The biochemical basis of microRNA targeting efficacy

The study addresses how miRNA targeting efficacy varies across miRNAs and target sites and whether repression can be quantitatively predicted from biochemical principles. miRNAs guide Argonaute (AGO) to mRNA sites, with the seed region (positions 2–8) driving recognition. Despite extensive biological roles for conserved miRNA families, quantitative prediction of target repression is limited by sparse measurements of miRNA–target binding affinities. Existing correlative models (e.g., TargetScan7) explain only a small fraction of variance in cellular responses (r^2 ≈ 0.14). The authors hypothesize that accurate measurements of AGO–miRNA binding affinities to diverse sequence contexts, coupled with a biochemical occupancy model, will substantially improve prediction of miRNA-mediated repression, including canonical and noncanonical interactions, and generalize across miRNAs via a neural network predictor.

Prior work established the dominance of seed pairing in target recognition and cataloged conserved targets across mammals. Limited biochemical affinity data existed for few miRNAs (e.g., let-7a, miR-21), with recent high-throughput imaging and slicing measurements informing a slicing model but not the more common non-slicing repression. Correlative predictors incorporate features such as site type, context, and conservation, but performance is modest (TargetScan7 r^2 = 0.14). Noncanonical binding modes (centered sites, pivot sites, 5mer matches) and crosslinking-based maps (CLIP, CLASH, PAR-CLIP) suggested alternative interactions but suffer from crosslinking biases and lacked quantitative affinities. Structural studies showed AGO remodels guide RNA energetics, affecting pairing contributions. Collectively, these studies motivated a comprehensive, quantitative affinity map to underpin a biochemical model of repression.

- AGO-RBNS adaptation: Purified human AGO2 loaded with a synthetic miRNA (initially miR-1) was incubated with a large RNA library containing a 37-nt random region. Five AGO2–miR-1 concentrations (7.3–730 pM, logarithmic) were used against 100 nM library. Protein-RNA complexes were captured via nitrocellulose filter binding to minimize loss of low-affinity binders, followed by reverse transcription, amplification, and sequencing. Replicates showed high reproducibility (k-mer enrichment r^2 = 0.86). - Quantification of affinities: Observed enrichments across the concentration series were fit with an equilibrium binding model accounting for AGO-independent background and site saturation. Maximum likelihood estimation simultaneously inferred relative KD values for site types, including a no-site class and later for any k-mer ≤12 nt. Robustness was assessed by leave-one-concentration-out fits (pairwise r^2 ≥ 0.994). - Unbiased site discovery: Iterative enrichment analysis of 10-nt k-mers at the highest AGO concentration identified enriched motifs; reads harboring discovered sites were removed iteratively until no ≥10-fold enriched 10-mer remained, yielding canonical and multiple noncanonical site types. Relative KD values were then fit for all identified sites. - Extension to additional miRNAs: AGO-RBNS was performed for let-7a, miR-7, miR-124, miR-155 (mammalian), and lsy-6 (nematode). Profiles included canonical and noncanonical sites; for some miRNAs, 3′-only sites with extensive pairing to the miRNA 3′ region were detected. - Flanking dinucleotides: For each canonical site type, the two flanking dinucleotides (5′ and 3′) were enumerated (256 contexts) to derive relative KD across 12-nt sequences, and a multiple linear regression model quantified per-position nucleotide effects across miRNAs and site types. Structural accessibility scores (probability unpaired for positions 1–14) were computed to relate accessibility to affinity. - Cellular repression assays: HeLa cells were transfected with miRNA duplexes; RNA-seq measured endogenous mRNA fold changes. For rare/non-abundant sites, a massively parallel reporter assay tested each site type in 184 3′ UTR contexts. Multiple linear regression attributed repression to specific site types. - Biochemical model of repression: For each mRNA and miRNA, predicted occupancy N was computed from KD values of all 12-nt k-mers with at least four contiguous seed matches in ORFs and 3′ UTRs, modulated by a fitted unbound AGO–miRNA concentration parameter (alpha_g). Repression was modeled as log2(1 + b*N) offset by background binding, with global parameters: b (per-bound-complex added decay), and an ORF-site penalty. 5′ UTR sites were excluded due to negligible effect. Model fit used HeLa transfection datasets for five miRNAs (excluding let-7a due to endogenous expression). - Context augmentation (biochemical+): Added linear adjustments (in log KD space) for structural accessibility, 3′ supplementary pairing propensity, and evolutionary conservation to approximate influences not captured by 12-nt affinities. - Benchmarking: Compared to TargetScan7 (original and retrained on 16 new datasets), and crosslinking-based predictors, across HeLa and test datasets (HCT116, HEK293/FT). - CNN for KD prediction: A convolutional neural network was trained to predict relative KD for 12-nt target k-mers from miRNA (first 10 nt) and target sequence pairs, using >1.5 million KD measurements (six AGO-RBNS datasets) and 16 transfection datasets (68,112 mRNA expression estimates). The CNN and biochemical model were trained jointly to fit KD and repression. Passenger strands were also considered. Generalization was tested on 12 new miRNAs transfected into HEK293FT cells (not present in training).

