Minor breaking change! Python environments of slendr are no longer automatically activated upon calling
library(slendr)! Using the coalescent msprime back end and slendr’s tree-sequence functions now requires making an explicit call to a new function
library(slendr)is executed. (PR #102)
Motivation for the change: A small proportion of users have been experiencing issues with broken conda environments and various other issues with Python virtual environments in general. It’s hard to guess how frequent this has been, but experience from workshops and courses suggests perhaps 1 in 20 of users experiencing Python issues which hindered their ability to use slendr .(Fun fact: the first user-submitted GitHub issue upon releasing the first version of the slendr R package was… a Python virtual environment issue).
Explanation: Activating Python environments automatically upon calling
library(slendr) has been a popular feature because it hid away most of the complexities of the R-Python interface that powers slendr’s tree-sequence functionality. This was particularly convenient for many slendr users, particularly those who have no experience with Python at all.
Unfortunately, in cases where a Python virtual environments with tskit/msprime/pyslim on a user’s system ended up corrupted (or if anything else at the Python level got broken), the automatic Python environment activation performed by the
library(slendr) call failed and slendr was not even loaded. Sadly, this completely pulled the rug from under slendr and there was nothing that could be done about it from its perspective (the issue happened at a low-level layer of embedded-Python before slendr could’ve been loaded into R). Solving these issues was not difficult for experienced users, but many slendr users have no experience with Python at all, they have never used conda, they don’t understand the concept of “Python virtual environments” or how the R-Python interface works. And nor should they! After all, slendr is an R package.
Splitting the Python virtual environment activation step into its own
init_env() function means that
library(slendr) now always succeeds (regardless of potential underlying Python issues on a user’s sytem), making it much easier to diagnose and fix Python problems from R once the package is loaded.
So, to recap:
library(slendr) no longer activates slendr’s isolated Python virtual environment. In order to simulate tree sequences and analyse them using its interface to tskit, it is necessary to call
init_env(). This function performs the same Python-activation steps that
library(slendr) used to call automagically in earlier slendr versions. No other change to your scripts is necessary.
When a named list is provided as a
sample_sets =argument to a oneway statistic function, the names are used in a
setcolumn of the resulting data frame even if only single samples were used. (#2a6781)
It is now possible to label groups of samples in slendr’s tskit interface functions which should make data frames with statistics results more readable. As an example, running
ts_f3(ts, A = c("p1_1", "p1_2", "p1_3"), B = c("p2_1", "p2_3"), C = c("p3_1", "p3_2", "p3_"))resulted in a following data-frame output:
> ts_f3(ts, A = c("p1_1", "p1_2", "p1_3", "p1_4", "p1_5"), B = c("p2_1", "p2_2", "p2_3"), C = c("p3_1", "p3_2", "p3_3", "p3_4")) # A tibble: 1 × 4 A B C f3 <chr> <chr> <chr> <dbl> 1 p1_1+p1_2+p1_3+p1_4+p1_5 p2_1+p2_2+p2_3 p3_1+p3_2+p3_3+p3_4 0.000130
This gets unwieldy rather quickly, especially when dozens or hundreds of samples are grouped together as populations. The new syntax allows the following shortcut via customised group names leveraging the standard named
list functionality in R:
> ts_f3(ts, A = list(group_one = c("p1_1", "p1_2", "p1_3", "p1_4", "p1_5")), B = list(group_two = c("p2_1", "p2_2", "p2_3")), C = list(group_three = c("p3_1", "p3_2", "p3_3", "p3_4"))) # A tibble: 1 × 4 A B C f3 <chr> <chr> <chr> <dbl> 1 group_one group_two group_three 0.000130
This is more readable and in line with some other tskit-interface functions of slendr which used this functionality via their
sample_sets = argument (
ts_diversity(), etc.). (#ac5e484)
- The default state of the
parent =argument of
"ancestor". This prevents silly surprising clashes in situation where some population’s name really is “ancestor”. The only change internally is that for populations which are ancestral, the
splitsdata frame element of a slendr model object which includes this population carries a formal “ancestral parent population” as
"__pop_is_ancestor"instead of just
"ancestor". Note that this is an internal implementation detail and not something that particularly has to involve the user. Still, if you have been somehow using slendr’s internal data structures, keep this in mind. (#f8a39a2)
CRAN release: 2022-09-30
msprime()function now makes sure that a given slendr model can fully coalesce to a single common ancestor population. Previously, having multiple ancestral populations created with
parent = "ancestor"would cause an infinite simulation when plugged into the
The initial size of a population which emerges from a split from another population is now printed in a population history summary in the R console. (#6525bf3)
A couple of fixes to support loading, processing, and plotting of “manually” created tree sequences have been implemented (see this). Not sure how practically useful, but it’s important to be able to load even “pure” tree sequences which are not from simulators such as SLiM and msprime. A set of unit tests has been added, making sure that a minimalist nodes & edges table can be loaded, as well as nodes & edges & individuals, plus tables of populations and sites & mutations. PRs with more extensive unit tests and bug reports of tree sequences which are failing to load would be appreciated! The code for handling cases of “manually-created” tree sequences which have missing individual table, missing populations table, etc. seems especially brittle at the moment (#79adf14).
-1value as a missing value indicator used in tskit is now replaced with the more R-like
NAin various tree-sequence tables (annotated by slendr or original through tskit itself) (#79adf14).
slendr models can now be optionally compiled without serialization to disk. This only works with the
msprime()coalescent back end but will be much faster in cases where a huge number of simulations needs to be run because for non-serialized models,
msprime()now calls the back end engine directly through the R-Python interface (rather than on the command line) and output tree sequences are not saved to disk, rather than passed through the Python-R interface directly in memory (PR #112).
