Changelog

Changelog#

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.1.6 - 2026-06-29#

Fixed#

  • Zero-shot in-silico perturbation (perturb_dataset) no longer hangs and runs substantially faster, especially on larger datasets and many-core machines.

Changed#

  • perturb_dataset now maps in a single process by default (nproc=1).

0.1.5 - 2026-06-28#

Added#

  • terra.__version__ exposes the installed package version.

Changed#

  • perturb_dataset matches a specific perturbed_cell_id on each cell’s own cell_id, so perturbing specific cells with a cell target no longer needs add_neigh_cell_ids. Only neighborhood-target perturbations on specific cells require it (to know each cell’s neighbors).

Fixed#

  • perturb_dataset no longer stalls on datasets that carry a neighborhood cell_ids column. The perturbation map no longer adds per-row flag columns (changing the output schema forced datasets to re-encode every column, including the large nested cell_ids); return_only_perturbed_cells now selects the affected cells computed directly from the perturbation table.

  • get_spatial_score now honors its compute_cosine_with_list argument (it was previously overwritten and ignored).

Documentation#

  • Expanded the tutorials (single- and multi-sample zero-shot quickstarts and a downstream-analysis walkthrough covering gene programs, spatial gene-pair scoring, EMD spatial structure, and in-silico perturbation) and user guide, and fixed numerous API-reference rendering issues.

0.1.4 - 2026-06-27#

Fixed#

  • Perturbing specific cells by ID now works on multi-sample data: the pipeline forwards add_neigh_cell_ids through its multi-sample branch, and the neighborhood cell_ids column is kept out of the torch format so it survives tokenization.

0.1.3 - 2026-06-27#

Fixed#

  • perturb_dataset no longer hangs on torch-formatted datasets. It now maps over the unformatted dataset (formatting the large nested token columns inside map could stall indefinitely) and restores the format afterwards.

Changed#

  • perturb_dataset(..., return_only_perturbed_cells=True) now subsets the perturbed cells with an index-based select instead of a full-table filter, avoiding a costly scan and rewrite of every tokenized row.

0.1.2 - 2026-06-27#

Fixed#

  • read_xenium_10x now aligns cell centroids correctly for older Xenium panels with an integer cell_id, which previously produced all-NaN spatial coordinates.

  • get_spatial_score and get_emd_distance no longer raise NameError: gene_embed_dataset — the helper they depend on is now imported.

  • perturb_dataset with perturbed_cell_id="all" now applies "all"-gene knockout/fold-change perturbations correctly (the all-cells fast path referenced an unbound variable for whole-panel perturbations).

0.1.1 - 2026-06-26#

Added#

  • Optional local_dir argument to download_pretrained to download a model bundle into a chosen folder instead of the Hugging Face cache.

Fixed#

  • Single-sample zero-shot inference (harmonize_tokenize_embed_pipeline with sample_key=None) no longer fails with KeyError: 'batch'; the whole AnnData is now treated as one batch and any existing batch column is kept.

0.1.0 - 2026-06-26#

First public release of TERRA, a JEPA-based foundation model for spatial transcriptomics, pretrained on HST-Corpus-112M (>100 million cells at single-cell resolution).

Added#

  • Pretrained models (TERRA-96M, TERRA-112M) distributed as self-contained bundles on the Hugging Face Hub, with download and publishing utilities (download_pretrained, terra-hub).

  • Zero-shot inference pipeline producing cell- and neighborhood-level embeddings for single- and multi-sample spatial data.

  • Downstream analyses: gene-level embeddings, spatial gene-pair scoring, EMD-based spatial structure, and in-silico perturbation.

  • Finetuning of the pretrained encoder with LoRA/PEFT.

  • Documentation, tutorials, and API reference.

Contents