Modeling In Vivo RNA Structure and Site Accessibility

30 concepts21 papers1 proposal

Traditional computational models of RNA secondary structure operate under a foundational assumption: that an RNA molecule folds in an isolated environment, seeking a state of minimum free energy or following patterns conserved across evolutionary history. However, a living cell is not an isolated test tube. It is a crowded, chaotic environment packed with molecular forces, chemical modifications, and active remodeling machinery like helicases and chaperones.

Because of this complex environment, a profound discrepancy exists between how an RNA molecule is predicted to fold in silico and how it actually behaves in vivo. To bridge this gap, a major research direction has emerged that integrates experimental in vivo structure-probing data directly into computational frameworks. By shifting from purely thermodynamic predictions to experimentally constrained models, researchers can accurately capture real-world RNA architecture and its downstream functional interactions.

The In Vivo Structural Discrepancy

When RNA is mapped inside a living cell, its physical shape often deviates significantly from theoretical models. As noted by Jianyu Zhou et al. and Xingli Guo et al., thermodynamic simulations fail to account for intracellular constraints, leading to inaccurate structural models.

Experimentation frequently reveals that RNAs inside living cells adopt much more open configurations than predicted. For instance, Qing-Jun Luo et al. demonstrated that intracellular RNAs exhibit larger terminal loops and unwound regions compared to theoretical models, pointing to the active role of unknown cellular remodeling factors. Furthermore, studying isolated RNA fragments in vitro misses the broader picture; Pan Li et al. showed that full-length viral genomes rely on long-range, genome-wide structural interactions that only manifest within the native host cell environment.

Compounding this complexity, the RNA landscape is highly dynamic. Chemical modifications can act as structural switches, flipping sequences between entirely different conformations, such as canonical stem-loops and non-canonical G-quadruplexes (Yicen Zhou et al.). Because purely sequence-based or thermodynamic models cannot anticipate these cellular interventions, they generate high rates of low-confidence predictions.

Accessibility as the Gatekeeper of Interaction

Why does this structural discrepancy matter so much? Because an RNA molecule's physical shape acts as the primary gatekeeper for its interactions with regulatory machinery, such as microRNAs and RNA-binding proteins (RBPs).

Historically, algorithms predicted RBP binding sites by scanning transcripts for linear sequence motifs. However, sequence matching alone is a notoriously poor predictor of actual in vivo binding. As Gerald Quon et al. and Junichi Iwakiri et al. explain, a protein cannot bind to a sequence motif if that motif is buried inside a double-stranded RNA helix. True binding depends on site accessibility—the probability that a target region remains single-stranded and physically exposed.

In Silico Prediction (Thermodynamic Minimum)
[Sequence Motif] ---> Locked inside a stable double-stranded stem ---> Inaccessible

In Vivo Reality (Cellular Remodeling)
[Cell Chaperones/Helicases] ---> Unwind the stem ---> [Sequence Motif Exposed] ---> RBP Binds

When models account for this physical shape, their predictive power improves dramatically. For example, Sanghamitra Bandyopadhyay et al. incorporated structural features like minimum free energy and loop configurations to filter out inactive "decoy" sites, drastically reducing false-positive rates in microRNA target prediction. Even when dealing with what was previously dismissed as the "unstructured" RBPome, Yaron Orenstein et al. proved that accounting for local structural preferences provides a statistically significant boost in identifying true protein binding sites.

Integrating Probing Data into Computational Models

To capture true cellular folding, researchers have developed data-driven frameworks that constrain computational models using empirical in vivo structure-probing data (such as SHAPE, icSHAPE-MaP, DMS, or PARIS sequencing). Instead of trying to simulate every complex force inside a cell from scratch, these models ingest experimental data as a proxy for the net effect of the cellular environment.

This integration takes several forms:

  • Bayesian Ensemble Modeling: The IRIS framework, introduced by Jianyu Zhou et al., combines thermodynamic principles with "fuzzy" and often contradictory in vivo PARIS data. By using a Bayesian model, it resolves experimental noise to predict the diverse ensembles of shapes an RNA molecule actually adopts inside a cell, frequently identifying biologically relevant structures that are thermodynamically sub-optimal.
  • Deep Learning and Pattern Recognition: Rather than relying on implicit learning of folding patterns from sequence data alone (Ying Xiong et al.), modern deep learning architectures explicitly ingest numerical profiles of in vivo secondary structures. PrismNet, developed by Kui Xu et al., integrates these experimental structural profiles with deep neural networks to predict how dynamic changes in RNA shape alter protein recruitment across different cellular conditions.

Why Experimentally Constrained Models Work

The reason experimentally constrained models outperform traditional methods comes down to context and geometry.

As reviewed by Junkang Wei et al., in vivo structural data inherently captures the biological state of the molecule. It bypasses the computational bottleneck of simulating molecular crowding or protein competition by using the cell's actual output. This real-world alignment is supported by Xiaomei Li et al., who demonstrated that solvent accessibility models show a significantly stronger correlation with functional regions when validated against in vivo DMS data rather than in vitro measurements.

Furthermore, explicitly providing structural context allows machine learning models to generalize across different cellular environments. Haoran Zhu et al. demonstrated that by utilizing numerical profiles of secondary structures, deep learning models can successfully predict how disease states or environmental shifts alter binding site accessibility.

This structural sensitivity is particularly vital for specific protein families. As Kui Xu et al. observed, the accuracy gains from in vivo data are most pronounced for proteins containing double-stranded RNA binding domains, where the physical geometry of the RNA backbone is the primary driver of interaction. By explicitly distinguishing between structural features like loops and stems, frameworks such as PaRPI (Lin Peng) can generalize to predict interactions for proteins and cell types they have never seen before.

Ultimately, moving from static, isolated folding models to experimentally constrained, context-aware architectures allows researchers to decode how the physical remodeling of RNA drives cellular function and disease pathology. Looking forward, the integration of biochemical data like SHAPE into broader workflows, such as multiple sequence alignments (Haruka Yonemoto et al.), promises to further refine our understanding of the non-coding transcriptome.

Key Foundational Work

Predicting in vivo binding sites of RNA-binding proteins using mRNA secondary structure (2010) — contributed 2 concept elements to this direction.
CentroidAlign-Web: A Fast and Accurate Multiple Aligner for Long Non-Coding RNAs (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_in_vivo_structural_remodeling_4f298177,
  title = {in vivo structural remodeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/in-vivo-structural-remodeling/rdr_4f298177b03843ac84d8efec3522a892},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_in_vivo_structural_remodeling_4f298177_v1,
  title = {in vivo structural remodeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/in-vivo-structural-remodeling/rdr_4f298177b03843ac84d8efec3522a892/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake