How SUICA Boosts Spatial Transcriptomics with Implicit Neural Representations

Spatial Transcriptomics Overview

Spatial Transcriptomics (ST) simultaneously records gene expression levels and spatial coordinates on a tissue slice, producing a high‑dimensional matrix that links molecular profiles to precise locations. This enables functional maps of cellular states and microenvironments within the tissue.

Motivation for Computational Enhancement

Resolution‑cost trade‑off: Higher probe density and deeper sequencing dramatically increase experimental expense (e.g., Stereo‑seq costs > $4,000 / cm²).

Signal sparsity and noise: Limited mRNA capture per spot leads to high dropout rates, causing loss of low‑abundance or regulatory genes.

Cross‑platform heterogeneity: Different platforms vary in probe layout, sequencing depth, and background noise, hindering integration across experiments.

Computational methods such as super‑resolution reconstruction, deep denoising, and missing‑value imputation can improve data quality without substantial cost increases, benefiting downstream analyses like cell‑cell communication, disease region annotation, and multimodal modeling.

SUICA: Unified Model Based on Implicit Neural Representations and Graph Autoencoders

SUICA treats each spatial spot as a node in a graph and constructs an adjacency matrix from spatial proximity. A graph‑convolutional encoder compresses the ultra‑high‑dimensional gene expression into a low‑dimensional latent space while preserving local spatial context. The latent embeddings are then input to an implicit neural representation (INR) network that learns a continuous mapping from spot coordinates to the latent space. The INR‑predicted embeddings are decoded by the graph autoencoder to reconstruct the full‑dimensional gene expression, effectively providing a continuous, high‑resolution expression field.

Experimental Validation

Benchmarking was performed on mouse embryo Stereo‑seq data and mouse brain Slide‑seq data. SUICA was compared against baseline implicit neural representation models (FFN, SIREN) on the task of predicting expression at unmeasured locations (super‑resolution). Across multiple quantitative metrics (e.g., reconstruction error, correlation with ground truth), SUICA outperformed baselines, accurately restoring spatial gene patterns and enhancing weak signals such as the developmental gene SEPT3 . Clustering of the generated expression profiles showed that SUICA‑derived cell types closely matched true cell identities and preserved finer organ‑level spatial structures.

Noise Reduction and Dropout Imputation

To assess denoising, Gaussian noise was added to the data; for imputation, 70 % of gene expression values were randomly set to zero. SUICA consistently achieved higher scores than competing methods on both tasks, demonstrating superior ability to reduce noise and recover true expression from dropout‑affected measurements.

Conclusion

SUICA leverages graph autoencoders to embed high‑dimensional, sparse spatial transcriptomics data and implicit neural representations to map continuous spatial coordinates to gene expression. The model delivers more accurate, denoised, and biologically meaningful expression maps without additional experimental cost.

Paper: https://go.hyper.ai/C6Zcl

Code example

来源：HyperAI 超神经
本文
约2600字
，建议阅读
9
分钟
东京大学郑银强老师组，麦吉尔大学丁俊老师组共同提出了一种针对空间转录组数据建模的方法 SUICA。