A number of algorithms are available for calculating the expression intensities on Affymetrix microarrays. Among them, MAS and RMA are two of the most widely used. Comparative studies using spike-in or dilution controls have suggested that RMA algorithms are more accurate than MAS [21, 22], but other experiments suggest that detection calls effectively filtered MAS data, removing the vast majority of false positive results, and that the filtered-MAS data yielded better results than RMA [23–25]. The contrasting results could be due to the different datasets, assessments, and assessment statistics used in different studies.
In this study, more differentially expressed genes between parental pairs were identified by RMA than by MAS, but smaller proportions of them showed fold changes greater than two. This supports the hypothesis proposed by previous studies [22, 25] that the RMA algorithm is more sensitive, particularly at low expression levels, but this may increase the proportion of false positive results, thereby increasing noise in the data . Given the conflicting results of previous experiments, we analyzed our data using both methods – MAS and RMA – to determine if the results we obtained were consistent across analysis methods.
The entries used in this study were previously tested in a field experiment [16, 18], which showed that the heterotic hybrids exhibited high parent heterosis for biomass yield and that these heterotic hybrids showed greater heterosis as the period of regrowth increased. Our growth chamber results indicated that the heterotic hybrids exhibited mid-parent heterosis, probably due to the shorter length of regrowth at harvest, which we limited to three weeks to avoid possible changes in gene expression due to flowering time differences, and/or to the very different environmental conditions in the chamber compared to the field. Mid-parent heterosis for biomass may not be useful for breeding applications, but it is meaningful for the genetic study of heterosis because the difference between the hybrid and the parental mean is the response variable to be related to nonadditive expression, not their absolute phenotypic performance.
We compared two hybrids expressing heterosis for biomass yield with a third hybrid that did not express heterosis. The heterotic hybrid families had a higher number and a higher proportion of genes exhibiting differential expression and nonadditive expression than did the non-heterotic family using either analysis method (RMA or MAS). Higher proportions of probe sets with expression outside of the parental range were also found in heterotic hybrids compared to a non-heterotic hybrid. At an FDR < 0.15, we found about 300 nonadditively expressed genes in heterotic hybrids based on MAS, about half of which had expression outside the parental range, compared to 16 in the non-heterotic hybrid. Similar patterns were seen with RMA. Our data suggest that genes that have non-additive expression in the hybrid and, more importantly, that have expression levels higher than the high parent or lower than the low parent could play a role in heterosis for biomass yield.
Although the two analysis methods produced broadly similar results, different numbers of probe sets were identified as differentially expressed by the two methods and only a small proportion of these probe sets overlapped. The algorithms use different background correction, normalization, and summarization methods , which could explain the non-concordance between them. Further investigation is needed to determine if one algorithm more accurately identified important genes in this experiment, although based on congruence with the RT-PCR results, MAS appeared to hold a slight advantage.
Our results stand in contrast to Stupar and Springer  who found very little evidence for hybrid gene expression that were nonadditive or that exceeded parental levels, and different from Uzarowska et al  who found a large proportion of genes showing nonadditive expression (90%) and expression outside the parental range (51%) in maize. Our results are broadly similar to those of Swanson-Wagner et al . However, comparisons among experiments for the percentage of nonadditively expressed genes need to be made cautiously for a number of reasons, including the use of different statistical methods and thresholds. Recently, a few studies compared the expression profiles of a set of hybrids simultaneously. Stupar et al  investigated the gene expression profile of six maize inbred-hybrid combinations with varying levels of better parent heterosis on five traits, and found a strong correlation between the number of differentially expressed genes and the level of genetic distance between inbred parents, while the proportions of nonadditive expression among the differentially expressed genes were similar among the hybrids. Interestingly, the hybrid with the smallest genetic distance – and the least high-parent heterosis for seedling traits – exhibited the greatest proportion of nonadditive expression. The authors proposed that nonadditive expression is not correlated with heterosis levels. Guo et al  found that heterosis was correlated with the proportion of additively expressed genes but not with the proportion of genes with expression levels outside of the parental range in a set of 16 maize hybrids.
Our study only analyzed three hybrids, limiting our ability to generalize these results to other hybrids. Perhaps more importantly, our results need to be interpreted cautiously given that we used non-inbred parents. Unfortunately, alfalfa suffers severe inbreeding depression, and true inbred lines are not available. To account for the heterogeneity of F1 hybrid indivduals, we pooled ten individuals for each hybrid. This can potentially lead to erroneous results, if alleles from the heterozygous parents are not present in the progeny in equal frequencies. In this case, the hybrid expression relative to the parental mean may be skewed – for example, if the progeny only received a highly expressing allele from one parent, then the overall hybrid expression level may be equal to or exceed the higher parent, even though the hybrid expression level should be additive. Without evaluating allele-specific expression patterns, this concern is difficult to allay. We examined the heterozygosity of the parents using 41 EST-SSR markers. WISFAL-6 (P1) had 1.92 alleles/marker, ABI408 (P2) had 1.95, and C96-513 had 2.15. Assuming that the SSR allele diversity mirrors the diversity of alleles producing different expression patterns, these results suggest that the three parents would have a similar chance to generate false expression results due to preferential allele inheritance. Therefore, we suggest that our comparisons among the three hybrids regarding the about the number and proportion of genes showing nonadditive expression are valid.
