Learning center

Model Selection: aligned sequences, fixed trees and reviewable comparisons

Model Selection is temporarily unavailable. This guide explains a locally reviewed workflow; it does not provide access or instructions to bypass that restriction.

Model Selection is temporarily unavailable. This guide explains a locally reviewed workflow; it does not provide access or instructions to bypass that restriction.

Use Model Selection to compare substitution models for an explicitly aligned sequence dataset on one supplied topology. The tree topology stays fixed while branch lengths and permitted model parameters are fitted. A preferred model is relative to the candidates and assumptions you supplied; it does not establish the true evolutionary history.

1. Prepare separate scientific inputs

RoleWhat to supply and check
AlignmentSelect DNA or Protein explicitly. Supply 4–32 uniquely named taxa, equal aligned lengths, at most 100,000 columns and 20,000 distinct patterns. DNA requires uppercase A/C/G/T and all four bases observed. Protein requires the 20 standard uppercase amino acids. No gaps, ambiguous symbols, RNA conversion, stop removal, translation, trimming or subsampling is performed by this contract.
TopologySupply one fully resolved unrooted Newick tree with exactly the same taxon labels, each once. Its representation root has three children; there are 2n−3 fitted branches for n taxa. Remove neither taxa nor sequence columns simply to force a match. Internal labels, support annotations and comments are unsupported. Any supplied initial lengths must be finite and nonnegative; a root stem is absent or zero. Initial lengths are not measured results.
Partitions, only when selectedSupply a separate UTF-8 CSV, bound to this alignment, with the exact header name,start,end,step. Use 2–16 named rows, 1-based inclusive coordinates, and a positive integer step (1 visits every site; 3 visits every third site). Every site must occur exactly once. Each partition needs all four observed bases. Replacing the alignment requires checking and rebinding its partitions.

In the reviewed candidate controls, choose the input kind, load the alignment through the launcher and use Load Newick tree for the separate topology. Partition and coding controls appear only when their methods are selected. Hosted candidate requests likewise bind named alignment, topology and conditional partition roles; a file’s name does not determine its role. These controls remain subject to the public gate.

2. Follow the complete public-data example

The local reference workflow used the PAML mtCDNA primate alignment at the pinned source revision: human, chimpanzee, bonobo, gorilla, orangutan, Sumatran and gibbon; 9,993 columns per taxon, 69,951 bases in total. The prepared FASTA preserves all seven declared records with whitespace/case normalization only. The upstream source itself describes prior stop-codon removal. This is a complete example alignment, not an unmodified whole genome.

The prepared FASTA’s SHA-256 is [recorded file identifier]. Retain the original source, conversion record and prepared file separately. This guide links to upstream data; it does not redistribute the alignment or claim its commercial redistribution has been cleared.

The local example supplied this topology separately, using 0.1 as initial branch lengths:

(((human:0.1,(chimpanzee:0.1,bonobo:0.1):0.1):0.1,gorilla:0.1):0.1,(orangutan:0.1,Sumatran:0.1):0.1,gibbon:0.1);
  1. Check the seven labels and 9,993-column length before selecting methods. The core alignment identity is distinct from the file-byte SHA-256.
  2. Choose DNA and select JC69, K80, HKY85, TN93 and GTR, plus the desired AIC, BIC or AICc ranking. Begin by reading this homogeneous-model example before adding separate extensions.
  3. In an authorized review environment, a completed run should retain all five fits, their 11 fitted branches, and parameter totals of 11, 12, 15, 16 and 19 respectively. Count branches as fitted parameters; do not use substitution parameters alone.
  4. Inspect the fit overview and requested criterion tables. The recorded homogeneous AIC comparison gave GTR a score of approximately 67,071.66304 and HKY85 67,243.90125, a difference of about 172.23821. These are recorded checks for this exact input/topology, not guaranteed values for another dataset or a claim of biological truth.
  5. Open the fitted tree and branch table. Compare branch identities using their taxon splits, rather than relying on row order or the apparent left-to-right drawing.

For an explicit partition exercise, the recorded position groups have 3,331 sites each:

name,start,end,step
position_group_1,1,9993,3
position_group_2,2,9993,3
position_group_3,3,9993,3

These supplied technical groups do not infer genes or a coding frame. Each partition fits its own branches and substitution parameters on the same topology. Inspect the complete site maps, per-partition fits and aggregate comparison. Linked/proportional branches and combined partition/rate-mixture fits are outside this route.

