BAli-Phy User's Guide v4.3

We typically run BAli-Phy on workstations with at least 16 GB of RAM and 4 cores. More cores will allow you to run more MCMC chains at once, and more RAM will allow you to run larger data sets. However, it is often easier and faster to run BAli-Phy on a (Linux) computing cluster, if you have one available.

If you have previously installed bali-phy, you do not have to remove the old version before installing the new version. Simply follow the installation instructions for the new version. If you are manually adding the new version of bali-phy to your PATH, just make sure that the new version comes before the old version in the PATH, or remove the old version from the PATH.

To remove an older version, delete its bali-phy-oldversion directory and remove that version from PATH. User-installed packages and the compiled-module cache are stored separately, normally under %LOCALAPPDATA%\bali-phy on native Windows or ~/.local/share/bali-phy on Unix-like systems. They are shared between installed versions and are not removed when you delete a program directory.

The preferred way to run BAli-Phy on Windows is through Windows Subsystem for Linux (WSL2). This lets you follow the Linux commands in this guide, including commands for running several analyses and processing their output. Alternatively, install the native Windows package and run it from PowerShell or Command Prompt.

PS> New-Item -ItemType Directory -Force "$HOME\Applications"
PS> tar -xf "$HOME\Downloads\bali-phy-4.3-win64.tar.gz" -C "$HOME\Applications"
PS> & "$HOME\Applications\bali-phy-4.3\bin\bali-phy.exe" --version

PS> $env:Path = "$HOME\Applications\bali-phy-4.3\bin;$env:Path"

C:\> set "PATH=%USERPROFILE%\Applications\bali-phy-4.3\bin;%PATH%"

Bioconda and Homebrew both provide convenient ways to install BAli-Phy. Bioconda is a natural choice if you already use Conda for bioinformatics software or want to keep BAli-Phy and its dependencies in a separate environment. Homebrew is a natural choice if you already use it to manage macOS software.

% conda create -n bali-phy bali-phy
% conda activate bali-phy

% bali-phy --version

% xcode-select --install

% brew install brewsci/bio/bali-phy

% bali-phy --version

% xattr -d com.apple.quarantine "$HOME/Downloads/bali-phy-4.3-mac-arm64.tar.gz"

% xattr -d com.apple.quarantine "$HOME/Downloads/bali-phy-4.3-mac-intel64.tar.gz"

% mkdir -p ~/Applications
% cd ~/Applications
% curl -LO https://github.com/bredelings/BAli-Phy/releases/download/4.3/bali-phy-4.3-mac-arm64.tar.gz
% tar -zxf bali-phy-4.3-mac-arm64.tar.gz

% mkdir -p ~/Applications
% cd ~/Applications
% curl -LO https://github.com/bredelings/BAli-Phy/releases/download/4.3/bali-phy-4.3-mac-intel64.tar.gz
% tar -zxf bali-phy-4.3-mac-intel64.tar.gz

% ~/Applications/bali-phy-4.3/bin/bali-phy --version

% brew install r

The PATH environment variable lists the directories in which your terminal searches for commands. Once the BAli-Phy bin directory is on PATH, you can type bali-phy, bpy-summarize, and the other command names without specifying their complete locations.

%USERPROFILE%\Applications\bali-phy-4.3\bin

PS> bali-phy --version

% export PATH=$HOME/Applications/bali-phy-4.3/bin:$PATH

% setenv PATH $HOME/Applications/bali-phy-4.3/bin:$PATH

% bali-phy --version

First copy an example alignment into a new working directory:

% cp ~/Applications/bali-phy-4.3/share/doc/bali-phy/examples/5S-rRNA/25.fasta .

PS> cp "$HOME\Applications\bali-phy-4.3\share\doc\bali-phy\examples\5S-rRNA\25.fasta" .

C:\> copy "%USERPROFILE%\Applications\bali-phy-4.3\share\doc\bali-phy\examples\5S-rRNA\25.fasta" .

If a package manager installed BAli-Phy elsewhere, copy examples/5S-rRNA/25.fasta from its documentation directory instead. Then run these commands in the same terminal:

% bali-phy --version
% draw-tree --help
% bpy-summarize --help
% bali-phy-pkg --help
% bali-phy 25.fasta --iterations=200
% bali-phy 25.fasta --iterations=200
% bpy-summarize 25-1 25-2

Then check that the file Results/index.html exists and can be opened in a web browser.

The main output from bpy-summarize is the report Results/index.html. Open this report in a web browser first: it contains summary trees and alignments, mixing diagnostics, and links to the underlying result files. The programs below are optional and are useful when you want to inspect those files more closely.

To check the viewers with the tutorial output, open 25-1/C1.log and 25-2/C1.log together in Tracer, Results/c50.PP.tree in a tree viewer, and Results/P1.consensus.pd-wsum.fasta in an alignment viewer.

Consensus trees produced by bpy-summarize, such as Results/c50.PP.tree, can be opened directly in TreeViewer, FigTree, or iTOL.

This example uses the 25-taxon 5S-rRNA data set. Create a new working directory and change into it. Using a new directory ensures that the four runs below save their results in directories named 25-1 through 25-4. Copy the example data into this directory:

% cp ~/Applications/bali-phy-4.3/share/doc/bali-phy/examples/5S-rRNA/25.fasta .

PS> cp "$HOME\Applications\bali-phy-4.3\share\doc\bali-phy\examples\5S-rRNA\25.fasta" .

C:\> copy "%USERPROFILE%\Applications\bali-phy-4.3\share\doc\bali-phy\examples\5S-rRNA\25.fasta" .

If a package manager installed BAli-Phy elsewhere, copy examples/5S-rRNA/25.fasta from its documentation directory instead.

Start four independent chains:

% bali-phy 25.fasta -S "GTR +> ASRV.Free +> Covarion.Huelsenbeck02" --iterations=1000 &

bali-phy 25.fasta -S "GTR +> ASRV.Free +> Covarion.Huelsenbeck02" --iterations=1000

Once the runs have produced some samples, you can summarize their results while they are still running. Use an available terminal in the directory containing 25.fasta, opening another terminal if necessary:

% bpy-summarize 25-1 25-2 25-3 25-4

You can rerun this command as more samples accumulate. Open or reload Results/index.html in a web browser to inspect the updated report. The report contains tree and alignment summaries, convergence warnings, mixing diagnostics, and links to the underlying result files.

