Method for predicting three-dimensional structure of circular RNA

Abstract

A method for predicting the three-dimensional structure of a circular RNA. The method is an IsRNAcirc method for predicting the three-dimensional structure of a circular RNA on the basis of coarse-grained molecular dynamics simulation. A predicted three-dimensional structure model of a circular RNA is scored using two knowledge-based scoring functions, i.e., rsRNASP and DFIRE-RNA, so as to evaluate the prediction precision of IsRNAcirc for predicting the three-dimensional structure of the circular RNA. The prediction precision of IsRNAcirc for the three-dimensional structure of a circular RNA is superior to that of template-based 3dRNA.

Description

A method for predicting the three-dimensional structure of circular RNA Technical Field

[0001] The invention relates to the field of computer-aided biomolecular design, and in particular to a method for predicting the three-dimensional structure of circular RNA based on coarse-graining of molecular dynamics simulation. Background Art

[0002] The 2023 Nobel Prize in Physiology or Medicine was awarded to scientists developing an mRNA COVID-19 vaccine. Circular mRNA, considered "linear mRNA 2.0," is predicted by Nature magazine to be the next generation of blockbuster drugs due to its numerous advantages, including enhanced stability and low immunogenicity. Because circular mRNA is highly stable and resistant to degradation, it can persist in the body for extended periods and continuously produce the encoded protein (Chinese Invention Patent 2023110345623, a circular mRNA drug delivery system; Chinese Invention Patent 2022104221991, a circular mRNA vaccine development platform for viral cancers; Chinese Invention Patent 2022101147883, a circular RNA technology platform for developing tumor immunotherapy drugs; Chinese Invention Patent 2022101149554, a circular RNA molecule and its application; Chinese Invention Patent 2022101310669, a circular mRNA tumor immunotherapy drug for colorectal cancer). Compared to traditional protein therapies, circular mRNA can provide longer-lasting therapeutic effects. At the same time, the therapeutic protein expression function of circular mRNA is modular and programmable (Chinese Invention Patent 2023108282537, a digital twin platform for intelligent manufacturing of circular RNA drugs; Chinese Invention Patent 2023104818780, a tumor digital twin platform for programmable drugs; Chinese Invention Patent 202310571620X, an artificial intelligence workstation for identifying carcinogenic factors). Simply by replacing the protein-coding sequence, it can direct cells to produce different peptides, enzymes, antibodies, channels, and receptors. It is well known that structure is the basis of function. The unique three-dimensional conformation of circular RNA determines its molecular behavior. If we imagine that the modular assembly and programming of circular mRNA can be guided by predicting the three-dimensional structure of circular RNA (Chinese invention patent 2023108282537, a digital twin platform for intelligent manufacturing of circular RNA drugs; Chinese invention patent 2023104818780, a tumor digital twin platform for programmable drugs; Chinese invention patent 202310571620X, an artificial intelligence workstation for judging carcinogenic factors), then it will be possible to predict the function of proteins encoded by circular mRNA and reduce the trial-and-error costs of circular RNA drug development.

[0003] Currently, the methods used to predict the three-dimensional structure of linear RNA include: comparative genomics methods, machine learning-based methods, molecular dynamics simulation methods, and secondary structure-based methods (Nucleic Acids Research, 2016: gkw973; Structure, 2020, 28(8): 963-976; PLoS One, 2014, 9(9): e107504; Bioinformatics, 2022, 38(16): 4042-4043; Methods, 2016, 103: 120-127; Computational and Structural Biotechnology Journal, 2020, 18: 2416-2423; RNA, 2009, 15(2): 189-199; Bioinformatics, 2008, 24(17): 1951-1952; Nucleic Acids Research, 2016, 44(7): e63; Journal of Chemical Theory and Computation, 2018, 14(4): 2230-2239; Journal of Chemical Theory and Computation, 2021, 17(3): 1842-1857; The Journal of Physical Chemistry B, 2021, 125(43): 11907-11915). However, due to the special structure and folding mechanism of circular RNA, the above prediction algorithms show significant differences when processing different types of circular RNA, resulting in the current prediction of circular RNA three-dimensional structure has not yet reached the ideal level of accuracy and reliability. Therefore, it is very necessary to develop a circular RNA three-dimensional structure prediction method with higher accuracy and reliability. Technical issues

[0004] In response to the problems of insufficient prediction accuracy in the prior art and the overly simplistic solutions for predicting the three-dimensional structure of circular RNA, the present invention provides a method for predicting the three-dimensional structure of circular RNA based on coarse-graining molecular dynamics simulation. Technical Solutions

[0005] The present invention is achieved in this way.

