Substitution Model¶

A phylogenetic probability model that assigns rates or probabilities to changes among a finite set of biological character states along evolutionary branches, converting branch duration into transition probabilities for likelihood, distance, simulation, and ancestral inference.

Version: v1 · 2026-08-30 · History
Domain-specific #: 2887
Origin domain: molecular evolution
Subdomain: statistical phylogenetics
Aliases: Model of Sequence Evolution

Core Idea¶

A substitution model is a probability model for changes among a finite set of biological character states along evolutionary lineages. The states may be nucleotides, amino acids, codons, or coded morphological character values. The model specifies how an ancestral state at one end of a phylogenetic branch becomes a descendant state at the other end as evolutionary opportunity accumulates. Combined with a phylogenetic tree, branch lengths, and a distribution at the root, it yields probabilities for the character patterns observed at the tips.^[1]^[2]

The standard mathematical implementation is a finite-state continuous-time Markov chain. An instantaneous rate matrix \(Q\) specifies relative or absolute rates of change, and matrix exponentiation converts it into a transition-probability matrix \(P(t)=e^{Qt}\) for a branch of length \(t\). This architecture corrects for unobserved multiple changes at one site: the same tip states can arise after zero, two, or more substitutions, and different tip states can conceal repeated replacement. It therefore supports phylogenetic likelihood, evolutionary-distance correction, ancestral-state reconstruction, simulation, and estimation of evolutionary parameters.

“Substitution Model” names the reusable model class, not one favored matrix. Jukes–Cantor, Kimura two-parameter, Felsenstein 1981, HKY, GTR, empirical amino-acid matrices, codon models, and the Mk model for morphology are instances with different alphabets and constraints.^[3]^[4]^[5]^[6] Reversibility, stationarity, homogeneous rates, independent sites, and a single rate category are frequent simplifying assumptions, but the broader abstraction also includes nonreversible, nonstationary, branch-heterogeneous, partitioned, and mixture models when they retain the finite-state evolutionary-transition role.

Structural Signature¶

Finite biological character alphabet + ancestral state + evolutionary branch with a duration or expected-change length + stochastic transition law + optional root/state-frequency distribution + composition across a phylogenetic tree -> probabilities of descendant states and observed tip patterns.

The mandatory roles are:

character state space \(\mathcal S\): a declared finite alphabet such as \(\{A,C,G,T\}\), twenty amino acids, 61 sense codons, or coded morphology;
lineage path: a phylogenetic branch on which change occurs;
time or branch scale \(t\): calendar time multiplied by a rate, or more commonly expected substitutions per site;
transition law: conditional probabilities \(P_{ij}(t)=\Pr[X(t)=j\mid X(0)=i]\);
composition rule: transition laws on adjacent branch segments compose consistently;
tree context: a topology and branch lengths that are inputs to, not components hidden inside, the substitution law;
root or ancestral distribution: probabilities assigned to states before transitions are propagated;
observation interface: tip states or site patterns whose probability can be evaluated.

For a homogeneous continuous-time model, off-diagonal entries of \(Q\) obey \(q_{ij}\ge 0\) for \(i\ne j\), and

\[ q_{ii}=-\sum_{j\ne i}q_{ij}, \qquad P(t)=e^{Qt}. \]

The row sums of \(Q\) are zero, so each row of \(P(t)\) sums to one. \(P(0)=I\), and the Markov semigroup property \(P(s+t)=P(s)P(t)\) makes adjacent evolutionary intervals composable. A common scale convention imposes

\[ -\sum_i \pi_i q_{ii}=1, \]

so \(t\) is measured in expected substitutions per site when \(\pi\) is the stationary distribution. Stationarity requires \(\pi Q=0\); reversibility additionally requires detailed balance, \(\pi_iq_{ij}=\pi_jq_{ji}\). These are distinct optional restrictions.

Recognition test. Identify the biological character alphabet, the conditional transition rule along a branch, the branch-length scale, and how the law contributes to probabilities of tip patterns or corrected evolutionary change. If an artifact only names a tree, an alignment, an optimization method, a distance matrix, or a software model code without a state-transition law, it is not itself a substitution model.

What It Is Not¶

It is not mutation as a mechanistic event. A mutation model may describe the origination of variants within populations; a phylogenetic substitution model usually describes changes that appear along lineages after population-level sorting and is often deliberately agnostic about the mutation, fixation, selection, and drift mechanisms that produced its effective rates.

