Phong BH. Nguyen

Blog

When Chromosomes Break

Somatic structural variants in cancer, seen from the clinic, the academic lab and big pharma

On this page Three people, one tumorWhat a structural variant isHow they drive cancerFinding them in the dataThe tools people useWhat people do with the callsAt the bedsideIn the academic labInside big pharmaLimits and workaroundsBack to the tumor board

Most talk about cancer genetics is about single letters: one base of DNA swapped for another, like a typo in a long book. But tumors also do something more violent. They tear whole pages out, photocopy chapters dozens of times, glue the end of one chapter onto the start of another, or put a paragraph back in upside down. These large-scale edits are called structural variants, and some of the best drugs in oncology exist because of them.

This post is about those edits as they happen in tumors (somatic structural variants), including the two kinds that get the most attention: copy number changes and gene fusions. It covers what they are, how they push a cell toward cancer, how software finds them in sequencing data, and where that software still falls short. Along the way three people will tell you why they care, and their reasons turn out to be quite different.

Three people, one tumor

Picture a molecular tumor board, the weekly meeting where a hospital’s specialists go through the hardest cases. Today’s case is a 46-year-old who has never smoked and has advanced lung adenocarcinoma. The first DNA panel result came back with no mutation in EGFR or KRAS, the usual suspects. An RNA test found something else: part of a gene called EML4 joined to part of a gene called ALK.

Three people at the table read that result differently.

The oncologist sees a treatment decision. An EML4–ALK fusion means the patient should get an ALK inhibitor, a pill that often controls this cancer for years, instead of chemotherapy.

The academic researcher, a computational biologist at a university cancer institute, sees a question. How did two genes on the same chromosome end up fused, what else broke at the same time, and could the shape of this genome predict how the tumor will escape the drug?

The pharma scientist, who works in translational oncology at a large drug company, sees a population. How many patients carry this fusion or its rarer cousins, how reliably can a test find them, and what will the tumor do when the current drug stops working, so that the next drug is ready?

Figure 1 One result on the table, three questions. The oncologist asks what to give the patient, the researcher asks how the genome broke, and the pharma scientist asks who else carries the same change.

Before we hear each of them out, we need the basics.

What a structural variant is

Your genome is about three billion letters long, packed into 23 pairs of chromosomes, one of each pair from each parent. A structural variant (SV) is a change that affects a stretch of at least 50 letters. That cutoff is a convention, not biology: shorter changes are called indels and are found with different methods. Real SVs in tumors range from a few hundred bases to whole chromosome arms.

The basic moves are few. A piece can be deleted. It can be duplicated, usually right next to itself (a tandem duplication), or copied many times over (an amplification). It can be flipped in place (an inversion), or moved somewhere else, often to another chromosome (a translocation). Cancers also produce messier events built from many of these at once. In chromothripsis, from the Greek for “chromosome shattering”, one region breaks into dozens of pieces in a single catastrophe and is stitched back together at random. In extrachromosomal DNA (ecDNA), pieces escape the chromosomes entirely and float in the nucleus as small circles, often carrying an oncogene.

Figure 2 The basic rearrangements, drawn as blocks of DNA. Notice which events change copy number and which only change the order.

Two terms come up again and again.

A copy number variant (CNV), or copy number alteration (CNA) in tumor jargon, is any SV that changes how many copies of a region the cell has. Deletions, duplications and amplifications are CNVs. Inversions and balanced translocations are not, because nothing is gained or lost, only rearranged. This difference matters a lot for detection, as we will see. Copy number changes also come at every scale, from one exon to an extra copy of a whole chromosome (aneuploidy), and many tumors double their entire genome at some point (whole-genome doubling).

A gene fusion is a consequence, not a separate type of event. When any SV puts its two break ends inside two different genes, the cell may read them as one hybrid gene. Translocations, inversions, deletions and tandem duplications can all make fusions. EML4–ALK, for example, comes from an inversion on the short arm of chromosome 2.

