TF-IDF and network construction

[BenxiaHu]

Hello Fulong,
SCAVENGE is a good tool deciphering the function of genetic variant at single-cell level.
I have 2 questions about the algorithm of SCAVENGE.
1: what is the binarized sparse matrix?
2: you used TF-IDF to calculate the weight for each feature. it seems that the IDF in your paper looks a little different from (log(N/(dfi+1))).

[fl-yu]

Hi, Thank you for your interest in our tool!

1. binarized sparse matrix is the read count matrix of feature(peak)-by-cell with all the counts more than 1 being valued as 1, as such, this matrix is binarized and the values in the matrix are either 1 or 0.

2. Yes, it is slightly different than standard idf. This version of idf calculation is commonly used in dimension reduction of scATAC-seq and performs well, for example, Cusanovich2018 and Stuart2021. Actually, there are many variants in tf-idf calculation link, I think the performance will be slightly different although I do not have a comprehensive benchmark.

Hope this is helpful. Please let me know if you have any other questions, Thanks!

[BenxiaHu]

thanks a lot for your explanation.
would you like to explain a little of how you build a nearest neighbor graph from the LSI matrix of N cells and d leading LSIs (d = 30)?
in your paper, it seems that you did not mention how to obtain LSI matrix. maybe I miss some important steps.
Best,

[fl-yu]

Great Happy to explain! I will convert it to the discussion as this is not an issue.

[fl-yu]

LSI matrix is calculated from the binarized count matrix by using TF-IDF and SVD techniques. Next, for each cell, the neighbor is evaluated by euclidean metric from LSI matrix and the top k cells were selected as the k-nearest neighbors (knn). We used a mutual knn graph rather than the original one by considering only the cells are the top k-nearest neighbors to each other, where we reasoned that it is more rigorous for cell-to-cell network construction.

[BenxiaHu]

hello Fulong,
I really appreciate your explanation.
I almost understand how to build the M-kNN.
Another question: do you know the principle of gchromVAR?
I try to read the method the authors described. (https://static-content.springer.com/esm/art%3A10.1038%2Fs41588-019-0362-6/MediaObjects/41588_2019_362_MOESM1_ESM.pdf)

it is not clear how do generate E matrix and correct confounding factors.

Best,