Hilary Parker | Math 155 Project

This web page was produced as an assignment for a course on Statistical Analysis of Microarray Data at Pomona College.

Significance Analysis of Microarrays (SAM) testing:
Three SAM tests were performed to assess significant miRNAs for the various tissue samples.

Comparison between all groups:
The first SAM test performed was comparing each of the groups individually. For this procedure, the ERMS sample and DDLPS sample were removed because SAM requires at least two arrays to perform testing.

Below is a list of the most significant miRNAs from the SAM comparison. H0 is that the miRNAs are not differentially expressed between all of the tissue samples.

miRNA spot number	miRNA name	q-value	reject H0 at alpha=0.05 level?	reject H0 with Bonferroni-adjusted alpha = 3.2e-05?
1146	hsa_miR_122a	0	Yes	Yes
1122	hsa_miR_122a	0	Yes	Yes
993	hsa_miR_145	0.001	Yes	No
969	hsa_miR_145	0.001	Yes	No
583	hsa_miR_143	0.001	Yes	No
75	hsa_miR_29b	0.003	Yes	No
51	hsa_miR_29b	0.004	Yes	No
353	hsa_miR_133b	0.006	Yes	No
523	ambi_miR_7510	0.006	Yes	No
353	hsa_miR_133b	0.006	Yes	No
499	ambi_miR_7510	0.006	Yes	No

Although the q-value is not completely the same as a p-value, it can be interpreted very similarly. The q-value for each miRNA represents the expected proportion of false positives incurred when calling that miRNA significant. Thus, because we don't want a false-positive rate higher than 0.05 (standard), we would not reject H0 for any miRNA with a q-value greater than 0.05. For each of the miRNAs above, however, the q-value is much less than 0.05, so H0 is rejected. Thus, we can conclude that these miRNAs are expressed differentially expressed between all of the tissue samples. However, when using the Bonferroni-adjusted cut-off value, we can only conclude that one miRNA (two tests) is significant.

The plot below shows a more visual representation of the SAM results. The significant genes are in green, and the genes we represented on the table are highlighted.

Comparison between NORM1 and SS:
For the second test, I looked at the NORM1 samples and SS samples (the same ones compared with limma here.

Below is a list of the most significant miRNAs from the SAM comparison.

miRNA spot number	miRNA name	q-value	reject H0 at alpha=0.05 level?	reject H0 with Bonferroni-adjusted alpha = 3.2e-05?
197	hsa_miR_214	0.176	No	No
221	hsa_miR_214	0.176	No	No
1421	hsa_miR_422b	0.176	No	No
1274	hsa_miR_378	0.176	No	No
1397	hsa_miR_422b	0.176	No	No
695	(blank)	0.176	No	No
1499	hsa_miR_1	0.176	No	No
1523	hsa_miR_1	0.176	No	No
719	(blank)	0.176	No	No

Here we do not reject H0 for any of the miRNAs with both the adjusted and un-adjusted cut-off levels, so we conclude that we do not have enough evidence to suggest that any of these miRNAs are expressed differently between the NORM1 tissues and SS tumors.

The plot below shows a more visual representation of the SAM results. The significant genes are in green, and the genes we represented on the table are highlighted.

Comparison of limma and SAM results:
The miRNAs identified using limma and using SAM are somewhat different. The common miRNAs to both lists are:
hsa_miR_1
spot 695
spot 719 (probably the same as 695)
hsa_miR_214

The high-ish level of miRNAs that are common to both shows that even though the SAM results are not "significant," the most significant genes are still consistent.

I am inclined to trust the SAM results slightly more than the limma results, mostly because SAM is a non-parametric method, so it seems like it would be more robust. I also think controlling FDR is important because the FDR is the most biologically appropriate thing to be controlling. We don't need "perfect" results (that is, we can tolerate some false-positives), but we do want to control how "extreme" our results are. If they cause a high FDR, then we probably aren't interested in these miRNAs biologically.

I also chose to use the Bonferroni adjustment for multiple comparisons, in large part because I find this test to be most intuitively satisfying and also slightly conservative.