The PDG-ACE algorithm uses an association of keywords with genes mined from Entrez Gene records. We have developed tools that build these associations in two ways: matching Entrez Gene text against a dictionary of keywords, and naïve recognition of phrases within the text. The first method finds all longest full matches to the dictionary. The second finds the longest non-stopword phrases within the text. In both cases, stopwords are filtered out, using a stopword list consisting of common English words.

We constructed three vocabularies. For each, we first derived an initial vocabulary, and then filtered the keywords to keep only those that are rare in Entrez Gene records. The first vocabulary is based on Medical Subject Headings (MeSH), from which we created a vocabulary by splitting headings to make phrases likely to be seen in text. We created the second vocabulary, meant to eliminate bias due to a particular dictionary, by extracting naïve keyphrases directly from Entrez Gene records. The third vocabulary was created to emphasize keywords related to neurological disorders. To do this, we extracted naïve keyphrases from OMIM 9 records containing the substring "neuro". Figure 2 illustrates the differences among the three vocabularies, which we refer to as the MeSH, NAÏVE and OMIM vocabularies.

Once the initial association is mined, we screen the vocabulary to eliminate keywords that are very rare or very common in Entrez Gene records. Keywords with fewer than three occurrences are eliminated. The threshold for eliminating common keywords uses an approximation to the statistical significance test used in the algorithm. Letting G be the total number of genes, and N be the total number of keywords, and assuming a Bonferroni correction of 0.05/N, we want keywords with at most 0.05⋅G2/N MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaWaaOaaaeaadaWcgaqaaiabicdaWiabc6caUiabicdaWiabiwda1iabgwSixlabdEeahnaaCaaaleqabaGaeGOmaidaaaGcbaGaemOta4eaaaWcbeaaaaa@356D@ occurrences. This narrows the vocabulary to words that are likely to be common across gene pairs and also pass the significance test for over-representation.

Our association-building tools are able to mine from different text elements of the Entrez Gene records. For the MeSH and OMIM vocabularies, we mined the official full name (gene-ref_desc), aliases (gene-ref_syn_E), summary (Entrezgene_summary), annotation from other databases (other-source_anchor), and Gene RIF (gene-commentary_text) elements. For the NAÏVE vocabulary we did not mine the synonyms and other sources, because of the large number of unique terms. Note that in preprocessing we build a list of genes and their locations from an authoritative source. Results presented here are based on hg18 data tables from the UCSC genome browser 10. Genes are also filtered to include only current Entrez Gene records.

The primary input to PDG-ACE is a pair of disease-associated loci and a delta in basepairs from each locus. These inputs define a pair of chromosomal regions from which genes are considered. The algorithm does one run using this observed pair of disease-associated loci, then performs permutations to determine the significance of the observed results.

In each run, each keyword is scored with the number of possible pairs of genes, across the loci, that the keyword describes. All keywords common to at least one gene in each region will have a nonzero score. The observation run assigns a score to each keyword at the observed interacting locus pair, and keywords that have a zero score are filtered prior to the permutation runs. The permutations are run on blocks consisting of the same number of sequential genes as the observed loci. A block is selected by randomly choosing a chromosome arm then randomly picking a block of sequential genes on that arm. If the arm is too small, then another arm is chosen until one that has enough genes is found.

As permutations are run, the rank of each observed keyword score is determined. If, on completion of the permutation runs, the score of a keyword ranks above a user provided threshold, the keyword, its rank, and the corresponding genes from both loci are reported. The p-value for a keyword is the proportion of scores for permutation runs that are greater than or equal to the observation run score. In post-processing, a Bonferroni correction can be applied so the threshold for significance is 0.05/N, where N is the number of keywords in the vocabulary.

We validated our approach using published studies as positive controls and randomly selected locus pairs as negative controls. Two control studies used microsatellite markers as loci, and the rest used genes.

For validation, the positive controls were from seven published studies showing statistically significant gene-gene interactions. These include two breast cancer studies 1112, and studies of osteoporosis 13, anorexia nervosa 14, colorectal cancer 15, asthma 16, and neural tube defects 17. Each of these studies found statistical evidence of gene-gene interactions. Our expectation was that PDG-ACE would find keywords that are over-represented and consistent with genetic interactions predisposing these diseases. The negative controls were pairs of randomly selected genes from Entrez Gene, with the expectation that PDG-ACE would not find over-represented common keywords.

For each locus pair, we tested loci defined by deltas from 10³ basepairs (KBP) to 10⁶ basepairs (MBP) from each gene's transcription start site. At each delta, we ran PDG-ACE in duplicate, and performed trials to ensure a sufficient sample as described below. Tests were performed in parallel, using all three vocabularies (OMIM, MeSH, and NAÏVE). In all but one case, results for deltas greater than 500 KBP showed no significant keywords; we report only smaller regions.

Several trials may be needed to determine the number of permutations at which the sample of the genome yields a consistent measure of significance for rare keywords. Each test is run in duplicate starting with one million iterations. The sample is considered sufficient if the top three keywords are identical, and in the same order in both runs. If that criterion is not met, we increase the number of permutations and re-run the test in duplicate until the criterion is met.

Table 1 shows hits for the positive controls and Table 2 shows hits for the negative controls, both using the MeSH vocabulary of 2531 keywords. Note that the pattern of hits in the positive controls is significantly different from the negative controls (χ² p-value < 0.01). In general, the strongest evidence for multi-gene effects is near the observed loci (+/-1 KBP), and the pattern of hits is consistent with p-values from the control studies. As expected, in most, but not all, cases, significantly over-represented, common keywords are consistent with disease etiology. For example, in the first breast cancer study, the COMT-CCND1 genetic interaction is significant (p-value 0.014 in the interaction study) and the over-represented, common keyword is "estradiol" (p-value 0.041). "Estradiol" is used in the same context at both loci, and may offer insight into hormone sensitive breast cancer etiology.