Bio Phylo Modern Tree Inference

Version Compatibility

Reference examples tested with: IQ-TREE 2.2+ / 2.3+, RAxML-NG 1.2+.

Before using code patterns, verify installed versions match. If versions differ:

• CLI: iqtree2 --version then iqtree2 --help to confirm flags
• CLI: raxml-ng --version then raxml-ng --help to confirm flags

If code throws an unrecognized-argument or model-parse error, introspect the installed tool and adapt the example to match the actual API rather than retrying.

IQ-TREE2 uses single-dash documented forms (-alrt, -bnni, -B); -B/-T are v2.x (v1.x used -bb/-nt). Do NOT write --alrt. The likelihood site-concordance flag --scfl requires IQ-TREE 2.2.2+ (older builds have only the parsimony --scf).

Modern ML Tree Inference -- ML Support Measures Repeatability, Not Correctness

"Build a maximum-likelihood tree with support from my alignment" -> Select a substitution model, search topology and branch lengths that maximize the likelihood, then attach support that quantifies repeatability and concordance that quantifies genealogical agreement.

• CLI: iqtree2 -s aln.fasta -m MFP -B 1000 -bnni -alrt 1000 (model selection + dual support, all built in)
• CLI: raxml-ng --all --msa aln.fasta --model GTR+G --bs-metric fbp,tbe (very large trees, transfer bootstrap, precise branch lengths)

Scope: ML estimation of topology, branch lengths, model selection, branch support, concordance factors, partitioning, topology tests, and LBA control. Model-free distance/NJ trees and distance correction -> distance-calculations. Posterior distributions, MCMC, and CAT-GTR -> bayesian-inference. Per-locus gene trees summarized into a species tree under ILS -> species-trees. Time-scaled trees -> divergence-dating.

The Single Most Important Modern Insight

An ML tree is the topology and branch lengths that maximize the likelihood under an ASSUMED substitution model, conditioned on a FIXED multiple-sequence alignment. It inherits every flaw of both: a misaligned column is a fabricated character the model dutifully fits, and a misspecified model biases the point estimate toward a wrong topology that more data only sharpens. The reported support measures REPEATABILITY under resampling, not correctness. Three load-bearing facts:

1. High support is consistent with being wrong. Bootstrap, UFBoot, SH-aLRT, and aBayes all ask whether a branch reappears when the data or tree is perturbed. Under model misspecification every replicate reproduces the same bias, so support climbs toward 100% precisely as the inference becomes more wrong. At genome scale, sampling error vanishes and UFBoot ~100 on most branches is the default, not a signal.
2. More data fixes variance, not bias. Adding sites shrinks sampling error and makes the estimate more confident, but the systematic error from saturation, compositional heterogeneity, and ILS is bias that CONCENTRATES with scale. The cure for confident-wrong trees is better models, better alignments, and better diagnostics -- never more bootstrap replicates.
3. Concordance factors are the honest measure at genome scale. gCF/sCF ask what FRACTION of genes or sites actually contains a branch. A node with UFBoot 100 and gCF 35 means the concatenated likelihood is certain but only ~35% of loci endorse that branch -- biologically unresolved (ILS or introgression), and the bootstrap answered a question nobody should have asked. Report CFs on every phylogenomic tree; treat bootstrap as necessary-not-sufficient.

Tool Taxonomy

IQ-TREE2 vs RAxML-NG (both hill-climb the same likelihood; they differ in built-in features and scaling, not correctness):

Common production pattern: model-select, compute CFs, and topology-test in IQ-TREE2; do heavy bootstrap on a mega-tree in RAxML-NG with --bs-metric fbp,tbe.

Model Selection

ModelFinder (Kalyaanamoorthy 2017) scores the substitution matrix and the rate-heterogeneity model jointly, including FreeRate categories the old jModelTest/ProtTest generation never tested, and ranks by BIC (the default; k*ln(n) penalty favors simpler models that generalize on large n).

• -m MFP -- ModelFinder Plus: test all models by BIC, then search with the winner. The standard. (-m MF selects only; avoid -m TEST/-m TESTONLY, the legacy jModelTest-style limited set.)
• +G vs +R. +G (discrete Gamma, one alpha) is the workhorse for single short genes. +R (FreeRate) freely estimates each category's rate AND weight, capturing the non-Gamma, often multimodal rate distributions of concatenated phylogenomic data; expect +R3 to +R6 to win on BIC at scale.
• The +I+G trap. The invariant-sites proportion (+I) and the Gamma shape (+G) describe overlapping rate distributions, so the likelihood surface has a flat ridge: the two estimates are individually near-meaningless and start-dependent. Prefer +R (its slowest category absorbs near-invariant sites); use +G alone unless BIC genuinely demands +I+G.
• Partition model selection. -m MFP+MERGE fits per-partition models AND greedily merges partitions that fit the same model (BIC-chosen scheme, the successor to PartitionFinder greedy). Pair with -rcluster 10 (relaxed clustering: only test the top 10% most-similar pairs) for many partitions.
• Site-heterogeneous mixtures for deep data. Empirical matrices (LG, WAG) assume one residue-frequency vector for the whole alignment; real proteins do not, and that across-site compositional heterogeneity is the chief driver of deep LBA. The ML answer is the C10..C60 profile-mixture series (LG+C60+F+G), made tractable by PMSF (Wang 2018): a guide-tree pass computes one posterior-mean profile per site, and the real search uses those frozen profiles. PMSF is the standard recommendation for deep / LBA-prone protein phylogenomics.