These short runs demonstrate how to run an analysis and summarize its results. They stop after 1,000 iterations, but reaching that limit does not establish convergence. Check the report's warnings and diagnostics before interpreting the estimates. See Section 12, “Convergence and Mixing: Is it done yet?” for help deciding whether longer runs are needed. That section also explains how to compare chains using tools such as Tracer.

See Section 4.2, “Posterior summaries” for details of the output files and Section 2.8, “Optional programs for inspecting results” for optional programs to inspect trees, alignments, and parameter traces.

When estimating the alignment, BAli-Phy accepts unaligned sequences: you do not need to align them first. If the input already contains an alignment, BAli-Phy removes the gaps and estimates a new alignment as part of the analysis.

When keeping the alignment fixed with -I none, provide aligned sequences of equal length, including gaps. Their columns define the alignment used throughout the analysis.

BAli-Phy can read in sequences and alignments in both FastA and PHYLIP formats. Filenames for FastA files should end in .fasta, .mpfa, .fna, .fas, .fsa, or .fa. Filenames for PHYLIP files should end in .phy. If one of these extensions is not used, then BAli-Phy will attempt to guess which format is being used.

FASTA format prefixes sequence names with ">":

>human       this is a comment and is not part of the sequence name
CTGACTCCTGAGGAGAAGTCTGCCGTTACTGCCCTGTGGGGCAAGGTGAACGTGGATGAA
GTTGGTGGTGAGGCCCTGGGCAGGCTGCTGGTGGTCTACCCTTGGACCCAGAGGTTCTTT
>tarsier     this is also a comment
CTGACTGCTGAAGAGAAGGCCGCCGTCACTGCCCTGTGGGGCAAGGTAGACGTGGAAGAT
GTTGGTGGTGAGGCCCTGGGCAGGCTGCTGGTCGTCTACCCATGGACCCAGAGGTTCTTT
>bushbaby
CTGACTCCTGATGAGAAGAATGCCGTTTGTGCCCTGTGGGGCAAGGTGAATGTGGAAGAA
GTTGGTGGTGAGGCCCTGGGCAGGCTGCTGGTTGTCTACCCATGGACCCAGAGGTTCTTT
>hare
CTGTCCGGTGAGGAGAAGTCTGCGGTCACTGCCCTGTGGGGCAAGGTGAATGTGGAAGAA
GTTGGTGGTGAGACCCTGGGCAGGCTGCTGGTTGTCTACCCATGGACCCAGAGGTTCTTC

If the sequence-name line contains a space, BAli-Phy treats everything after the space as a comment.

To analyze only part of the input, append a range to the filename:

% bali-phy sequences.fasta:1-30,90-100

This selects positions 1–30 and 90–100 in the input sequences, counting from 1 and including both endpoints. For aligned input, these positions are alignment columns. Selection takes place before any gaps are removed for alignment estimation.

Sensible defaults are supplied for command line options that are not specified. For example, if sequences.fasta contains DNA sequences, then

% bali-phy sequences.fasta

% bali-phy sequences.fasta -A DNA -S TN93 -I RS07

% bali-phy sequences.fasta -S "LG +> ASRV.Gamma +> Inv"

Use bali-phy --help or bali-phy help for commonly used options. For more specialized options, use bali-phy help advanced, bali-phy help expert, or bali-phy help developer. Each level includes the options from the preceding levels.

The same BAli-Phy options, such as -S and --iterations, work in all the shells covered here. Native Windows also accepts slash forms such as /iterations:1000, but this guide uses the shared forms.

Put quotes around model expressions containing spaces, parentheses, or square brackets. Single quotes work in Unix shells and PowerShell; in Command Prompt, use double quotes instead. For example:

bali-phy sequences.fasta -S 'HKY85(kappa=2)'

bali-phy sequences.fasta -S "HKY85(kappa=2)"

Models containing double-quoted strings are easier to specify in a configuration file. For DNA data, save this line in model.config:

:smodel GTR(pi={"A":0.1, "C":0.2, "G":0.3, "T":0.4})

Then use the same command in any shell:

bali-phy sequences.fasta -c model.config

The configuration file contains the model expression itself, without outer shell quotes. See Section 3.5, “Configuration files” for other configuration options.

For these expressions on the command line, Unix shells and PowerShell 7.3 or later with its default native argument handling preserve double quotes inside single quotes. Windows PowerShell 5.1 and earlier PowerShell 7 releases can lose those inner quotes when passing arguments to a program; use the configuration-file form there or in Command Prompt. To check your PowerShell version, run $PSVersionTable.PSVersion.

For interactive PowerShell commands on Windows, cp is a short form of Copy-Item, and start is a short form of Start-Process. To open a report, use:

PS> start Results\index.html

C:\> start "" Results\index.html

Put paths containing spaces inside double quotes. Keep double quotes around PowerShell paths using $HOME or other variables so their values are expanded. Invoking an executable by a quoted path in PowerShell also requires & before the path, as in the installation instructions.

In addition to using the command line, you may specify options and model-language definitions in a configuration file. Configuration options use the long form of command-line options. Each option is given on its own line using the syntax ":option value" instead of "--option value". The value can be blank if the option does not take an argument. The align option indicates sequence files. Other non-comment lines are combined into model-language definitions and may occur before, between, or after option lines. Lines whose first non-whitespace character is # are comments, and blank lines are ignored.

For example, consider the following configuration file:

# sequence data for 3 genes/partitions
:align ITS1.fasta
:align 5.8S.fasta
:align ITS2.fasta

# linked substitution model for 1st and 3rd partition
:smodel 1,3:TN93  +>  ASRV.Free(n=3)

# substitution model for 2nd partition
:smodel 2:TN93

# indel model for second partition
:imodel 2:none

# linked scale for 1st and 3rd partition
:scale 1,3:

# choose a name for output directories
:name ITS-analysis1

Configuration files are specified with the -c configuration_file option. For example:

% bali-phy -c config.txt

% bali-phy -c config.txt --name ITS-analysis1b

BAli-Phy does not read an implicit configuration file from your home directory; pass each configuration file explicitly with --config.

Running bali-phy on a computing cluster is not necessary, but can speed up the analysis dramatically. This is because a cluster allows you to run several independent MCMC chains simultaneously and pool the resulting samples. You can run multiple chains simultaneously simply by starting several different instances of bali-phy. Each instance of bali-phy runs only one chain and does not require using MPI or special command-line options.