[0006] A method for predicting the three-dimensional structure of a circular RNA comprises the following steps: a first step, obtaining sequence information of a circular RNA for which three-dimensional structure prediction is required; a second step, using cRNAsp12 (website: http: / / xxulab.org.cn / crnasp12 / ) to predict the circular RNA's secondary structure information, including free energy, represented by dotted brackets, and folding stability; a third step, using the circular RNA sequence in the first step and the secondary structure represented by dotted brackets with the lowest free energy obtained in the second step as input files, and using the RNAcomposer tool (website: http: / / rnacomposer.ibch.poznan.pl / ) to predict a three-dimensional initial structure model of the circular RNA; a fourth step, using the circular RNA sequence in the first step, the information represented by the secondary structure represented by the brackets obtained in the second step, and the three-dimensional initial structure model obtained in the third step as input files, and inputting them into IsRNAcirc; a fifth step, inputting IsRNAcirc (Iterative Simulated method for Predicting circular RNAs 3D) into the fourth step. The two ends of the initial three-dimensional structure model of the circular RNA in Structures were connected, and the unreasonable structures predicted by the template method were initially eliminated by increasing the simulation temperature. At the same time, the connection between the 3' and 5' ends was gradually strengthened during the simulated annealing process. The conformational ensemble generated during the simulated annealing process was then clustered. In the sixth step, the clustered results in the fifth step were subjected to replica exchange molecular dynamics simulations using 10 temperatures between 260 and 480 K. The top 10% of structures with the lowest potential energy in the collected snapshots were clustered based on the root mean square deviation (RMSD) between each pair. At the same time, the top five coarse-grained circular RNA three-dimensional structures were reconstructed by all-atom reconstruction for user visualization. In the seventh step, QRNAS (https: / / bmcstructbiol.biomedcentral.com / articles / 10.1186 / s12900-019-0103-1) was used to refine the three-dimensional structure of the circular RNA obtained in the sixth step. In the eighth step, two knowledge-based scoring functions, rsRNASP (https: / / www.sciencedirect.com / science / article / pii / S0006349521009851) and DFIRE-RNA (https: / / pubmed.ncbi.nlm.nih.gov / pubmed / DFIRE-RNA) were used.gov / 31638408 / ) to score the circular RNA three-dimensional structure model refined in the seventh step, and compare the prediction accuracy with the results predicted by the template-based 3dRNA prediction method (URL: https: / / www.sciencedirect.com / science / article / pii / S2001037020303706).

