💻 bioinformatics

Inferring ecDNA Evolution from Allele Frequency Spectra

This paper introduces ecEvol, a computational method that analyzes allele frequency spectra from bulk whole-genome sequencing to infer the timing, selection strength, and clonal dominance of extrachromosomal DNA (ecDNA) emergence, revealing that early ecDNA formation driven by oncogenes correlates with poorer patient survival and that ecDNA dynamics vary significantly between untreated and advanced tumors.

Original authors: Dameracharla, B., Prasad, G., Luebeck, J., Amin, S., Kirkland, J., Chapman, O. S., Chavez, L., Verhaak, R., Huang, W., Mischel, P. S., Werner, B., Bafna, V.

Published 2026-10-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dameracharla, B., Prasad, G., Luebeck, J., Amin, S., Kirkland, J., Chapman, O. S., Chavez, L., Verhaak, R., Huang, W., Mischel, P. S., Werner, B., Bafna, V.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Cancer is not a single event but a long, chaotic story of cells changing and competing. Within a tumor, different groups of cells, called subclones, arise as they accumulate genetic mistakes. Usually, these mistakes happen on the chromosomes, the long strands of DNA that act as the cell's instruction manual. These chromosomal changes are passed down to daughter cells in a predictable way, like a family heirloom. However, some aggressive tumors carry a different kind of genetic troublemaker: extrachromosomal DNA. These are circular rings of DNA that float freely inside the cell, separate from the main chromosomes. Because they lack the anchor that holds chromosomes in place during cell division, these rings are tossed into daughter cells randomly. One cell might end up with dozens of rings, while its neighbor gets none. This randomness allows the rings to multiply rapidly, often carrying genes that make the cancer grow faster and resist treatment.

For years, scientists have known that these floating rings are a sign of a dangerous cancer, but they have been unable to tell exactly when the rings first appeared in a patient's tumor history. Did they show up at the very beginning, driving the cancer from its start? Or did they appear much later, perhaps as a desperate reaction to chemotherapy? Without knowing the timing, it is difficult to understand how these rings shape the disease or why some patients survive while others do not. Now, a team of researchers has developed a new way to look back in time. By analyzing the patterns of genetic mutations found in a single tumor sample, they can infer when these floating rings emerged, how strongly they helped the cancer grow, and how much of the tumor they now control.

The researchers, led by computational scientists at the University of California, San Diego, created a tool called ecEvol. To build it, they first had to teach a computer what the genetic signature of a tumor with floating rings looks like. They wrote a computer program that simulated the life of a tumor, starting with a single cell and letting it divide millions of times. In these simulations, they introduced the floating rings at different times and gave them different levels of "fitness," or the ability to help the cell divide faster than its neighbors. They watched how the genetic mutations spread through the population of cells. They found that the presence of these rings leaves a unique fingerprint in the data. While normal chromosomal mutations create a smooth, predictable pattern of frequencies, the floating rings create a distinct spike of very high-frequency mutations and a gap where medium-frequency mutations should be. This happens because the rings copy themselves so many times that they drown out the normal chromosomes in the cells that carry them.

Using these simulated fingerprints, the team trained a sophisticated computer model to recognize the patterns in real-world data. They tested the model on thousands of simulated tumors where they knew the exact answers, and the model learned to predict the timing of the rings' appearance, their growth advantage, and their mutation rate with high accuracy. The model proved to be far better at this task than previous methods, which were designed for standard chromosomal mutations and could not see the unique signal of the floating rings. Crucially, the team also taught the model to handle the messy reality of real patient samples, where the tumor tissue is often mixed with healthy tissue. By training the model on data that included these imperfections, they ensured it would work even when the exact amount of healthy tissue in a sample was unknown.

When the researchers applied this new tool to real tumor samples from large cancer databases, they uncovered a clear story about the life cycle of these floating rings. In primary tumors, those that had not yet been treated with chemotherapy, the floating rings usually appeared very early in the cancer's development. They often emerged within the first few dozen cell divisions, long before the tumor was large enough to be detected. These early arrivals were also the most successful; they grew so strongly that they came to dominate the tumor, making up nearly all of the cells in the sample. The study found that when the rings carried known cancer-driving genes, they appeared even earlier and took over the tumor even more completely than rings that did not carry these specific genes.

