Research map / Architectures and efficiency / RNNs and normalization
STINR: Deciphering Spatial Transcriptomics via Implicit Neural Representation
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 STINR, an implicit neural representation (sine-MLP) that maps 3D spot coordinates (x, y, slice index) to gene expression, jointly trained with a single-cell-reference Poisson deconvolution encoder-decoder for imputation, denoising, spatial domain detection, and cell-type deconvolution of spatial transcriptomics.
Problem
Integrating spatial correlations with gene expression in spatial transcriptomics (ST) is hard: graph-based methods are computationally costly for large spot graphs and inflexible for misaligned slices or sparse spots.
Approach
Replace spatial graph construction with an implicit neural representation whose Lipschitz smoothness implicitly encodes spatial and slice-wise correlations; jointly train it with a gene encoder-decoder that deconvolutes cell-type proportions using single-cell references under a Poisson likelihood.
Key results
- On mouse brain sections, STINR achieved the lowest average MSE/MTD across sampling rates and noise levels versus GraphST and GNTD (e.g., section 1 SR 0.6 MSE 0.156 vs GNTD 0.158), at about 10 s versus 24-39 s for graph baselines.
- On DLPFC spatial domain detection, STINR achieved the highest ARI on all 8 slices (e.g., #151673 ARI 0.653 vs GraphST 0.575; #151674 0.713 vs 0.606).
- On mouse embryo E11.5 (30124 spots), STINR ran in 27 s versus 718 s for GraphST while achieving lower MSE (0.244 vs 0.284 at SR 0.6).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STINR for spatial domain detection | DLPFC · DLPFC #151673 | ARI: 0.575 -> 0.653 | Helped |
| Use STINR for gene imputation | Mouse brain section 1 · Mouse brain section 1, sampling rate 0.7 | MSE: 0.173 -> 0.168 | Helped |
| Use STINR for gene imputation | Mouse embryo E9.5 · Mouse embryo E9.5, sampling rate 0.7 | MSE: 0.215 -> 0.198 | Helped |
| Use STINR for DLPFC spatial domain detection | DLPFC · DLPFC sections | ARI | Helped |
| Use a suitable network configuration such as width 200 and depth 4 (ablation) | STINR · Mouse embryos, sampling rate 0.7 | MSE: 0.223; 0.222; 0.220; 0.221; 0.225; 0.231; 0.234 | No clear effect |
| Use a suitable network configuration such as width 200 and depth 4 (ablation) | STINR · Mouse embryos, sampling rate 0.7 | MTD: 0.070; 0.069; 0.068; 0.069; 0.071; 0.073; 0.074 | No clear effect |
| Add the reported loss components and jointly model multiple slices (ablation) | STINR · Mouse embryos, sampling rate 0.7 | MSE: 0.282; 0.280; 0.226; 0.275; 0.224; 0.273; 0.221 | Helped |
Surprising findings
- Modeling multiple slices jointly in one INR outperformed per-slice fitting, with the TV loss giving the largest individual gain in ablation.
- GNTD, a graph-based method, is competitive with STINR at some settings, while STINR's advantage is mostly in MTD and runtime.
Limitations
- Graph-based baselines are not used for every task, and the mouse embryo comparison omits GNTD because slices are too misaligned for it.
- Evaluation of imputation/denoising relies on synthetic dropout and Gaussian noise rather than ground truth for real sparse data.
- Hardware and compute details are not stated in the main text.
Experimental setup
Datasets: 10x Genomics mouse brain coronal sections Stereo-seq mouse embryo organogenesis (STDS0000058) DLPFC (spatialLIBD) Human squamous cell carcinoma (GSE144240) Human lymph node multi-omics (GSE263617) Mouse brain single-cell data (mousebrain.org) GSE121891 GSE144136 GSE144236
Benchmarks: Mouse brain coronal sections (imputation/denoising) Mouse embryo organogenesis E9.5-E11.5 DLPFC spatial domain detection Human SCC spatial domain detection Human lymph node multi-omics DLPFC cell-type deconvolution
Metrics: MSE MTD (mean Tweedie deviance) ARI AUC running time (s)
Baselines: KNR Decision tree GraphST GNTD STAGATE IRIS STitch3D SpatialGlue
Builds on: implicit neural representation GraphST STAGATE STitch3D
Closest papers by meaning
- SUICA: Learning Super-high Dimensional Sparse Implicit Neural Representations for Spatial Transcriptomics (ICML 2025 · similarity 0.90)
- Preserving DEG Rankings for Gene Discovery in Histology-Based Spatial Gene Expression Prediction (NeurIPS 2026 · similarity 0.88)
- Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images (ICLR 2025 · similarity 0.87)
- Global Context-aware Representation Learning for Spatially Resolved Transcriptomics (ICML 2025 · similarity 0.87)
- SToFM: a Multi-scale Foundation Model for Spatial Transcriptomics (ICML 2025 · similarity 0.87)
- Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.