A fully-open structure-guided RNA foundation model for robust structural and functional inference

A fully open structure-guided RNA foundation model for robust structural and functional inference | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results A fully open structure-guided RNA foundation model for robust structural and functional inference View ORCID Profile Heqin Zhu , Ruifeng Li , Ao Chang , Haobin Chen , Feng Zhang , View ORCID Profile Fenghe Tang , Tong Ye , Xin Li , Yunjie Gu , View ORCID Profile Peng Xiong , View ORCID Profile S. Kevin Zhou doi: https://doi.org/10.1101/2025.08.06.668731 Heqin Zhu 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Heqin Zhu Ruifeng Li 3 College of Computer Science and Technology, Zhejiang University , Hangzhou, Zhejiang, 310027, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Ao Chang 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Haobin Chen 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Feng Zhang 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Fenghe Tang 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Fenghe Tang Tong Ye 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Xin Li 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Yunjie Gu 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Peng Xiong 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China 5 Biomedical Basic Research Center (BBRC) of Jiangsu Province , Suzhou, Jiangsu, 215123, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Peng Xiong For correspondence: xiongxp{at}ustc.edu.cn skevinzhou{at}ustc.edu.cn S. Kevin Zhou 1 School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC) , Hefei, Anhui, 230026, China 2 Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), Suzhou Institute for Advance Research, USTC , Suzhou, Jiangsu, 215123, China 4 Jiangsu Provincial Key Laboratory of Multimodal Digital Twin Technology , Suzhou, Jiangsu, 215123, China 5 Biomedical Basic Research Center (BBRC) of Jiangsu Province , Suzhou, Jiangsu, 215123, China 6 State Key Laboratory of Precision and Intelligent Chemistry, USTC , Hefei, Anhui, 230026, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for S. Kevin Zhou For correspondence: xiongxp{at}ustc.edu.cn skevinzhou{at}ustc.edu.cn Abstract Full Text Info/History Metrics Supplementary material Data/Code Preview PDF Abstract RNA language models have achieved strong performances across diverse downstream tasks by leveraging large-scale sequence data. However, RNA function is fundamentally shaped by its hierarchical structure, making the integration of structural information into pre-training essential. Existing methods often depend on noisy structural annotations or introduce task-specific biases, limiting model generalizability. Here, we propose structRFM, a structure-guided RNA foundation model that is pre-trained on millions of RNA sequences and secondary structures data by integrating base pairing interactions into masked language modeling through a novel pair matching operation. We further introduce MUSES (multi-source ensemble of secondary structures) to mitigate model bias, and a dynamic masking ratio to balance the structure-guided mask and nucleotide-level mask. structRFM learns joint knowledge of sequential and structural data, producing versatile representations, including classification-level, sequence-level, and pairwise matrix features, that support a broad spectrum of downstream adaptations. structRFM ranks among the top models in zero-shot homology classification across seventeen biological language models, and sets new benchmarks for secondary structure prediction. structRFM further derives Zfold, which enables robust and reliable tertiary structure prediction, with consistent improvements in estimating 3D structures and their accordingly extracted 2D structures, achieving a pronounced about 20% performance gain compared with baselines and comparable performances with AlphaFold3 on CASP15-natural, CASP16, and RNA-Puzzles datasets. In functional tasks such as internal ribosome entry site identification, structRFM achieves a whopping 48% performance gain in F1 score. Furthermore, state-of-the-art performances in extensive experiments across novel RNA families and long non-coding RNAs indicate the robustness and generalizability of structRFM. These results demonstrate the effectiveness of structure-guided pre-training and highlight a promising direction for developing multi-modal RNA language models in computational biology. To support the broader scientific community, we have made the 21-million sequence-structure dataset and the pre-trained structRFM model fully open-source, facilitating the development of multimodal foundation models in biology. 1 Introduction Ribonucleic acid (RNA), which plays an essential role in molecular biology, not only serves as a carrier of genetic information to bridge the central dogma of deoxyribonucleic acid (DNA) transcription and protein translation, but also folds into various structures and functions as a versatile architect of cellular processes [ 1 ]. Beyond its classical roles such as ribosomal RNA (rRNA) and transfer RNA (tRNA) in gene expression, numerous non-coding RNAs (ncRNAs) have emerged as essential regulators across cellular contexts, such as the internal ribosome entry site (IRES) for 5’ cap-independent translation in circular RNAs (circRNA) [ 2 ], small nuclear RNA (snRNA) for splicing pre-mRNA to produce mature mRNA [ 3 ]. These profound functions of ncRNAs are encoded within their distinguished sequence patterns and structures, which demonstrate hierarchical relationships: primary sequence motifs determine base pair patterns and folding tertiary structures, which further dictate specific biochemical activities and interactions. Unraveling these complex relationships, linking sequence to structure and ultimately to biological function, is crucial for understanding fundamental RNA biology and advancing the development of next-generation RNA-centric therapeutics. Decades of innovation in nucleic acid sequencing have produced an unprecedented volume of RNA sequence data [ 4 ], revolutionizing the human ability to map the complexity of the transcriptome [ 5 , 6 ]. Inspired by advances in natural language processing (NLP), RNA language models, pre-trained in a task-agnostic, self-supervised manner, have demonstrated great potential in extracting representative patterns and capturing contextual relationships from large-scale RNA sequential data, offering a considerable promise for systematically decoding RNA’s molecular language, predicting tertiary structures [ 7 , 8 ], specializing in functional tasks [ 9 , 10 ], and advancing both fundamental research and translational applications. Most existing RNA language models are primarily pre-trained on non-coding RNA sequences for general-purpose applications [ 11 ]. Examples include RNA-FM [ 12 ] and RiNALMo [ 13 ], which employ standard BERT architectures; RNAErnie [ 14 ], which incorporates motif patterns during pre-training; and AIDO.RNA [ 15 ], which scales the model to 1.6B parameters. These models exhibit strong generalizability across diverse downstream tasks. In contrast, other models focus on specialized RNA sequences for specific objectives, such as PlantRNA-FM [ 16 ] to identify functional RNA motifs in plants, UTR-LM [ 9 ] to analyze untranslated regions and related functions of mRNA, and SpliceBERT [ 10 ] to improve prediction of RNA splicing sites. As RNA structure plays a central role in defining molecular properties, mediating interactions with other biomolecules, and ultimately determining biological function [ 17 , 18 ], incorporating structural information into the pre-training of language models holds substantial potential beyond sequence alone. Several recent RNA language models integrate encoded structural data. For instance, RNABERT [ 19 ] incorporates aligned seed structures from the Rfam database into base embeddings to capture structural differences across RNA families. ERNIE-RNA [ 20 ] introduces structural features via a pairwise position matrix embedded within the attention mechanism. MP-RNA [ 21 ] and UTR-LM [ 9 ] adopt an auxiliary loss during pretraining to reconstruct masked secondary structures. While these approaches mark significant progress, they also present key limitations. RNABERT depends on potentially inaccurate seed alignments, which may hinder RNA family discrimination. ERNIE-RNA’s pairwise matrix, based on simplified interaction scores (e.g., 3 for GC, 2 for AU, 0.8 for GU pairs), may inadequately capture the complexity of RNA structural features. MP-RNA and UTR-LM rely on ground-truth structures from ViennaRNA [ 22 ] and incorporate task-specific supervised losses, diverging from task-agnostic self-supervised learning and potentially biasing the model toward secondary structure prediction. To overcome these limitations and support robust and generalizable downstream adaptation, it is essential to develop a new pre-training strategy that incorporates accurate structural information while maintaining task agnosticism. Motivated by this, we introduce structRFM , a fully open structure-guided RNA foundation model, pre-trained using an elaborately designed structure-guided masked language modeling (SgMLM) strategy. This approach implicitly encodes base-pairing interactions through informed masking patterns and jointly models structural and sequential dependencies. structRFM effectively harnesses structural priors while maintaining task-agnostic pre-training, enabling robust generalization across a broad spectrum of downstream tasks, including zero-shot prediction, secondary and tertiary structure prediction, and multiple function inferences ( Fig. 1a ). Our work makes three key contributions toward integrating structural knowledge into RNA foundation modeling. Download figure Open in new tab Fig. 1: Overview of structRFM. a structRFM as an RNA foundation model for zero-shot or downstream-adapted structural and functional inferences. The model outputs classification feature, sequence feature, and pairwise matrix feature, supporting tasks such as zero-shot homology classification, zero-shot 2D structure prediction, 2D and 3D structure prediction, splice site prediction, IRES identification, and ncRNA classification. b Structure-guided pre-training workflow of structRFM. RNA sequences from RNAcentral are filtered by length, and their secondary structures are obtained using the proposed MUSES strategy, which incorporates multi-source ensemble of secondary structure to construct a sequence-structure library for SgMLM pre-training. c The structure-guided masked language modeling (SgMLM) strategy. RNA base-pairing information is used to generate an integrated mask that combines random nucleotide-level and structure-guided masking, with the proportion of structure-guided masking dynamically increasing during pre-training. First, we design SgMLM, a structure-guided pre-training strategy ( Fig. 1c ), featuring two core components: structure-guided masking and dynamic masking balance. The first component selectively masks input tokens corresponding to canonical base pairs within local structural contexts, encouraging the model to recover base-pair interactions based on neighboring loop regions. This implicitly guides structRFM to capture both RNA sequential patterns and hierarchical structural regularities without relying on task-specific objectives, uncovering diverse nucleotide-to-nucleotide relationships, facilitating the transfer of sequential and structural knowledge to downstream structure and function prediction tasks. The secondary component dynamically increases the proportion of structure-guided masking within the overall mask during pre-training, gradually shifting the model’s focus from sequence-level to structure-informed representations, and finally balances nucleotide-wise and structure-wise masking. Furthermore, to address the inherent complexity of RNA folding, we introduce MUSES (Multi-source ensemble of secondary structures). This strategy integrates three complementary methodological paradigms for RNA secondary structure prediction: ViennaRNA [ 22 , 23 ] (thermodynamics-based), CON-TRAfold [ 24 ] (probability-based), and BPfold [ 25 ] (deep learning-based, leveraging base pair motif energy priors). By synergizing these diverse approaches, MUSES miti-gates biases inherent to reliance on a single predictor, more faithfully captures the full spectrum of RNA conformations, thereby enhances the robustness of structural annotations for pre-training, directly facilitating improved generalizability of the resulting foundation model. Second, we comprehensively evaluate structRFM across zero-shot assessments, downstream-adapted structure and function inference