[0007] Furthermore, the three-dimensional structural model of circular RNA described in this invention utilizes a coarse-grained model. Because large RNA structural models require long simulation times, computational methods based on all-atom simulations are challenging. To address these challenges, numerous researchers have previously developed various coarse-grained models (Frontiers in Molecular Biosciences, 2021, 8). Coarse-grained methods combine several atoms into a single bead and use only a few beads to model nucleotides, or employ a more coarse-grained representation by treating helices and loops as vertices and edges. Because the number of atoms after coarse-graining is smaller, the energy landscape in coarse-grained models is smoother, allowing for faster global energy minima and computational speed. Coarse-grained models offer advantages in RNA structure prediction and folding simulations because they significantly improve the efficiency of sampling the conformational space, particularly for large RNA molecules. In recent years, numerous coarse-grained models using different nucleotide representations have been developed to predict RNA three-dimensional structures and study RNA folding behavior. Both the YUP model and the NAST model use a single bead to represent nucleotides. The former uses one bead to represent the nucleotide at atom P and uses a harmonic potential function to describe bond stretching, bond angle bending, and torsion angle distortion (Journal of Chemical Theory and Computation, 2006, 2(3): 529-540). The NAST model simplifies the nucleotide using only the C3' atom (RNA, 2009, 15(2): 189-199). The iFoldRNA model uses three beads to represent the nucleotide. Two beads are located at the center of mass of the phosphate group and the five-atom sugar ring. The third bead is located at the center of the six-atom ring of the nucleobase (Bioinformatics, 2008, 24(17): 1951-1952). The TOPRNA model uses three beads to describe the phosphate, sugar, and base parts of the nucleotide. In the oxRNA model, nucleotides are considered as rigid bodies with five interaction sites, including backbone, hydrogen bonding, cross-stacking, and 3′ and 5′ stacking interactions (The Journal of Physical Chemistry B, 2014, 118(10): 2615-2627). The SimRNA model reduces the 20-23 heavy atoms of nucleotides to five, namely the atoms P, C4, C2, N1, and C4′ of pyrimidines, and N9 and C6 of purines (Nucleic Acids Research, 2016, 44(7): e63).In the HiRE-RNA model, the coarse-grained representation of nucleotides consists of six or seven beads: five beads represent the backbone (atoms P, O5', C5', C4', C1'), and one bead (for pyrimidines) or two beads (for purines) are placed at the center of mass of the heavy atoms in the base ring (Journal of Chemical Theory and Computation, 2015, 11(7): 3510-3522; The Journal of Physical Chemistry B, 2013, 117(27): 8047-8060). Ernwin is a helix-based coarse-grained model that maps helices as reduced cylinders and loops as edges connecting helices or connecting two helices. The relative orientation and position of two helices are determined by the loop connecting them (RNA, 2015, 21(6): 1110-1121).

[0008] Furthermore, the IsRNAcirc described in the present invention is developed based on the IsRNA2 software for predicting the three-dimensional structure of linear RNA. However, IsRNA2, which predicts the three-dimensional structure of linear RNA, adopts a coarse-grained folding model. The main chain of the linear RNA is represented by two CG beads P and S located on the P and C4' atoms, which define the phosphate group and the ribose sugar ring, respectively. The IsRNA2 model uses three CG beads for purines and pyrimidines, and the Watson Crick edge, Hoogsteen edge and sugar edge of the base are fully preserved, thereby better explaining the non-canonical base pairing interactions in large linear RNAs. The CG beads of all bases are located at the center of mass of the aggregated heavy atoms (Journal of Chemical Theory and Computation, 2018, 14(4): 2230-2239; Journal of Chemical Theory and Computation, 2021, 17(3): 1842-1857; The Journal of Physical Chemistry B, 2021, 125(43): 11907-11915). Predicting the three-dimensional structure of linear RNA using IsRNA2 requires the sequence information of the linear RNA, the secondary structure of the linear RNA expressed in dot-bracket form, and an initial three-dimensional structure as the initial input files. Molecular dynamics simulations of the IsRNA2 model were performed in the modified LAMMPS software, with Langevin dynamics simulations performed with an integration time step of Δt = 1 fs. To predict or refine the three-dimensional structure of the linear RNA, 10 replica exchange molecular dynamics (REMD) simulations were performed at temperatures between 200 and 425 K to improve sampling efficiency in the three-dimensional conformational space. Each simulation lasted 50 nanoseconds and was repeated three times using different initial structures. After sufficient relaxation, structural snapshots were collected at 50 picosecond intervals for the final 25 nanoseconds of the simulation. To obtain the predicted structure, the top 10% of the collected snapshots (a total of 5000 snapshots) with the lowest potential energy were clustered based on their pairwise root mean square deviation (RMSD). The centroid structures of the top clusters (sorted by size) provided the predicted top 3D structure.

