A high performance test of differential gene expression for oligonucleotide arrays

Table 1 Summary of statistical test results

	Incidences						IQR for ranks			Known positives achieving rank
	TP	FP	TN	FN	PPV	Sens	1st Q	Median	3rd Q
Affymetrix Latin Square (known positives per comparison = 14, total, all comparisons = 1274) (rank = 14, maximum achievable positives = 1274)
MAS5	950	13,641	1,134,051	324	7%	75%	13	36	130	335
dChip PM	1,068	14,390	1,133,302	206	7%	84%	6	19	63	558
RMA	1,098	10,406	1,137,286	176	10%	86%	5	12	38	734
logit-Exp	1,037	12,311	1,135,381	237	8%	81%	6	15	53	636
logit-ExpR	1,002	11,667	1,136,025	272	8%	79%	6	15	69	619
logit-t	1,110	345	1,147,347	164	76%	87%	4	8	13	1066
Gene Logic Spike (known positives per comparison = 10, total, all comparisons = 210)
MAS5	24	1,305	263,631	186	2%	11%	151	456	1,283	10
dChip PM	38	1,729	263,207	172	2%	18%	72	255	1,505	19
RMA	91	1,860	263,076	119	5%	43%	21	112	450	41
logit-t	106	79	264,857	104	57%	50%	4	8	21	151
Gene Logic AML (known positives per comparison = 11, total, all comparisons = 605)
MAS5	86	3,473	690,352	519	2%	14%	172	816	3,952	21
dChip PM	84	3,790	690,035	521	2%	14%	64	703	5,296	44
RMA	199	3,504	690,321	406	5%	33%	15	330	5,166	139
logit-t	263	107	693,718	342	71%	43%	4	8	738	349
Gene Logic Tonsil (known positives per comparison = 11, total, all comparisons = 726)
MAS5	239	5,854	826,736	487	4%	33%	35	127	681	81
dChip PM	307	3,760	828,830	419	8%	42%	12	49	418	180
RMA	398	4,693	827,897	328	8%	55%	7	33	653	251
logit-t	490	1,752	830,838	236	22%	67%	3	7	15	524
Gene Logic Tonsil - except two comparisons
MAS5	234	4,540	802,820	470	5%	33%
dChip PM	295	2,378	804,982	409	11%	42%
RMA	381	3,263	804,097	323	10%	54%
logit-t	473	116	807,244	231	80%	67%

Spiked in RNAs are considered positives and all others considered negatives. In the Tonsil dataset, two comparisons resulted in scores much worse than others for all methods (0.75 versus 2 pM and 0.75 versus 75 pM). The last block of results had these comparisons removed. In each block, the first three rows tally t-test results on the MAS5, dChip and RMA indexes with positives having p value < 0.01. Last row tallies t-values of the Logit-t method with positives having |t| > threshold, based on df. Threshold t values correspond approximately to p < 0.01. The first four columns tally calls: TP, true positive; FP, false positive; TN, true negative; FN, false negative. The next two columns indicate performance measures: PPV, positive predictive value TP/(TP+FP); Sens, sensitivity TP/(TP+FN); IQR, Interquartile range for ranks. Ranks of statistics demarking the 1st quartile, median and 3rd quartiles for known positives. The last column shows the number of known positives achieving rank at or below the number of known positives in a comparison. The maximum achievable number for the final column is the total number of known positives for all comparisons.

ISSN: 1474-760X