Force, Not Just Free Energy: Closing the Mechanical Property Gap in Cap-Independent Translation

Materials engineers have a well-documented blind spot: a new composite gets optimized and certified against one headline number — usually axial compressive strength — while tensile capacity, flexural stiffness, and behavior under seismic or corrosive loads are never measured. The result is a literature full of members whose capacity is known to within a few percent in the tested regime and not at all outside it, as when Jun Wei et al. report a precise 20% gain in axial compression capacity for concrete-filled steel tubes while seismic, fire, and corrosive loading go explicitly uncharacterized. I will argue that RNA biology has the same blind spot, in a form that can be tested and fixed.

When a ribosome initiates translation without the usual 5' cap structure — so-called cap-independent translation, via internal ribosome entry sites (IRESs) or covalently closed circular RNAs — the existing models treat the transcript either as a bag of sequence words or as a static thermodynamic folding. Gritsenko et al. regress IRES activity on positional $k$-mers across 28,669 fragments and explain almost none of the variance; IRESeek adds equilibrium base-pairing probabilities and reaches only modest classification accuracy. But the ribosome is a motor. It grips the RNA, tracks along it, and physically tears apart hairpins and G-quadruplexes as it goes — the very structures that Koch et al. show are strictly required for initiation, and that act as "speed bumps" forcing initiation at non-standard codons in the mechanisms characterized by Tanaka et al. and Pöyry et al.. No study in the compiled literature measures the mechanical property envelope of these transcripts: how much force it takes to unwind a given hairpin in the direction the ribosome actually pulls, how stiff the loop is, or what torsional strain circularization imposes. That is the mechanical characterization gap, transplanted from structural engineering to molecular biology.

Core Claim

Cap-independent translation activity is governed more by the mechanical resistance of structural RNA elements — the force a ribosome must exert to dismantle them, plus the torsional strain of a closed circular topology — than by their equilibrium folding free energy. Two hairpins can have identical thermodynamic stability but very different unfolding resistance, and current models cannot tell them apart. If the claim is right, adding a small set of force-space descriptors to existing models should close a large share of the variance they currently leave on the table, and deliberately engineering mechanical stiffness (while holding free energy fixed) should tune translation output in the laboratory.

Research Program

The program has three linked parts: a computational pipeline that turns sequences into mechanical descriptors, a predictive test on the field's existing benchmarks, and a designed wet-lab perturbation study that can refute the hypothesis outright.

Mechanical descriptor pipeline. For every sequence in the benchmarks below, we compute three descriptor families. First, per-structural-element critical unfolding force, estimated from the folding-energy landscape along a 5'-to-3' pulling coordinate — the direction a translocating ribosome would unwind. Following the standard Bell-type relation, we define

$$F_c \approx \Delta G^\ddagger / \Delta x^\ddagger,$$

where $\Delta G^\ddagger$ is the barrier height from the folded state to the nearest partially unfolded intermediate along that coordinate, and $\Delta x^\ddagger$ is the RNA extension gained in crossing it. We derive both from the same base-pair probability ensembles that IRESeek already computes (via RNAfold-style folding), so the new features cost no extra thermodynamic machinery — they are a re-projection of existing information into force space, which is exactly what makes the comparison to IRESeek a clean test of mechanics rather than of added inputs. Second, bending stiffness and effective persistence length per element, from stacking patterns. Third, a torsional-strain descriptor for covalently closed transcripts, capturing the writhe and supercoiling a circularized RNA must absorb. This last one is the direct analog of the hoop prestress that confinement induces in circular concrete-steel composite sections, which Wei et al. and Li et al. study at the structural scale: closing a member into a ring changes its stiffness even when its material composition is unchanged. IRESeek trains on artificially linearized 174-nucleotide fragments and so models none of this.

Predictive test. We graft the mechanical descriptors onto the existing strongest models, holding architectures and training data fixed, and measure whether prediction improves on three benchmarks: the IRESeek circRNA IRES dataset, the Gritsenko native-fragment regression, and the ROSE ribosome-stalling benchmark. This borrows the methodological pattern from the materials side of the file: Yan et al. show that a model can hit $R^2 > 0.98$ on the single axis it was trained on (compressive strength of lightweight concrete, Elman network RMSE 0.160) while remaining brittle against every unmodeled property; the diagnostic is always whether the uncharacterized axes carry signal once measured.

