Fig. 3

Versatility test scenario for method validation. a Seven signatures are simulated as depicted, representing various interactions or laboratory effects (definitions in Supplementary Table 1). To test dissection of dominating signatures, #1 was chosen to mimic two different experimental protocols affecting all genes. #2 simulates a strong binary cluster for only 1% of all measured genes; it could represent gender specific differences. To test detection of signatures that only affect sample subclasses, #3 mimics a gradual interaction that only exists in 50% of all samples. #4 simulates a one-sided interaction that is typical for disease-specific real world signatures like the stromal-1 signature². To test detection specificity for small signatures, #5 contains 20 genes with only one being anti-correlated. To test detection sensitivity with respect to noise, #6 simulates a weak signal. Finally, #7 tests sensitivity for narrow expression offsets in only 5% of all samples. b Signals for all seven signatures are superposed for random gene and sample orders. Normal noise is added (SD = 0.5). The resulting signal (displayed in an arbitrary gene and sample reference order) is input to signal dissection by correlation maximization (SDCM). c All seven signatures and zero false positives were detected. Detected signature signals are displayed in their known simulated gene and sample orders for a direct visual comparison with (a). All major gradual changes were reconstructed. d Residual signal after seven dissection iterations. e Similarities between simulated and detected signatures (columns represent simulated gene or sample axes for patterns #1-#7, rows correspond to detected axes in best-matching order; formulae in Supplementary Note 8). Red diagonals indicate high detection sensitivity. Black off-diagonals indicate high detection specificity.