The Family Trees That Tell Tumor Secrets
The Hidden Signature Inside a Tumor's Family Tree
Every tumor has a family tree. Not the kind tracked at reunions, but something far more consequential: a phylogenetic record written in mutations, passed down through billions of cell divisions. And just as you can learn about your own ancestry from genetic tests, scientists can now reconstruct the evolutionary history of a tumor by sequencing its cells and tracing their genetic relationships backward through time.
But here's the puzzle that has quietly troubled cancer researchers: two tumors of identical size can tell radically different stories. One might have grown explosively, doubling every few weeks before hitting some wall. Another might have crept along slowly for years, struggling against its own cramped microenvironment. Both could end up with the same number of cells today—but their internal genealogies would be profoundly different. The shape of those trees, it turns out, carries a fingerprint of how they grew.
A new paper by Wang, Weng, Foo, and Wang at the Chinese University of Hong Kong and the University of Minnesota offers the most systematic analysis yet of what these genealogical signatures actually look like—and the findings are striking. Tumor cells that expanded rapidly, unconstrained by their environment, show a characteristic pattern of branching times in their reconstructed family trees. Tumors that grew more slowly, bumping up against limiting factors like space, nutrients, or blood supply, leave a distinctly different signature. And crucially, as constrained tumors mature, their branching-time patterns shift in predictable ways, moving through distinct phases like a tumor aging in real time (Wang et al., 2026).
This isn't merely a theoretical curiosity. Understanding a tumor's growth history could reshape how we assess its prognosis, predict its future behavior, and choose treatments. A tumor that grew explosively before plateauing carries a different evolutionary legacy than one that crept along steadily—and that difference might matter for whether it metastasizes, develops resistance, or responds to therapy.
The Science: Reading Family Trees Written in Mutations
To understand this work, you need to grasp what a phylogenetic tree actually is—and why reconstructing one from tumor cells is both possible and profoundly useful.
When a tumor cell divides, it copies its DNA. Most of the time, that copy is perfect. But occasionally, a random mutation slips in—a single letter changed, a small insertion, a deletion. When that daughter cell divides again, it passes those mutations to its descendants. Over time, this creates a nested hierarchy of genetic differences: cells that share a recent common ancestor will have nearly identical mutation profiles, while cells whose shared ancestor lies far in the past will have accumulated more differences between them.
Modern sequencing technologies can detect these somatic mutations—the changes that happen in the body after conception, as opposed to inherited variants. By comparing mutations across many cells sampled from a tumor, bioinformaticians can infer the branching structure of the cell population: which cells are most closely related, which lineages split off earlier, which branches died out. The result is a rooted tree, with the original tumor-founding cell at the bottom and the sampled cells at the tips (Gerstung et al., 2020).
Each internal node of this tree represents a division event in the tumor's past—one cell that divided into two lineages, both of which left descendants in the final sample. The timing of these branching events, and how they're distributed across the tumor's history, is what Wang and colleagues set out to analyze.
The key quantity in their paper is the internal branching-time distribution: a summary of when retained division events occurred, scaled relative to the total time from tumor initiation to sampling. If most internal nodes cluster near the root of the tree, that tells you something different than if they cluster near the tips—or if they're spread across the full timeline.
To study how different growth regimes imprint on this distribution, Wang and colleagues used mathematical models of tumor cell populations. Specifically, they analyzed continuous-time birth-death processes, which simulate tumors as populations where each cell independently faces random chances to divide or die at any moment.
They compared two fundamental growth regimes:
The first is exponential growth, where cells divide at a constant rate regardless of how many cells exist. This represents unconstrained expansion—the kind of rapid doubling that characterizes early tumor development, before the tumor runs into space limits, nutrient shortages, or hostile microenvironmental conditions. In mathematical terms, the per-cell birth rate λ(n) and death rate μ(n) stay constant: λ(n) ≡ λ₀ and μ(n) ≡ μ₀. The net growth rate r₀ = λ₀ - μ₀ drives exponential expansion, and the expected population size follows E[N(t)] = N₀e^(r₀t).