It is not a phylogenetic tree. The tree supplies topology and branch paths; the substitution model supplies conditional probabilities on those paths. The same tree can be evaluated under many models, and the same model can be used on many trees.

It is not an alignment or a sequence-similarity score. An alignment proposes positional homology. A scoring matrix such as BLOSUM can reward plausible alignment matches, while a substitution rate matrix defines an evolutionary stochastic process. Some empirical protein models inform both practices, but the inferential contracts differ.

It is not simply a transition matrix observed at one lag. A continuous-time model supplies a generator \(Q\), a family \(P(t)\), and consistent composition across branch lengths. Nor is it merely a corrected distance formula: distance correction is one consequence of an underlying transition model.

It is not necessarily stationary, time-reversible, homogeneous across branches, independent across sites, or neutral. Those assumptions characterize common tractable families. Treating them as universal would exclude legitimate nonstationary, nonreversible, covarion, mixture, and selection-aware codon models.^[2]^[5]

Finally, a best-fitting model is not a claim that the evolutionary process is literally true in every detail. Model selection ranks candidates under a criterion; adequacy testing asks whether even the selected model can reproduce important features of the data.^[7]

Scope of Application¶

The home scope is molecular evolution and statistical phylogenetics. Nucleotide models range from the equal-rate, equal-frequency Jukes–Cantor model through transition/transversion-sensitive and unequal-frequency models to the general time-reversible family. Amino-acid models use a twenty-state alphabet, often estimating exchangeabilities empirically from large protein collections. Codon models use sense codons and can parameterize synonymous versus nonsynonymous change, transition bias, codon frequencies, and lineage- or site-specific selective regimes.^[5]

The abstraction also operates in morphological phylogenetics. Lewis's Mk framework models changes among coded discrete character states and corrects ascertainment when datasets omit constant characters.^[6] Binary presence–absence traits, multistate anatomy, and certain behavioral characters can therefore enter a likelihood or Bayesian tree analysis through a substitution-model interface, provided the coding and ascertainment assumptions are explicit.

Substitution models support several distinct tasks: evaluating tree likelihoods, estimating topology and branch lengths, correcting pairwise distances, reconstructing ancestral states, simulating data under a null or fitted process, testing molecular-clock or selection hypotheses, and calculating expected site-pattern frequencies. The same transition kernel is reused across these tasks, which is evidence that the candidate is an autonomous abstraction rather than a software option.

Scope limits matter. Insertions, deletions, recombination, gene-tree/species-tree discordance, duplication and loss, and alignment uncertainty usually require additional models. A substitution model conditioned on an alignment does not make those processes disappear. Likewise, a model fitted to one genomic partition may fail on another because composition, rates, selection, or site dependencies differ.

Clarity¶

The abstraction clarifies what “the sequence evolved by this tree” means probabilistically. A topology alone says which lineages connect; it does not say how likely an \(A\to G\) change is, how repeated hits accumulate, or what a branch length measures. The substitution model supplies that missing law.

It also cleanly separates three layers that software syntax often compresses into one label. The base model constrains exchange rates and state frequencies; a rate-heterogeneity model varies the scalar rate among sites or categories; and a partition or mixture scheme assigns different parameter sets or latent classes. Thus GTR+G is not simply a new alphabet: it combines a reversible nucleotide generator with gamma-distributed rate variation.^[8]

A reader can diagnose misuse by asking whether every parameter has a probabilistic role. Exchangeabilities, equilibrium or root frequencies, branch lengths, mixture weights, and rate-category scalars affect different parts of the generative story. Treating a program's compact model name as an uninterpreted quality badge conceals those commitments.

Manages Complexity¶

The Markov transition law compresses an enormous space of possible histories. Even one site may change many times on many branches, but likelihood calculation need not enumerate every continuous-time event history. Matrix exponentiation integrates over the unobserved substitutions on each branch, and the pruning algorithm sums over possible ancestral states at internal nodes.^[1]

For a node whose state is \(i\), with child conditional-likelihood vectors \(L_c\), the recursion has the form

\[ L(i)=\prod_c\left(\sum_j P_{ij}(t_c)L_c(j)\right). \]

At the root, the site likelihood is \(\sum_i \rho_iL(i)\), where \(\rho\) is the root distribution. This dynamic program turns a sum over exponentially many ancestral assignments into local matrix-vector operations on the tree. Across conditionally independent sites, site likelihoods multiply and log-likelihoods add.

