Companion diagnostics: tissue assays for drug eligibility

Most of medicine accepts imperfect predictions; targeted oncology cannot afford them. Drugs aimed at molecular machinery — a growth-factor receptor, a DNA-repair pathway — deliver benefit almost exclusively inside the population whose tumours carry that machinery, while exposing everyone else to toxicity without hope. This asymmetry created a novel regulatory object: the companion diagnostic, an assay whose own validation data ship welded to the drug’s label. Prescribe without testing, outside trials, and the prescription itself falls out of the approved universe. Tissue-based platforms built this category; their science repays attention precisely because they look deceptively simple.

A stain is an amplifier wearing an antibody

Immunohistochemistry runs amplification-by-catalysis on a glass slide. An engineered antibody binds the target protein wherever it sits; a secondary reagent then deposits coloured polymer at those coordinates, minute by minute, so each captured molecule recruits thousands of dye particles. Read under a microscope, the tumour appears self-reporting: membrane rings, cytoplasmic washes, nothing. Since 1998, when breast-cancer therapy first demanded receptor confirmation before use, that visual grammar has decided eligibility for billions of treatment courses. The measurement hides in training the reader: pathologists score semiquantitatively — none, weak, intermediate, strong — because the underlying chemistry yields a continuum that clinical rules must bin.

What chromosomes say that proteins conceal

Protein abundance can mislead in both directions, which is why fluorescence hybridisation survives as arbiter. Fluorescently labelled DNA probes are flooded across nuclei and bind their matching loci; counting glowing spots per cell then measures gene dosage directly. Receptor overexpression sometimes traces to dozens of copied genes driving output, and probe ratios expose that architecture cleanly — amplified clusters versus paired control-chromosome counts settle ambiguous stains with different physics entirely: statistics on hundreds of nuclei rather than colourimetric intensity on one section. Where protein staining reports outcome state, copy-number counting reports cause.

Cut-offs belong to trials, not to nature

Beneath every routine number lies an uncomfortable truth: the boundary separating “eligible” from “not” is negotiated in the drug’s pivotal study, not derived from tissue physiology. Expression runs continuously through the population; some line selects a band where benefit-to-risk worked, frozen thereafter into hospital workflow. Consequences follow relentlessly. Different antibody clones grade the very same slide differently; different indications adopt different thresholds even within one cancer type, each empirically anchored to its own trial; scores near boundaries shift with fixation protocols, staining runs and human eyes, which quality systems police but never abolish. And deeper limits hum underneath: biopsies sample one region of heterogeneous disease, positive stains predict modest response probabilities rather than certainties, and immune-facing biomarkers particularly depend on features no single marker captures. Companion diagnostics replaced guesswork with calibrated uncertainty — an honest bargain, and still only a bargain.