Chronotherapy is the practice of timing a drug to the body's 24-hour rhythms. Its logic for anti-cancer drugs depends on one assumption that has almost never been checked in humans: that tumor cells still divide more at some hours than others. You cannot biopsy a patient's tumor every two hours. Hospital archives, however, already hold tens of thousands of tumor specimens, each removed at a recorded clock time. Each one is a snapshot of how many cells were dividing at that moment. We propose to pool these snapshots across patients and fit a daily rhythm to them, a "population time series" of human tumor cell division. The observer problem has blocked this so far. Counting dividing cells (mitotic figures) is one of the least reproducible tasks in pathology, and the variation between pathologists would swamp any daily signal. Automated detectors of dividing cells now make the counts objective enough to try.
The payoff runs in two directions. For chronotherapy, the method gives a direct tissue-level answer to whether division in a given human tumor type is still gated by time of day. Current personalization approaches cannot answer this, because they read the patient's rhythm from wrist activity, which reflects the host rather than the tumor. For diagnostic grading, if a daily rhythm exists, then part of what looks like disagreement between pathologists, or between repeat biopsies, is actually the time of day the tissue was taken. Correcting for it should make mitotic grades both more reproducible and more predictive of long-term outcome. We will judge that predictive gain only against long-term outcomes, because short-window prognostic studies are exactly where spurious grading gains hide.
Core Claim
Routine pathology archives, scored by a detector that separates the stages of cell division, contain enough temporal information to estimate the amplitude and peak hour of daily mitotic rhythms in human tumors. Two results would make the work new: such rhythms are detectable, or they are credibly absent, above both rater noise and handling artifacts; and removing them improves the long-term prognostic value of mitotic grading. No work in our file measures human tumor division rhythms this way, and none of the pathology benchmarks record acquisition time.
The argument rests on three supporting results. The age-structured cell-cycle models of Circadian rhythm and cell population growth show that time-periodic control of division can beat constant control. That is theory, and it assumes the tumor's division machinery still responds to periodic signals. Our study tests that premise empirically rather than building on it. The micro-trench single-cell tracking in A single-cell micro-trench platform for automatic monitoring of cell division and apoptosis after chemotherapeutic drug administration shows that vincristine's potency depends strongly on cell-cycle position: the dose that inhibits half the cells (IC50) is 9 nM in chemically synchronized cells versus 652 nM in unsynchronized ones. That is cell-cycle synchrony in a leukemia line with no body clock, not circadian timing. It therefore motivates why tumor synchrony would matter for dosing, but it does not show that human tumors are synchronized. Finally, Robust Individual Circadian Parameter Estimation for Biosignal-based Personalisation of Cancer Chronotherapy personalizes timing from four days of wrist activity using an asymmetric two-half-wave rhythm model, an "extended Cosinor." That model fits cancer patients only modestly (adjusted $R^2$ of 0.198 versus 0.173 for the standard fit). We reuse its asymmetric rhythm shape as a method. Our target signal is the tumor rather than the wrist.
Research Program
The first component is an objective division counter. We start from the two-stage pipeline of A Subphase-Labeled Mitotic Dataset for AI-powered Cell Division Analysis. That pipeline has a Mask R-CNN detector with a ConvNeXt backbone, followed by a hierarchical classifier. The classifier first rejects look-alikes such as apoptotic bodies, then assigns prophase through telophase or marks the figure as atypical. Stage labels matter here for a specific reason. The time spent in each stage is short and roughly fixed, so the mix of stages in a specimen carries information about how synchronized the dividing population is, beyond the raw count. A tumor whose divisions are clock-gated should show a stage mix that shifts with the hour of removal. We will train on stMIDOG++ and check transfer on LUNG-MITO before deployment. We then apply the detector to a timestamped institutional archive, which must be assembled and is not in our file. That archive will cover colorectal, breast and lung resections plus endoscopic Barrett's biopsies. For each case it records the procedure clock time, the time from removal to fixation (ischemia time), the procedure type, patient covariates, and linked cancer-registry outcomes with at least five years of follow-up.
The second component is a population rhythm model. For specimen $i$ removed at clock time $t_i$ (hours), let $m_i$ be the detector's mitotic density per square millimetre of tumor. We fit
$$\log m_i = x_i^\top \beta + g(t_i;\, A, \phi, \omega_1, \omega_2) + u_{s(i)} + \varepsilon_i ,$$
where $x_i$ collects stage, grade, ischemia time and procedure type with coefficients $\beta$. The term $g$ is the asymmetric extended-Cosinor curve with amplitude $A$, peak hour $\phi$, and separate rising and falling frequencies $\omega_1, \omega_2$. The term $u_{s(i)}$ is a random effect for the surgeon, scanner and staining batch, and $\varepsilon_i$ is residual noise. A parallel multinomial model fits the stage proportions against the same time term. The quantity of interest is $A$ with its confidence interval, estimated per tumor type. A credible null is as valuable as a positive finding, because it would tell chronotherapy trials not to expect tumor-side gating in that tissue.
The third component is the observer-variability lens. On a subset of about 500 slides, three pathologists will count mitoses independently. Following the logic of More is more: leveraging multi-rater information for whole slide images grading via virtual expert panel, we model individual raters instead of collapsing them to a consensus. That paper showed that aggregating per-rater models improves both F1 and calibration over a consensus-trained model on Barrett's grading. We use the same idea to split the variance of each slide's count into three parts: variance between raters, daily variance from $g$, and variance between specimens. This decomposition tells us directly whether a daily rhythm could ever have been seen in manual counts. The blinded RECIST double-read review RECIST 1.1 assessments variability: a systematic pictorial review of blinded double reads documents a similar pattern: small, unmodeled measurement variation there flips response categories. We ask whether time of acquisition plays that role for mitotic grade.