The second is logistic growth, which incorporates density-dependent constraints. As the population grows, resources become scarce and competition intensifies. The net growth rate declines as population size approaches a carrying capacity κ—some theoretical maximum that the microenvironment can support. Wang and colleagues analyzed two variants of logistic growth: one where density dependence acts primarily through increasing cell death (the microenvironment becomes more hostile as cells crowd together), and one where it acts through suppressing cell birth (proliferation is squeezed by space and nutrient limitations).
Both logistic models share the same deterministic population trajectory, but they assign density dependence differently—through mortality or through fertility—and this turns out to affect the genealogical signatures they produce.
The authors then simulated tumor genealogies under both regimes, sampled cells at various observation times, pruned the trees to retain only lineages represented in the sample, and analyzed the distribution of internal branching times. They supplemented these simulations with coalescent point process theory, a mathematical framework that links forward-time population dynamics to the backward-time structure of sampled genealogies.
What They Found: Two Worlds of Branching Times
Exponential Growth: A Single Window Near the Root
Under exponential growth, Wang and colleagues found that the branching-time distribution is unimodal—it has one peak—and that peak sits near the root of the tree, representing early branching events close to tumor initiation.
This finding emerged both from mathematical analysis and from extensive simulations. Under a coalescent point process approximation appropriate for large, growing populations, each internal branching time τ can be expressed as:
where r₀ is the net growth rate, m is the sample size, W_S is a random variable following an exponential distribution, and the U_{S,i} terms are independent logistic variables describing the spread around the characteristic scale.
The authors proved that under this approximation, the marginal density of each τ_{S,i} is log-concave, which implies unimodality. In simpler terms: there's one characteristic time window for retained branching events, and internal nodes cluster around it rather than spreading across the full timeline.
The intuition is elegant. During rapid exponential expansion, the population grows so quickly that early divisions have enormous advantages: they give their lineages more time to reproduce and leave descendants in the sample. Every division is a lottery, and early tickets are worth more because they buy more rounds of reproduction. When the researchers pruned the full genealogy to just the sampled lineages, early branching events were far more likely to have both daughter lineages survive to the sample. Recent divisions, by contrast, created daughter lineages that represented only a tiny fraction of the final population—and when those lineages didn't happen to contribute sampled cells, their division events disappeared from the reconstructed tree.
In simulations with varying net growth rates, the median mode of the branching-time distribution shifted predictably. For a net growth rate of 0.25, the mode sat at 0.58 on the scaled time axis (where 0 is tumor initiation and 1 is sampling time). Slower growth pushed the mode later: at r₀ = 0.15, it moved to 0.646, and at the slowest growth rate tested (r₀ = 0.05), it reached 0.726. Slower growth means the characteristic time scale lengthens, so retained branching events occur later relative to the total observation window. But crucially, across all parameter settings, the distribution remained unimodal—a single peak near the rootward side of the tree.
Figure 1 from the paper shows branching-time distributions under constant-rate exponential growth. The solid curve represents the median kernel density estimate across 20 independent simulations, with darker bands showing the 25th-75th percentile range and lighter bands showing the 5th-95th range. Across all three growth rates tested, the distributions are unimodal with a single peak near the root of the tree.
Logistic Growth: Three Phases of Genealogical Transformation
The story under logistic growth is far more intricate—and far more interesting.
When tumor growth becomes density-dependent, constrained by carrying capacity, the branching-time distribution doesn't stay single-peaked. Instead, Wang and colleagues identified a systematic transition through three distinct phases as the tumor ages:
Phase 1: Expansion-Dominated. Early in tumor development, when the population is still far below carrying capacity, logistic growth behaves almost identically to exponential growth. The net growth rate remains positive and substantial, the population expands rapidly, and retained branching events cluster near the root. The branching-time distribution is unimodal and rootward-shifted, just like the exponential case.
