xMINT: A Multimodal Integration Transformer for Xenium Gene Imputation
Keywords: Multimodal, Transformer, Imputation, Xenium, Pathology Images
TL;DR: Introducing xMINT, a novel gene imputation approach for Xenium data that integrates spatial gene expressions with pathology images using Transformer models, enhancing imputation accuracy while maintaining efficiency.
Abstract: Xenium provides multimodal data with pathology images and corresponding spatial gene expressions to enhance biomedical studies. However, its limited ability to sequence only around 500 genes introduces complexity in panel design and restricts its capacity for exploration analysis. To address this challenge, some methods are developed to impute genes based on external single-cell RNA sequencing (scRNA-seq) data; however, they have neglected the rich cellular morphology and location information available in the Xenium pathology images. We introduce xMINT $(\underline{\textbf{M}}$ultimodal $\underline{\textbf{In}}$tegration $\underline{\textbf{T}}$ransformer for $\underline{\textbf{X}}$enium), a novel gene imputation method utilizing both gene expression profiles and corresponding pathology images to enhance imputation accuracy for Xenium data. xMINT is small and efficient; yet it has a superior imputation accuracy compared to competing methods.
Submission Number: 20
Loading