Response to: An alternative explanation for apparent epistasis
Hemani et al.
We thank Wood et al. for their interesting observations but do not believe that their overall conclusions are consistent with the results presented. First, although we replicate our results in large, independent samples, they do not replicate 19/30 of our reported interactions (Table 1 in [1]) in the InCHIANTI dataset (N=450) at a type-I error rate of 0.05/30=0.002, including none of our reported cis-trans interactions. Despite having insufficient data to draw conclusions on the cis-trans effects, Wood et al. claim that this alternative explanation implies that there remains ‘no compelling evidence for widespread epistasis in humans’.
Table 1.
Meta-analysis of results from discovery and replication cohorts. The analysis followed that of Wood et al. In each cohort the effect of the imputed IncSeq SNP was regressed against the probe levels and the residuals used as an adjusted phenotype. Interaction effects were estimated following Hemani et al. and the results combined using Fisher’s method (see Hemani et al.) using results from all three datasets or just the two replication datasets. Two IncSeq SNPs were either not in the 1000 Genomes reference panel or did not pass imputation quality control. Remaining imputed IncSeq SNPs had imputation accuracy R2 > 0.98 in the Brisbane Systems Genetics Study (BSGS). Of the remaining 26, 24 had interaction p values < 0.05/26 = 1.9e−3.
|
cis/ trans |
Gene (chr) | SNP1 (chr) | SNP2 (chr) | IncSeq SNP from imputed data |
Interaction −log10 P value (three studies) |
Interaction −log10 P value (two studies) |
|---|---|---|---|---|---|---|
| cis | ADK (10) | rs2395095 (10) | rs10824092 (10) | rs67594352 | 3.25 | 2.9 |
| cis | ATP13A1 (19) | rs4284750 (19) | rs873870 (19) | NA | NA | NA |
| cis | C21ORF57 (21) | rs9978658 (21) | rs11701361 (21) | rs11702450 | 6.62 | 5.57 |
| cis | CSTB (21) | rs9979356 (21) | rs3761385 (21) | rs35285321 | 1.64 | 1.63 |
| cis | CTSC (11) | rs7930237 (11) | rs556895 (11) | rs56375235 | 10.53 | 7.88 |
| cis | FN3KRP (17) | rs898095 (17) | rs9892064 (17) | NA | NA | NA |
| cis | GAA (17) | rs11150847 (17) | rs12602462 (17) | rs4889970 | 11.85 | 8.29 |
| cis | HNRPH1 (5) | rs6894268 (5) | rs4700810 (5) | rs10078796 | 10.82 | 4.91 |
| cis | LAX1 (1) | rs1891432 (1) | rs10900520 (1) | rs2185079 | 1.01 | 1 |
| cis | MBLN1 (3) | rs16864367 (3) | rs13079208 (3) | rs67903230 | 4.19 | 3.23 |
| trans | MBLN1 (3) | rs7710738 (5) | rs13069559 (3) | rs67903230 | 3.42 | 2.97 |
| trans | MBLN1 (3) | rs2030926 (6) | rs13069559 (3) | rs67903230 | 5.31 | 3.96 |
| trans | MBLN1 (3) | rs2614467 (14) | rs13069559 (3) | rs67903230 | 3.12 | 2.88 |
| trans | MBLN1 (3) | rs218671 (17) | rs13069559 (3) | rs67903230 | 4.85 | 2.84 |
| trans | MBLN1 (3) | rs11981513 (7) | rs13069559 (3) | rs67903230 | 6.49 | 5.75 |
| cis | MBP (18) | rs8092433 (18) | rs4890876 (18) | rs470929 | 4.08 | 3.27 |
| cis | NAPRT1 (8) | rs2123758 (8) | rs3889129 (8) | rs10093709 | 4.07 | 2.95 |
| cis | NCL (2) | rs7563453 (2) | rs4973397 (2) | rs13019380 | 3.48 | 3.24 |
| cis | PRMT2 (21) | rs2839372 (21) | rs11701058 (21) | rs4819255 | 15.80 | 12.16 |
| cis | SNORD14A (11) | rs2634462 (11) | rs6486334 (11) | rs2354863 | 5.01 | 3.66 |
| cis | TMEM149 (19) | rs807491 (19) | rs7254601 (19) | rs28656784 | 4.82 | 3.57 |
| trans | TMEM149 (19) | rs8106959 (19) | rs6926382 (6) | rs28656784 | 3.14 | 2.91 |
| trans | TMEM149 (19) | rs8106959 (19) | rs914940 (1) | rs28656784 | 3.47 | 3.12 |
| trans | TMEM149 (19) | rs8106959 (19) | rs2351458 (4) | rs28656784 | 4.77 | 4.01 |
| trans | TMEM149 (19) | rs8106959 (19) | rs6718480 (2) | rs28656784 | 4.86 | 3.69 |
| trans | TMEM149 (19) | rs8106959 (19) | rs1843357 (8) | rs28656784 | 3.34 | 3.14 |
| trans | TMEM149 (19) | rs8106959 (19) | rs9509428 (13) | rs28656784 | 3.06 | 2.73 |
| cis | VASP (19) | rs1264226 (19) | rs2276470 (19) | rs4803827 | 4.41 | 3.27 |
Second, applying their method in our discovery and replication datasets [1] fails to abrogate the statistical evidence for epistasis. Specifically, the meta-analysis of these results shows that interaction effects remain for 24/26 epistasis pairs after correcting for effects of the IncSeq SNP (Table 1). For the remaining two pairs (at CSTB and LAX1) we cannot rule out a haplotype effect such as postulated by Wood et al. and this may indeed be a more parsimonious explanation for these two pairs. Haplotype effects are known to be confounding factors in cis-cis interactions, as stated in Hemani et al.
Third, Wood et al. ignore the possibility that the IncSeq SNP is either one of the epistatic causal loci, or in higher LD with the causal loci than the genotyped epistatic SNP and assume that a direct comparison of the interaction p-value before and after linear adjustment of the IncSeq SNP provides evidence for their alternative explanation.
Fourth, for 11 of the cis-cis pairs that were replicated by Wood et al. there is evidence for additional cis-genetic variation to that explained by the IncSeq SNPs [2]. Hence the IncSeq SNPs are not the only (causal) variants in cis and therefore the additive effect of the IncSeq SNPs may contain additive effects of additional variants. Furthermore, these probes are within the 95th percentile of non-additive genetic variation estimated using a pedigree-based method that is completely orthogonal to SNP based methods [3] (Table 2).
Table 2.
Correlation coefficients are calculated between relative pairs in BSGS [4]. PP = parent-parent, PO = parent-offspring, DZ = dizygotic twins, SIB = Sibling pairs not including DZ and MZ twins, MA = monozygotic twins. Estimates of additive (h2) and non-additive (d2) variance components estimated from pedigree data [3]. All probes are within the top 90th percentile of h2 estimates and the 95th percentile of d2 (from 17,994 probes).
| ILMN_GENE | PROBE_ID | PP | PO | DZ | SIB | MZ | h2 | d2 |
|---|---|---|---|---|---|---|---|---|
| ADK | ILMN_2358626 | 0.01 | 0.14 | 0.12 | 0.09 | 0.38 | 0.41 | 0.12 |
| ATP13A1 | ILMN_2134224 | −0.02 | 0.16 | 0.14 | 0.20 | 0.61 | 0.67 | 0.16 |
| C21ORF57 | ILMN_1795836 | −0.02 | 0.15 | 0.17 | 0.23 | 0.47 | 0.51 | 0.08 |
| CSTB | ILMN_1761797 | −0.06 | 0.16 | 0.15 | 0.17 | 0.30 | 0.25 | 0.04 |
| CTSC | ILMN_2242463 | 0.12 | 0.14 | 0.20 | 0.16 | 0.37 | 0.27 | 0.08 |
| FN3KRP | ILMN_1652333 | −0.07 | 0.17 | 0.14 | 0.21 | 0.43 | 0.31 | 0.11 |
| GAA | ILMN_2410783 | −0.05 | 0.16 | 0.14 | 0.13 | 0.39 | 0.39 | 0.06 |
| HNRPH1 | ILMN_2101920 | 0.01 | 0.15 | 0.12 | 0.13 | 0.24 | 0.17 | 0.05 |
| LAX1 | ILMN_1769782 | −0.06 | 0.14 | 0.17 | 0.19 | 0.36 | 0.27 | 0.04 |
| MBNL1 | ILMN_2313158 | 0.02 | 0.18 | 0.16 | 0.18 | 0.42 | 0.18 | 0.11 |
| NAPRT1 | ILMN_1710752 | −0.06 | 0.19 | 0.21 | 0.28 | 0.51 | 0.37 | 0.14 |
| NCL | ILMN_2121437 | −0.02 | 0.14 | 0.18 | 0.14 | 0.40 | 0.31 | 0.08 |
| PRMT2 | ILMN_1675038 | −0.04 | 0.20 | 0.19 | 0.18 | 0.40 | 0.34 | 0.06 |
| SNORD14A | ILMN_1799381 | 0.03 | 0.17 | 0.14 | 0.13 | 0.52 | 0.43 | 0.14 |
| TMEM149 | ILMN_1786426 | 0.06 | 0.27 | 0.23 | 0.17 | 0.49 | 0.41 | 0.09 |
| VASP | ILMN_1743646 | 0.00 | 0.14 | 0.27 | 0.18 | 0.52 | 0.38 | 0.13 |
Fifth, there is evidence of interaction variation for pairs of SNPs that include the IncSeq SNPs themselves. Due to lower minor allele frequencies of the IncSeq SNPs many of the pairwise genotype classes are missing, meaning epistatic effects cannot be tested between well-imputed IncSeq SNP and genotyped SNPs in our discovery data. However, in 3/4 pairs for which epistatic effects can be tested there is evidence for interaction variation between the imputed IncSeq SNP and the SNP from the original pair that was in least LD with it (Table 3).
