The team began where microbiology has always begun: with morphology and biochemistry. All three representative isolates proved to be Gram-positive, catalase-positive and Voges-Proskauer-positive, a classic Bacillus signature that reflects a thick peptidoglycan cell wall, the ability to decompose hydrogen peroxide, and the capacity to ferment sugars into acetoin and related neutral products. Yet the tests were not merely confirmatory. The Methyl Red test, which detects stable mixed-acid fermentation, separated AMU-K-1 from AMU-UP-1 and AMU-MY-1, revealing genuine metabolic heterogeneity among bacteria that looked superficially alike. This kind of phenotypic split matters for anyone hoping to deploy soil bacilli in agriculture, because fermentation end products such as acetoin and 2,3-butanediol are directly relevant to industrial bioprocesses and to plant-microbe signaling in the rhizosphere.

Antimicrobial susceptibility testing added a second layer of differentiation. In agar well diffusion assays, the isolates produced clear inhibition zones against tetracycline, amoxicillin and azithromycin, indicating susceptibility to three clinically important antibiotic classes. But when the researchers moved to broth microdilution, a quantitative method that determines minimum inhibitory concentrations rather than relying on visible zones, they found substantial variability in resistance profiles between isolates. The discrepancy is a useful caution for the field: diffusion-based assays can be influenced by diffusion rates, growth conditions and inoculum density, and the authors’ decision to pair them with microdilution follows current best practice in antimicrobial susceptibility testing. For soil-derived bacteria intended as plant growth promoters or biocontrol agents, antibiotic resistance profiles are not an academic detail, since resistance genes carried by beneficial strains can potentially spread to pathogens in the environment.

The molecular core of the study rests on the 16S rRNA gene, the workhorse of bacterial taxonomy. The researchers amplified roughly 1,500-base-pair fragments of the gene and obtained high-quality Sanger sequencing reads of 1,048 bases for AMU-K-1 and 1,102 bases for AMU-UP-1. BLAST searches against public databases identified AMU-K-1 as closely related to Bacillus safensis, a species originally isolated from spacecraft cleanrooms but now recognized in soils worldwide. AMU-UP-1, however, presented a familiar taxonomic headache: its sequence showed 100 percent identity to multiple closely related taxa within the Bacillus subtilis and Bacillus velezensis species complex. The 16S rRNA gene, for all its utility, simply evolves too slowly in this clade to separate species that diverged recently, and the authors were candid that unambiguous species-level resolution was not possible from this gene alone.

To place the isolates in an evolutionary framework, the team reconstructed phylogenetic trees using the Tamura 3-parameter model with invariant sites, a substitution model designed for sequences with strong transition-transversion bias and skewed guanine-cytosine content. The resulting trees confirmed that both isolates fall within well-supported Bacillus lineages. Nucleotide substitution analysis across more than 1,300 aligned positions revealed a pronounced excess of transitions over transversions, meaning that purine-to-purine and pyrimidine-to-pyrimidine changes dominate the mutational landscape of these ribosomal sequences. That pattern is expected in conserved RNA-coding genes, where structural constraints penalize the more disruptive transversions, but quantifying it provides a baseline for future comparative work on Indian soil bacilli and reinforces the choice of the T92+I model for this dataset.

The most distinctive element of the study is its thermodynamic treatment of the 16S rRNA molecule itself. Using RNA secondary structure prediction tools, the researchers computed minimum free energy structures for the sequenced regions. AMU-K-1 yielded a minimum free energy of minus 530.80 kilocalories per mole, while AMU-UP-1 yielded minus 523.80 kilocalories per mole, both indicating highly stable ribosomal conformations. More intriguing was the ensemble diversity, a measure of how many alternative folding states a sequence can adopt at equilibrium: AMU-K-1 scored 357.82 against 151.53 for AMU-UP-1, suggesting that the ribosomal RNA of the B. safensis-like isolate explores a much broader conformational space in silico. The authors are appropriately careful here, noting that whether such computational differences translate into physiological adaptability requires direct experimental verification. Still, the approach hints at a future in which the folding energetics of housekeeping genes become part of the standard characterization toolkit.

