Table 4.
Comparison between TICA, TACO, CENTDIST, and NMF predictions
| Predictor | Cell line | Recall | Specificity | Geometric mean performance | Enrichment |
|---|---|---|---|---|---|
| TICA | K562 | 0.421 | 0.807 | 0.583 | 2.181 |
| TACO | K562 | 0.140 | 0.938 | 0.362 | 2.258 |
| TICA ∪ TACO | K562 | 0.526 | 0.760 | 0.632 | 2.192 |
| TICA | HepG2 | 0.278 | 0.857 | 0.488 | 1.944 |
| CENTDIST | HepG2 | 0.390 | 0.720 | 0.530 | 1.393 |
| TICA ∪ CENTDIST | HepG2 | 0.585 | 0.643 | 0.613 | 1.639 |
| TICA | GM12878 | 0.424 | 0.611 | 0.509 | NA* |
| NMF | GM12878 | 0.238 | 0.911 | 0.468 | NA* |
| TICA | K562# | 0.202 | 0.792 | 0.400 | NA* |
| NMF | K562 | 0.214 | 0.835 | 0.423 | NA* |
Note: Union of predictors is defined as predicting a positive interaction if and only if it is predicted positive by at least one of TICA and TACO/CENTIDIST (respectively). An interaction is predicted negative if and only if it is predicted negative by both methods. Comparison was performed only on the cell lines indicated (K562 for TACO, HepG2 for CENTDIST, GM12878, and K562 for NMF [23]). * indicates that there is no software available for database-wide comparison. # indicates that only a subset of TFs predicted by NMF to be in complexes are used for comparison. NMF, nonnegative matrix factorization method by Giannopoulou and colleagues [30].