Navigating gene networks

Combined results

For the third part, the list of gene links made using text mining was read in in Matlab to do some more statistical analysis. There were different lists of gene links made, each with a different 'treshold'. As explained on the text mining page, this treshold represents the value the strength of the links has to exceed to be included in the network. The goal was again to check the differences between disease and non-disease genes.

Scatter plots: neighbours

First, scatter plots were made, showing the number of neighbours in the gene network. It was expected that disease genes are more linked to other genes in general and to other disease genes in particular. This was the case, and the figures also suggest there is some fixed number of disease neighbours for regular genes which is largely lost for disease genes.

Scatter plot, percentage of disease neighbours versus total number of neighbours, using GAD and a treshold of 2

This image was made using the GAD list of disease genes. You can see two 'peaks' of genes, one for the disease genes (+) and one for the non-disease genes (o). The disease genes are a bit more to the left of the plot, i.e. have a slightly lower percentage of disease neighbours. This may indicate that disease genes have a higher number of neighbours overall, but the same number of disease neighbours. To check these hypotheses, boxplots and an additional scatter plot were created.

Boxplots visualizing the difference in number of general and disease neighbours, both using GAD and a treshold of 2

Analyzing these boxplots, the following conclusion seems probable: disease genes have got more neighbours in general and more disease neighbours, but because 'more disease neighbours' only means an increase of one or two gene links and 'more general neighbours' a higher increase, the disease genes move to the left of the scatter plot.

Scatter plot, number of disease neighbours versus total number of neighbours, using GAD and a treshold of 2

In these scatter plots, the number of disease neighbours vs. the number of neighbours in general is shown. This suggests some linear connection between the two in regular, non-disease genes. For disease genes, higher abberations are visible, perhaps visualizing the partial loss of this linear connection.

General feature comparison

After that, we compared the first neighbours of disease genes with the disease genes themselves and with other 'regular' non-disease genes. To do this, the homology comparisons were repeated, together with the comparisons of gene and exon length.

For all the values, the median of the set was taken. P-values were calculated using the ranksum Matlab command (from the Statistics Toolbox), performing a Mann-Whitney U-test. This tests two data sets for the assumption that they come from an identical distribution. An additional advantage of this test is the possibility to compare data sets of a different sizes.

The three P-values are subsequently for the disease vs. neighbour, the disease vs. non-disease and the neighbour vs. non-disease comparisons.

ENSEMBL - OMIM - no treshold
	Disease genes	NB of disease genes	Non-disease genes	P-values
Gene length	35345	25413	25705	3.6e-4, 1.4e-4, 0.9661	boxplot
Sum of exon lengths per gene	5432	4299	4169	1.5e-8, 1.1e-10, 0.2628	boxplot
Exon length	130	136	133	1.4e-9, 3e-5, 4.6e-4	boxplot
CDS length	108	110	107	0.0590, 0.7174, 0.0079	boxplot

ENSEMBL - OMIM - treshold 5
	Disease genes	NB of disease genes	Non-disease genes	P-values
Gene length	34179	25764	24021	0.2250, 0.0169, 0.3421	boxplot
Sum of exon lengths per gene	6166	4417	4166	0.0028, 6.9e-7, 0.1985	boxplot
Exon length	125	141	134	4.4e-15, 2.9e-9, 1.3e-4	boxplot
CDS length	101	113	107	5.2e-10, 1e-7, 0.0088	boxplot

ENSEMBL - OMIM - treshold 10
	Disease genes	NB of disease genes	Non-disease genes	P-values
Gene length	34528	23863	23297	0.2400, 0.0676, 0.9262	boxplot
Sum of exon lengths per gene	5540	4029	3929	0.0258, 0.0014, 0.8941	boxplot
Exon length	127	150	139	1e-9, 2.8e-10, 1.7e-4	boxplot
CDS length	106	122	103	5e-7, 0.3564, 1e-6	boxplot

ENSEMBL - GAD - treshold 5
	Disease genes	NB of disease genes	Non-disease genes	P-values
Gene length	24287	27980	24021	0.9254, 0.7075, 0.6907	boxplot
Sum of exon lengths per gene	4626	4166	4356	0.0818, 0.1056, 0.7273	boxplot
Exon length	135	135	133	0.5174, 0.9030, 0.5877	boxplot
CDS length	110	107	103	0.2849, 0.0003, 0.0290	boxplot