Model families manage inferential complexity by exposing controlled degrees of freedom. JC69 ties all nucleotide exchange rates and frequencies; richer families release selected constraints. Protein matrices can borrow estimates from large reference datasets when an individual alignment cannot support hundreds of free rates. Codon models impose biological structure rather than estimate an unconstrained \(61\times61\) generator. Partition and mixture models trade parsimony for better representation of heterogeneous evolution.

Abstract Reasoning¶

The simplest distance-correction example shows why the abstraction is needed. Under JC69, if \(p\) is the observed fraction of differing nucleotide sites between two sequences, the estimated evolutionary distance is

\[ d=-\frac{3}{4}\log\left(1-\frac{4p}{3}\right), \qquad 0\le p<\frac34. \]

For small \(p\), \(d\approx p\); at larger divergence, \(d>p\) because multiple substitutions are hidden. As \(p\) approaches the model's saturation value \(3/4\), the estimated distance diverges and the data cease to identify arbitrarily long history. The formula is therefore both a correction and a warning about information loss.^[3]

The general reasoning pattern is:

choose a state alphabet and model family appropriate to the characters;
declare root-frequency, stationarity, reversibility, homogeneity, and site-dependence assumptions;
scale \(Q\) so branch lengths have a stated unit;
obtain \(P(t)\) on each branch;
compose branch transitions and sum over unobserved ancestral states;
fit or integrate parameters;
compare models and test adequacy;
report which biological conclusions change under plausible alternatives.

Several deductions follow. Detailed balance implies stationarity for an irreducible model, but stationarity does not imply reversibility. A reversible model used with its stationary root distribution cannot by itself identify the root position from sequence likelihood—the familiar pulley symmetry—whereas nonreversible or nonstationary structure may carry rooting information. Branch length and rate are confounded unless one fixes a scale or supplies temporal calibration. A better likelihood can reward complexity, so model comparison must penalize parameter count or integrate over uncertainty rather than equate fit with truth.

Knowledge Transfer¶

Within evolutionary biology, the same abstraction transfers literally across nucleotide, amino-acid, codon, binary, and morphological alphabets. The matrix dimension, parameterization, and biological interpretation change, but the roles remain: finite states, branch duration, conditional transition probabilities, composition over a tree, and an observation likelihood.

Transfer also occurs across inference tasks. The kernel used to estimate a tree can simulate alignments for method validation, reconstruct ancestral states on a fixed tree, calculate corrected distances, and form null distributions for goodness-of-fit tests. That recurrence is stronger than merely sharing vocabulary.

Outside evolutionary character change, finite-state continuous-time Markov chains appear in queueing, epidemiology, reliability, and chemical kinetics. Those are literal instances of the parent Markov Process, not of the domain-specific Substitution Model. Calling customer churn or machine failure a “substitution model” would import biological lineage and character-state vocabulary unnecessarily. The portable mathematical skeleton belongs to the prime parent; the evolutionary tree and homologous-character interpretation keep this node domain-specific.

Examples¶

JC69 nucleotide model. The alphabet is four nucleotides, all equilibrium frequencies are \(1/4\), and all six unordered exchange classes are tied. It is useful as a transparent baseline and produces the multiple-hit distance correction above, but it is often too restrictive for real sequences.^[3]

GTR nucleotide model. Six reversible exchangeability parameters combine with four stationary nucleotide frequencies subject to a sum-to-one constraint. Detailed balance makes the model reversible. Rate normalization removes one additional scale degree of freedom. GTR is a flexible family, not “the true model,” and commonly appears with site-rate or partition extensions.^[4]^[9]

Empirical amino-acid model. A twenty-state exchangeability matrix and amino-acid profile estimated from a reference corpus can be used on a new protein alignment. The transfer is useful when the new alignment cannot estimate all exchangeabilities, but fit depends on whether the reference proteins resemble the target process.^[2]

Codon model. Goldman and Yang define a Markov process over sense codons, structuring rates by transition/transversion bias, codon frequencies, and synonymous versus nonsynonymous effects. This lets the substitution model estimate biologically interpretable selective constraints instead of treating all nucleotide changes independently.^[5]

Mk morphology model. Coded discrete morphological states evolve along tree branches under an equal-rate Markov model, with an ascertainment correction when constant characters were excluded from the dataset. This preserves the transition-law architecture while moving beyond molecular sequences.^[6]

Nonexample: an alignment score matrix. A BLOSUM score rewards or penalizes residue pairings for alignment. Unless it is embedded in a normalized evolutionary transition process over branch duration, it is not by itself the substitution model used to generate phylogenetic site-pattern probabilities.

