Computational methods for enhancing clinical use of single-cell data - Project Summary Assessing cell distributions is a core technology across medicine. Common clinical technologies for this (flow cytometry, etc.) routinely generate rich single-cell data distributions (scData), but the value of this data remains largely untapped. Clinical scData are often aggregated into coarse summary markers, which hinders our ability to detect early signatures of disease, and to monitor patient dynamics. We believe that more careful use of clinical scData can greatly enhance patient care. However, to achieve this, there is a need for novel scData analysis methods, which are tailored to clinical needs and workflows. In particular, we identify three clinical needs that existing methods inadequately address: generating clinically actionable biomarkers from scData; tracking patient-specific cellular dynamics throughout the clinical course; and multi-modal modelling of scData alongside broader clinical inputs. We will address these gaps through three methodologic aims. In each, we will also investigate a well-defined, significant clinical challenge which scData may be able to help address. To support and power our analysis we will leverage a unique data source – a large-scale archive of raw flow cytometry files, underlying all blood count testing at a major US medical center. This archive will support development of robust methods and applications for enhanced use of the complete blood count – the most commonly collected blood test in the world. To ensure generality, we will also leverage multiple open-source clinical single-cell datasets, reflecting diverse patient samples, experimental analyses, and clinical challenges. In aim 1, we will develop a novel framework for scData biomarker generation. This will use unsupervised clustering to identify cell sub-populations, and perform biomarker generation via statistical modelling and use of self-supervised embedding models. Generated markers will be validated based on their efficacy across multiple outpatient disease screening tasks. In aim 2, we will use biophysics models to capture scData temporal patterns, and estimate patient-specific dynamics parameters. These markers will be used to monitor patient recovery during major physiologic disruptions, including pregnancy, thoracic surgery, and hematologic malignancies. In our final aim, we will develop multi-modal scData methods for clinical decision support. We will first establish baseline model frameworks for isolated use of scData for risk prediction, and then use early, late, and intermediate fusion to create multimodal integrations between scData and broader clinical features. We will apply these methods to the meaningful clinical challenge of early prediction of blood and urine culture results, to inform antibiotic prescription and use. Collectively, this work will provide novel, clinically-oriented methods for enhanced scData use, that will enhance our ability to generate clinically actionable biomarkers, estimate hidden cellular dynamics features, and directly support clinical decision making. Our proposed clinical studies will culminate in full readiness to undertake a prospective clinical trial of risk-informed care.
