A 58-year-old never-smoker. Metastatic lung cancer. Genomic profiling returns a KRAS G12C mutation — an actionable alteration, with approved drugs. She starts sotorasib. The tumor shrinks for four months. Then it grows back. Down the hall, a second patient with the exact same mutation gets the same drug. Zero shrinkage from day one. Same biomarker, same drug, two different failures. The match was real in both cases. The response was not. That gap is everything this chapter is about.

For most of oncology's history, drugs were assigned by tissue type. This chemotherapy for everyone with this cancer, regardless of molecular features. Some responded. Others didn't. The difference was called biological variability — a phrase that meant the field didn't know why. Precision oncology is the reframing of that variability as molecular. The insight: the cancer's genetic and protein features, more than its tissue of origin, predict which drugs will work. Sequence the tumor. Find the molecular feature. Prescribe the matched drug.

An actionable alteration is the unit of precision oncology: a molecular feature for which a matched therapy exists and has demonstrated benefit. Actionable features include point mutations like EGFR exon-19 deletion and BRAF V600E, gene fusions in ALK, ROS1, RET, and NTRK, copy-number changes like HER2 amplification, and functional signatures like microsatellite instability and tumor mutational burden. The word actionable is load-bearing. It does not mean present. A mutation you can detect but cannot exploit is interesting biology, not an actionable alteration.

The single most important distinction in this chapter — the one students most consistently collapse — is between a predictive biomarker and a prognostic biomarker. A prognostic biomarker tells you how the patient will fare, regardless of treatment. It forecasts natural history. A predictive biomarker tells you whether a specific therapy will work. EGFR exon-19 deletion predicts response to osimertinib. BRCA1 and BRCA2 germline mutations predict response to PARP inhibitors through synthetic lethality. Precision oncology runs on predictive biomarkers. Conflating the two produces a specific clinical error: treating a marker of aggressive biology as if it were a drug target. The feature predicts bad outcomes not because it drives a blockable pathway, but because it marks an inherently aggressive tumor type. Treating it does nothing.

The most evolved expression of predictive-biomarker logic is tumor-agnostic therapy: a drug approved for any cancer carrying a specific molecular feature, regardless of where in the body it originated. The first tumor-agnostic approval was pembrolizumab for microsatellite-instability-high tumors in 2017 — approved based on response rates of roughly 30 to 50 percent spanning colorectal, endometrial, gastric, and other sites. Then larotrectinib for NTRK fusion-positive tumors, with response rates above 75 percent across all histologies. The claim embedded in every tumor-agnostic approval: the biology is what matters; the tissue is incidental. That claim is powerful. It is also not without limits.

Precision oncology needed new trial designs because the molecular subgroups it works with are small and scattered across many cancer types. Basket trials enroll patients by molecular alteration across many tumor types — the design that produces tumor-agnostic claims. Umbrella trials work in the opposite direction: they enroll one cancer type and subdivide patients into arms by molecular feature. Lung-MAP is the classic example. Platform trials maintain permanent infrastructure and add or drop treatment arms over time as evidence accumulates. Reading a precision-oncology result requires knowing which design produced it. A basket result supports a tumor-agnostic claim. Citing a basket result to justify treatment in a tumor type not represented in the basket is an extrapolation, not an evidence application.

The productive reframing separates the target from the tumor's dependence on that target over time. At baseline, the cancer was driven by KRAS G12C signaling; sotorasib blocked it, and the tumor shrank. The hypothesis — this tumor depends on KRAS G12C — was confirmed by the response. Resistance emerged not because the drug stopped binding, but because the tumor evolved. Treatment is selection pressure. It acts on a population under selection, selecting for cells capable of routing around the blockade. New downstream mutations restored the MAPK signal; a subclone adopted a different driver entirely. The drug still engaged its target. The cancer found another way to grow.

Three possible outcomes across the whole population of KRAS G12C patients: durable benefit, transient benefit followed by resistance, and no benefit at all from day one. The match identified a candidate dependency. Whether each patient's tumor actually depended on it — and kept depending on it — was an empirical question that the drug answered. This is why matched and responded are different measurements. The report says what alteration is present. It does not say whether the tumor depends on it. It does not say whether that dependency will persist under selective pressure. The field is working toward pre-treatment tools that would close this gap: organoid testing, AI-integrated multi-omic models, functional assays of pathway dependence. None has yet reached reliability. Until they do, the match remains the hypothesis, and the response remains the test.