The story changed, however, when the researchers looked at advanced tumors, particularly those from patients who had already received treatment. In these cases, the floating rings appeared much later in the tumor's history. The data suggests that the stress of chemotherapy or the progression of the disease itself triggered the formation of new rings. These late-arriving rings were also very fit, growing rapidly to take over a significant portion of the tumor, but because they started later, they had not yet had time to dominate the entire mass as completely as the early rings did. The researchers also tracked patients who had biopsies taken at different times. In these cases, they saw that the proportion of cells carrying the rings increased over time, confirming that these lineages continued to expand even after the initial detection.

Perhaps most importantly, the timing of these rings mattered for the patient's survival. The study found that patients whose tumors had floating rings that emerged very early tended to have worse outcomes than those whose rings appeared later. This suggests that the earliest appearance of these rings gives the cancer a head start, allowing it to build a massive, dominant population of aggressive cells before treatment can even begin. The researchers also noted that the strength of the rings' growth advantage and the fraction of the tumor they controlled were linked to survival, but the timing of their arrival was the strongest signal.

This work does not just offer a new way to measure cancer; it offers a new way to see its history. By looking at the distribution of genetic mutations in a single biopsy, the researchers can now reconstruct the evolutionary timeline of the tumor's most dangerous component. They have shown that the floating rings are not just a random feature of advanced cancer but a strategic player that often sets the stage for the disease's aggression right from the start. While the study relied heavily on computer simulations to build its model, the patterns it found in real patient data were consistent with what is known about how these tumors behave. The findings suggest that the evolutionary history of a tumor is encoded in its genetic makeup, waiting to be read by the right tools. For doctors, this could eventually mean that a single test could reveal not just what a cancer is, but how it began and how it might behave in the future, offering a clearer path for understanding the most aggressive forms of the disease.

Technical Summary: Inferring ecDNA Evolution from Allele Frequency Spectra

Problem Statement

Extrachromosomal DNA (ecDNA) is a driver of tumor aggressiveness, oncogene amplification, and poor patient survival. While ecDNA is known to arise at various stages of tumor development—sometimes in premalignant tissue and other times in response to therapy—the precise chronology of its emergence and the strength of its selective advantage remain unknown. Current genomic sequencing methods cannot determine when an ecDNA arose or quantify its evolutionary parameters (selection strength, mutation rate, emergence time, and clonal dominance) from a single bulk tumor biopsy. Existing methods for inferring tumor evolution are designed for chromosomal mutations and fail to account for the distinct allele frequency spectrum (AFS) generated by the stochastic, non-Mendelian segregation of ecDNA.

Methodology

The authors developed ecEvol, a simulation and deep learning framework designed to infer ecDNA evolutionary parameters directly from bulk whole-genome sequencing (WGS) data.

1. Forward Simulation Model

The authors implemented an efficient forward-time simulator to generate training data. Key features include:

Stochastic Segregation: Explicitly models the random segregation of ecDNA copies during mitosis, distinct from chromosomal inheritance.
Dual Mutation Tracking: Separately tracks chromosomal mutations and ecDNA-specific mutations (including "kataegis" bursts often associated with ecDNA formation).
Fitness Dynamics: Models a selection coefficient ( $s$ ) where ecDNA-bearing cells divide at a rate $1+s$ relative to founder cells.
Efficiency: Utilizes efficient data structures to scale as $\sim N \log(N)$ , allowing the simulation of populations up to 100,000 cells in minutes.
Data Generation: Generated a corpus of 1.26 million simulated tumors with ground-truth parameters ( $s$ , emergence time $g$ , mutation rates $\mu$ and $\mu_e$ ) across varying sequencing depths (30x, 50x, 100x, 200x).

2. Dual-Spectrum Input Strategy

The method leverages two distinct AFS derived from the same sample:

Chromosomal AFS (chAFS): Derived from the diploid genome, reflecting standard clonal and subclonal mutation patterns.
ecDNA AFS (ecAFS): Derived from the ecDNA-overlapping region. Due to high copy numbers ( $K$ ) and random segregation, this spectrum exhibits a unique signature: a prominent cluster of high-frequency mutations (frequency $\approx K/(K+2)$ ) and a depletion of intermediate-frequency variants.