Structural Tensions¶

T1 — Tractability versus biological realism. Tied rates, stationarity, and independent sites make likelihood calculation manageable but may miss lineage composition shifts, context dependence, or selection. Diagnostic: compare fit and predictive adequacy under targeted relaxations, not maximal complexity by default.

T2 — Fit versus identifiability. A richer model may fit better while leaving rates, branch lengths, or mixture components weakly identified. Diagnostic: inspect information, profiles, posterior uncertainty, and sensitivity rather than reporting only the optimum.

T3 — Common process versus heterogeneity. Pooling sites gives power; forcing one process across unrelated genes, codon positions, or structural classes creates bias. Diagnostic: partition or mix only when the data and biological design support the extra classes.

T4 — Effective substitution versus population mechanism. A fitted \(q_{ij}\) summarizes lineage-level replacement but may conflate mutation, fixation, drift, selection, and context. Diagnostic: do not interpret an effective rate as a direct mutation rate without a population-genetic bridge.

T5 — Reversibility versus rooting information. Reversibility simplifies computation and avoids choosing ancestral direction on a branch, but sacrifices directional information about the root. Diagnostic: state whether rooting comes from an outgroup, clock, nonreversible model, or external evidence.

T6 — Model selection versus model adequacy. The best candidate in a menu can still fail to reproduce composition, site-pattern, or heterogeneity features. Diagnostic: supplement relative criteria with absolute or posterior-predictive checks.^[7]

Structural–Framed Character¶

Substitution Model is mixed-structural and strongly formal. Its rate-matrix, semigroup, conditional-probability, and pruning architecture is mathematically explicit and discovered wherever finite-state evolutionary change is modeled. Yet the node is framed by biological commitments: homologous characters, evolutionary lineages, phylogenetic branches, substitutions rather than arbitrary state transitions, and inferential tasks such as ancestral reconstruction.

The abstraction is not an institutional label or evaluative category. Competing model families can be compared by explicit assumptions and empirical adequacy. Its domain framing comes from what the states and paths mean, not from a community's arbitrary classification.

Structural Core vs. Domain Accent¶

The structural core is a finite-state Markov transition family indexed by branch duration, with conditional probabilities that compose along paths and support marginalization over hidden states. This core is already carried by Markov Process, State and State Transition, Stochastic Process, and Statistical Model.

The domain accent supplies homologous biological characters, ancestral and descendant lineages, phylogenetic topology, substitutions per site, equilibrium or root character frequencies, multiple-hit correction, and likelihood of tip site patterns. Those commitments create the autonomous domain-specific residual.

Removing the domain accent does not yield a broader “substitution model” prime; it yields the existing Markov Process abstraction. Conversely, adding a DNA alphabet and a tree to a generic Markov chain does not automatically create a scientifically adequate model: the substitution interpretation, branch scaling, data-generation assumptions, and empirical boundaries must all hold.

Substitution Model strictly specializes Markov Process. Its future character state is conditionally governed by the current state through a transition law, while the phylogenetic branch provides the evolution index. The proposal-only DAG therefore uses one subsumption/specializes/strict edge to live prime:markov_process.

It also relates to State and State Transition, Stochastic Process, Probability, Representation, and Statistical Model. Those nodes explain ingredients or broader formal roles, but Markov Process is the narrowest literal live genus. Natural Selection and Phylogenesis are biologically adjacent, yet many substitution models are effective neutral or phenomenological laws and do not explicitly model selection or taxon origination.

Relationships to Other Abstractions¶

Current abstraction Substitution Model Domain-specific

Parents (1) — more general patterns this builds on

Substitution Model is a kind of Markov Process Prime

Substitution Model strictly specializes Markov Process.

