The study addresses the fact that viral RNA contains information beyond the order and abundance of its four nucleotide bases. RNA molecules fold into complex three-dimensional structures through interactions between neighboring nucleotides and complementary bases. These secondary structures can include hairpins, stems and paired regions, and they can differ among viruses even when their overall nucleotide composition appears remarkably similar. Raman spectroscopy detects characteristic molecular vibrations when a sample interacts with light. In the experiment, the resulting signals reflected contributions from the RNA backbone, ribose rings, phosphate linkages and nitrogenous bases, allowing the investigators to characterize patterns associated with RNA organization rather than reading the sequence letter by letter.

For the experiment, the team extracted RNA nucleotide strands from 20 SARS-CoV-2 strains. The panel comprised the original Japanese isolate, seven named non-Omicron variants—Alpha, Beta, Gamma, Delta, Lambda, Theta and Mu—and 12 Omicron subvariants. The researchers systematically acquired the Raman spectra of these samples and mathematically deconvoluted them to identify their individual molecular components. All samples shared characteristic signals associated with common RNA chemistry, including the backbone, ribofuranose rings, phosphate groups and nitrogenous bases. At the same time, subtle but reproducible differences emerged among the strains and subvariants.

Those differences were then encoded as Raman barcodes, defined in the study as characteristic patterns of Raman signals that capture distinctive molecular features of viral RNA. The approach was intended to preserve structural information that might not be apparent from a simple comparison of nucleotide composition. In the authors’ conception, a “Raman genome” represents an RNA genome through its molecular architecture and spectral fingerprint rather than through a direct reading of its nucleotide sequence. This is a conceptual extension of genome characterization: the barcode captures molecular features associated with RNA folding and conformation, including the physical contributions of the sugar-phosphate framework, ribose and the bases.

A central finding was that the observed spectral distinctions could not be explained only by variation in the relative abundance of the four RNA bases. When the genome sequences were used to assess those proportions, the investigated strains had highly similar base fractions. Their Raman spectra nevertheless exhibited markedly different molecular signatures. This separation between sequence-based composition and measured spectra indicates that the Raman measurement was sensitive to information beyond base abundance. The authors investigated the molecular origins of the differences by relating particular Raman bands to features of RNA secondary structure, including backbone conformation, nucleotide-chain characteristics, ribofuranose dynamics, hairpin formation, stem length and base-pairing environments.

The resulting barcodes discriminated all 20 strains and subvariants examined, including multiple Omicron subvariants. Conventional principal component analysis, by comparison, provided only limited separation among the same material. The distinction between these two analyses is important to the reported result: the successful classification was based on explicitly encoded Raman spectral features, rather than on the limited discrimination produced by the conventional PCA approach. The finding also shows why the spectral differences matter. They were not merely an incidental effect of broad base composition; instead, selected bands could be associated with interpretable structural features of the RNA strands and used to distinguish the tested panel.

Technical interpretation of the spectra depends partly on changes in Raman signal intensity caused by molecular interactions. Hyperchromic and hypochromic responses involving nitrogenous bases provide clues about pairing environments and RNA secondary structure. Bands associated with ribose vibrations, phosphate linkages and RNA backbone conformation also contribute to the encoded barcode. These physically interpretable features connect the classification result to molecular architecture rather than leaving it as an unexplained statistical separation. The work therefore combines two aspects of the proposed method: a Raman fingerprint capable of differentiating samples and an analysis linking parts of that fingerprint to secondary-structure characteristics such as stem formation, hairpins and the environments in which bases are paired.

The researchers present the result as a first step toward a broader strategy for characterizing viral RNA through spectroscopy. They propose that future development could combine physically interpretable Raman markers with automated spectral analysis, theoretical calculations and artificial intelligence. Such systems could potentially provide a rapid layer of molecular information alongside sequencing, particularly during outbreaks, epidemics and emergency or critical-care situations where timely characterization may be important. The release does not report the implementation of such a clinical system. The reported evidence concerns an experimental comparison of RNA strands from a defined set of 20 strains and subvariants, not prospective surveillance or routine diagnostic validation.