tasks. structRFM exhibits strong generalization ability, ranking among the top-performing models in zero-shot homology classification and zero-shot secondary structure prediction ( Fig. 2c ) within seventeen biological language models. For secondary structure prediction, structRFM attains the best F1 score of 0.873 and 0.641 on ArchiveII600 [ 26 ] and bpRNA-TS0 [ 27 ] datasets ( Fig. 3c and d ), respectively. In tertiary structure prediction, structRFM (derived method termed “Zfold”) outperforms ablation version, trRosettaRNA [ 28 ], and other language model integrated method on the CASP15 [ 29 ], CASP16 [ 30 ], and RNA-Puzzles datasets ( Fig. 3h ), across metrics such as ϵ RMSD [ 31 ], TM-score [ 32 ], and GDT-TS [ 33 ], respectively. These gains are corroborated by consistent increases in the interaction network fidelity (INF) metric [ 34 ] of secondary structures extracted from the predicted 3D structures ( Fig. 3i ), affirming the effectiveness and robustness of SgMLM strategy for structural adaptation. Notably, structRFM also surpasses AlphaFold3 [ 35 ] in TM-score, yielding an improvement of 25.0% on CASP15-natural and 23.4% on RNA-Puzzles. Besides, structRFM shows remarkable ability for function prediction ( Fig. 6 ), achieving state-of-the-art performances on splice site prediction, IRES identification, and ncRNA classification, benchmarking a top-k accuracy of 0.416 on SpliceAI dataset [ 3 ], an F1 score of 0.701 on IRES dataset [ 36 , 37 ], and an F1 score of 0.979 on nRC dataset [ 38 ], respectively. Furthermore, structRFM shows great generalizability and robustness on new RNA families and long sequences, obtaining state-of-the-art performance (0.65-0.70 in INF) on new RNA families from Rfam for RNA secondary structure prediction, and F1 > 0.95 for long non-coding RNAs classification. Download figure Open in new tab Fig. 2: Feature visualization and zero-shot evaluation of structRFM. a Visualization of RNA embeddings from structRFM and a random baseline using t-SNE on a subset of RNAcentral (23,994 RNAs), showing that structRFM makes clear separations of RNA families. Also shown are two examples of matrix features extracted from structRFM and corresponding contact maps. b Cosine similarity distributions between homologous and non-homologous RNA pairs for different models on Rfam (24,523 RNAs) and ArchiveII600 (3,911 RNAs) datasets, and c the overlap ratio (OR) of the two distributions is demonstrated, showing remarkable homology clarification ability of structRFM. The mean value is annotated for each method and each dataset. d Assessment of structRFM on zero-shot secondary structure prediction. The first figure shows the INF metrics across different layer and head of attention maps extracted from structRFM, the second figure shows the INF metrics of the attention map from layer 11 and head 7 after postprocessing across thresholds from 0 to 1 with an interval of 0.001, and the third figure shows an example attention heatmap. Download figure Open in new tab Fig. 3: structRFM for structure prediction. a, b Neural network architecture for secondary and tertiary structure prediction, respectively. c The pipeline for analyzing the sequence similarity from test dataset against the training dataset. d, e The INF of secondary structure prediction among structRFM and other models on ArchiveII600 (3,911 RNAs) and bpRNA-TS0 (1,305 RNAs) datasets. The white line in each violin denotes the medium number, while the mean value is displayed above each violin. f INF comparison of structRFM and structRFM − on the nine families of ArchiveII600. The number of samples of each RNA family is displayed at the top of the figure. g, h The relationship between the INF metric and the sequence similarity of ArchiveII and bpRNA-TS0 dataset against their training dataset RNAStrAlign and bpRNA-TR0, respectively. The line in blue is the linear regression fit. i ϵ RMSD, TM-score, GDT-TS, and INF performance mparison for tertiary structure prediction on CASP15-natural (8 RNAs), CASP16 (13 RNAs) [ 30 ] and RNA-Puzzles (20 RNAs), where the error bar in gray represents 95% confidence index. j Examples from predicted tertiary structures. Native structure is in lightpink. Third, we release the entire suite of resources to the community, including a 21-million sequence-structure dataset with according base-pairing interactions of structure ensembles, the pre-trained structRFM model, and the fine-tuned state-of-the-art models for structural and functional tasks. These open resources are intended to democratize the development of RNA foundation models and support future work in multi-modal biological modeling. Conclusively, structRFM demonstrates its effectiveness and versatility as a fully open RNA foundation model for robust downstream adaptation. Moreover, the SgMLM pre-training strategy and the collected sequence-structure dataset introduced here are readily extensible and can be seamlessly integrated into other pre-training frameworks, marking a substantial step toward empowering RNA foundation models with multi-modal structural knowledge and advancing broader biological research. 2 Results In this section, we conduct a series of evaluations to demonstrate the versatility and effectiveness of the proposed structRFM foundation model that is pre-trained on 21 millions RNA sequences and secondary structures. We begin with feature visualization and zero-shot tasks, where the model is tested without labeled training data. structRFM produces RNA sequence feature embeddings that encode family-level and base-pair interaction information, which are validated on zero-shot RNA homology classification and zero-shot secondary structure prediction. We further explore the model’s capability in structure prediction, showing that structRFM consistently and reliably improves the prediction accuracy of RNA secondary and tertiary structures. We also evaluate the geometric and topological correctness of the predicted tertiary structure by structRFM. Next, we fine-tune structRFM on several supervised function prediction tasks, including splice site prediction, IRES identification, and ncRNA sequence classification, demonstrating its strong transferability across nucleotide-level and sequence-level applications. Finally, to explore the generalizability of structRFM on novel RNA families and unseen long non-coding RNA sequences, we assess the performance of structRFM for RNA secondary structure on new RNA families from Rfam and for lncRNA sequence classification on two datasets of human and mice lncRNA. We denote structRFM without SgMLM as structRFM − , respectively. The overall results on these tasks are displayed in Extended Data Table 1 . Summary of all used datasets in this study is displayed in Extended Data Table 2 2.1 structRFM extracts category-discriminative and structure-encoded features The task-agnostic self-supervised pre-training enables structRFM to capture the intrinsic characteristic of any RNA sequence, while the SgMLM strategy allows structRFM to extract structure-encoded feature embeddings. For an input RNA sequence of length L , structRFM extracts classification feature F cls ∈ ℝ 1× D and sequence feature F seq ∈ ℝ L × D that represent the RNA category and nucleotide-wise sequential embeddings, respectively. Based on this, we can further obtain the matrix feature F mat ∈ ℝ L × L by applying averaging to the 12 attention maps , i = 11, j = {0, 1, …, 11} of the 12 heads from the last layer, which encodes the nucleotide-to-nucleotide relationships. To qualitatively assess the learned sequential and structural knowledge, we visualize the classification feature and matrix feature using data from the pre-training dataset. In detail, we randomly sample 23,994 RNA sequences across 19 RNA families and feed them into structRFM to extract the sequence features and matrix features. Using the t-SNE [ 39 ] algorithm, we project the classification feature F cls of structRFM into two dimensions, contrasting it with a randomly initialized model without SgMLM pre-training. As illustrated in Fig. 2a , structRFM exhibits strong category-discriminative capability, effectively clustering RNA families, projecting different categories of RNA features into distinct clusters, whereas the random model fails to do so, projecting the RNA sequences into mixed vector representations of latent space. Furthermore, we visualize the matrix feature as a heatmap, comparing it to the native contact map. The high-response regions (in blue) highlight structRFM’s ability to capture structural information, particularly base-pairing interactions. Collectively, these results demonstrate that structRFM successfully learns RNA representations that encode both sequential and structural properties, enhancing its adaptability for structure- and function-related tasks. 2.2 structRFM enables effective zero-shot prediction Following previous work [ 11 ], we perform zero-shot assessments, i . e ., zero-shot homology classification and zero-shot secondary structure prediction to evaluate whether the pre-trained model acquires generalizable structural and functional knowledge from unlabeled data, such as structural regularities, functional motif information, and base-pairing formation rules. To evaluate whether the learned sequence embeddings encode homology-level information, we conduct zero-shot RNA homology classification using the Rfam and ArchiveII datasets. Specifically, we randomly sample 100,000 homologous and 100,000 non-homologous sequence pairs from each dataset. In Rfam, homologous pairs come from the same RNA family, where non-homologous pairs are from different families. For ArchiveII, homologous sequences share the same RNA type. We extract the sequence embeddings and compute the cosine similarity between each pair of sequences. The cosine similarity distributions are visualized, and the overlap ratio (OR) of these two distributions is computed for comparison, where a lower OR indicates less overlap between distributions and thus better discriminative ability. As shown in Fig. 2b, c , structRFM achieves state-of-the-art results, ranking top 4 among 17 DNA, RNA, and protein language models on the Rfam dataset (OR=0.10) and ArchiveII dataset (OR=0.45), showing a strong separation between homologous and non-homologous embeddings, despite AIDO.RNA costs 10 times more of the parameters. More distributions of other models are visualized in Supplementary Figs. 1 and 2. In contrast, DNA language models (e.g., DNABERT and NT) and specialized RNA language models (e.g., SpliceBERT and UTR-LM) perform poorly, particularly on ArchiveII (OR > 0.6), suggesting that general task-agnostic pre-training better captures of family-level semantics. Extended Data Table 3 lists the detailed information of structRFM with other eight RNA language models. Beyond classification, we further assess whether the model captures RNA secondary structures. Inspired by previous findings in RNA-MSM [ 40 ], where 2D attention maps encode base-pairing patterns, we construct a test set of 116 RNA sequences from the PDB database and extract corresponding features using structRFM for zero-shot evaluation. In this experiment, 144 attention maps are extracted from the 144 heads from 12 layers of structRFM, with each layer containing 12 heads. Each attention map is recognized as a base-to-base pairing contact map. After post-processing the attention map using a range of thresholds from 0 to 1 with an interval of 0.001, we calculate the INF metric to assess zero-shot secondary structure prediction. As shown in Fig. 2d , in the grid search from each layer and each head, only a limited number of layers and heads are effective for zero-shot secondary structure prediction, while other layers and heads may focus on the sequential data and joint modeling. In the threshold selection experiment, the attention map of head 5 from layer 11 achieves the highest INF of 0.44 with a threshold of 0.545, indicating a moderate ability to capture secondary structure pairing features without any task-specific supervision or labeled data. The sharp peak around the optimal threshold highlights the sensitivity of prediction to threshold choice, especially in zero-shot settings, where no calibration is available. 