Review papers with raw data transparency

Paper review (critical, skeptical, evidence-based)

• Representation learning: train a graph-transformer (CellTransformer) on local cell neighborhoods (radius-box in microns) using a masked center-cell gene prediction objective with a negative-binomial likelihood for MERFISH probe counts Masked neighborhood prediction objective.
• Domain discovery: compute neighborhood embeddings for every reference cell across sections, concatenate across sections, then cluster with GPU-accelerated k-means to obtain discrete spatial domains at multiple resolutions (k values like 25/354/670; additional stability-based choice extended to ~1300) Clustering + stability selection.
• Performance: report improved spatial homogeneity/smoothness and better similarity to Allen CCF (NMI/ARI and Pearson-correlation-based region matching) relative to selected scalable baselines (e.g., CellCharter, SPIRAL; plus gene- and neighborhood-based k-means baselines) Reported metric improvements.
• Scalability + integration: claim nearly perfect consistency for up to ~100 spatial domains across 4 mice with millions of cells and hundreds of tissue sections, and demonstrate generalization to Slide-seqV2 with different gene counts and a larger model Multi-animal + Slide-seqV2 generalization.

• Neighborhood tokens are computed from both cell-type identity and local gene expression, then aggregated through an attention pooling / register token to produce a fixed-size “neighborhood representation” for each reference cell Encoder + pooling architecture.
• Self-supervised objective forces the encoder/decoder to capture predictive structure in neighborhood context that helps reconstruct (masked) center-cell expression distributions, which is a pragmatic way to learn spatially structured latent factors without explicit spatial coordinates used as labels Masked prediction training scheme.
• Graph-transformer framing: the model’s attention is restricted to within-neighborhood adjacency (radius-defined neighborhood graphs), which aligns the inductive bias with physical proximity constraints rather than arbitrary fully-connected attention across tissue Local attention constraints.

• Discrete-domain framing may hide gradients: the paper explicitly notes its objective is not normative, and domains may reflect either discrete regions or gene-expression gradients; clustering via k-means imposes a discretization prior. This is an epistemic choice that could favor CCF-like parcellations even when biology is continuous Discrete vs continuous biological framing.
• Evaluation anchored to CCF: NMI/ARI and Pearson-correlation matching are compared to CCF labels; mis-registration or atlas convention differences can affect scores. The paper itself attributes low ARI/NMI magnitude to registration challenges when comparing to dense MRIs Registration affects NMI/ARI.
• Radius hyperparameter: the neighborhood graph uses a user-specified distance cutoff (paper uses 85 μm for MERFISH datasets). If the optimal physical scale varies across brain structures, model learned features and discovered domain boundaries may shift with this hyperparameter Neighborhood radius used.
• k-means dependence and initialization: domain discovery relies on k-means; the paper uses a stability criterion across multiple random initializations, but this still leaves open whether the metric selects “biologically meaningful granularity” vs an optimization-stable granularity. The authors note stability increases with k and propose ~1300 based on second-derivative crossing of averaged inertia + instability How k=1300 is chosen.
• Smoothing might erase fine boundaries: they report an optional Gaussian smoothing step prior to clustering and observe slight erosion of fine laminar boundaries in cortex consistent with smoothing. That suggests a tradeoff: improved coherence vs boundary fidelity at higher granularity Smoothing erodes fine boundaries.
• Dependence on cell-type labels: the training objective “requires only cell-type labels” and uses cell-type conditioning in encoder/decoder; they also test ablations without cell-type in decoder or without cell-type entirely. Still, downstream embeddings inherit biases from the upstream cell-type taxonomies and mapping used Cell-type conditioning and ablations.

• Data and code availability: the paper states code is publicly available and MERFISH and Slide-seq data are available through Allen/Cellxgene/BrainCellData portals. Reproducibility should be reasonably high if preprocessing and QC steps are precisely specified, but exact hyperparameter details (especially neighborhood construction and embeddings aggregation specifics) must match the paper’s methods exactly Data and code availability.
• Cross-dataset alignment risk: the strongest claims (cross-animal consistency, CCF similarity) depend on registration quality to CCF or between datasets. The paper acknowledges NMI/ARI interpretation is affected by registration difficulty Registration affects atlas metric magnitudes.