Blinded comparison of natural killer cell detection with AI-driven nanophotonic imaging versus traditional spectral flow cytometry: Development of the CB-Scout assay.

Abstract

e14539 Background: Chimeric antigen receptor natural killer (CAR-NK) cells and in vivo CAR-T therapies are emerging as promising alternatives to traditional CAR-T treatments. However, concerns regarding post-treatment persistence require ultra-sensitive analytical methods to reliably detect rare circulating cells. Spectral flow cytometry (SFC) typically detects events at ~1 in 10⁴ cells but is limited by background noise and reduced phenotypic clarity at low concentrations. Next-generation sequencing (NGS) achieves sensitivities approaching 1 in 10⁶ cells but lacks single-cell phenotypic resolution. Our objective was to develop an assay combining the workflow and phenotypic resolution of SFC with the sensitivity of NGS. CB-Scout employs AI-driven nanophotonic imaging to detect rare cells at frequencies of ≤1 in 10⁶ with high phenotypic resolution. We present initial analytical validation of CB-Scout for NK cell detection compared to SFC in PBMCs spiked with NK-92 cells. Methods: Activated NK-92 cells were spiked into a single-donor HLA-mismatched PBMC sample at concentrations of 0, 70, 200, 1,200, 6,000, or 9,000 cells per ~10 M PBMCs. Split aliquots were analyzed using CB-Scout (CellsBin; Vega 1.2 system) or SFC (Cytek Aurora) by blinded operators to assess sensitivity, linearity, specificity, and functional phenotyping. Cells were stained with a multicolor panel including surface (CD45, CD3, CD56, CD16, HLA) and intracellular markers (IFN-γ, TNF-α, Granzyme B, Perforin, CD107a). Ground-truth phenotyping was established using ~100K activated NK-92 cells. Performance metrics included limit of blank (LoB), detection (LoD), and quantification (LoQ) per CLSI EP17-A2. Results: CB-Scout achieved a LoD of ~0.61 cells/M vs. ~30 cells/M for SFC and a LoB < 0.05 cells/M vs. ~1 cell/M for SFC. At the 35-cell spike level (~6.4 M PBMCs), CB-Scout maintained a ~17% CV, meeting LoQ criteria, whereas SFC required ~350 cells to achieve comparable precision. CB-Scout demonstrated higher sensitivity and specificity, with no false positives in > 21 M blank PBMCs, while SFC produced ~1 false-positive per 1 M blank PBMCs. Linear quantification was observed across three orders of magnitude, with consistent single-cell functional marker resolution, whereas SFC showed false positives and distorted functional readouts at low concentrations (~1 in 10 5 ). Conclusions: Using an SFC-like workflow, CB-Scout provides superior sensitivity and phenotypic resolution for therapeutic NK cells, enabling ultra-sensitive detection below 1 cell per million. By matching or exceeding NGS sensitivity while preserving single-cell functional profiling, CB-Scout may offer an optimized approach for monitoring emerging cell therapies, including CAR-NK and in vivo CAR-T.