Closing the gap between tool development and biology to decode extrachromosomal DNA in cancer

ecDNA formation is one of the most paradigm-defying things a tumour can do to its genome. When ecDNA forms, the tumour genome relocates oncogenes from chromosomes and carries them instead on circular DNA molecules that don't obey the rules of normal inheritance or gene regulation. Cancers can amplify these driver genes to high copy numbers, and create rates of adaptation that chromosomal amplification can't match. Studies over the years have found patients with tumours bearing ecDNA or even other types of focal amplifications tend to have worse outcomes.

ecDNA leaves a distinctive fingerprint in whole-genome sequencing data, but decoding that fingerprint means combining copy number and structural rearrangement calls across many genomic regions that have shattered into dozens of pieces and come together into a larger structure. Those combined data can be transformed computationally into a graph structure with nodes and edges. However, a reconstruction on its own doesn't directly tell you whether what you're looking at is the product of an ecDNA-based mechanism or something else. We knew that the field needed a toolkit that was tested, reproducible, and able to take some of the subjectivity out of interpreting and classifying these complex genome structures.

Analysing the structure of the cancer genome is computationally expensive, since it involves processing DNA sequencing data that quickly sums to terabytes across large collections of samples. Some of those datasets, like The Cancer Genome Atlas (TCGA) and the Cancer Cell Line Encyclopedia (CCLE), are used routinely across the cancer genomics community. It made little sense for every group to keep repeating the same expensive computation when they needed to reference those results. So alongside our reproducible toolkit, we built a community-editable platform where anyone can find ecDNA predictions across these datasets, and contribute their own results for the rest of the field to use.

Our driver workflow is called “AmpliconSuite-pipeline”, and it takes the analysis from end to end. Sequencing data goes in, and a classified focal amplification comes out, labelled as ecDNA, a breakage-fusion-bridge cycle, or one of several chromosomal forms. At its core is AmpliconArchitect, which reconstructs the local genome structure and now runs much faster than the version we first released in 2019. On top of that sits AmpliconClassifier, which applies the rules that turn a reconstruction into a focal amplification call.

A new module in the workflow, ecContext, asks a different question: does an ecDNA's structure tell us anything about how it formed? Circles that arise from a clean excision look structurally different from circles assembled out of a chromothriptic shattering event, and those differences are visible in the reconstructions.

Across 6,506 tumours, ecDNA carried oncogenes at 2.5 times the density of the reference genome. This tells us that sequences that become amplified on ecDNA in the tumour must have a reason to be retained.

The pattern that surprised us most was co-amplification. CDK4 and MDM2 usually sit more than 10 megabase-pairs apart on chromosome 12. However, when they are amplified together, 72% of the time ecDNA seems to have played a role in bringing those distant loci together – even if they ultimately end up embedded back in a chromosome as some sort of neochromosomal structure. Elsewhere we saw ecDNA repeatedly packaging chromatin remodellers alongside driver oncogenes, as though assembling a self-contained support system. EcDNA formation mechanisms also vary by tissue, with prostate and brain tumours leaning toward simple excision while skin cancers and sarcomas skew chromothriptic.

Infrastructure work rarely produces a headline finding on its own, but science doesn't work well when data and results stay siloed on individual servers. AmpliconRepository.org, is a collaboration between Jill Mesirov’s lab and Vineet Bafna’s lab, which Ted Liefeld, Edwin Huang and I have been developing since 2021. It now hosts ecDNA predictions from nearly 20,000 publicly available samples across 33 projects, drawn from sixteen distinct publications. Anyone can create a project, upload their own results, and query across everyone else's.

The site is built to support real analysis rather than just browsing. Users can search by sample name, project, gene, cancer type or amplicon class, combining terms with boolean operators, then download an entire matching collection to analyze themselves. Each sample comes with an interactive copy-number profile, the underlying genome reconstructions, and an embedded genome browser for inspecting the genes an amplicon carries. A co-amplification analysis tool lets users ask which genes travel together on ecDNA across whole cohorts. Projects are versioned and can be kept private while a study is still in progress, results can be uploaded by completing a simple web form, and retrieved from the command line, and the site now accepts long-read reconstructions alongside short-read ones.

Much of what the site can do was inspired by the scientific questions asked by our team members in eDyNAmiC. Working within a large team means tool developers and the people using the tools are in direct contact, so a bottleneck someone hits in their own analysis becomes a feature request, then a release, becoming the edge that makes the next study possible. We have run that loop for years now: building something for a problem, watching how people actually use it, finding out what is missing, and refining. I don't think the tools would work as well as they do had they been designed at a distance from the biology they serve.

I'm from the Pacific Northwest, and I studied at the University of Washington in Seattle before moving south to start a PhD in Bioinformatics & Systems Biology at University of California San Diego. My thesis work, with Vineet Bafna, was on computational methods for reconstructing ecDNA, which at the time was a considerably smaller field than it is now. As I was finishing, our research team, led by Paul Mischel, was selected for a Cancer Grand Challenges award, and I decided to stay on in Vineet’s lab and support the bioinformatics and computational biology work rather than move elsewhere, which turned out to be one of the better decisions I've made. Team eDyNAmiC is now in its fifth and final year, and most of the past decade of my career has gone into growing these tools into an end-to-end research platform. I'm currently an Associate Project Scientist at the University of California San Diego, living and working remotely from the Seattle area.

What draws me to tool-building is that the work compounds. An analysis script helps one paper, while a method other people rely on supports dozens of studies. Working alongside cytogeneticists, clinicians, mathematicians, and patient advocates who each see ecDNA from a completely different angle has shaped the analysis tools as much as any algorithmic idea has.