2.3 structRFM achieves new benchmarks for secondary structure prediction As shown in Fig. 3a , for secondary structure prediction, we utilize structRFM as the backbone to extract the sequence feature F seq . We concatenate extracted feature with the one-hot embedding of the input RNA sequence, and further adopt a structure module - MXfold2 [ 41 ] to predict four predefined stem-loop structural patterns. The final predicted secondary structure is obtained by processing the predicted patterns using Zuker-style dynamic programming approach. Following RNAErnie [ 14 ], we fine-tune structRFM and baseline methods on RNAStrAlign [ 42 ] (19,313 RNAs) and bpRNA-1m [ 27 ] (12,114 RNAs), and evaluate on ArchiveII600 [ 26 ] (3,911 RNAs) and bpRNA-TS0 [ 27 ] (1,305 RNAs), respectively. All models are evaluated under the same configuration as RNAErnie for a fair comparison. As Fig. 3c, d , and Supplementary Table 1 demonstrate, structRFM achieves the best performance across all methods, including fine-tuned RNA language models, specialized learning-based methods, and traditional non-learning methods. Specifically, structRFM reaches an INF score of 0.890 on ArchiveII600 and an INF score of 0.669 on bpRNA-TS0, much better than RiNALMo-Mega (0.858 on ArchiveII600, p = 6.17 × 10 −15 , two-sided t-test; 0.638 on bpRNA-TS0, p = 0.019, two-sided t-test), also outperforming other non-LM deep learning methods that are tailored for secondary structure prediction, such as MXfold2 [ 41 ] (0.768 on ArchiveII600, 0.558 on bpRNA-TS0), Contextfold [ 43 ] (0.842 on ArchiveII600, 0.546 on bpRNA-TS0). When comparing structRFM with its ablated variant, structRFM − , which lacks the structure-guided pre-training component, structRFM consistently outperforms structRFM − on ArchiveII600 by 4.46% and bpRNA-TS0 by 6.19% with statistical significance ( p = 2.74 × 10 −20 for ArchiveII600 and p = 1.10 × 10 −3 for bpRNA-TS0, two-sided t-test. These results indicate that structure-guided pre-training effectively imparts prior knowledge that facilitates downstream secondary structure prediction. In addition, we evaluate family-level performance on the nine detailed families within the ArchiveII600 dataset ( Fig. 3e , Supplementary Table 2). structRFM achieves superior results across eight families with the largest number of samples. However, structRFM − slightly outperforms structRFM on the 23S rRNA family, which is attributed to lack of pre-training sequences (23S rRNA consists of about 2,900 nt, much longer than the pre-trained sequence length of 512 nt) and poor behavior of existing RNA secondary structure prediction methods on long sequences as demonstrated in Extended Data Fig 1b,c . Furthermore, we analyze sequence similarity between the training and test datasets. As Fig. 3c shows, for each sequence in both datasets, we adopt rMSA [ 44 ] and blastn [ 45 ] to generate multiple sequence alignment (MSA) from the RNAcentral [ 46 ] and Nucleotide (NT) [ 47 ] databases, then compute pairwise MSA similarities to construct a similarity matrix. For each test sequence, its similarity to the training dataset is defined as the maximum value in this matrix. Fig. 3g,h shows ArchiveII has higher similarity to its training set than bpRNA-TS0 to its respective set. The Pearson correlation gives R 2 values of 2.57% (ArchiveII) and 8.60% (bpRNA-TS0), indicating an extremely weak linear relationship between sequence similarity and INF. Yet both are highly significant (p < 0.01), ruling out random variation. 2.4 structRFM enables robust and reliable tertiary structure prediction RNA structure folds in a hierarchical manner, where tertiary structure is largely determined by its secondary structure. We hypothesize that accurate base pairing interactions can boost tertiary structure prediction, and the secondary structure extracted from predicted tertiary structures provides a complementary assessment of their reliability. Moreover, unlike sparse representations of secondary structure (e.g., dot bracket notation or binary contact maps), feature embeddings extracted from RNA language models encode rich sequential and structural information that may better support tertiary structure modeling. In this experiment, we evaluate whether structRFM pre-trained with base-pairing interactions can consistently improve the accuracy and reliability of both tertiary structures and secondary structures derived from them, showing remarkable robustness in structural adaptation. As illustrated in Fig. 3b , we adopt trRosettaRNA [ 28 ] as the structure prediction framework and use structRFM as the backbone for extracting a matrix feature representing base-to-base relationships. The original trRosettaRNA pipeline takes multiple sequence alignment (MSA) generated by rMSA [ 44 ] of input RNA sequence and the contact map of secondary structure predicted by SPOT-RNA [ 48 , 49 ] as input to predict 1D and 2D geometries (e.g., 1D orientations, 2D contacts, distances, and orientations). These are then converted into structural restraints for 3D folding via energy minimization. In our variant, we replace the input contact maps with matrix features from structRFM to assess the impact of structural embeddings on tertiary structure prediction. We follow the original trRosettaRNA setup by using a training set comprising 3,633 RNA chains from the PDB [ 50 , 51 ], filtering with a cutoff date of 2022-01 and excluding self-distillation data. We adopt trRosettaRNA framework and structRFM backbone for fine-tuning, where the resulting method is termed “Zfold” (we continue to use “structRFM” in this paper for clarity and comparison). Evaluation is performed on three benchmark datasets, namely CASP15 [ 29 ] (12 RNAs, 8 natural and 4 synthetic), CASP16 [ 30 ] (13 RNAs), and RNA-Puzzles [ 52 ] (20 RNAs). Secondary structures are extracted from predicted tertiary structures using RNAview [ 53 ], and visualized using tools R2DT [ 54 ] and Forna [ 55 ]. For evaluation metrics, we utilize nucleotide interaction centric ϵ RMSD [ 31 ] (lower is better), TM-score [ 32 ], and GDT-TS to evaluate the predicted tertiary structure, while the extracted secondary structures are evaluated by Interaction Network Fidelity (INF, pair) [ 34 ]. As shown in Fig. 3i and Supplementary Table 3, on CASP15-natural, CASP16, and RNA-Puzzles datasets, structRFM achieves excellent performances in ϵ RMSD (1.26, 1.49, 1.26), TM-score (0.45, 0.33, 0.58), GDT-TS (0.50, 0.32, 0.70), and INF (0.84, 0.60, 0.72), respectively, outperforming RiNALMo-Mega, trRosettaRNA and structRFM − by a huge gap, demonstrating superior ability in extracting representative feature than other RNA language models and in encoding of structural information into the pre-training of RNA sequences than structRFM − . Compared to AlphaFold3, a unified model leveraging diverse molecules (protein, RNA, etc.), structRFM performs slightly better on CASP15-natural and RNA-Puzzles datasets in TM-score and GDT-TS metrics. In contrast, the ablation variant structRFM − , pre-trained without the SgMLM strategy and lacking structural awareness, falls to supply knowledge of base pairing interactions for tertiary structure prediction, yielding poorer performances: ϵ RMSD of 1.48 (CASP1-natural), 1.69 (CASP16), 1.48 (RNA-Puzzles), and INF values of 0.73, 0.34, 0.69, respectively. On CASP15, CASP16, and RNA-Puzzles datasets, structRFM outperforms structRFM − significantly (p value: 0.031, 0.057, and 0.012 for ϵ RMSD; 0.041, 0.074, and 1.90 × 10 −5 for TM-score; 0.081, 0.012 and 0.64 for INF.). Fig. 4a shows RNA-Puzzles has higher average sequence similarity than the other two datasets, thus structRFM obtains better overall performances in TM-score and INF. However, for about one third of targets with less than 0.4 similarity from CASP15 and CASP16, structRFM still achieves satisfying performances (INF > 0.7, TM-score > 0.4), demonstrating the generalizability of structRFM in the situation of unseen data and missing structural templates. Notably, structRFM outperforms structRFM − across all datasets and metrics, especially at low sequence similarity. Compared to trRosettaRNA, structRFM improves secondary structure INF by 16.7% on CASP15-natural, 5.26% on CASP16, and 10.8% on RNA-Puzzles, particularly effective in no homology and difficult prediction scenarios. Representative examples (R1149 from CASP15, R1296 from CASP16, and PZ15 from RNA-Puzzles) are visualized in Fig. 3j , showing that structRFM achieves more accurate predictions, especially when homology or a template is available, while other models tend to fail under such conditions. Download figure Open in new tab Fig. 4: Analysis of predicted structures from CASP15, CASP16, and RNA-Puzzles datasets and generalizability of structRFM a The relationship between INF and TM-score metrics and the sequence similarity of CASP15 (12 RNAs), CASP16 (13 RNAs), and RNA-Puzzles (20 RNAs) against the PDB training dataset, respectively. The straight line is the linear regression fit while the surrounding curves are the 95% percentile intervals. b The concordance between the secondary structures predicted by the secondary-structure fine-tuned models (on RNAStrAlign and bpRNA1m, respectively) and those reconstructed from the tertiary structures predicted by structRFM from the three datasets. c The entanglement analysis of predicted structures by structRFM on the three datasets. d The stereochemical quality analysis of the predicted structures by structRFM on the three datasets. e The validation of structRFM generalizability on novel new RNA families (Rfam v15.0, 436 RNAs) for secondary structure prediction. f The validation of structRFM generalizability on long non-coding RNA sequences of human (18,105 RNAs) and mouse (7,401 RNAs) species. To further validate the benefits of structure-guided pre-training, we perform head-to-head comparisons of structRFM against four other methods for each structure target ( Fig. 2a ), showing consistent and target-wise results. We also report the RMSD results for all methods across the three datasets. Note that the RMSD metric may be ambiguous when its value is too high, and it will be affected by the sequence length. As Fig. 2b shows, structRFM achieves RMSDs of 9.97Å on CASP15-natural and 5.32Å on RNA-Puzzles, better than AlphaFold3 (15.7Å on CASP15-natural and 7.89Å on RNA-Puzzles, respectively). On the CASP16 dataset, structRFM achieves an RMSD of 28.6 Å. Generally, structRFM outperforms other methods across all datasets in RMSD. To provide a comprehensive case-level analysis, we visualize all predicted 3D and extracted 2D structures from structRFM, together with native structures from CASP15 in Fig. 5 . Notably, structRFM benchmarks an ϵ RMSD of 1.26 on the 8 natural targets of CASP15, much better than that (1.42) of the 4 synthetic targets. This performance gap is likely due to synthetic RNAs having longer sequences and denser long-range interactions, which are underrepresented in the training distribution. While structRFM achieves strong overall performance, the extracted 2D structures still exhibit some deficiencies, particularly in folding short local stems (green), occasionally forming larger, incorrect multi-loops (red). More visualization of CASP16 is displayed in Extended Data Fig. 3 . The performance of structRFM gets worse when the sequence length extends, indicating the current challenge in long sequence modeling. Download figure Open in new tab Fig. 5: Visualization of structRFM (Zfold) predicted structures and native structures on the CASP15 dataset. The ϵ RMSD, TM-score, GDT-TS of tertiary structures, together with the INF metric of secondary structures extracted from tertiary structures are annotated. In the secondary structure, stems are in green, multiple loops are in red, interior loops are in yellow, hairpin loops are in blue, while 5’ and 3’ unpaired regions are in orange. Generally, structRFM performs better on natural targets than synthetic targets that consist of dense long-distance interactions, but sometimes fails to predict local short stems (in green) and result in large multiple loops (in red). Download figure Open in new tab Fig. 6: structRFM for function prediction. a, b Neural network architecture for splice site prediction, IRES identification, and RNA sequence classification. c Strategies for processing long RNA sequences, including truncation, non-overlapping sliding window, and overlapping sliding window. d The top-k accuracy of structRFM and structRFM − for splice site prediction (16,505 RNAs) on acceptor, donor, and the averaged performance. e, f The radar visualization in seven classification metrics of structRFM, ablation version structRFM − , and other methods for ncRNA classification on nRC dataset (2,600 RNAs) and IRES identification (1,164 RNAs). g The averaged prediction response of structRFM across nucleotides for splice site prediction. h F1 comparison of structRFM with other models on the detailed 13 ncRNA families of nRC dataset. structRFM and RNA-FM achieve state-of-the-art performances, leading ahead of other models by a huge gap. i , j PR curve and ROC curve for IRES identification (1,164 RNAs). k, l The max MSA similarity of the ncRNA and IRES test dataset against corresponding training dataset. The relationship between the prediction confidence (probability) and the