[0009] Furthermore, circular RNA is different from linear RNA and is a unique type of RNA. It has a covalently closed structure and lacks a 5' cap and a 3' poly (A) tail. There is a cyclization site at the head-to-tail junction of the circular RNA sequence, also known as the back-splicing junction (BSJ) region (Cell, 2019, 177 (4): 865-880). When using the IsRNAcirc described in the present invention to predict the three-dimensional structure of circular RNA: first, the sequence information of the circular RNA needs to be obtained, and then cRNAsp12 is used to predict the secondary structure information of the circular RNA in the form of dotted brackets. Among them, the cRNAsp12 server provides a user-friendly web interface to predict the secondary structure and folding stability of the circular RNA sequence (International Journal of Molecular Sciences, 2023, 24 (4): 3822). After the user enters the relevant parameters such as the sequence information of the circular RNA, the folding temperature and the number of output structures, the prediction is submitted. At the end of the prediction, a considerable number of secondary structures can be obtained according to the number of output structures set, and these secondary structures are represented in the form of dotted brackets. Among them, each secondary structure is not only followed by the free energy of the structure, but also these predicted structures are sorted from top to bottom according to the free energy from small to large.

[0010] Furthermore, the secondary structure represented by the lowest free energy dot bracket is used for the next step of three-dimensional initial structure prediction: the circRNA sequence and the lowest free energy secondary structure predicted by cRNAsp12 are used as input files, and the RNAcomposer tool is used to predict the 3D initial structure of the circRNA. RNAComposer is a user-friendly and freely available server that can be used to predict the 3D structure of RNAs up to 500 nucleotide residues. After inputting the circRNA sequence information and the secondary structure information obtained above, and setting the temperature and the number of 3D structure models to be output, the task is submitted for execution (Methods, 2016, 103:120-127). After the task is completed, the 3D structure information output by RNAcomposer is obtained.

[0011] Furthermore, the sequence of the circular RNA, the secondary structure information represented by the bracketed form of the circular RNA, and the three-dimensional structure model of the circular RNA predicted by RNAcomposer are used as input files and input into IsRNAcirc.

[0012] Furthermore, since the IsRNAcirc was developed based on the IsRNA2 software for predicting the three-dimensional structure of linear RNA, its prediction process for the three-dimensional structure of circular RNA has certain similarities with that of linear RNA (The Journal of Physical Chemistry B, 2021, 125(43): 11907-11915). However, due to the covalent closure of the 3' and 5' ends of circular RNA, there are significant structural differences with linear RNA (Cell, 2019, 177(4): 865-880). Therefore, the first step of IsRNAcirc's prediction of the three-dimensional structure of circular RNA is to connect the two ends of the input initial three-dimensional structure. In this step, the simulated annealing method is used to preliminarily eliminate the unreasonable structure predicted by the template method by increasing the simulation temperature, and gradually strengthen the connection between the 3' end and the 5' end during the simulated annealing process. The conformational ensemble in the simulated annealing process is then clustered.

[0013] Furthermore, the second step in predicting circular RNA 3D structures using IsRNAcirc is to predict or improve the 3D structure of the circular RNA. During the prediction process, the results from the first step were clustered using 10 temperatures between 260 and 480 K and subjected to replica exchange molecular dynamics (REMD) simulations to improve sampling efficiency in the 3D conformational space. Each simulation lasted 100 nanoseconds. To obtain the predicted structure, the top 10% of structures with the lowest potential energy from the collected snapshots were clustered based on their pairwise root mean square deviation (RMSD). Simultaneously, the top five coarse-grained circular RNA structures were subjected to all-atom reconstruction for user visualization.

[0014] Furthermore, since coarse-grained simulations are used to improve the efficiency of conformational space sampling during the prediction of circular RNA three-dimensional structures using IsRNAcirc, the simplification during the coarse-graining process is accompanied by the loss of fine structural details. The stereochemical characteristics of the predicted three-dimensional structures of circular RNAs are not perfect, and there may also be unnatural bond lengths or steric conflicts. In order to eliminate some of the problems caused by coarse-graining, such as steric hindrance and bond breakage, which may exist in the all-atom reconstruction process, QRNAS was used for structural refinement (BMC Structural Biology, 2019, 19(1)).

