When a tumor or a healthy tissue is sequenced in bulk, millions of cells are ground together and their DNA is read as a single average. This averaging hides which cells carry which mutations. It also hides something almost no one measures: when each group of related cells (a subclone) split off from its ancestors. Biologists who date species splits have a standard tool for this question, the molecular clock. A molecular clock turns a count of accumulated mutations into elapsed time, and the "relaxed" version lets the mutation rate differ from branch to branch of the family tree. Somatic cells keep a clock of the same kind, because their mutations pile up roughly with age. Applying such a clock to a bulk tumor sample breaks down for a specific reason. Mutations held by only a minority of cells fall below the detection limit, so the branches leading to small subclones look shorter than they are, and those subclones appear younger.
We propose a relaxed somatic clock whose likelihood models the masking process directly. Branch mutation counts are treated as censored observations, with each branch's detection probability set by its clone's cell fraction, the tumor purity, the copy number, and the sequencing depth. Bulk data and a small number of noisy single cells enter one model, and each does a different job. The bulk sample measures the long shared trunk of the tree precisely. As few as 10 to 25 single cells reveal the branch-to-branch rate differences that bulk averaging washes out. The output is a calibrated timeline of subclone origins, in years since conception and with credible intervals. Current somatic tree methods do not produce this: they return either a branching order or a yes/no test of clock-like evolution.
Core Claim
Our central claim has two parts. First, dating subclones from bulk or mixed samples with a standard clock gives estimates that are biased in a predictable direction: minor subclones come out too young, and the bias grows as purity-adjusted depth falls. Second, an explicit detection-censoring term in the clock likelihood removes most of that bias. Pairing that term with a handful of single cells also recovers the selection-driven rate shifts that a strict clock misses.
The evidence for the gap comes from two papers. Single-cell phylogenies reveal deviations from clock-like, neutral evolution in cancer and healthy tissues builds Poisson-weighted single-cell trees and tests whether root-to-tip mutation counts are equal. It rejects a strict clock in 10 of 13 cancer and 5 of 9 healthy datasets, and it ties the rejections to driver mutations on internal branches. That paper tells us that rate heterogeneity is real and that bulk neutrality tests cannot see it. It stops at rejection and estimates no times. Integrative inference of subclonal tumour evolution from single-cell and bulk sequencing data (B-SCITE) shows that bulk allele frequencies give the "vertical" ordering of mutations while single cells give the "horizontal" separation of branches. It reaches good trees with about 25 cells. Its output is a mutation tree without branch durations, and its search is MCMC over tree space.
We reuse B-SCITE's division of labor with a new reading: bulk calibrates time, and single cells calibrate rate. The constraint comes from Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction. That benchmark of 31 bulk reconstruction methods finds that accuracy degrades below roughly 32 reads per chromosome copy (NRPCC), and that algorithm choice explains more performance variance than tumor biology. Any dating method that inherits a bulk tree uncritically will inherit that fragility. Our detection model is built around exactly this depth threshold.
Research Program
The model works on a clone tree with infinitely many possible mutation sites. We keep that assumption to stay comparable with B-SCITE and CTC-SCITE, and we handle copy-number changes through per-segment copy states as bulk callers do. The root is fixed at conception (age 0), and each sample is tip-dated at the patient's age when it was taken. This is the same use of sampling timestamps that pathogen studies such as Long-term evolution and transmission dynamics of swine influenza A virus rely on, and longitudinal biopsies add further tip dates. The baseline mutation rate per year gets an informative prior estimated from matched normal tissue. This is a form of internal calibration in the spirit of Calibrating the Human Mutation Rate via Ancestral Recombination Density in Diploid Genomes, but it uses an age-linked rate instead of recombination density.
