az7jh2/SDePER: SDePER v2.0.0

This release represents a major update to SDePER, featuring substantial algorithmic improvements and optimization enhancements.

Updates:

• For optimization with graph Laplacian constraints, we transitioned from an ADMM-based framework to a gradient descent-based optimization strategy, as the ADMM approach was computationally expensive and inefficient for large-scale problems.

• The adaptive lasso penalty has been revised to support spot-wise hyperparameters for penalty strength, instead of a single global hyperparameter shared across all spots. This functionality is currently under comprehensive testing and will be fully enabled in a later release.

• The estimation precision of $\sigma^2$ has been improved from 1e-2 to 1e-6, leading to more accurate variance estimation.

• In the two-stage optimization framework, if a hyperparameter is explicitly set to zero by the user in stage 1 (adaptive lasso for variable selection) or stage 2 (graph Laplacian constraint), the corresponding stage will now be skipped directly, rather than executing an additional optimization round without the penalty.

• During cell type proportion $\theta$ estimation, a minimum boundary value of 1e-9 is introduced to ensure numerical stability during optimization. After optimization, boundary values are now reset to exactly zero to maintain sparsity.

Bug Fixes:

• In differential analysis for cell type-specific marker genes, the option --sortby_fc set to true previously prioritized p-values before fold changes. This behavior has been corrected so that results are now properly sorted by fold change as intended.