sequence similarity of each test dataset against their training dataset, respectively. The line in blue is the linear regression fit. In addition, to assess the concordance between secondary structures predicted by the structure-fine-tuned model and those extracted from predicted tertiary structures, we calculate the INF metric for pairwise comparison. As shown in Response Fig. 4b , structRFM achieves averaged INF values 0.720, 0.630, and 0.597 across CASP15, CASP16, and RNA-Puzzles (maximum of the two training datasets - RNAStrAlign and bpRNA1m), respectively, indicating its strong generalizability and transferability. In contrast, the INF values of secondary structures extracted from the tertiary predictions are 0.822 (CASP15), 0.597 (CASP16), and 0.725 (RNA-Puzzles). Notably, there are gaps between the two training datasets, demonstrating that the fine-tuning dataset has a negligible impact on final performance for this task. Finally, we evaluate the geometric and topological properties of tertiary structures predicted by structRFM across the three datasets. Fig. 4c quantifies the distribution of nine entanglement categories [ 56 ] (D&D, D&L, D(D), D(L), D(S), L&L, L(D), L(L), L(S)) for each target. Notably, nearly all targets exhibit zero entanglement across most categories; among datasets, CASP16-predicted structures have relatively fewer entanglements than those from CASP15 and RNA-Puzzles. A small subset of targets from CASP15 and RNA-Puzzles show entanglements (D&D, D(D), D(L), L(D)), primarily due to incorrectly predicted dinucleotide steps. We also check the topological knots [ 57 ] for each target and found that no target exhibits such artifacts. Moreover, Fig. 4d quantifies the distribution of six stereochemical error types [ 58 ] (close, bond length, bond angle, planar, chirality, polymer). Almost all predicted structures contain bond angle and planar errors, while no target exhibits polymer or chirality errors. These results align with the visualized tertiary and extracted secondary structures in Fig. 5 and Fig. 3 . For example, structRFM fails to accurately predict the CASP15 target R1138, resulting in extensive long-range loop regions and about 50 errors each for planar and bond angles. Thus, deep learning-predicted structures may benefit from optimization with physical force fields post-processing to reduce these stereochemical artifacts. In summary, structRFM leverages structure-guided pre-training to enhance the robustness and reliability of RNA structural predictions, consistently improving the accuracy of both tertiary and secondary structure modeling. Challenges remain in modeling synthetic RNAs, long-range interactions, and local stem formation, presenting opportunities for future improvement. 2.5 structRFM transfers knowledge for splice site prediction Splicing is a key step in processing pre-mRNAs into mature mRNAs and is essential for gene expression and protein translation. The splice site prediction task is to classify each nucleotide of a pre-mRNA into one of three categories: donor site, acceptor site, or non-site. As shown in Fig. 6a , we construct the neural network that uses structRFM as the backbone to extract sequence features, followed by a simple linear layer that maps the concatenated input features to nucleotide-level splice site scores, where each category is represented as 0 for non-site, 1 for donor, and 2 for acceptor. We adopt SpliceAI dataset [ 3 ] to fine-tune and evaluate structRFM and structRFM − , which consists of 162,706 RNA sequences for training and 16,505 RNA sequences for testing, with all sequences in a uniform length of 100. Since these sequences are pre-mRNAs, substantially different from ncRNA sequences for pretraining structRFM, we enable trainable encoder of structRFM to make effective fine-tuning and knowledge transfer for splice site prediction. We follow the same configurations as BEACON [ 59 ] and adopt the top-k accuracy as measurement. As Fig. 6d demonstrates and BEACON [ 59 ] benchmarks, structRFM achieves a top-k accuracy of 0.42, surpassing most existing RNA language models, such as RNA-MSM (0.38), UTR-LM (0.37), RNA-FM (0.35), and falls a little behind the specialized SpliceBERT (0.45) that is tailored for splice site prediction, thus validating the efficiency of structRFM. Moreover, structRFM outperforms structRFM − for both donor and acceptor prediction with improvements of 1.09% and 13.5%, respectively, in accuracy, showing the robustness and versatility of structRFM and SgMLM strategy for transferring the learned sequential and structural knowledge from ncRNAs to mRNAs. Fig. 6g further demonstrates the averaged response of the model outputs across all sequences for donor, acceptor, and non-site, showing the minority and consistency of site occurrence compared to non-site. 2.6 structRFM generalizes well for IRES identification Internal ribosome entry site (IRES) is a special structure that mediates the 5’ cap-independent translation in circular RNAs, which produces a stable and low-free-energy structure and leads to promising applications in RNA therapy. Therefore, accurate identification of IRES is pivotal in developing RNA-centric therapeutics and biological innovations. Here we adopt structRFM to extract the classification feature F cls for IRES identification, illustrated in Fig. 6b . The classifier is a simple feed-forward network that takes the concatenated input features and outputs the distribution probability of each class. To deal with the imbalanced data, we further introduce oversampling and undersampling strategies to sample different training datasets and fine-tune models on these datasets. The final model is an ensemble of these high-precision and high-recall models. We evaluate structRFM, its variants and two state-of-the-art methods, i . e . IRES-finder [ 60 ], DeepIRES [ 61 ], on an IRES dataset [ 36 , 37 ] that contains 12,998 RNA sequences for training (3,949 positive and 9,034 negative samples) and 1,164 RNA sequences for testing (582 positive and 582 negative samples), with a uniform length of 174. As shown in Fig. 6i , j, and Supplementary Table 6, the precision-recall curve (PRC) and receiver operating characteristic curve (ROC) indicate that structRFM achieves the best performance, with an area under PRC (AUPRC) of 0.694 and an area under ROC (AUROC) of 0.713, surpassing structRFM − which obtains 0.642 in AUPRC and 0.684 in AUROC. This highlights the effectiveness of the SgMLM pretraining in capturing structural patterns relevant to IRES functionality. Furthermore, we visualize the seven classification metrics, i . e ., F1 score, recall, precision, accuracy, AUPRC, AUROC, Matthews correlation coefficient (MCC), in Fig. 6f and Extended Data Fig. 4a , for an overall comparison against other methods. As visualized in the radar plot, structRFM consistently outperforms all baselines methods across most metrics, particularly excelling against non-language model methods, namely IRES-finder and DeepIRES. For instance, structRFM benchmarks an F1 score of 0.701 with an improvement of 47.5% than IRESfinder (F1 = 0.475) and 146% over DeepIRES (F1 = 0.284). Fig. 6k shows the sequence similarity against the training set, which indicates no linear relationship (Pearson correlation: R 2 =0.0490% between the similarity and the classification confidence (probability)). 2.7 structRFM achieves highly accurate ncRNA classification As demonstrated in the previous subsection, structRFM has shown a remarkable ability for zero-shot homology classification without relying on labeled training data. To further assess its downstream adaptability, we evaluate structRFM on a supervised non-coding ncRNA classification task using the nRC dataset [ 38 ], which comprises 13 primary ncRNA classes. The model architecture also consists of two components: the structRFM backbone for extracting classification features, and a linear classifier for outputting the class probability distribution, where the predicted category is selected based on the maximum probability. All models are fine-tuned under the same configurations as RNAErnie for fair comparison. As demonstrated in Fig. 6e and Extended Data Fig. 4b , structRFM reaches the best performance and highest accuracy for ncRNA classification, achieving new benchmarks of F1 score=0.979, MCC=0.977, AUROC=0.999, and AUPRC=0.994, outperforming structRFM − by a huge gap, suggesting that SgMLM strategy in the pre-training and trainable encoder in the fine-tuning largely enhance the performance of structRFM by encoding and transferring structural information, respectively. RNABERT and RNA-MSM extract less representative features for classifying different categories of ncRNAs, obtaining an F1 score of 0.727 and 0.934, respectively. Notably, RNA-FM attains comparable performance with structRFM, regardless of its weaker performances for secondary structure prediction and splice site prediction than structRFM, which indicates it may capture classification-relevant features more effectively than sequential or structural features. More benchmark results of other methods are displayed in Supplementary Table 7, where RNAErnie [ 14 ] and tailored non-language model method ncRDense [ 62 ] are ahead of other methods, reaching an F1 score of 0.964 and 0.951, respectively. Furthermore, we visualize the F1 scores of the seven models across the thirteen detailed RNA categories. As Fig. 6h demonstrates, structRFM performs consistently accurate classifications on all RNA types, while structRFM − obtains lower F1 scores on rare categories (in contrast with rRNA that weights 66.3% of the pre-training dataset), such as miRNA, HACA-box, and scaRNA, showing the generalizability and transferability of the sequential and structural knowledge learned by structRFM through the structure-guided pre-training strategy. Extended Data Fig. 4c further displays the PRC and ROC curves, confirming robust classification across all RNA families. Fig. 6l shows the most sequences in the test dataset share low sequence similarity against the training set, resulting no statistically significant linear relationship (Pearson correlation: R 2 =0.540%, p value=0.01). 2.8 Evaluating structRFM on new RNA families and long sequences Validating the generalizability of deep learning-based methods represents a critical benchmark for their utility. To assess structRFM’s performance in secondary structure prediction for novel RNA families unseen during pre-training, given that our pre-training relies on RNAcentral v24.0 (released 2024-06-25), we curate a distinct validation set comprising 436 sequences from the newly released sequences from Rfam v15.0. We report INF alongside the maximum MSA similarity relative to the two fine-tuning datasets (RNAStrAlign and bpRNA1m). As shown in Fig. 4e , structRFM generalizes robustly to these previously unseen families, achieving INF scores of 0.648 (RNAStrAlign-fine-tuned model) and 0.698 (bpRNA1m-fine-tuned model), both state-of-the-art results for secondary structure prediction [ 25 ]. This strong performance is notable given the low sequence similarity between the test set and fine-tuning data (0.198 vs. RNAStrAlign; 0.491 vs. bpRNA1m). For the bpRNA1m dataset (comprising Rfam v12.3-14.10 sequences), we observe a weak linear correlation between INF and sequence similarity (R2 = 12.0%). Collectively, these results confirm structRFM’s robust generalizability to novel Rfam families. Modeling long RNA sequences remains an unresolved challenge in the field, with existing secondary structure prediction methods exhibiting a significant performance drop as sequence length increases ( Extended Data Fig. 1 ). To validate structRFM’s robustness and effectiveness for long sequences, we conduct lncRNA classification experiments on two datasets: lncRNA-Human (86,419 training and 21,605 test RNAs) and lncRNA-Mouse (41,961 training and 10,491 test RNAs) [ 63 ]. Owing to computational constraints, these sequences are filtered to a maximum length of 3,000 nt. This yields 72,768 training and 18,105 test sequences for lncRNA-Human, and 29,625 training and 7,401 test sequences for lncRNA-Mouse. As outlined in Fig. 6c , we implement three strategies to handle long RNA sequences: (1) Truncation: Sequences exceeding the model’s maximum input length (512 nt) are truncated. This strategy is only suitable for sequence-wise tasks and not applicable to nucleotide-level prediction tasks. (2) Non-overlapping sliding window: A fixed 512-nt window is slid across full-length sequences with no overlap. This approach is compatible with nucleotide-level tasks. (3) Overlapping sliding window: In contrast to strategy (2), the 512-nt window is slid with 50% overlap (half of the window width) to preserve contextual continuity across segments. Experimental results ( Fig. 4f ) show that the two sliding window strategies yield markedly higher performance than truncation, with a substantial margin of improvement. Notably, structRFM using the overlapping sliding window strategy achieved an F1 score of 0.959 for lncRNA-Human and 0.9455 for lncRNA-Mouse. These results demonstrate high accuracy for lncRNA classification and highlight structRFM’s strong application potential for modeling long RNA sequences. 