This approach to parallel computation is sometimes more efficient than MCMCMC-based parallelism involving heated chains. It is equivalent to running MCMCMC with no temperature difference between chains, with the exception that it allows results from all chains to be used, instead of just results from the single "cold" chain. Thus, if you run 10 independent chains in parallel, then you may gather samples 10 times faster than a single chain.

Bayesian inference programs must run for many iterations to complete an analysis. A data set is considered "too large" if waiting for it to complete takes "too long".

The bpy-summarize script summarizes the posterior samples to create posterior summaries for the alignment, tree, and parameters. It creates an HTML page Results/index.html that summarizes the posterior distribution.

You may run bpy-summarize inside the output directory, like this:

% bpy-summarize --skip=iterations

% bpy-summarize --skip=iterations directory-1/ directory-2/

% bpy-summarize --skip=iterations --verbose

This section is primarily about summarizing the posterior to extract estimates from posterior samples, not about assessing convergence. See Section 12, “Convergence and Mixing: Is it done yet?” for methods of determining effective sample sizes, and for checking mixing and convergence.

% trees-consensus dir-1/C1.trees dir-2/C1.trees > c50.PP.tree

% trees-consensus -s 1000 -x 10 dir-1/C1.trees dir-2/C1.trees > c50.PP.tree

% trees-consensus -s20% -x10 --consensus=0.66 dir-1/C1.trees dir-2/C1.trees > c66.PP.tree

% trees-consensus -s20% -x10 dir-1/C1.trees dir-2/C1.trees --consensus=0.66:c66.PP.tree

% trees-consensus -s20% -x10 dir-1/C1.trees dir-2/C1.trees --consensus=0.5:c50.PP.tree,0.66:c66.PP.tree

% tree-tool c50.PP.tree --strip-internal-names > c50.tree

% trees-consensus --skip=burnin dir-1/C1.trees dir-2/C1.trees --greedy-consensus=greedy.tree

% trees-consensus --skip=burnin dir-1/C1.trees dir-2/C1.trees --map-tree=MAP.tree

% trees-bootstrap dir-1/C1.trees dir-2/C1.trees

% statreport dir-1/C1.log dir-2/C1.log > Report 

% cut-range dir-1/C1.Pp.fastas dir-2/C1.Pp.fastas --skip=burn-in | alignment-chop-internal --tree tree | alignment-max > Pp-max.fasta

% cut-range dir-1/C1.Pp.fastas dir-2/C1.Pp.fastas --skip=burn-in | alignment-chop-internal --tree tree  | alignment-gild alignment.fasta tree  > alignment-AU.prob
% alignment-draw alignment.fasta --AU alignment-AU.prob > alignment-AU.html

The default tree distribution for BAli-Phy is unrooted. This works fine for reversible substitution models, but if you specify a non-reversible substitution model then you will get an error:

bali-phy: Error! phyloCTMC: the tree is unrooted, but the model is not reversible!

--tree '~UniformRootedTree(taxa)'

The default substitution model for DNA and RNA is TN93.

GTR(pi={"A":0.1, "C":0.2, "T":0.3, "G":0.4})

GTR(pi=Frequencies.uniform)

The default substitution model for proteins is LG.

WAG +> F({"A":0.047, "R":0.19,...})

WAG +> F(pi=Frequencies.uniform)

WAG +> FE

The doublets alphabet consists of 16 RNA dinucleotides. It is used to model RNA stems, where two nucleotides matched in the RNA secondary structure are highly correlated.

The default substitution model for doublets is TN93_sym +> x2_sym +> F.

HKY85(pi={"A":0.1, "C":0.2, "T":0.3, "G":0.4}) +> x2

HKY85_sym +> x2_sym +> F({"AA":0.01, "AC":0.01, "AG":0.01, "AU":0.22, "CA":0.01, "CC":0.01, "CG":0.22, "CU":0.01, "GA":0.01, "GC":0.22, "GG":0.01, "GU":0.01, "UA":0.22, "UC":0.01, "UG":0.01, "UU":0.01})

The triplets alphabet is similar to the codons alphabet, except that stop codons are included. Unlike the codons alphabet, the triplets alphabet has no knowledge of the genetic code.

The default substitution model for triplets is TN93 +> x3.

HKY85(pi={"A":0.1, "C":0.2, "T":0.3, "G":0.4}) +> x3

HKY85_sym +> x3_sym +> F(pi=F1x4)            // nucleotide frequencies are estimated

HKY85_sym +> x3_sym +> FE

Branch lengths for triplet models measure expected nucleotide substitutions per nucleotide site, as for DNA/RNA models. A mutation changing two or three nucleotides counts as two or three substitutions, respectively.

The default substitution model for codons is GY94.

GY94(pi={"AAA":0.01, "AAC":0.02,...})
MG94(pi={"A":0.1, "C":0.2, "T":0.3, "G":0.4})

GY94(pi=F1x4)              // nucleotide frequencies are estimated

% bali-phy sequence-file -S GY94 -A Codons
% bali-phy sequence-file -S GY94 -A "Codons(RNA)"

% bali-phy sequence-file -S GY94 -A "Codons(DNA,mt-vert)"
% bali-phy sequence-file -S GY94 -A "Codons(,mt-vert)"

Complex substitution models in BAli-Phy are constructed as mixtures of reversible CTMC models that run at different rates (e.g. \(\Gamma_4 + Inv\)) or have different parameters (e.g. an M2a codon model).

These models attempt to model the fact that evolutionary rates may change over time within a single column. These models are sometimes called "covarion" models, based on the idea that changes in rate might be caused by changes in an unspecified covarying site.

These models are "Markov modulated" models that create multiple different states for each letter by augmenting each letter with some unobserved hidden state. They attempt to model the fact that substitution processes might not be Markov on the letters, but might become more Markov given the hidden state.

Models, probability distributions, and functions are treated the same in BAli-Phy because all of them have parameters or arguments. Parameters have names in BAli-Phy. Parameter values are specified using parentheses as follows:

HKY85(kappa=2)            // model
log(x=2)                  // function
Normal(mean=0,sigma=1)    // probability distribution

HKY85(2)
log(2)
Normal(0,1)

Normal(0,sigma=1)   // OK
Normal(mean=0,1)    // not OK

% bali-phy help HKY85

% bali-phy file.fasta -S 'HKY85(kappa=2)'
% bali-phy file.fasta -S 'mixture([TN93,HKY85(2)])'

Models in phylogenetics literature are often combined using +. For example, the model WAG + F + G4 + I starts with the WAG amino-acid model, and places several modifiers, like " + G4" on the right.

