Chemotherapy kills ninety-nine percent of the tumor's cells. The scans look clear. The oncologist uses the word remission. Then, months or years later, the tumor returns — and this time it is harder to treat than the original. The cancer stem cell hypothesis proposes an explanation: that one percent that survived was not a random remnant. It was the engine. A small population of cells with the capacity to self-renew, to differentiate into the full range of cell types in the tumor, and to regenerate the entire hierarchy from scratch. Killing the bulk without killing the apex produces a response that looks like a cure and functions like a pause. That is the central claim. It is one of the most consequential ideas in contemporary oncology, and it is contested in ways that matter.

Normal tissue renewal runs on division of labor. A small stem cell compartment at the top performs two functions: self-renewal, producing daughter cells that remain stem cells, and differentiation, producing the bulk of functional working cells. Those bulk cells do the tissue's job. They are numerous. They cannot regenerate the organ. The cancer stem cell hypothesis says tumors borrow this architecture. At the apex is a small cancer stem cell population with self-renewal and differentiation capacity. Below it is a bulk of non-regenerating progeny — the cells that fill the tumor mass and that a standard cytotoxic drug can kill. The stakes follow directly from the architecture: if the hierarchy is real, response to therapy must be evaluated not by how many cells die, but by whether the apex cells were among them. A drug that clears the bulk and spares the apex produces remission. A drug that kills the apex can produce a cure, even if the bulk takes longer to clear.

The AML finding was not an isolated result. Similar hierarchy evidence emerged in solid tumors over the following decade. In breast cancer, a CD44-positive CD24-negative population was identified where as few as two hundred cells initiated tumors in xenografts — in experiments where twenty thousand cells from the bulk population could not. In glioblastoma, CD133-positive cells formed neurospheres in culture, initiated tumors serially in mice, and were markedly more resistant to radiation and chemotherapy than CD133-negative cells. In colorectal cancer, Lgr5-positive cells — the same surface marker that identifies stem cells in the normal intestinal crypt — marked the tumor-initiating compartment. A consistent phenotypic profile emerged across cancer types: high aldehyde dehydrogenase activity, Hoechst dye exclusion via ABC transporters, slow cell cycling, non-adherent sphere formation in suspension culture, and disproportionate tumor-initiating capacity in xenograft assays.

The assay problem is an empirical challenge. Plasticity is a conceptual one. Non-stem breast cancer cells — CD44-negative, CD24-positive, with no tumor-initiating capacity in standard assays — placed in conditions that mimic the tumor niche can reacquire the CD44-positive CD24-negative phenotype and regain tumor-initiating capacity. The state transition works in both directions. In glioblastoma, selectively depleting the CD133-positive cancer stem cell fraction does not produce the expected result. The remaining CD133-negative cells, exposed to appropriate microenvironmental signals, convert back into CD133-positive cells and reconstitute the apex. Plasticity breaks the simple hierarchy model in an important way. If stemness is a reversible cell state rather than a fixed identity, then targeting a surface marker eliminates only the cells that currently express it. The tumor re-promotes its stem population from the bulk that survives. The marker keeps reappearing because the state that produces it keeps being re-induced. You are not eliminating a cell type. You are competing with a dynamic equilibrium.

The current synthesis sits between the strong original claim and full rejection of the concept. Most cancers do show some degree of hierarchical organization: a smaller population with disproportionate regenerative capacity, identifiable by functional assays, not simply an artifact of xenograft stringency. That much appears solid. What is revised is the model of fixed, irreversible cell identity. Stemness is better understood as a cell state — shaped by genetic programs and niche signals, and at least partially reversible in response to microenvironmental cues. The degree of hierarchy varies substantially across cancer types and across individual patients with the same cancer type. Some tumors are tightly hierarchical with a clearly defined apex population. Others are flat, with most cells capable of tumor initiation under the right conditions. No currently available clinical assay can tell an oncologist which category a particular patient's tumor falls into. That gap between the concept and the clinical tool is where most of the unsolved problems live.

The preferential survival of cancer stem cells through cytotoxic therapy is not a coincidence. Several overlapping mechanisms converge on the same functional outcome. First, quiescence. Most standard chemotherapy agents target actively dividing cells — they disrupt DNA replication, spindle formation, or both. A cell sitting in G-zero, the quiescent state outside the cell cycle entirely, is largely invisible to these drugs. Cancer stem cells in many tumors are slow cycling or transiently quiescent, which makes them intrinsically resistant to cycle-dependent cytotoxins. Second, ABC transporter-mediated drug efflux. The same membrane pumps responsible for excluding Hoechst dye — the basis of the side-population assay — actively pump chemotherapy agents out of the cell before intracellular drug concentrations reach lethal levels. Third, more efficient DNA repair and elevated BCL-2 family expression, which raises the apoptotic threshold. And fourth, the niche: surrounding endothelial cells supply Notch ligands, hypoxia drives HIF-1-alpha to maintain stemness transcription factors, and stromal Wnt signals maintain self-renewal. Even a drug that reaches the cancer stem cell must overcome these layered resistances.

Cytotoxic therapy does not kill cancer cells randomly. It kills the cells most vulnerable to it — which are the actively cycling, drug-sensitive cells of the bulk population. The bulk dies. The response is dramatic. The scan looks clear. But the therapy has simultaneously selected for the fraction least vulnerable to it: the quiescent, drug-effluxing, repair-proficient, niche-supported stem-like cells. By clearing the competition, therapy enriches the residual tumor for exactly the cells with highest regenerative capacity. The regrown tumor is not a random sample of the original. It has been filtered. And having been selected through a round of chemotherapy, it is now enriched for resistance mechanisms that were already present, simply rare. Remission followed by relapse with progressively harder-to-treat disease is the expected clinical signature of this process. Minimal residual disease monitoring in leukemia is the clinical acknowledgment of this dynamic: MRD-positive patients in morphological remission — no tumor visible by microscopy — relapse because a stem-like population below the detection threshold is already rebuilding the hierarchy.

The cancer stem cell concept is real, contested, and clinically necessary to hold with appropriate nuance. Real: most cancers show hierarchical organization, with a smaller population having disproportionate regenerative capacity, and this is not simply an artifact of the xenograft assay. Contested: the degree of hierarchy varies enormously across cancer types and patients, plasticity undermines simple marker-targeted strategies, and no available clinical tool measures how hierarchical a particular patient's tumor actually is. Clinically necessary to hold with nuance: because the implication for therapy design is direct. A drug program evaluated only by overall response rate may be selecting against its own durability. Trials that measure minimal residual disease, that track cancer stem cell population dynamics through treatment, that distinguish hierarchical from non-hierarchical tumors — these ask the right questions. The unsolved problem is not lack of targets. Notch inhibitors, Wnt inhibitors, Hedgehog inhibitors, and ALDH inhibitors all exist. The unsolved problem is combining apex-targeting with anti-plasticity strategy so that eliminating the stem fraction does not simply trigger re-promotion from the surviving bulk.