Finally, somatic is the key word in this post. Germline variants are the ones you inherited, present in every cell of your body. Somatic variants appear later in one cell and are passed only to its descendants. A tumor is a family of descendants of one cell that went wrong, so its somatic variants are a record of what went wrong. To find them, you usually sequence the tumor and a normal sample from the same person (often blood), then keep only what the tumor has and the normal lacks.

How broken chromosomes drive cancer

Every cell’s DNA breaks now and then. Repair crews usually fix it correctly, but sometimes they join the wrong ends, or copy a stretch twice while repairing it. Cancer cells are often worse at repair and under more stress when copying DNA, so errors pile up. Some tumors leave tell-tale patterns: cancers that lose the repair gene BRCA1 tend to collect many small tandem duplications, while those that lose BRCA2 collect many small deletions.

Oh, I forgot to mention: cancer biologists talk about cells as if they were cars, and the whole field has borrowed the language of driving. So here is a quick lesson from the driving instructor before we go on.

Oncogenes are the accelerator. In a healthy cell they tell it to grow and divide, but only when asked. If one is mutated, copied too many times or welded into a fusion, the pedal jams to the floor. One faulty copy is enough, because one stuck accelerator is all it takes.

Tumor suppressors are the brakes. They pause division, fix damage, or tell a cell that is beyond saving to self-destruct. You carry two copies of each, one from each parent, so a tumor usually has to wreck both before the brakes fail. A car with one working brake still stops.

Then there are the people in the car. A tumor carries a crowd of genetic changes, and nearly all of them are passengers: they happened to be in the cell that took off, they get carried along for the ride, and they do nothing but look out of the window. A handful are drivers: the changes that actually help the cell grow, survive or spread, usually by flooring an oncogene or cutting a tumor suppressor’s brake lines. Telling the two apart is one of the hardest jobs in cancer genomics, because passengers vastly outnumber drivers and nobody in the car is wearing a name tag.

Right. Back to the broken chromosomes.

Most of these SVs are passengers. A few are drivers, and they work in four main ways.

Changing the dose of a gene. More copies of an oncogene often means more of its protein. ERBB2 (better known as HER2) is amplified in roughly one in five breast cancers, and the drug trastuzumab, approved in 1998, was built to block it. Other common amplified oncogenes include MYC, EGFR, MET, CCND1 and MDM2. The reverse also matters: losing both copies of a tumor suppressor, such as CDKN2A or PTEN, takes away a brake on cell growth.

Breaking a gene. A break in the middle of a tumor suppressor can disable it even when no copy number changes.

Making a fusion. The most famous SV in medicine is the Philadelphia chromosome, first seen in 1960 in chronic myeloid leukemia and later shown to be a swap between chromosomes 9 and 22. It fuses BCR to ABL1 and makes a kinase, a signaling enzyme, that is permanently switched on. The drug imatinib, approved in 2001, blocks it and turned a deadly leukemia into a manageable condition for most patients. Kinase fusions work like EML4–ALK: the front partner contributes a promoter that keeps the gene expressed and often a region that makes two copies of the protein stick together, and the back partner contributes the kinase. Other fusions swap a promoter instead. In about half of prostate cancers, TMPRSS2–ERG puts the cancer-causing transcription factor ERG under the control of an androgen-responsive promoter, so male hormones drive its expression.

Figure 3 EML4–ALK, the fusion from our tumor board, from DNA to tumor. Step through the three states: the flip in the DNA, the jammed accelerator it creates, and the drug that releases it.

Moving a switch. Genes are controlled by enhancers, stretches of DNA that can sit far away. An SV can move a strong enhancer next to an oncogene, called enhancer hijacking. In Burkitt lymphoma, a translocation puts MYC next to the antibody gene enhancer, which is extremely active in B cells. No fusion protein is made at all, and the gene’s sequence is untouched, yet it is dramatically overexpressed.

The complex events add something more. Chromothripsis can create several drivers in one stroke. Amplified oncogenes on ecDNA are split unevenly between daughter cells, because the circles lack the machinery that shares chromosomes out fairly. That lets a tumor’s oncogene dose drift fast from cell to cell, which helps it adapt under treatment.

