The single most useful organizing distinction in pharmacogenetics is pharmacokinetics (PK) versus pharmacodynamics (PD). PK genes govern what the body does to the drug — how fast it is metabolized and cleared, and therefore the concentration that reaches the target. PD genes govern what the drug does to the body — the sensitivity of the receptor, channel, or immune system the drug acts on. The two failure modes look different at the bedside: a PK variant shifts a familiar dose-response curve left or right (too much or too little drug), whereas a PD variant changes the shape of the response itself (paradoxical worsening, an immune catastrophe). Most actionable neurology pharmacogenetics is PK (CYP450, UGT), but the highest-stakes interactions — sodium-channel blockers in Dravet, HLA-mediated skin reactions, POLG and valproate — are PD or immunogenetic, and crucially are not rescued by lowering the dose.

Why metabolizer phenotypes matter. For PK genes the body inherits two alleles, and their combined enzyme activity sorts patients into a phenotype:

• Poor metabolizer (PM) — two loss-of-function alleles; drug accumulates, risking toxicity at standard doses
• Intermediate metabolizer (IM) — one functional and one reduced/absent allele
• Normal/extensive metabolizer (NM/EM) — the reference state the standard dose was designed for
• Ultra-rapid metabolizer (UM) — extra functional gene copies; drug is cleared so fast that levels are subtherapeutic, risking treatment failure

The key inference is directional: knowing only the metabolizer class tells you which way to expect the concentration to move, but not by exactly how much — which is why pharmacogenetics complements, rather than replaces, therapeutic drug monitoring. A second subtlety is the prodrug reversal: for drugs that must be enzymatically activated (clopidogrel, codeine), a poor metabolizer has too little active drug and a UM has too much, exactly inverting the usual PM-equals-toxicity intuition.

Four CYP450 enzymes carry most of the neurological drug load: CYP2C9 (phenytoin, valproate, losartan), CYP2C19 (clopidogrel, clobazam, diazepam, omeprazole), CYP2D6 (tricyclics, opioids, atomoxetine, antipsychotics), and CYP3A4/5 (carbamazepine, oxcarbazepine, statins). They are not interchangeable from a pharmacogenetic standpoint, and understanding why predicts how strongly genotype matters for each.

Two features decide whether genotype is clinically actionable. First, the therapeutic index of the substrate: when a drug with a narrow window (phenytoin, tricyclics) is handled by a polymorphic enzyme, a modest genotype-driven shift in clearance crosses the line from therapeutic to toxic. Second, whether the enzyme is the rate-limiting step or just one of several parallel routes. CYP2D6 and CYP2C19 are often the dominant clearance pathway for their substrates, so a loss-of-function genotype has a large, predictable effect. CYP3A4/5 is the workhorse of drug metabolism but is also broadly redundant and heavily modulated by inducers and inhibitors; its genetic variants are therefore usually swamped by drug–drug interactions, which is why CYP3A5 genotype rarely changes prescribing.

CYP2D6 is the most polymorphic and the most architecturally complex. Its activity ranges across a continuum captured by an activity score, and — uniquely among these enzymes — it commonly varies by gene copy number: whole-gene deletions produce poor metabolizers, while duplications and multiplications (*1xN) produce ultra-rapid metabolizers. SNP-based panels read the alleles but can miss these structural variants and gene–pseudogene hybrids, a recurring blind spot discussed in the testing section.

Ancestry shapes the prior probability. Loss-of-function CYP2C19 alleles are far more common in East Asian populations (poor-metabolizer frequency ~15–20% vs ~2–5% in Europeans), while CYP2D6 poor metabolizers are ~7–10% of Europeans. Allele frequencies differ enough across ancestries that a panel optimized for one population is not automatically valid for another.

HLA associations are a fundamentally different kind of pharmacogenetics from the CYP story, and the difference is mechanistic. CYP variants are pharmacokinetic — they scale concentration, so the danger rises and falls with dose. HLA-mediated reactions are immunogenetic and behave like an idiosyncratic, immune-mediated event: the severe cutaneous reactions — Stevens-Johnson syndrome (SJS), toxic epidermal necrolysis (TEN), and DRESS — are essentially all-or-none and dose-independent.

Why dose-independent? Class I HLA molecules present peptides to cytotoxic CD8 T cells. A small-molecule drug (or its metabolite) binds within the peptide-binding groove of a specific HLA allotype, altering the repertoire of self-peptides displayed or directly engaging the T-cell receptor. In a carrier of the at-risk allele, this turns the drug into a potent neoantigen and triggers a clonal cytotoxic T-cell attack on keratinocytes. Because the trigger is molecular recognition rather than cumulative exposure, even a small, standard, or test dose can set off the cascade — there is no "safe low dose," and the only safe move is to avoid the drug entirely in a carrier. This is why HLA results are among the few genuine pharmacogenetic contraindications, in contrast to the dose-adjustment logic of CYP genotypes.

The archetype is HLA-B\*15:02 and carbamazepine, identified in a landmark Taiwanese study in which essentially all carbamazepine-induced SJS/TEN patients carried the allele while drug-tolerant controls did not (Chung et al. 2004). Pre-prescription HLA typing in the right population is therefore a true prevention strategy — but it is allele- and ancestry-specific, so the test only helps where the at-risk allele is plausibly present.