Phase 2: Early-Recent Bimodal. As time passes and the population approaches carrying capacity, something remarkable happens: the single peak splits into two. One mode remains near the root (representing old branching events from the expansion phase), but a second mode emerges near the tips of the tree, reflecting more recent divisions. The distribution becomes bimodal—two peaks where there was one.
This bimodality arises because density-dependent growth creates a temporally heterogeneous genealogical environment. During the expansion phase, early branching events are retained in the sampled tree just as in exponential growth. But as growth slows and the population stabilizes near carrying capacity, the dynamics change. The population stops expanding, so lineage extinction becomes more likely. Recent divisions create daughter lineages that compete in a crowded, stable environment rather than a growing one. Some of these recent branches survive in the sample—and when they do, their branching events appear near the tips of the reconstructed tree.
Phase 3: Recent-Dominated. Given enough time since tumor initiation, the distribution completes its transformation. The early mode fades, and retained branching events concentrate in a single recent window. The distribution returns to unimodality, but now the peak sits near the tips of the tree rather than near the root.
Wang and colleagues characterized this phase transition quantitatively, tracking how the branching-time intensity—the rate at which retained branching events occur at different times—evolves as a logistic tumor matures.
Figure 2 illustrates this temporal evolution under density-dependent death (Model 1). The columns show four different observation times, increasing from left to right (runtime multipliers M = 1, 100, 500, and 1000, where larger M means more time has passed since the tumor initiated). The upper row shows sampled genealogies at each time point; the lower row shows the corresponding branching-time distributions. You can see the progression from unimodal (M=1) through bimodal (M=100, 500) to recent-dominated (M=1000).
The two logistic models—density-dependent death versus density-dependent proliferation suppression—produced similar overall patterns but with notable differences in detail. Under density-dependent death, the microenvironment becomes more lethal as the population grows, accelerating extinction of marginal lineages. Under density-dependent proliferation suppression, the environment becomes more static rather than hostile, producing somewhat different branching-time dynamics. Both models, however, exhibited the three-phase transition: expansion-dominated → bimodal → recent-dominated.
The researchers also examined how population-level statistics evolved. The mean population size followed the classic logistic trajectory: exponential-like growth initially, then a slowing rate of increase, then a plateau near carrying capacity. Simultaneously, the genealogical structure—encoded in the branching-time distribution—transformed in lockstep. The birth rate (mean number of births per unit time, averaged across the population) peaked during the expansion phase and declined as the population approached carrying capacity, consistent with the intuition that rapid population growth creates more branching events in the genealogy.
Why This Changes Things: Reading Tumor History from Samples
The implications of this work extend far beyond mathematical biology.
Today, when oncologists sequence a tumor, they reconstruct its phylogenetic tree and look for patterns that might guide treatment: driver mutations, subclonal diversity, evolutionary trajectories. The dominant frameworks focus on tree topologies—whether tumors are shaped like lilies (mostly linear evolution) or bushes (branching diversification)—and on the accumulation of mutations over time (Lewinsohn et al., 2023).
What Wang and colleagues add is a new lens: the temporal distribution of internal branching events, and what that distribution can reveal about the tumor's growth history. This matters because the past growth trajectory of a tumor encodes information that isn't directly visible from its current state.
Consider two tumors of similar size and similar subclonal composition. One grew explosively for six months, then plateaued. The other grew slowly and steadily for three years. Both might show similar patterns of genetic diversity measured at a single time point. But their genealogical histories—how many cell generations they've been through, how long they've been exposed to microenvironmental pressures, how much evolutionary time has passed—could be radically different.
This difference might matter for prognosis. A tumor that's been slowly adapting to a crowded, hostile microenvironment for years has undergone a different selective regime than one that simply exploded and then stalled. The slow-grower has had more opportunities for adaptive mutations that thrive in constrained conditions. It may have already evolved resistance mechanisms. Its recent-dominated genealogical structure reflects a mature, stabilized population rather than a recently expanded one.