How common is all of this? In the Pan-Cancer Analysis of Whole Genomes (PCAWG), a 2020 study of 2,658 tumor genomes across 38 cancer types, SVs turned up in nearly every tumor, and a follow-up analysis found chromothripsis in about 29 percent of them, and in more than half of some cancer types (PCAWG, Cortés-Ciriano et al.). Some cancers, such as many sarcomas and leukemias, are defined by one specific fusion. Others, such as high-grade serous ovarian cancer, are dominated by chaotic copy number changes with few recurrent point mutations.

“I used to think of the genome as a sequence. Cancer taught me it is also a shape.” (the academic researcher)

Finding them in sequencing data

Here is the problem. A standard sequencer, the kind used for almost all clinical and research work, reads DNA in short pieces of about 150 letters, called reads. It reads both ends of each DNA fragment, usually a few hundred letters long, giving a read pair. Software then lines up every read against the reference genome, a standard human genome, to see where it came from. That gives you tens to hundreds of millions of short pieces, each placed on a map. A structural variant is a change in how the map is arranged, and you have to infer it from clues left by pieces that do not fit.

There are four main clues.

  1. Read depth. If a region is deleted, fewer reads land there. If it is amplified, more do. Count reads in windows along the genome and copy number changes show up as steps up or down.
  2. Discordant read pairs. The two ends of a fragment should land a predictable distance apart, facing each other. Pairs that land too far apart, face the wrong way, or land on different chromosomes point to a rearrangement between them.
  3. Split reads. A read that crosses a breakpoint aligns in two pieces: its first half maps to one place and its second half to another. Split reads pin the breakpoint down to the exact base.
  4. Assembly. Instead of mapping reads one by one, gather the odd reads near a suspected breakpoint and piece them together into a longer sequence, which then shows the junction directly.

A fifth clue comes from long reads, from instruments made by Pacific Biosciences or Oxford Nanopore, which read tens of thousands of letters at a time. A long read can span a whole rearrangement and show it outright.

Figure 4 The same four clues for four kinds of event. Switch to the inversion and note that read depth, the clue copy number callers rely on, stays perfectly flat.

Different callers combine these clues differently, and each family of variants has its own pipeline.

Calling breakpoints (SV callers)

An SV caller looks for the junctions themselves. It scans the alignments for clusters of discordant pairs and split reads that agree with each other, often assembles the reads around each cluster, and reports each junction as two break ends and how they are joined. Then it scores each candidate on how many reads support it, how well they map, and whether the same reads show up in the matched normal sample. A junction with support in the tumor but none in the normal is called somatic.

This approach can see every type of SV, including balanced ones that change no copy number. Its weakness is repeats. About half the genome is made of repeated sequence, and SVs love to form between repeats, exactly where short reads cannot be placed with confidence.

Calling copy number (CNV callers)

A copy number caller works with depth rather than junctions. The steps are roughly the same everywhere:

  1. Count reads in windows (whole genome) or per targeted region (gene panels and exomes).
  2. Correct biases. Regions rich in G and C bases, or hard to map, get more or fewer reads for reasons that have nothing to do with copy number.
  3. Normalize against the matched normal, or against a panel of normals, a pool of normal samples sequenced the same way, to get a log ratio of tumor depth over expected depth.
  4. Segment the noisy log ratios into stretches of constant level, using methods such as circular binary segmentation or hidden Markov models.
  5. Convert to copy numbers. This is the hard step. A tumor sample is never pure: it is a mix of tumor cells and normal cells such as immune cells and connective tissue. The tumor may also have an unusual average number of copies (its ploidy). Better callers estimate purity and ploidy together and use a second signal, the B-allele frequency (BAF). At positions where the patient inherited different letters from each parent, the share of reads carrying each letter shows the balance between the two parental copies. A loss shows up in BAF even when depth barely changes, and BAF reveals copy-neutral loss of heterozygosity, where one parent’s copy is lost and the other is duplicated, so depth is normal but half of the inherited variation is gone.

