System-level measurement, modeling, and manipulation of RNA

In the two and a half decades since the sequencing of the human genome, we have witnessed an explosion in technologies designed to interrogate how genetic information unfolds within a human body in space and time, during development and aging, across healthy and diseased tissues. We can now simultaneously measure the abundance, localization, modification status, and protein output of all RNAs expressed in a single cell for millions of cells, as well as carry out functional screens of hundreds of thousands of sequence variants. These measurements almost always rely on identifying nucleic acid sequences and their variants, usually by sequencing. Naturally, the large volumes of data that are generated within individual experiments can only be interpreted with appropriate computational analysis methods. In this area too, there has been tremendous development, especially since the broad adoption of machine learning and AI approaches specifically designed to take advantage of large, comprehensive data sets.

Our aim for this Special Issue of the RNA Journal was to highlight perspectives from leaders in the field on the most recent efforts to obtain quantitative data pertaining to processes that involve RNA, on application of novel technologies and on exciting future directions of this research. Important questions that can now be addressed are whether data sets obtained with different technologies and at different levels of gene expression are consistent with each other and whether they provide sufficiently accurate representations of the biological systems that we study, representations that enable the discovery of complex “grammar” rules of biological mechanisms and the prediction of perturbation effects.

Initially developed to sequence full length nucleic acid molecules, the nanopore technology has been taken in a variety of new directions. One of these is the direct sequencing of RNA to enable the detection of RNA modifications. Fleming and Burrows highlight this application, in particular the analysis of dwell times, which has proven informative for calling RNA modifications. While the sequence context-dependency of the signal makes it still difficult to estimate the level of a modification in a given sample, changes between conditions can be reliably detected, as the authors have demonstrated for Escherichia coli rRNA upon heat shock, cold shock, and nutrient stress, and human rRNAs upon inflammatory stress. Cross talks between modifications are also currently explored.

One of the big unknowns in predicting the fate of RNAs within cells is their structure, which impacts and is also modulated by their interactions with proteins, with all the downstream consequences. Selective 2′-hydroxyl acylation analyzed by primer extension (SHAPE) and dimethyl sulfate (DMS) have been used for large-scale structure probing, but these experiments are technically demanding, and therefore, accurate RNA structure prediction could be an important alternative. The piece by Szikszai, Aviran and colleagues contrasts thermodynamics-based with deep learning approaches for RNA structure prediction, emphasizing that the thermodynamic folding is as relevant as ever, while deep learning approaches struggle to generalize. One of the contributing factors, initially overlooked but now addressed more judiciously when training deep learning (DL) models, is the clear, homology-based separation between training and test sets. Furthermore, very promising results have been obtained by bridging physics-based models and deep learning with differentiable folding. Continuous representation of sequences (a la weight matrices) enables the specification of a probability distribution over sequences and subsequently of a differentiable form for the ensemble free energy (EFE) function. The latter allows gradients to be computed with respect to the input probability distribution over sequences, making the optimization problem amenable to deep learning techniques. Interestingly, by maximizing the probability of observed sequence-structure pairs, more accurate parameters of thermodynamic folding models can be obtained without the need to carry out melting experiments.

Reflecting the great interest in the DL methods for predicting RNA properties and RNA design, the piece by Sanguinetti and colleagues provides a very concise summary of this very rapidly expanding field, going over the history of RNA structure prediction, and recent efforts in both data generation and model building. They emphasize that further areas of exploration include the importance of RNA modifications, which drastically expand the “RNA alphabet,” the prediction of pseudo-knots, noncanonical base pairs and structures of large RNAs, the context in which RNAs fold and act, whether in terms of interactions with ions, small molecules and proteins, or of the broader environment of individual cell types.