Why does this matter beyond taxonomy? Bacillus species are among the most commercially important bacteria in agriculture. Bacillus velezensis in particular has emerged as a model plant-associated beneficial bacterium, producing lipopeptides that suppress fungal pathogens and colonizing roots in ways that enhance nutrient uptake. Bacillus safensis and related strains have been associated with drought tolerance in crops such as sugarcane, and Bacillus-based inoculants are central to the push for sustainable crop protection. Accurate identification is the precondition for all of these applications, because a mislabeled strain can carry unwanted traits, including antibiotic resistance determinants or pathogenicity factors, into field deployments. The Indian isolates, deposited in NCBI GenBank under accession numbers PZ619164 and PZ619172, now form a verified reference resource for such efforts.

The study also sits within a broader methodological debate. Full-length 16S rRNA sequencing, including nanopore-based approaches, has been shown to achieve species-level resolution in some bacterial groups, and whole-genome comparisons are increasingly regarded as the gold standard for distinguishing members of the B. subtilis complex. By combining phenotypic, biochemical, phylogenetic and thermodynamic evidence, the authors demonstrate a pragmatic middle path for laboratories without routine access to genome sequencing: extract the maximum information from a single locus, quantify its evolutionary dynamics, and treat computational predictions as hypotheses to be tested rather than conclusions. The explicit caveat that in silico ensemble diversity does not equal physiological flexibility is a model of how such integrative papers should be written.

For India specifically, the work contributes to a growing catalog of regionally characterized soil bacilli, joining earlier studies of Bacillus subtilis and Priestia megaterium from eastern Indian farmlands and Bacillus licheniformis and Bacillus cereus from sorghum rhizospheres. Soil biodiversity is under pressure from land-use change and a warming climate, and documenting native microbial diversity before it erodes has become a quiet urgency in environmental microbiology. Each verified isolate with published sequences, biochemical profiles and susceptibility data is a permanent, reusable asset for bioprospecting, whether the target is a biocontrol agent, a bioremediation workhorse or an industrial enzyme producer.

The eight isolates characterized here are unlikely to be the last word on Indian soil Bacillus, and the authors themselves frame the study as a foundation for future exploitation in plant growth promotion, biological control, environmental bioremediation and microbial biotechnology. What the paper ultimately delivers is a template: a demonstration that old-fashioned petri dish biochemistry, when fused with careful sequencing, model-based phylogenetics and even RNA folding thermodynamics, can still generate new knowledge about the organisms beneath our feet. As the search for sustainable alternatives to chemical fertilizers and pesticides accelerates, the humble soil isolate, rigorously identified and honestly annotated, remains one of biotechnology’s most valuable raw materials.

Subject of Research: Integrated phenotypic, molecular, phylogenetic and thermodynamic characterization of soil Bacillus isolates from India

Article Title: Integrated morpho-biochemical, molecular, phylogenetic, and thermodynamic characterization of Bacillus isolates from ecologically distinct Indian soils

Article References: Mohanty, A., Sahoo, J. P., Mohapatra, A., Samal, P. P., Sahoo, L., Mahapatra, S. S., & Hossain, S. N. (2026). Integrated morpho-biochemical, molecular, phylogenetic, and thermodynamic characterization of Bacillus isolates from ecologically distinct Indian soils. Molecular Biology Reports, 53(1), Article 1692. https://doi.org/10.1007/s11033-026-12911-7

Image Credits: AI Generated

DOI: 10.1007/s11033-026-12911-7

Keywords: Bacillus, soil microbiology, 16S rRNA, phylogenetics, Bacillus subtilis, Bacillus velezensis, Bacillus safensis, antimicrobial susceptibility, RNA secondary structure, plant growth promotion, India, Sanger sequencing

What makes ASOs so attractive compared with conventional small-molecule drugs and monoclonal antibodies is their programmability. Because they bind their targets through straightforward Watson–Crick base pairing, researchers can in principle design a molecule against almost any RNA sequence with high specificity, relatively simple synthesis, and reproducible manufacturing. Unlike small molecules, which must find a pocket on a protein surface, ASOs target RNA directly, offering a route to diseases long considered undruggable. The review emphasizes that this specificity, combined with chemical modifications that improve stability and cellular uptake, positions ASOs as a leading platform for personalized therapies, particularly for rare genetic disorders where no other treatment options exist.