[0015] The above steps complete the process of predicting circular RNA 3D structures using IsRNAcirc. Finally, we obtain PDB files for five predicted circular RNAs. Furthermore, we can adjust the number of predicted circular RNA 3D structures by modifying the parameters in the config file.

[0016] Furthermore, using IsRNAcirc to predict the 3D structure of circular RNAs requires several input files, including the circRNA sequence information and the corresponding secondary structure (in dot-bracket format). It should be noted that the circRNA sequence and secondary structure information must be placed in the same file, as well as the starting 3D structure (in standard PDB format) and a configuration file containing the basic prediction parameters to be performed (e.g., simulation step size, clustering cutoff, number of predicted models output). The standard output of IsRNAcirc mainly includes the predicted 3D structural model of the circular RNA (in standard PDB format), trajectory files for replica exchange molecular dynamics simulations, and log files for circularization and simulation runs.

[0017] Furthermore, the three-dimensional structure of circular RNA in the dataset of Chen Lingling's research group (Cell, 2019, 177(4): 865-880) was selected as the test set. The test dataset contains 34 circular RNAs with lengths ranging from 161 to 435 nucleotides. The secondary structure of RNA can be divided into two major categories: helix and loop. According to the characteristics of RNA secondary structure, there are many types of circular structures, such as hairpin loops, internal loops, protrusion loops and multi-way junction loops. At the same time, based on the location of BSJ in the secondary structure of circular RNA predicted by cRNAsp12 (International Journal of Molecular Sciences, 2023, 24(4): 3822), these 34 circular RNAs were divided into four categories: helix-circular RNAs, hairpin-circular RNAs, internal-circular RNAs, and junction-circular RNAs. The number of these four categories is 11, 5, 13, and 5, respectively.

[0018] To verify the prediction effect of IsRNAcirc, the present invention predicted the three-dimensional structures of 34 circular RNAs. Among them, the secondary structures and initial three-dimensional structures of the 34 circular RNAs were predicted using cRNAsp12 and RNAcomposer (Methods, 2016, 103: 120-127), respectively. cRNAsp12 can predict multiple secondary structures based on the sequence of the circular RNA and sort them according to free energy, selecting the secondary structure with the lowest free energy as input to predict the initial three-dimensional structure (International Journal of Molecular Sciences, 2023, 24(4): 3822). In the user interface of RNAcomposer, the number of output predicted structures was set to 3, and the predicted structure file contained three prediction models of the initial three-dimensional structure. In addition, the root mean square deviation (RMSD) of the initial three-dimensional structure prediction of the circular RNA and the three-dimensional structure predicted by IsRNAcirc was calculated, and the number of local unreasonable structures in the initial three-dimensional structure predicted by RNAcomposer was counted. The results showed that the local unreasonable structures can be roughly divided into four categories, which can be named: I, II, III, and IV.

[0019] Furthermore, IsRNAcirc's prediction of the 3D structures of 11 circular RNAs within helix-circular RNAs revealed significantly fewer irrational structures than the initial 3D structure models. In particular, FCHO2 (268 nt) and KIAA0368 (435 nt) showed better predictions than similar circular RNAs, while DHX34 (218 nt) showed poorer predictions. These results suggest that IsRNAcirc is effective in predicting the 3D structures of helix-circular RNAs.

[0020] Furthermore, IsRNAcirc's prediction of the 3D structures of five circular RNAs in the hairpin-circular RNAs group revealed significantly fewer irrational structures than the initial 3D structure models. In particular, IsRNAcirc's predictions for LDLRAD3 (346 nt) and RELL1 (434 nt) were significantly better than those for similar circular RNAs. These results demonstrate the effectiveness of IsRNAcirc's 3D structure prediction for hairpin-circular RNAs.

[0021] Furthermore, IsRNAcirc's prediction of the 3D structures of 13 circular RNAs within internal-circular RNAs revealed significantly fewer irrational structures than the initial 3D structure model. In particular, IsRNAcirc achieved better prediction results for CNNB1 (378 nt) and FGFR1-2 (315 nt) than for similar circular RNAs. These results demonstrate the effectiveness of IsRNAcirc's 3D structure prediction for internal-circular RNAs.

