SUICA: Learning Super-high Dimensional Sparse Implicit Neural Representations for Spatial Transcriptomics

TL;DR

SUICA addresses the challenge of modeling super-high dimensional, sparsely observed spatial transcriptomics data by learning continuous, compact representations via Implicit Neural Representations guided by a Graph Autoencoder. The method decouples coordinate-to-embedding mapping (INR) from embedding-to-expression decoding, employing a decoding head and a regression-by-classification loss (with Dice regularization) to preserve sparsity and numerical fidelity. Across Stereo-seq MOSTA, Slide-seqV2, Visium, and MERFISH, SUICA achieves superior numerical fidelity, stronger bio-conservation, and robust imputation under various degradations, demonstrating a degradation-agnostic, platform-agnostic approach. This work offers a scalable, reference-free pathway to enhance spatial resolution and gene signal recovery in ST analyses, with broad implications for downstream cellular and tissue biology.

Abstract

Spatial Transcriptomics (ST) is a method that captures gene expression profiles aligned with spatial coordinates. The discrete spatial distribution and the super-high dimensional sequencing results make ST data challenging to be modeled effectively. In this paper, we manage to model ST in a continuous and compact manner by the proposed tool, SUICA, empowered by the great approximation capability of Implicit Neural Representations (INRs) that can enhance both the spatial density and the gene expression. Concretely within the proposed SUICA, we incorporate a graph-augmented Autoencoder to effectively model the context information of the unstructured spots and provide informative embeddings that are structure-aware for spatial mapping. We also tackle the extremely skewed distribution in a regression-by-classification fashion and enforce classification-based loss functions for the optimization of SUICA. By extensive experiments of a wide range of common ST platforms under varying degradations, SUICA outperforms both conventional INR variants and SOTA methods regarding numerical fidelity, statistical correlation, and bio-conservation. The prediction by SUICA also showcases amplified gene signatures that enriches the bio-conservation of the raw data and benefits subsequent analysis. The code is available at https://github.com/Szym29/SUICA.

Figures (7)

• Figure 1: Starting with the discretely sampled spots (a) of ST, SUICA performs continuous modeling (b) by aid of the great approximation power of INRs. This approach enables complete profiling of cell heterogeneity, as visualized in the UMAP (c), which further facilitating the discovery of new biology.
• Figure 2: The overall pipeline of SUICA. At training-time, a GAE based on cell graphs is trained, with whose pre-trained decoder concatenated to a INR. The INR then maps spot coordinates to the corresponding gene expressions. SUICA is capable of performing spatial imputation, gene imputation, and denoising.
• Figure 3: Spectral analysis with the embeddings of AE and GAE. GAE yields structure-aware and disentangled embeddings with high-frequency details. GTV: Graph Total Variation.
• Figure 4: Visual comparisons of predicted marker gene expressions on MOSTA dataset chen2022spatiotemporal and Slide-seqV2 Mouse hippocampus dataset stickels2021highly, with the descriptions of the markers attached below the results.
• Figure 5: Spatially visualized comparison on bio-conservations of predicted spots on MOSTA mouse embryo E16.5, Slide-seqV2 mouse hippocampus, and Visium-Mouse brain.