Changelog#

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

[1.0.0] — 2026-06-04 - Release#

This is the first stable release of Cellina, graduating from the 0.99.x pre-release series.

subgraph_type parameter on CellinaGCN and GraphJointDataSplitter ("induced" | "directional"); directional mode cuts counterfactual-inference VRAM by ~40% compared to the induced default.
End-to-end tutorial notebook shipped in the repository at docs/tutorial.ipynb, covering a full CRC counterfactual workflow from data loading to perturbation analysis.

__version__ is now resilient to missing package metadata (returns "unknown" instead of raising PackageNotFoundError).
README updated to reference docs/tutorial.ipynb as the primary getting-started resource.

Vectorized SupCon loss for improved training speed; updated tests accordingly.
Hops and layer mismatches will raise a warning. From tests, this has a negligable effect if batch_size or the number of neighbors is large enough.
Added log1p for spatial_x in case not normalized.

Added minus 1.0 back to multiplicative perturbation to avoid zeroing out genes with large negative logFCs; updated tests accordingly.

This is the first public version of Cellina: a dual-encoder variational autoencoder for spatial transcriptomics with adversarial domain forgetting.

Cellina / CellinaModule: MLP-based dual-encoder VAE with adversarial domain classifier and discriminator for batch-effect removal
CellinaGCN / CellinaGCNModule: GCN-based variant that encodes spatial context via a graph convolutional network (s_encoder) alongside the count encoder (z_encoder); supports link prediction for edge-level tasks
CellinaAdversarialTrainingPlan: unified two-step adversarial training plan shared by both module types
GraphJointDataSplitter / InferenceBatchLoader / JointBatchLoader: graph-aware data loading built on PyTorch Geometric NeighborLoader and LinkNeighborLoader
spatial_neighbors: builds spatial connectivity graphs (kNN + kernel weighting) with support for gaussian, exponential, and linear kernels, per-library graph construction via library_key, and the new test_indices parameter
test_indices in spatial_neighbors / _spatial_neighbors_core: isolates test cells from the spatial graph at build time via coordinate displacement, producing all-zero rows and columns for those cells without post-hoc masking
compute_spatial_features, make_neighbor_perturbation, make_perturbed_expression: spatial feature computation and in-silico perturbation utilities