BAli-Phy follows this convention by treating A +> B as an abbreviation for B(A). When there are multiple '+>' symbols they associate to the left, so that A +> B +> C is understood to mean (A +> B) +> C, which is equivalent to C(B(A)). For example:

HKY85 +> ASRV.Gamma              // rewritten to ASRV.Gamma(HKY85)
HKY85 +> Inv                      // rewritten to Inv(HKY85)
WAG +> F                          // rewritten to F(WAG)
WAG +> F +> ASRV.Gamma +> Inv    // rewritten to Inv(ASRV.Gamma(F(WAG)))

Some function arguments have default values. For example, the ASRV.Gamma parameter n has a default value of 4. Thus the following are equivalent:

HKY85 +> ASRV.Gamma(n=4) +> Inv
HKY85 +> ASRV.Gamma +> Inv

When the default value is random, then the argument has a default prior. For example, the kappa parameter of HKY85 has a default value of ~LogNormal(log(2),0.25), so the following are equivalent:

HKY85(kappa~LogNormal(log(2),0.25))
HKY85

Every function has a result type, as well as an argument type for each argument. The argument type specifies what kind of arguments are acceptable, and the result type specifies what kind of result the function produces. Types include Int for integers, Double for double-precision floating point numbers, and String for text strings. Integer arguments are implicitly converted to Double when the argument type is Double.

Some types contain parameters. For example List<Int> indicates a list of integers and List<Double> indicates a list of real numbers. In order to indicate a list of unknown type, we use a type variable a and write List<a>. Type variables always begin with a lower-case letter. They are able to match any specific type, and their value is found by pattern-matching. For example, the function x+y takes two arguments of type a and has a result of type a. Thus:

1 + 2        // arguments are a=Int, so result is of type Int
1.0 + 2.0    // arguments are a=Double, so result is of type Double

Types for components of substitution models are often parameterized by type of the alphabet. For example, HKY85 has a result type of CTMC<a>, where a could be DNA or RNA. The use of alphabet types in substitution models prevents combining substitution models with mismatched alphabets.

You should analyze multiple genes under different evolutionary models by putting each one it its own data partition. Placing different genes in different partitions means that their alignments vary independently. It also prevents sequences in one gene from being aligned against sequences in another gene.

Different partitions share the same tree topology and a common set of unscaled branch lengths. However, branch lengths are scaled by a different factor in each partition, since some genes may evolve faster than others.

To put different genes in different partitions, you can place the sequences from each partition in a different FASTA or Phylip file. The sequence names in files for all partitions should be the same.

% bali-phy gene1.fasta gene2.fasta

% bali-phy sequences.fasta:3-350 sequences.fasta:351-570

By default, each partition will have its own substitution model, insertion/deletion model, and scaled tree length. For example, even if all partitions are assigned a TN93 substitution model, their base frequencies will all be estimated independently. When parameters are estimated separately for two partitions, we say that the parameters for those partitions are "unlinked".

A substitution model or insertion-deletion model that is specified without qualification will apply to every partition. However, each partition will receive its own copy of each model with unlinked parameter values:

% bali-phy sequence-file1 sequence-file2 -S TN93 -I RS07

You can select partition-specific values for 4 options: -S, -I, -A, and --scale. For example, to specify different substitution models but the same alphabet:

% bali-phy sequence-file1 sequence-file2 -S 1:TN93 -S 2:GTR -A DNA

You can fix the alignment and ignore insertion/deletion information in one partition, while allowing the alignment to vary and using insertion/deletion information in another partition:

% bali-phy sequence-file1 sequence-file2 -I 2:none

Specifying -I none fixes the alignment in all partitions:

% bali-phy sequence-file1 sequence-file2 -I none

You can also specify that two partitions share a single copy of a single substitution model or indel model. For example, if two partitions both have a TN93 model, linking these models would force the partitions to have the same nucleotide frequencies and substitution rates. Linking partitions reduces the number of parameters that need to be estimated, and also pools information between the partitions:

% bali-phy sequence-file1 sequence-file2 -S 1,2:TN93 -I 1,2:RS07

% bali-phy sequence-file1 sequence-file2 --scale 1,2:

The --link command is provided to allow specifying a model for each partition separately, and then afterwards choose which partitions to link.

% bali-phy sequence-file1 sequence-file2 -S 1:TN93 -S 2:TN93 --link=1,2 -t
% bali-phy sequence-file1 sequence-file2 -S TN93             --link=1,2 -t   

% bali-phy sequence-file1 sequence-file2 -S 1:TN93 --link=1,2 -t
bali-phy: Error! Partitions 1 and 2 cannot be linked because they have differing values 'TN93' and ''

You can also specify which of the 3 attributes "smodel", "imodel", and "scale" are being linked:

% bali-phy sequence-file1 sequence-file2 --link=1,2:smodel,scale -t

Start with a model that uses one ω for all sites and branches:

% bali-phy coding.fasta -A Codons -S 'GTR +> MNM +> MutSelAA +> dNdS'

GTR describes nucleotide mutation. MNM allows mutations to change one, two, or three nucleotides within a codon. MutSelAA estimates amino-acid preferences shared across sites, and dNdS adds the ω multiplier. Here ω describes nonsynonymous change beyond the effect of those preferences; it is not a raw ratio of observed changes. Setting ω to one leaves those preferences intact, rather than removing all selection.

MutSelAA assigns equal fitness to synonymous codons. To allow separate codon preferences, use MutSel instead:

% bali-phy coding.fasta -A Codons -S 'GTR +> MNM +> MutSel +> dNdS'

Both models estimate preferences shared across sites. The examples below use MutSelAA because MutSel currently mixes substantially more slowly and thus needs longer runs for reliable estimates. See Section 10.6, “Codon frequencies and the background model” and Section 10.8, “Multinucleotide mutations” for the reasons behind these background-model choices.

The default LogNormal(0,1) prior gives equal probability to ω below and above one. To restrict ω to values below one, use dNdS(omega=~Uniform(0,1)). Run independent chains, check convergence, and summarize them as in Section 4.2, “Posterior summaries”. Read the ω estimate and credible interval in the parameter report. A single ω cannot identify selected sites, and averaging across the sequence can hide selection affecting only a few sites.

A site model allows ω to differ among codon sites. Each site belongs to an unknown category with its own ω, which applies on every branch. The analysis estimates the category proportions, any unknown ω values, and each site's probability of belonging to each category.

