The goal of the test is to benchmark the performance and outputs interpretability of different clustering/matrix decomposition methods using let-418, hpl-2, lin-61 and lin-13 replicates dataset.

Test data includes chrI tracks binned in 100bp and z-scored:

• LET418_N2_L3_IL026^IL025
• HPL2_N2_L3_AA381^AA382^AA162
• LIN61_N2_L3_AA005^AA156^AA160
• LIN13_NA_NA_AA314^AA207^PK013

##  num [1:150724, 1:4] -1.17 -1.17 -1.16 -1.15 -1.08 ...
##  - attr(*, "dimnames")=List of 2
##   ..$ : NULL
##   ..$ : chr [1:4] "LET418" "HPL2" "LIN61" "LIN13"

##         LET418       HPL2      LIN61      LIN13
## [1,] -1.166359 -0.7580633 -1.0294355 -0.8053523
## [2,] -1.166359 -0.7580633 -1.0294355 -0.8053523
## [3,] -1.163778 -0.7464810 -1.0248851 -0.7908842
## [4,] -1.152165 -0.6199542 -0.9705874 -0.7114266
## [5,] -1.081366 -0.6224677 -0.7709940 -0.6868739

• k-means
• Hierarchical clustering
• Correlation based clustering
• PCA
• NMF
• FA
• SOM
• NSFA

We want to compare other models to NSFA. NSFA run on test data (15 iterations) produced 13 factors, so we would use same number of clusters/components/factors for other methods, unless the method can infer the number of clusters from the data.

• Cluster 1: 27.66%
• Cluster 2: 3.91%
• Cluster 3: 11.07%
• Cluster 4: 1.32%
• Cluster 5: 0.11%
• Cluster 6: 3.69%
• Cluster 7: 9.91%
• Cluster 8: 32.43%
• Cluster 9: 0.02%
• Cluster 10: 1.26%
• Cluster 11: 0.3%
• Cluster 12: 0.43%
• Cluster 13: 7.88%