Several limitations follow directly from that experimental scope. The panel was limited to 20 SARS-CoV-2 strains and subvariants, so the result demonstrates discrimination within the tested collection rather than universal performance across the full diversity of SARS-CoV-2 or other viruses. The researchers used RNA nucleotide strands extracted from the strains, meaning the measurements were performed downstream of sample preparation and did not by themselves demonstrate direct identification in unprocessed patient material. The news release also does not report prospective clinical comparisons, blinded diagnostic performance, turnaround time, cost, sensitivity, specificity or reproducibility across laboratories. It likewise does not describe testing under changing sample conditions or a clinical workflow for detecting newly emerging variants.

Even with these qualifications, the experiments offer a technically distinct route for viral science. Conventional sequencing determines the order of nucleotides, whereas the proposed Raman method captures a molecular fingerprint associated with RNA structure and conformation. By successfully discriminating 20 strains and subvariants whose evaluated base fractions were highly similar, the study shows that structural variation supplied information that was not captured by simple composition estimates alone. The authors do not claim that Raman analysis will replace RNA sequencing. Instead, they describe it as a complementary approach that may eventually support faster variant screening and broaden molecular surveillance. The work was financially supported by a Japan Society for the Promotion of Science grant, and the authors declared no conflicts of interest.

Subject of Research: Experimental Raman spectroscopy of SARS-CoV-2 RNA secondary structure for distinguishing variants and subvariants

Article Title: “Raman Genome” for SARS-CoV-2: new Raman technology distinguishes viral variants by their molecular RNA architecture

Article References: “Raman Genome” for SARS-CoV-2: new Raman technology distinguishes viral variants by their molecular RNA architecture. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: SARS-CoV-2, Raman spectroscopy, Raman barcodes, RNA structure, viral variants, Omicron subvariants, molecular diagnostics, viral surveillance, secondary structure, Advanced Science, Raman, Genome

The Editorial, published on 9 July 2026, argues that the most productive open questions in lncRNA and circRNA biology lie not in yet another downstream target study, but in the intrinsic biology of the RNA molecules themselves. What controls their expression, what shape they take, how they are produced, how much of them is actually present in a cell, and which exact isoform is being studied. The authors call this a shift in perspective, and they argue it is not merely of biochemical interest but ultimately determines the clinical relevance of the entire field.

The evidence for this concern comes from a retrospective analysis of 35 lncRNA and circRNA research papers published in Experimental & Molecular Medicine over the past decade, from June 2016 to May 2026. Of these, 15 papers, or 43 percent, proposed competitive endogenous RNA or miRNA sponging mechanisms, while another 16 papers, or 46 percent, described lncRNA-protein binding events characterized primarily through pull-down and luciferase assays. Only one paper addressed any of the five underexplored areas highlighted in the Editorial, and not a single paper made biogenesis, RNA structure, absolute quantification or isoform-specific function a primary mechanistic claim.

The competitive endogenous RNA hypothesis, which proposes that an RNA can act as a molecular sponge for microRNAs, was challenged on quantitative grounds more than 10 years ago. Many lncRNAs are present at copy numbers far too low to titrate cellular miRNA pools meaningfully. Single-molecule fluorescence in situ hybridization and absolute quantification studies have shown that roughly 40 percent of lncRNAs are present at ten or fewer copies per cell. Most canonical sponge or scaffold mechanisms simply cannot operate at those concentrations. Yet the template has remained remarkably common in the literature, with a new RNA reported as differentially expressed in a disease, knocked down to produce a phenotype in one or two cell lines, confirmed in an animal model, shown in a small patient cohort, and then assigned a predicted microRNA target confirmed by luciferase assay.