At their core, ASOs are chains of nucleotides joined by phosphodiester bonds, but unmodified DNA degrades rapidly in the body. The field has therefore developed successive generations of chemical modifications. First-generation phosphorothioate backbones replace a non-bridging oxygen atom with sulfur, conferring nuclease resistance and a longer plasma half-life, though sometimes at the cost of pro-inflammatory side effects. Second-generation chemistry adds 2′-O-methyl or 2′-O-methoxyethyl sugar modifications, and chimeric gapmer designs combine a central DNA region capable of recruiting RNase H with modified flanks that resist degradation. Third-generation chemistries, including locked nucleic acids and phosphorodiamidate morpholino oligomers, push binding affinity and stability even further, though each comes with its own trade-offs in uptake and mechanism.

The review details three principal modes of action. In RNase H1-mediated silencing, the ASO forms a DNA–RNA hybrid with its target messenger RNA, and the enzyme RNase H1 recognizes that hybrid and cleaves the RNA strand, destroying the transcript and halting protein synthesis. In splice modulation, ASOs bind pre-mRNA regulatory sequences to correct abnormal exon inclusion or exclusion; the approved drug nusinersen uses this strategy in spinal muscular atrophy by promoting inclusion of exon 7 in the SMN2 transcript, while eteplirsen induces exon skipping in Duchenne muscular dystrophy to restore the dystrophin reading frame. Finally, steric block ASOs physically obstruct ribosomes or RNA-binding proteins without triggering degradation, offering a gentler way to tune gene expression. The authors contrast ASOs with small interfering RNAs, noting that siRNAs are double-stranded molecules that operate through the RNA-induced silencing complex and Argonaute cleavage, and that studies such as Bilanges and Stokoe’s comparison of PDK1 knockdown show the two platforms produce distinct transcriptional consequences.

Designing an effective ASO is a multi-parameter optimization problem, and this is where computational tools earn their keep. The review lays out the critical design variables: oligonucleotide length, typically 15 to 25 nucleotides; GC content between 40 and 60 percent to balance hybrid stability against overly tight, nonspecific binding; binding energy of roughly −8 kcal/mol or better; minimal cross-reactivity with other transcripts; and careful target-site selection that accounts for mRNA accessibility. Regions of RNA buried inside stable stems or hairpins are poor targets, whereas loops and bulges expose nucleotides for hybridization. Thermodynamic parameters such as melting temperature determine whether the ASO–mRNA duplex survives physiological conditions, and secondary structure prediction helps designers avoid oligos that fold onto themselves and lose their intended activity.

Several dedicated software platforms now support this workflow. Sfold uses statistical RNA folding to evaluate target-site accessibility, making it a workhorse for general ASO design. PFRED offers a user-friendly computational platform for both siRNA and antisense design, integrating stability and potency parameters. ASOptimizer applies deep learning to minimize off-target effects and cytotoxicity, as demonstrated in work on IDO1 gene regulation. lncASO brings machine learning to the special case of long non-coding RNA targets, while OptiRNAi and siExplorer serve the RNA interference community. High-throughput bioinformatics pipelines built around these tools can screen hundreds of candidate sequences in silico, ranking them by melting temperature, hybridization efficiency, and predicted off-target interactions, dramatically reducing the cost and labor of experimental screening.

Structural prediction remains the hardest computational problem. Unlike proteins, where decades of experimental data have enabled powerful folding models, single-stranded DNA adopts a bewildering variety of conformations, including hairpins, bulges, and G-quadruplexes, and its polyanionic character is strongly influenced by metal ions. The review highlights a workflow proposed by Jeddi and Saiz that repurposes RNA 3D prediction tools for ssDNA: predict the DNA secondary structure, model the corresponding ssRNA tertiary structure, then convert ribose to deoxyribose and uracil to thymine. Tools such as Mfold, RNAfold from the Vienna RNA WebSuite, CentroidFold, RNAstructure, and KineFold handle secondary structure prediction with complementary thermodynamic and centroid-based approaches, while RNA Composer builds large RNA 3D models from secondary structure information. Conversion utilities like PyMOL, ChimeraX, 3DNA, and the authors’ own RNA2DNA script then translate RNA structures into their DNA counterparts.

Downstream, molecular docking platforms such as HADDOCK, HDOCK, and HNADOCK can generate structural hypotheses for ASO–protein interactions, and molecular dynamics engines like Schrödinger’s DESMOND and GROMACS simulate how ASO–target complexes behave over time under biological conditions. The review is careful to note the caveats: docking scores for flexible, chemically modified single-stranded oligonucleotides should be treated as hypothesis-generating rather than definitive, and experimental validation through biochemical binding assays, crosslinking–mass spectrometry, or high-resolution structural methods remains essential. Current in silico models also struggle with non-canonical base pairing such as pseudoknots and G-quadruplexes, with the dynamic role of RNA-binding proteins, and with cellular environmental factors like ionic conditions and competing endogenous RNAs, all of which can shift ASO behavior away from computational predictions.

