Changed

• Fixed bugs inVIPRSGridSearch andVIPRSBMA models, specifically how they were handling_log_var_tau,
  and the hyperparameters objects after selecting best models or performing model averaging.
• Fixed bug in howviprs_fit handles validationgdls when the user passes genotype data.
• Updated interfaces inHyperparameterSearch script to make it more flexible and efficient. Primarily,
  I added shared memory object for the LD matrix to avoid redundant memory usage when fitting multiple
  models in parallel. (** WORK IN PROGRESS **).
• Updated implementation ofpseudo_r2 to use square of pseudo correlation coefficient instead. The previous
  implementation can be problematic with highly sparsified LD matrices.
• Updated implementation ofVIPRSGrid to be better integrated with theVIPRS class. The new implementation
  also allows for fitting the grid in apathwise fashion (now default behavior), where we use
  parameter estimates from previous grid points as warm-start initialization for the current grid point.
• RemovedVIPRSGridSearch andVIPRSBMA classes for now. These functions are implemented ingrid_utils.py instead
  and they can be applied generically to anyVIPRSGrid model.

Added

• Addedviprs-cli-example.ipynb notebook to demonstrate how to use theviprs commandline interface.
• Added documentation page for Downloading LD matrices.
• Added new utility functioncombine_coefficient_tables to combine the output from multiple VIPRS models.
• Added more thorough tests for the various models + CLI scripts.
• AddedPeakMemoryProfiler toviprs_fit to more accurately track peak memory usage. Temporary solution,
  this will be moved tomagenpy later on.
• Added support for splitting GWAS sumstats to training/validation sets and exposed appropriate interfaces
  in the base classBayesPRSModel.
• AddedIterationConditionCounter class to keep track of the number of consecutive iterations
  where a certain condition is met. This is used to monitors convergence of the optimization routine.

Changed

• Fixed bug in implementation of.fit method of VIPRS models. Specifically,
  there was an issue with thecontinued=True flag not working because theOptimizeResult
  object wasn't refreshed.
• Replacedprint statements withlogging where appropriate (still needs some more work).
• Updated way we measure peak memory inviprs_fit
• Updateddict_concat to just return the element if there's a single entry.
• Refactored pars ofVIPRS to cache some recurring computations.
• UpdatedVIPRSBMA &VIPRSGridSearch to only consider models that
  successfully converged.
• Fixed bug inpsuedo_metrics when extracting summary statistics data.
• Streamlined evaluation code.
• Refactored code to slightly reduce import/load time.
• Fixed bug inviprs_evaluate

Added

• Added SNP position to output table from VIPRS objects.
• Added measure of time taken to prepare data inviprs_fit.
• Added option to keep long-range LD regions inviprs_fit.
• Added convergence check based on parameter values.
• Addedmin_iter parameter to.fit methods to ensure CAVI is run for at leastmin_iter iterations.
• Added separate method for initializing optimization-related objects.
• Added regularization penaltylambda_min.
• Added Spearman R and residualized R-Squared metrics to continuous metrics.

Additional files

Attached are LD matrices for 6 continental ancestry groups, as defined by thePan-UKB project. The LD matrices are estimated from unrelated samples in the UK Biobank using block-diagonal masks.

Changed

• Fixed bugs in the E-Step benchmarking script.
• Re-wrote the logic for finding BLAS libraries in thesetup.py script. 🤞
• Fixed bugs in CI / GitHub Actions scripts.

Added

• Dockerfiles for bothcli andjupyter modes.

A large scale restructuring of the code base to improve efficiency and usability.

Changed

• Moved plotting script to its own separate module.
• Updated some method names / commandline flags to be consistent throughout.
• Updated theVIPRS class to allow for more flexibility in the optimization process.
• Removed theVIPRSAlpha model for now. This will be re-implemented in the future,
  using better interfaces / data structures.
• Moved all hyperparameter search classes/models to their own directory.
• Restructured theviprs_fit commandline script to make the code cleaner,
  do better sanity checking, and introduce process parallelism over chromosomes.

Added

• Basic integration testing withpytest and GitHub workflows.
• Documentation for the entire package usingmkdocs.
• Integration testing / automating building with GitHub workflows.
• New self-contained implementation of E-Step inCython andC++.
  • UsesOpenMP for parallelism across chunks of variants.
  • Allows for de-quantization on the fly of the LD matrix.
  • Uses BLAS linear algebra operations where possible.
  • Allows model fitting with only
• Benchmarking scripts (benchmark_e_step.py) to compare computational performance of different implementations.
• Added functionality to allow the user to track time / memory utilization inviprs_fit.
• AddedOptimizeResult class to keep track of the info/parameters of EM optimization.
• New evaluation metrics
  • pseudo_metrics has been moved to its own module to allow for more flexibility in evaluation.
  • New evaluation metrics for binary traits:nagelkerke_r2,mcfadden_r2,
    cox_snell_r2liability_r2,liability_probit_r2,liability_logit_r2.
  • New function to compute standard errors / test statistics for all R-Squared metrics.