Phenotypes

The Phenotypes engine combines multiple gating results into cell type assignments:

1. Define phenotypes — Each phenotype has a name and a Boolean expression using gate results. Example: "Cytotoxic T" = CD3_positive AND CD8_positive AND CK_negative.
2. Evaluate — For each cell, all gate results are known (positive/negative per marker). The engine evaluates each phenotype definition against the cell's gate profile.
3. Assign — The first matching phenotype (in priority order) is assigned. If no defined phenotype matches, the cell is labeled "Unclassified" or assigned to a default category.
4. Summarize — Population statistics are computed: total count and percentage for each phenotype in each ROI. These summaries are the primary quantitative output for many clinical applications.

Classification as partition: MIT's Statistical Models chapter frames classification as partitioning feature space into decision regions. Each phenotype definition creates a region in the multi-dimensional marker space — the set of all cells with the specified marker combination. The phenotype definitions collectively partition the marker space: every cell belongs to exactly one phenotype. Well-designed phenotype panels produce partitions that align with biologically distinct cell populations.

The combinatorial explosion: With n binary markers, there are 2ⁿ possible marker combinations. A 6-marker panel has 64 possible phenotypes; a 10-marker panel has 1,024. In practice, most combinations are biologically meaningless (a cell cannot be CD3+ and CD20+ simultaneously in normal biology) or extremely rare. Effective phenotype design focuses on the 5-15 biologically meaningful combinations rather than trying to characterize all possible ones.

Statistical methods for threshold optimization (Dilbilir): The quality of phenotype assignments depends entirely on the quality of individual marker gates. Each gate threshold determines the boundary between positive and negative for one marker. Dilbilir's statistical framework emphasizes that threshold optimization should minimize classification error across the population — not just separate the most obvious positive from the most obvious negative cells. Cells near the threshold (the "dim positive" or "equivocal" population) are the most error-prone and most important to get right.

Phenotyping a 6-plex immuno-oncology panel (CD3, CD8, CD20, FOXP3, PD-L1, CK):

• Tumor PD-L1+: CK+ AND PD-L1+ (16% of all cells)
• Tumor PD-L1−: CK+ AND PD-L1− (34%)
• Cytotoxic T: CD3+ AND CD8+ AND CK− (8%)
• Helper T: CD3+ AND CD8− AND CK− (6%)
• Regulatory T: CD3+ AND FOXP3+ AND CK− (2%)
• B cell: CD20+ AND CD3− AND CK− (4%)
• Other: remaining cells (30%)

These 7 phenotypes capture the major cell populations relevant to immunotherapy response prediction. The 50/50 split of PD-L1+/− tumor cells and the 4:1 ratio of cytotoxic to regulatory T cells are directly clinically informative.