Cellular Transformer: Topology & Imaging

• Cellular Transformer is a family of transformer models that operate on cell complexes and 3D imaging data, integrating high-order topological and spatial attention.
• The architecture employs specialized pairwise and general attention mechanisms alongside advanced positional encodings to capture complex relationships.
• Empirical results demonstrate improved performance in molecular regression, classification, and cell membrane segmentation with robust training strategies.

The term "Cellular Transformer" (CT) denotes distinct, state-of-the-art transformer-based architectures in two major domains: topological deep learning on cell complexes (Ballester et al., 2024) and 3D cell membrane tracking with subcellular-resolved quantification in embryology (Li et al., 16 Dec 2025). Both approaches leverage transformer models but target fundamentally different data structures and scientific questions. The following entry systematically addresses both interpretations at a technical level.

1. Mathematical and Topological Foundations

In topological deep learning, the Cellular Transformer is built to operate natively on cell complexes, which are topological spaces generalizing graphs and simplicial complexes. A (regular) 2-dimensional cell complex is a triple

$X = (X_0, X_1, X_2)$

where $X_0$ comprises 0-cells (vertices), $X_1$ contains 1-cells (edges), and $X_2$ the 2-cells (faces). Each edge $e \in X_1$ is an ordered pair $[v_1, v_2]$ of vertices; each face $\sigma \in X_2$ is a cyclically ordered sequence of edges forming a closed, non-self-intersecting path.

Boundary operators $\partial_k : C_k(X) \to C_{k-1}(X)$ are realized as signed incidence matrices $B_k$ with

$$[B_k]_{i,j} = \begin{cases} +1 & \text{if the } j\text{-th }k\text{-cell ends in the }i\text{-th }(k-1)\text{-cell}, \ -1 & \text{if it begins there}, \ 0 & \text{otherwise} \end{cases}$$

and unsigned incidence matrices $X_0$0. The $X_0$1-th combinatorial Hodge Laplacian is given by

$X_0$2

with lower and upper Laplacians $X_0$3 and $X_0$4.

In contrast, the bioimage analysis CTransformer pipeline processes 4D (3D + time) fluorescent microscopy volumes. Here, the focus is on segmenting cell membranes and tracking cell lineages in living embryos (Li et al., 16 Dec 2025). The data is volumetric and temporal, given as $X_0$5, with each time point processed as a 3D volume.

2. Architecture and Attention Mechanisms

Topological CT for Cell Complexes

The Cellular Transformer layer operates on tuples of $X_0$6-cochains for $X_0$7:

$X_0$8

Two principal attention schemes are defined:

• Pairwise Cellular Attention: For source $X_0$9 and target $X_1$0, single-head attention is:

$X_1$1

with learned projections $X_1$2, $X_1$3, $X_1$4, neighbourhood matrix $X_1$5 (e.g., incidence or adjacency), and $X_1$6 indicating matrix addition (self) or entrywise product (incidence).

• General Cellular Attention: All cochains are concatenated. Shared $X_1$7 and rank-specific $X_1$8 are used:

$X_1$9

Each layer employs prenorm ordering: LayerNorm, multi-attention, residual sum, feedforward, and another residual.

CTransformer for Cellular Imaging

CTransformer methods utilize a transformer-U-Net backbone (TUNETr). The model sequentially processes 3D volumes with patch embedding, multi-head self-attention (Swin-Transformer blocks performing intra-window and shifted-window attention), encoder-decoder structure with skip connections, and a Euclidean Distance Transform feature regression (EDT-GFR) segmentation head. Patch-embedded tokens are partitioned into windows, with positional biases and attention computed as:

$X_2$0

where $X_2$1 encodes learnable position biases. Patch merging and expanding perform spatial down/upsampling. For instance segmentation, membrane predictions are binarized, nuclei are inferred (by direct channel or GAN), and Delaunay-watershed merging produces final cell instances.

3. Topological and Geometric Positional Encodings

For topological CT, positional encodings exploit high-order structure:

• Barycentric-Subdivision PE (BSPe): The 1-skeleton of the barycentric subdivision $X_2$2 is built; the $X_2$3 smallest eigenvectors of its Laplacian are used per cell.
• Random Walk PE (RWPe, RWBSPe): Random walks on adjacency graphs (either subdivision or order-specific) yield features as functions of power iterates' diagonal entries.
• Topological Slepian PE: Slepian eigenproblems on $X_2$4 within a frequency band, concentrated on a subset of cells, provide spectral-localized encoding.