M1a has two categories: ω below one and ω equal to one. M2a adds a category with ω above one. M7 uses a beta distribution for ω between zero and one, approximated by a finite set of categories. M8a adds a category with ω equal to one; M8 instead adds one with ω above one.

Use M2a_test to compare M1a with M2a, or M8a_test to compare M8a with M8. In each comparison, the selection hypothesis includes a category with ω above one; the null hypothesis does not. Each command allows the chain to move between the two hypotheses:

% bali-phy coding.fasta -A Codons \
  -S '|w:GTR +> MNM +> MutSelAA +> dNdS(omega=w)| +> M2a_test'
% bali-phy coding.fasta -A Codons \
  -S '|w:GTR +> MNM +> MutSelAA +> dNdS(omega=w)| +> M8a_test'

The expression |w:...| builds the codon model for a supplied ω value. The site model supplies these values and their priors, rather than using the single-ω default prior. The nucleotide and amino-acid preference parameters are shared across categories. To include positive selection throughout the analysis without comparing hypotheses, replace M2a_test with M2a, or M8a_test with M8.

After summarizing independent runs with bpy-summarize, open the positive-selection section of Results/index.html. First read the posterior probability of the selection hypothesis. These tests give the two hypotheses equal prior probability by default, so posterior odds equal the Bayes factor in favor of selection. For example, posterior probability 0.9 gives odds of 9:1, or a Bayes factor of 9. These are Bayesian comparisons, not PAML likelihood-ratio tests or p-values; results depend on the parameter priors as well as the data.

In the parameter log, posSelection records whether the selection hypothesis is active. Averaging PrPosSelection can estimate its posterior probability more accurately than averaging this indicator. statreport uses LogOddsPosSelection to summarize posterior log odds accurately even when probabilities are near zero or one. Do not average LogOddsPosSelection directly.

Then examine the site table. Its probabilities describe whether each codon has ω above one. Overall posterior includes uncertainty about whether positive selection is present in the sequence at all. Posterior with selection uses only samples under the selection hypothesis. For example, if a codon has probability 0.95 of being selected under that hypothesis, but the hypothesis has probability 0.20, its overall probability is 0.19. Mean dNdS is an estimated ω value, not a selection probability.

In the alignment viewer, choose posSelection. The same two summaries are called Model-averaged and Conditional on positiveSelectionInModel.

With an uncertain alignment, a displayed column can contain letters with different probabilities. For each column, the main table reports the letter most likely to be positively selected, assuming the selection hypothesis. If that estimate is unavailable, it uses the highest overall probability instead. The sequence name and position identify the chosen letter. The table shows at most 20 columns passing the display threshold; All rows (TSV) exports a row for every column, regardless of the threshold. These thresholds control what is displayed; they are not significance tests.

Estimates based on few samples under the selection hypothesis may be imprecise. With no such samples, conditional estimates are unavailable.

A branch model assigns branches to numbered categories, each with one ω shared across all sites. Several branches can share a category. Fix the topology and label branches in a Newick tree containing all your sequences. Save it as branches.tree. This four-sequence example is an unrooted tree and labels the branch separating A and B from C and D:

((A,B):[&foreground=1],C,D);

The label follows a colon because it belongs to the branch, not the node. This assigns that branch to category 1; unlabelled branches use category 0. Branch lengths are still estimated:

% bali-phy coding.fasta -A Codons --fix topology=branches.tree \
  -S '|w:GTR +> MNM +> MutSelAA +> dNdS(omega=w)| +> BranchModel'

The Branch models report summarizes the category ω values. A category number identifies a shared parameter, not a particular branch. Different branch groups can have different ω values while all remain below one.

To test whether the labelled branch needs a separate ω, change its label to [&foreground={0,1}] and replace BranchModel with BranchModel_test in the command. In this example, hypothesis 0 uses category 0 everywhere; hypothesis 1 uses category 1 on that branch.

Each entry in the label gives that branch's category under a hypothesis: the first entry is for hypothesis 0, the second for hypothesis 1, and so on. Thus, [&foreground={1,0}] reverses the assignments on that branch. Unlabelled branches use category 0 under every hypothesis.

The report gives the probability of each hypothesis and separate ω summaries using only samples from that hypothesis. These tables omit unused categories. Summaries across all hypotheses include prior draws whenever a category is unused. This tests whether ω differs among branches, not specifically whether it exceeds one.

A branch-site model distinguishes branches of interest, called foreground branches, from the remaining background branches. Some sites can have ω at or below one on background branches but ω above one on foreground branches. BranchSite compares this with a null hypothesis in which those sites have ω equal to one on the foreground branches. The analysis estimates which sites are affected and the ω shared by these sites on foreground branches.

Use a fixed topology with [&foreground=1] on foreground branches, as in the first tree above, not the {0,1} label used for BranchModel_test. Leave background branches unlabelled:

% bali-phy coding.fasta -A Codons --fix topology=branches.tree \
  -S '|w:GTR +> MNM +> MutSelAA +> dNdS(omega=w)| +> BranchSite'

Read model support and the two posterior views as for site models. In the alignment viewer, foreground-posSelection gives selection probabilities and foreground-dNdS gives ω estimates on foreground branches. background-posSelection and background-dNdS describe the same sites on background branches. Fixing the topology does not fix the alignment. Choose foreground branches from the biological question before examining selection results. Searching many branch sets and reporting only the strongest result can exaggerate the evidence.

A false positive is an inference of positive selection when it is absent. Evidence for selection depends on which changes the model expects without it. If the background model assigns those changes too little probability, increasing ω can partly compensate for the error. This can improve the fit without correctly explaining the process: ω increases only nonsynonymous rates, rather than correcting the background model.

A better background can strengthen or weaken evidence for selection. The following subsections explain the background choices in the examples, how alignment uncertainty is handled, and how to add synonymous rate variation. These choices improve robustness to particular sources of systematic error; they do not prevent all false positives.

Goldman–Yang (GY) models weight a change by the frequency of the destination codon. Muse–Gaut (MG) models instead use the frequency of the nucleotide being introduced. For example, for AGA → AGT, a GY model uses the frequency of AGT, whereas an MG model uses the frequency of T. With codon frequencies calculated as products of nucleotide frequencies, the GY rate therefore includes factors for the unchanged A and G as well. These different backgrounds can lead to different conclusions about selection, including false positives or missed selection (Yap et al. 2010).

