Evolution
Molecular Clock
Neutral substitutions accumulate at ~constant rate per lineage — calibrating the tree of life
The molecular clock uses the steady accumulation of neutral nucleotide and amino-acid substitutions in DNA and protein sequences to estimate how long ago two lineages shared a common ancestor. Proposed by Emile Zuckerkandl and Linus Pauling in 1962-1965 from hemoglobin comparisons across vertebrates, the idea was given a mechanistic basis by Motoo Kimura's 1968 neutral theory: if most substitutions are selectively neutral, the fixation rate equals the mutation rate and is independent of population size. Cytochrome c diverges at about 1 amino-acid substitution per site every 20 million years; HIV evolves roughly a million times faster than mammalian nuclear genes. Modern relaxed-clock Bayesian methods allow each branch its own rate and remain the standard tool for dating divergences when the fossil record is sparse.
- ProposedZuckerkandl & Pauling 1962-65
- Theoretical basisKimura neutral theory 1968
- Cytochrome c rate~1 sub / site / 20 Myr
- Mammal nuclear synonymous~2-3 × 10-9 / site / yr
- HIV-1 rate~2.5 × 10-3 / site / yr (106× mammals)
- Modern methodRelaxed clock + Bayesian (BEAST)
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
Why the molecular clock matters
- Dates events the fossil record cannot. Soft-bodied organisms, deep microbial lineages, and most rapid radiations leave no usable fossils. The split of birds and crocodiles is fossil-anchored at ~250 Mya, but the split of placental mammal orders, the origin of HIV-1 group M (estimated ~1908 in Kinshasa from sequence dating), and the age of the last universal common ancestor (~3.8-4.1 Gya) all lean heavily on molecular clocks.
- Quantifies viral and pathogen emergence. Tip-dated phylogenies of SARS-CoV-2, influenza, and Ebola use sampling dates as calibration points and recover most-recent-common-ancestor dates within weeks of the true outbreak. SARS-CoV-2's MRCA was placed late November 2019 from a few hundred early genomes, before epidemiology confirmed it.
- Independent of generation count. Under Kimura's derivation, fixation rate = mutation rate × neutral fraction, with population size canceling out. This is why a single clock can compare species with wildly different effective population sizes — bacteria (Ne ~108) and humans (Ne ~104) — for genes under similar selection regimes.
- Mitochondrial vs nuclear gives two clocks. Mammal mtDNA ticks 5-10× faster than nuclear DNA because of higher mutational load near the replication fork and weaker repair. Comparing the two within the same species lets you cross-check divergence times and detect introgression — Neanderthal mtDNA differs by ~200 substitutions from modern human mtDNA (split ~500 Kya) while autosomes show a much shallower split (~700 Kya with introgression in non-Africans).
- Detects positive selection. dN/dS ratios — nonsynonymous substitutions per nonsynonymous site divided by synonymous substitutions per synonymous site — quantify selection across branches. dN/dS < 1 is purifying, =1 is neutral, >1 is positive selection. Influenza HA shows dN/dS > 1 on antigenic loops; lysozyme in colobine monkeys (foregut fermenters) shows dN/dS spikes consistent with adaptation.
- Anchors deep phylogenomics. Multi-locus relaxed-clock analyses with 30-100 fossil calibrations now place crown-group placental mammals at 80-90 Mya (just before K-Pg), crown angiosperms at 140-180 Mya, and crown animals at 600-800 Mya. The numbers are model-dependent but the order of magnitude is robust across implementations.
- Cheap relative to paleontology. A high-coverage genome costs ~$200; an expedition for vertebrate fossils in Madagascar costs ~$200,000. Once a clock is calibrated, every new sequenced species adds tip data essentially for free.
Common misconceptions
- The clock is universal. It is not — it is per-gene and per-lineage. Histones tick at ~10-10 per site per year, fibrinopeptides at ~10-8. Asking "what is THE molecular clock rate" is like asking "what is THE temperature of Earth." You must specify the site class and the genomic region.
- Constant rate means constant ticks per year. Constant-rate models are constant per generation, not per year. Mice with one-year generations and humans with 25-year generations cannot share a single rate without violating the assumption — this is the generation-time effect that motivated relaxed-clock methods.
- A linear regression of distance on time gives the clock. Sequence distance saturates: as substitutions pile up, multiple hits at the same site become the rule, and observed differences plateau. Phylogenetic methods correct this with substitution models (Jukes-Cantor, Kimura, GTR+gamma) that estimate true distances from observed ones.