Hierarchy paths (4) — routes to 4 parentless roots

Substitution Model → Markov Process → Stochastic Process

Show alternative paths (3)

Neighborhood in Abstraction Space¶

Substitution Model sits in a sparse region of the domain-specific corpus (86^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Variation Strategies concerns deliberately generating or retaining alternatives; it does not specify a lineage-indexed probabilistic transition kernel. Phylogenesis is the evolutionary origin and diversification of taxa; a substitution model describes character change conditional on branches and need not itself model speciation. Convergent Evolution names independent emergence of similar outcomes, whereas a substitution model assigns probabilities to state changes and can be used to infer—but is not identical to—convergence.

Genetic Assimilation is an evolutionary process by which an environmentally induced phenotype becomes genetically encoded. Coevolution is reciprocal evolutionary change among interacting lineages. Invasive Species concerns ecological spread. None supplies the finite character alphabet, generator, transition family, or tree likelihood.

Lexical neighbors such as Substitution Effect, Liskov Substitution Principle, Attribute Substitution, Relevance Substitution, and Substitutability use “substitution” for replacement in economics, software contracts, cognition, or function. Their roles do not overlap the phylogenetic model.

Within the home domain, distinguish the general class from JC69, K80, F81, HKY, GTR, WAG, LG, GY94, and Mk, which are model families or named instances; from gamma rate heterogeneity, which modifies rates across sites; from Felsenstein's pruning algorithm, which computes likelihood under a model; and from ModelFinder or MODELTEST, which compare candidate models.

References¶

[1] Felsenstein, Joseph. “Evolutionary Trees from DNA Sequences: A Maximum Likelihood Approach.” Journal of Molecular Evolution 17, no. 6 (1981): 368–376. Establishes computationally feasible phylogenetic likelihood from nucleotide transition probabilities and tree pruning. registry ↩a ↩b

[2] Arenas, Miguel. “Trends in Substitution Models of Molecular Evolution.” Frontiers in Genetics 6 (2015): 319. Authoritative review of nucleotide, codon, amino-acid, heterogeneity, selection, and misspecification boundaries. registry ↩a ↩b ↩c

[3] Jukes, Thomas H., and Charles R. Cantor. “Evolution of Protein Molecules.” In Mammalian Protein Metabolism, vol. 3, edited by H. N. Munro, 21–132. Academic Press, 1969. Introduces the equal-rate correction now called JC69. registry ↩a ↩b ↩c

[4] Tavaré, Simon. “Some Probabilistic and Statistical Problems in the Analysis of DNA Sequences.” In Some Mathematical Questions in Biology—DNA Sequence Analysis, edited by R. M. Miura, 57–86. American Mathematical Society, 1986. General finite-state reversible nucleotide parameterization associated with GTR. registry ↩a ↩b

[5] Goldman, Nick, and Ziheng Yang. “A Codon-Based Model of Nucleotide Substitution for Protein-Coding DNA Sequences.” Molecular Biology and Evolution 11, no. 5 (1994): 725–736. Defines a Markov codon model with transition bias, codon usage, and protein-level constraint. registry ↩a ↩b ↩c ↩d

[6] Lewis, Paul O. “A Likelihood Approach to Estimating Phylogeny from Discrete Morphological Character Data.” Systematic Biology 50, no. 6 (2001): 913–925. Extends model-based likelihood to discrete morphology and addresses ascertainment. registry ↩a ↩b ↩c

[7] Goldman, Nick. “Statistical Tests of Models of DNA Substitution.” Journal of Molecular Evolution 36, no. 2 (1993): 182–198. Establishes adequacy testing for components of phylogenetic substitution models. registry ↩a ↩b

[8] Yang, Ziheng. “Maximum Likelihood Phylogenetic Estimation from DNA Sequences with Variable Rates over Sites: Approximate Methods.” Journal of Molecular Evolution 39, no. 3 (1994): 306–314. Introduces practical discrete-gamma rate categories. registry ↩

[9] IQ-TREE Development Team. “Substitution Models.” IQ-TREE Manual, accessed 28 August 2026. Current official technical documentation for DNA, protein, codon, binary, morphological, partition, mixture, and rate-heterogeneity implementations. registry ↩

[10] “Substitution model,” Wikipedia, frozen revision 1351496616, 28 April 2026. Discovery provenance only; identity, formulas, boundaries, and classification were independently reconstructed from primary and authoritative sources. registry