The clinical track record justifies the investment. Fomivirsen (Vitravene) became the first FDA-approved ASO in 1998 for cytomegalovirus retinitis, followed by mipomersen (Kynamro) for homozygous familial hypercholesterolemia in 2013. Nusinersen (Spinraza) transformed the outlook for spinal muscular atrophy, inotersen (Tegsedi) treats hereditary transthyretin-mediated amyloidosis, and the exon-skipping drugs eteplirsen and golodirsen address Duchenne muscular dystrophy. Each success rests on the same triad of rational sequence design, protective chemical modification, and delivery engineering, whether through lipid nanoparticles, polymeric carriers, cell-penetrating peptides, or GalNAc conjugates that home in on hepatocytes. The review argues that integrating computational screening with cell-based assays, exon-skipping experiments, high-throughput sequencing, and RNA immunoprecipitation creates a hybrid pipeline that narrows the design space before any expensive laboratory work begins.

The authors conclude that in silico approaches are now unavoidable in ASO development, offering speed, cost savings, and accuracy that no purely experimental campaign can match, while cautioning that computational predictions will never fully capture biological complexity on their own. Future progress, they suggest, depends on feeding high-resolution experimental data from cryo-electron microscopy, X-ray crystallography, and small-angle X-ray scattering back into predictive models, and on developing algorithms that handle non-canonical structures and dynamic protein–RNA interactions. As ASO platforms expand beyond rare genetic diseases into oncology, infectious disease, and neurodegeneration, the fusion of next-generation computational design with rigorous experimental validation is expected to accelerate the arrival of safer, more specific, and more personalized gene-silencing medicines.

Subject of Research: Computational design and engineering of antisense oligonucleotide therapeutics

Article Title: Silencing with precision: In silico strategies for antisense oligonucleotide engineering

Article References: Nagaraj, A., Kavana, C. P., Harendra, B., Srinivasa, S., Dharmashekara, C., Kollur, S. P., Shreevatsa, B., & Shivamallu, C. (2026). Silencing with precision: In silico strategies for antisense oligonucleotide engineering. Molecular Biology Reports, 53(1), Article 1639. https://doi.org/10.1007/s11033-026-12635-8

Image Credits: AI Generated

DOI: 10.1007/s11033-026-12635-8

Keywords: antisense oligonucleotides, gene silencing, in silico drug design, RNase H1, splice modulation, molecular docking, molecular dynamics, RNA secondary structure, nucleic acid therapeutics, nusinersen, personalized medicine, machine learning

The motivation behind the work lies in the limitations of the current diagnostic toolkit. Next-generation sequencing is the gold standard for decoding viral genomes and tracking mutations, but it is a complex, multi-step process that can take weeks and may occasionally miss or misinterpret variants. Reverse transcription polymerase chain reaction, by contrast, is fast and widely deployed from bench to bedside, yet it can only detect sequences that are already known: if mutations arise in the primer-binding regions, the assay may fail entirely, and designing new primers for an emerging variant can take nearly as long as full sequencing. Raman spectroscopy offers a third path — one that does not read the base sequence at all, but instead captures the structural consequences of sequence variation as they are written into the RNA molecule’s vibrational behavior.

In the study, viral variants were propagated in VeroE6/TMPRSS2 cells, and RNA was extracted from culture supernatants using a commercial kit. Raman spectra were collected with a confocal LabRAM HR800 spectrometer excited at 532 nanometers, achieving spectral resolution better than one wavenumber with an internal neon lamp reference for calibration. Ten spectra were averaged for each RNA sample across an area of roughly two square millimeters, and the resulting spectra were baseline-corrected and deconvoluted into Gaussian-Lorentzian sub-band components using an automatic solver built on machine learning algorithms. Crucially, the optimization preserved the spectral positions and bandwidths of elementary molecular components while treating relative intensities as structure-dependent variables, a constrained approach that dramatically reduces the risk of overfitting in spectrally congested regions.

