Run time of chess sim module

[vaquerizaslab]

Run time of chess sim module

[nickmachnik]

Hi Misha,
I don't think we have ever systematically checked this, @kaukrise @sgalan , please correct me if I am wrong.

Of the factors that you mentioned, the matrix resolution has the strongest influence on the runtime, simply because of the quadratic growth of the amount of data with decreasing bin size. The runtime should just approximately linearly depend on the number of windows (controlled by step and genome size). I am unsure about the effect of the window size; I think this mostly influences the runtime of the comparison step (probably quadratically), which involves subsetting the data that is already in memory and computing the similarity score; I guess the time here grows quadratically with the window size, but this step is very fast, so the overall runtime might not change so much because of this.
@kaukrise , do you have anything more helpful on this?

[nickmachnik]

In the paper we have one example in Extended Data Figure 6. chess sim took 100 seconds to process human chr19 on a single cpu on a Intel Xeon W, 3GHz with 128 gb ram, with a 1 mb window and 500 kb step size.
The data was from Bonev et al 2017, binned at 5 kb.

[kaukrise]

I don't have too much to add, except that the O/E calculation is by far the most time-consuming step. If you want to run many CHESS calculations on the same dataset, it is strongly recommended to convert your data beforehand. Either to FAN-C (you could use fanc from-cooler or fanc from-txt) or Juicer format, which is very flexible, or using the chess oe command if you are working with sparse TXT matrices.

[magnitov]

Ok, thanks.

I am currently trying this on 20 and 50 kb resolution for the whole human genome, so let's see. What I noticed so far is that the longest part is loading reference/query contact matrices.

@kaukrise, yes, I had to switch to .hic format, since .cool is indeed very time-consuming. What I can actually suggest about this, is to use pre-computed obs/exp matrix for .cool format, which can be easily generated by cooltools: https://github.com/open2c/cooltools. Maybe you can consider this in a future.

[kaukrise]

If you already have O/E matrices, even in .cool format, you should be able to use the --oe-input flag to skip the O/E conversion, isn't that right @nickmachnik?

[magnitov]

@kaukrise, as far as I know, .cool format does not store obs/exp counts, only raw/corrected matrix itself. Therefore, it is rather hard to overcome this and create a new .cool with obs/exp counts only.

[kaukrise]

Okay, then I misunderstood your comment. I thought cooltools allowed you to generate a .cool file that only contains O/E values (as matrix entries).