Abstract
Motivation. BLAST-based screening of small interfering RNAs (siRNAs) frequently produces false positives in the "grey zone", where alignment statistics are marginally significant but functional affinity is low. RNA language model embeddings carry functional information, but it is unclear which geometric features of these embeddings are predictive. In this work we show that a two-point feature — the Euclidean distance between siRNA and mRNA embeddings — adds predictive information that is not available from one-point features alone.Results. On a curated dataset of 101 siRNAs targeting the PCSK9 mRNA (NM_174936.4) with experimentally determined IC50 values, the Euclidean distance dE(vs, vm) between sense-strand and mRNA-window embeddings adds ΔR² = +0.142 beyond classical features (GC content, site position, cluster membership), raising R² from 0.285 to 0.427 (bootstrap 95% CI for the increment: [+0.031, +0.245]). Because the sense strand matches the targeted mRNA segment, the sense-strand distance primarily measures the sensitivity of the RNA-FM embedding to context length; the antisense-strand distance provides a more direct proxy for the guide strand, though it is still a distance between two embeddings rather than a physical interaction, and it is available for only 69 of 101 siRNAs. The hyperbolic distance dH does not outperform the Euclidean one (ΔR² = +0.105 relative to the same classical baseline), and the two are partly redundant. Within the PCSK9 3'UTR hot-spot region (3520–3570 nt, n = 52), dE captures a within-region signal: ρ(dE, log10 IC50) = +0.477 (bootstrap 95% CI [+0.245, +0.660]). Adding dE to site position raises R² from 0.025 to 0.326. The increment is positive in 96.8% of bootstrap resamples, but its two-sided 95% CI [−0.018, +0.627] includes zero at n = 52; we therefore report this within-region result as suggestive one-sided evidence, not as a confirmed effect. The signal is not reducible to the siRNA norm (ρ(dE, ‖vs‖ | cluster A) = +0.05, p = 0.72), to direction (cosine similarity ρ = +0.13, non-significant), or to absolute position. For the 69 siRNAs with available antisense sequences, the antisense distance dE_anti yields a within-region correlation of ρ = 0.593 (p = 4.6 × 10⁻³, computed on the n = 22 antisense-available siRNAs in cluster A; the small sample size implies a wide confidence interval), stronger than the sense-strand correlation on the same region. As a negative control, we test the radial finite-scale index Gε(r) from our earlier theoretical work; we show analytically and empirically that when applied to Euclidean embeddings via the standard radial projection, Gε(r) collapses to a monotone function of the PCA-norm of the embedding (ρ = 0.997). Adding Gε(r) on top of position and dE does not improve performance (ΔR² = −0.019, two-sided 95% CI [−0.076, +0.034], i.e. within noise). Cross-gene validation on the Huesken dataset (n = 2361, 34 genes) shows that dE does not improve over thermodynamic features under zero-shot evaluation. Within the hot-spot region the mRNA window is nearly constant, so the within-region signal should be read as evidence that the sense-strand embedding carries local efficacy information, rather than as evidence for a pairwise guide–target interaction.Conclusion. A two-point feature — the distance between siRNA and mRNA-window embeddings — is a genuinely new geometric predictor for siRNA efficacy on a single gene, not reducible to a norm, to direction, or to absolute position, and confirmed by the complementary antisense representation. Its usefulness is currently limited to the PCSK9 hot-spot region and does not extend to cross-gene prediction with zero-shot embeddings. We position these findings as a rigorous case study of which geometric features of RNA language model embeddings are predictive and which are not.