[0022] Furthermore, IsRNAcirc's prediction of the 3D structures of five circular RNAs in junction-circular RNAs showed significantly fewer unreasonable structures overall compared to the initial 3D structure models. In particular, EPHB4 (362 nt) showed superior prediction results compared to similar circular RNAs. These results demonstrate the effectiveness of 3D structure prediction for junction-circular RNAs. It is worth noting that in some embodiments of the present invention, the number of 3D structure models predicted by IsRNAcirc and the initial structures used for comparison differed. In some embodiments of the present invention, using RNAcomposer as input to predict three structural models, IsRNAcirc typically outputs five structural models for each circular RNA. Furthermore, in some embodiments of the present invention, the dot-bracket secondary structure predicted for the FGFR1-1 3D structure was not the secondary structure with the lowest free energy. For various reasons, using the secondary structure with the lowest free energy was not sufficient to predict the initial 3D structure in some embodiments of the present invention.

[0023] Furthermore, in order to estimate the time it takes for IsRNAcirc to predict the three-dimensional structure of a circular RNA, in some embodiments of the present invention, the runtimes of 34 circular RNAs tested were given. These circular RNAs range in length from 161 to 435 nucleotides. The results show that the runtime for predicting the three-dimensional structure of a circular RNA increases linearly with the length of the circular RNA. In some embodiments of the present invention, the runtime for the shortest circular RNA is about 100 hours. In some embodiments of the present invention, for circular RNAs greater than 400 nt in length, the runtime is approximately within 300 hours.

[0024] Furthermore, due to the complexity of circular RNA structure and the scarcity of experimental structures, it is currently impossible to directly evaluate the prediction accuracy of IsRNAcirc provided by the present invention for circular RNA. However, in some embodiments of the present invention, a scoring function is used to evaluate the prediction accuracy. In some embodiments of the present invention, the three-dimensional structure predicted by the template-based method 3dRNA (Computational and Structural Biotechnology Journal, 2020, 18: 2416-2423; Journal of molecular biology, 2022, 434(11): 167452; Scientific Reports, 2012, 2(1)) and the three-dimensional structure predicted by IsRNAcirc described in the present invention are scored separately. From the overall results of the two scoring functions, the prediction effect of IsRNAcirc provided by the present invention on circular RNA is better than that of 3dRNA. Beneficial effects

[0025] The present invention provides a method for predicting the three-dimensional structure of circular RNA using a coarse-grained model based on molecular dynamics simulations. Due to the lack of experimental structures, it is impossible to directly assess the effective prediction accuracy of IsRNAcirc for the three-dimensional structure of real circular RNAs. However, comparing the prediction accuracy of the three-dimensional structures predicted by the template-based method 3dRNA with that of the IsRNAcirc of the present invention using two scoring functions, the IsRNAcirc of the present invention outperforms 3dRNA in predicting circular RNA. In other words, the present invention addresses the shortcomings of the existing state of the art in predicting the three-dimensional structure of circular RNAs, as well as the overly simplistic solutions for predicting the three-dimensional structure of circular RNAs. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1. Coarse-grained representation of RNA and the workflow for linear RNA 3D structure prediction using IsRNA2. The left side of Figure 1 (A) shows the all-atom model of RNA (PDB ID: 1Z5C). The center of Figure 1 (A) shows details of the CGMD (coarse-grained matrix) representation of four nucleotides: the RNA backbone is represented by two coarse-grained beads, P and S, located at the P and C4' atoms, defining the phosphate group and the ribose sugar ring, respectively. Purines and pyrimidines are represented using three coarse-grained beads, with each base's coarse-grained bead located at the center of mass of the heavy atom group represented. The right side of Figure 1 (A) shows the coarse-grained RNA model. Figure 1 (B) shows the workflow for linear RNA 3D structure prediction using IsRNA2. IsRNA2 takes the sequence information of the required linear RNA, the secondary structure information expressed in dot-bracket form, and the initial three-dimensional structure as input. IsRNA2 first samples the three-dimensional conformational space by running replica exchange molecular dynamics simulations. Secondly, it selects the top 10% structures with the lowest potential energy from the simulated trajectory to construct a conformational ensemble. Then, it clusters the structures with the lowest energy according to the set RMSD threshold. The coarse-grained structural model is then reconstructed into an all-atom structural model and then energy minimized. Finally, the top five predicted three-dimensional structural models are output.

[0027] Figure 2. IsRNAcirc's workflow for predicting circular RNA 3D structures. Figure 2 (A) obtains the circRNA sequence information and uses cRNAsp12 to predict the circRNA's secondary structure information (denoted by dotted brackets). Figure 2 (B) uses RNAcomposer to predict the initial circRNA 3D structure based on the circRNA sequence and the secondary structure with the lowest free energy predicted by cRNAsp12. The circRNA sequence information, the secondary structure information (denoted by dotted brackets), and the PDB file of the initial 3D structure model predicted by RNAcomposer are input into IsRNAcirc. Figure 2 (C) first circularizes the 3' and 5' ends of the circRNA, followed by simulated annealing and clustering to obtain the circularized RNA structure. Figure 2 (D) uses replica exchange molecular dynamics simulation, clustering, and all-atom reconstruction to obtain the top five circRNA 3D structures. Figure 2 (E) utilizes QRNAS to refine the structure to eliminate coarse-grained issues such as steric hindrance and bond breakage that may occur during the all-atom reconstruction. Part (F) in Figure 2 shows the three-dimensional structural models of the five circular RNAs after refinement.

[0028] Fig. 3 Root mean square deviation (RMSD) between the initial structures of four types of circular RNA and the predicted three-dimensional structures of IsRNAcirc.

[0029] Figure 4 shows the main types and number of local irrational structures in the initial 3D structure model of the input IsRNAcirc. Figure 4 (A), (B), (C), and (D) show the main types of local irrational structures in the initial 3D structure model of the input IsRNAcirc. Figure 4 (A): The initial 3D structure of the input IsRNAcirc was predicted using a template-based approach. Template-based approaches require a library of original RNA 3D structures. This type of local irrational structure is primarily caused by the lack of available structures in the template library during the 3D structure prediction process. Figure 4 (B): This type of local irrational structure is a tangle of the RNA backbone. Figure 4 (C): This type of local irrational structure occurs when the RNA backbone passes between two bases at other positions. Figure 4 (D): This type of local irrational structure occurs when the RNA backbone passes between two bases at other positions. Figure 4 (E): This type of local irrational structure is a statistical analysis of the number of the four main types of local irrational structures. Type 3 has the highest number of irrational structures, followed by types 1 and 2, with type 4 having the lowest number.

[0030] Figure 5 Prediction results for POLR2A (336 nt) (circRNA ID: hsa_circ_0000741) in the helix-circular RNA class. Figure 5 (A): Superposition of the initial 3D structure and the predicted 3D structure of IsRNAcirc, with the 3D structures represented in light gray and dark gray, respectively. Figure 5 (B): The BSJ of the circular RNA is located in the helix region of the secondary structure. Figure 5 (C): This type of locally unreasonable J structure is of the first type, arising from the lack of a natural RNA 3D template. The left side shows the 3' and 5' ends of the initial structure before circularization. Figure 5 (D): After predicting the circular RNA using IsRNAcirc, the locally unreasonable structure in the initial 3D model has been optimized. The right side shows the 3' and 5' ends of the RNA after circularization. Part (E) of Figure 5: IsRNAcirc was used to predict the three-dimensional structures of 11 circular RNAs in helix-circular RNAs. The overall number of local unreasonable structures was much smaller than that of the initial three-dimensional structure model. In particular, FCHO2 (268nt) and KIAA0368 (435nt) had better prediction results than similar circular RNAs, but the prediction effect for DHX34 (218nt) was poor.

