For most of the sequencing era, cancer genomics has been told in single letters. Point mutations in KRAS, TP53 and PIK3CA dominate driver catalogues, partly because they are the variants our instruments detect most easily. Many tumor genomes, however, look less like a manuscript with typos and more like one that has been torn apart, reassembled out of order and partially photocopied.

Structural variants include deletions, duplications, inversions, translocations and the complex composites that combine them. They alter far more base pairs than point mutations ever do. They fuse coding sequences from distant loci, relocate enhancers next to oncogenes, and shift gene dosage across entire chromosome arms. Pan-cancer analyses indicate that most tumors carry rearrangement-driven changes, and in some cancer types rearrangements rather than substitutions are the principal engine of transformation.

From an engineering standpoint, this is evolution with blunt instruments. Breakage, misrepair and missegregation generate variation, and relentless selection filters it. Understanding the process requires three things. The first is technologies that can resolve rearranged genomes. The second is mechanistic models of how fusions acquire oncogenic function. The third is a dynamic view of how copy number is selected over time. Each also points toward a therapeutic lever, and toward lessons for anyone trying to direct genome evolution deliberately.

Short-read sequencing infers structure indirectly. Reads are roughly 150 bases and fragments a few hundred bases long, so a rearrangement is detected only through its fingerprints: read pairs that map at unexpected distances or orientations, split reads spanning a breakpoint, and shifts in read depth. These signals work well for simple events in unique sequence. They fail where cancer genomes are often most interesting, inside segmental duplications, satellite arrays and mobile-element repeats where breakpoints tend to cluster.

Long-read platforms turn an inference problem into an observation problem. PacBio HiFi reads span 15 to 25 kilobases at high accuracy, while Oxford Nanopore ultra-long protocols produce molecules exceeding 100 kilobases. A single read can traverse several adjacent junctions, revealing the order and orientation of rearranged segments rather than leaving them to be stitched together computationally. Long reads also carry native methylation signal, so altered regulatory states can be read alongside the rearrangements that produced them.

Optical genome mapping works at a coarser but complementary scale. Fluorescently labeled sequence motifs are imaged along linearized DNA molecules hundreds of kilobases long, producing barcodes that can be aligned against a reference. Some large insertions, balanced translocations and inversions are cryptic to sequencing because their breakpoints sit in unmappable sequence. On an optical map, they appear as unmistakable disruptions in label patterns. In haematological malignancies, the approach is increasingly evaluated as a single-assay alternative to combined karyotyping, FISH and microarray workflows.

These methods matter most for chromothripsis, the catastrophic shattering of one or a few chromosomes followed by largely random religation. First described in 2011, it leaves a distinctive signature: copy number oscillating between two states, dense clusters of breakpoints, and fragments joined in all four orientations. Later work traced its origin to micronuclei and chromatin bridges, where missegregated chromosomes are damaged in isolation. Pan-cancer surveys showed it is far from rare and is present in more than half of cases in some sarcomas and melanomas.

Short reads can detect chromothripsis but rarely resolve it. Reconstructing the derivative chromosome means establishing which fragments survived, in what order, and on which haplotype. That requires reads or molecules spanning multiple junctions. The reconstruction has practical consequences. It distinguishes events that merely disrupted a tumor suppressor from those that assembled a new oncogenic configuration. It also exposes the complex amplicons and extrachromosomal circles that frequently emerge from the same catastrophe.

Takeaway

A variant caller finds only the variant classes it can see. The apparent dominance of point mutations in cancer catalogues is partly a measurement artefact, and every gain in read length redraws the map of what drives tumors.

The Philadelphia chromosome, observed in 1960 and later resolved as a reciprocal t(9;22) translocation, established the template for oncogenic fusion. The breakpoint joins the 5′ portion of BCR to most of ABL1. BCR contributes a coiled-coil oligomerization domain that drives the ABL1 kinase into constitutive activity. The fusion also loses the autoinhibitory N-terminal cap that normally restrains ABL1. The result is a signaling protein that no longer needs a signal.

Most kinase fusions follow the same logic. EML4-ALK arises from an inversion on chromosome 2p in a subset of lung adenocarcinomas, and NTRK fusions occur across dozens of tumor types. Both pair a kinase domain with a partner that supplies dimerization and a strong, broadly active promoter. The partner is often interchangeable; the kinase is the constant. This modularity explains why larotrectinib earned tumor-agnostic approval: the selected function is defined by the domain, not the tissue.

