Research map / Probabilistic and scientific modeling / Molecules and proteins

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

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative latent models for single-cell data · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SUICA, an implicit neural representation pipeline that maps spatial coordinates to graph-autoencoder embeddings and decodes to super-high-dimensional sparse spatial transcriptomics profiles, trained with MSE, L1 and a Dice loss.

Problem

Modeling spatial transcriptomics data, which is spatially irregular, extremely high-dimensional (over 20,000 genes), and zero-inflated, in a continuous and compact way for spatial imputation, gene imputation and denoising.

Approach

Pre-train a GCN-encoder graph autoencoder to compress gene expression into a low-dimensional embedding, fit an INR (FFN or SIREN) from coordinates to these embeddings, then freeze the INR and fine-tune the decoder with a combined MSE/L1/Dice loss to recover sparsity and non-zero fidelity.

Key results

• On MOSTA spatial imputation, SUICA reaches MSE 0.85 (x1e-2) and cosine 0.797 vs. 1.11 and 0.732 for the next best (STAGE), and ARI 0.343 vs. 0.289 (SIREN).
• On MOSTA gene imputation and denoising, SUICA achieves the best MAE, MSE, Pearson and Spearman in denoising (MSE 0.934 vs 1.66 for STAGE).
• Ablation on Embryo E16.5 shows MSE falls from 2.35 (vanilla INR) to 1.47 (with AE, Dice and graph), and on Visium human brain from 9.33 to 5.67 (x1e-3).

Measured effects reported in the paper

Surprising findings

• Adding the autoencoder alone worsened performance on the sparse Visium human brain data (MSE 9.33 to 11.27), with gains only after adding Dice loss and graph context.
• Replacing the learned GAE with PCA lost the sparsity pattern entirely (IoU of zero maps dropped to 0), while SUICA maintained IoU around 0.9.

Limitations

• Case-by-case training: a new model must be trained from scratch for each ST slice, as in NeRF.
• Performance gains depend on high-quality data (spatial density and sequencing depth); on MERFISH with only 238 genes SUICA degenerates to a vanilla INR-like performance.
• Compared with histology-aided methods, TRIPLEX was reported as sensitive to dataset characteristics, and the paper lacks ablation over several INR backbones beyond FFN and SIREN.

Experimental setup

Datasets: MOSTA Stereo-seq Slide-seqV2 mouse hippocampus Visium human brain Visium mouse brain MERFISH human heart

Benchmarks: MOSTA Stereo-seq Slide-seqV2 mouse hippocampus Visium human brain Visium mouse brain MERFISH human heart

Metrics: MAE MSE cosine similarity Pearson correlation Spearman correlation ARI

Baselines: FFN SIREN STAGE TRIPLEX UNI v2 NGP

Builds on: SIREN FFN (Random Fourier Features) Dice loss graph convolutional network STAGE

Closest papers by meaning

• STINR: Deciphering Spatial Transcriptomics via Implicit Neural Representation (CVPR 2025 · similarity 0.90)
• Multi-View Hierarchical Alignment Learning for Spatial Transcriptomics (CVPR 2026 · similarity 0.88)
• Global Context-aware Representation Learning for Spatially Resolved Transcriptomics (ICML 2025 · similarity 0.87)
• Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.86)
• Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.86)
• SToFM: a Multi-scale Foundation Model for Spatial Transcriptomics (ICML 2025 · similarity 0.86)