Designed perturbation library. The decisive experiment uses the toolbox of Koch et al. on their own substrate: the 35-nucleotide P4 stem-loop of the Hoxa9 IRES, whose disruption abolishes initiation. We design variant pairs that are matched in minimum free energy (within a small tolerance) but differ in computed critical unfolding force — achieved by repositioning weak pairs and internal bubbles along the stem to change barrier height at constant overall stability — and, conversely, pairs matched in unfolding force but differing in free energy. Each variant is assayed as a circular RNA reporter (verified by RNase R digestion, which enriches true circular species 8- to 38-fold in Koch et al.), with promoterless NLB fusion reporters to rule out cryptic transcription, and isoform-sensitive polysome-qPCR for ribosome engagement. If protein output tracks computed unfolding force across the $\Delta G$-matched pairs, thermodynamic stability is confounded out and the mechanical claim stands; if it tracks $\Delta G$, we are wrong and will say so. A second library extends Gritsenko et al.'s synthetic multimerization result — they showed that stacking 1 to 8 copies of a recruitment element linearly increases output ($p < 0.003$) — by inserting mechanically tuned spacer hairpins between copies, testing whether output can be shaped by stiffness independently of copy number.

Baselines

The naive reference point is a majority-class classifier with GC-content-matched shuffled sequences as negatives on the IRESeek test set, establishing that the benchmark is not trivially separable and giving the chance-level AUROC of about 0.5. The strongest existing competitor on circRNA IRES identification is the full IRESeek model — a dual-branch network coupling a 1D-convolution/Transformer sequence path with a graph network over thermodynamic base-pair probabilities and motif energies — which reports an AUROC of 0.7223 on its held-out test set of 582 positive and 582 negative 174-nucleotide sequences (IRESeek, NAR 2025). Because IRESeek's own ablations show its structural branch is its decisive ingredient (recall improvements of 1.5%–58.4% over baseline methods), beating it with force-space features specifically tests whether mechanics adds information beyond equilibrium structure. We will retrain both IRESeek and our mechanical variant on the identical published split (3,949 positive / 9,034 negative training sequences) with the same class-balance handling, the same random-search tuning budget and GPU-hours, and identical evaluation code. Beating it means a bootstrap-corrected AUROC improvement of at least 0.03 over 0.7223 with recall improved at matched precision. On the native-fragment task, the strongest competitor is the Gritsenko gradient-boosted $k$-mer regression, which reports a test $R^2$ of 0.18 across the full 28,669-fragment dataset and an AUC-ROC of 0.77 for active-versus-inactive classification (Gritsenko et al., 2017); we use their fragments and split conventions unchanged, and count a material increase in $R^2$ above 0.18 as success. Finally, two ablations isolate the key ingredient: our own architecture with the mechanical descriptors removed (sequence plus thermodynamics only), and a scrambled-mechanics ablation in which the descriptors are randomly permuted across sequences, showing that any gain comes from the descriptor values rather than from the extra parameters or added capacity.

Why not just…?

Why not just add more thermodynamic folding features? This is the most serious alternative, and the field has already run it: IRESeek is precisely the best current version of "feed richer folding information to a deep network," and it plateaus at an AUROC of 0.7223; earlier, Gritsenko et al. found that minimum-free-energy metrics added no significant power over plain $k$-mers ($R^2$ of 0.18). The limitation is structural, not incidental: an equilibrium ensemble assigns identical descriptions to two hairpins with the same free energy but different barrier heights, and it is blind to the torsional state of a covalently closed ring. Our $\Delta G$-matched reporter pairs exist precisely because this alternative could win — if activity tracks free energy across those pairs, the mechanical hypothesis is refuted, and we have designed the experiment so that it can refute us.

Why not just scale up sequence-based deep learning on a bigger oligo library? Weingarten-Gabbay et al. already screened 55,000 designed oligos and expanded the known catalog of cap-independent sequences roughly 50-fold, yet the follow-on regression on 28,669 fragments still explained under a fifth of the variance, and ROSE — a 64-model convolutional ensemble predicting codon-level stalling from ±30-codon sequence context, reaching AUROCs above 83% on ramp regions — can say where ribosomes pause but not whether a novel hairpin can physically stop one. Koch et al.'s P4 result sharpens this: small deletions that break the physical stem abolish translation while leaving most sequence words intact, so a sequence-only model mis-scores exactly the variants that matter. Still, this is a real threat — a sufficiently large library with enough capacity could implicitly learn mechanics — which is why our fixed-architecture feature-addition test, and the deliberately out-of-distribution $\Delta G$-matched pairs that no sequence model has seen, are built to separate implicit learning from explicit mechanical characterization.