ENSEMBL - GAD - treshold 10
	Disease genes	NB of disease genes	Non-disease genes	P-values
Gene length	23863	23607	23297	0.4837, 0.7835, 0.5794	boxplot
Sum of exon lengths per gene	4403	3864	4293	0.0981, 0.8523, 0.1803	boxplot
Exon length	139	134	137	0.3221, 0.9714, 0.3490	boxplot
CDS length	110	108	102	0.7220, 0.0025, 0.0041	boxplot

Conclusion: there is a somewhat strange result in the ENSEMBL - OMIM treshold 5 and 10 tables. The exon and CDS length are significantly higher for neighbours of disease genes than for disease or non-disease genes. This effect is not observed when using the GAD database for disease genes.

Homology

In these pictures, a distinct difference is visible between disease and non-disease genes. It seems that disease genes are more common than others for all the different organisms studied. This is stastically confirmed by a Whilcoxin rank sum test in Matlab (the same test as above). All of the homology differences pass the test, i.e. reject the hypothesis that the two samples have the same median. The difference is less pronounced when the test is done using the GAD database of disease genes. Probably, the reason for this is the larger number of 'false positive' disease genes.

Homology - OMIM - no treshold
	Disease genes	NB of disease genes	Non-disease genes	P-values
M.musculus	86.76	83.76	82.76	8.6e-4, 6.3e-6, 0.0588	boxplot
R.norvegicus	84.76	81.76	79.76	0.0054, 1.4e-4, 0.0932	boxplot
D.melanogaster	0	0	0	0.5293, 0.1545, 0.0002	boxplot
G.gallus	61.76	57.76	57.76	0.0900, 0.0506, 0.7092	boxplot
P.troglodytes	98.76	98.76	98.76	0.7843, 0.8369, 0.8780	boxplot
C.elegans	0	0	0	0.9133, 0.0623, 0.0016	boxplot
X.tropicalis	52.76	47.76	48.88	0.1228, 0.1097, 0.9814	boxplot

Homology - OMIM - treshold 2
	Disease genes	NB of disease genes	Non-disease genes	P-values
M.musculus	86.76	82.76	81.76	0.0171, 0.0024, 0.4545	boxplot
R.norvegicus	83.76	79.76	78.76	0.0151, 0.0043, 0.6758	boxplot
D.melanogaster	0	0	0	0.3252, 0.8215, 0.0670	boxplot
G.gallus	61.76	57.76	53.76	0.1131, 0.0070, 0.1796	boxplot
P.troglodytes	98.76	98.76	97.76	0.7808, 0.4669, 0.5222	boxplot
C.elegans	0	0	0	0.8987, 0.2238, 0.0588	boxplot
X.tropicalis	52.76	47.76	46.76	0.1881, 0.0987, 0.7659	boxplot

Homology - OMIM - treshold 5
	Disease genes	NB of disease genes	Non-disease genes	P-values
M.musculus	84.76	81.76	79.76	0.1494, 0.0122, 0.4133	boxplot
R.norvegicus	82.76	77.76	76.76	0.0152, 0.0007, 0.6023	boxplot
D.melanogaster	0	0	0	0.3651, 0.8883, 0.2908	boxplot
G.gallus	56.76	57.76	50.76	0.7066, 0.0833, 0.1608	boxplot
P.troglodytes	98.76	98.76	97.76	0.7930, 0.3825, 0.5360	boxplot
C.elegans	0	0	0	0.7338, 0.2156, 0.3547	boxplot
X.tropicalis	53.76	48.76	44.76	0.2291, 0.0156, 0.2736	boxplot

Conclusion: the homology is always higher for disease genes than for non-disease genes, with the neighbour genes in between. The P-values are higher for higher tresholds, because the gene network is smaller in this cases.

Another observation is that the non-disease homology values are higher than in the 'statistical analysis' part. Perhaps part of the explanation is to be found in the fact that for the statistical analysis, the non-disease genes were just what was left over in the genome database after selecting the disease genes. In this part, the non-disease genes are already in the gene network.

by text mining and statistical analysis

Combined results

Scatter plots: neighbours

General feature comparison

Homology