Transcriptome-wide prediction of miRNA targets in human and mouse using FASTH

Abstract

Transcriptional regulation by microRNAs (miRNAs) involves complementary base-pairing at target sites on mRNAs, yielding complex secondary structures. Here we introduce an efficient computational approach and software (FASTH) for genome-scale prediction of miRNA target sites based on minimizing the free energy of duplex structure. We apply our approach to identify miRNA target sites in the human and mouse transcriptomes. Our results show that short sequence motifs in the 5' end of miRNAs frequently match mRNAs perfectly, not only at validated target sites but additionally at many other, energetically favourable sites. High-quality matching regions are abundant and occur at similar frequencies in all mRNA regions, not only the 3'UTR. About one-third of potential miRNA target sites are reassigned to different mRNA regions, or gained or lost altogether, among different transcript isoforms from the same gene. Many potential miRNA target sites predicted in human are not found in mouse, and vice-versa, but among those that do occur in orthologous human and mouse mRNAs most are situated in corresponding mRNA regions, i.e. these sites are themselves orthologous. Using a luciferase assay in HEK293 cells, we validate four of six predicted miRNA-mRNA interactions, with the mRNA level reduced by an average of 73%. We demonstrate that a thermodynamically based computational approach to prediction of miRNA binding sites on mRNAs can be scaled to analyse complete mammalian transcriptome datasets. These results confirm and extend the scope of miRNA-mediated species- and transcript-specific regulation in different cell types, tissues and developmental conditions.

Figure 1. Number of predicted miRNA target sites for selected miRNAs in human RefSeq mRNAs distributed through a range of free energy scores.

Numbers of predicted target sites per miRNA and its control sequences for (A) miR-1 and its controls with WC nt 2–8; if miR-1 hybridized with perfect WC complementarity this would yield −30.8 kcal/mol (see Methods); (B) let-7a and imposing only the requirement of WC base pairs within nucleotide positions 2–8; let-7a perfect WC complementarity would yield −33.2 kcal/mol; (C) miR-17-5p and its controls with WC nt 2–8; perfect WC complementarity would yield −44.5 kcal/mol; (D) miR-324-3p and its controls with WC nt 2–8; perfect WC complementarity would yield −52.8 kcal/mol; and (E) miR-129 and its controls with WC nt 2–8; perfect WC complementarity would yield −41.4 kcal/mol. Blue bars show distributions for native miRNAs, red bars for mononucleotide-shuffled controls (MS), and green bars for first-order Markov controls (FOM) (see Supplementary Text).

Figure 2. Number of predicted miRNA target sites and signal-to-noise ratio in human RefSeq mRNAs distributed through a range of free energy scores.

Numbers of predicted target sites (A) imposing only the requirement of WC base pairs within nucleotide positions 2–8; (B) requiring WC base pairs at positions 2–8, plus <6 mismatches and GU pairs at position 15 and beyond (see text). For (A) and (B), blue bars show distributions for native miRNAs, red bars for mononucleotide shuffled (MS) controls, and green bars for first-order Markov (FOM) controls (see Methods). Signal-to-noise ratio based on MS controls (red bars) and FOM controls (green bars), (C) requiring WC base pairs at nucleotide positions 2–8 alone, and (D) requiring WC base pairs at positions 2–8, plus <6 mismatches and GU pairs at position 15 and beyond.

Figure 3. Calculated free energy of duplex formation at the miRNA target sites predicted by our approach, versus the minimum possible free energy for that miRNA binding with perfect WC base-pair complementarity to a (theoretical) target site.

Free energy scores of predicted target sites (y-axis) are plotted against the free energy score of each miRNA, where each of 313 human miRNAs binds to a target site with perfect WC base pair complementary, imposing the requirement(s) of (A) WC base pairs within nucleotide positions 2–8; and (B) WC base pairs within positions 2–8, plus <6 mismatches-and-GU pairs at position 15 and beyond (see text for details). The red line is the non-parametric local fitted line.