Why not just validate with standard bicistronic dual-luciferase plasmids? Because they generate false positives through cryptic promoters and aberrant splicing, as Koch et al. document; they had to build RNase R-verified circular reporters and promoterless reporters precisely to escape this trap. We adopt their entire validation stack rather than re-litigating it.

Why not just measure the mechanics directly with single-molecule force spectroscopy? Direct pulling experiments on a handful of constructs are the gold standard for calibration, and within the compiled literature no study reports direct force measurements on any IRES element — but they are far too low-throughput to characterize the 14,000-plus sequences in the benchmarks. Our computed descriptors scale; single-molecule measurements on a small calibration subset of the P4 variant library, checking that computed unfolding forces rank-order correctly, would anchor the pipeline. No reported result found on Lacuna for existing single-molecule unfolding data on these elements.

Proposed Experiments and Decisive Checks

The computational experiments run on three benchmarks with reported numbers. First, the IRESeek circRNA IRES dataset (4,531 positive, 9,616 negative uniformly 174-nt sequences; published 582/582 test split), metric AUROC against the reported 0.7223, plus recall and MCC at matched precision against reported improvements of 1.5%–58.4% in recall and 2.6%–39.1% in MCC over baseline methods. Second, the Gritsenko native-fragment dataset (28,669 fragments), metrics test $R^2$ and AUC-ROC against the reported 0.18 and 0.77, with the per-category breakdown compared against their dsRNA-virus peak of 30%. Third, the ROSE stalling benchmark on human and yeast ribosome-profiling datasets, metric AUROC on ramp regions against the reported values above 83.0%, testing whether adding mechanical unfolding barriers of downstream structures improves pause prediction; additional profiling data can be drawn from RPFdb (777 samples across 82 studies and 8 species), with Xtail (AUC of 0.96 on simulated data with $n = 3$) as the reference statistical pipeline for decoupling footprint shifts from mRNA abundance in any perturbation experiment. The wet-lab experiments are the P4 $\Delta G$-matched and force-matched variant libraries in circular reporters, with readouts in protein output, RNase R enrichment factors, and polysome association, as specified above.

The decisive check is a single scatter plot: computed critical unfolding force versus measured translation output across $\Delta G$-matched P4 variants. A tight relationship there — with no corresponding relationship in the free-energy-matched direction — would establish that the mechanical property characterization gap is real in this system, and would hand RNA engineers a tuning knob (stiffness) that sequence-only design, as in the multimerization strategy of Gritsenko et al., does not currently have. The applications are concrete and already in the file: circular RNAs are translated in vivo to functional microproteins, from lipid-metabolism regulators in Drosophila (Sheng et al.) to immune-activated microproteins in T cells (Yang et al.), against a growing non-canonical ORF catalog (TransCODE Consortium); every one of those efforts currently designs transcripts without knowing how hard the ribosome will have to pull.

References

  1. Study on the Nonlinear Behavior and Factors Influencing the Axial Compression of High-Durability Fibrous Concrete Wrapped Steel Tube Composite Members
  2. Sequence features of viral and human Internal Ribosome Entry Sites predictive of their activity
  3. IRESeek: structure-informed deep learning method for accurate identification of internal ribosome entry sites in circular RNAs
  4. A versatile toolbox for determining IRES activity in cells and embryonic tissues
  5. Protein translational control and its contribution to oncogenesis revealed by computational methods
  6. Further Characterisation of the Translational Termination-Reinitiation Signal of the Influenza B Virus Segment 7 RNA
  7. A Constitutive Model for Circular and Square Cross-Section Concrete Confined with Aramid FRP Laminates
  8. Estimation of compressive strength of ultra-high performance lightweight concrete (UHPLC) using neural network
  9. Systematic discovery of cap-independent translation sequences in human and viral genomes
  10. ROSE: A Deep Learning Based Framework for Predicting Ribosome Stalling
  11. RPFdb: a database for genome wide information of translated mRNA generated from ribosome profiling
  12. Genome-wide assessment of differential translations with ribosome profiling data
  13. The circular RNA circbabo(5,6,7,8S) regulates lipid metabolism and neuronal integrity via TGF-β/ROS/JNK/SREBP signaling axis in Drosophila
  14. Proteogenomics Reveals Microproteins in Activated T Cells
  15. An expanded reference catalog of translated open reading frames for biomedical research