counts for peak i across all cells. The TF-IDF matrix is then computed as TF × IDF. We found that, when applied to scATAC-seq data, this implementation often results in nonzero values in the TF-IDF matrix having low variance and a mean very close to zero and a poor ability to discriminate between cell types. We developed a simple modification to the TF-IDF weighting scheme that improves the results of LSI when applied to single-cell chromatin data. In our modified method we compute the inverse document frequency as IDF = N/ni and TF-IDF as log(1 + (TF × IDF) × 104).
