PARM v0.2.6

Minor update to fix compatibility with newer versions of pytorch

Full Changelog: v0.2.5...v0.2.6

PARM v0.2.5

What's Changed

• Add L_max parameter to prediction function; Handle test-set predictions for datasets without FEATname by @luciabarb
• Fix motif indexing in find_hits_and_make_logo by @vhfsantos

Full Changelog: v0.2.0...v0.2.5

PARM v0.2.0

Change default hyperparameters for training

Now the default hyperparameters for training PARM models are set to match those that we used in our paper

Additional changes

• Clean up code for PARM_train
• Update citation message

v0.1.44

Fix issues when starting up PARM mutagenesis and PARM predict from last version

v0.1.39

Optimise PARM startup time

This version's main change is to improve PARM's startup time by implementing lazy imports in its functions.

Additional changes

• Matching the input format to the preprocessing pipeline.
• Adding feature names in the output of test-set predictions.
• Allowing training from defined weights, instead of starting from random.

v0.1.27

Improve support for user-trained PARM models

Key changes:

• Organised output directories in PARM_train.py by creating subfolders (temp_models and performance_stats) for temporary model files and performance metrics, respectively.
• Enhanced validation loop in PARM_train.py to generate and save scatter plots showing predicted vs. measured values for each validation epoch.
• Change the filename of the output model. Now, it follows the basename of the output directory.
• Added support for test fold predictions in PARM_predict.py (via --predict_test_fold argument), allowing evaluation of trained models using HDF5 test fold data. This includes generating measured vs. predicted plots and calculating Pearson correlation coefficients.
• Add instructions on the README on how to train the models.

---

Full Changelog: v0.1.0...v0.1.27

PARM v.0.1.0

We made a significant rework on PARM and now introduce version 0.1.0.

Key changes

• Make available all nine pre-trained models: AGS, HAP1, HCT116, HEK116, HepG2, K562, LNCaP, MCF7, and U2OS.
• Update the settings for input model: now, PARM deals with one cell/model at a time.
• Implement a batch system that significantly speeds up the prediction time.

Other changes

• Add extra parameters for user-trained models (type_loss, filter_size)
• Improve logging and progress bar

v0.0.7

What's Changed

Some important changes in the parm train task:

• Add the column name of the measurement data as an argument: now, the user needs to specify which input data column should be used for training.
• Make the model weight file be named as the cell type: before, the output model was always called model.parm. Now, it is automatically set to the name of the cell type set by the user.

Small changes:

• Apply attribution_range also to importance track, not only for the matrix
• improve log messages and hide the progress bar for stdout

Full Changelog: v0.0.6...v0.0.7

v0.0.6

What's Changed

• PARM now checks if the input fasta contains any sequence longer than the L_max
• Added L_max parameters for the tasks, in case of custom models are used for them
• Improved the verbosity and small fixes of the PARM train

Full Changelog: v0.0.5...v0.0.6

v0.0.5

What's Changed

• Add --min_relative_attribution argument to parm plot, so that the user can filter out motif hits by defining a minimum percentage of the highest letter the motif's attribution should have.
• All the error raises were changed to sys.error to make it more integrated as a command line
• Small bugs fixed in defining the required arguments and the argument types (thanks @magnitov!)

Full Changelog: v0.0.4...v0.0.5