AI-powered single-cell imaging and morphology-guided label-free enrichment reveal tumor heterogeneity, genomic relatedness, and divergence in malignant effusions

ABSTRACT Background: Malignant effusions are clinically accessible yet underutilized sources of metastatic tumor cells. However, abundant background immune cells complicate tumor cell isolation and molecular characterization. We integrate AI-powered single-cell imaging and morphology-guided label-free tumor cell enrichment with genomic profiling to characterize phenotypic and genomic heterogeneity in malignant effusions. Methods: We applied a pre-trained deep learning model to brightfield images of single cells from malignant effusions of 10 patients with metastatic breast cancer to extract morphological features, quantify tumor cell fraction (TCF), and guide label-free microfluidic enrichment. We profiled copy-number alterations and somatic mutations in enriched tumor cell fractions and compared them with available matched effusion cell-free DNA and solid tumors. Results: AI-powered single-cell morphological analysis identified putative tumor cells in all 10 patients (median TCF, 5.4%; range, 0.51-15.8%; n = 25 samples). AI estimates of major immune-cell populations varied across patients and samples and correlated with flow cytometry measurements (n = 21 paired samples, Pearson r = 0.69-0.75). Unsupervised analysis identified four tumor cell morphotypes, revealing intra- and inter-patient morphological heterogeneity and longitudinal variation in morphotype composition. Genomic profiling of enriched tumor cell fractions by low-pass whole-genome sequencing (n = 26 samples) and whole-exome sequencing (n = 24 samples) identified copy-number alterations and somatic mutations, with tumor fraction influencing genomic signal strength. Comparisons of genomic profiles from tumor cell-enriched fractions with matched effusion cell-free DNA and solid tumors revealed shared genomic alterations, providing orthogonal support for tumor origin and relatedness, while differences across analytes revealed genomic divergence and tumor heterogeneity. Conclusion: AI-powered analysis leverages single-cell morphology for tumor cell identification and label-free enrichment, while integration with genomic profiling enables simultaneous interrogation of phenotypic and genomic heterogeneity in malignant effusions. This approach expands the potential of malignant effusions for liquid biopsy applications and biomarker development and discovery.