How Many Bits Does a MicroRNA Target Score Need?

Computational tools that predict which genes a microRNA silences output fine-grained confidence scores. MicroRNAs are short RNAs that dampen the expression of the genes they bind. The conclusions people actually draw from these scores are coarse and collective: "miR-X represses the unfolded protein response" or "miR-Y acts on invasion genes." Researchers who build reconfigurable intelligent surfaces face a structurally similar situation. These surfaces are wall-sized arrays of cheap reflectors that steer radio waves. Engineers there have shown that each element can be restricted to one or two bits of phase control with almost no loss in total signal power. The reason is that per-element rounding errors are independent and cancel when summed over many elements. We propose to test whether microRNA target predictions behave the same way when they are summed over the genes in a pathway. We also propose a correction for the case where they do not.

The central hypothesis is that crude, low-bit target calls are sufficient for pathway-level inference once one condition holds: the errors must be independent across genes. They fail precisely because microRNA prediction errors are not independent. Seed-match false positives are shared, because they depend on properties such as transcript length and site count that are common to many genes and many microRNAs. This shared component sets a ceiling on how much signal a pathway can carry, and no amount of added score precision or pathway size can raise it. If the hypothesis holds, the field has been working on the wrong problem when it refines per-site scores. The useful quantity to estimate and remove is the error that recurs across microRNAs.

Core Claim

Discrete phase-shift optimization supplies the analytical template here, and microRNA function inference is the application. On the Optimality of Reconfigurable Intelligent Surfaces (RISs) shows that received power keeps its quadratic growth in the number of elements even under 1-bit phase control. Secure Outage Analysis of RIS-Assisted Communications with Discrete Phase Control quantifies the price: about 60% more elements offsets 1-bit quantization. Several studies place the saturation point at 2 to 4 bits: Energy Efficient Multi-User MISO Communication using Low Resolution Large Intelligent Surfaces, Boya Di et al., and Yali Chen et al.. Each of these results rests on the assumption that quantization errors are independent and zero-mean. We carry over the analysis, not the physics. We ask for the bit-depth at which pathway-level inference saturates, and for the "extra genes" rule that compensates for 1-bit calls. The template also predicts where it breaks down.

Write $q_b(s_g)$ for a $b$-bit quantized target score of gene $g$. Write $t_g$ for its true repression. Let the error $e_g = q_b(s_g) - t_g$ have variance $\sigma^2$ and average pairwise correlation $\rho$ across the $N$ genes of a pathway. Suppose the per-gene true signal has mean $\mu$. The signal-to-noise ratio of the summed pathway score is then

$$\mathrm{SNR}(N) = \frac{N\mu^2}{\sigma^2\,(1 + (N-1)\rho)},$$

which follows directly from the variance of a sum of equicorrelated variables. When $\rho = 0$, SNR grows linearly in $N$ and bit-depth barely matters, which is the surface-array regime. When $\rho > 0$, SNR saturates at $\mu^2/(\sigma^2\rho)$ however large the pathway is. Finer bits reduce $\sigma^2$ but leave the ceiling in place unless they also reduce $\rho$. This clarifies an apparent conflict in the file. Network biologists report that false positives propagate through pathway analyses (Overcoming False Positives in MicroRNA Target Prediction; Kowarsch et al.). Array theory says local errors are harmless. Both can be correct, and $\rho$ determines which regime applies.

Research Program

The first step is measurement. Perturbation data record the gene expression change after a microRNA is added. Using these data, we quantize existing continuous target scores to 1, 2, 3 and 4 bits and to continuous values. We sweep pathway size $N$ using Gene Ontology sets. We then fit the saturation curve above and estimate $\rho$ from it. We cross-check $\rho$ with a direct estimate: for unrelated microRNAs, we compute the correlation of gene-level residuals. An error that recurs across microRNAs with different seeds is systematic by construction.

The second step is correction. It is analogous to the way surface designers estimate the channel before configuring the array. We estimate the shared error component using held-out microRNA perturbation experiments. This is a gene-level offset explained partly by covariates such as 3'-UTR length and total seed-site count, with the remainder left as a free latent term. We subtract this component before quantizing and aggregating. The prediction is that after correction, 1–2 bit calls with a modest increase in pathway size match continuous scores. We will report the empirical compensation factor as the counterpart of the 1.6x rule.