For CTransformer in imaging, relative position bias in Swin blocks encodes spatial relationships between 3D patches, enhancing the ability to integrate local and non-local context.

4. Implementation and Training Specifics

Topological CT

Graphs are lifted to 2D cell complexes via cycle-filling (TopoX). Vertex, edge, and face features are initialized with original features or summary statistics. Sparse block matrices support batched computation. Typical hyperparameters are: layers $X_2$5, hidden dims $X_2$6–$X_2$7, heads $X_2$8–$X_2$9, dropout $e \in X_1$0, AdamW optimizer, and cosine annealing scheduler.

CTransformer

Training employs 71 manually annotated 3D volumes for TUNETr and 16 evaluation volumes; batch size is 1 due to 3D memory requirements. Data augmentations include intensity scaling, random flips, and sub-volume crops. Optimization uses Adam (lr=$e \in X_1$1, weight decay=$e \in X_1$2, AMSGrad), cosine learning rate decay, up to 5000 epochs.

Segmentation is supervised via geometric loss $e \in X_1$3 (MSE between predicted/probability and GT mask) and topological constraint loss $e \in X_1$4 (p-Wasserstein between persistent homology diagrams of predicted and GT EDT maps), giving $e \in X_1$5 with $e \in X_1$6 ramped after initial epochs.

5. Empirical Evaluation and Ablation

Topological CT

Performance is assessed on GCB (lifted to complexes), ZINC (molecular regression), and ogbg-molhiv (molecular classification). CT attains or exceeds SOTA on GCB Accuracy ($e \in X_1$7 vs. prior $e \in X_1$8). For ZINC MAE, CT achieves $e \in X_1$9; for molhiv AUC-ROC, $[v_1, v_2]$0.

Ablations spanning 30 attention and positional encoding combinations show pairwise attention with local PE excels on tasks with heterogeneous features (molecular), while general attention with global PE dominates for homogeneous features (GCB). Global PE (BSPe) is consistently among the top three.

CTransformer

Segmentation quality is measured using Dice, Jaccard, and Hausdorff metrics. CTransformer achieves Dice $[v_1, v_2]$1, Jaccard $[v_1, v_2]$2, and Hausdorff $[v_1, v_2]$3 μm, outperforming SwinUNETR and CShaper++. Lineage tracing reaches over $[v_1, v_2]$4 accuracy at the 550-cell stage, far surpassing prior cell loss rates.

GAN module yields high-quality pseudo-nucleus images (PSNR $[v_1, v_2]$5–$[v_1, v_2]$6 dB, SNR $[v_1, v_2]$7–$[v_1, v_2]$8). Molecular quantification achieves single-cell and interface-level analysis of marker expression (e.g., E-cadherin gradients).

6. Applications and Biological Insights

Topological CT is designed to generalize transformer architectures to higher-order complex domains. It seamlessly incorporates true high-order relationships (e.g., edge-face, vertex-edge), removing the need for virtual nodes or graph rewiring, and flexibly encodes global and local positional/topological information. Potential extensions include linear-time attention for large complexes, adaptation to higher-dimensional cell complexes (3D, 4D), and integration with generative models or manifold/sheaf-theoretic learning (Ballester et al., 2024).

CTransformer enables spatiotemporally resolved, lineage-aware quantification of molecular markers in entire embryos. It has elucidated key developmental mechanisms in C. elegans: e.g., anterior–posterior E-cadherin adhesion gradients (correlation $[v_1, v_2]$9, $\sigma \in X_2$0), sublineage inheritance asymmetry, tight CV$\sigma \in X_2$1 of adhesion across embryos, and contact-specialized expression patterns, linking them to Wnt and Notch signaling roles (Li et al., 16 Dec 2025).

7. Code and Reproducibility

CT code and resources for the cell complex transformer are available as stated in the published work, although direct URLs are not quoted. For CTransformer in bioimage analysis, all annotated data, trained models, segmentation outputs, and source code (including TUNETr, m2nGAN, MolQuantifier modules) are supplied at https://doi.org/10.6084/m9.figshare.27085657, together with a GUI application for plug-and-play deployments (Li et al., 16 Dec 2025).