Branch Support -- the Heart

Five measures, three different questions; the cardinal sin is cross-comparing their cutoffs.

UFBoot2 (Hoang 2018) uses the RELL trick (resample site log-likelihoods, reuse a candidate tree set) to run hundreds of times faster than the Felsenstein bootstrap (1985), and its values are CLOSER to unbiased clade probabilities -- which is exactly why the strong-support cutoff is 95, not the conservative-bootstrap 70. Treating UFBoot 70 as "good" is a category error. -bnni re-optimizes each replicate tree by NNI to rein in the inflation that model violation causes; use -B 1000 -bnni routinely. UFBoot is on a DIFFERENT scale from the standard bootstrap -- never read it with the BP-70 rule.

SH-aLRT (Guindon 2010) does not resample data; for each branch it tests whether the ML likelihood beats its two best NNI rearrangements. UFBoot (data perturbation) and SH-aLRT (tree perturbation) have different failure modes, so the community-standard joint criterion requires both:

A branch is strongly supported iff SH-aLRT >= 80% AND UFBoot >= 95%.

For large rogue-taxon-prone trees, the binary Felsenstein bootstrap lets a single wandering tip crush an otherwise-recovered deep branch; transfer bootstrap (TBE, Lemoine 2018, RAxML-NG --bs-metric tbe) replaces the in/out indicator with a gradual transfer distance and rescues those branches, at the cost of being more permissive.

Concordance Factors

Bootstrap quantifies statistical confidence given the concatenated data; concordance factors (Minh 2020) quantify how much of the actual data carries a branch.

• gCF (gene concordance factor): the percentage of decisive single-locus gene trees that contain the exact branch. Needs per-locus gene trees.
• sCF (site concordance factor): the percentage of decisive sites supporting the branch, from sampled quartets; works on a single concatenated alignment with no gene trees.
• sCFL (likelihood sCF, Mo 2023): uses ancestral-state likelihoods rather than parsimony quartet counting, substantially reducing (not abolishing) homoplasy and taxon-sampling bias. Prefer --scfl over the old --scf on IQ-TREE 2.2.2+.

# one gene tree per locus from a directory of locus alignments (-S = separate, no concatenation)
iqtree2 -S loci_dir -m MFP -B 1000 -T AUTO --prefix loci   # -B 1000 = UFBoot per gene tree, needed to contract weak branches before ASTRAL

# gene + likelihood site concordance against a fixed concatenated tree (-te fixes the tree)
iqtree2 -te concat.treefile -s concat.fasta --gcf loci.treefile --scfl 100 -T 4 --prefix concord
#  --gcf loci.treefile   per-locus gene trees for gCF
#  --scfl 100            100 sampled quartets per branch for likelihood sCF (higher = more stable)

Outputs concord.cf.tree (Newick with gCF/sCF labels) and concord.cf.stat (per-branch gCF, gDF1, gDF2, gDFP, sCF). A node with UFBoot 100 but gCF ~35 (gDF1 ~33, gDF2 ~30) is genes split three ways: the concatenated point estimate barely edges the alternatives and the node is biologically unresolved -- the signature of ILS or introgression, not a clade. Report CFs alongside support on every phylogenomic tree.

Topology Tests

For testing an a-priori hypothesis ("can I reject that X and Y are monophyletic?") against the ML tree, by comparing a set of fixed trees. Build the constrained tree with -g constraint.tree, then evaluate both trees:

# trees.nex holds the unconstrained ML tree + the constrained/alternative trees
iqtree2 -s aln.fasta -m <model> -z trees.nex -n 0 -zb 10000 -au --prefix autest
#  -z trees.nex   trees to compare        -n 0   no fresh search, just evaluate
#  -zb 10000      RELL replicates (>=1000) -au    add the AU test (must accompany -zb)

The AU test (Shimodaira 2002) uses multiscale bootstrap resampling to correct both the selection bias of the KH test (invalid on the data-selected ML tree) and the over-conservatism of the SH test (which rejects less as the candidate set is padded). Use AU by default. p-AU < 0.05 means that tree is REJECTED; p-AU >= 0.05 means it is in the 95% confidence set (failure to reject is not acceptance -- weak data fails to reject many trees). Report SH/KH only for completeness.