The first seed plants, from the Late Devonian roughly 365 million years ago, left a lopsided fossil record. Their tough parts survive: ovules (the structures that become seeds), the cup-like cupules that hold them, and woody stems. Their pollen organs, the male structures that shed pollen, almost never do. That imbalance matters for dating. Methods that turn DNA differences into calendar dates need fossils fixed in advance as age anchors, and in practice those anchors come from whatever organ happens to fossilize. So the age assigned to the origin of seed-plant male reproduction is really a statement about ovules.
We propose to treat the scarce pollen-organ record as a measurement problem with a known structure. The same lineage is seen through several "channels": stems, ovulate cupules, and pollen organs. Each channel has its own chance of being preserved, and a missing record in one channel can mean either "not preserved" or "not yet evolved." Single-cell biology has met this exact problem. RNA and protein are two imperfect readouts of one cell state, and a zero in an RNA count may be a capture failure rather than true silence. We borrow that modeling machinery to infer the hidden lineage history jointly from all organ channels. We then turn the result into a fossil age anchor for the male lineage that accounts for preservation bias, and measure how much it moves molecular date estimates.
Core Claim
A joint, zero-inflated model of organ-specific fossil occurrence can recover the origin time of a poorly preserved organ (pollen organs) by borrowing strength from well-preserved organs of the same lineages. The calibrations it produces will cover the true origin time more often than the two current practices. The first current practice uses the oldest pollen organ found as a hard minimum. The second inherits ovule-based minimums for all nodes. The new ingredient is not modeling preservation in itself. It is coupling organ channels through a shared lineage latent state, so that dense cupule and stem records tell us about the sparse pollen record.
Why This Gap Is Real
The paleobotanical work motivates the gap.
- Wang et al.'s description of Placotheca minuta shows a fused, synangiate pollen organ already in the Famennian. This means male organs were morphologically complex early, so the true origin of the male program may sit well before its first preserved occurrence.
- Li et al. (2025) analyze about 250 specimens of the ovulate cupule Zaijunia biloba. They place it in a 111-character, 40-taxon morphological matrix and run airflow simulations, but they explicitly lack its pollen organ and do no time calibration.
- Wang (2015) shows that stem anatomy alone (the arrangement of the central vascular cylinder) can assign Yiduxylon to seed-plant groups without reproductive organs. That is precisely the kind of abundant-channel signal we want to exploit.
The dating literature constrains the design. The a priori versus a posteriori calibration study shows, by forward simulation, that fossil calibrations must enter a Bayesian relaxed clock as priors. Here a "relaxed clock" means a model letting mutation rate vary across branches. Rescaling relative dates afterwards, as RelTime does, fails when rates shift. Our output must therefore be a prior distribution, not a post-hoc correction. Applied studies such as Jin et al.'s Juglandaceae plastome dating use four fossil anchors in BEAST and illustrate the standard practice we are improving on: well-preserved fossils taken at face value as minimum bounds.
The method comes from single-cell multi-omics. scChord separates technical dropout from biological variation using a zero-inflated negative binomial decoder. A zero-inflated negative binomial is a count distribution with an extra probability mass at zero for "missed" observations. scChord then maps a noisy abundant modality onto a scarcer one. The motivation for treating readouts as decoupled is the observation, summarized in the human proteome blueprint, that transcript levels leave roughly half of protein variance unexplained.
We use this direction strictly as a methodological lens. We do not claim that transcript-protein discordance in pollen tissue distorts molecular clocks. That idea appears in the research file, but nothing in it supports the claim, and testing it would require plant proteomes the file does not contain.
Research Program
The core object is an occurrence tensor. Its count $y_{\ell k o}$ is the number of specimens of lineage $\ell$, in stratigraphic time bin $k$, of organ type $o$. The organ types are vascular axis, ovulate cupule or ovule, and pollen organ. We model
$$ y_{\ell k o} \sim \rho_{ko}\,\delta_0 + (1-\rho_{ko})\,\mathrm{NB}\!\left(z_{\ell k}\,\lambda_{\ell k}\,\pi_{ko},\ \phi_o\right), $$
where:
- $z_{\ell k}\in\{0,1\}$ indicates whether the lineage, or the organ's character state, existed in bin $k$;
- $\lambda_{\ell k}$ is the lineage's latent abundance, shared across all organ channels;
- $\pi_{ko}$ is the organ- and bin-specific preservation rate;
- $\rho_{ko}$ is the extra probability of a structural zero, for example bins with no suitable rock;
- $\mathrm{NB}(\mu,\phi)$ is a negative binomial count with mean $\mu$ and dispersion $\phi$;
- $\delta_0$ is a point mass at zero.
The origin time of the male character is the first bin in which $z_{\ell k}=1$ for the pollen-organ state. It sits upstream of the observed first occurrence. We place a birth-style prior on $z$ (once present, a state persists until extinction) and a hierarchical prior on $\pi_{ko}$. That hierarchy lets the model learn, across lineages, that pollen organs preserve one to two orders of magnitude less often than cupules, without fixing the ratio by hand.
Isolated organs are found detached, so lineage assignment is itself uncertain. The second component therefore assigns each specimen a probability over candidate lineages, rather than a hard label. For organs with images, we use spatially aware dictionary coding as in Kong et al., which handles sparse, randomly oriented microscopy better than standard deep networks. For anatomical fragments, we score characters from the Rothwell and Stockey matrix as extended in Li et al.. These assignment probabilities enter the occurrence model as soft counts, so uncertainty about which plant an organ came from propagates into the date.
