Fig. 3. CellSighter learns spatial expression features.

A Left: Calprotectin (green) and cell boundaries (white) in one FOV. Right: Zoom-in on boxes 1–3. Shown are cells classified as Neutrophils by XGBoost (top, orange) or CellSighter (bottom, blue). Arrows show cells that were classified as neutrophils by XGBoost, but not CellSighter. B The proportion of cells classified as neutrophils (y-axis) as a function of the fraction of pixels in the cell which stain for Calprotectin (x-axis). CellSighter (blue) classifies as neutrophils less cells with low overlap with Calprotectin. C Example images in which CD4 (green) is simulated as a membranous signal and CD20 (red) as a patch, partially overlapping with the cell membrane. D Simulations of 549 cells as in (C) repeated 100 times. Shown is the fraction of cells classified as B cells (y-axis) by XGBoost and CellSighter. Asterix denote a two tailed t-test, P < 10⁻²⁰. Boxplots show median, first and the third quartile. Whiskers reach up to 1.5∙(Q₃–Q₁) from the end of the box. Dots denote outliers. E Simulations of 549 cells as in (C). Shown is the fraction of cells classified as B (y-axis) as a function of the models’ confidence (x-axis). CellSighter (blue) has lower confidence in erroneous classifications relative to XGBoost (orange). F Shown is the normalized sum of gradients (y-axis) as a function of the normalized distance from the cell center (x-axis) for 1353 T-helper and 42 Tregulatory cells. Gradients for FoxP3 (nuclear protein, solid line) and CD4 (membranous, dashed line) reach 60% of their maximum at ±40% and ±75% of the cell, respectively. Mean ± (SD/2) are shown by lines and gray area. G Example images in which CD4 (green) is simulated as a membranous signal and CD20 (red) as a membranous signal with differential overlap of the membrane, ranging from 100% to 12.5%. H Simulations of 200 cells as in (G). Shown is the fraction of cells classified as B (y-axis) as a function of the percent of membranous CD20 (x-axis). Despite similar signal levels across the simulations, CellSighter (blue) was less likely to classify cells as B cells when the overlap with the membrane was small. Source data are provided as a Source Data file.