3. Deep Learning Architecture

The inference pipeline consists of two models trained on the simulated spectra:

ecEvol-CNN: A 1D convolutional neural network that estimates the descendant fraction ( $\phi$ ), representing the proportion of tumor cells descended from the ecDNA-founding ancestor.
ecEvol-CA: A cross-attention multi-task model that jointly infers selection strength ( $s$ ), mutation rate ( $\mu$ ), and emergence time ( $g$ ). This model uses bidirectional cross-attention to learn dependencies between the chAFS and ecAFS, allowing it to detect relative distortions between the two spectra.

4. Robustness to Purity Misestimation

To address the challenge of tumor purity misestimation in real clinical data, the authors implemented purity-aware data augmentation. During training, simulated spectra were perturbed by introducing random errors in purity estimation. This ensures the model learns to recover evolutionary signals even when the input spectra are distorted by imperfect purity correction.

Key Results

1. Model Performance and Validation

Accuracy: ecEvol outperformed Approximate Bayesian Computation (ABC) baselines across all parameters. At 100x sequencing depth, it achieved $R^2 \approx 0.53$ for selection strength, $R^2 \approx 0.97$ for mutation rate, and 80% accuracy for emergence time (within one generation).
Robustness: Models trained with purity augmentation maintained high performance under purity misestimation, whereas models trained on clean data failed significantly.
Ablation Studies: Confirmed that the dual-spectrum input and cross-attention mechanism were critical for performance. Removing purity augmentation caused a collapse in selection recovery under realistic noise conditions.

2. Application to Primary Tumors (PCAWG/TCGA)

Applied to 288 treatment-naive primary tumors:

Early Emergence: The median estimated emergence time was $g = 3.67$ doublings, indicating ecDNA typically arises early in tumor evolution.
Strong Selection: The median selection coefficient was $s = 1.91$ , with 85% of samples showing strong positive selection ( $s > 1.5$ ).
Oncogene Impact: ecDNAs carrying known oncogenes emerged significantly earlier ( $g=3.32$ vs $5.23$) and reached higher clonal dominance ( $\phi=0.96$ vs $0.65$) compared to those without known oncogenes.
Tissue Specificity: Emergence times varied by cancer type, with CNS tumors and sarcomas showing the earliest emergence, consistent with their aggressive biology.

3. Application to Advanced/Treated Tumors (Hartwig Medical Foundation)

Applied to 913 samples from the Hartwig Medical Foundation (HMF) cohort (mostly pretreated or advanced/metastatic):

Later Emergence: The median emergence time was significantly later ( $g = 7.02$ ) compared to primary tumors ( $g = 3.67$ ).
Higher Fitness, Lower Dominance: Despite higher inferred selection coefficients ( $s=2.15$ ), the descendant fraction was lower ( $\phi=0.76$ ) than in primary tumors, suggesting new ecDNA arises later in the disease course but has not yet fully colonized the population.
Longitudinal Consistency: In paired longitudinal samples, emergence time and selection strength remained stable, while the descendant fraction increased over time, supporting the continued expansion of ecDNA-bearing lineages.

4. Clinical Correlation

Survival: In the PCAWG cohort, patients whose ecDNA emerged earliest (lowest $g$ tertile) had significantly worse overall survival. Composite indices combining emergence time, selection, and dominance further stratified survival risk.

Significance and Claims

The paper claims to provide the first quantitative framework for reconstructing the evolutionary history of ecDNA from a single tumor biopsy. By demonstrating that allele-frequency spectra encode the timing and strength of ecDNA emergence, the authors establish that:

Chronology is Inferrable: It is possible to determine when ecDNA arose and how strongly it was selected, even without longitudinal data.
Distinct Evolutionary Dynamics: ecDNA evolution follows a distinct trajectory compared to chromosomal subclones, characterized by early emergence and strong selection in primary tumors, and later emergence in advanced, treated disease.
Prognostic Value: The timing of ecDNA emergence is a significant predictor of patient survival, with earlier emergence correlating with poorer outcomes.
Methodological Advance: The ecEvol framework, utilizing dual-spectrum inputs and cross-attention deep learning, offers a robust tool for characterizing tumor evolution in the presence of complex copy-number alterations.

The authors note that their model assumes a single dominant ecDNA event and constant selection, serving as an "effective" measure of these parameters. They acknowledge that future work may refine fitness functions to account for copy-number-dependent selection and multiple competing ecDNA populations.