The output is allele-specific copy number: for each segment, how many copies came from each parent, for example “2+1” for a single-copy gain.

Figure 5 Slide purity down and watch the copy number steps shrink toward zero. Copy-neutral LOH is invisible in the top track at any purity; only the allele frequency track reveals it.

Calling fusions

A fusion starts as a break in the DNA, so DNA sequencing can see it. Whole-genome sequencing finds the break with the same split reads and read pairs as any other SV (Figure 4). The trouble is the targeted DNA panels most clinics run. They capture exons, the coding pieces of genes, and skip most introns, the stretches of non-coding sequence between them. A fusion break almost always falls inside an intron, and a read can only reveal it if it covers the break itself. Panels add probes across the introns of key genes, which works when the intron is short: most ALK breaks fall in one short intron, so panels usually catch EML4–ALK. Other genes have introns tens of thousands of letters long and full of repeats, where probes are impractical and reads cannot be placed. A break there, or on the side of a partner nobody thought to probe, never shows up. And even a DNA break that is found does not prove the fused gene is switched on, or spliced into a protein that works.

That is why clinical tests usually look for fusions in RNA. Splicing removes the introns, so in the messenger RNA the two genes always meet at clean exon boundaries, wherever in the intron the DNA broke. The RNA also shows that the fusion is actually expressed, and a busy fusion gene makes many copies, which helps when the sample holds little tumor. A fusion caller aligns RNA-seq reads and collects two kinds of evidence: junction reads that cross the fusion point, and spanning pairs with one end in each gene. It then filters out the usual false alarms, such as similar genes that reads cannot tell apart, neighboring genes read through as one transcript, and artifacts created in the lab. Finally it annotates whether the fusion keeps the reading frame (the three-letter rhythm in which RNA is translated into protein) and which protein domains survive.

Figure 6 Try different breakpoints. Many combinations produce a real fusion transcript, but only a few produce a protein that could drive a tumor.

Clinics also use non-sequencing methods. Fluorescence in situ hybridization (FISH) uses glowing probes on either side of a gene such as ALK and looks for the two colors coming apart. Immunohistochemistry stains tumor tissue for a protein, such as ALK, that should not normally be there. Targeted RNA panels amplify only known fusion partners, so they need very little material.

The tools people use

There are hundreds of tools. These are the ones you are most likely to meet, grouped by job. The list is not exhaustive, and no single tool wins at everything.

Job Data Common tools Notes
Somatic SV breakpoints Short-read WGS, tumor and normal Manta, DELLY, GRIDSS2, SvABA, LUMPY, BRASS GRIDSS2 and SvABA lean on assembly. Many pipelines run two or three and merge the calls
SV breakpoints Long reads Severus, nanomonsv, SAVANA (built for tumor and normal pairs); Sniffles2, cuteSV (mostly germline) Best in repeats and for complex events
Allele-specific copy number, purity, ploidy WGS, or SNP arrays for ASCAT ASCAT, Battenberg, FACETS, Sequenza, PURPLE, ABSOLUTE Battenberg also models subclones. PURPLE pairs with GRIDSS2 and LINX
Copy number Gene panels, exomes CNVkit, GATK CNV, FACETS, PureCN Panel of normals matters more than the algorithm
Copy number from cell-free DNA Low-coverage WGS of plasma ichorCNA Estimates tumor fraction in blood
Copy number from single cells scRNA-seq or single-cell DNA inferCNV, CopyKAT, CHISEL Separates clones within one tumor
Fusions RNA-seq STAR-Fusion, Arriba, FusionCatcher, JAFFA, CICERO FusionInspector re-checks candidates. Arriba is fast and tuned for clinical use
Complex events WGS calls plus copy number ShatterSeek (chromothripsis), AmpliconArchitect (ecDNA), LINX (chains of junctions) Need both breakpoints and copy number as input
Recurrence and interpretation Many tumors, knowledge bases GISTIC2 (recurrent copy number), AnnotSV, OncoKB, CIViC, COSMIC, Mitelman database Which events are drivers, and which have a drug

For benchmarking, the best somatic resources are well-studied cancer cell lines sequenced with many technologies, such as COLO829 (melanoma) and HCC1395 (breast cancer, from the FDA-led SEQC2 project). Comparing calls between tools is its own task, handled by tools such as Truvari and SURVIVOR, because two callers can describe the same event in slightly different coordinates.

