Beyond Invariable Sites: Using Evolutionary Stasis to Map Multilayered Constraints on the Evolution of Viral and Mammalian Genomes Authors: Sergei L Kosakovsky Pond, Hannah Verdonk, Steven Weaver, Gallean Brown, Danielle Callan, Anton Nekrutenko, Darren P Martin, Barbara Holland Published: 2026 ABSTRACT AbstractThe quantification of genomic conservation has progressed from foundational statistical modeling of evolutionary rates to state-of-the-art deep learning architectures. However, a major resolution gap remains at the zero-rate origin, where standard selection inference tools fail to distinguish between sites that are invariant due to chance (stochastic invariance) or low substitution opportunity and those that are invariant due to extreme purifying selection. We present B-STILL (Bayesian Significance Test of Invariant Low Likelihoods), a hierarchical Bayesian framework designed to resolve the selective landscape of protein-coding genes near the zero-rate limit. By leveraging gene-level rate distributions (prior calibration) and modeling codon-site-specific substitution opportunities (determined by genetic-code degeneracy and nucleotide substitution biases), B-STILL quantifies the statistical significance of observed stasis. We define a rate-based stasis threshold to identify evolutionary stasis anchors (ESAs)—sites where the upper bound on the evolutionary rate is statistically constrained relative to the background rate of the gene due to extreme purifying selection. Validation against clinical and pathogen datasets confirms that ESAs are strong predictors of biological fitness and pathogenicity. Applying B-STILL across viral and mammalian genomes, we identify thousands of significantly clustered ESAs that map to known functional domains and uncharacterized structural motifs. These results establish B-STILL as a scalable, statistically rigorous framework for high-resolution genomic annotation, converting previously uninformative invariant sites into precise markers of extreme evolutionary constraint.Keywords: comparative genomics, codon models, purifying selection, evolutionary stasis, B-STILL Distinguishing genome positions under extreme purifying selection from those that remain unchanged by chance alone is a major challenge in evolutionary biology. We present B-STILL (Bayesian Significance Test of Invariant Low Likelihoods), a computational framework that differentiates these scenarios by modeling site-specific substitution opportunity. By analyzing viral and mammalian genomes, we demonstrate that B-STILL identifies evolutionary stasis anchors—highly conserved sites that serve as key indicators of critical functional and structural constraints. Introduction Historically, invariable nucleotide sites have been viewed as uninformative for phylogenetic reconstruction, as positions lacking variation provide no signal for resolving tree topology. Consequently, these sites were often relegated to background covariates—most notably the “+I” parameter in general time reversible (GTR) + I + Γ models—designed to correct for rate heterogeneity and prevent the artificial underestimation of branch lengths. This approach, while computationally convenient, effectively acts as a black box that discards site-specific information, obscuring a primary signature of purifying selection: the absolute persistence of specific nucleotide states at a genomic position across deep time. The field of genomic functional annotation has since transitioned through three distinct analytical epochs. The first was defined by the mathematical formalization of phylogenetic likelihoods, yielding foundational tools like phyloP (Davydov et al. 2010; Pollard et al. 2010) that provided essential, site-specific maps of purifying selection. While effective for identifying conserved regulatory elements, these methods often encounter a “significance plateau” at invariable sites in deep alignments. In the limit of absolute stasis—defined here as the complete absence of observed nucleotide substitutions at a site in the alignment—standard likelihood ratio tests (LRTs) saturate at a maximal significance ceiling. Under these conditions, LRTs lose the resolution necessary to distinguish between sites that are invariant due to low evolutionary opportunity (where substitutions are unlikely to occur under neutral evolution) and those that are invariant or nearly invariant due to extreme, site-specific purifying selection (evolutionary constraint). For a detailed technical review of existing methodologies and their respective limitations in the context of site-level conservation, see the Supplementary material. The second epoch expanded the dimensionality of conservation, utilizing sophisticated null models to identify multilayered and overlapping regulatory architectures (Sealfon et al. 2015). Currently, the field is entering a third epoch dominated by phylogeny-aware deep learning architectures and genomic language models (gLMs) such as GPN-Star (Ye et al. 2025; Avsec et al. 2026). These models shift the paradigm from raw nucleotide identity to “semantic matching,” capturing hidden regulatory grammar across hundreds of millions of years. However, these advanced neural networks often act as black boxes, making it difficult to isolate the underlying phylogenetic signal from complex biochemical and contextual features. We present B-STILL (Bayesian Significance Test of Invariant Low Likelihoods), a framework that bridges these eras by resolving the selective landscape of protein-coding sequence data around the limit