3 Discussion In this paper, we introduce structRFM, a model designed to integrate structural information into RNA language modeling, enabling the joint pre-training of sequential and structural knowledge. This integration produces discriminative features that incorporate structural insights, enhancing both structural and functional predictions. structRFM takes an important step towards explicitly exploring multimodal structural data without introducing task-specific supervised objectives that may bias the model. Our experimental results demonstrate the model’s effectiveness, robustness, and transferability across diverse downstream tasks, achieving top-tier performance in zero-shot prediction, tertiary structure prediction, and splice site prediction. Additionally, structRFM sets new benchmarks in secondary structure prediction, internal ribosome entry site (IRES) identification, and non-coding RNA (ncRNA) classification. Evaluations of structRFM on new RNA families and long sequences further validate its generalizability for unseen data. Notably, the multimodal 21-million-sequence-structure dataset we constructed, along with our structure-guided pre-training strategy, can be adapted to other biological language models, improving the understanding of hierarchical structures and the relationship between sequence and structure. The constructed large-scale multimodal dataset, code, and pre-trained structRFM are publicly available and fully open-source. Despite these achievements, several avenues for improvement remain. First, for long sequences, structRFM employs an overlapping sliding window strategy to enable modeling, which has been validated as effective for lncRNA classification with extremely high accuracy. However, this approach fails to capture long-range base interactions and global sequence relationships, potentially limiting its performance in structure-related tasks. A direct solution is to pre-train and fine-tune structRFM on longer RNA sequences, but this incurs substantial computational costs without guaranteeing significant performance gains. Second, while the MUSES strategy integrates three RNA secondary structure predictors to ensemble multi-source annotations, it focuses solely on canonical base pairs. Future work could extend this ensemble framework to capture the dynamic nature of RNA conformations, incorporating non-canonical base pairs, stacking interactions, and motif-level structural knowledge. Third, although the structure-guided pre-training strategy has demonstrated effectiveness in extracting structure-encoded features and strong generalizability across downstream tasks, its pre-training efficiency and scalability remain unvalidated. Fourth, as shown in Supplementary Table 8, structRFM fine-tuned secondary structure prediction model outperforms BPfold, ViennaRNA RNAfold, and CONTRAfold, showing the possibility of exploring a self-distillation framework that leverages structRFM itself to iteratively guide subsequent rounds of pre-training, which is promising for further improving model performance. Looking ahead, an exciting direction is the development of multimodal language models for biological research. structRFM employs structure-guided pre-training by implicitly integrating base-pairing interactions within a single backbone encoder, jointly extracting sequential and structural features. A promising extension would be to explicitly integrate multimodal structural data by using separate encoders for each modality, enabling contrastive sequence-structure pre-training. Such an approach would align RNA sequence and structural data through large-scale contrastive learning, enhancing generalization across downstream tasks. Contrastive learning in this context could facilitate robust predictions of RNA structure, function, and interactions, even in data-scarce conditions. For efficient modeling of long sequences, replacing the transformer [ 64 ] architecture (quadratic time complexity) with recently developed frameworks such as Mamba [ 65 ] (linear time complexity) holds great promise. Furthermore, another emerging direction is to build unified frameworks that integrate diverse biomolecular modalities. For instance, AlphaFold3 [ 35 ] expands structural prediction to proteins, nucleic acids, and small molecules within a single architecture, while models such as evo2 [ 6 ] and LucaOne [ 5 ] aim to unify molecular language modeling across biomolecular types. These efforts reflect a growing trend toward general-purpose, multimodal foundation models capable of capturing the complexity of biological systems. We anticipate that the convergence of contrastive learning, structure-guided pre-training, and unified multi-modal modeling will catalyze the next wave of advances in computational biology, offering deeper interpretability and broader applications. 4 Methods 4.1 Overview of structRFM In this work, we present structRFM ( Fig. 1a ), an RNA foundation model that is pre-trained on multi-modal non-coding RNA sequences together with corresponding secondary structures for general-purpose downstream adaptation. We first construct the pre-training sequence-structure dataset by filtering RNA sequences from the RNAcentral dataset [ 46 ] and collecting ensemble of secondary structures. Next, as illustrated in Fig. 1b , we design a structure-guided masked language modeling (SgMLM) strategy to integrate RNA secondary structures into RNA sequence modeling when pre-training task-agnostically. After structRFM is well pre-trained, it is capable of extracting representative structural feature of an input RNA sequence to further play a vital role in various downstream tasks, including structure and function tasks, such as secondary structure prediction, tertiary structure prediction, non-coding RNA classification, IRES identification in circular RNAs, and splice site prediction. 4.2 Pre-training sequence-structure dataset construction Starting with the largest non-coding RNA dataset RNAcentral version 24.0 [ 46 ], similar to existing RNA language models [ 12 , 14 ], we filter 21,477,078 RNA sequences from the active part of the RNAcentral dataset with a maximum sequence length of 512, which is different from some methods [ 15 , 66 ] that collect RNA sequences from both active part and inactive part that are no longer maintained. After collecting RNA sequences, we apply the MUSES strategy (described below) to ensemble RNA secondary structures from three different methodological paradigms. Based on these vast sequence-structure pairs, we randomly split them into training and test datasets, with the number of samples being 20,403,225 and 1,073,855, respectively. This multimodal dataset consists of a large number of RNA sequences and RNA secondary structures, relieving the shortage of structural data in a computational way at the secondary structure level and facilitating the development of multi-modal structure-enhanced RNA foundation models. Statistics of sequence length, family number, base pair interactions, and confidence index are visualized in Extended Data Fig. 5 . 4.3 Multi-source ensemble of secondary structures (MUSES) Relying on a single secondary structure predictor risks imprinting its biases onto the foundation model. RNA folding is inherently defined by rugged energy landscapes with numerous metastable states, strong dependence on environmental conditions, and frequent pseudoknots that are rarely captured by standard predictors. Representing each sequence by a sole deterministic structure therefore under-represents the full conformational diversity of RNA. To address this limitation, we introduce an ensemble framework termed MUSES ( Multi-source ensemble of secondary structures ). Instead of relying on single predictor, MUSES integrates hypotheses from diverse sources, including ViennaRNA RNAfold [ 22 , 23 ], CONTRAfold [ 24 ], and BPfold [ 25 ], synergizing three distinct methodological paradigms. First, ViennaRNA RNAfold provides a thermodynamics-based prediction using its well-established minimum free energy model. Second, CONTRAfold contributes a statistically learned perspective, leveraging conditional random fields trained on structural databases to identify probable motifs. Finally, BPfold introduces a physical-prior-integrated deep learning approach, which is robust and accurate by means of high-quality and full-coverage canonical base pair motif library, for forecasting secondary structure with confidence index to reveal the potential interactions between canonical base pairs, including G-C, C-G, A-U, U-A, G-U, and U-G. The consensus structures are derived through a weighted integration of these individual predictions, prioritizing base pairs with high accuracy across multiple frameworks. This tripartite approach mitigates the inherent biases and limitations of any single method, ensuring a more accurate and reliable structural model for further applications. Specifically, each method produces a pairwise secondary structure matrix M i ∈ [0, 1 ] L × L , where M i ( j, k ) denotes the base pair probability (∈ { [0, 1 ]}) or pairing status (∈ {0, 1}). The deterministic dot-bracket outputs of BPfold are converted into calibrated binary matrices, while probabilistic predictors such as ViennaRNA RNAfold and CONTRAfold directly yield base pair probability. Formally, the ensemble base pair probability matrix is defined as: where w i reflects method-specific confidence (e.g., higher weight for pseudoknot-aware or thermodynamically consistent predictors). From M ens , a consensus secondary structure is derived using a maximum expected accuracy (MEA) decoder: where 𝒮 is the set of valid secondary structures under RNA pairing constraints. This yields a consensus structure that reflects agreements across predictors while maintaining conformational diversity. Extended Data Fig. 1a presents our evaluation of MUSES across three datasets, with varying weights from 1 to 10 assigned to its three constituent methods. Ternary simplex plots illustrate performance variations across datasets and weight combinations, highlighting the limitations of relying on a single predictor. The final weights used for constructing the pre-training dataset is set as 4:1:7 according to the best MUSES performances across the three datasets. 4.4 Structure-guided masked language modeling strategy To improve the ability of learning base pairing interactions and capturing structural patterns for RNA language model, we elaborately design a structure-guided masked language modeling strategy (SgMLM) to replace the standard masked language modeling for pre-training structRFM with inspirations from base pair interactions. By doing so, we can leverage the structure-level knowledge and extract discriminative representations of RNA sequences to serve the downstream tasks. Specifically, given an RNA sequence X = { x 1 , x 2 , …, x L } ∈ {A, U, G, C, N} L , the standard masked language modeling strategy generates a random mask M r ∈ {0, 1} L on this RNA sequence in a Bernoulli distribution with a probability of 15%, where M r ( i ) = 1 indicates the token of base i is masked as “[MASK]” and M r ( i ) = 0 indicates the token of base i is kept the same as original. Based on this, we introduce the canonical base pairs of the RNA secondary structure into this procedure. The corresponding secondary structure of RNA sequence with a length of L can be represented in format of a binary adjacent matrix C ∈ {0, 1} L × L , where C ( i, j ) = 1 indicates that bases i and j form a base pair and C ( i, j ) = 0 indicates that bases i and j do not form a base pair. As Fig. 1b shows, we can obtain a secondary structure mask M s ∈ {0, 1} L from the secondary structure where M s ( i ) = 1 indicates that base i is a part of a base pair and vice versa. The structure-guided masking strategy is to mask RNA bases with more paired interactions instead of isolated bases. Therefore, we apply an operation of matching canonical base pairs, denoted as “PairMatch”, to make up for these isolated bases that have formed base pairing interactions in the random mask M r . The final structured-guided mask M f comprises two parts: 1. “PairMatch” of the “logical and” of random mask M r and secondary structure mask M s , 2. the “logical and” of random mask M r and the “logical not” of secondary structure mask M s , which can be formulated as follows: In the whole training process, to conduct data augmentation and enhance the diversity of training data, among the structure-guided masked tokens, we replace the “[mask]” with random tokens at 10% of the time and keep the masked input tokens unchanged at another 10% of the time. Furthermore, to enhance the pre-training stability and enable structRFM to learn both token-wise and pair-wise knowledge, we introduce a warm-up mechanism by gradually increasing the ratio r of structure-guided masking against standard masking from 0 to 0.5, which can be formulated as follows: where T is the total training steps where t is current training steps. Therefore, the language model is finally pre-trained with sequence-level masking and structure-level masking, jointly modeling the relationship of RNA sequences with base pairing interactions, resulting in enhanced capability of understanding RNA structural patterns. During masked language modeling, the proxy task is to recover the masked tokens as the original sequence base tokens with a cross-entropy loss. By applying structure-guided masking, structRFM can infer the masked matched base pairs in the unmasked unpaired surrounding base loops, intrinsically capturing the base pair interactions according to input RNA sequences, which is beneficial for structural prediction. 