For a 3D image with totally N voxels, | E|=13| V| and | V|=κ N, where κ (0<κ≤1) is a subsampling factor controlling if only a portion of image voxels are used in the computation.
The Init_Scan module initializes the current scan, setting the components that will be compressed in that scan and the length of the restart interval, according to the subsampling factor and of the number of components in the scan.
Good results were obtained with the following NeuronJ settings: Hessian smoothing scale = 1.0, cost weight factor = 0.7, snap window size = 1 × 1, path-search window size = 800 × 800, tracing smoothing range = 10, tracing subsampling factor = 1, and line width = 1.
It calculates this offset depending on how the system is configured: size of components, subsampling factors, restart intervals, number of scans, and so on.
For identifying taxa present at 1% or greater, we recommend a subsampling size of 250 random reads and a sampling size of 2,500 reads for identifying taxa present at 0.1% (see Additional file 2: Table S1b-c).
Two Markov chains were run in parallel with a subsampling frequency of 100 until convergence was reached (population effective size of 242, maximum difference of 0.105175, mean difference of 0.00381158) (supplementary fig. S3, Supplementary Material online).
Parameters for establishing posterior probability of the Bayesian clusters were based on a chain length of 1′100′000, with a burn-in length of 100′000 and a subsampling frequency of 200.
If the user wants to accurately quantify taxonomic groups present in the sample at a level of 10% or greater, we recommend a subsampling size of 25 100 random reads (see Additional file 2: Table S1a).
The results for a subsampling value of p = 66% confirms that the number of connectivity matrices one needs to compute is more relevant than the percentage of the data matrix actually used to compute them.