What people do with the calls

A list of breakpoints is raw material. Typical analyses on top of it:

  • Driver discovery. Look for regions amplified or deleted more often than chance across many tumors (GISTIC2 does this), or genes hit by breakpoints more often than expected. The trap is fragile sites, regions that break often but harmlessly, so good methods model where breaks happen anyway.
  • Fusion catalogs. Pool fusions across cohorts, keep the in-frame ones with intact functional domains, and rank partners that recur.
  • Signatures. Just as mutational signatures summarize patterns of single-letter changes, SV and copy number signatures summarize patterns of rearrangements. They reflect the processes behind them, such as defective repair or replication stress, and two large 2022 studies built pan-cancer copy number signature sets (Drews et al., Steele et al.).
  • Homologous recombination deficiency (HRD) scores. Tumors that cannot repair breaks properly leave scars: long stretches of lost heterozygosity, imbalances at chromosome ends, and many large-scale breaks. Scores built from these scars, such as those in HRDetect, CHORD or commercial HRD tests, predict response to PARP inhibitors.
  • Timing and evolution. Combining allele-specific copy number with single-letter mutations shows which events came early and which came late, and when the genome doubled. Sequencing several regions or time points shows how clones branch and which one survives treatment.
  • Linking to expression and 3D structure. Matching SVs to RNA expression reveals enhancer hijacking. Chromosome conformation capture (Hi-C) shows how SVs rewire which DNA touches which.
  • Monitoring in blood. Tumor DNA in plasma shows copy number changes, and a fusion junction unique to one patient makes an excellent marker to track their disease.

Nobody uses all of this, and our tumor board shows why. Hand the list above to the oncologist and they will skim straight past signatures, timing and Hi-C to look for one thing: is there a variant on the report with a drug attached? An HRD score that points to a PARP inhibitor counts, and so does a blood test that shows the tumor coming back. The academic researcher lives in the middle of the list. Driver discovery, signatures, evolution and 3D structure are their whole job, and a single patient’s report matters to them mainly as one data point among thousands. The pharma scientist reads it for numbers they can plan around: how often a fusion turns up in the catalogs, whether an HRD test is reliable enough to choose patients for a trial, and how blood monitoring can catch resistance early. Same breakpoints, three reading lists. The next three sections let each of them explain theirs.

At the bedside

Here is how the oncologist sees it.

“I don’t run variant callers. I read the report, and I need it to answer three questions. Is there something I can target? Does it change the diagnosis? Is there a trial for this patient?

“Structural variants answer the first question more often than people realize. In lung adenocarcinoma alone, ALK, ROS1, RET and NTRK fusions each have approved targeted drugs, and MET exon 14 skipping, which can come from a deletion, has its own. NTRK fusions are rare in most cancers but have a drug, larotrectinib, approved in 2018 for any tumor that carries one, wherever it started. Breast cancer has HER2 amplification. Ovarian cancer has HRD scores that steer PARP inhibitor use. When a tumor carries one of these and we miss it, the patient gets a worse treatment, and we often do not find out.

“In blood cancers and sarcomas, fusions are part of the diagnosis itself. PML–RARA makes a leukemia acute promyelocytic leukemia, which is a medical emergency with its own treatment. EWSR1–FLI1 points to Ewing sarcoma. Labs that once relied on the microscope and karyotyping are moving to sequencing and optical genome mapping for this reason.

