In a weighted gene co-expression network analysis (using WGCNA), the soft-threshold power is recommended as a noise filtering. It consists on raising the correlation to a certain number. To decide this power the scale-free topology is estimated for some powers. The function to estimate this prints:
Power SFT.R.sq slope truncated.R.sq mean.k. median.k. max.k.
1 1 0.9300 3.110 0.996 2960.0 3060.0 3970
2 2 0.7510 1.010 0.964 1750.0 1780.0 2900
3 3 0.1730 0.258 0.806 1170.0 1150.0 2280
4 4 0.0942 -0.183 0.713 833.0 782.0 1870
5 5 0.3800 -0.463 0.777 623.0 559.0 1580
6 6 0.5350 -0.656 0.834 481.0 412.0 1360
7 7 0.6270 -0.797 0.872 381.0 312.0 1190
8 8 0.6870 -0.910 0.900 307.0 241.0 1050
9 9 0.7270 -1.000 0.918 252.0 189.0 936
10 10 0.7490 -1.080 0.928 210.0 150.0 841
11 12 0.7850 -1.190 0.948 150.0 98.0 693
12 14 0.8090 -1.280 0.958 111.0 65.9 582
13 16 0.8290 -1.360 0.968 84.0 45.6 497
14 18 0.8410 -1.410 0.973 65.2 32.1 429
15 20 0.8490 -1.450 0.977 51.6 23.0 375
The recommendations of the FAQ indicate that a SFT.R.sq value should be above
0.8 for reasonable powers (less than 15 for unsigned or signed hybrid networks, and less than 30 for signed networks)
and the mean connectivity below the hundreds.
Others have used the power just as noise filtering without caring much about the scale-free topology fit. I would pick the first power even if the mean connectivity is in the order of thousands because the scale-free topology fit is pretty high, however the slope is puzzling me.
How should the soft-threshold power be selected?
Based on an example question