3. Choose compatible analyses

Method familyMeaning and limits
Homogeneous DNA and rankingsJC69, K80, HKY85, TN93 and GTR. Rankings require at least two selected models on the same alignment and topology. Lower AIC/BIC/AICc is preferred within that candidate set. AICc requires n > k+1, including within partitions. n is alignment columns, not taxa × columns.
Hierarchy, averaging, LRT, summary, decisionHierarchy compares selected converged fits. Averaging reports conditional AIC weights, not averaged biological truth. Nested LRT uses asymptotic chi-square tails and Holm correction at 0.05 for the displayed family. Decision uses BIC-weighted Euclidean loss between matched branch vectors; it is not a posterior probability or KL divergence. Summary needs a ranking criterion and preserves ties.
Gamma and invariant extensionsGamma adds separate +G4 fits: four equally weighted conditional-mean rate categories, fitted alpha between 0.05 and 100, boundary estimates refused. Invariant adds separate +I fits and requires constant and variable columns. Observed constant columns differ from the inferred zero-rate fraction. Selecting both gives separate +G4 and +I fits, never +I+G. Existing rankings cover homogeneous fits only. +I branches use mean-all-sites scaling with variable-component rate 1/(1−p).
Codon ModelsExplicitly confirm a homologous coding alignment, frame 1 and NCBI code 1 (Standard) or 2 (Vertebrate mitochondrial). Require complete codons, no stops under that code, and at most 4,000 codon patterns. GY94-F1X4 and GY94-F3X4 fit one kappa, one omega and branches; F3X4 requires every base at every codon position. Empirical frequency parameters count toward AIC/BIC. Sample size is codon columns; branches are nucleotide substitutions per codon. Scores stay separate from DNA rankings. One omega is not a positive-selection test, site/branch classification or confidence interval.
Protein ModelsProtein mode fits JTT, WAG and LG with fixed published exchangeabilities and normalized published frequencies. AIC/BIC/AICc use amino-acid columns and fitted branches. No automatic DNA translation, +F, mixture, topology search or protein adequacy inference is included. DNA controls are disabled in Protein mode.

4. Read adequacy and uncertainty differently

Model Adequacy: choose a homogeneous DNA reference model already selected, 19–999 simulation/refit replicates and a seed from 0 to 2⁶⁴−1. Each full-size simulated alignment is refitted on the supplied topology. The multinomial-deviance tail probability is a parametric-bootstrap diagnostic, not posterior predictive inference. Check observed deviance, exceedances, replicate count, minimum reportable probability and the conditional Monte Carlo interval. A large value does not establish adequacy. It does not assess the additional gamma, invariant, partition or codon fits.

Model Uncertainty: choose at least two homogeneous DNA models, one or more ranking criteria, 19–999 replicates and an explicit seed. Columns are sampled with replacement, using the same column indices for every taxon; all selected models are refitted. The complete record is limited to 2,000,000 resampled columns. Sole-best and tied-best frequencies are separate; ties are not broken by selection order. The intervals describe Monte Carlo error conditional on the setup, not biological confidence. Exchangeable independent sites are assumed; topology, block, codon and partition resampling are not included.

Preserve integer seeds as exact decimal text. Hosted candidate inputs use decimal text for unsigned 64-bit seeds. Spreadsheet automatic number conversion can round long seeds even when the exported CSV is exact.

5. Treat refusals as incomplete work

The public availability refusal is expected. In local candidate reviews, malformed or mismatched inputs, duplicate labels, unsupported symbols, incomplete or overlapping partitions, invalid coding choices and insufficient AICc sample size are refused. The numerical workflow also refuses boundary estimates, disagreeing or unconverged optimizer starts, failed replicates, cancellation and timeout. No failed replicate is silently omitted and no provisional fit set becomes a completed result.

The complete calculation shares a time budget: the ordinary candidate fit budget is five minutes and the uncertainty workflow uses ten minutes including earlier fits. Larger inputs or more replicates can exceed the budget. Reducing replicate count changes Monte Carlo precision; disclose it. Preserve the refusal and parameters rather than reporting a partial ranking or zero-filled result. Two optimizer starts and reference agreement do not prove a global optimum.

6. Preserve complete numerical outputs

  1. Review the fit overview, criterion tables, branch lengths, rate matrices, frequencies and fitted trees. Use horizontal scrolling and the full-value field for long values; rounded display cells are not the stored numbers.
  2. Use the candidate viewer’s export action to preserve typed results: table CSV files, fitted Newick .nwk files, text/reproducibility records, plot PNG files and the artifact manifest. The native reproducibility record is JSON text within the exported record; hosted candidate output also carries complete structured JSON. Retain each interface’s actual package rather than assuming identical filenames.
  3. Open a numerical table and fitted tree outside the viewer. Check all model rows, complete branch and site maps, input identities, parameter counts and settings. Newick export retains 17 significant digits. A drawn tree or screenshot cannot replace numerical branch data.
  4. For resampling, retain every replicate’s fit, exact seed and source-column indices where applicable. Preserve partition definitions, coding code/frame, rate categories and bounds when selected. Keep the raw and prepared input identities with the run.
  5. Check manifest filenames, byte counts and SHA-256 digests. An integrity manifest certifies file identity, not scientific validity. Extremely small float64 weights may underflow to zero; inspect score differences and, where reported, log weights and underflow flags before interpreting them.

Worked examples describe particular inputs, methods and recorded results. Check current availability and input requirements before planning an evaluation.

← Back to the learning center