Cellina

Cellina is a dual-encoder variational autoencoder for predicting how a cell's gene expression changes under altered spatial contexts — a class of queries we call tissue graph counterfactuals.

In tissues, a cell's transcriptional state is shaped by its local neighborhood: the composition of nearby cells and the signals they emit. Existing perturbation methods typically treat cells as independent and apply perturbations uniformly. Cellina addresses this gap by explicitly separating a cell's intrinsic state (z, encoding cell identity) from its spatial context (s, encoding microenvironmental influence), then uses s as a conditioning input to render counterfactual predictions under two types of intervention:

• Edge perturbation — rewire a cell's neighborhood (replace neighbors with those from a different domain)
• Node perturbation — modify the expression of existing neighbors (e.g. pathway activation or knockout)

Getting started

Follow the worked tutorials for end-to-end examples on colorectal cancer tissue: Cellina and Cellina-GAT (or run them locally from docs/tutorial.ipynb and docs/tutorial_gat.ipynb).

How it works

Generative model. Cellina is a VAE with two latent variables. An MLP encoder $\text{Enc}_z$ maps raw counts to $z \sim q(z \mid x)$; a spatial encoder maps the cell's neighborhood to $s \sim q(s \mid \mathcal{N}(v))$. A shared decoder reconstructs counts from $[z;, s]$ under a Negative Binomial likelihood. Both latents have standard normal priors.

Supervised disentanglement. Optimizing the ELBO alone does not prevent $z$ from absorbing spatially-driven variation. Cellina adds auxiliary objectives:

• A cell-type classifier on $z$ anchors it to transcriptional identity.
• An adversarial discriminator is trained to predict spatial domain from $z$; the encoder is then trained to fool it, routing microenvironmental variation to $s$ by elimination.
• A graph-supervised contrastive loss $s$ (CellinaGCN only, optional), as a biologically grounded inductive bias that promotes similarity within local neighbourhoods. Enabled by setting link_prediction_weight > 0.

Training alternates between a discriminator step (encoder frozen) and a VAE step (discriminator frozen), following a standard adversarial schedule.

Two variants differ in how the spatial encoder is implemented:

Code class	Paper name	Spatial encoder
Cellina	Cellina	Degree-normalized weighted pseudobulk aggregation of neighbor expression → MLP
CellinaGCN	Cellina-GAT	Multi-layer GATv2 on the local subgraph; self-loops excluded so $v$'s own expression is captured by $z$ alone; modified contrastive loss on $s$

The two variants perform on par. Cellina decouples neighborhood construction from training and scales similarly to non-spatial baselines; CellinaGCN learns attention over each subgraph at additional cost per step.

Tissue graph counterfactuals

Cellina supports two post-training interventions on the spatial graph $\mathcal{G}$:

Edge perturbation replaces a cell's spatial neighbourhood with donors sampled from a target tissue domain, while keeping the cell's own expression fixed:

$$\mathcal{N}(v) := \mathcal{N}'$$

Node perturbation modifies the feature vectors of $v$'s neighbours while preserving graph topology. For a target gene set $\mathcal{S}$ and a gene-specific transformation $T_g$:

$$x_{u,g}^{\mathrm{cf}} = \begin{cases} T_g(x_{u,g}) & g \in \mathcal{S} \ x_{u,g} & g \notin \mathcal{S} \end{cases}$$

$T_g$ can encode any intervention (additive shift, knockout, overexpression, or learned counterfactual values).

See the Cellina and Cellina-GAT tutorials for full worked examples.

Release notes

See the changelog.

Installation

Cellina ships two conda environments: environment.yml for the full GPU/CUDA setup, and env_minimal.yml for a lightweight CPU-only install. Create one with conda env create, then follow the tutorials above.

Citation

Citation coming soon.

Built on scvi-tools.

Contact

If you found a bug, please use the issue tracker.

Copyright (c) 2026, PMBio