The spectra, recorded between 400 and 1800 wavenumbers, were morphologically similar across all 20 strains but differed in telling details. Ring-breathing markers for the four RNA bases appeared at characteristic positions: about 724 wavenumbers for adenine, 778 for cytosine, 672 for guanine and 791 for uracil. Backbone signals included the O-P-O stretching modes near 807 and 824 wavenumbers and a C3′-O-P bending band at 432 wavenumbers, while ribofuranose sugar rings contributed a strong C-O/C-C stretching signal at roughly 1045 wavenumbers. When the researchers compared the fractional intensities of the base signals with the actual base fractions computed from genomic data, the two disagreed substantially — pyrimidine signals systematically overestimated their base fractions and purine signals underestimated them. These mismatches could not be explained by base composition, which is nearly homogeneous across strains, but instead arise from differences in Raman cross-sections and from hyperchromic and hypochromic effects tied to RNA secondary structure.

That structural sensitivity is precisely what makes the approach powerful. Single-stranded viral RNA folds into stems, hairpin loops, pseudoknots, bulge loops, internal loops and multibranched junctions, and each motif modulates the Raman spectrum in its own way. Double-stranded stems enhance base stacking and hydrogen bonding, sharpening and intensifying purine ring signals; hairpin loops bend the phosphodiester backbone and alter its vibrational response; bulges and internal loops distort local geometry and reduce stacking-related intensity. To rationalize these effects, the team predicted the secondary structures of all 20 genomes with the ViennaRNA package using sequences from the GISAID database, computing positional descriptors including base-pairing probability, cumulative minimum-free-energy nesting depth, Shannon entropy and a custom composite “strain score” flagging nucleotides that are structurally engaged yet conformationally unstable.

Several quantitative relationships emerged. The ribofuranose ring signal at 1045 wavenumbers showed a clean linear dependence on genome length — the longer the RNA chain, the stronger the signal — because the sugar ring’s two additional dihedral angles affect only the pucker conformation and are not directly strained by backbone bending. Phosphodiester backbone signals, by contrast, scattered far more widely, since the six backbone dihedral angles are directly perturbed by kinks introduced through non-canonical base pairing. Searches for AUUCU repeats and kink-turn motifs, known sources of backbone deformation, found that their counts varied only marginally among the strains and could not alone account for the observed scatter. More strikingly, the 432-wavenumber C3′-O-P bending signal displayed an exponential hyperchromic dependence on the number of hairpin loops, with the ancestral strain — which had the most hairpins, 592 — showing the strongest intensity. The researchers interpret this as evidence that RNA backbone vibrations become more Raman-active as secondary-structure complexity increases, likely because tight structural constraints reduce vibrational energy dissipation and make the bending mode more cooperative.

A second key marker appeared near 1460 wavenumbers, a composite band dominated by C=N and C=C stretching in guanine and adenine together with C-H and N-H bending. Its relative intensity rose linearly with average stem length, tracking the increased base stacking and hydrogen bonding of longer double-stranded regions. Neither of these trends could be reproduced for neighboring bands of similar origin, underscoring the unique sensitivity of the 1460-wavenumber signal to base-stacking effects. Because hairpins and stems regulate viral infectivity — stabilizing the genome against host nucleases, enabling translation initiation of the ORF1a/ORF1ab replicase polyproteins through conserved stem-loop structures in the 5′ untranslated region, and hiding immunostimulatory double-stranded motifs from innate immune sensors such as RIG-I, MDA5, PKR and the OAS family — the two Raman markers can be read as proxies for replication efficiency and immune evasion, respectively.

Plotted as a rescaled intensity timeline across the 20 strains, these markers tell an evolutionary story. The ancestral variant scored highest on both functions. Successive variants showed a general decline in the translation/replication indicator, with the Theta variant as a notable exception, its secondary structure scoring closest to the ancestor’s — consistent with reports of concerning spike mutations in that lineage. From the Mu variant onward, the immune-evasion indicator climbed sharply through the first wave of Omicron subvariants, in agreement with the known immune-escape character of Omicron, and a further, smaller spike appeared in a later Omicron subvariant. The authors note that such trade-offs between immune escape and replication fitness are well documented in HIV, influenza and hepatitis C, and that Omicron’s emergence indeed traded replication efficiency for immune evasion.

