NEWS

glmGamPoi 1.19 (2024-11-26)

glmGamPoi is using the new tatami library instead of the old beachmat interface. These libraries helped to work with different kind of matrices (dense, HDF5-backed). One important change is that glmGamPoi now requires C++17 (instead of C++14). (PR#66,thanks Aaron Lun)

'test_de' can now compute the standard error of the log2-fold change (PR#63, thanks @jackkamm)
'predict' now considers the 'ridge_penalty' when calculating the standard errors.
Detect if smart subset of MatrixGenerics' 'rowMeans2' and 'rowSums2' can be used to speed-up 'pseudobulk' calculation.

Implement a likelihood ratio test based on the Chi-squared distribution, if 'test_de' is called after setting 'overdispersion_shrinkage = FALSE'. Note that this test is less reliable than than the quasi-likelihood F test that is run for 'overdispersion_shrinkage = TRUE'.

Breaking change: rename 'pseudobulk_sce' to 'pseudobulk'
Add a new vignette explaining how and why pseudobulking is a powerful concept for single cell data analysis
Depcreate 'pseudobulk_by' argument in 'test_de'. Use the 'pseudobulk' function instead.
Add a new argument 'max_lfc' to to 'test_de' to avoid impractically large log fold changes for lowly expressed genes.
Support rlang quosures for the contrast argument in 'test_de'
Add a helper function called 'fact' that simplifies specification of contrast for complex experimental designs
Add 'use_assay' argument to 'glm_gp'
Add 'vctrs' as dependency. The package is necessary to replicate the 'group_by' behavior from 'dplyr'.
Add 'size_factors = "ratio"' to emulate the behavior of DESeq2's size factor calculation
Make sure that the 'ignore_degeneracy' argument is propagated to 'test_de'

Breaking change to the way that non-standard evaluation parameters are handled. Variables in arguments such as 'pseudobulk_by' or 'subset_to' which evaluate to a single string are no longer interpreted as referring to a column. This change makes the handling of NSE more consistent.
Add new function 'pseudobulk_sce' to easily form pseudobulk samples

Choose a more reasonable scale for global overdispersion estimate
Make code more robust accidental internal NA's
Add fallback mechanism in case the Fisher scoring fails to converge. Instead of returing NA, try again using the BFGS algorithm.
Better error message if the design contains NA's

Ridge regularization framework. glmGamPoi now supports regularizing the coefficient estimates using a quadratic penalty function. Furthmore, more advanced regularization schemes, such as regularizing towards a specific value and full Tikhonov regularization are implemented.
New predict() function. Also supports estimating the standard error of the mean estimate.
Make sure that Fisher scoring does not converge to unrealistically large values of mu
Fix minor bug in test_de() concerning the calculation of the degree of freedom
Fix minor bug in calculation of working and Pearson residuals, which used to return NaN if mu was 0. Now, they are 0.
Improve vignette/Readme: add section on differential expression analysis with Kang et al. (2018) as example data
'glm_gp' returns the Offset matrix and bug fix for test_de() if a offset was specified
Add CITATION file
Make sure that residuals are pristine (when the input was a DelayedArray)
Set dimnames of residuals
Improve error message if input is a sparse matrix

Remove dual likelihood functions for overdispersion estimation. Instead merge functionality into conventional_***. This should cause no user facing changes, however should make it easier to maintain the package
Make conventional_score_function_fast() more robust to extreme inputs. Avoid numerically imprecise subtractions and employ bounds based on series expansions for very small input
If dispersion estimate quits because there is no maximum or all y are 0, return iterations = 0
Add limits (1e-16 / 1e16) for nlminb estimates of the dispersion. This protects against errors due to NA's in the conventional_likelihood_fast
Automatically set 'size_factors = FALSE' for input with 0 or 1 row. This will change the estimated beta, but not the mu's
Rename gampoi_overdispersion_mle() -> overdispersion_mle()
Store data in the object returned by glm_gp()
Remove Y from the interface of residuals.glmGamPoi, because I can just get it directly from fit$data
Add function test_de() that does a quasi-likelihood ratio test to detect differentially expressed genes
Add functionality to make a pseudobulk test directly from test_de() by aggregating the data around one column
In group-wise beta estimation, fall back to optimize() if the Newton method fails
Change the default size factor estimation method from "poscounts" to "normed_sum" and provide an easy way to call scran::calculateSumFactors()
New "global" mode for dispersion estimation