Changelog#

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.2.0 (Unreleased)#

Support for multiple additive terms.
Constant pseudo-prior. When used for weights, this can be used to project new data into an already existing latent space.
GSFA prior for analysis of CRISPR perturbation screens.
Single AnnData objects can now be used as input data. MOFA-FLEX will assume exactly one view for this type of input.
MuData objects with axis=1 can now be used as input data. MOFA-FLEX will treat each modality as a group and use the group_by argument, if given, to select a column in .var to split the data into views.
pl.variance_explained gained a factor_filter argument to plot only factors whose names satisfy a predicate (e.g. the annotation-informed factors of an InformedHorseshoe prior, dropping the uninformed dense ones).
It is now possible to disable the progress bar during training and to control its update interval.

The show_featurenames argument to pl.factor is now called show_samplenames to better reflect what it actually does.
The API has received an overhaul. Make sure to re-familiarize yourself with the tutorials.
R2 estimation for non-Gaussian likelihoods should be more robust.
The on-disk format for trained models has changed. Files created with mofaflex 0.1 cannot be read by 0.2 and vice versa.
The Gaussian process prior can now also be used for weights.
The spike and slab prior now has option to make the background distribution a Gaussian.
Training with sparse inputs and minibatching is about 1.5 times faster.
pl.factor_significance now ranks factors by the variance they explain within the selected views/groups, rather than the overall variance, so restricting views reorders the plot accordingly.

MOFAFLEX.load without an explicit map_location now uses the device the model was trained on (stored in the file), so methods that need it (e.g. pl.factor_significance/PCGSE, GP priors) work on reloaded models.

Using a MuData .obs column as a guiding variable now works.
When using guiding variables together with annotations, the data frame returned by get_annotations now has a correct row index.
Compatibility with Pandas 3.
Using independent lengthscales in the GP prior now works.

Gaussian processes with dynamic time warping and a custom reference group now actually use the the set reference group instead of the first warped group as reference.
The PCGSE test now also works if only a single annotated factor is present.