The authors are careful to say that these mechanisms are not wrong. The principles of ceRNA and lncRNA-protein interactions are real biological phenomena that have advanced the field. The concern is not with the ceRNA concept itself, but with how often it is invoked as a default mechanism without sufficient quantitative or mechanistic support. A study that repeats this familiar flow under a new RNA name, even with the now standard combination of cell line, animal and patient sample data, adds relatively little that is conceptually new. Multi-system validation has become the baseline expectation rather than a mark of distinction, and for research to be truly novel, more is required than ever before.

The first underexplored area is upstream regulation of long non-coding RNAs. There is a clear asymmetry in the literature. Many papers describe what an lncRNA regulates, but very few ask what regulates the lncRNA itself. Genome-wide analyses have shown that lncRNA promoters carry distinct transcription factor binding patterns, which implies that lncRNAs operate under their own regulatory programme. For most disease-associated lncRNAs, however, we still do not know which transcription factors actually drive them, which signals turn them on, or why their expression is restricted to particular tissues or disease states. The post-transcriptional layer is similarly underdeveloped. LncRNAs are typically short-lived, and their turnover by the nuclear and cytoplasmic surveillance machinery is what really sets their steady-state abundance, but this is rarely the focus of mechanistic studies.

The second area is biogenesis and the host gene relationship of circular RNAs. Almost every circRNA paper asks what the circRNA does, but very few ask why the back-splicing event happened in the first place. The answer involves cis elements such as flanking inverted repeats and intron length asymmetry, together with a growing set of trans-acting RNA-binding proteins, but the catalogue is far from complete. Biogenesis is also highly context dependent. Stresses such as hypoxia, DNA damage, and immune activation reshape the back-splicing landscape, often in dramatic ways, and the mechanisms behind those shifts are mostly unknown. The relationship between a circRNA and its host gene is a second open question. We rarely know whether the two are coupled or independent, whether circRNA production comes at the expense of linear mRNA from the same locus, or whether the circRNA feeds back on its own host. The same set of questions applies to divergent, antisense and intronic lncRNAs, where it is often unclear whether the act of transcription, the RNA product, or both, carries the function.

The third area is structure and function. Protein biochemistry accepted long ago that mechanism cannot be claimed without structure, but research on lncRNAs and circRNAs has been slower to adopt this standard. Most papers infer function from sequence and binding partners, with no information about the secondary or tertiary structure that actually supports those interactions. The tools have been available for years. In-cell chemical probing methods now map RNA structure at nucleotide resolution across the transcriptome, cryogenic electron microscopy of ribonucleoprotein particles is now routine for many complexes, and computational structure prediction has improved enough to be a useful starting point. Studies that map the structural modules of an lncRNA or circRNA, identify the elements that drive its protein interactions, and test how RNA modifications change folding will move the field towards a genuine mechanism, and will also establish the foundation that any RNA-targeted therapeutic effort will eventually need.

The fourth area is quantitative biology. Mechanism is a quantitative claim. A proposed sponge or scaffold function requires the regulator to be present at a concentration high enough to engage its targets. The same level of quantitative rigour applied for years in studies of transcription factors and signalling pathways could valuably be extended to lncRNA and circRNA research. The fifth area is isoform identity. A single lncRNA locus often produces several isoforms with different ends, splice patterns and polyadenylation sites, and a single circRNA host gene can give rise to dozens of distinct circular products. The literature still routinely refers to a given lncRNA or circRNA as if it were a single, uniform molecule. Defining which isoform actually does the work, in which cell type and under which conditions, should be treated as a prerequisite for a mechanistic claim, not as an optional refinement. Long-read sequencing, isoform-specific perturbation and isoform-resolved expression analysis have made this practical.

The broader point is that none of this implies classical ceRNA and descriptive lncRNA-protein interaction studies no longer have a place. Rather, such studies are most valuable when they meet a higher bar than has often been applied. A proposed mechanism is more convincing when it is quantitatively plausible, not merely consistent with what is already known. Binding interactions are more compelling when shown to be direct rather than inferred from prediction tools alone. The authors hope that work taking up any of the directions sketched above, or other emerging directions aligned with this shift towards the lncRNA or circRNA itself, will help to shape where the field goes next. The essential question for authors is what makes their work a substantive advance over the existing literature, rather than a familiar story with a new RNA name attached.