Classification posed its own challenge. Principal component analysis, a standard chemometric tool in biomedical spectroscopy, succeeded in separating only a minority of the 20 RNA samples, a partial failure the researchers attribute to the extreme complexity of RNA secondary structure. Instead, they converted each deconvoluted spectrum into a Raman barcode, assigning to each sub-band a line whose thickness and spacing encode the band’s width and area. Every one of the 20 barcodes proved unique, univocally locating the molecular structure of each sample, including the spectroscopic imprint of its secondary structure. Compared with DNA barcoding initiatives that require sequencing short genetic markers, the Raman approach skips sequence analysis altogether, potentially enabling on-site pathogen identification in a time dictated solely by RNA extraction, once a comprehensive spectral library and standardized deconvolution protocols are in place.

The authors are careful to frame the work as a proof of concept with clear limitations. The study used purified viral RNA from controlled cell cultures rather than clinical specimens containing host RNA, proteins and debris; it cannot identify individual nucleotide substitutions or determine primary sequence, and should be regarded as complementary to genomic sequencing rather than a replacement; and the barcode algorithm will require inter-laboratory validation, standardized acquisition protocols and expanded spectral libraries before clinical translation. Its ability to resolve newly emerging variants differing by only a few mutations remains to be systematically tested. Nevertheless, the study establishes that Raman spectroscopy captures structural information encoded within viral RNA that conventional compositional analyses cannot reach — amplified spectral differences arising from small genomic ones — and demonstrates that a Raman library of viral RNA barcodes could one day give virologists a rapid, low-cost route to variant identification in emergency and critical-care settings, precisely where timely classification of emerging strains matters most.

Subject of Research: Raman spectroscopic fingerprinting of SARS-CoV-2 viral RNA secondary structure for variant classification

Article Title: Raman Analysis of RNA Nucleotide Strands From SARS‐CoV‐2 Variants and Subvariants: A Step Forward in the Definition of “Raman Genome”

Article References: Pezzotti, G., Yasukochi, Y., Angiola, G., Ueno, T., Ikegami, S., Okawa, R., Adachi, T., Zhu, W., Mazda, O., Grillo, A., Higasa, K., & Okuma, K. (2026). Raman Analysis of RNA Nucleotide Strands From SARS‐CoV‐2 Variants and Subvariants: A Step Forward in the Definition of “Raman Genome”. Advanced Science, Article e77679. https://doi.org/10.1002/advs.77679

Image Credits: AI Generated

DOI: 10.1002/advs.77679

Keywords: SARS-CoV-2, Raman spectroscopy, viral RNA, Raman genome, variants and subvariants, RNA secondary structure, Raman barcode, molecular diagnostics, machine learning, ViennaRNA, immune evasion, viral evolution

The analogy is more than a metaphor. DNA is written in an alphabet of four nucleotide letters, proteins in an alphabet of twenty amino acids, and RNA in a four-letter code of its own. Just as words derive meaning from their context in a sentence, the function of a nucleotide or amino acid depends on its position within a sequence and on the sequences that surround it. The transformer architecture, introduced in 2017 in the landmark paper Attention Is All You Need, excels precisely at capturing such long-range contextual dependencies. By applying self-attention mechanisms across millions of biological sequences, models can learn statistical patterns that reflect evolutionary constraints, structural motifs, and functional sites without any explicit programming.

The review organizes the field into biological foundation models, general-purpose systems pretrained on massive unlabeled sequence databases, and specialized models tailored to particular problems. On the protein side, models such as ProtTrans and the Evolutionary Scale Modeling family demonstrated that unsupervised learning on hundreds of millions of sequences yields internal representations from which structure and function can be read off. This line of work culminated in AlphaFold and its successors, which achieved atomic-level accuracy in protein structure prediction and, with AlphaFold 3, extended to predicting interactions between proteins, nucleic acids, and small molecules. Generative protein LLMs such as ProGen and ProGen2 have gone further, producing entirely novel amino acid sequences that fold into functional proteins, a capability the review highlights as one of the most consequential advances of the past decade.

DNA language models have followed a parallel trajectory. The Nucleotide Transformer, built and evaluated on human and multi-species genomes, showed that robust foundation models could be trained at nucleotide resolution. Subsequent work demonstrated that such models are powerful predictors of genome-wide variant effects, meaning they can estimate whether a mutation in a patient’s genome is likely to be harmful, a task central to clinical genetics. The review also points to Evo and its successor Evo 2, genome-scale models capable of modeling and designing sequences across all domains of life, from bacteria to humans. These systems operate at a scale that was unthinkable for genomics only a few years ago, processing contexts long enough to encompass entire genes and their regulatory neighborhoods.