4.5 Pre-training structRFM with secondary structures In the pre-training process, structRFM makes use of the constructed sequence-structure dataset. For an input RNA sequence, structRFM generates a structured-guided mask for it according to its secondary structure, tokenizes the input sequence, and embeds it as feature vectors. Consequently, structRFM utilizes transformer [ 64 ] as its backbone, effectively models the long-range relationship of RNA sequences, and aims at learning the sequential and structural representations by the proxy task of recovering the masked tokens. Specifically, given an RNA sequence X = { x 1 , x 2 , …, x L } ∈ {A,U,G,C,N} L that is composed of “A”, “U”, “G”, and “C” base nucleotides (other unknown base is denoted as “N”), after applying structure-guided masking, the masked tokens are replaced with “[mask]” token. Additionally, we add a “[CLS]” token and an “[SEP]” token at the beginning and end of the sequence, respectively, which is helpful for fine-tuning on classification tasks and separating multiple input RNA sequences in situations of dealing with RNA-RNA interaction and multiple chains. To keep the sequence length uniform in a batch, we pad it with “[PAD]” token at the end of the sequence. These tokens are further encoded as numbers and processed by an embedding layer to form the input sequence embedding F 0 ∈ ℝ L × D . As demonstrated in Fig. 1a , structRFM employs N consecutive transformer neural network blocks as its backbone to process the input sequence embedding. The transformer block consists of a multi-head self attention module (MHSA) and a feed-forward network (FFN), which is composed of linear layers and GELU activation. After the input feature is processed by the MHSA and FFN module, respectively, it is further processed by a residual operation and a LayerNorm neural network layer. The whole process can be formulated as follows: where i = 0, 1, 2, …, N − 1 denotes the i -th transformer block with learnable parameter θ i . 4.6 structRFM feature embeddings for multi-task inferences After pre-training, given an arbitrary RNA sequence of length L as input, structRFM is capable of outputting a representative embedding F in the shape of ℝ (1+ L )× D with structural information encoded, which consists of a classification feature and a sequence feature. Classification feature F cls ∈ ℝ 1× D is derived from the “[CLS]” token placed in the first row of the output embedding F , helpful for sequence-level classification tasks. Sequence feature F seq ∈ ℝ L × D encodes each nucleotide of RNA sequence, powerful for nucleotide-level classification and regression tasks. Moreover, we derive a matrix feature F mat ∈ ℝ L × L by averaging the attention maps from the 12 attention heads in the last transformer layer ( i = 11, j ∈ {0, 1, …, 11}). The resulting matrix encodes pairwise nucleotide-to-nucleotide relationships in a latent feature space and can be interpreted as a learned adjacency matrix, which is particularly important for modeling interactions between nucleotide pairs in structure prediction tasks. Formally, F mat is computed as follows: The above three kinds of feature embeddings, with discriminative structural information, are powerful representations as the transitional intermediates to be further processed by a simple multiple-layer perceptron or other downstream modules for various RNA downstream tasks, including structural and functional predictions with task-specific fine-tuning. For sequence-level classification tasks, such as non-coding RNA classification, the classification feature F cls is used. For sequence-level regression tasks, the sequence feature F seq is adopted by calculating its mean value in the sequence-wise dimension. Besides, the sequence feature F seq is intrinsically suitable for nucleotide-level regression tasks, such as splice site prediction. For RNA secondary structure and tertiary structure prediction, regrading the difficulties of modeling interactions between base pairs, we employ the matrix feature F mat for providing sufficient base-to-base prior knowledge. 4.7 Fine-tuning structRFM for downstream adaptation After pre-training structRFM with RNA sequences and corresponding secondary structures in a task-agnostic self-supervised manner, we are able to adapt it for various downstream tasks by zero-shot prediction or fine-tuning it with labeled data in a task-specific supervised manner. Here, we demonstrate the supervised fine-tuning strategies in terms of fine-tuning different parts of task-specific neural networks and the ways to construct neural architectures for various structural and functional tasks. Fine-tuning strategy The learnable parameters of structRFM for downstream task adaptations consist of two parts: the pre-trained backbone (including sequence embedding layers and feature encoder) and task-specific heads or modules, which naturally derives two common fine-tuning strategies for transferring the learned knowledge of structRFM. The first fine-tuning strategy is to freeze the pre-trained backbone and fine-tune the task-specific heads and modules, which costs few computational resources and training time. The other fine-tuning strategy is to fine-tune whole neural network including backbone and other parts, which is usually resource-consuming but more powerful. In our experiments, regrading the distribution shift of input RNA sequences, we take the secondary fine-tuning strategy and comparison the vanilla “structRFM” with its ablation variant “structRFM − ” that is pre-trained without base pairing knowledge and SgMLM strategy. Neural architecture for downstream tasks Given an input RNA sequence, structRFM, as a foundation model, processes and extracts sequence-related and encoded structural feature embeddings, which are further fed into the prediction head for customizing different output representations. Tasks such as RNA sequence classification only need a task-specific head - a fully connected layer while some tasks outputting complex dense predictions like secondary structure prediction and tertiary structure prediction need post-processing procedures and structural modules. Secondary structure prediction RNA secondary structure demonstrates base pair patterns and stem-loop regions, playing vital roles in modeling RNA tertiary structure and exploration of functional RNA motifs within non-coding genomic regions. Most existing methods predict the canonical base pairs ( i . e ., A-U, G-C, G-U) of a given input RNA sequence of length L . The secondary structure can be represented as an adjacent matrix in the shape of L L , in which each element stores the pairing relationship of base i and base j, i . e ., 0 for unpaired and 1 for paired. For predicting such dense connections in a scheme of nucleotide-wise classification, complex postprocess procedures such as symmetry operation, sharp loop, and overlap removal are applied [ 67 , 68 ], and thermodynamic priors are adopted [ 25 , 41 ] in previous methods. In our experiments, similar to RNAErnie [ 14 ], we adopt the light-weighted MXfold2 neural network [ 41 ] as the output structural module. Specifically, as demonstrated in Fig. 3a , structRFM extracts sequence feature from the input RNA sequence and then concatenates it with the one-hot encoding of the RNA sequence, forming the enhanced feature embeddings, which further utilizes the MXfold2 neural networks to process feature embeddings and computes four kinds of folding scores at base pair level: helix stacking, unpaired region, helix opening and helix closing. At last, a Zuker-style dynamic programming approach [ 69 ] is applied for postprocessing to predict the secondary structures. Tertiary structure prediction Regarding the widely-participant cellular activities such as protein translation and gene regulation, RNA tertiary structures are essential for understanding their functions in biological processes. Predicting RNA tertiary structure is still challenging due to its dynamic nature and lack of high-quality experimental data. More and more methods [ 7 , 8 ] take use of RNA language model to generate representative sequential feature as the input formation of tertiary structure prediction methods. In our experiments, as illustrated in Fig. 3a , we employ trRoset-taRNA [ 28 ] as the structure module to replace the direct secondary structure input with the matrix feature produced by structRFM, aiming to examine the effectiveness of the structural matrix feature generated by structRFM. The structural module takes multiple sequence alignments (MSA) and matrix feature as inputs and output 1D and 2D geometries, which is further refined by PyRosetta [ 70 ] to generate the final tertiary structure. Splice site prediction Pre-mRNA splicing is an essential step in eukaryotic gene expression. Introns (non-coding regions) must be accurately removed, and exons (coding regions) are merged together by the spliceosome complex to form mature mRNA, which is then translated into functional protein. Therefore, it is important to accurately detect the exon-intron boundaries - the splice sites. Splice site prediction is a nucleotide-wise classification task that aims to recognize whether each base is a donor site, an acceptor site, or non-site. As Fig. 6a shows, we simply add a linear layer as the prediction head to the structRFM backbone to construct the neural network architecture for splice site prediction. To compare with RNA language models benchmarked in BEACON [ 59 ], we follow the same configuration by passing the sequence feature of the first layer of structRFM to a linear layer to make nucleotide-wise predictions. IRES identification The internal ribosome entry site (IRES) mediates the cap-independent translation process in circular RNAs, widely found in viral genomes and eukaryotic cells, exhibiting high stability and tissue specificity. Therefore, accurate identification of IRES aids in developing RNA therapeutics and targeted cancer interventions. As illustrated in Fig. 6a , the neural network architecture for IRES identification is composed of the structRFM backbone and a simple classifier consisting of a linear fully connected layer. Regarding the sequence-wise classification task, we adopt the classification feature extracted from the RNA sequence by structRFM, and concatenate it with the one-hot encoding of the input sequence as the input feature for the classifier, making it neat and effective. Furthermore, to deal with the imbalanced data, we fine-tune different models on oversampling (IRES) dataset and undersampling (Non-IRES) dataset, respectively, and ensemble these models to make soft voting predictions. Sequence classification Non-coding RNAs play various roles and demonstrate diverse functions in gene regulation and cellular processes. Clear classification of ncRNA is meaningful for advancing biological research and developing RNA-based therapy, and improving disease diagnosis and treatment. In this downstream task, we simply use the classification feature provided by structRFM and a light-weighted classifier head linear fully connected layer as the neural network architecture to predict varieties of ncRNA categories, such as rRNA, tRNA, mRNA, and so on. 4.8 Experimental configurations At the pre-training stage, we pre-train structRFM for 1,627,466 steps with a batch size of 96 on two NVIDIA Tesla A100 80GB graphics processing units for approximately 400 hours. structRFM adopts