCLC, HCC) based on morphology alone. The model accurately clustered these cells based on their morphological features, even identifying differences between cell lines of the same type. In silico analysis demonstrated high accuracy in identifying low-abundance cells (e.g., AUC > 0.97 for positive selection of NSCLC, HCC, and fnRBC from PBMCs). Experimental validation confirmed COSMOS's ability to enrich target cells from mixtures, achieving high purities even at 1:100,000 dilution. For example, with A549 cells spiked into whole blood at 40 cells/mL and 400 cells/mL, final purities of 35-80% and fold enrichment of 10,900-33,500x were achieved following RBC lysis, CD45 depletion, and COSMOS processing. COSMOS sorting had minimal impact on cell viability and transcriptomic profiles, as shown by scRNA-Seq analysis, resulting in minimal alteration of gene expression profiles compared to unprocessed samples. The application of COSMOS to dissociated solid tumor biopsies (NSCLC) successfully enriched malignant cells, enhancing the detection of mutations and copy number variations. In the case of DTC samples, the purity of EpCAM+/CD45- cells increased from 6.71% to 94.16% post-sorting.

COSMOS offers a significant advance in image-based cell sorting by enabling label-free, real-time analysis and sorting based on high-dimensional morphological profiles. Unlike existing methods, COSMOS does not require sample pre-processing (like fluorescent labelling), allowing for the analysis of biomarker-negative cells and preserving cell viability for downstream applications. This platform facilitates the discovery and analysis of cell populations with unknown phenotypic or molecular characteristics. The ability to sort viable, unaltered cells allows for ex vivo culturing, functional assays, and drug testing. The cloud-enabled database allows for continuous data analysis and potential identification of additional cell phenotypes. This is crucial for research into cellular heterogeneity and response to treatments.

COSMOS provides a powerful platform for label-free cell sorting based on high-dimensional morphology. Its ability to handle complex mixtures, isolate viable cells, and integrate seamlessly with downstream analyses offers significant advantages over existing techniques. Future improvements could focus on increased throughput, enhancing interpretability of AI predictions, and extending the platform’s capabilities to subcellular analysis. The integration of COSMOS with other single-cell technologies will likely generate considerable insights across various biological fields.

COSMOS relies on cells in suspension, requiring a dissociation step for solid tissues, potentially altering cell morphology. The current system has lower throughput compared to conventional sorters due to gentle sorting and computationally intensive real-time image analysis. While the model accurately classifies cells, interpreting the AI-derived morphological descriptors in terms of traditional biological features remains a challenge. Sorting live cells at very low concentrations (e.g., 1:100,000) at high throughput may be limited by the current system’s capabilities.