Long-Branch Attraction

Long-branch attraction (LBA) is a systematic bias in phylogenetic inference whereby two or more distantly related evolutionary lineages that have undergone rapid sequence substitution—or share convergent nucleotide/amino-acid compositional bias—are erroneously reconstructed as closely related sister taxa (or pulled artificially toward a distant outgroup root). First proved mathematically by Joseph Felsenstein in 1978 in his landmark analysis of four-taxon trees, long-branch attraction differs fundamentally from random sampling error: whereas random noise diminishes as more genes or sites are sequenced, an unmodeled systematic bias is statistically inconsistent—meaning that adding larger genome-scale supermatrices (phylogenomics) causes the inference algorithm to converge on the wrong tree topology with 100% bootstrap support or 1.0 Bayesian posterior probability.

Throughout the history of molecular systematics, long-branch attraction has generated some of the most famous controversies in evolutionary biology—including the artificial basal placement of Microsporidia (intracellular fungal parasites once misclassified as ancient amitochondriate "Archezoa"), the "acoelomorph" and "myzostomid" bilaterian root debates, the nematode–arthropod (Ecdysozoa vs. Coelomata) controversy, and the topological placement of Asgard archaea (Heimdallarchaeia) relative to the eukaryotic nuclear lineage during eukaryogenesis.

To demonstrate how a phylogenetic method can become statistically inconsistent, Felsenstein (1978) analyzed an unrooted four-taxon tree ((A, B), (C, D)) connected by an internal branch of length q (representing the true shared ancestry of clade A+B), where two non-sister terminal branches (A and C) evolve rapidly with high substitution probability p, while their true sister lineages (B and D) evolve slowly with low substitution probability q (p ≫ q):

Why Maximum Parsimony Fails in the Felsenstein Zone

Maximum Parsimony (MP) selects the tree topology that minimizes the total number of character-state steps without estimating branch lengths or correcting for multiple hits at the same site. In the parameter space where p² / (K − 1) exceeds q (where K is the alphabet size: K = 4 for DNA, K = 20 for amino acids)—a region now known universally as the "Felsenstein Zone"—parallel convergent substitutions along the two long branches A and C (homoplasy) occur more frequently by chance alone than true single substitutions along the short internal branch q (synapomorphy). As sequence length L → ∞, the law of large numbers guarantees that parsimony will infer the false tree ((A, C), (B, D)) 100% of the time.

When Felsenstein introduced Maximum Likelihood (ML) phylogenetics on DNA sequences (1981), he showed that if the substitution model matches the true evolutionary generating process, ML is statistically consistent inside the Felsenstein Zone because it uses estimated branch lengths (p) to discount expected parallel hits along long branches. However, in real biological sequences, standard ML and Bayesian models (GTR+Γ, LG+Γ) are simplified approximations of biochemistry and remain vulnerable to LBA whenever two major forms of model misspecification occur:

1. Across-Site Compositional Heterogeneity (Overestimating the Effective Alphabet Size K_eff): Standard site-homogeneous amino acid matrices (WAG, LG, JTT) assume that every column in a protein alignment can freely sample all K = 20 amino acids according to a single global equilibrium vector. In reality, structural and biophysical constraints restrict any individual protein site to an effective alphabet of only K_eff ≈ 2 to 5 biochemically compatible residues (for example, only Ile, Val, Leu in a tightly packed hydrophobic core, or only Lys, Arg at a salt bridge). Because the probability of chance convergence between two long branches scales inversely with the effective alphabet size (∼ 1 / (K_eff − 1) rather than 1 / 19), site-homogeneous LG+Γ models underestimate the rate of parallel amino-acid convergence by 4- to 9-fold, causing long branches to attract even under Maximum Likelihood.
2. Across-Branch Compositional Bias (GC/AT vs. Thermophilic IVYWREL Enrichment): Lineages that independently adapt to similar environments—such as AT-rich endosymbionts/parasites (Microsporidia, Buchnera, Plasmodium, mitochondrial genomes) enriching for Phe, Tyr, Met, Ile, Asn, Lys (FYMINK codons), or hyperthermophilic archaea and bacteria enriching for purine-loaded (G+A) and hydrophobic/charged IVYWREL residues—converge on identical nucleotide or amino-acid compositions across unrelated sites, pulling hyperthermophilic or AT-rich branches together on the tree.
3. Heterotachy (Covarion Shift): When the evolutionary rate of specific sites shifts across lineages over time (for example, when a structural loop becomes an active-site binding interface in one clade), a single site-specific gamma rate multiplier (r_i) cannot capture the branch-by-branch rate shift.

