🧬 biology

Smoothly Time-Varying Continuous Time Markov Chains in Phylogenetics

This paper introduces a novel "spline clock" model that utilizes inhomogeneous continuous-time Markov chains and cubic B-splines to flexibly and accurately reconstruct time-varying evolutionary rates from molecular sequence data, outperforming existing methods in both simulations and real-world applications involving foamy virus and SARS-CoV-2.

Original authors: Pratyusa Datta, Philippe Lemey, Marc A. Suchard

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pratyusa Datta, Philippe Lemey, Marc A. Suchard

Original paper licensed under CC BY 4.0 (http://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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Imagine you are trying to reconstruct the history of a family, but instead of looking at old photo albums, you are looking at their DNA. Scientists use these DNA sequences to build a "family tree" (called a phylogeny) that shows how different species or viruses are related. To make this tree tell a story about when things happened, they need to know how fast the DNA changes over time. This speed is called the "molecular clock."

For a long time, scientists assumed this clock ticked at a steady, unchanging pace, like a grandfather clock. Later, they realized the clock sometimes speeds up or slows down, but they still treated it like a clock that only changes its speed in big, sudden jumps—like switching from "slow" to "fast" at a specific hour.

The Problem: The "Jumpy" Clock
The authors of this paper argue that nature isn't so rigid. Evolutionary rates often change smoothly and gradually, like a car accelerating or braking, rather than snapping instantly from one speed to another. Previous methods that tried to model these changes often created "jumpy" or "bumpy" estimates that didn't look real, mixing genuine signals with artificial noise.

The Solution: The "Spline Clock"
The team introduces a new tool called the Spline Clock Model. Here is how it works, using some everyday analogies:

The Smooth Curve: Imagine you want to draw the path of a roller coaster. The old way was to connect dots with straight lines, creating a jagged, bumpy ride. The Spline Clock uses a flexible ruler (called a "cubic B-spline") to draw a perfectly smooth, flowing curve that fits the data. This allows the evolutionary speed to change gradually and naturally over time.
The Math Magic (Gauss-Legendre Quadrature): To calculate how much "distance" the DNA traveled along this smooth curve, you have to do some heavy math (integrals). Usually, this is like trying to measure the area under a wavy line by counting tiny squares—it's slow and can be inaccurate. The authors use a clever shortcut called Gauss-Legendre quadrature. Think of this as a super-precise ruler that knows exactly where to place its measurement points to get the perfect answer with very few steps, making the calculation fast and accurate.
The Safety Net (Gaussian Markov Random Field): Because the model is so flexible, it could potentially get "too creative" and invent fake patterns in the data (overfitting). To stop this, the authors add a "safety net" (a statistical prior). This acts like a gentle hand on the artist's shoulder, encouraging the curve to stay smooth and not wiggle wildly unless the data strongly demands it.

What They Found
The team tested their new clock in two ways:

The Simulation Test: They created fake DNA data where the speed of evolution was known to change in a specific, smooth way. When they ran their new model, it found the true speed almost perfectly, drawing a smooth line that matched the truth. The old models either missed the trend or drew a jagged, inaccurate line.
Real-World Examples:
- Foamy Virus: They looked at a virus that has been evolving for over 100 million years. The Spline Clock revealed that the virus's speed of change has fluctuated wildly over deep time, speeding up and slowing down by huge amounts. It showed that the virus evolves much faster recently than it did millions of years ago, but the pattern wasn't a simple, straight-line decline as some previous theories suggested.
- SARS-CoV-2 in Europe: They tracked how the coronavirus spread across 10 European countries in 2020. The model acted like a speedometer for the virus's travel. It clearly showed the speed dropping to near zero during strict lockdowns, picking up speed as restrictions eased in the spring, peaking in the summer, and then slowing down again as the second wave hit and new measures were put in place. The model perfectly mirrored real-world public health events.

The Bottom Line
The Spline Clock Model is a new, flexible way to measure how fast evolution happens. Instead of forcing nature into rigid boxes or jagged steps, it allows scientists to see the smooth, flowing changes in speed that actually occur in the natural world. This helps researchers get a clearer, more accurate picture of evolutionary history and how diseases spread over time.

Technical Summary: Smoothly Time-Varying Continuous Time Markov Chains in Phylogenetics

Problem Statement
Reconstructing evolutionary histories from molecular sequence data is central to evolutionary biology and infectious disease research. Standard phylogenetic inference relies on homogeneous continuous-time Markov chains (CTMCs) rescaled by molecular clock models. While relaxed clock models allow evolutionary rates to vary across branches, they often fail to capture systematic, time-dependent rate variation. Empirical studies on rapidly evolving viruses (e.g., foamy virus, SARS-CoV-2) have demonstrated that evolutionary rate estimates are highly sensitive to the sampling timeframe, a phenomenon known as the time-dependent rate phenomenon. Previous attempts to model this, such as the "polyepoch clock" (Datta et al., 2025), utilize piecewise constant rate functions. While computationally tractable, these models introduce artificial discontinuities at epoch boundaries, leading to spurious local oscillations in rate estimates and wider credible intervals, particularly in regions of dense taxon sampling. There is a need for a model that accommodates time-varying rates while enforcing smoothness and continuity across the temporal domain.

