These observations call for computational models of enhancer regulation, in which multiple enhancers may work in concert to enhance the expression of a gene. To this end, we focused on 2,206 RefSeq TSSs for which the joint expression of nearby enhancers (the closest ten enhancers within 500kb) is highly predictive of the gene expression. Model shrinkage showed that in most cases, only 1-3 enhancers are necessary to explain the expression variance observed in the linked gene, and generally proximal enhancers are more predictive than distal ones (Fig. 6a, Supplementary Fig. 29b-d, and Supplementary Text). One hypothesis explaining the function of multiple enhancers driving the same expression pattern is that they might confer higher transcriptional output of a gene25,26. Indeed, the number of highly correlated (redundant) enhancers close to TSSs (Supplementary Methods Online) increased with the observed maximal TSS expression over all libraries (Fig. 6b), implying that these enhancers are redundant in terms of transcription patterns but additive in terms of expression strength. Expression redundancy is also common in genomic clusters of closely spaced enhancers (24% of 815 identified genomic clusters,