It might also matter for treatment. If a tumor's branching-time signature suggests it recently transitioned from expansion to plateau, that could signal that microenvironmental pressures are beginning to constrain it—and that further progression might require escaping those constraints (through metastasis, angiogenesis, or immune evasion). If the signature shows an old, stable, recent-dominated pattern, the tumor may have already navigated that transition and found ways to persist.
The researchers note that their framework applies not just to tumors but to any evolving biological population modeled by birth-death processes. Viral populations, bacterial communities, even immune cell repertoires—all could exhibit these signatures of growth-regime transitions. The conceptual move is general: rather than treating phylogenetic trees as static snapshots, it asks what their temporal structure reveals about the population dynamics that shaped them.
This connects to a broader movement in phylogenetics. Methods like Bayesian skyline plots and birth-death skyline approaches use the timing of coalescent events to reconstruct population-size trajectories over time (Drummond et al., 2005; Stadler et al., 2013). Wang and colleagues' work complements these approaches by asking a complementary question: instead of inferring population size from tree structure, what can tree structure directly tell us about growth regime, independent of detailed trajectory reconstruction?
The unimodal versus bimodal versus recent-dominated distinction provides a qualitative signature that could be read off from reconstructed trees without requiring full coalescent analysis. A clinician examining a tumor phylogeny could potentially assess whether the branching-time distribution suggests explosive early growth, recent plateau, or something in between. This isn't yet a ready-to-deploy diagnostic—more work is needed to translate these theoretical results into practical inference methods—but it establishes a conceptual framework that could eventually inform clinical interpretation.
What's Next: Open Questions and Practical Horizons
Wang and colleagues have mapped out a theoretical landscape. What remains is exploration.
From theory to practice. The biggest open question is how these signatures perform on real tumor data. The paper is primarily mathematical and simulation-based; it establishes what patterns should arise under idealized conditions. Real tumors are messier: they have spatial structure, heterogeneous microenvironments, temporally varying mutation rates, and sampling biases. Whether the clean three-phase transition survives contact with real-world complexity is an empirical question.
Inference methods. The paper characterizes branching-time distributions in population-level terms—what the distribution looks like across many trees or many samples. Clinicians, however, work with single tumors and single samples. Developing inferential methods that can distinguish exponential from logistic growth signatures from a single reconstructed tree would be a significant contribution. This might involve summary statistics of the branching-time distribution—skewness, modality tests, comparisons of rootward versus tipward clustering—that could be computed from a single phylogeny.
Model extensions. The paper focuses on two growth regimes: constant-rate exponential and density-dependent logistic. Real tumors likely experience more complex trajectories: phases of expansion, followed by bottlenecks, followed by renewed growth, punctuated by metastasis and treatment response. Extending the branching-time signature framework to multi-phase growth histories—transient expansion, selective sweeps, population bottlenecks—would increase its clinical relevance.
Spatial structure. The current model assumes well-mixed populations where every cell competes equally. Real tumors have spatial architecture: cells at the periphery have different microenvironments than cells in the core, and this spatial heterogeneity could leave additional signatures in the genealogy. Wang and colleagues acknowledge this as a limitation; incorporating space into the model is a natural next step.
Connecting to existing cancer phylogenetics. The cancer genomics community has developed sophisticated tools for reconstructing tumor phylogenies from multi-region or single-cell sequencing data (Navin et al., 2011; Gerstung et al., 2020). Integrating the branching-time signature framework with these existing methods could create a new dimension of interpretation: not just "what is the tree structure?" but "what does the tree structure imply about growth history?"
Prognostic validation. The ultimate test will be whether branching-time signatures correlate with clinical outcomes. Does a recent-dominated signature (suggesting old, stable, constrained growth) predict different prognosis than an expansion-dominated signature? Do these signatures predict differential responses to therapies that target proliferation (like chemotherapy) versus therapies that target microenvironmental dependencies (like anti-angiogenics)? Prospective studies linking genealogical signatures to clinical outcomes would test the framework's practical value.
