Gao Lab GLUE Versions Save

Bug fixes:

• Fixed "real_cross" loss in PairedSCGLUETrainer and SCCLUETrainer.

Bug fixes:

• Fixed NaN loss in PairedSCGLUETrainer.
• Restored rna_anchored_prior_graph as a deprecated function (to be replaced by rna_anchored_guidance_graph).

New features:

• New tutorial and functions for regulatory inference (Resolves #15, #41).
• New tutorial for training on partially paired data (Resolves #24).

Enhancements:

• Modularized scglue.dx.integration_consistency to allow for non-raw-count input (Resolves #30).
• Added documentation translation in Chinese.

Minor improvements and bug fixes

Bug fixes:

• Data frame in obsm no longer triggers an error during model training (Resolves #32).

Enhancements:

• scglue.data.transfer_labels uses a new strategy with SNN-based estimation of transfer confidence (Resolves #23).
• Allow setting custom bedtools path via scglue.config.BEDTOOLS_PATH (Resolves #22).

Minor improvements and bug fixes

Bug fixes:

• Device detection is now more reliable (Resolves #17).

Enhancements:

• Custom encoders and decoders can now be registered without changing package code (Resolves #14).

Minor improvements and dependency fixes

New features

• Added fit_SCGLUE function to simplify model training
  • Incorporates weighted adversarial alignment by default, with increased robustness on datasets with highly-skewed cell type compositions
• Added support for batch effect correction, which can be activated by setting use_batch in configure_dataset
• Added a model diagnostics metric “integration consistency score”

Enhancements

Support for training directly on disk-backed AnnData objects, scaling to almost infinite number of cells

Bug fixes

Fixed a bug where the graph dataset was not shuffled across epochs

Experimental features

• A partially paired GLUE model for utilizing paired cells whenever available
• The CLUE model that won the NeurIPS 2020 competition in multimodal integration (modality matching task) is here!

First public release