Skip to main content
Back to timeline
arXivSource publication:

Sliced spatial transcriptomics stops stitching slice by slice: HoloGene conditions on sparse Z, lifting whole-brain ARI to 0.467

Synopsis

Serial-section spatial transcriptomics is dense in the XY plane but sparse along Z, and standard coordinate implicit neural representations encode all axes symmetrically, so they do not reflect that sampling asymmetry. HoloGene takes in-plane coordinates as the main input, modulates every layer with a depth condition, and aligns cross-slice covariance within pseudo-label groups derived from training-section expression. On held-out measured sections it reaches the lowest MSE and the highest Pearson, masked Spearman and ARI on MOSTA E16.5 and whole mouse brain, raising whole-brain ARI from the strongest baseline's 0.428 to 0.467.

Source-provided article image: HoloGene: Learning to Lift Sliced Spatial Transcriptomics to Holistic 3D Gene Fields
arXiv

Interpretation

HoloGene reframes estimation of a continuous 3D gene-expression field as a factorization matched to acquisition geometry: in-plane coordinates drive synthesis, and depth enters only as conditioning. Earlier coordinate implicit neural representations encoded the three axes symmetrically, did not explicitly separate dense in-plane measurement from sparse axial measurement, and lacked a cross-slice alignment objective to separate slice-specific technical variation from biology. Under a held-out measured-section protocol, HoloGene attains the lowest MSE and the highest Pearson, masked Spearman and ARI on both MOSTA E16.5 and whole mouse brain, with whole-brain ARI rising from the strongest baseline's 0.428 to 0.467.

In practice, high-dimensional expression is first compressed into a graph-autoencoder latent, the synthesis network regresses only that latent field, and a pretrained decoder maps it back to gene space. The graph-autoencoder compression follows the manifold-learning route used by SUICA-style methods, replacing regression of sparse counts with fitting a smooth latent manifold. The spatial neighbor graph is built from training sections only, test sections take no part in graph construction or message passing, the latent dimension is typically small, and the objective is mean squared error over valid spots.

Depth information enters through per-layer FiLM modulation, while a signature network predicts a Leiden pseudo-label distribution from 3D coordinates as coarse tissue-state guidance. FiLM here turns the depth condition into per-layer scale and shift parameters, with the signature condition and the depth condition each contributing to those two parameter sets. In ablations, replacing the isotropic variant that feeds XYZ coordinates directly with FiLM-conditioned synthesis improves MSE and Pearson on both primary datasets.

Cross-slice batch differences are constrained by a signature-stratified CORAL regularizer that aligns second-order statistics only for signature groups with enough spots in at least two slices. A coordinate regression objective alone does not explicitly separate slice effects from independently sequenced sections, and CORAL adds that alignment. On whole mouse brain, enabling CORAL lowers slice-identity leakage from 0.1966 to 0.1770 and raises annotation retention from 0.4929 to 0.5102.

Perspective

Evaluation uses a slice-wise measured-section holdout: sections are split 80% training and 20% testing along Z, all spots from held-out sections are excluded from graph construction, pretraining and synthesis training, and the main results are the MOSTA E16.5 mouse embryo and whole mouse brain data, with visual cortex and breast cancer as supplementary robustness checks. The method estimates expression within a supplied pre-registered coordinate domain and does not jointly solve slice registration, so queries at intermediate Z positions are model-based interpolation estimates.

The model produces point estimates without uncertainty quantification and requires pre-registered coordinates; performance degrades under larger residual misalignment and very sparse axial coverage, with the sparsest whole-brain midpoint reaching MSE 2.8928 and Pearson 0.4004. Signature supervision relies on Leiden pseudo-labels, expression reconstruction stays stable over the evaluated resolution and label-corruption ranges while annotation-referenced domain recovery is more sensitive, and behavior beyond those ranges remains uncharacterized.