For branch $b$ with duration $\Delta t_b$ years, the number of mutations we actually detect on it is modeled as
$$n_b \sim \text{Poisson}\big(\mu \, r_b \, \Delta t_b \, \bar d_b\big),$$
with each symbol defined as follows. $\mu$ is the baseline mutation rate per genome per year. $r_b$ is a branch-specific rate multiplier drawn from a lognormal prior, which is the "relaxed" part of the clock. The prior adds a mixture component that allows a rate jump on branches carrying a candidate driver mutation. $\bar d_b$ is the probability that a mutation arising on that branch is detected in the data. For a bulk sample,
$$\bar d_b = \mathbb{E}_{D}\Big[\Pr\big(\text{Binomial}(D,\ \rho f_b m / c) \ge k\big)\Big],$$
with these symbols. $D$ is the read depth at the site, averaged over the sample's depth distribution. $\rho$ is tumor purity. $f_b$ is the fraction of tumor cells descending from branch $b$. $m$ is the number of mutated copies. $c$ is the local total copy number. $k$ is the caller's minimum count of reads supporting the variant.
For the single cells, detection is multiplied by one minus the allelic dropout rate, following the noise model of B-SCITE. Each mutation is detected if any modality catches it, so the bulk and single-cell terms combine through a union. Mutations acquired after a subclone's last common ancestor form a low-frequency tail. Their expected detected count follows from the same formula integrated over the fractions they reach under exponential growth, so the tail adds information about the subclone's age instead of acting as noise.
Inference alternates between clone fractions and the tree, which are initialized from B-SCITE and resampled, and the continuous parameters ($\mu$, $r_b$, $\Delta t_b$). The continuous parameters are sampled by Hamiltonian Monte Carlo, because once the tree is fixed the likelihood is smooth in them. Uncertainty in tree and fractions therefore carries through into the dates. The program has three parts. We build the model and a timed simulator. We derive and check the predicted direction and size of the dating bias as a function of NRPCC and clone fraction. We then apply the model to real hybrid datasets.
Baselines
The naive reference point is a strict clock applied to the bulk clone tree. Each node's age is set to the patient's age times the fraction of the deepest root-to-tip mutation count accumulated by that node, with no rate variation and no detection correction. This is the reading a practitioner would make by eye, and we expect it to place minor subclones systematically too late.
The strongest existing competitor is a two-stage pipeline. A clone tree and fractions are inferred first, with B-SCITE for hybrid data and the best per-task bulk method from SMC-Het (PhylogicNDT) for bulk-only data. A standard lognormal relaxed clock with the same tip dates and rate prior is then fitted to the raw branch mutation counts. This competitor is strong because it already uses the best available trees and a full relaxed clock. Our method differs from it only in the detection term and in carrying uncertainty through jointly. Both systems get identical reads, the same mutation calls, the same priors, and matched MCMC budgets in effective samples per parameter.
No reported result found on Lacuna for node-age error of any subclonal dating method, so no published number serves as the target. The closest reported figures concern tree inference, not dates. B-SCITE is reported to outperform SCITE, OncoNEM and ddClone with about 25 cells, but the file gives no numbers. The PT test reaches over 90% power to reject the clock with 10 cells. We will count it as beating the two-stage pipeline if median relative error in node age falls by at least a third below NRPCC 32, and if 95% credible-interval coverage stays near nominal there while the pipeline's coverage collapses. At high depth the two methods should tie.
Four ablations isolate the ingredients: removing $\bar d_b$ (set to 1), forcing $r_b=1$ (strict clock), dropping the single cells, and dropping the bulk sample. A single-cell-only arm with a relaxed clock on the PT-style Poisson tree serves as a further competitor.
Why not just…?
Why not just run the two-stage pipeline with a standard relaxed-clock tool? This is the real threat, and the experiments are built to settle it rather than argue around it. In principle a relaxed clock can absorb censoring by inflating the rates on short-looking branches. We predict the absorption fails in a specific way: the clock spreads the shortfall as spurious rate decreases on minor-clone branches and reports confident, wrong dates. Our design tests this prediction directly. If the two-stage pipeline matches our error and coverage on the pseudo-bulk experiment below NRPCC 32, the detection term adds nothing. We would then report that result and reframe the contribution as the bias analysis alone.