RNA has proven a particularly fertile ground because its structure and function are notoriously difficult to predict from sequence alone. Classical approaches relied on thermodynamic folding algorithms, but deep learning models such as UFold and RNAformer have shown that learned representations can match or exceed physics-based methods for secondary structure prediction. More recently, language model-based deep learning approaches have achieved accurate RNA 3D structure prediction, and motif-aware pretraining strategies have produced multi-purpose RNA models that handle splicing prediction, binding site identification, and function annotation within a single framework. The review notes that self-supervised learning on millions of primary RNA sequences from dozens of vertebrate species has substantially improved sequence-based splicing prediction, with direct implications for interpreting disease-causing mutations in untranslated regions and splice sites.

Beyond the three canonical sequence types, the review surveys a rapidly growing ecosystem of specialized applications. Interaction prediction models forecast which proteins bind which RNAs, which RNAs pair with other RNAs, and how mutations perturb these contacts. Structure prediction systems now cover proteins, RNA, and biomolecular complexes at near-experimental accuracy. Perhaps most striking is the single-cell revolution: foundation models such as scGPT, scBERT, and Geneformer treat gene expression profiles across thousands of individual cells as a kind of language, learning universal representations that support cell type annotation, perturbation response prediction, and the integration of massive transcriptomic datasets. Studies assessing GPT-4 for cell type annotation suggest that general-purpose chatbots can also contribute, though specialized models currently hold the edge on most benchmarks.

The authors are candid about the obstacles. Hallucination, the tendency of generative models to produce plausible but false outputs, is a documented failure mode of natural language generation that carries serious risks when outputs inform drug design or clinical interpretation. Benchmarking remains immature: evaluation suites such as DART-Eval for regulatory DNA and PertEval-scFM for perturbation prediction reveal that performance gains do not always transfer across datasets and cell types. Computational cost is another barrier, since genome-scale attention models demand enormous memory, motivating architectural innovations such as the Hyena hierarchy and Mamba-based state space models that the review identifies as promising directions for long-context biological modeling. Data quality, batch effects in single-cell experiments, and the interpretability of learned representations round out the open challenges.

What distinguishes this review from many surveys of artificial intelligence in biology is its educational dimension. Because it appears in a journal focused on digital education, the authors devote substantial attention to integrating LLMs into genomics teaching and learning. They propose practical projects in which students fine-tune pretrained models on small datasets, use conversational agents to simplify bioinformatics workflows, and learn to critically evaluate model outputs. Tools such as BioMANIA, which allows researchers to interrogate bioinformatics pipelines through natural language conversation, and BioCoder, a benchmark for bioinformatics code generation, illustrate how LLMs can lower the technical barriers that traditionally separate biologists from computational analysis. The authors argue that fluency in these tools is becoming as essential to the modern life scientist as statistical literacy.

The trajectory the review describes is remarkably compressed. In roughly four years, language models have progressed from proof-of-concept embeddings to systems that simulate 500 million years of evolutionary history, design functional proteins, predict the pathogenicity of genetic variants, and annotate cell atlases containing millions of cells. The Wuhan University team frames the current moment as an inflection point: the same scaling laws and reasoning enhancements that drive progress in general-purpose AI, including reinforcement learning approaches that sharpen chain-of-thought reasoning, are poised to accelerate biological discovery further. If the challenges of reliability, evaluation, and interpretability can be met, the authors suggest, the fusion of language models and genomics may redefine both how biology is researched and how the next generation of scientists is trained, turning the genome from an inscrutable code into a text that machines can read, annotate, and even rewrite.

Subject of Research: Applications of large language models in genomics, including biological foundation models for protein, DNA, and RNA sequence analysis

Article Title: AI-Empowered Genome Decoding: Applications of Large Language Models in Genomics

Article References: Li, S., Fan, W., & Zhou, Y. (2025). AI-Empowered Genome Decoding: Applications of Large Language Models in Genomics. Frontiers of Digital Education, 2(1), Article 14. https://doi.org/10.1007/s44366-025-0051-1

Image Credits: AI Generated

DOI: 10.1007/s44366-025-0051-1

Keywords: large language models, genomics, deep learning, protein structure prediction, DNA language models, RNA secondary structure, single-cell analysis, foundation models, AlphaFold, transformer architecture, bioinformatics education, variant effect prediction