Subject of Research: Rebalancing the non-coding RNA research agenda towards upstream regulation, biogenesis, structure, quantitative biology and isoform identity

Article Title: Rebalancing the non-coding RNA research agenda: future directions for long non-coding RNA and circular RNA research

Article References: Kim, Y.-K., & Kook, H. (2026). Rebalancing the non-coding RNA research agenda: future directions for long non-coding RNA and circular RNA research. Experimental & Molecular Medicine, 58(9), 2809-2811. https://doi.org/10.1038/s12276-026-01832-x

Image Credits: AI Generated

DOI: 10.1038/s12276-026-01832-x

Keywords: lncRNA, circRNA, non-coding RNA, ceRNA hypothesis, RNA structure, RNA biogenesis, quantitative biology, isoform identity, Experimental & Molecular Medicine, RNA therapeutics, Rebalancing, non-coding

The core challenge that PRADA addresses is a long-standing bottleneck in proximity labeling. The most widely used tools, such as APEX2 and TurboID, genetically fuse an engineered enzyme to a protein of interest. The enzyme then converts a supplied small molecule into a reactive species that tags nearby biomolecules, which can later be fished out and identified by mass spectrometry or sequencing. The catch is the trigger. APEX2 requires a bolus of exogenous hydrogen peroxide, a reactive oxidant that damages cells and diffuses poorly through intact tissues. TurboID requires biotin, which works in vivo but generates its own background labeling. Neither approach has proven ideal for the delicate, three-dimensional environment of a living organism.

The Shanghai- and Singapore-based team, led by Shuo Han of the Chinese Academy of Sciences together with Yue Wan of the Genome Institute of Singapore, took a different route: instead of supplying hydrogen peroxide directly, they engineered the system to make it on the spot. PRADA pairs a genetically fused peroxidase with an engineered D-amino acid oxidase, or DAAO, derived from the yeast Rhodotorula gracilis. When researchers deliver a nonproteinogenic D-amino acid such as D-phenylalanine or D-alanine — molecules that animal cells largely ignore — the oxidase converts them into hydrogen peroxide right next to the peroxidase. That locally generated oxidant then activates the peroxidase, which converts phenol probes such as biotin-phenol into short-lived radicals that covalently tag whatever proteins or nucleic acids sit within roughly a few tens of nanometers.

The engineering itself was nontrivial. The team used structural modeling and rational mutagenesis to delete a flexible C-terminal loop of the R. gracilis oxidase, producing a monomeric variant that fuses cleanly with APEX2 without disrupting either enzyme’s activity. Extended data show the researchers systematically tested fusion orientations, split-enzyme designs and D-amino acid concentrations, using fluorescence assays to confirm that the cascade produces hydrogen peroxide efficiently and that the labeling signal colocalizes precisely with the tagged protein’s known subcellular address. When the two enzymes were expressed separately rather than fused, labeling dropped sharply, confirming that the reaction is genuinely confined to the immediate vicinity of the fusion protein.

That spatial confinement is what makes the method safe enough for living animals. Because the hydrogen peroxide is generated in situ at nanometer scales and consumed almost immediately by the peroxidase, it never accumulates to toxic levels. The researchers measured malondialdehyde levels, a marker of lipid peroxidation, and protein carbonylation after PRADA labeling and found both below detection limits or unchanged compared with controls. Mitochondrial superoxide, a sensitive indicator of oxidative stress, was actually lower in PRADA-labeled cells than in cells labeled conventionally with APEX2 and exogenous peroxide. Cell viability assays showed no measurable harm from the D-amino acid treatment itself. In effect, PRADA converts the most dangerous step of proximity labeling into a self-limiting, locally contained reaction.

