Automated robust normalization and context-aware rescue of TReg cells from multiplex immunofluorescence
Presenter: Joao Paulo Oliveira da Costa, DDS;PhD Session: Digital Pathology 3 Time: 4/21/2026 9:00:00 AM → 4/21/2026 12:00:00 PM
Authors
Joao Paulo Oliveira da Costa , Alina Ainbinder , Kenneth Trieu , Henry Reinhart , Yury Sheikin Pathology, Takeda Pharmaceutical Company Ltd. (U.S.), Cambridge, MA
Abstract
CCR8⁺FoxP3⁺ regulatory T cells (Tregs) are a key immunosuppressive subset in solid tumors, but their accurate quantification in multiplex immunofluorescence (mIF) is hindered by slide-to-slide staining variability, dim FoxP3 signal, and background-driven CCR8 false positives. We developed a fully automated, background-aware R pipeline for robust cross-slide calling of CCR8⁺FoxP3⁺ cells from HALO single-cell data across NSCLC, CRC, GEJ, and HNSCC.Single-cell intensities for CCR8 and FoxP3 were transformed using inverse hyperbolic sine and normalized per slide using robust z-scores (rz) computed from medians and MADs, with fallbacks when FoxP3⁻ or CCR8⁻ reference populations were sparse. FoxP3 positivity was reassessed using a two-stage approach: (i) “soft” FoxP3 calls integrating HALO labels with high-rz outliers, and (ii) a strict rescue rule in which each cell’s FoxP3 rz was evaluated against the empirical FoxP3⁻ distribution, requiring both a minimum rz floor and low FoxP3 background-z, preventing inflation of false positives. CCR8 calling used a spatially explicit background model: slides were partitioned into fine spatial bins, and CCR8 background medians and dispersions were estimated from local CCR8⁻ cells, with smoothing across sparse regions. This produced a background z-score (bg_z) and a background-adjusted CCR8 rz (rz_bg). From these, we defined a lenient background-aware CCR8 call and a strict CCR8 keeper enforcing strong evidence and artifact filters (cytoplasm completeness, CK negativity, nucleus quality). Final CCR8⁺FoxP3⁺ assignments used a FoxP3-aware rule: FoxP3⁺ cells were evaluated with lenient CCR8 criteria to recover dim true positives, whereas FoxP3⁻ cells required the strict CCR8 keeper to prevent spurious double-positive inflation. Agreement analyses showed that the strict CCR8 criteria substantially reduced Halo-only artifacts while adding very few “rescued-only” events. Conversely, FoxP3 soft-rescue recovered dim nuclei primarily in low-background regions. Processing was optimized using Arrow to avoid loading full datasets into memory, enabling analysis of >10 million cells on a standard workstation. The combined FoxP3 normalization, empirical rescue, and CCR8 background-aware gating produced stable distributions across indications and improved both sensitivity and specificity for CCR8⁺FoxP3⁺ Treg quantification.
Disclosure
J. Oliveira da Costa, Takeda Pharmaceuticals Employment. A. Ainbinder, Takeda Pharmaceuticals Employment. K. Trieu, Takeda Pharmaceuticals Employment. H. Reinhart, Takeda Pharmaceuticals Employment. Y. Sheikin, Takeda Pharmaceuticals Employment.
Cited in
Control: 6372 · Presentation Id: 3131 · Meeting 21436