Multi-task benchmarking of spatially resolved gene expression simulation models

Computational methods for spatially resolved transcriptomics (SRT) are often developed and assessed using simulated data. The effectiveness of these evaluations relies on the ability of simulation methods to accurately reflect experimental data. However, a systematic evaluation framework for spatial simulators is currently lacking. Here, we present SpatialSimBench, a comprehensive evaluation framework that assesses 13 simulation methods using ten distinct STR datasets. We introduce simAdaptor, a tool that extends single-cell simulators by incorporating spatial variables, enabling them to simulate spatial data. SimAdaptor ensures SpatialSimBench is backwards compatible, facilitating direct comparisons between spatially aware simulators and existing non-spatial single-cell simulators through the adaption. Using SpatialSimBench, we demonstrate the feasibility of leveraging existing single-cell simulators for SRT data and highlight performance differences among methods. Additionally, we evaluate the simulation methods based on a total of 35 metrics across data property estimation, various downstream analyses, and scalability. In total, we generated 4550 results from 13 simulation methods, ten spatial datasets, and 35 metrics. Our findings reveal that model estimation can be influenced by distribution assumptions and dataset characteristics. In summary, our evaluation framework provides guidelines for selecting appropriate methods for specific scenarios and informs future method development.