Although higher proportions of the nonadditive expression and expression higher or lower than either parent were found in heterotic hybrids compared to a non-heterotic hybrid in our study, the majority of genes showed additive expression in all hybrid families. We may have underestimated the numbers of genes with nonadditive expression due to limitations in our statistical power for this experiment. However, in maize, although the F1 hybrid between Mo17 and B73 showed significant high parent heterosis for seedling growth, only 22% of differentially expressed genes had nonadditive expression and only a small proportion of them showed expression outside of the parental range, similar to our results . Springer and Stupar  proposed that heterosis could result from the additive expression of multiple genes, whereby particularly low or high expression values that are generally detrimental to the plant are modulated in the hybrid, which expresses an average expression level in a moderate, but more biologically functional range. While this may be true in some cases, the clear differences in expression patterns between hybrid types in our experiment suggests that nonadditive expression may also be important for heterosis expression.
What is heterosis? Heterosis simply represents the manifestation of a phenotype in a hybrid that is different from the expectation of a parental average value for that phenotype, be it yield, plant height, or any other trait. The manifestation of the phenotype – particularly of quantitatively inherited traits like yield – results from the complex actions of many components, including the timing of the expression of various genes, the magnitude and location of their expression, and the interaction of their gene products. The genetic hypothesis for the cause of heterosis that has the most empirical support at the current time is that each parent contains a set of dominant alleles at loci controlling the trait and that at some loci, the other parent has recessive alleles at those loci; thus, hybridization brings these dominant alleles together, with the parents complementing each other and giving the hybrid a larger set of dominant (and desirable) alleles than either parent. Complementary expression patterns – each parent contributing alleles that show higher expression than those at the relevant loci in the other parent – could have the same effect. Under this model, hybrids expressing heterosis should have more nonadditive expression, as we have shown in our alfalfa example. Given that control of complex traits likely involves many genes and given that the expression level of most genes is additive, this model does not exclude the possibility that additivity also plays a role in heterosis, under the model suggested by Springer and Stupar .
Conceivably, only a subset of genes may need to deviate from additivity of expression in order to produce a heterotic phenotype. The extent of nonadditive expression at different development stages and different tissues may vary and across the life cycle of the plant, the expression patterns cumulatively produce the observed heterotic response. Arabidopsis allotetraploids had little overlap between the set of genes exhibiting nonadditive expression in leaves and that in flowers, suggesting a role of developmental stages and tissue types on nonadditive gene regulation . If nonadditively expressed genes truly do underlie heterosis, this result suggests that different genes contribute to heterosis in different tissues and at different developmental stages. Thus, for integrative phenotypes like yield, the cumulative effect of these different genes acting at different places and times could result in heterosis. If this is the case, then the nonadditive expression observed at a single timepoint and in a single tissue, as we assayed here, would only give a small part of the overall picture of how gene expression may affect the ultimate expression of the yield phenotype. Finally, genetic divergence between the parental lines appears to result in more differential expression between parents. Both in our study and in that in Arabidopsis by , a higher proportion of nonadditive expression occurred in hybrids whose parents showed divergent expression levels than in hybrids whose parents had similar expression levels. This suggests that there could be more nonadditive expression in the crosses between more distantly related parents, exactly the type of situation in which agronomically useful heterosis levels are also commonly observed. However, recent results in maize suggest that this may not be the case .
The expression levels of individual genes are themselves controlled by other genes, acting in cis or trans [8, 30]. Thus, heterosis for an ultimate phenotype, in this case, biomass yield, must be controlled by multiple genes exhibiting some level of dominance, with some residing in each parental genome . The genes themselves may also be controlled by a number of other genes, and this control can result in expression levels ranging from additivity to some level of non-additivity. Genes controlling transcript levels have been inferred from experiments mapping eQTL, that is, quantitative trait loci that control the expression of a transcript [5, 30, 31]. Interestingly, no eQTL could be mapped for some genes with highly heritable transcript levels in yeast, suggesting that many loci of small effect and/or epistasis among loci controls their expression .
We know that biomass yield, like many other agronomically important traits, is quantitatively inherited, suggesting that it is controlled by many loci (and possibly by multiple interactions among them), and infer that directional dominance plays a role in its control, at least in the certain hybrids that express heterosis. As a means of understanding the nature of the genetic mechanisms underlying biomass yield and yield heterosis, we identified a suite of genes whose expression in hybrids is phenotypically nonadditive, in some cases falling outside of the parental range, and a subset of which only show that expression pattern in heterotic hybrids. But expression of each individual gene is itself the result of a number of gene interactions, and hence, the regulation of expression of any single gene may also have a complex genetic basis. This complexity shows that the genetic control of quantitative traits is difficult to untangle because many levels of interactions, from genes to gene expression profiles to proteins and metabolites, occur to produce the ultimate phenotype.