In this linear regression model, H3K36me3 and H4K20me1 have the highest regression coefficients, suggesting a heavy influence of these two histone modifications in governing the exon expression levels. Consistently, recent studies have demonstrated significantly more H3K36me3 and H4K20me1 on exons than introns, and H3K36me3 has been validated as a regulator of alternative splicing. Since exon expression is highly correlated with its associated gene expression, we needed to exclude the possibility that we were still modeling the 
correlation between gene expression and histone modifications. Therefore, the model for gene expression was used to predict the expression of cassette exons based on the histone modification levels on cassette exons. The accuracy was 0.63, which is about 12% lower than the model for cassette exons. This suggests that the relationship between histone modifications and gene expression could be different from that between histone modifications and exon expression. A general quantitative relationship between histone modifications and exon expression has been presented above, but it cannot be interpreted as a direct interaction, since the quantitative correlation does not provide a way of distinguishing between direct and indirect associations. Some works have been reported to infer the relationship among histone modifications, non-histone proteins and gene expression, or the interplay among exon splicing, conserved sequence and splicing factors. Those studies used clustering-based Bayesian network learning Folinic acid calcium salt pentahydrate methods to recover the interaction relationships, but the clustering procedure might cause loss of information, and different procedures could yield different network structures. In addition, the expression of alternative exons is to a great extent determined by the expression of the corresponding gene. Thus, to investigate whether a specific histone modification could result in differentiation between exon expression and gene expression, it is necessary to remove the transcription effect from exon expression. Genelevel-normalized exon intensity, which is defined as the ratio of exon expression to gene expression, has been widely used for studying alternative splicing. However, owing to the high-level of inherent noise, some studies using this approach have reported low validation rates for the identification of alternative splicing events. Considering these facts, we applied the partial correlations to remove the transcription effect from exon expression and deduce the putative direct interaction between histone modifications and exon inclusion. Partial correlation has been widely utilized to model gene co-expression network and protein-protein interaction network. A recent study employed partial correlation to study exon co-splicing networks, and achieved a higher statistical power than the approach based on gene-levelnormalized exon intensity. The partial correlation coefficient is the correlation that remains between two variables when the effects of the other variables are regressed away. For example, in order to exclude the possibility that a high correlation between one histone modification and exon expression is due to the association between that histone modification and gene expression, we calculated the partial correlation coefficient between the histone modification and exon expression conditional on gene expression. If the partial correlation remained high, it could be claimed that there is an association between the histone modification and exon expression and this association represents a putative direct regulatory relationship. In addition, the links between different histone modifications on exonic 4-(Benzyloxy)phenol regions were studied, where a high correlation between two modifications is not due to their association with a third histone modification.