Figure 4. Number of predicted miRNA target sites with perfect WC base-pair complementarity at the 5′ end and at the 3′ end of miRNAs in human RefSeq mRNAs.

Numbers of predicted target sites and ratios (the number of predicted target sites with perfect WC base pairs at 5′ end divided by the number of predicted target sites with perfect WC base pairs at 3′ end) for known miRNAs, and mononucleotide shuffled (MS) and first-order Markov (FOM) controls (see Methods), with perfect WC base pair complementary at the 5′ end (yellow bars) and 3′ end (green bars) of miRNAs, imposing the requirement(s) of (A) WC base pairs within nucleotide positions 2–7; (B) WC base-pairs within positions 2–8; (C) WC base pairs within positions 2–7. plus <6 mismatches-and-GU pairs at position 15 and beyond, and 40% free energy threshold; and (D) WC base pairs within positions 2–8, plus <6 mismatches and GU pairs at position 15 and beyond, and 40% free energy threshold (see text for details). ‘WC base pairs at 5′ end’ indicates the number of target sites when positions are counted from the 5′ end of the miRNAs, and ‘WC base pairs at 3′ end’ indicates these values when positions are counted from the 3′ end of the miRNAs (i.e. enforcing perfect WC base pairs at the 3′ end of miRNAs).

Figure 5. Number of predicted miRNA target sites with perfect WC base-pair complementarity at the 5′ end and at the 3′ end of miRNAs in human RefSeq mRNAs, and their signal-to-noise ratios.

Numbers of predicted target sites and signal-to-noise ratios for known miRNAs (blue bars) and mononucleotide shuffled and first-order Markov controls (red bars), with perfect WC base pair complementarity at the 5′ end (yellow bars) and 3′ end (green bars) of miRNAs, imposing the requirement(s) of (A) WC base pairs within nucleotide positions 2–7; (B) WC base pairs within positions 2–8; (C) WC base pairs within positions 2–7, plus <6 mismatches-and-GU pairs at position 15 and beyond, and 40% free energy threshold; and (D) WC base pairs within positions 2–8, plus <6 mismatches-and-GU pairs at position 15 and beyond, and 40% free energy threshold (see text for details). ‘WC base pairs at 5′ end’ indicates the number of target sites when positions are counted from the 5′ end of the miRNAs, and ‘WC base pairs at 3′ end’ indicates these values when positions are counted from the 3′ end of the miRNAs (i.e. enforcing perfect WC base pairs at the 3′ end of miRNAs).

Figure 6. Effects of other binding parameters.

Number of predicted target sites and signal-to-noise ratio of prediction with different seed lengths: WC base pairs at nucleotide positions 2–7 and 2–8 inclusive, (A) on mononucleotide-shuffled (MS) controls, and (B) on first-order Markov (FOM) controls. (C) Number of predicted target sites (blue bars) and signal-to-noise ratio (red bars) when different parameters are used, based on MS controls. Parameter 1: WC base pairs with at most one GU pair within nt 2–7, plus <6 mismatches and GU pairs at nt ≥15. Parameter 2: WC base pairs within nt 2–7, plus <6 mismatches and GU pairs at nt ≥15. Parameter 3: WC base pairs with at most one GU pair within nt 2–8, plus <6 mismatches and GU pairs at nt ≥15. Parameter 4: WC base pairs within nt 2–8, plus <6 mismatches and GU pairs at nt ≥15. Parameter 5: WC base pairs within nt 2–8, maximum of one loop at nt 9–14, plus <6 mismatches and GU pairs at nt ≥15. A 40% free energy threshold was applied in every case.

Figure 7. Proportion of predicted miRNA target sites (for 313 human miRNAs) as a function of free energy score, in each segment (5′ UTR, CDS and 3′ UTR) of human RefSeq mRNAs.

Line smoothed for ease of interpretation.

Figure 8. Experimental validation of predicted target sites.

Y-axis: relative luciferase activity (see Methods). (A) has-miR-15a on TSPYL2 and BCL2; (B) has-miR-324-3p on CREBBP, DBL2 and WNT9B; and (C) has-miR-17-5p on TNFSF12.