The third step is a stress test in the regime where the systematic error should be largest. Sequencing of Argonaute-bound microRNA/mRNA hybrids reveals regulation of the unfolded protein response by microRNA-320a shows that in Drosha-knockout cells, miR-320a accounts for about 65% of Argonaute interactions. These interactions mostly occur through non-seed pairing in coding sequences. Canonical seed-based scores are therefore biased there, not merely noisy. The theory predicts that $\rho$ rises sharply in this setting, that bit-depth becomes irrelevant, and that only the correction step recovers pathway signal such as the reported unfolded-protein-response enrichment.

Baselines

The naive reference is a 1-bit seed-match call, meaning "has a 7-mer or 8-mer site in the 3'-UTR." It is aggregated by simple counting over the pathway. This is what the surface literature says should nearly suffice, so it tests the hypothesis directly.

The strongest existing method is PRIMITI (PRIMITI). It is a gradient-boosted tree model over more than 150 site features, with carefully filtered negatives. Its transcript-level mode reports an AUC up to 0.96 on the miRTarget RNA-seq repression data. It also outperforms other methods on the unseen Linsley transfection microarray, but no reported result is found on Lacuna for its numerical value there. We will use PRIMITI's continuous transcript scores as the full-precision condition and aggregate them over pathways exactly as the quantized versions are aggregated. PRIMITI is not trained at the pathway level, so the comparison asks whether its precision survives aggregation. Beating it means either of two outcomes. The first is that corrected 1–2 bit scores match or exceed continuous PRIMITI on pathway-level agreement with measured repression, on held-out microRNAs. The second is that corrected continuous scores exceed uncorrected continuous ones by a margin larger than the across-fold spread.

Two network-level competitors round out the comparison. The first is PUMA (PUMA: PANDA Using MicroRNA Associations). It refines sequence priors by passing messages over GTEx tissue co-expression, and 86% (334 of 387) of its predicted microRNA functions overlap with miRdSNP annotations. The second is miRLAB's correlation and ensemble methods (miRLAB). In miRLAB, Hoeffding's D matched perturbation data best, and the file gives no numerical value for it. Every method receives the same gene sets, the same held-out microRNA split and the same tuning budget, with hyperparameters chosen on training microRNAs only. Two ablations isolate the key ingredient. One removes the correction step. The other randomly permutes errors across genes within each microRNA, which destroys $\rho$ while preserving $\sigma^2$. Under the theory, this permutation should make quantization harmless.

Why not just…?

Why not just use the best continuous scores and skip quantization? If $\rho$ is substantial, extra precision cannot raise the pathway ceiling. The saturation fit tests exactly this. Quantization serves as the measuring instrument, and a cheaper predictor is a side benefit. If continuous PRIMITI scores keep improving with pathway size while quantized ones saturate, the hypothesis is wrong and we will report that.

Why not just adjust gene-set tests for 3'-UTR length? This alternative poses a real threat. Covariate-adjusted enrichment may remove most of $\rho$. We will include it as a named baseline. The proposal adds value only if the latent residual component, estimated across microRNAs, explains error that the covariates miss. The comparison between covariate-only correction and full correction settles this directly.

Why not just rely on PUMA's tissue context? PUMA reweights edges using co-expression, but its priors are canonical seed predictions. In the miR-320a setting, those priors are systematically wrong, and co-expression alone cannot identify that. PUMA also needs large tissue cohorts. The Drosha-knockout test shows whether PUMA's outputs carry the high-$\rho$ signature.

Why not just measure binding with crosslinking assays? Binding does not guarantee repression (Bader et al.; Liu et al.), and qCLASH covers one cell line in one condition. We treat qCLASH hybrids as a separate prediction channel with its own quantization and error structure, rather than as ground truth.

Proposed Experiments

The primary data are the miRTarget RNA-seq perturbation set and the Linsley transfection microarray used by PRIMITI. TCGA-derived MCC and PRAD cohorts, accessed through miRLAB, provide an observational check, and the HCT116 wild-type, Drosha-knockout and Dicer-knockout qCLASH data serve as the stress test. The splits hold out whole microRNAs, and seed families are kept within one fold to prevent leakage through shared seeds. The metrics are the Spearman correlation between predicted and measured mean log fold-change per pathway, and pathway-level AUC for discriminating repressed from unaffected sets. Gene-level AUC is reported for continuity with PRIMITI.

Three results would be decisive. The first is whether the fitted $\rho$ is significantly above zero and whether the SNR curve flattens as the formula predicts. The second is whether the permutation ablation restores linear growth. The third is whether, after correction, the bit-depth saturation point falls at 1–2 bits, as it does in surfaces, together with a measured compensation factor. A flat, $\rho \approx 0$ result would also be informative. It would mean that false-positive propagation is overstated at the pathway level and that coarse predictors are sufficient.