The mutation–selection models used here separate nucleotide mutation biases from amino-acid or codon preferences. Rates depend on mutation rates and fitness differences, rather than directly multiplying by the destination-codon frequency. This separates two causes of unequal codon usage and motivates using MutSelAA or MutSel in the examples. It does not remove all assumptions about the background process.

Aligning unrelated codons can create apparent amino-acid substitutions and spurious evidence of selection. BAli-Phy jointly estimates alignment and selection, allowing each hypothesis to be evaluated over plausible alignments instead of treating one alignment as certain. In branch-site simulations, this removed the excess false positives caused by alignment error while detecting selection nearly as often as when the true alignment was known (Redelings 2014).

The examples above already sample alignments. Omit -I none to retain this behavior. Fixing the topology for a branch or branch-site analysis does not fix the alignment. See Section 9, “Character properties” for interpreting site results across sampled alignments.

Two or three differences within a codon can arise from one mutation. A model allowing only single-nucleotide mutations must explain them through several events, potentially exaggerating the evidence for positive selection. This effect has been demonstrated for branch-site tests (Venkat et al. 2018).

MNM, already included in the examples, estimates rates of double and triple mutations as part of the background model, with little added computational cost. It allows these events within codons, not across codon boundaries. Replace it with x3 to restrict mutations to single-nucleotide events, as in Section 10.10, “Comparing with CODEML”.

Codons can differ in their baseline substitution rates, affecting synonymous as well as nonsynonymous changes. This is distinct from variation in ω. Ignoring baseline-rate variation can distort evidence for selection (Rubinstein et al. 2011). For comparable nonsynonymous change, a lower inferred baseline rate can strengthen evidence for selection; a higher baseline can weaken it.

Append ASRV.Gamma to add a gamma distribution of rates among codons:

% bali-phy coding.fasta -A Codons \
  -S '|w:GTR +> MNM +> MutSelAA +> dNdS(omega=w)| +> M8a_test +> ASRV.Gamma'

Each codon now has a rate category as well as an ω category. Its rate multiplier scales synonymous and nonsynonymous rates together without changing ω, and applies on every branch. Append ASRV after the completed selection model, outside |w:...|, as shown. The same distribution of rate multipliers applies to all ω categories and both hypotheses. This construction works with the other models above.

ASRV.Gamma approximates the gamma distribution with four categories by default. Alternatively, ASRV.Free estimates category rates and proportions without requiring a gamma shape. Start with two categories:

% bali-phy coding.fasta -A Codons \
  -S '|w:GTR +> MNM +> MutSelAA +> dNdS(omega=w)| +> M8a_test +> ASRV.Free(n=2)'

The number of categories is fixed by n, not inferred. If the two estimated rates are clearly separated, try n=3 and check whether the data distinguish a third rate and mixing remains adequate. The default of four free-rate categories introduces more parameters. Both ASRV extensions increase computation by combining rate categories with selection categories, so they are optional rather than part of the baseline examples.

The model families above correspond to analyses available in PAML's CODEML program, but the background choices differ. Many earlier analyses used GY94-type models. To compare with these, use GY94 and match their codon-frequency specification. For example, GY94(pi=F3x4) uses separate nucleotide frequencies at each codon position. In a site model, supply ω explicitly:

% bali-phy coding.fasta -A Codons -I none --fix topology=tree.tree \
  -S '|w:GY94(omega=w,pi=F3x4)| +> M8a_test'

The Álvarez-Carretero et al. (2023) tutorial instead uses FMutSel (CodonFreq=7). The corresponding base model replaces GTR with HKY85, MNM with x3, and MutSelAA with MutSel:

% bali-phy coding.fasta -A Codons -I none --fix topology=tree.tree \
  -S 'HKY85 +> x3 +> MutSel +> dNdS'

Using MutSelAA instead corresponds to the FMutSel0 preference model (CodonFreq=6). CODEML also supports GTR mutation rates through hkyREV=1, but its codon models restrict instantaneous changes to a single nucleotide; they do not include the MNM extension.

Use the same alignment, topology, genetic code, foreground branches, and retained sites in both programs. Omit ASRV if the CODEML analysis did not include it. Matching the model family does not make the analyses identical.

The tutorial uses observed codon frequencies (estFreq=0), whereas the examples here infer preferences. BAli-Phy also uses priors, and the programs approximate beta mixtures differently. BAli-Phy reports posterior estimates and Bayesian model support, not CODEML's maximum-likelihood estimates, likelihood-ratio tests, or empirical-Bayes site probabilities.

A CODEML result with p < 0.05 does not imply a posterior probability of positive selection above 95%. A p-value measures how unusual the test statistic would be under the hypothesis of no positive selection; it is not the probability that this hypothesis is true. A posterior probability measures support for a hypothesis given the data, model, and priors. Thus, CODEML can give a significant result when BAli-Phy reports posterior support below 95%, even with the same likelihood function. These are different criteria for assessing evidence, not equivalent thresholds.

For some models, CODEML uses Bayes empirical Bayes (BEB) to account for uncertainty in the parameters describing variation in ω among sites when calculating site probabilities (Yang, Wong, and Nielsen 2005). Other parameters, such as branch lengths, remain fixed at their maximum-likelihood estimates. BAli-Phy averages over uncertainty in all parameters being inferred, for site probabilities and other estimates alike, so no separate BEB correction is needed. Quantities fixed by the user, such as the topology, remain fixed.

BAli-Phy can reconstruct ancestral sequences for all internal nodes. Unlike some programs, BAli-Phy explicitly infers the presence and absence of characters in ancestral sequences. This means that if the ancestral sequence has no character for a column, the reconstructed ancestor will have a gap there. BAli-Phy reconstructs ancestors for fixed-alignment partitions as well as variable-alignment partitions, but it won't write out fixed alignment samples unless you add the flag --set write-fixed-alignments=true. Additionally, if you have an ambiguous character such as N in an observed sequence BAli-Phy will impute this character.

BAli-Phy can reconstruct ancestral sequences for a given tree topology and (leaf sequence) alignment. This is similar to the ancestor-reconstruction that is usually done for fixed-alignment analyses. However, it is not quite the same, because of uncertainty in the tree and the alignment. When computing the probability of an ancestral residue, this summary averages over uncertainty in the topology, the alignment, and the ancestral state itself.