of absolute evolutionary stasis. Crucially, a site may appear invariant in an alignment for three distinct reasons: (i) it is under extreme purifying selection at the amino-acid level, where any nonsynonymous change is deleterious; (ii) it has low evolutionary opportunity (ie it is “evolutionarily cold”) due to local DNA repair dynamics, low tree length (insufficient evolutionary depth), or codon structure that restricts synonymous substitutions; or (iii) it is stochastically invariant, where substitutions could have occurred neutrally but did not by random chance. We define sites in the first category as evolutionary stasis anchors (ESAs). The core intuition of B-STILL is that the significance of an invariant site is conditioned on its cumulative substitution opportunity. By leveraging the site-specific and gene-specific synonymous substitution rate distributions as an internal control for the neutral background substitution rate, B-STILL distinguishes between sites that are invariant because they are evolutionarily cold and those that are truly anchored (absolute sequence stasis despite synonymous opportunity). By pooling information across the entire gene to estimate data-specific empirical priors, B-STILL calculates an empirical Bayes factor (EBF) that identifies ESAs as positions where strict sequence constraint is statistically anomalous relative to the global selective regime for the entire gene. To identify larger-scale signatures of constraint, we complement these site-level inferences with a nonparametric Hypergeometric Scan Statistic that detects regional Stasis Clusters—contiguous genomic intervals exhibiting a significantly high density of ESAs relative to the gene-wide background. Materials and Methods The B-STILL Framework B-STILL is built upon the FUBAR framework (Murrell et al. 2013), which utilizes a Dirichlet process prior to model rate heterogeneity across sites. The underlying evolutionary process is modeled using a standard Muse–Gaut (MG94) codon model modified with a General Reversible (REV) nucleotide substitution matrix (Muse and Gaut 1994; Kosakovsky Pond and Frost 2005). The rate of substitution from codon i to codon j (qij) is defined by where α and β represent the site-specific synonymous and nonsynonymous substitution rates, θ is the nucleotide exchangeability (relative to θAG=θGA=1, and π is the target nucleotide frequency. The equilibrium frequency over sense codons is obtained using the CF3×4 estimator (Kosakovsky Pond et al. 2010). In the B-STILL implementation, the nucleotide exchangeability parameters (θ) and branch lengths are first estimated using a standard GTR model on the entire alignment to provide a robust background fit. These parameters are then held fixed during the subsequent Bayesian inference phase, allowing the model to focus exclusively on site-specific rate variation. Fixed-Grid Bayesian Inference (FUBAR) To achieve computational efficiency, B-STILL follows the FUBAR approach of approximating the continuous rate distribution using a discrete K×K grid of (α,β) points. Let Θ={(αk,βk)}k=1K2 be the set of grid points. The likelihood of the data at site s given a specific grid point k is Ls(αk,βk), which is calculated using the standard Felsenstein’s pruning algorithm. FUBAR avoids the computational burden of traditional MCMC sampling over branch lengths and substitution parameters by leveraging this fixed grid. The prior probability of each grid point, π={πk}k=1K2, is estimated by pooling information across all N sites in the alignment. In B-STILL, we typically employ a 0th order variational Bayes (VB0) approximation, which is mathematically equivalent to the expectation–maximization algorithm with a Dirichlet prior on the grid weights, or a collapsed Gibbs sampler to estimate the prior weights π that maximize the marginal log-likelihood: Once the gene-wide prior π is established, the posterior probability of site s belonging to rate regime k is calculated as This hierarchical sharing allows B-STILL to calibrate the surprise of invariance at a single site against the background frequency of conservation observed across the entire gene. Standard selection scans can have limited resolution at the boundary of zero substitutions because linear or log-spaced grids lack the density required to resolve the likelihood surface near the origin. B-STILL addresses this by implementing a high-resolution quadratic grid in the near-zero regime. For a grid of K points per dimension, the rates for the kth point are defined as This quadratic clustering provides the statistical sensitivity needed to distinguish between near-zero purifying selection and absolute evolutionary immobilization. The primary statistical metric in B-STILL is the EBF, which quantifies the surprise of observing a specific selective regime relative to the gene-wide prior: Exact invariance (α=0,β=0): Quantifies evidence for absolute sequence constraint. Exact invariance (α=0,β=0): Quantifies evidence for absolute sequence constraint. Synonymous stasis (EBFsyn, α=0): Quantifies evidence for constraint at the nucleotide level, regardless of the nonsynonymous rate. Synonymous stasis (EBFsyn, α=0): Quantifies evidence for constraint at the nucleotide level, regardless of the nonsynonymous rate. Nonsynonymous invariance (β=0): Quantifies evidence for constraint at the protein level, regardless of the synonymous rate. Nonsynonymous invariance (β=0): Quantifies evidence for constraint at the protein level, regardless of the synonymous rate. Proximal Stasis (EBFprox, E[S]