HoloGene: Learning to Lift Sliced Spatial Transcriptomics to Holistic 3D Gene Fields

Yumin Zheng, Qingtian Zhu, Ziyan Zhu, Kailu Song, Yinqiang Zheng, Jun DingarXiv 2026 · arXiv

TL;DR

HoloGene is a novel deep learning framework that reconstructs continuous 3D gene expression fields from a series of disconnected 2D spatial transcriptomics (ST) slices. By using an "acquisition-aware" architecture that treats the dense in-plane data differently from the sparse depth data, it effectively bridges the gaps between tissue sections while correcting for technical "batch effects" between slices.

What Problem Does This Paper Address?

Spatial Transcriptomics (ST) is a revolutionary technology that allows scientists to see which genes are active at specific locations within a tissue. However, capturing a true 3D volume is difficult; most datasets are actually stacks of independent 2D slices. This creates three major hurdles for researchers:

  1. Physical Anisotropy: Data is dense and quasi-continuous within a slice ($x, y$ plane) but extremely sparse along the depth ($z$-axis). Standard neural networks treat all three dimensions the same, which doesn't match the physical reality of the sampling.
  2. Inter-slice Batch Effects: Because each slice is processed and sequenced independently, technical variations (noise) can appear as fake biological signals, making it hard to align the "look" of one slice to the next.
  3. The Curse of Dimensionality: Each spot in a tissue can have measurements for over 20,000 genes, most of which are "zero" due to technical limitations (sparsity).

HoloGene aims to "lift" these 2D slices into a holistic, queryable 3D field that preserves biological structures and allows for smooth interpolation between measured sections.

Methods

The authors propose a multi-stage framework that moves away from standard "isotropic" (uniform) coordinate modeling:

  • Manifold Compression: Instead of predicting 20,000+ genes directly, HoloGene uses a Graph Autoencoder (GAE) to compress gene expression into a compact, 64-dimensional latent space. This denoises the data and makes the learning task manageable.
  • Pioneer Signature Learning (PSN): A small network learns to predict coarse "tissue signatures" (derived from unsupervised clustering) based on 3D coordinates. This provides a high-level "map" to guide the more detailed reconstruction.
  • Anisotropic Latent Synthesis: This is the core innovation. Rather than feeding $(x, y, z)$ into a standard MLP, the network treats $(x, y)$ as the primary coordinates and uses $z$ (depth) as a conditioning signal via FiLM (Feature-wise Linear Modulation). This forces the network to learn a "canonical" 2D manifold that evolves smoothly as you move through the $z$-axis.
  • CORAL Regularization: To fight batch effects, the authors use Correlation Alignment (CORAL). This loss function ensures that the statistical distribution of latent features for a specific tissue type remains consistent across different physical slices.

Datasets

The model was evaluated on several high-profile 3D ST datasets:

  • MOSTA E16.5: A comprehensive volumetric atlas of a mouse embryo.
  • Whole Mouse Brain: A large-scale dataset covering complex brain architecture.
  • 3D Visual Cortex & Breast Cancer: Used as supplementary benchmarks for robustness.

The evaluation used a "held-out slice" protocol: the model was trained on 80% of the physical slices and asked to "predict" the gene expression of the 20% it had never seen.

Key Insights and Ideas

  • Decoupling is Key: The "Aha!" moment is that $z$ is not like $x$ and $y$ in ST data. By treating depth as a modulator (FiLM) rather than just another input coordinate, the model avoids the "isotropic bias" that plagues general-purpose neural representations.
  • Signature-Stratified Alignment: Batch correction (CORAL) is much more effective when applied to specific tissue types (signatures) across slices rather than the entire slice at once. This prevents the model from "averaging out" real biological differences between different parts of the organ.
  • Continuous Querying: Because HoloGene learns a continuous function, researchers can query the model at any coordinate, effectively generating "virtual slices" at resolutions higher than what was physically sequenced.

Key Results

HoloGene consistently outperformed general-purpose coordinate networks (like FFN and SIREN) and specialized ST baselines (like STAGE and STINR):

  • Reconstruction Accuracy: Achieved the lowest Mean Squared Error (MSE) and highest Pearson/Spearman correlations on the primary mouse embryo and brain datasets.
  • Biological Preservation: The model was significantly better at preserving downstream biological readouts, such as Cell-Cell Communication (CCC) patterns and Differential Gene Expression (DEG) signatures.
  • Domain Recovery: On the Whole Mouse Brain, it raised the Adjusted Rand Index (ARI)—a measure of how well predicted clusters match real anatomical regions—from 0.428 (best baseline) to 0.467.

Limitations and Future Work

  • Registration Dependency: HoloGene assumes the 2D slices have already been aligned (registered) into a 3D coordinate system. It does not solve the initial 3D alignment problem itself.
  • Point Estimates: The current model provides a single "best guess" for gene expression but doesn't yet offer uncertainty quantification (i.e., "how sure am I about this virtual slice?").
  • Sparse Axial Coverage: While it handles sparsity well, performance inevitably degrades if the physical gap between slices becomes too large to bridge biologically.

Future work could involve integrating the registration step directly into the learning process or adding probabilistic outputs to model technical noise more explicitly.

Summary updated

Abstract

Spatial transcriptomics (ST) provides spatially resolved gene expression measurements, but most 3D ST datasets are acquired as independently processed 2D sections. Computational approaches span slice registration and integration, explicit or generative interpolation, and coordinate-field modeling. Within the last family, implicit neural representations (INRs) offer a continuous, coordinate-queryable parameterization, but standard isotropic coordinate encodings do not explicitly reflect that serial-section measurements are dense in-plane and sparse along the axial direction. We present HoloGene, an anisotropic conditional INR framework for estimating continuous 3D gene-expression fields from pre-registered ST sections. Its central contribution is an acquisition-aware factorization that couples dense in-plane synthesis with sparse axial evolution using established GAE, FiLM, pseudo-label supervision, and CORAL components. HoloGene compresses high-dimensional expression profiles into a graph-autoencoder latent manifold, uses pseudo-label-derived signature conditioning to provide coarse intra-section context, and applies depth-dependent FiLM modulation to model axial variation. A signature-stratified correlation-alignment regularizer encourages cross-slice covariance alignment in latent features. Using held-out measured sections as cross-sectional references, HoloGene improves reconstruction accuracy and biological preservation metrics compared with general-purpose INRs and adapted spatial-transcriptomics baselines. These results support anisotropic conditional modeling as a practical strategy for estimating 3D expression fields from discontinuous ST sections. Code is available at https://github.com/Szym29/HoloGene.
BibTeX
@misc{zheng2026hologene,
  author = {Yumin Zheng and Qingtian Zhu and Ziyan Zhu and Kailu Song and Yinqiang Zheng and Jun Ding},
  title = {{HoloGene: Learning to Lift Sliced Spatial Transcriptomics to Holistic 3D Gene Fields}},
  year = {2026},
  eprint = {2611.20368},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20368},
}
Figure