“So my concerns are practical. Did the test cover it? Many DNA panels miss fusions whose breakpoints sit in long introns, which is why guidelines push for RNA testing when a lung cancer comes back negative on DNA. Was there enough tumor in the sample? A biopsy that is 15 percent tumor can hide an amplification completely. How long will it take? A test that returns in four weeks is useless if the patient must start treatment in two. And can the report tell me what the variant means? A fusion between two genes nobody has heard of, or an amplification just above the cutoff, puts me in the grey zone, and that is where the tumor board earns its lunch.

“There is also resistance. When an ALK-positive tumor progresses on treatment, a new biopsy or blood test may show a mutation in ALK itself, or a bypass route such as MET amplification. Each points to a different next step.”

In the academic lab

The academic researcher has different worries.

“I care about the same variants, but my question is ‘how?’ and ‘what else?’, not ‘what drug?’.

“Somatic SVs are where cancer genomics still has the most open questions. Why does chromothripsis happen? We think one route is a chromosome trapped in a small separate compartment of the nucleus after a faulty cell division, where it breaks apart. How do ecDNA circles form, and why do they cluster together in the nucleus? How much of a tumor’s behavior is set by its copy number landscape, the overall pattern of gains and losses, compared with its point mutations? In many solid tumors, copy number chaos explains outcomes better than any single mutation.

“Much of my time goes into methods, because the methods are not done. We compare callers on benchmark cell lines and find they disagree a lot, especially for small events, events in repeats and complex rearrangements. We build tools that combine breakpoints and copy number into a single graph of the tumor genome, because a junction list without copy number, or copy number without junctions, is half a picture. We work with long reads, single-cell DNA and spatial methods to see SVs in the clones and tissue regions where they actually happen.

“Big cohorts made this field. PCAWG, the Hartwig Medical Foundation’s whole-genome data on metastatic tumors, and Genomics England’s clinical genomes gave us thousands of tumors sequenced the same way. With that, rare events start to look recurrent, and signatures become visible.

“What I want from the field is shared truth sets, reusable pipelines and honest reporting of what a method misses. Half of the SV papers I review claim to beat everything else, on a benchmark the authors picked.”

Inside big pharma

The pharma scientist works on a longer timeline and at a larger scale.

“My job is to make sure the right drug reaches the right patients, which means my questions are about numbers, reliability and what comes next.

“Structural variants feed the pipeline at every stage. In target discovery, a recurrent in-frame kinase fusion is close to the ideal target: the tumor depends on it, normal cells do not have it, and the test for it is clear-cut. That logic gave us drugs for ALK, ROS1, RET, NTRK and FGFR2 fusions. Copy number gives us targets too, sometimes indirectly. When CDKN2A is deleted, its neighbor MTAP often goes with it, and cells without MTAP become sensitive to a particular kind of PRMT5 inhibitor, now in clinical trials. Tumors with CCNE1 amplification are being tested against inhibitors of cell-cycle checkpoint proteins.

“In preclinical work, we screen hundreds of cancer cell lines. A known trap: in CRISPR screens, cutting a highly amplified region kills a cell simply because there are so many cuts, whatever the gene does. Without correcting for copy number, amplified genes look like great targets when they are not.

“In clinical development, a rare fusion means a rare patient. If a fusion occurs in 1 percent of lung cancers, we screen a hundred patients to enroll one, so tissue-agnostic basket trials and good diagnostic partners matter a great deal. The test becomes a companion diagnostic, approved alongside the drug, and its definitions become regulatory facts. What exactly counts as amplified: six copies, a ratio of two over the chromosome average, something else? Different assays answer differently, and the answer decides who gets the drug.

“In life-cycle management, we study why patients stop responding. Resistance often comes through SVs: amplification of the target, amplification of a bypass gene such as MET, or oncogenes on ecDNA that let the tumor raise its dose quickly. Blood-based testing lets us follow this across thousands of patients in trials and real-world clinical-genomic databases.

“So my wish list is different from the academic’s. I need assays that are reproducible across labs and countries, thresholds that hold up, data from formalin-fixed tissue that actually exists in hospitals, and calls I can defend to a regulator. A beautiful method that only works on fresh-frozen, high-purity whole genomes does not help my trial.”