The versatility of the platform is striking. In cultured cells, the team targeted PRADA to the nucleus, nucleolus, endoplasmic reticulum, plasma membrane and mitochondria, and in each case recovered proteomic profiles consistent with the known composition of those compartments. They also showed that PRADA can drive functional polymer assembly — the genetically targeted chemical assembly of conducting or insulating materials inside living cells — extending the technique beyond mapping into materials synthesis. And, crucially, they demonstrated that the peroxidase step can label RNA as well as protein, opening a door that most proximity-labeling systems leave closed.

That RNA reactivity became the foundation of the study’s most inventive application. When peroxidase radicals oxidize RNA, they leave behind chemical lesions that cause reverse transcriptase to misread bases during cDNA synthesis. The team realized these misincorporations could be read as a molecular fingerprint of RNA structure: nucleotides that are chemically accessible — that is, single-stranded and exposed — accumulate more mutations than nucleotides buried inside base-paired helices. Building on the logic of mutational profiling sequencing, they developed PRADA-MaPseq, a strategy that converts peroxidase labeling directly into a spatiotemporally resolved map of RNA secondary structure inside living cells.

They benchmarked PRADA-MaPseq against the well-established mitochondrial DMS-MaPseq datasets, showing that uracil and guanine reactivity scores accurately reproduce known structures of mitochondrial messenger RNAs, with area-under-the-curve values reaching 0.60 to 0.82 across transcripts and inter-experimental correlation coefficients as high as 0.96 for individual transcripts. The approach worked with both 4-thiouridine and 6-selenoguanosine metabolic labeling, and the team optimized reverse transcriptase choice to maximize mutational signal. In short, they turned a proximity-labeling enzyme into a structure-probing reagent that reports on RNA folding in a defined cellular compartment at a defined moment in time.

The payoff came in a mouse xenograft model. The researchers implanted HEK293T cells expressing mitochondria-targeted PRADA into immunodeficient mice, then administered D-amino acids and phenol probes systemically. Labeling proceeded efficiently inside the tumors in living animals. Proteomic analysis recovered the expected mitochondrial proteome, transcriptomic enrichment captured mitochondrial RNAs with high specificity, and PRADA-MaPseq yielded the first in vivo structurome of mitochondrial RNA within a living mammalian tumor. Comparing these in vivo structure maps with in vitro measurements of the same transcripts revealed systematic differences — evidence that RNA folding inside a living organism is actively shaped by its environment rather than being a fixed property of the sequence.

Those differences carried biological meaning. The team focused on LRPPRC, a mitochondrial RNA-binding protein known to organize the folding of the mitochondrial transcriptome. When they knocked down LRPPRC in the xenografts, changes in RNA structure propagated to changes in mitochondrial gene expression, demonstrating that RNA architecture functions as a regulatory layer controlling how mitochondrial genes are read. This finding elevates the structurome from a descriptive atlas to a mechanistic insight: the three-dimensional shapes of RNAs inside mitochondria help determine how much protein the organelle makes, and those shapes are tunable in vivo.

Beyond mitochondria, the breadth of validated applications suggests PRADA could become a workhorse across biology. The team demonstrated labeling in the cytoplasm and nuclei of body wall muscle cells in Caenorhabditis elegans, recovering nuclear-enriched transcripts and detecting retained introns as expected; in rat neurons and Drosophila; and in zebrafish. Because the trigger molecule — a simple D-amino acid — is cell-permeable, inexpensive and largely inert, the method should scale to tissues and organisms where peroxide bolus delivery is impractical, such as dense neural tissue, developing embryos or intact tumors. The authors have filed patent applications, and the analysis code for RNA structuromics is freely available on GitHub under an MIT license, with sequencing and proteomic data deposited in public repositories.