This example uses the original run files and discards iterations 0–999 from each run. Replace 1000 with a burn-in appropriate for your runs. For older PowerShell versions, see Section 3.4.1, “Pipelines and redirected output” before running these pipelines. First, construct the tree topology to annotate:

% trees-consensus --skip=1000 dir-1/C1.trees dir-2/C1.trees | tree-tool - --strip-internal-names --name-all-nodes > c50.tree

Construct the leaf-sequence alignment to annotate using posterior decoding:

% cut-range dir-1/C1.P1.fastas dir-2/C1.P1.fastas --skip=1000 | alignment-chop-internal --tree c50.tree | alignment-max --out P1-max.fasta

Finally, reconstruct the ancestral sequences on the given tree and alignment:

% summarize-ancestors P1-max.fasta -A dir-1/C1.P1.fastas -T dir-1/C1.trees -A dir-2/C1.P1.fastas -T dir-2/C1.trees --skip=1000 -n c50.tree > P1.ancestors.fasta

The default --subsample=10 matches one saved alignment to every tenth tree before applying burn-in. Use --until to set the last retained iteration and --thin to thin the eligible pairs. Do not remove internal sequence names or tree labels from the original sample files.

For codon data, pass the same --alphabet option to alignment-max and summarize-ancestors, for example --alphabet="Codons(DNA,standard)". Use the genetic code from your analysis.

BAli-Phy uses an alignment estimate (here, P1-max.fasta) as a template to construct a consensus alignment with ancestral sequences. BAli-Phy doesn't condition on the alignment columns, because (i) many columns occur only once in a posterior sample and (ii) conditioning on the column gives too much weight to the template alignment.

Because the alignment is uncertain, residues in the same column of the template alignment may end up in different columns in an MCMC sample. Therefore, in a given MCMC sample, different leaf residues in the same column may have different ancestors at the same internal node! However, in our ancestral reconstruction, a given column may only display a single ancestral residue.

BAli-Phy addresses this problem by averaging across the different ancestral residues in each column of the template alignment. When identifying the ancestral character for column C from a sampled alignment A, we randomly select a residue in C and use it to select a column from A. This procedure has the nice property that it will yield the traditional ancestral residue prediction if the alignment column is fixed.

Ancestral sequences are written as part of the alignment matrix in each iteration. Ancestral sequences are given names starting with the letter A. For example, in the following alignment, the sequences A5, A6, and A7 are reconstructed ancestors:

>Halobacterium
-T-TAAGGCGGCCATAGCGGTGGGGTTACTCCCGTAC
>Pyrococcus
GG-TACGGCGGTCATAGCGGGGGGGCCACACCCGGTC
>Sulfolobus
GC-CCACCCGGTCACAGTGAGCGGGCAACACCCGGAC
>Homo
GTCTACGGC---CATACCACCCTGAACGCGCCCGATC
>Escherichia
TG-CCTGGCGGCCGTAGCGCGGTGGTCCCACCTGACC
>A5   
GG-CAAGGCGGCCATAGCGGGGGGGCCACACCCGGCC
>A6   
GT-CAAGGCGGCCATAGCGGGGGGGCTACACCCGGTC
>A7   
GT-CAAGGCGGCCATAGCGGGGGGGCTACACCCGGTC

Sampled alignments for the nth partition are in the file. C1.Pn.fastas.

Ancestral states in these alignments are randomly sampled from their joint posterior and do not represent the most probable ancestral state. The alignment of ancestral sequences is also inferred, so these sequences may contain gaps. The length of ancestral sequences may vary between samples when the length of the ancestral sequence is uncertain.

Each sampled alignment matrix corresponds to a tree in the file C1.trees that is written in the same iteration. This tree specifies the phylogenetic location of each ancestral sequence by labelling the internal nodes of the tree. For example, the tree below shows where the internal nodes A5, A6, and A7 are located on the tree:

(Halobacterium:0.213240,((Escherichia:0.435762,Pyrococcus:0.122678)A5:0.114725,Sulfolobus:0.427210)A6:0.042527,Homo:0.427026)A7;

While the tree is written every iteration, the alignment is only written every 10 iterations (by default) in order to save disk space. One method for extracting the trees that correspond to saved alignments is to extract every 10th tree with the program bpy-subsample:

bpy-subsample 10 < C1.trees > C1.10.trees

PS> cmd /c "bpy-subsample 10 < C1.trees > C1.10.trees"

Instead of constructing consensus ancestral sequences, you can also summarize the sampled alignments and their ancestral sequences directly. This approach involves performing a downstream analysis on each sampled (alignment,tree) pair, yielding the posterior distribution of the downstream analysis. Averaging these results then yields the posterior mean analysis result.

When this approach is feasible, it is more statistically rigorous than analyzing the consensus ancestral sequence alignment. The consensus ancestral sequence alignment does not account for uncertainty in the ancestral sequences, and is not a joint reconstruction. In contrast, analyzing the posterior samples accounts for uncertainty in the ancestral sequences, the alignment, and the tree. Furthermore, it also summarizes joint reconstructions instead of a marginal reconstruction.

In Bayesian phylogenetic analyses, the tree is not fixed. Therefore the internal node corresponding to the ancestral sequence you wish to reconstruct may not exist in every posterior sample. The standard Bayesian approach to tree uncertainty is to reconstruct the ancestor for each node by conditioning on the existence of that node in the tree. This allows the reconstructed ancestor for each node to average over uncertainty about the existence of other nodes.

BAli-Phy additionally allows the researcher to condition on branches, since a branch condition is less restrictive. BAli-Phy does not run a separate MCMC chain with a tree constraint for each node, but instead performs conditioning by selecting samples from a single run that satisfy the condition.

% extract-ancestors -A post-1.fastas -T post-1.trees -A post-2.fastas -T post-2.trees --subsample=1 -n c50.tree -g c50.tree > P1.ancestors.fastas

name = taxon1 taxon2 ... taxonN

Convergence refers to the tendency of a Markov chain to "forget" its starting value and become typical of its equilibrium distribution. Note that convergence is a property of the Markov chain itself, not of individual runs of the Markov chain. Ideally a number of individual runs should be examined in order to determine how many initial iterations to discard as "burnin".

In MCMC, each sample is not fully independent of previous samples. In fact, even after a Markov chain has converged, it can get "stuck" in one part of the parameter space for a long time, before jumping to an equally important part. When this happens, each new sample contributes very little new information, and we need to obtain many more samples to get good precision on our parameter estimates. In such a case, we say that the chain isn't "mixing" well.