Table 3.
Epistatic effects between the IncSeq SNP and the genotyped SNP with the lowest LD in BSGS data. IncSeq SNPs were imputed (imputation accuracy R2 > 0.99) against the 1000 Genomes reference panel. There were only 4 pairs that had sufficient data (all 9 genotype classes and a minimum genotype class size of 5 individuals) existing between the IncSeq SNP and corresponding original epistasis SNP with the lowest LD with the IncSeq SNP (denoted with *). Of these one is CSTB that shows no interaction effect. The remaining three have strongly significant effects, and explain more genetic variance than the original interactions in two cases.
| Original analysis (SNP1 and SNP2) Hemani et
al. |
Analysis between IncSeq SNP and * original
SNP |
|||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gene | Probe | Original epistatic SNP1 |
Original epistatic SNP2 |
IncSeq SNP rs id |
4df P value | 8df P value | 8df R2 | 4df R2 | 4df P value | 8df P value | 8df R2 | 4df R2 |
| CSTB | ILMN_1761797 | rs9979356* | rs3761385 | rs35285321 | 12.0 | 17.2 | 0.1 | 0.07 | 0.8 | 25.5 | 0.14 | 0.01 |
| HNRPH1 | ILMN_2101920 | rs6894268* | rs4700810 | rs10078796 | 15.4 | 17.1 | 0.1 | 0.08 | 9.6 | 30.8 | 0.16 | 0.06 |
| MBP | ILMN_2398939 | rs8092433* | rs4890876 | rs470929 | 5.4 | 16.9 | 0.1 | 0.03 | 6.5 | 37.1 | 0.19 | 0.04 |
| VASP | ILMN_1743646 | rs1264226* | rs2276470 | rs4803827 | 5.1 | 15.6 | 0.1 | 0.03 | 7.9 | 81.9 | 0.32 | 0.05 |
Finally, we did not report that epistasis was ‘widespread’ and in fact pointed out that for gene expression additive genetic variation explains much more of the total genetic variation than non-additive variation [1, 3].
References
- 1.Hemani G, Shakhbazov K, Westra H, Esko T, Henders AK, et al. Detection and replication of epistasis influencing transcription in humans. Nature. 2014 doi: 10.1038/nature13005. In Press. [DOI] [PMC free article] [PubMed] [Google Scholar] [Retracted]
- 2.Westra HJ, Peters MJ, Esko T, Yaghootkar H, Schurmann C, et al. Systematic identification of trans eQTLs as putative drivers of known disease associations. Nat Genet. 2013;45:1238-U1195. doi: 10.1038/ng.2756. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Powell JE, Henders AK, McRae AF, Kim J, Hemani G, et al. Congruence of Additive and Non-Additive Effects on Gene Expression Estimated from Pedigree and SNP Data. PLoS Genet. 2013;9 doi: 10.1371/journal.pgen.1003502. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Powell JE, Henders AK, McRae AF, Caracella A, Smith S, et al. The Brisbane Systems Genetics Study: genetical genomics meets complex trait genetics. PLoS One. 2012;7:e35430. doi: 10.1371/journal.pone.0035430. [DOI] [PMC free article] [PubMed] [Google Scholar]