History

1.6.0 (2026-07-08)

  • Added the ProMERGE co-embedding algorithm, accessible via --algorithm promerge. ProMERGE co-embeds a query context against a base (anchor) context, disentangling shared from context-specific signal using either a learned mutual-information estimator (MINE) or direct subtraction. Adds the cellmaps_coembedding.promerge package and ProMERGECoEmbeddingGenerator, along with --lambda_disentangle, --lambda_l2_disentangle, and --lambda_var flags.

  • Bug fix: corrected a syntax error (== in place of =) in the ProMERGECoEmbeddingGenerator constructor that prevented runner from importing.

  • Bug fix: added promerge to the --algorithm choices so the algorithm is reachable from the command line.

  • Added a test suite for ProMERGE covering generator and CLI wiring, the CoEmbed model forward pass, save_results, and end-to-end training smoke tests.

1.5.0 (2025-10-23)

  • Deprecated cellmaps_coembedding.protein_gps module in favor of cellmaps_coembedding.proteinprojector. Updated documentation accordingly.

1.4.0 (2025-09-10)

  • Expose k parameter as flag

1.3.1 (2025-07-03)

  • Fix for UD-3116, missing header 0 row

1.3.0 (2025-05-29)

  • Create flags for parameters of proteingps coembedding algorithm

1.2.2 (2025-05-15)

  • Updated to PEP 517 compliant build system

  • Bug fixes: update constants, add back L2 normalization and fix separator issue for proteinGPS

1.2.1 (2025-04-14)

  • Fix scipy package version

1.2.0 (2025-03-19)

  • Added functionality to generate umap of embeddings (in cellmaps_coembedding.utils)

1.1.0 (2025-03-05)

  • Added functionality to evaluate embeddings using statistical analysis and visualization (functions get_embedding_eval_data and generate_embedding_evaluation_figures in cellmaps_coembedding.utils).

  • Update defauls (EPOCHS and DROPOUT)

1.0.0 (2025-01-28)

  • Rename auto coembedding name and proteinGPS. –algorithm auto option is depreacted and –algorithm proteingps should be used. The coembedding implementation was moved to ProteinGPSCoEmbeddingGenerator class and AutoCoEmbeddingGenerator is deprecated and calls proteingps. The package name was renamed from autoembed_sc to proteingps.

  • Removed l2 normalization

  • Constants updated in ProteinGPSCoEmbeddingGenerator (triplet_margin=0.2) and in proteingps’s fit_predict (triplet_margin=0.2, lambda_reconstruction=5.0, lambda_triplet=5.0)

  • Bug fix: add missing a .to(device) call to ensure tensors are correctly moved to the appropriate device.

  • Update version bounds of required packages

0.4.0 (2024-12-02)

  • Added README generation.

  • Refactor code.

  • Update defaults (EPOCHS)

0.3.1 (2024-09-13)

  • Bug fix: raise more informative error when no embeddings overlap.

0.3.0 (2024-09-06)

  • Added --provenance flag to pass a path to json file with provenance information. This removes the necessity of input directory to be an RO-Crate.

0.2.0 (2024-07-17)

  • Added a new coembedding algorithm accessible via flag --algorithm auto. This algorithm utilizes neural networks to generate latent embeddings, optimizing both reconstruction and triplet losses to improve embedding accuracy by learning intra- and inter-modality relationships.

0.1.0 (2024-02-12)

  • First release on PyPI.