Limits and how to work around them

All three hit the same walls, from different sides.

Limitation Who feels it most What helps
Repeats. Short reads cannot be placed in repeated regions, where many SVs form Academic researcher Long reads, assembly-based callers, newer reference genomes (T2T-CHM13) and pangenome references that include common variation
Low purity and subclones. Normal cells and minor clones dilute every signal (Figure 5) Oncologist Pathologist review of tumor content, deeper sequencing, callers that model purity and subclones, BAF alongside depth
Callers disagree. Each tool sees a different subset, and their coordinates differ Everyone Run several callers and merge them, benchmark on COLO829 or HCC1395, confirm important calls another way (PCR, FISH, long reads, RNA)
No matched normal. Common with panels, cell lines and archived samples Pharma scientist Panels of normals, population SV databases such as gnomAD-SV to remove inherited variants, careful filters
Tumor in the “normal”. In blood cancers the blood sample may contain tumor cells Oncologist Use skin or saliva as the normal, and callers that tolerate tumor-in-normal
Formalin-fixed tissue. Fixation fragments DNA and RNA and creates fake junctions and noisy depth Oncologist, pharma scientist Targeted RNA methods that work on short fragments, artifact filters built for fixed tissue, panels of normals made from fixed samples
Panels miss things. Breakpoints in uncovered introns, unknown partners, balanced events Oncologist RNA-based fusion testing, tiling probes across long introns of key genes, whole-genome sequencing where turnaround allows
Ploidy guesses can be wrong. A tumor that doubled its genome can fit two solutions equally well Academic researcher Use point mutation allele frequencies to constrain the fit, flag ambiguous solutions, review by an analyst
Low tumor fraction in blood. Copy number vanishes when tumor DNA is a few percent of plasma DNA Pharma scientist, oncologist Deeper targeted sequencing, assays designed around a patient’s own tumor, signals from DNA fragment patterns
Interpretation. Telling drivers from passengers, setting amplification cutoffs, fusions with unknown partners Everyone Frame and domain annotation, curated knowledge bases (OncoKB, CIViC), shared tiering guidelines (AMP/ASCO/CAP), functional tests in the lab
Inconsistent formats. The same event written in different ways in VCF files Academic researcher, pharma scientist Standard breakend notation, tools like Truvari for comparison, shared pipelines such as nf-core/sarek or Hartwig’s open-source pipeline

A few broader trends are easing these limits. Long-read sequencing keeps getting cheaper and now works on clinical samples, and it resolves many of the repeat and complexity problems in one go. Whole-genome sequencing is becoming routine care in a few health systems, which gives SVs and copy number for free alongside point mutations. And graph-based representations of tumor genomes, which tie breakpoints and copy numbers into one consistent model, are slowly replacing separate lists that do not agree with each other.

Back to the tumor board

The patient starts an ALK inhibitor. The oncologist notes the result, books a follow-up scan and plans a blood test if the cancer progresses. The academic asks whether the leftover sample can be long-read sequenced, to see what else rearranged around the inversion. The pharma scientist adds one more patient, anonymized, to a mental count of how often these fusions show up in never-smokers, and of how often, months or years later, MET gets amplified.

They are looking at the same broken chromosome. For one it is a prescription, for another a mechanism, and for the third a market and a trial design. Somatic structural variants are some of the most useful things we can find in a tumor, and some of the hardest to find well. Getting better at them takes all three points of view: methods built by people who care how genomes break, tested on the samples that exist in hospitals, and judged by whether a patient gets the right drug.

A note on sources. The patient and the three narrators are fictional composites. Biology, drug approvals and method descriptions are written from general knowledge of the field; the pan-cancer figures come from PCAWG and Cortés-Ciriano et al., and the copy number signatures from Drews et al. and Steele et al.. Frequencies are approximate and vary between studies. The tool list reflects common practice at the time of writing and is not an endorsement. The figures are simplified, illustrative models.

Comments