A second class works through regulatory substitution. In roughly half of prostate cancers, an interstitial deletion or translocation on 21q22 places ERG under the androgen-responsive promoter of TMPRSS2. The ERG protein is largely intact; what changes is when and how much of it is made. Androgen signaling itself appears to bring the two loci into proximity and recruit topoisomerase II beta, generating the double-strand breaks that fuse them. In effect, a hormone helps write its own oncogenic rearrangement.

The extreme case requires no fused transcript at all. In Burkitt lymphoma, t(8;14) juxtaposes MYC with immunoglobulin heavy-chain enhancers, driving expression at levels no normal regulatory context would permit. This kind of enhancer hijacking has since been documented well beyond lymphoid tumors, including activation of GFI1 family genes in medulloblastoma. Rearrangements that delete or invert topologically associating domain boundaries achieve the same effect, letting enhancers contact genes they were previously insulated from.

Nuclear architecture therefore constrains which fusions are possible, while selection determines which persist. Loci that are co-regulated or spatially adjacent rearrange more often, so recurrent fusions reflect both elevated formation rates and strong fitness advantages. To an engineer, the process resembles a crude domain-shuffling screen. A tumor population generates many random junctions, and only those that create constitutive signaling or deregulated transcription expand clonally.

Takeaway

An oncogenic fusion is rarely a new invention. It is usually an existing function released from regulation, either by giving a protein a permanent partner or by giving a gene the wrong instructions.

Copy number alteration is the most pervasive form of structural change in cancer. Most solid tumors are aneuploid, and focal amplifications and homozygous deletions concentrate on a remarkably consistent set of loci across tissues. MYC, CCND1, EGFR and ERBB2 are commonly gained, while CDKN2A, PTEN and RB1 are commonly lost. Recurrence on this scale is not random accumulation. It is the statistical signature of positive selection acting on dosage.

Arm-level changes are harder to interpret because each affects hundreds of genes. One influential model proposes that their selection reflects cumulative dosage effects. Arms rich in tumor suppressors and poor in oncogenes tend to be lost, while the inverse tend to be gained. Tissue-specific patterns follow, such as 3p loss in clear-cell renal carcinoma and recurrent 1q and 8q gains across many carcinomas. Here the arm, not the gene, becomes the unit of selection, trading precision for magnitude of effect.

Whole-genome doubling reshapes the evolutionary landscape. It occurs in roughly a third of tumors and is associated with poorer prognosis. Doubling buffers cells against the deleterious consequences of later chromosome losses and loss of heterozygosity. Doubled genomes can then shed chromosomes freely, exploring copy-number states that would be lethal in a diploid background. In this sense, doubling functions less as a driver than as an evolvability enhancer, widening the range of tolerated variation.

Extrachromosomal DNA takes this further. Circular ecDNA amplicons lack centromeres and segregate unevenly at mitosis, so copy number can differ dramatically between daughter cells. Present in a substantial minority of tumors and especially common in glioblastoma, ecDNA lets populations respond to selection within a few divisions. Glioblastoma cells have been shown to lose EGFR-bearing ecDNA under EGFR inhibition and regain it once therapy is withdrawn. This is a reversible, dosage-based form of resistance.

Therapeutically, copy number is both a target and a vulnerability. ERBB2 amplification defines the patient population for HER2-directed therapy. Deletions offer a subtler opportunity. CDKN2A loss frequently takes the neighboring MTAP gene with it, so the metabolite MTA accumulates and partially inhibits PRMT5. MTA-cooperative PRMT5 inhibitors exploit this collateral damage selectively, turning a passenger deletion into a therapeutic window that is absent from normal tissue.

Takeaway

Tumors tune gene dosage rather than gene sequence, and the passengers swept along by large deletions can become as valuable to therapy as the drivers that selected them.

Structural variation reframes cancer as an evolutionary process with unusually large step sizes. Point mutations explore sequence space one residue at a time. Rearrangements recombine domains, reassign regulatory elements and rescale dosage in single events, sometimes dozens at once, as chromothripsis shows.

The consequences for genetic technology are concrete. Long-read and optical methods are making complete, haplotype-resolved tumor genomes feasible. With them come clinical assays that capture fusions, ecDNA and complex amplicons alongside conventional variants. Longitudinal sampling could then track copy-number trajectories under therapy, anticipating resistance rather than discovering it.

The deeper lesson concerns control. Engineers who design synthetic genomes or run directed-evolution campaigns can learn from tumors which architectural changes evolve most readily and which leave lasting dependencies. Predicting cancer's next rearrangement may prove as important as targeting its last.