Modern phylogenomic studies apply a four-part diagnostic and corrective protocol to detect and suppress long-branch attraction:

• Site-Heterogeneous Mixture Models (CAT-GTR, PMSF, and LG+C20..C60): Introduced by Nicolas Lartillot and Hervé Philippe (2004), the Bayesian Dirichlet-process CAT model (and its Maximum Likelihood approximations: LG+C60 empirical profile mixtures and Posterior Mean Site Frequency [PMSF] models in IQ-TREE 2) assigns each alignment column to its own biochemical equilibrium profile (K_eff ≈ 2–6), accurately predicting high rates of site-specific homoplasy and dissolving LBA artifacts across deep eukaryotic and archaeal phylogenies.
• Dayhoff-6 and SR4 Physicochemical Amino-Acid Recoding: Recoding the 20 amino acids into 6 Dayhoff exchangeability groups ({A,G,P,S,T}, {D,E,N,Q}, {H,K,R}, {F,W,Y}, {I,L,M,V}, {C}) or 4 Susko–Roger (SR4) groups strips away superficial intra-group substitutions caused by AT/GC mutational pressure or thermophilic Ile ↔ Val / Lys ↔ Arg swaps, retaining only deep inter-group structural transversions.
• Taxon Sampling Expansion (Breaking Up Long Branches) and Outgroup/Fast-Site Stripping: Adding intermediate sister taxa bisects long unbroken branches of length p into shorter segments (p₁, p₂) whose product of parallel convergence drops substantially; complementary sensitivity sweeps (slow-fast site stripping removing the fastest-evolving 10%–50% of sites, or long-branch taxon jackknifing) test whether a clade collapses when distant outgroups are removed.

How did Long-Branch Attraction create the "Archezoa" hypothesis for Microsporidia?

In the 1980s and early 1990s, early 16S/18S small-subunit ribosomal RNA trees placed Microsporidia (obligate intracellular eukaryotic parasites lacking canonical cristae-bearing mitochondria) at the very base of the eukaryotic tree alongside Giardia and Trichomonas, inspiring Thomas Cavalier-Smith's "Archezoa" hypothesis—the idea that these organisms were living relics that diverged before the endosymbiotic acquisition of the mitochondrion. Subsequent multi-gene protein phylogenies and site-heterogeneous models proved that this basal placement was a classic long-branch attraction artifact: because Microsporidia have compact, rapidly evolving genomes with extreme substitution rates, their ultra-long 18S rRNA branch was artificially pulled down to the equally long prokaryotic (archaeal/bacterial) outgroup branch used to root the eukaryotic tree. When analyzed with slow-evolving protein markers (RPB1, α/β-tubulin, Hsp70) under gamma-corrected and site-heterogeneous models—and confirmed by the discovery of remnant mitochondrial organelles (mitosomes) inside microsporidian cells—Microsporidia were robustly resolved as highly derived intracellular Fungi.

How did resolving Long-Branch Attraction establish the Ecdysozoa (molting animal) clade?

Before 1997, traditional morphology and early uncalibrated 18S rRNA trees grouped Arthropoda (insects, crustaceans) with Annelida (segmented worms) in an "Articulata" clade, while placing Nematoda (roundworms, such as C. elegans) near the base of Bilateria as primitive "pseudocoelomates." Anna Marie Aguinaldo, James Lake, and colleagues (1997) demonstrated that Caenorhabditis elegans and most model rhabditid nematodes evolve 2- to 3-fold faster in their 18S rRNA gene than other bilaterians, causing the long C. elegans branch to be attracted toward the distant basal outgroup. By specifically sequencing slowly evolving nematodes (such as the trichinellid Trichinella spiralis), they broke up the long nematode branch and showed that nematodes cluster directly with arthropods, tardigrades, and priapulids inside Ecdysozoa—the superphylum of animals that molt a three-layered cuticle under ecdysteroid hormone control.

1. Felsenstein, J. (1978). "Cases in which parsimony or compatibility methods will be positively misleading." Systematic Zoology, 27(4), 401–410. https://doi.org/10.2307/2412923
2. Lartillot, N., & Philippe, H. (2004). "A Bayesian mixture model for across-site heterogeneities in the amino-acid replacement process." Molecular Biology and Evolution, 21(6), 1095–1109. https://doi.org/10.1093/molbev/msh112
3. Bergsten, J. (2005). "A review of long-branch attraction." Cladistics, 21(2), 163–193. https://doi.org/10.1111/j.1096-0031.2005.00059.x
4. Wang, H. C., Minh, B. Q., Susko, E., & Roger, A. J. (2018). "Modeling site heterogeneity with posterior mean site frequency profiles accelerates accurate phylogenomic estimation." Systematic Biology, 67(2), 216–235. https://doi.org/10.1093/sysbio/syx068