The Deeper Significance: Time, History, and the Stories Tumors Tell
There's something profound about the fact that a tumor's family tree carries a record of its own growth history. We're accustomed to thinking of genetic information as encoding current traits—how a cell looks, how it behaves, what vulnerabilities it has. But phylogenetic trees also encode temporal information: how long ago lineages split, how many generations have passed, how the population's environment shaped its trajectory.
Wang and colleagues' work is part of a broader intellectual movement that takes time seriously in biology. The coalescent theory that underlies their analysis was itself revolutionary: it showed that the patterns of genetic variation in a population are sculpted by its demographic history, that "null models" of constant population size produce predictable genealogical structures, and that deviations from those predictions reveal demographic events in the population's past (Kingman, 1982).
The same logic, applied to tumors, suggests that phylogenetic trees aren't just ancestry diagrams—they're dynamical records. The branching-time distribution is a kind of demographic archaeology: a window into the population pressures that shaped the observed sample.
This matters beyond cancer, too. The framework developed here—branching-time signatures of growth regime in birth-death processes—applies wherever phylogenetic trees can be reconstructed from sampled individuals. Viral phylodynamics has used these ideas for decades, inferring epidemiological dynamics from viral sequences (Stadler et al., 2013). Bacterial populations, immune cell repertoires, even ancient DNA from fossils—all could be interrogated with similar methods to ask not just "what is the tree?" but "what growth regime produced this tree?"
For cancer specifically, the clinical stakes are high. Understanding a tumor's growth history could inform prognosis, guide treatment selection, and illuminate why some tumors behave aggressively while others indolently persist. The branching-time signature isn't just a mathematical curiosity—it's a potential new tool for reading the stories tumors tell about themselves.
The paper ends with a note of intellectual modesty: these results are a theoretical foundation, not a finished clinical tool. But the foundation is solid. The unimodal signature of exponential growth, the three-phase transition of logistic growth, the differences between density-dependent death and density-dependent proliferation suppression—these are now established facts about how genealogical structure reflects population dynamics.
What remains is the work of turning those facts into tools. And in the meantime, there's something quietly remarkable about the idea that the shape of a tumor's family tree encodes the story of how it grew—that time, as it passes, leaves marks not just on cells but on the records we use to reconstruct their history.
References:
Benzekry, S., et al. (2014). Classical mathematical models for description and prediction of experimental tumor growth. PLOS Computational Biology.
Cheek, D. (2022). Coalescent theory for birth-death processes. Theoretical Population Biology.
Drummond, A. J., et al. (2005). Bayesian coalescent inference of past population dynamics from molecular sequences. Molecular Biology and Evolution.
Gerlee, P. (2013). The model muddle: In search of tumor growth laws. Cancer Research.
Gerstung, M., et al. (2020). The evolutionary history of 2,658 cancers. Nature.
Johnson, V. E., et al. (2023). Coalescent-based estimators of clonal growth rates. PNAS.
Kingman, J. F. C. (1982). The coalescent. Stochastic Processes and their Applications.
Lambert, A., & Stadler, T. (2013). Birth-death models and coalescent point processes. Theoretical Population Biology.
Lewinsohn, M., et al. (2023). Phylodynamic analysis of spatial tumor evolution. Cell.
Navin, N., et al. (2011). Tumor evolution inferred by single-cell sequencing. Nature.
Noble, R., et al. (2022). Role of ecological interactions in cancer. Nature Reviews Cancer.
Slatkin, M., & Hudson, R. R. (1991). Pairwise comparisons of mitochondrial DNA sequences in stable and exponentially growing populations. Genetics.
Stadler, T., et al. (2013). Birth-death skyline plots reveal temporal patterns of diversification and extinction. Frontiers in Ecology and Evolution.
Wang, M., Weng, Z., Foo, J., & Wang, Z. (2026). Branching-time signatures of growth regime in tumor birth-death models. arXiv.