the vanilla BERT architecture [ 71 ] with hidden dimension being 768 and the number of transformer layers being 12, and is pre-trained on our constructed sequence-structure dataset with a maximum sequence length of 512, which is derived from RNAcentral [ 46 ] and RNA secondary structure prediction method - BPfold [ 25 ]. We utilize PyTorch [ 72 ] framework to implement structRFM, and the optimizer is an AdamW optimizer with a learning rate of 2 × 10 −4 . We design the SgMLM strategy and utilize the cross-entropy loss to enable structRFM to capture sequential and structural patterns. We evaluate structRFM on two zero-shot assessments and five downstream fine-tuning tasks, namely: zero-shot homology classification, zero-shot secondary structure prediction, secondary structure prediction, tertiary structure prediction, splice site prediction, IRES identification, and ncRNA classification. The overall performances are displayed in Extended Data Table 1 . These experimental datasets include Rfam [ 73 , 74 ], PDB [ 50 ], ArchiveII [ 26 ], bpRNA-1m [ 27 ], CASP15 [ 29 ], RNA-Puzzles [ 52 , 75 ], Splice [ 3 ], IRES [ 36 , 37 ] and nRC [ 38 ], detailed in Supplementary Table 1. Data availability The pre-training sequence-structure dataset, pre-trained, and fine-tuned structRFM are publicly available at Zenodo [ 76 ]. The training and test datasets are publicly available as follows: RNAcentral [ 46 ]: https://ftp.ebi.ac.uk/pub/databases/RNAcentral/releases/24.0/ ;Rfam [ 73 , 74 ]: https://rfam.org/ ;ArchiveII [ 26 ]: https://rna.urmc.rochester.edu/publications.html ;RNAStrAlign [ 42 ]: https://github.com/mxfold/mxfold2/releases/tag/v0.1.0 ;bpRNA [ 27 ]: https://bprna.cgrb.oregonstate.edu/download.php#bpRNA ;PDB [ 50 , 51 ]: https://rcsb.org/ ;CASP15 [ 29 ]: https://predictioncenter.org/download_area/CASP15 ;CASP16 [ 30 ]: https://predictioncenter.org/download_area/CASP16 ;RNA-Puzzles [ 52 ]: https://www.rnapuzzles.org/puzzles/list/ ;SpliceAI [ 3 ]: https://github.com/Illumina/SpliceAI ; IRES [ 36 , 37 ]: https://bitbucket.org/alexeyg-com/irespredictor/src/v2/data/ ;nRC [ 38 ]: http://tblab.pa.icar.cnr.it/public/nRC/paper_dataset/ ;lncRNA-Human [ 63 ]: https://www.gencodegenes.org/human/ ;lncRNA-Mouse [ 63 ]: https://www.gencodegenes.org/mouse/ . Source data are provided with this paper. Code availability The source code of structRFM is publicly available at Zenodo [ 77 ] and GitHub ( https://github.com/heqin-zhu/structRFM ). Author contribution S.K.Z and P.X conceived the idea, supervised the study. H.Z. designed and implemented the framework, conducted pre-training, structure prediction, and ncRNA classification experiments, analyzed the results and wrote the paper. R.L. conducted zero-shot experiments, wrote the corresponding section, and revised the overall writing. A.C. analyzed the results of sequence similarity, the geometric and topological correctness of the predicted structures. H.C. processed the sequence similarity between training and test datasets, and visualized the CASP15 and CASP16 datasets. F.Z. conducted the IRES identification experiments. F.T. improved the overview of structRFM. T.Y. visualized the tertiary structures. X.L. designed the experiment of evaluating model on various sequence lengths. Y.G improved the t-SNE visualization. All authors read, contributed to the discussion, and approved the final paper. Competing interests The authors declare no competing interests. Additional information Supplementary information The supplementary for this paper is publicly available. Download figure Open in new tab Extended Data Fig. 1: MUSES strategy for RNA secondary structure annotation and benchmarking of existing prediction methods across variable sequence lengths. a Ternary simplex visualization of the MUSES performance landscape across weighted combination configurations. Each vertex corresponds to a single standalone method (BPfold, ViennaRNA or CONTRAfold), and interior points represent weighted ensemble models. We normalized weighting coefficients for proportional equivalence and applied linear interpolation over a dense grid to smooth the INF metric distribution. Color gradients (blue to red) represent INF values with contour lines for enhanced resolution, which intuitively illustrates how ensemble weighting modulates MUSES performance for robust secondary structure annotation and identifies optimal weight coefficient regimes with associated trade-offs. b Bench-mark evaluation of six established multi-source RNA secondary structure prediction methods [22–25, 41, 48, 49, 78] on the Rfam [ 73 , 74 ] and ArchiveII [ 26 ] datasets, stratified by three sequence length categories ( 1000 nt). c Linear extrapolation fitting of prediction performance for the six methods on both datasets. A substantial performance degradation is observed with increasing sequence length, highlighting the current challenge of modeling RNA secondary structures for long RNA sequences. Download figure Open in new tab Extended Data Fig. 2: Detailed comparison of structRFM, structRFM − , RiNALMo-Mega [ 13 ], trRosettaRNA [ 28 ], and AlphaFold3 [ 35 ] for tertiary structure prediction on CASP15-natural [ 29 ] (8 RNAs), CASP16 (13 RNAs) [ 30 ], and RNA-Puzzles [ 52 ] (20 RNAs) datasets. a Head-to-head comparison of structRFM with other models in ϵ RMSD [ 31 ], TM-score [ 32 ], GDT-TS [ 33 ], and INF [ 34 ] metrics. The size of the circle depends on the absolute difference value between the two methods. b Violin visualization of the RMSD performances. Download figure Open in new tab Extended Data Fig. 3: Visualization of structRFM (Zfold) predicted structures and native structures on the CASP16 dataset. The ϵ RMSD, TM-score, GDT-TS of tertiary structures, together with the INF metric of secondary structures extracted from tertiary structures are annotated. In the secondary structure, stems are in green, multiple loops are in red, interior loops are in yellow, hairpin loops are in blue, while 5’ and 3’ unpaired regions are in orange. Generally, structRFM performs better on viral and aptamer targets, while obtains worse performances on multimer and bacterial targets. Targets marked with blue circles have largely different sequences between the predicted CASP16 targets and the ground truth from PDB, which is due to the unmodeled region from the experimental structure. Download figure Open in new tab Extended Data Fig. 4: Detailed visualization of IRES identification and ncRNA classification. a Overall comparison of methods for IRES identification in seven metrics. b Overall comparison of methods for ncRNA classification in seven metrics. c The PR curve and ROC curve of the detailed 13 families in nRC dataset for ncRNA classification. Download figure Open in new tab Extended Data Fig. 5: Overall statistics of our constructed pre-training sequence-structure dataset. a Distribution of sequence lengths, with most RNA sequences centered around 250 nucleotides. b Number count of RNA families (categories). Distribution of RNA families. rRNA and tRNA dominate the dataset, accounting for 66.3% and 16.7%, respectively. c Composition of base-pairing interactions in secondary structures. Canonical base pairs represent approximately half of all interactions, including GC (26.5%), AU (14.4%), and GU (7.2%) pairs. d Distribution of confidence scores for predicted secondary structures. The majority exhibit high confidence, supporting the reliability of base-pairing annotations. View this table: View inline View popup Download powerpoint Extended Data Table 1: Overall performance of evaluating structRFM on zero-shot, structure, and function inference tasks. View this table: View inline View popup Download powerpoint Extended Data Table 2: Summary of used datasets. View this table: View inline View popup Download powerpoint Extended Data Table 3: Comparative overview of structRFM with eight RNA language models. Seq. is short for sequence and Stru. is short for secondary structure. The performances of overlap ratio (OR, lower is better) for zero-shot homology classification on Rfam [ 73 , 74 ] dataset are displayed. Acknowledgements We thank Dr. Zaixi Zhang (Postdoctoral researcher, Princeton University) for valuable comments and discussions that improved this work. This work is supported by National Natural Science Foundation of China (62271465 to S.K.Z., 32370581 to P.X.) and Suzhou Basic Research Program (SYG202338 to S.K.Z.). Funder Information Declared National Natural Science Foundation of China , 62271465 , 32370581 Suzhou Basic Research Program , SYG202338 Footnotes 1. Secondary structure annotation with MUSES 2. Validated generalizability on long sequences and new RNA families 3. Analysis of training-test set similarity 4. RNA structure analysis, add CASP16 https://github.com/heqin-zhu/structRFM https://doi.org/10.5281/zenodo.18272040 References [1]. ↵ Cooper , T. A. , Wan , L. & Dreyfuss , G. Rna and disease . Cell 136 , 777 – 793 ( 2009 ). OpenUrl CrossRef PubMed Web of Science [2]. ↵ Fan , X. , Yang , Y. , Chen , C. & Wang , Z. Pervasive translation of circular rnas driven by short ires-like elements . Nature Communications 13 , 3751 ( 2022 ). OpenUrl PubMed [3]. ↵ Jaganathan , K. et al. Predicting splicing from primary sequence with deep learning . Cell 176 , 535 – 548 ( 2019 ). OpenUrl CrossRef PubMed [4]. ↵ Ozsolak , F. & Milos , P. M. Rna sequencing: advances, challenges and opportu-nities . Nature reviews genetics 12 , 87 – 98 ( 2011 ). OpenUrl CrossRef PubMed Web of Science [5]. ↵ He , Y. et al. Generalized biological foundation model with unified nucleic acid and protein language . Nature Machine Intelligence 7 , 942 – 953 ( 2025 ). URL doi: 10.1038/s42256-025-01044-4 . OpenUrl CrossRef [6]. ↵ Brixi , G. et al. Genome modeling and design across all domains of life with evo 2 . BioRxiv 2025 – 02 ( 2025 ). [7]. ↵ Li , Y. , Feng , C. , Zhang , X. & Zhang , Y. Ab initio rna structure prediction with composite language model and denoised end-to-end learning . bioRxiv 2025–03 ( 2025 ). [8]. ↵ Shen , T. et al. Accurate rna 3d structure prediction using a language model-based deep learning approach . Nature Methods 1–12 ( 2024 ). [9]. ↵ Chu , Y. et al. A 5 utr language model for decoding untranslated regions of mrna and function predictions . Nature Machine Intelligence 6 , 449 – 460 ( 2024 ). OpenUrl PubMed [10]. ↵ Chen , K. et al. Self-supervised learning on millions of primary rna sequences from 72 vertebrates improves sequence-based rna splicing prediction . Briefings in bioinformatics 25 , bbae163 ( 2024 ). OpenUrl CrossRef PubMed [11]. ↵ Wang , H. , Zhang , Y. , Chen , J. , Zhan , J. & Zhou , Y. A comparative review of rna language models . arXiv preprint arxiv: 2505.09087 ( 2025 ). [12]. ↵ Chen , J. et al. Interpretable rna foundation model from unannotated data for highly accurate rna structure and function predictions . arXiv preprint arxiv: 2204.00300 ( 2022 ). [13]. ↵ Penić , R. J. , Vlašić , T. , Huber , R. G. , Wan , Y. & Šikić , M. Rinalmo: General-purpose rna language models can generalize well on structure prediction tasks . Nature Communications 16 , 5671 ( 2025 ). OpenUrl PubMed [14]. ↵ Wang , N. et al. Multi-purpose rna language modelling with motif-aware pre-training and type-guided fine-tuning . Nature Machine Intelligence 6 , 548 – 557 ( 2024 ). OpenUrl CrossRef [15]. ↵ Zou , S. et al. A large-scale foundation model for RNA function and structure prediction ( 2024 ). [16]. ↵ Yu , H. et al. An interpretable rna foundation model for exploring functional rna motifs in plants . Nature Machine Intelligence 6 , 1616 – 1625 ( 2024 ). OpenUrl PubMed [17]. ↵ Cao , X. , Zhang , Y. , Ding , Y. & Wan , Y. Identification of rna structures and their roles in rna functions . Nature Reviews Molecular Cell Biology 25 , 784 – 801 ( 2024 ). OpenUrl CrossRef PubMed [18]. ↵ Mortimer , S. A. , Kidwell , M. A. & Doudna , J. A. Insights into rna structure and function from genome-wide studies . Nature Reviews Genetics 15 , 469 – 479 ( 2014 ). OpenUrl CrossRef PubMed [19]. ↵ Akiyama , M. & Sakakibara , Y. Informative rna base embedding for rna structural alignment and clustering by deep representation learning . NAR genomics and bioinformatics 4 , qac012 ( 2022 ). OpenUrl [20]. ↵ Yin , W. et al. Ernie-rna: an rna language model with structure-enhanced representations . Nature Communications 16 , 10076 ( 2025 ). OpenUrl PubMed [21]. ↵ Al-Onaizan , Y. , Bansal , M. & Chen , Y. Yang , H. & Li , K. Al-Onaizan , Y. , Bansal , M. & Chen , Y. (eds) MP-RNA: unleashing multi-species RNA foundation model via calibrated secondary structure prediction . (eds Al-Onaizan , Y. , Bansal , M. & Chen , Y. ) Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16 , 2024 , 5278 – 5296 ( Association for Computational Linguistics , 2024 ). OpenUrl [22]. ↵ Hofacker , I. L. et al. Fast folding and comparison of rna secondary structures . Monatshefte fur chemie 125 , 167 – 167 ( 1994 ). OpenUrl CrossRef Web of Science [23]. ↵ Lorenz , R. et al. Viennarna package 2.0 . Algorithms for molecular biology 6 , 1 – 14 ( 2011 ). OpenUrl PubMed [24]. ↵ Do , C. B. , Woods , D. A. & Batzoglou , S. Contrafold: Rna secondary structure prediction without physics-based models . Bioinformatics 22 , e90 – e98 ( 2006 ). OpenUrl CrossRef PubMed Web of Science [25]. ↵ Zhu , H. et al. Deep generalizable prediction of RNA secondary structure via base pair motif energy . Nature Communications 16 , 5856 ( 2025 ). URL doi: 10.1038/s41467-025-60048-1 . OpenUrl CrossRef PubMed [26]. ↵ Sloma , M. F. & Mathews , D. H. Exact calculation of loop formation probability identifies folding motifs in rna secondary structures . RNA 22 , 1808 – 1818 ( 2016 ). OpenUrl Abstract / FREE Full Text [27]. ↵ Danaee , P. et al. bprna: large-scale automated annotation and analysis of rna secondary structure . Nucleic acids research 46 , 5381 – 5394 ( 2018 ). OpenUrl CrossRef PubMed [28]. ↵ Wang , W. et al. trrosettarna: automated prediction of rna 3d structure with transformer network . Nature communications 14 , 7266 ( 2023 ). OpenUrl PubMed [29]. ↵ Das , R. et al. Assessment of three-dimensional rna structure prediction in casp15 . Proteins: Structure, Function, and Bioinformatics 91 , 1747 – 1770 ( 2023 ). OpenUrl [30]. ↵ Kretsch , R. C. et al. Assessment of nucleic acid structure prediction in casp16 . Proteins: Structure, Function, and Bioinformatics ( 2025 ). [31]. ↵ Bottaro , S. , Di Palma , F. & Bussi , G. The role of nucleobase interactions in rna structure and dynamics . Nucleic acids research 42 , 13306 – 13314 ( 2014 ). OpenUrl CrossRef PubMed [32]. ↵ Xu , J. & Zhang , Y. How significant is a protein structure similarity with tm-score= 0.5? Bioinformatics 26 , 889 – 895 ( 2010 ). OpenUrl CrossRef PubMed Web of Science [33]. ↵ Zemla , A. Lga: a method for finding 3d similarities in protein structures . Nucleic acids research 31 , 3370 – 3374 ( 2003 ). OpenUrl CrossRef PubMed Web of Science [34]. ↵ Parisien , M. , Cruz , J. A. , Westhof , E. & Major , F. New metrics for comparing and assessing discrepancies between rna 3d structures and models . Rna 15 , 1875 – 1885 ( 2009 ). OpenUrl Abstract / FREE Full Text [35]. ↵ Abramson , J. et al. Accurate structure prediction of biomolecular interactions with alphafold 3 . Nature 630 , 493 – 500 ( 2024 ). OpenUrl CrossRef PubMed [36]. ↵ Chen , C.-K. et al. Structured elements drive extensive circular rna translation . Molecular Cell 81 , 4300 – 4318 .e13 ( 2021 ). OpenUrl CrossRef PubMed [37]. ↵ Weingarten-Gabbay , S. et al. Systematic discovery of cap-independent translation sequences in human and viral genomes . Science 351 , aad4939 ( 2016 ). OpenUrl Abstract / FREE Full Text [38]. ↵ Fiannaca , A. , La Rosa , M. , La Paglia , L. , Rizzo , R. & Urso , A. nrc: non-coding rna classifier based on structural features . BioData mining 10 , 1 – 18 ( 2017 ). OpenUrl PubMed [39]. ↵ Van der Maaten , L. & Hinton , G. Visualizing data using t-sne . Journal of machine learning research 9 , 2579 – 2605 ( 2008 ). OpenUrl [40]. ↵ Zhang , Y. et al. Multiple sequence alignment-based rna language model and its application to structural inference . Nucleic Acids Research 52 , e3 – e3 ( 2024 ). OpenUrl CrossRef PubMed [41]. ↵ Sato , K. , Akiyama , M. & Sakakibara , Y. Rna secondary structure prediction using deep learning with thermodynamic integration . Nature communications 12 , 941 ( 2021 ). OpenUrl PubMed [42]. ↵ Tan , Z. , Fu , Y. , Sharma , G. & Mathews , D. H. Turbofold ii: Rna structural alignment and secondary structure prediction informed by multiple homologs . Nucleic acids research 45 , 11570 – 11581 ( 2017 ). OpenUrl CrossRef PubMed [43]. ↵ Zakov , S. , Goldberg , Y. , Elhadad , M. & Ziv-Ukelson , M. Rich parameterization improves rna structure prediction . Journal of Computational Biology 18 , 1525 – 1542 ( 2011 ). OpenUrl CrossRef PubMed [44]. ↵ Zhang , C. , Zhang , Y. & Pyle , A. M. rmsa: a sequence search and alignment algorithm to improve rna structure modeling . Journal of Molecular Biology 435 , 167904 ( 2023 ). OpenUrl CrossRef PubMed [45]. ↵ Altschul , S. F. , Gish , W. , Miller , W. , Myers , E. W. & Lipman , D. J. Basic local alignment search tool . Journal of molecular biology 215 , 403 – 410 ( 1990 ). OpenUrl CrossRef PubMed Web of Science [46]. ↵ Sweeney , B. A. et al. Rnacentral 2021: secondary structure integration, improved sequence search and new member databases . Nucleic Acids Res . 49 , D212 – D220 ( 2021 ). OpenUrl CrossRef PubMed [47]. ↵ Coordinators , N. R. Database resources of the national center for biotechnology information . Nucleic acids research 41 , D8 – D20 ( 2012 ). OpenUrl PubMed Web of Science [48]. ↵ Singh , J. , Hanson , J. , Paliwal , K. & Zhou , Y. Rna secondary structure prediction using an ensemble of two-dimensional deep neural networks and transfer learning . Nature communications 10 , 5407 ( 2019 ). OpenUrl PubMed [49]. ↵ Singh , J. et al. Improved rna secondary structure and tertiary base-pairing prediction using evolutionary profile, mutational coupling and two-dimensional transfer learning . Bioinformatics 37 , 2589 – 2600 ( 2021 ). OpenUrl CrossRef PubMed [50]. ↵ Berman , H. M. et al. The protein data bank . Nucleic acids research 28 , 235 – 242 ( 2000 ). OpenUrl CrossRef PubMed Web of Science [51]. ↵ Rose , P. W. et al. The rcsb protein data bank: integrative view of protein, gene and 3d structural information . Nucleic acids research gkw1000 ( 2016 ). [52]. ↵ Cruz , J. A. et al. Rna-puzzles: a casp-like evaluation of rna three-dimensional structure prediction . Rna 18 , 610 – 625 ( 2012 ). OpenUrl Abstract / FREE Full Text [53]. ↵ Yang , H. et al. Tools for the automatic identification and classification of rna base pairs . Nucleic Acids Research 31 , 3450 – 3460 ( 2003 ). OpenUrl CrossRef PubMed Web of Science [54]. ↵ Sweeney , B. A. et al. R2dt is a framework for predicting and visualising rna secondary structure using templates . Nature Communications 12 , 3494 ( 2021 ). OpenUrl PubMed [55]. ↵ Kerpedjiev , P. , Hammer , S. & Hofacker , I. L. Forna (force-directed rna): Simple and effective online rna secondary structure diagrams . Bioinformatics 31 , 3377 – 3379 ( 2015 ). OpenUrl CrossRef PubMed [56]. ↵ Popenda , M. et al. Entanglements of structure elements revealed in rna 3d models . Nucleic acids research 49 , 9625 – 9632 ( 2021 ). OpenUrl CrossRef PubMed [57]. ↵ Gren , B. A. , Antczak , M. , Zok , T. , Sulkowska , J. I. & Szachniuk , M. Knotted artifacts in predicted 3d rna structures . PLOS Computational Biology 20 , e1011959 ( 2024 ). OpenUrl [58]. ↵ Carrascoza , F. , Antczak , M. , Miao , Z. , Westhof , E. & Szachniuk , M. Evaluation of the stereochemical quality of predicted rna 3d models in the rna-puzzles submissions . Rna 28 , 250 – 262 ( 2022 ). OpenUrl Abstract / FREE Full Text [59]. ↵ Ren , Y. et al. Beacon: Benchmark for comprehensive rna tasks and language models . Advances in Neural Information Processing Systems 37 , 92891 – 92921 ( 2024 ). OpenUrl [60]. ↵ Zhao , J. et al. Iresfinder: Identifying rna internal ribosome entry site in eukaryotic cell using framed k-mer features . Journal of genetics and genomics= Yi chuan xue bao 45 , 403 – 406 ( 2018 ). OpenUrl CrossRef [61]. ↵ Zhao , J. et al. Deepires: a hybrid deep learning model for accurate identification of internal ribosome entry sites in cellular and viral mrnas . Briefings in Bioinformatics 25 , bbae439 ( 2024 ). OpenUrl CrossRef PubMed [62]. ↵ Chantsalnyam , T. , Siraj , A. , Tayara , H. & Chong , K. T. ncrdense: a novel computational approach for classification of non-coding rna family by deep learning . Genomics 113 , 3030 – 3038 ( 2021 ). OpenUrl CrossRef PubMed [63]. ↵ Frankish , A. et al. Gencode 2021 . Nucleic acids research 49 , D916 – D923 ( 2021 ). OpenUrl CrossRef PubMed [64]. ↵ Guyon , I. et al. Vaswani , A. et al. Guyon , I. et al. (eds) Attention is all you need . (eds Guyon , I. et al. ) In 31th Conference on Neural Information Processing Systems , 5998 – 6008 (NeurIPS, 2017 ). [65]. ↵ Gu , A. & Dao , T. Mamba: Linear-time sequence modeling with selective state spaces . arXiv preprint arxiv: 2312.00752 ( 2023 ). [66]. ↵ Zhang , Z. et al. Rnagenesis: Foundation model for enhanced rna sequence generation and structural insights . bioRxiv 2024 – 12 ( 2024 ). [67]. ↵ Fu , L. et al. Ufold: fast and accurate rna secondary structure prediction with deep learning . Nucleic acids research 50 , e14 – e14 ( 2022 ). OpenUrl CrossRef PubMed [68]. ↵ Wallach , H. M. et al. Chen , X. , Li , Y. , Umarov , R. , Gao , X. & Song , L. Wallach , H.M. et al. (eds) RNA secondary structure prediction by learning unrolled algorithms . (eds Wallach , H. M. et al. ) 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30 , 2020 , 8024 – 8035 (OpenReview.net, 2020 ). URL https://openreview.net/forum?id=S1eALyrYDH . OpenUrl [69]. ↵ Zuker , M. & Stiegler , P. Optimal computer folding of large rna sequences using thermodynamics and auxiliary information . Nucleic acids research 9 , 133 – 148 ( 1981 ). OpenUrl CrossRef PubMed Web of Science [70]. ↵ Chaudhury , S. , Lyskov , S. & Gray , J. J. Pyrosetta: a script-based interface for implementing molecular modeling algorithms using rosetta . Bioinformatics 26 , 689 – 691 ( 2010 ). OpenUrl CrossRef PubMed Web of Science [71]. ↵ Burstein , J. , Doran , C. & Solorio , T. Devlin , J. , Chang , M. , Lee , K. & Toutanova , K. Burstein , J. , Doran , C. & Solorio , T. (eds) BERT: pre-training of deep bidirectional transformers for language understanding . (eds Burstein , J. , Doran , C. & Solorio , T. ) Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , 4171 – 4186 ( Association for Computational Linguistics , 2019 ). [72]. ↵ Wallach , H. M. et al. Paszke , A. et al. Wallach , H. M. et al. (eds) Pytorch: An imperative style, high-performance deep learning library . (eds Wallach , H. M. et al. ) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada , 8024 – 8035 ( 2019 ). URL https://proceedings.neurips.cc/paper/2019/hash/bdbca288fee7f92f2bfa9f7012727740-Abstract.html . [73]. ↵ Kalvari , I. et al. Non-coding rna analysis using the rfam database . Current protocols in bioinformatics 62 , e51 ( 2018 ). OpenUrl [74]. ↵ Kalvari , I. et al. Rfam 14: expanded coverage of metagenomic, viral and microrna families . Nucleic acids research 49 , D192 – D200 ( 2021 ). OpenUrl CrossRef PubMed [75]. ↵ Miao , Z. et al. Rna-puzzles round iv: 3d structure predictions of four ribozymes and two aptamers . Rna 26 , 982 – 995 ( 2020 ). OpenUrl Abstract / FREE Full Text [76]. ↵ Zhu , H. Data and checkpoints for structrfm ( 2025 ). URL doi: 10.5281/zenodo.16754363 . OpenUrl CrossRef [77]. ↵ Zhu , H. heqin-zhu/structrfm: v0.0.3 ( 2025 ). URL doi: 10.5281/zenodo.16754337 . OpenUrl CrossRef [78]. Huang , L. et al. Linearfold: linear-time approximate rna folding by 5’-to-3’dynamic programming and beam search . Bioinformatics 35 , i295 – i304 ( 2019 ). OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted May 14, 2026. Download PDF Supplementary Material Data/Code Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following A fully open structure-guided RNA foundation model for robust structural and functional inference Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share A fully open structure-guided RNA foundation model for robust structural and functional inference Heqin Zhu , Ruifeng Li , Ao Chang , Haobin Chen , Feng Zhang , Fenghe Tang , Tong Ye , Xin Li , Yunjie Gu , Peng Xiong , S. Kevin Zhou bioRxiv 2025.08.06.668731; doi: https://doi.org/10.1101/2025.08.06.668731 Share This Article: Copy Citation Tools A fully open structure-guided RNA foundation model for robust structural and functional inference Heqin Zhu , Ruifeng Li , Ao Chang , Haobin Chen , Feng Zhang , Fenghe Tang , Tong Ye , Xin Li , Yunjie Gu , Peng Xiong , S. Kevin Zhou bioRxiv 2025.08.06.668731; doi: https://doi.org/10.1101/2025.08.06.668731 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Bioinformatics Subject Areas All Articles Animal Behavior and Cognition (7618) Biochemistry (17636) Bioengineering (13860) Bioinformatics (41847) Biophysics (21401) Cancer Biology (18536) Cell Biology (25424) Clinical Trials (138) Developmental Biology (13353) Ecology (19860) Epidemiology (2067) Evolutionary Biology (24287) Genetics (15583) Genomics (22463) Immunology (17701) Microbiology (40300) Molecular Biology (17141) Neuroscience (88434) Paleontology (666) Pathology (2825) Pharmacology and Toxicology (4813) Physiology (7633) Plant Biology (15107) Scientific Communication and Education (2042) Synthetic Biology (4285) Systems Biology (9808) Zoology (2268)