Antiepileptic drugs (AEDs) are where every layer of pharmacogenetics converges, and sorting the interactions by mechanism is what makes them clinically usable.

Pharmacodynamic — the drug target itself is altered. In Dravet syndrome, SCN1A loss-of-function reduces Nav1.1, a channel expressed predominantly on inhibitory interneurons. Sodium-channel-blocking AEDs (carbamazepine, oxcarbazepine, lamotrigine, phenytoin) further suppress those already-failing interneurons, lifting the brake on the network and paradoxically worsening seizures, sometimes precipitating status. This is a PD contraindication: like the HLA reactions, it is not fixed by lowering the dose, and the offending drug must be avoided. Oxcarbazepine and lamotrigine deserve special caution because they are so commonly reached for before a genetic diagnosis is in hand.

Mitochondrial/immunogenetic — POLG and valproate. POLG encodes the catalytic subunit of mitochondrial DNA polymerase γ; biallelic mutations (Alpers-Huttenlocher and the broader POLG spectrum) leave hepatocytes with a precarious capacity to replicate mtDNA and regenerate. Valproate is independently mitotoxic — it depletes carnitine, interferes with β-oxidation and the respiratory chain, and impairs liver regeneration — so layering it onto a POLG-deficient liver can tip a marginally compensated organ into fulminant, frequently fatal hepatic failure that responds poorly even to transplantation. Prospective POLG testing identifies these high-risk individuals before exposure (Stewart et al. 2010). Practically: screen for POLG (or its red-flag features — developmental regression, refractory seizures, hepatic involvement, mitochondrial signs) before starting valproate in a young child.

Pharmacokinetic — and a reminder that environment can mimic genotype. Lamotrigine clearance is dominated by UGT1A4 glucuronidation, and that enzyme is strongly induced by estrogen. Pregnancy and estrogen-containing contraceptives can roughly halve lamotrigine levels, causing breakthrough seizures, with an abrupt reversal after delivery that swings the patient toward toxicity if the pregnancy dose is not tapered. The point: an acquired, hormonally driven shift in enzyme activity can reproduce the clinical picture of a metabolizer phenotype, which is why level monitoring — not genotype alone — anchors lamotrigine management. Separately, CYP2C9 poor metabolizers clear phenytoin slowly and reach toxic levels (nystagmus, ataxia, lethargy) at ordinary doses, a narrow-therapeutic-index drug where genotype-guided dose reduction is well supported. Valproate also inhibits the urea cycle and can unmask hyperammonemic encephalopathy in partial urea-cycle-defect carriers — worth considering before valproate in anyone with unexplained hyperammonemia or protein aversion.

The right mental model is that pharmacogenetic evidence is a steep gradient, not a flat field — a handful of drug-gene pairs are genuinely actionable, while the broad panels marketed around them coast on the credibility those few pairs earned. Interpreting a report well means reading where on that gradient each result actually sits.

Where the evidence is strong — a small number of specific drug-gene pairs carry CPIC level-A guidance and/or FDA labeling: HLA-B\15:02/carbamazepine, CYP2C9/phenytoin, CYP2C19/clopidogrel, and POLG/valproate. These share a structure that explains their strength: a single* gene with large effect size, a clear mechanism (immunogenetic contraindication or a narrow-therapeutic-index PK shift), and a concrete action (avoid, or adjust dose). They are high-yield precisely because the genotype maps onto one decision.

Why combinatorial panels are oversold — broad combinatorial panels marketed to guide psychotropic selection (e.g., GeneSight) blend many small-effect variants into a proprietary, color-coded composite. That design dilutes rather than concentrates signal: depression response is overwhelmingly polygenic and psychosocial, so the metabolic genes the panel reads explain only a sliver of who responds to which antidepressant. The published trials are mixed, often unblinded, and frequently funded by test makers, and major guidelines do not endorse routine combinatorial panel testing to choose antidepressants. The proprietary algorithm also obscures which gene drove a recommendation — so a clinician cannot audit it the way they can a single CYP2C19 result. Treat the marketing skeptically.

Avoid over-interpretation — categorical, color-coded bins are frequently misread as verdicts. The collapse from a continuous activity score into three colors discards exactly the gradation a prescriber needs; a 'red' or 'use with increased caution' bin almost always means consider a dose adjustment or closer monitoring, not an absolute contraindication. True 'do not use' results are mechanistically specific and few — the immunogenetic and mitochondrial pairs (HLA-B\*15:02/carbamazepine, POLG/valproate), not the metabolizer phenotypes. Pharmacogenomics informs a minority of prescribing decisions and never replaces clinical judgment, phenotype, or therapeutic drug monitoring.

Preemptive vs. reactive, and the analytic blind spots — reactive testing (ordered at the moment of prescribing) can delay treatment by days; preemptive panels store results in the EHR so they are waiting when needed. But the format matters more than the gene list: EHR-integrated clinical decision support that fires an alert at the point of order is what actually changes prescribing — a passive PDF in the chart rarely does. Finally, the assays themselves have limits. SNP-based panels are built around common European variants and read poorly across structural variation, so CYP2D6 copy-number changes and gene–pseudogene hybrids can be missed, and sensitivity falls in non-European ancestries — a normal-metabolizer result can therefore be a false reassurance rather than a true one.