- Bayesian relaxed clocks remove all uncertainty. They quantify it but do not remove it. Posterior 95% highest-density intervals for deep nodes routinely span 50-100 Myr. Calibration density choice (uniform vs lognormal) and clock model choice (UCLN vs autocorrelated) frequently shift point estimates by 20-30%.
- Faster genes are better clocks. Faster genes saturate faster. Cytochrome c is "slow" precisely because that lets it date splits hundreds of millions of years old. For dating Holocene events you want fast-evolving microsatellites or whole mtDNA; for dating the origin of eukaryotes you want ribosomal RNA.
- Molecular and fossil dates must agree. They often disagree by 20-50% on deep nodes. The "rocks vs clocks" controversy for crown placental mammals — fossils say ~65 Mya, clocks say ~85-100 Mya — has lasted three decades and is partly resolved by recognizing fossil dates are minimums and clock dates have wide credible intervals that overlap once both uncertainties are honest.
How the molecular clock works
The starting point is a multiple sequence alignment of homologous DNA or protein from several species. Pairwise sequence differences are converted into evolutionary distances using a substitution model that corrects for multiple hits — Jukes-Cantor for symmetric four-state DNA, Kimura two-parameter for transition/transversion bias, GTR+gamma for site-rate heterogeneity, and codon-aware models like Goldman-Yang for protein-coding sequences. The corrected distances are then fit to a tree topology by maximum likelihood or Bayesian inference. Without a clock the branch lengths are in units of expected substitutions per site; the clock is what converts those into time.
Under a strict clock the branch lengths are constrained so that all paths from the root to the tips have equal expected length — a single rate parameter scales the tree. With a relaxed clock each branch gets its own rate, drawn from a prior distribution. Uncorrelated lognormal (UCLN) treats branch rates as independent log-normal draws around a mean; autocorrelated models constrain a child branch's rate to be close to its parent's. Calibration densities — typically lognormal or uniform — are placed on a handful of internal nodes anchored to fossils, and Markov chain Monte Carlo (MCMC) integrates over branch lengths, rates, calibrations, and tree topology. The output is a posterior distribution of node ages.
Modern pipelines (BEAST 2, MCMCTree in PAML, RevBayes) handle phylogenetic uncertainty, fossil calibrations, lineage-specific rate variation, and demographic histories simultaneously. Convergence is checked with effective sample size diagnostics; nodes with ESS < 200 or with prior density that overwhelms data are flagged as poorly informed. Tip-dating methods, where viral or ancient-DNA samples have known collection dates, sidestep the fossil-calibration problem entirely and have made the molecular clock the workhorse of pandemic surveillance.
Strict vs relaxed clock
| Aspect | Strict clock | Relaxed clock (UCLN) |
|---|---|---|
| Rate parameter | Single rate for all branches | One rate per branch, drawn from a prior |
| Assumption | Constant substitution rate across the tree | Rate varies; mean and variance of variation are estimated |
| Fits short timescales (within-genus) | Often acceptable | Acceptable but over-parameterized |
| Fits deep phylogenies (vertebrates, plants) | Rejected by likelihood-ratio test | Standard choice |
| Statistical test | Felsenstein 1981 LR test against unconstrained tree | Pass through, no test needed (it is the unconstrained model) |
| Number of free parameters | 1 | 2B + 2 where B is number of branches |
| Computation | Fast, closed-form for some priors | Slow MCMC, hours to days |
| Software example | r8s, MEGA's RelTime | BEAST 2, MCMCTree, RevBayes |
| Generation-time effect | Cannot accommodate | Captures via branch-specific rates |
Famous case studies
- Hemoglobin alpha-beta gene duplication (Zuckerkandl & Pauling 1962). The original molecular clock paper. By comparing alpha and beta hemoglobin chains across vertebrates, they estimated the gene duplication that produced the two chains at ~500 Mya — an estimate that has held up under modern analyses to within ~20%.
- Origin of HIV-1 group M. Bette Korber and colleagues (2000) used a tip-dated phylogeny of HIV-1 sequences sampled from 1959 to 1998 to estimate the most recent common ancestor of group M at ~1931 (95% CI 1915-1941). Subsequent analyses with archived 1959 Kinshasa sample DRC60 and 1960 ZR59 narrowed the date to ~1908 in southwestern Cameroon or DRC.
- Hominoid slowdown (Wen-Hsiung Li 1987). Li and colleagues showed that synonymous substitutions in human-rodent comparisons exceed those in human-primate comparisons by a factor of 2-4, which under a strict clock should be impossible. This empirical result drove the field to develop relaxed-clock models in the 1990s.
