BPReveal 5.2.2

This is a minor bugfix release that corrects an issue with how numpy is installed if you use the conda environment scripts.

Happy training,
Charles.

Zenodo release

No changes to the code, this is just a new release to trigger the Zenodo archiving machinery.

BPReveal 5.2.1

Lots of goodness in this release.

• Added an option to include a few low-counts regions when you use prepareBed.
• A bunch of tweaks and little new features for PISA plots, such as new annotation shapes and better control over the y-axes.
• interpretFlat now lets you only calculate profile or counts scores instead of both, saving time in cases where you only want one type of importance score.
• Fixed a nasty bug when making predictions from a fasta file that would crash if you didn't supply coordinates.
  There should be no breaking changes with this release, and it is tested on Tensorflow 2.20.

Happy training,
Charles.

BPReveal 5.2.0

Tooling-wise, this is a pretty minor update. You can now use the interpretation tools on CPU, which is nice for quick one-off stuff. A few bug fixes here and there also made it in. Most of the changes were under the hood, and the way that the project builds itself changed significantly.

More importantly, though, BPReveal is now on pip! Installing should now be as easy as pip install bpreveal! There may be some funny business as I start the whole pip adventure, but it should become the default way to install BPReveal for most use cases.

Happy training,
Charles.

BPReveal 5.1.0

This release brings a much more flexible interpretation framework. You can now supply your own custom metrics for shap value calculation, opening up new possibilities for the sort of effects that we can explain. I've also added an interpretation backend that uses scanning mutagenesis instead of DeepSHAP to get importance scores. It's slow, but lets you perturb the sequence to your heart's content without having to write a whole bunch of boilerplate.

In other news, the gaOptimize library got support for random mutations (both SNPs and indels) which is handy for systematic mutation experiments like marginalization.

easyPredict can now handle input sequences that are longer than a model's input length. In this case, it will make enough predictions to tile the whole input window and concatenate them seamlessly.

More detailed changes can be found in the changelog, as usual.

Happy training,
Charles.

BPReveal 5.0.2

This minor bugfix release fixes a rare segfault that could occur with extremely large training data sets. This bug was introduced in 4.1.1. If you need this fixed in a 4.x release, copy over the new libslide.c file into src/internal and run make in src to re-build libslide with the segfault fix.

This release has no breaking changes.

BPReveal 5.0.1

Version 5.0 of BPReveal moves from the old Keras 2.0-based code to Keras 3.0 through most of the repository. All existing configurations should still be valid, but note that models are now saved with a .keras extension instead of .model. If you are using the BPReveal API, those calls should all work just as before, but if you were manually loading a model, that will probably need to change.

If your current BPReveal install is working fine, there's not much reason to upgrade to 5.0, but new features will not be backported.

Note that 5.0.0 should not be installed, since Tensorflow 2.18 has a regression that totally breaks the GPU on my machine and therefore I'm not able to test it. 5.0.1 (this release) locks the Tensorflow version to 2.17 and that should be used instead.

Happy training!

BPReveal 4.3.0

• Fixed a long-standing bug where it was impossible to interpret combined or transformation models. Fixing this required a slight change to the internals of how a transformation model works, but this should be transparent for most users.
• There's a motif curation program in the tools/ directory, and it will migrate to the main repository in the next release. It's called bestMotifsOnly.py.
• You can also now have interactive PISA plots, which is handy for exploring your data.
• You can render the documentation as man pages now.

Documentation is available as a PDF attached to this announcement, or online at https://bpreveal.readthedocs.io/en/v4.3.0/

Happy training!

BPReveal 4.2.0

This release brings the python version to 3.12 and Tensorflow to 2.16, but keeps using the old keras (TF_ENABLE_LEGACY_KERAS). If you're using Keras directly, you'll have to import tf_keras instead of keras on its own.
Once the deepshap code can use Keras 3, then the whole repo will migrate to keras 3, which will likely introduce more breaking changes.

This version introduces plotting.py, which can create handsome graphics like this:

As always, everything is documented to the nines, and you can read the documentation on readthedocs, or in the attached four-and-a-half megabyte, two hundred fifty-nine page PDF.

Happy training!

BPReveal 4.1.4

Locked the tensorflow version to 2.15, since the change in Keras version in 2.16 caused all sorts of breakage.
Due to the tensorflow update being a breaking change, old releases of BPReveal will no longer create a valid environment.

Added a tool to shift PISA data, mostly useful for mnase data.

Added a very performant metapeak calculator to bedUtils.

Lots of little bug fixes.