Avoid the unnecessary
arraytype of tskit results returned via reticulate. Numeric vectors (columns of data frames with numerical results) obtained in this way are simple R numeric vector (#5101b39).
One-way and multi-way statistics results are now returned as simple numerical vectors. Previously, results were returned as a type
arraydespite “looking” as vectors (this is how values are returned to R from the reticulate-Python layer), which caused unnecessary annoyances and type-conversions on the R side of things and was not even intended (#403df3b).
Computing population genetic statistics on named samples that are not present in a tree sequence (most likely typos) is now correctly caught and reported as an error (#da7e0bb).
CRAN release: 2022-08-19
SLiM 4.0 is now required for running simulations with the
slim()engine. If you want to run slendr simulations with SLiM (spatial or non-spatial), you will need to upgrade you SLiM installation. SLiM 3.7.1 version is no longer supported as the upcoming new slendr spatial features will depend on SLiM 4.x and maintaining two functionally identical yet syntactically different back ends is not feasible (PR #104).
At the same time as the SLiM 4.0 release, new versions of Python modules msprime, tskit and pyslim have also been released. In fact, to be able to work with SLiM 4.0 tree sequences properly, those Python modules must be upgraded as well. Next time you load
library(slendr), you will be prompted to setup a new updated Python environment which you can do easily by running
Experimental support for running coalescent msprime simulations and analysing tree-sequence data using tskit on the Windows platform has now been implemented (PR #102).
CRAN release: 2022-08-09
slendr is now on CRAN!
Big changes to the way tree-sequence outputs are handled by slendr by default. See this comment for an extended description and examples of the change. (PR #100). Briefly, simulation functions
msprime()now return a tree-sequence object by default (can be switched off by setting
load = FALSE), avoiding the need to always run
ts <- ts_load(model)as previously. At the same time, a parameter
output =can be now used in
msprime()to specify the location where a tree-sequence file should be saved (temporary file by default).
slendr’s tree-sequence R interface to the tskit Python module has been generalized to load, process, and analyze tree sequences from non-slendr models! This means that users can use the slendr R package even for analyzing tree sequences coming from standard msprime and SLiM scripts, including all spatial capabilities that have been only available for slendr tree sequences so far. Please note that this generalization is still rather experimental and there might be corner cases where a tree sequence from your msprime or SLiM script does not load properly or leads to other errors. If this happens, please open a GitHub issue with the script in question attached. (PR #91)
Removed functions and some function arguments originally deprecated during the renaming phase of the pre-preprint refactoring. This affects
shrink. Similarly, deprecated
dirargument of the
gene_flow, and the
_distsuffix was removed from
dispersal_dist. If you get an error about a missing function or a function argument in code which used to work in an ancient version of slendr, this is why. (#985b451)
When setting up an isolated Python environment using
setup_env(), slendr now makes a decision whether to install Python dependencies using pip (critical on osx-arm64 for which the conda msprime/tskit are unfortunately currently broken) or with conda (every other platform). This can be still influenced by the user using the
pip = <TRUE|FALSE>argument, but we now change the default behavior on ARM64 Mac. (#54a413d)
The name of the default slendr Python environment is now shortened even more, and the redundant
_pandasprefix is now dropped. Users will be notified upon calling
library(slendr)that a new environment should be created. This is OK, it’s not a bug. (#54a413d)
The format of the default slendr Python environment is now
msprime-<version>_tskit-<version>_pyslim-<version>_pandas, dropping the
slendr_prefix. This paves the way towards a future non-slendr tskit R package, which will share the same Python environment with slendr (because both R packages will go hand in hand). This isn’t really a user-facing change, except that calling
setup_env()will suggests creating a new Python environment and
library(slendr)will appear as if a slendr environment is not yet present. Calling
setup_env()and creating a new Python environment from scratch will solve the problem. (#eb05180)
setup_env()function for creating dedicated mini Python environments for slendr now installs packages using pip by default. Reason: The rate of conda failures and dependency conflicts (even in the trivial case of installing nothing more than msprime + tskit + pyslim + pandas) is too high to rely on it. The option to use conda for package installations with
setup_env()is still there, but the users must explicitly call
setup_env(pip = FALSE)to get this behavior. Note that conda is still used as a means to install Python itself! This change only affects the way how Python modules are installed into a dedicated slendr Python environment, not the installation of Python itself. (#81be1a7)
The name of the automatically created slendr-specific Python environment is now composed from the names and versions of Python modules installed. This makes it possible to naturally upgrade both slendr and its Python dependencies in case the tskit / msprime / pyslim folks upgrade some of those packages. In that case, if a slendr user upgrades the slendr package (and that new version requires newer versions of Python modules), slendr will simply recommend to create a new Python environment without additional effort on our part. (#81be1a7)
The code of
setup_env()was simplified to bare essentials. Now it only serves as a way to auto-setup a dedicated, isolated Python installation and slendr environment. The interface to install Python modules into custom-defined Python environment created outside R has been removed because this functionality is not necessary – these custom environments can be easily activated by calling
If some Python users want to use custom Python environments with msprime, tskit, and pyslim, they can silence the suggestion to use
setup_env()printed by the
library(slendr)call by setting
options(slendr.custom_env = TRUE). (#30f24b9)
sim_length =is now renamed to
simulation_length =. Both are accepted for the moment and using the old name will simply inform the user of the future deprecation. (#56491fb)
Extensive set of runnable examples including figures and a built-in pre-compiled example model have been added to the documentation. (#395df62c)
- First numbered version of slendr to celebrate its bioRxiv preprint. 🥳 🎉