Limitations remain, as with any new technology. The labeling window is measured in tens of minutes rather than seconds, so fast molecular events may still escape capture. The peroxidase chemistry, while gentler than exogenous peroxide, still relies on radical intermediates whose diffusion radius sets the spatial resolution. And the D-amino acid oxidase must be carefully engineered for each fusion context to avoid perturbing the protein under study. Nevertheless, PRADA represents a genuine conceptual advance: it decouples proximity labeling from externally supplied oxidants, unifies protein, RNA and polymer chemistries under one enzymatic umbrella, and — for the first time — brings RNA structure mapping into living animals. If the platform generalizes as its developers hope, the spatial organization of biomolecules in health and disease may soon be readable not just in culture dishes, but in the bodies of living organisms, one D-amino acid at a time.

Subject of Research: An engineered D-amino-acid-activated peroxidase cascade for in vivo multiomic proximity labeling of proteins and RNA structures

Article Title: Multiomic proximity labeling in vivo by D-amino-acid-activated peroxidase reaction

Article References: Liu, J., Han, J., Wang, Y., Song, M., Zhu, J., Liang, Y., Chen, F., Liu, Z., Yan, X., Wang, Z., Peng, W., Tu, R., Zhang, Z., Zhong, B., Sun, H., Yin, J., Chang, J., Long, Y., Men, Y., … Han, S. (2026). Multiomic proximity labeling in vivo by D-amino-acid-activated peroxidase reaction. Nature Chemical Biology. https://doi.org/10.1038/s41589-026-02336-5

Image Credits: AI Generated

DOI: 10.1038/s41589-026-02336-5

Keywords: proximity labeling, PRADA, D-amino acid oxidase, APEX2, RNA structure, structurome, mitochondria, spatial proteomics, chemical biology, in vivo imaging, xenograft, LRPPRC

The system, called NucleicBERT, applies a transformer-based language model to ribonucleic acid sequences, training it to predict masked positions in nucleotide strings drawn from large public databases. The approach deliberately avoids labels. Instead of being told which sequences are ribozymes, which are microRNAs, or which bind particular proteins, the model is simply asked to fill in the blanks across millions of natural sequences. In doing so, it is forced to internalise the statistical regularities of real RNA: which nucleotides tend to co-occur, which motifs recur across distant branches of life, and which combinations essentially never appear. Those patterns, the authors contend, encode a compressed representation of the physical and evolutionary constraints that shape functional RNA.

The technical foundation is the bidirectional encoder architecture popularised by models such as BERT. In natural language, such models read text in both directions and learn contextual embeddings, so that the meaning of a word depends on its neighbours. NucleicBERT imports that idea wholesale into molecular biology. Each nucleotide in an RNA sequence is treated as a token, and the encoder produces a vector for every position that reflects its biological context. A cytosine embedded in a stem-loop of a transfer RNA acquires a different representation from the same cytosine sitting in the loop of a riboswitch, even though the raw letter is identical. This context sensitivity is precisely what hand-crafted features, position-weight matrices and simple motif searches have historically lacked.

Pretraining proceeds with a masked-language objective. Random positions in each training sequence are hidden, and the model must reconstruct them from surrounding context. Because the training corpus spans diverse RNA families and organisms, the network cannot succeed by memorising shallow patterns; it must capture deeper regularities such as compensatory mutations in paired regions, conserved loops, and the compositional biases of different RNA classes. The resulting embeddings can then be transferred downstream: a relatively small amount of labelled data is sufficient to fine-tune the pretrained network for specific prediction tasks, a strategy that has transformed fields from computer vision to protein biochemistry.

The practical payoff comes in the form of benchmark performance on tasks that matter to RNA biologists. According to the paper, NucleicBERT embeddings improve predictive accuracy on problems including the classification of non-coding RNA families, the identification of RNA-binding protein sites, and the assessment of sequence variants that disrupt splicing or translation. In each case the pretrained model outperforms baselines trained from scratch on the same labelled data, and the advantage is largest precisely where labelled examples are scarcest. That pattern is the classic signature of useful pretraining: the model arrives at a task already fluent in the underlying vocabulary, so supervision only needs to teach the final grammar.

