GPattern-Bench: Benchmarking Gene Spatial Pattern Classification in Subcellular Spatial Transcriptomics

Jinghao Wang, Xilin Dang, Jiezhong Qiu, Pheng-Ann Heng

Keywords: benchmark, spatial transcriptomics, gene spatial pattern classification

TL;DR: we propose a benchmark, GPattern-Bench, for Gene Spatial Pattern Classification and a correspondent model GPSNet

Abstract: Subcellular transcriptomics technologies have revolutionized our ability to study gene expression and its spatial context at single-cell resolution. One fundamental yet underexplored task is \textit{\textbf{gene spatial pattern classification}}, which involves predicting localization patterns for genes within a single cell. To this end, we introduce \textit{GPattern-Bench}, a novel benchmark for this task that unifies evaluation across four established baselines on three diverse datasets, comprising 43 million RNA molecules across 101,000 cells. Given the suboptimal performance of existing machine learning methods, we also develop \textit{GPSNet}, a transformer-based architecture tailored for efficient modeling of spatial transcriptomics data. To address the computational challenges of modeling thousands of RNA molecules in a single cell, we propose a KNN-attention mechanism as a plug-in module for the transformer architecture, enabling the model to efficiently capture spatial dependencies. Extensive experiments on GPattern-Bench demonstrate that GPSNet outperforms existing methods by a significant margin in both accuracy and inference speed, achieving an average F1-macro score of 70\% across the three datasets, a relative improvement of over 30\% compared to the best baseline. We believe GPattern-Bench will facilitate future research in this area, and GPSNet can serve as a strong deep-learning baseline for future methods. We will publicly release GPattern-Bench and GPSNet to the community.

Primary Area: datasets and benchmarks

Submission Number: 2728