The third component converts the posterior on origin time into a calibration prior for a relaxed-clock analysis. We then compare node ages under this prior with node ages under conventional calibrations.
Baselines
Naive reference: inherit the ovule minimum. This is the do-nothing reference and the de facto status quo. The stem seed-plant node takes a minimum-bound calibration from the oldest cupule or ovule, and male characters inherit it. It will be run on identical simulated and empirical data in the same BEAST relaxed-clock setup. Beating it means better coverage of the true male-origin time at equal or smaller interval width.
Hard minimum from the oldest pollen organ. The Placotheca-age minimum uses the oldest pollen organ found as a hard minimum, with a conventional diffuse tail. It is the most direct use of the scarce record. It is run identically, and beating it means reducing the systematic bias toward too-young origins that we expect from rare-organ truncation.
Ablation of the key ingredient: independent per-organ preservation. This is the same zero-inflated model, but each organ channel gets its own lineage abundance, with no shared $\lambda_{\ell k}$. It isolates whether coupling channels, rather than merely modeling preservation, is what helps. It gets the same priors, the same sampler, and the same number of posterior draws. If the full model does not beat this ablation at high preservation asymmetry, the core claim fails.
Image-assignment baselines. The strongest reported figure is SACO-II at 86.13% accuracy on three Picea pollen species, against 65.11% for VGG19 with a support vector machine. Both figures come from Kong et al.. We treat SACO as the assignment method to match rather than beat, since assignment is not our contribution.
Calibration-quality numbers. For calibration quality itself, no reported result was found on Lacuna for coverage or node-age error under organ-asymmetric preservation. The calibration-strategy study reports its a priori advantage only qualitatively in the available material.
Why not just…?
Why not just use a fossilized birth-death model with a single preservation rate? Models that treat fossil sampling as a process along the tree are the serious competitor, and in spirit they are our independent-preservation ablation with one shared rate. A single rate is dominated by abundant cupules and stems, so it will underestimate how long pollen organs can go unrecorded. Lineage-specific rates without cross-organ coupling lose the information that dense cupule records carry about the same lineage. The ablation comparison settles whether this alternative is good enough. If the uncoupled model matches ours, we will report that and recommend it as the simpler tool.
Why not just rescale an uncalibrated tree afterwards with the sparse pollen dates? The calibration-strategy simulations show that post-hoc rescaling misestimates dates and rate shifts whenever rates vary across branches, which is expected across a major reproductive transition. We therefore do not include it as a main competitor.
Why not just assign organs with an off-the-shelf convolutional network? On sparse, randomly oriented pollen microscopy, VGG19 features reached 65.11% against 86.13% for spatially aware coding (Kong et al.). Devonian samples are smaller still. In any case, our method consumes assignment probabilities from any classifier, so a better network would only help.
Why not just extend the gap with a linear morphological clock? Derkarabetian et al. found that morphological divergence tracked time almost linearly ($R^2 = 0.952$) in cave harvestmen under constant selection. The early fusion seen in Placotheca suggests a burst rather than steady change, so a linear extrapolation is likely to be biased. We include it as a secondary baseline in simulations that contain a burst.
Decisive Experiments
Simulation. Following the replicate design of the calibration-strategy study, we simulate lineage histories with known origin times for a male character. Organ-specific preservation is set at cupule-to-pollen ratios of 1, 10, and 100, with 100 replicates per condition. We then simulate sequences under rate-shifting relaxed clocks. The metrics are:
- coverage of the true male-origin time by 95% intervals;
- median absolute error in origin time;
- interval width;
- downstream node-age error after BEAST dating with each calibration.
Decisive check. The full model must hold nominal coverage at ratio 100, where the hard-minimum baseline should under-cover. It must also beat the independent-preservation ablation on error. If it does not beat the ablation, the cross-organ idea is unnecessary.
Empirical test. We compile Famennian South China occurrences from the cited descriptions: the Wutong Formation Zaijunia collection, Placotheca, and Yiduxylon. We then run a hold-out test. We remove Placotheca and ask whether the model, using only stem and cupule channels, places substantial posterior mass for the pollen-organ state at or before its actual horizon. A model that cannot predict a withheld rare occurrence from abundant ones is not trustworthy for extending ghost ranges.
Report. Finally, we report how far the stem seed-plant date shifts under the new calibration relative to the ovule-inherited minimum. This shift is the practical quantity that dating studies of seed plants would need to absorb.
References
- A Late Devonian Fertile Organ with Seed Plant Affinities from China
- New ovulate cupule further informs the relationships among early seed plants and their adaptation to wind pollination
- A new Late Devonian genus with seed plant affinities
- Performance of A Priori and A Posteriori Calibration Strategies in Divergence Time Estimation
- Plastome evolution of Engelhardia facilitates phylogeny of Juglandaceae
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- A high-stringency blueprint of the human proteome
- Spatially Aware Dictionary Learning and Coding for Fossil Pollen Identification
- Repeated and Time-Correlated Morphological Convergence in Cave-Dwelling Harvestmen (Opiliones, Laniatores) from Montane Western North America