Continuing on the theme of RNA–RNA interactions, Song and Van Nostrand address the recent progress in crosslinking and immunoprecipitation-based approaches to map interaction sites of RNAs with proteins, focusing specifically on protein-associated RNA–RNA interactions. These have been used to characterize the interactomes of small regulatory RNAs such as miRNAs, snoRNAs, and piRNAs, revealing much interesting biology. Remaining challenges for the future are to improve the efficiency of individual experimental steps (crosslinking, chimera formation), filter false positives resulting from postlysis artifacts, and improve the accuracy of computational analyses. The lack of ground truth data sets has been a persistent, difficult problem in this field, potentially addressable by coupling biophysical models built on in vitro data with measurements of interaction site occupancy obtained in vivo.

In a similar vein, the piece by Aigner et al. addresses recent progress in the quantification of miRNAs and target 3′UTRs, interaction partners with particularly important gene expression regulatory roles in the nervous system. It is expected that spatial transcriptomics techniques can be adapted for the simultaneous quantification mRNAs and target 3′UTRs within the same tissue.

The review by Ma and Tian discusses the diverse molecular strategies available for modulating a particular step of RNA processing, namely the cleavage and polyadenylation at specific poly(A) sites. They include antisense oligonucleotides, CRISPR-based systems, and transcriptional elongation control. Alteration of gene expression via alternative polyadenylation isoform expression opens the door for therapeutic applications, for example, in the context of genetic diseases due to mutations in polyadenylation signals.

An issue that all high-throughput approaches have to contend with is that of noisy or missing data in very high-dimensional spaces. Often, denoised, lower-dimensional manifolds are used for downstream tasks such as data exploration and dynamic inferences. Kobak and colleagues discuss representation learning approaches that have been proposed for single cell RNA-seq data (scRNA-seq). They provide a taxonomy of the methods using five axes: the training objective, model architecture, batch handling, pretraining and fine-tuning, and count transformations. Ground truth data sets and appropriately-designed, community-driven benchmarks are identified as crucial for future progress in the representation learning of scRNA-seq data, as indeed, for many other tasks. The authors also point out that one type of analysis does not necessarily suit all objectives, for example, cell typing and differential gene expression analysis are more naturally formulated in the high-dimensional space of genes rather than latent variables, without a meaningful physical interpretation.

Finally, the piece by Iwasaki, Stasevich and colleagues takes a first step in the direction of bridging across scales, from the single-molecule to the genome-wide investigation of mRNA translation. Nascent chain tracking (NCT) is a live-cell approach to visualize translation of individual mRNA reporters. The technique has revealed the bursty nature of translation, and allows the measurement of initiation and elongation rates for individual mRNAs. Ribosome stalling and subcellular localization-dependent processes can also be studied. A drawback of the technique is its reliance on various constructs, making it relatively low-throughput and susceptible to biases. On the other hand, ribosome profiling (Ribo-seq) enables the determination or ribosome occupancy of all mRNAs in parallel, to reveal global trends in the translation dynamics as well as novel translated regions (translons). Coupled with specific perturbations such as the inhibition of translation initiation, Ribo-seq allows the quantification of average elongation rates. It is important that developers of complementary approaches are actively exploring ways to bridge different types of measurements to ultimately converge on a quantitative description of translation of individual mRNAs. This will ultimately enable accurate prediction of perturbation outcomes.

We hope that this Special Issue of the RNA Journal conveys the intense efforts of the community to combine large-scale data generation with both simple and sophisticated computational modeling approaches to accelerate progress in understanding the dynamics of processes involving RNAs. Ultimately, this will allow a more reliable design of approaches to target or generate RNAs with specific properties, supporting the efforts to develop RNA-based therapies.

We thank all the contributors to this Special Issue for sharing their insights and hope that these will catalyze further discussions in the field. In particular, it seems clear that ground truth, large-scale data sets and community-organized benchmarking challenges are needed for a variety of tasks, and we hope that more scientists will become involved in these efforts.

The unconditional support of the editorial team of RNA has made putting together this issue extremely rewarding. Special thanks go to Ann Marie Micenmacher for the always informative and on-time responses to our questions.

• © 2026 Zavolan and Yeo; Published by Cold Spring Harbor Laboratory Press for the RNA Society