- Comprehensive affinity landscapes: For six miRNAs, canonical 8mer, 7mer-m8, 7mer-A1, and 6mer sites were the highest-affinity within ≤10 nt sequences; additional noncanonical sites were identified per miRNA, often with single mismatches, wobbles, or bulges in the seed region. Some miRNAs (miR-155, miR-124, lsy-6) exhibited 3′-only sites (9–11 nt pairing to the 3′ region), while centered sites showed little affinity advantage, questioning their general functionality. - miRNA-specific canonical differences: Relative contributions of A1 and m8 features varied widely between miRNAs (e.g., miR-155 7mer-A1 nearly matched 7mer-m8; miR-124 7mer-A1 >9-fold weaker than 7mer-m8). Double-mutant-cycle analysis showed A1 and m8 contributions act largely independently across miRNAs. - AGO remodeling of energetics: Affinity increased with predicted seed-pairing stability, but slopes were far shallower than expected from solution duplex thermodynamics, indicating AGO dampens intrinsic stability differences among miRNAs; lsy-6 had the weakest affinities but less disparity than predicted by solution models. - Strong affinity–efficacy correspondence: Relative KD values correlated tightly with repression for endogenous 3′ UTR sites upon transfection (r^2 ≈ 0.80–0.97 across six miRNAs), for both canonical and evaluable noncanonical sites. Reporter assays confirmed efficacy of rare, high-affinity noncanonical and 3′-only sites. - Flanking dinucleotides: Immediate flanking dinucleotides modulated affinity by ~100-fold; A/U enhanced, G reduced, and C had intermediate effects. The 5′ flanking dinucleotide contributed about twice as much as the 3′. Affinity tracked predicted structural accessibility (r^2 = 0.82), implicating local RNA structure as the primary mechanism underlying flanking-sequence effects. In favorable contexts, some 6mers exceeded poorly flanked 8mers in affinity. - Biochemical repression model: Using measured KD values, the model explained r^2 ≈ 0.34 of mRNA fold-change variance across five HeLa transfections, outperforming TargetScan7 (original r^2 = 0.14; retrained r^2 = 0.28). Incorporating context features (biochemical+) improved to r^2 = 0.37. The model generalized to other cell types/miRNAs (e.g., let-7c in HCT116, miR-124/miR-7 in HEK293), surpassing TargetScan and CLIP-based approaches in multiple benchmarks. - Canonical vs noncanonical: Including noncanonical sites modestly improved predictions; shuffling noncanonical KD values across miRNAs reduced performance, supporting miRNA-specific noncanonical contributions. However, canonical sites dominate repression due to higher abundance despite some noncanonical sites having high affinity. - CNN generalization: CNN-predicted KD values recapitulated the affinity–efficacy relationship for canonical sites across 12 new miRNAs (r^2 = 0.76) and stratified site efficacy across different miRNAs (e.g., site-type-specific differences, r^2 = 0.52). When fed into the biochemical(+ ) model, predictions exceeded TargetScan for individual mRNAs in HEK293FT (overall r^2 ≈ 0.18–0.21; biochemical+ ≈ 0.20), including for mRNAs not expressed in HeLa. - Quantitative parameter insights: Fitted b ≈ 1.8 implies one bound AGO complex nearly triples decay rate (~60% abundance reduction). ORF-site penalty indicated ~5.5-fold reduced efficacy relative to 3′ UTR sites. Performance remained robust over wide deviations in alpha_g, suggesting applicability with approximate miRNA abundance estimates. Estimated alpha_g correlated with lower target abundance and weaker general affinity.

The results demonstrate that miRNA-mediated repression can be quantitatively modeled as a function of site occupancy determined by binding affinity. Direct measurements of AGO–miRNA affinities across exhaustive local sequence contexts reveal substantial miRNA-specific differences in both canonical and noncanonical recognition that are not captured by generic duplex thermodynamics. The strong correspondence between in vitro relative KD and in-cell repression across diverse miRNAs and site types validates the occupancy-based framework and clarifies why canonical sites dominate cellular repression: they provide the most efficient binding and are vastly more prevalent, while high-affinity noncanonical sites are rarer. The pronounced effect of immediate flanking dinucleotides, largely through structural accessibility, explains long-standing context preferences (AU-richness) and can exceed differences between site types. The biochemical model, leveraging measured or CNN-predicted KD values and basic cellular parameters, consistently outperforms correlative and crosslinking-based approaches, generalizing across cell types and miRNAs. Parameter estimates provide biological insight into the per-site impact on decay and diminished ORF site efficacy, while the observation that AGO dampens seed thermodynamic disparities suggests a mechanism for more equitable targeting across miRNAs in vivo.

This work replaces correlation-based miRNA target prediction with a principled biochemical framework grounded in comprehensive affinity measurements. By adapting RBNS to AGO–miRNA complexes, discovering miRNA-specific noncanonical sites, quantifying the profound influence of flanking dinucleotides via structural accessibility, and building an occupancy-based repression model, the authors substantially improve predictive accuracy. A CNN extends these insights to miRNAs without experimental affinity maps, enabling robust predictions across contexts and explaining about half to sixty percent of the variability attributable to direct targeting. Future research should expand affinity profiling to more miRNAs to refine CNN predictions, improve estimation of intracellular free AGO–miRNA concentrations, and incorporate distal mRNA features via end-to-end machine-learning models that operate on full transcript sequences in conjunction with biochemical affinity estimates to further close the gap to theoretical ceilings set by noise and secondary effects.

- Predictive ceiling: Even the best model achieved r^2 ≈ 0.37 on total variance due to experimental noise and secondary effects; after accounting for these, about 50–60% of direct-effect variability is explained, indicating remaining unexplained factors. - Sequence window: Affinity estimation was limited to 12-nt k-mers, which underrepresents contributions from 3′ supplementary pairing and more distal RNA structural contexts. - Context approximation: Structural accessibility, 3′ pairing, and conservation were approximated via heuristic scores rather than direct biophysical measurements; only partially capture distal mRNA features. - Generalization variance: CNN and biochemical model performance varied among miRNAs (e.g., miR-190a when including noncanonical sites), indicating remaining gaps in modeling noncanonical recognition for some guides. - Parameter estimation: The unbound AGO–miRNA concentration (alpha_g) is dataset- and miRNA-specific and requires fitting; accurate a priori estimation in new contexts remains challenging. - Biological scope: 3′-only sites were observed for a subset of miRNAs; centered sites showed limited binding under these conditions, but may exist for other miRNAs or contexts not assayed.