Personalized diagnostics

A diagnostic result is an inference, and inference has two inputs: the measurement and the prior. Population diagnostics fix the prior at the population average; personalized diagnostics change the prior. That is the entire conceptual move, and it has three biological mechanisms behind it — metabolic, immunological and oncological — each of which makes the same analyte mean something different in a different person.

The metabolic mechanism is pharmacogenomics. Drug-destroying enzymes come in versions: the CYP2C19 loss-of-function alleles that make clopidogrel a prodrug some patients barely convert, the CYP2D6 copy-number variation that spans ultrarapid to nonfunctional metabolizers. A genotype does not measure disease; it changes the dose-to-effect curve on which every later measurement of “therapeutic level” is graded. This is also why genotype-guided dosing is the oldest genuinely personalized test — it needs no new instrument, only the recognition that one number on the lab report is conditioning information for another.

The immunological mechanism is idiosyncrasy with a molecular address. HLA-B*57:01 carriers hypersensitize to abacavir; the drug fits that particular MHC presentation pocket and the immune system reacts to the complex. Before genotyping, such reactions were “allergies” — unpredictable by definition. After it, they became predictable in exactly the population that carries the allele, which is the cleanest demonstration that personalization is not about testing more, but about conditioning on the variant that carries the causal weight.

The oncological mechanism is heterogeneity inside the lesion. A tumour’s behaviour is driven by a minority of altered genes — drivers — embedded in a mass of passenger damage; sequencing changes the diagnostic question from “what is this tissue” to “which circuit is broken”, and the answer reclassifies cancers that look identical under the microscope. Liquid biopsy extends the same logic in time, because the tumour’s genotype evolves under therapy — the mechanism liquid biopsy pages cover in detail, and the reason a single baseline genotype has a shelf life.

The statistical ceiling is what keeps honesty in the field. Polygenic risk scores aggregate thousands of variants, yet each contributes little variance, and scores trained on one ancestry lose calibration on another — a measurement-equity problem with a statistical root, not merely a dataset shortage. Environment does the rest: the genotype sets distributions, not fates, and the same score means different risk at different ages and exposures. Personalized diagnostics is therefore best read not as a promise of certainty but as systematically better priors — the same arithmetic that powers companion diagnostics platforms, applied to the patient’s own genome rather than to a drug label, and increasingly combined with the omics layer described under multi-omics diagnostic panels.