Why not just count only clock-like mutation types and use a strict clock? Restricting to mutation classes that accumulate steadily with age reduces rate variation that comes from mutational processes. It cannot fix the issue that the PT study identifies, which is lineage-specific acceleration from selection and changes in division rate on driver-carrying branches. It also does nothing about detection censoring, and it discards most of the mutations, which worsens the low-depth problem. We include it as a baseline variant on the same simulator, where we control both the class-specific rates and the selection, to show which source of error dominates.
Why not just date from single cells alone? High-throughput single-cell DNA data sits at ultra-low coverage. SECEDO reports that the earlier method SBMClone reached an adjusted Rand index of about 0 at 0.03× coverage, against SECEDO's roughly 0.6. At that depth most mutations on long shared branches go unobserved in any given cell, so trunk lengths, and with them absolute dates, are poorly determined. Single cells are also typically few, so the most recent subclones may be absent. Our single-cell-only ablation measures how much the bulk sample contributes to trunk calibration.
Why not just extend a neural generative tree model to output branch lengths? Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes shows that diffusion-based simulation inference beats likelihood tools such as IQ-Tree when the likelihood is intractable, as with insertions and deletions. It explicitly leaves branch lengths to future work and is tested only up to 65 taxa. Our observation model remains tractable under infinite sites, so a likelihood method is both more data-efficient and easier to audit. Amortized simulation inference becomes the right tool only when infinite sites is dropped, for example with recurrent mutation or loss of heterozygosity. We list that as a follow-up rather than the core.
Proposed Experiments
The decisive experiment uses pseudo-bulk masking of real single-cell data. For each of the 22 datasets in the PT-test cohort, we first date the full single-cell tree with the relaxed clock, using patient ages. We treat these dates as the reference. We then hide the cell-level structure by pooling reads from random subsets of cells into synthetic bulk samples. The sweep covers NRPCC from about 8 to 64, purity from 0.3 to 0.9, and 0, 10 or 25 retained single cells. This is the same logic by which the Kassandra profiles used in POPPY build bulk from sorted cells. We measure relative node-age error, interval coverage, and the sign of the bias for each method. The core claim predicts that the uncorrected methods are biased toward younger minor subclones, more strongly as NRPCC falls, and that our method's bias stays flat.
The second experiment is a controlled simulation. We re-simulate the structural settings of the SMC-Het BAMSurgeon tumors, including purity, copy-number events and subclonal whole-genome doubling, but attach known node ages and branch rates, with rate jumps on driver branches. Here we score date recovery and the precision and recall of rate-shift branches against the simulated drivers. Those drivers are the quantity the PT test detects but cannot locate.
The third experiment covers real hybrid data. We apply the model to B-SCITE's leukemia, breast and colorectal cases, and check two things: whether every inferred subclone origin falls inside the patient's lifetime, and whether childhood leukemia shows the compressed timelines expected from its short history. As an application, we apply the bulk-mode model to the oligoclonal circulating-tumor-cell clusters from Phylogenetic inference reveals clonal heterogeneity in circulating tumor cell clusters. These clusters cannot be split into single cells, which makes them an extreme case of masking. We ask whether divergent clones within a cluster split long before dissemination or shortly before it. The comparison between fast- and slow-growing xenograft clusters provides a built-in contrast.
The project fails cleanly if the two-stage pipeline matches us on the pseudo-bulk sweep, or if our credible intervals undercover on real data where the patient-age bounds are known.
References
- Single-cell phylogenies reveal deviations from clock-like, neutral evolution in cancer and healthy tissues
- Integrative inference of subclonal tumour evolution from single-cell and bulk sequencing data
- Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction
- Long-term evolution and transmission dynamics of swine influenza A virus
- Calibrating the Human Mutation Rate via Ancestral Recombination Density in Diploid Genomes
- SECEDO: SNV-based subclone detection using ultra-low coverage single-cell DNA sequencing
- Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Phylogenetic inference reveals clonal heterogeneity in circulating tumor cell clusters