The fourth component is long-horizon validation. We compare raw mitotic density with the time-corrected version, $\log m_i - \hat g(t_i)$, as predictors of recurrence and cancer-specific survival. Because Direction 2 of our file shows repeatedly that short windows mislead, we deliberately report results at one, three and five-plus years. A correction that helps only at one year will be treated as unproven, following the warnings in A prospective evaluation of cardiovascular magnetic resonance measures of dyssynchrony in the prediction of response to cardiac resynchronization therapy and Efficacy and safety of geptanolimab (GB226) for relapsed or refractory peripheral T cell lymphoma: an open-label phase 2 study (Gxplore-002) that early readouts mask late behavior.
Baselines
The naive reference is a flat model with $A = 0$ that keeps all covariates and random effects. Any claimed rhythm must beat it on held-out likelihood and survive a permutation test that shuffles clock times within surgeon and procedure-type strata. The stratification is essential, because unstratified shuffling would break real scheduling confounds and overstate significance.
For detection, the strongest existing method is the refined ConvNeXt pipeline of Iván et al. It reports an overall F1 of 0.794 and a class-wise F1 of 0.600 on stMIDOG++, and an overall F1 of 0.730 on LUNG-MITO, against 0.683 and 0.570 for a ResNet-50 Mask R-CNN. We will retrain both on identical splits with the same augmentation and hyperparameter search budget. We adopt our detector only if it matches or exceeds these published numbers. The goal is a counter at least as good as the published one, not a new detector.
The ablation that isolates our key ingredient replaces stage-aware counts with a single-class mitosis count from the same detector. Stage information "earns its place" if the stage-proportion model detects rhythm at a lower false-positive rate under permutation, or narrows the confidence interval on $\phi$. A second ablation replaces the asymmetric curve with a standard symmetric Cosinor, as a test of whether Schott et al.'s asymmetry transfers from wrist to tumor.
For prognosis, the baseline is the raw automated mitotic index in a Cox model with standard clinicopathologic covariates, compared on concordance index at each follow-up horizon. Beating it means a concordance gain whose bootstrap interval excludes zero at five years. A gain only at one year does not count.
Why not just…?
Why not just have pathologists count mitoses on timestamped slides? Manual counting is the source of the problem. If the variance between raters is comparable to the daily amplitude, a manual study cannot see a rhythm, and it cannot scale to the tens of thousands of slides needed to estimate an amplitude across time. The three-rater subset measures this directly. If it shows that manual counts alone recover $A$, the detector becomes a convenience rather than a necessity, and we will say so.
Why not just use Ki67 staining, which is less subjective? Ki67 marks cells across most of the cycle, which lasts many hours (about 20 hours in the MOLM-13 line tracked by Chatzopoulou et al.). It therefore averages over the day. A mitotic figure is a snapshot of roughly an hour. We predict Ki67 will show a much smaller daily amplitude than mitotic density on the same cases. That comparison is itself a check that any rhythm we find reflects division timing rather than a handling artifact that would affect both markers.
Why not just time chemotherapy from wrist activity, as current personalization does? Wrist activity measures the host's rhythm. The closing limitation in Schott et al. is that very sick patients have disrupted rhythms, and nothing in that work shows the tumor follows the wrist. Our tissue estimate can reveal whether tumor gating is absent, or shifted relative to typical host activity peaks. A chronotherapy scheme timed to the wrist would miss either case.
Why not just run a prospective study that randomizes biopsy time? That is the gold standard, and it is the real threat to a purely retrospective design. We treat the archive study as the way to obtain effect sizes and choose tumor types. We then propose a small prospective check in Barrett's surveillance endoscopy, where scheduling time is already arbitrary and ischemia time is short, to confirm $\hat\phi$ on independently timed biopsies.
Decisive Checks
The central threat is confounding. Morning and afternoon cases differ in patient mix and urgency, and delayed fixation changes how mitotic figures look. Three checks carry the weight. First, the rhythm must persist after adjusting for ischemia time, and it must appear within endoscopic biopsies, where fixation is near-immediate. Second, a placebo clock, the hour at which the slide was scanned, must show no effect. Third, the peak hour must be consistent across independent hospitals with different operating schedules. A rhythm that tracks each hospital's schedule rather than clock time is an artifact.
We have found no reported result for human tumor circadian mitotic amplitude, so we make no numerical prediction. Instead, we will run a power analysis from the rater-subset variances to state the smallest amplitude the archive can detect. The study succeeds if it either estimates a replicated nonzero $A$ that improves five-year prognostic concordance, or rules out amplitudes above that detectable floor. Either outcome informs whether tumor-directed chronotherapy is worth trialing in that tissue.
References
- Circadian rhythm and cell population growth
- A single-cell micro-trench platform for automatic monitoring of cell division and apoptosis after chemotherapeutic drug administration
- Robust Individual Circadian Parameter Estimation for Biosignal-based Personalisation of Cancer Chronotherapy
- A Subphase-Labeled Mitotic Dataset for AI-powered Cell Division Analysis
- More is more: leveraging multi-rater information for whole slide images grading via virtual expert panel
- RECIST 1.1 assessments variability: a systematic pictorial review of blinded double reads
- A prospective evaluation of cardiovascular magnetic resonance measures of dyssynchrony in the prediction of response to cardiac resynchronization therapy
- Efficacy and safety of geptanolimab (GB226) for relapsed or refractory peripheral T cell lymphoma: an open-label phase 2 study (Gxplore-002)