To assess the performance of our modified TF-IDF method in comparison to alternative methods for computing LSI and the currently top-performing methods for scATAC-seq dimension reduction, we downsampled the total counts for the multimodal PBMC dataset ran a variety of scATAC-seq dimension reduction methods. We performed LSI using our modified method, the original LSI method4 and a recently proposed ‘log-TF’ method (Methods). Furthermore, we included a comparison with cisTopic28 and SnapATAC34 as these methods were highlighted as top-performing in a recent benchmarking study52, as well as the recently developed SCALE method30. Overall, we found that our modified LSI method (LSI (Signac)), the log-TF method (LSI (log-TF)) and SCALE were the top-performing methods, whereas SnapATAC and cisTopic struggled to separate cell types in the downsampled datasets (Fig. 3a). Furthermore, we observed substantial differences in the runtimes for different methods, with cisTopic and SCALE having the longest runtimes (Fig. 3b), although the runtime for SCALE can be improved by using a graphics processing unit with sufficient memory30.
