Cross-Species Morphology Learning Enables Nucleic Acid-Independent Detection of Live Mutant Blood Cells

Abstract

In both neonates and adults, the presence of malignancy-associated mutations in peripheral blood (PB) correlates with an elevated risk of future neoplastic transformation, with certain mutations, such as KMT2A rearrangements, exhibiting near-complete penetrance. If feasible, pre-malignant screening could enable early intervention and even disease prevention. However, nucleic acid sequencing- and hybridization-based mutation detection have limited cost-efficiency, constraining their use in screening. Here, we introduce a computer vision platform that can identify mutant cells in fresh PB samples that carry KMT2A-MLLT3 (a frequent mutation in pediatric and adult leukemias and detectable in newborn blood samples) or JAK2V617F (a frequent mutation in myeloproliferative neoplasms and clonal hematopoiesis). This is achieved by high-throughput single-cell imaging and mutation detection by machine learning (ML)-powered morphology recognition. The ML models were developed by cross-species learning of conserved features between mutant cells from mouse genetic models and from human samples, enabling a cost-effective approach for detecting mutations in live blood cells. This platform holds promise for pre-malignant screening in asymptomatic neonates and adults with KMT2A-MLLT3 or JAK2V617Fmutation and is potentially generalizable to the detection other malignancy-associate mutations. Our platform provides a novel single-cell morphological data modality that complements existing single-cell genomics.