cyMAEv2: learning robust cell representations in mass cytometry without supervision 4582

Abstract Description Batch effects in cytometry lead to inconsistent protein expression measurements and prevent researchers from integrating data from the most well-controlled studies. To address these challenges, we present cyMAEv2, an enhanced self-supervised learning model to generate robust cell representations in mass cytometry data. Compared to existing approaches, cyMAEv2 introduces 1) Enhanced Data Diversity and 2) Learning Cell Population Distributions. The model learns both protein expression distributions and their percentile distributions of eight distinct cohorts without cell type label. Validation was performed using two testing sets from 8 cohorts 1) a healthy donor set (HD), containing data with similar biological variations but different technical variations and; 2) an all-cohort set (ALL), which includes both healthy and diseased samples, reflecting differing biological variations. We used simplified single-cell integration benchmarking (scIB) metrics, to quantify conservation of biological variance (AvgBIO) and the removal of batch effects (AvgBATCH). For the HD set, cyMAEv2 achieved AvgBIO=0.833 and AvgBATCH=0.780, showing gains of + 0.066 AvgBIO and +0.175 AvgBATCH compared to using the original data. For the ALL set, cyMAEv2 achieved AvgBATCH=0.685, showing +0.071 gain from original data. These results demonstrate its ability to preserve biological integrity while mitigating batch effects. We believe this study marks a significant step forward for cytometry data integration. Funding Sources Supported by NIH AI082630, QuantumLeap Healthcare Collaborative fund, the University of Pennsylvania Perelman School of Medicine COVID Fund. Topic Categories Computational and Systems Immunology (COMP)