- The Cambrian explosion debate. Molecular clocks consistently place crown-bilaterian origin at ~600-700 Mya, well before the Cambrian fossil record (~540 Mya). The discrepancy is partly explained by missing Precambrian fossils (small, soft-bodied, non-preserved) and partly by clock acceleration during the explosion itself, which violates clock assumptions.
- SARS-CoV-2 MRCA dating. Within weeks of the first sequences, Bayesian tip-dating placed the most recent common ancestor in late November 2019 (95% HPD interval roughly Oct-Dec 2019). Subsequent analyses with thousands of genomes have not substantially shifted this estimate, demonstrating the robustness of the clock when sampling is dense.
Frequently asked questions
Who first proposed the molecular clock?
Emile Zuckerkandl and Linus Pauling first articulated the idea in a 1962 Festschrift paper for Albert Szent-Gyorgyi, then elaborated it in their 1965 essay Evolutionary Divergence and Convergence in Proteins. By comparing hemoglobin sequences across vertebrates they noticed that the number of amino-acid differences scaled almost linearly with the time since common ancestry estimated from fossils — alpha and beta globin chains differ by about the same amount in human-horse and human-cow comparisons. They coined the phrase chemical paleogenetics and conjectured that proteins were timekeepers. Motoo Kimura's 1968 neutral theory provided the mechanistic underpinning: if most substitutions are selectively neutral, fixation rate equals mutation rate, which is approximately constant per generation.
How fast is the molecular clock?
Rates vary by orders of magnitude depending on the gene, the lineage, and the type of site. Synonymous sites in mammalian nuclear DNA tick at about 2 to 3 x 10^-9 substitutions per site per year; nonsynonymous sites are about ten times slower because purifying selection removes most amino-acid changes. Cytochrome c, a deeply conserved respiratory protein, diverges at roughly 1 substitution per site per 20 million years. RNA viruses evolve about a million times faster than vertebrate nuclear genes — HIV-1 substitutes about 2.5 x 10^-3 per site per year, accumulating intra-host diversity in weeks. Mitochondrial DNA in mammals ticks 5 to 10x faster than nuclear DNA. The clock is genome-region-specific, not universal.
What is the difference between a strict and a relaxed clock?
A strict clock assumes one rate parameter for the whole tree — every branch ticks at the same speed. This is the original Zuckerkandl-Pauling formulation and is rejected by likelihood-ratio tests on most empirical datasets because lineages with shorter generation times (rodents) accumulate substitutions faster than lineages with longer generation times (great apes). A relaxed clock allows each branch its own rate drawn from a prior distribution — uncorrelated lognormal in the popular UCLN model, or autocorrelated Brownian motion in the older method. Bayesian programs like BEAST, MCMCTree, and RevBayes integrate over rate variation while constraining a few internal nodes with fossil calibrations, then output posterior age distributions for every node.
How are clocks calibrated with fossils?
Fossils provide minimum age constraints — a 56-million-year-old early primate fossil tells you the primate crown group is at least that old, but not how much older. Clock calibration places a probability distribution on selected internal nodes (gamma, lognormal, or uniform with a hard minimum from the fossil and a soft maximum from a guess about the geologic context). The MCMC then samples branch lengths in time units rather than substitutions, with rate parameters as nuisance variables. A common pitfall: with a strict clock and a single calibration point, all node ages scale linearly with that one number — a 10% miscalibration shifts every estimate by 10%. Multiple well-spread calibrations are essential.
Why does the molecular clock work at all?
Kimura's neutral theory provides the cleanest answer. If a fraction f of mutations are selectively neutral and the per-generation mutation rate at a given site is mu, then in a population of size N the per-generation fixation probability of any one new mutation is 1/(2N) for a diploid. The expected number of new neutral mutations per generation is 2N x mu x f, so the fixation rate is just mu x f — independent of population size. Population fluctuations and drift cancel out. The clock thus ticks at the per-site mutation rate times the neutral fraction, which biology keeps roughly stable for housekeeping genes. The clock breaks down when selection regimes shift, generation times change, or DNA repair efficiency varies.
What is the generation-time effect and why does it matter?
Most mutations arise during DNA replication, so species that copy their genomes more often per unit time accumulate more mutations per unit time. Mice (1 generation per year) substitute roughly 2 to 4x faster per million years than humans (25 years per generation) at synonymous sites, and rate differences of 5x have been measured between rodents and primates. This violates the strict clock and was the central reason relaxed-clock methods were developed in the late 1990s. The hominoid slowdown — apes ticking slower than Old World monkeys — was first measured by Wen-Hsiung Li in 1987 and remains a textbook example of why a single clock rate cannot be applied across a vertebrate phylogeny.