Potential Scale Reduction Factors check that different runs have similar posterior distributions. Only numerical variables may have a PSRF. To calculate the PSRF for each numerical parameter, you may run:

% statreport dir-1/C1.log dir-2/C1.log ... dir-n/C1.log > Report 

We compare the PSRF based on the length of 80% credible intervals (Brooks and Gelman 1998) and report the result as PSRF-80%CI. For integer-valued parameters, we add 1 to both the pooled and within-run interval widths before taking their ratio, since an integer interval from \(a\) to \(b\) contains \(b-a+1\) values.

We also report a new PSRF that is more sensitive for integer distributions. For each individual distribution, we find the 80% credible interval. We divide the probability of that interval (which may be more than 80%) by the probability of the same interval under the pooled distribution. The average of this measure over all distributions gives us a PSRF that we report as PSRF-RCF.

This convergence diagnostic gives a criterion for detecting when a parameter value has stabilized at different values in several independent runs, indicating a lack of convergence. This situation might occur if different runs of the Markov chain were trapped in different modes and failed to adequately mix between modes.

Show basic information about the alignment:

% alignment-info file.fasta
% alignment-info file.fasta file.tree

To select columns from an alignment:

% alignment-cat -c1-10,50-100,600- file.fasta > result.fasta
% alignment-cat -c5-250/3 file.fasta > first_codon_position.fasta
% alignment-cat -c6-250/3 file.fasta > second_codon_position.fasta

To concatenate two or more alignments:

% alignment-cat file1.fasta file2.fasta > all.fasta

Remove columns without a minimum number of letters:

% alignment-thin --min-letters=5 file.fasta > file-thinned.fasta

Remove sequences by name:

% alignment-thin --remove=seq1,seq2 file.fasta > file2.fasta

Remove short sequences:

% alignment-thin --longer-than=250 file.fasta > file-long.fasta

Remove sequences with <= 5 differences from the closest other sequence:

% alignment-thin --cutoff=5 file.fasta > more-than-5-differences.fasta

Like --cutoff, but stop when we have the right number of sequences:

% alignment-thin --down-to=30 file.fasta > file-30taxa.fasta

Protect some sequences from being removed:

% alignment-thin --down-to=30 file.fasta --protect=seq1,seq2 > file-30taxa.fasta

Remove sequences that are missing conserved columns:

% alignment-thin --remove-gappy=10 file.fasta > file2.fasta

Draw an alignment to HTML, optionally coloring residues by AU.

%  alignment-draw file.fasta --show-ruler --color-scheme=DNA+contrast > file.html
%  alignment-draw file.fasta --show-ruler --AU=file-AU.prob --color-scheme=DNA+contrast+fade+fade+fade+fade > file-AU.html

Find the last (or first) FastA alignment in a file.

% alignment-find --first < file.fastas > first.fasta
% alignment-find < file.fastas > last.fasta

Turn columns from a template alignment into alignment constraints:

% alignment-indices template.fasta > constraints.txt
% alignment-indices -c100-110,200,300- template.fasta > constraints.txt

Each line in this file corresponds to one alignment column.

Remove internal-node ancestral sequences from an alignment. (This probably only works for alignments output by bali-phy.)

% alignment-chop-internal < file.fastas > file-chopped.fastas

This program summarizes the tree sample contained in file. It reports the MAP topology, the supported taxa partitions (including partial partitions), and the majority consensus topology.

Usage: trees-bootstrap file1 [file2 ... ] --predicates predicate-file [OPTIONS]

This program summarizes the tree samples contained in file1, file2, etc. It gives the support of each tree sample for each predicate in predicate-file, and reports a confidence interval based on the block bootstrap.

Each predicate is the intersection of a set of partitions, and is specified as a list of partitions or (multifurcating) trees, one per line. Predicates are separated by blank lines.

Usage: trees-to-SRQ predicate-file [OPTIONS] trees-file

This program summarizes the tree samples contained in trees-file. It uses them to produce an SRQ plot for each predicate in predicate-file. Plots are produced in gnuplot format, with one point per line and with plots separated by a blank line.

If --mode sum is specified, then a "sum" plot is produced instead of an SRQ plot. In this plot, the slope of the curve corresponds to the posterior probability of the event. If the --invert option is used then the slope of the curve correspond to the probability of the inverse event. This is recommended if the probability of the event is near 1.0, because the sum plot does not distinguish variation in probabilities near 1.0 well.

In order to compile BAli-Phy, you need

• a C++23 compiler
• Meson (version >= 1.6)

% sudo apt-get install g++ git libcairo2-dev pandoc libboost-all-dev libcli11-dev

% sudo apt-get install meson
% meson --version
1.6.0

% conda create -n devel -c conda-forge --strict-channel-priority
% conda activate devel
% conda install meson gxx boost-cpp cmake pkg-config cairo
% export BOOST_ROOT=$CONDA_PREFIX

% sudo apt-get install python3 python3-pip python3-venv ninja-build
% python3 -m venv meson
% source meson/bin/activate
% pip3 install meson

% xcode-select --install

% brew install git meson cairo pandoc

% pacman -Syu

% pacman -S --needed \
  git \
  mingw-w64-x86_64-boost \
  mingw-w64-x86_64-cairo \
  mingw-w64-x86_64-cereal \
  mingw-w64-x86_64-eigen3 \
  mingw-w64-x86_64-fmt \
  mingw-w64-x86_64-gcc \
  mingw-w64-x86_64-meson \
  mingw-w64-x86_64-range-v3 \
  mingw-w64-x86_64-zstd

First check out the code using git:

% git clone https://github.com/bredelings/BAli-Phy.git
% cd BAli-Phy

Then run meson to configure the build process:

% meson setup build --prefix="$HOME/Applications/bali-phy-4.3/" --buildtype=release

Finally, build and install the software:

% ninja -C build test
% ninja -C build install

You can select the C++ compiler by setting the CXX variable. A useful example of this is to use g++-14 on systems where g++ invokes a compiler that is too old:

% CXX=g++-14 meson setup build --prefix="$HOME/Applications/bali-phy-4.3" --buildtype=release

% CXXFLAGS="-mtune=native -march=native" meson setup build --prefix="$HOME/Applications/bali-phy-4.3"

% CPPFLAGS="-I/usr/local/include" LDFLAGS="-L/usr/local/lib" meson setup build --prefix="$HOME/Applications/bali-phy-4.3" --buildtype=release