Spatial Information Based Imputation for Spatially Resolved Transcriptomics Data

Abstract

New technologies in Spatially Resolved Transcriptomics (SRT) Data have achieved near single-cell resolution while consisting of spatial data that has an immense impact in discovering various biological insights. High-resolution SRT data are sparse and contain dropouts, which may hinder the accurate interpretation of spatial domains. Hence, computationally available imputation methods are often necessary to impute the dropouts from such datasets. Numerous state-of-the-art (SOTA) imputation methods already exist that are used in general tabular data, dedicated single-cell RNA data, as well as the SRT datasets. However, there is no benchmarking of these imputation methods on these newer SRT technologies’ datasets. In this work, we select seven SOTA imputation methods and provide detailed benchmarking results on five SRT technologies consisting of 23 datasets. We show that there are no single imputation methods that consistently outperform others, demonstrating poor imputation capability in most cases, and also poor performance in identifying valid dropouts. Thus, analyzing these SOTA methods’ performance, we propose a new imputation method, ‘SpaMean-Impute’, designed for SRT datasets that utilize spatial information while mitigating dropouts. Furthermore, the proposed method can detect valid dropouts, unlike the SOTA benchmarked methods. Due to integration of spatial information-based dropout location detection, ‘SpaMean-Impute’ only imputes a valid location while preserving the inherent biological relations of the datasets, unlike SOTA methods. Our proposed method outperforms the SOTA imputation methods across evaluation metrics such as: Adjusted Rand Index(ARI), Normalized Mutual Information (NMI), Adjusted Mutual Information(AMI), and Homogeneity (HOMO). In case of ARI, the proposed method outperforms the SOTA methods on average 16.15%, whereas 18.45% improvement in NMI, 18.96% improvement in AMI, and 13.98% improvement in the case of HOMO. Furthermore, the proposed method is computationally efficient compared to other SOTA methods. For example, compared to the SOTA deep-learning-based imputation methods, the proposed method is approximately 33 times faster and requires, on average, 1500 MB less memory during imputation.