What makes the work conceptually significant is not merely the benchmark numbers but the interpretability experiments layered on top of them. The authors probe what the model has learned by examining attention patterns and embedding geometry. Sequences with related structures and functions cluster together in the embedding space even when their nucleotide identities differ substantially, suggesting the model has discovered homology that raw sequence comparison misses. Attention heads, the internal components that let a transformer weigh relationships between positions, turn out to concentrate on regions that biologists recognise as structurally or functionally meaningful, such as paired stems and conserved catalytic motifs. In effect, the network rediscovers, from raw data alone, some of the hard-won knowledge that RNA biochemists assembled over half a century.

The study also confronts one of the central puzzles of RNA biology: the sheer size of sequence space. An RNA molecule of only 100 nucleotides has 4 to the power of 100 possible sequences, a number that dwarfs the number of atoms in the observable universe. Natural RNA occupies a vanishingly sparse subset of that space, organised into families shaped by common ancestry and common physics. Language models are, in a formal sense, tools for modelling the distribution of data, and NucleicBERT can therefore be read as a statistical map of where functional RNA lives within the vast combinatorial wilderness. Sequences the model assigns high likelihood are, heuristically, sequences that look like biology; sequences it assigns low likelihood are candidates for exotic synthetic designs, or for failure.

That map has immediate applications in engineering. RNA therapeutics, from messenger RNA vaccines to small interfering RNAs and antisense oligonucleotides, all depend on the properties of sequence: how stably a molecule folds, how efficiently it is translated, how recognisable it is to the innate immune system, and how long it survives in the cell. The authors report that NucleicBERT representations correlate with measurable properties such as secondary-structure stability and expression level, offering drug developers a way to screen and optimise candidate sequences in silico before expensive synthesis and testing begin. The same representations can guide the design of synthetic riboswitches and regulatory elements for synthetic biology, where designers currently iterate through costly build-and-test cycles.

The researchers are candid about limitations. RNA databases are biased towards well-studied model organisms and abundant RNA classes, so the model’s fluency is strongest where data are richest and weakest for rare transcripts and poorly characterised clades. The masked-language objective captures linear sequence context directly and higher-order structure only indirectly, so tasks that hinge on detailed three-dimensional folding may still require complementary physics-based or structure-specific models. And like all deep networks, NucleicBERT offers correlations rather than mechanisms: its embeddings are a powerful substrate for prediction, but turning them into causal explanations of why a particular fold catalyses a particular reaction remains future work. The authors frame the model not as a replacement for biochemical experiment but as a hypothesis engine that tells experimentalists where to look.

Even with those caveats, the arrival of a mature nucleic-acid language model marks a turning point in how the life sciences approach sequence data. For twenty years, genome annotation has leaned on alignment-based tools that compare new sequences against known ones, a strategy that fails for molecules with no recognisable relatives. Self-supervised models offer a different epistemology: knowledge distilled from the totality of observed sequences, applicable even to orphans with no evolutionary cousins. As sequencing technologies continue to generate data far faster than any human can annotate them, systems like NucleicBERT are likely to become standard equipment in the computational biology toolkit, reading the genome’s least understood language at a pace no human reader could match and pointing the way to RNA molecules that biology has not yet invented.

Subject of Research: Self-supervised language modelling of RNA sequence space with the NucleicBERT neural network

Article Title: NucleicBERT interprets RNA sequence space through self-supervised language modelling

Article References: Upadhyay, U., Herold, J., Götz, M., & Schug, A. (2026). NucleicBERT interprets RNA sequence space through self-supervised language modelling. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01295-9

Image Credits: AI Generated

DOI: 10.1038/s42256-026-01295-9

Keywords: NucleicBERT, RNA, self-supervised learning, language models, machine learning, transformers, non-coding RNA, RNA therapeutics, sequence biology, computational biology, embeddings, RNA structure