Methodology
The authors propose the spline clock model, which integrates inhomogeneous continuous-time Markov chains (ICTMCs) with a flexible, nonparametric rate parameterization.

ICTMC Framework: The sequence substitution process is modeled as an ICTMC acting along the branches of a phylogeny. The authors assume the time-varying rate function $f(t)$ scales all elements of the infinitesimal generator matrix $Q$ equally, such that $Q(t) = f(t)Q$ . Under this assumption, the transition probability matrix $P(t_0, t)$ for a branch reduces to a function of the integral of the rate function over that branch: $P(t_0, t) = \exp\left(\int_{t_0}^t f(\tau) d\tau \cdot Q\right)$ .
Spline Parameterization: To ensure smoothness, the log-transformed rate is modeled as a cubic B-spline basis expansion: $g(t) = \beta_0 + \sum \beta_i B_{i,d}(t)$ , where $d=3$ . The rate function is defined as $f(t) = \exp(g(t))$ . This choice ensures the rate is positive and allows for a log-linear interpretation, avoiding the identifiability issues associated with squared link functions.
Numerical Integration: The integral of $\exp(g(t))$ required for likelihood evaluation does not have a closed-form solution. The authors approximate these integrals efficiently using Gauss-Legendre (GL) quadrature. By exploiting the piecewise polynomial nature of the B-splines, the GL rule is applied locally on each knot interval overlapping with a branch. This approach provides high accuracy with a deterministic, fixed computational cost, avoiding the iterative convergence issues of adaptive integrators.
Bayesian Inference: The model employs a Bayesian framework. A Gaussian Markov Random Field (GMRF) prior is placed on the spline coefficients to penalize roughness and prevent overfitting. Specifically, a first-order random walk prior with a precision parameter $\tau$ is used. To ensure posterior propriety (since the likelihood does not decay to zero as coefficients approach infinity), the authors utilize a proper GMRF prior with a modified precision matrix ( $D - \rho A$ with $\rho=0.99$ ). The hyperparameters, including the smoothing precision, are estimated from the data. Inference is performed via a Metropolis-within-Gibbs MCMC scheme.

Key Contributions

Novel Model Architecture: The introduction of the spline clock model, which replaces the piecewise constant assumption of previous epoch-based models with a continuously differentiable cubic B-spline basis.
Computational Efficiency: The development of a piecewise Gauss-Legendre quadrature strategy that allows for the efficient and accurate numerical integration of the rate function required for ICTMC transition probabilities, bypassing the computational prohibitions of fully nonparametric frameworks.
Regularization Strategy: The implementation of a proper GMRF prior on spline coefficients that balances flexibility with smoothness, controlled by a data-driven precision parameter, ensuring stable inference without arbitrary knot constraints.

Results

Simulation Study: The model was tested on simulated data with a known time-varying rate trajectory (log-linear increase followed by a sharp decline).
- The spline clock model recovered the true rate trajectory with high accuracy and tighter credible intervals compared to the uncorrelated relaxed clock (UCLD) and the polyepoch clock (PCM).
- The UCLD model failed to capture systematic temporal trends, while the PCM exhibited roughness and discontinuities due to its discretized nature.
Foamy Virus (FV) Analysis: Applied to a dataset of 14 FV pol sequences spanning ~100 million years.
- The model revealed strong time-varying signals, with rate estimates varying over four orders of magnitude.
- Unlike previous power-law decay models, the spline clock recovered a non-monotonic trajectory, including an elevated rate near the root (approx. 90 MYA) and a subsequent drop, followed by an increase in recent times. This suggests complex rate heterogeneity potentially driven by both time and lineage-specific factors.
SARS-CoV-2 Spatial Diffusion: Applied to 3,959 genomes from 10 European countries collected in 2020 to model the rate of spatial diffusion.
- The inferred diffusion rates closely mirrored real-world public health interventions.
- The model detected a decrease in diffusion rates during the first wave lockdowns (Jan–April), a surge coinciding with the relaxation of measures (April–July), and a subsequent decline during the second wave (July–October).
- The analysis successfully identified the origins and spread of specific variants (B.1.160/20A.EU2 and B.1.177/20E(EU1)).

Significance and Claims
The paper claims that the spline clock model fills a critical methodological gap between rigid parametric clock models and computationally prohibitive fully nonparametric alternatives. By providing a principled way to account for systematic temporal variation in evolutionary rates, the model yields smoother, more accurate estimates with improved precision. The authors assert that the framework is not limited to nucleotide substitution rates but extends naturally to other time-varying discrete processes, such as spatial diffusion in phylogeography. The ability to recover strong time-varying signals in both deep evolutionary timescales (FV) and rapid epidemic dynamics (SARS-CoV-2) demonstrates the model's versatility and robustness. The authors suggest that future work could extend this framework to jointly capture rate variation across both time and lineages.

Technical Summary: Smoothly Time-Varying Continuous Time Markov Chains in Phylogenetics

More like this