[0031] Figure 6 Prediction results for FKBP8 (259 nt) (circRNA ID: hsa_circ_0000915) in hairpin-circular RNAs. Figure 6 (A): Superposition of the initial 3D structure and the predicted 3D structure of IsRNAcirc, with the 3D structures represented in light gray and dark gray, respectively. Figure 6 (B): The BSJ of the circular RNA is located in the hairpin loop region of the secondary structure. Figure 6 (C): This type of localized irrational structure is a tangle of the RNA backbone. The right side shows the 3' and 5' ends of the initial structure before circularization. Figure 6 (D): After predicting the circular RNA using IsRNAcirc, the localized irrational structure in the initial 3D model has been optimized. The top shows the 3' and 5' ends of the circularized RNA. Part (E) of Figure 6: The three-dimensional structure prediction results of five circular RNAs in the hairpin-circular RNAs using IsRNAcirc show that the overall number of unreasonable structures is much smaller than that of the initial three-dimensional structure model. In particular, LDLRAD3 (346nt) and RELL1 (434nt) have better prediction results than similar circular RNAs.

[0032] Figure 7 Prediction results for MBOTA2 (224 nt) (circRNA ID: hsa_circ_0007334) among internal-circular RNAs. Figure 7 (A): Superposition of the initial 3D structure and the predicted 3D structure of IsRNAcirc, indicated in light brown and light blue, respectively. Figure 7 (B): The BSJ of the circular RNA is located in the internal loop region of the secondary structure. Figure 7 (C): This type of local irrational structure is a sharp bend in the RNA backbone. The right side shows the 3' and 5' ends of the initial structure before circularization. Figure 7 (D): After using IsRNAcirc to predict the circular RNA, the local irrational structure in the initial 3D model has been optimized. The boxed image above shows the 3' and 5' ends of the circularized RNA. Part (E) of Figure 7 shows the prediction results of the three-dimensional structures of 13 circular RNAs among internal-circular RNAs using IsRNAcirc. The overall number of unreasonable structures is much smaller than that of the initial three-dimensional structure model. In particular, CNNB1 (378nt) and FGFR1-2 (315nt) have better prediction results than similar circular RNAs.

[0033] Figure 8: Prediction results for SLC22A23 (259 nt) (circRNA ID: hsa_circ_0075504) in junction-circular RNAs. Figure 8 (A): Superposition of the initial 3D structure and the predicted 3D structure of IsRNAcirc, indicated in light gray and dark gray, respectively. Figure 8 (B): The BSJ of the circular RNA is located in the junction region of the secondary structure. Figure 8 (C): This type of local irrational structure occurs when the RNA backbone passes between two bases at other locations. The framed image above shows the 3' and 5' ends of the initial structure before circularization; the two ends are far apart. Figure 8 (D): After predicting the circular RNA using IsRNAcirc, the local irrational structure in the initial 3D model has been optimized. The right side shows the 3' and 5' ends after circularization. Part (E) of Figure 8 shows the three-dimensional structure prediction results of five circular RNAs in junction-circular RNAs using IsRNAcirc. The results show that the overall number of unreasonable structures is much smaller than that of the initial three-dimensional structure model. In particular, EPHB4 (362nt) has a better prediction effect than similar circular RNAs.

[0034] Figure 9: Scoring results for the three-dimensional structures of circular RNAs predicted by the DFIRE-RNA scoring function for 3dRNA and IsRNAcirc. Figure 9 (A): Scoring results for helix-circular RNAs. With the exception of TBCD, IsRNAcirc scores outperformed 3dRNA. Figure 9 (B): Scoring results for helix-circular RNAs: IsRNAcirc outperformed 3dRNA, particularly EZH2. Figure 9 (C): Scoring results for internal-circular RNAs: IsRNAcirc scores outperformed 3dRNA overall, while IsRNAcirc scores for all three-dimensional structures of PHF21A and SDHAF2 were lower than those of 3dRNA. Figure 9 (D): Scoring results for junction-circular RNAs: With the exception of SNHG4, IsRNAcirc scores outperformed 3dRNA.

[0035] Figure 10 shows the results of scoring the three-dimensional structures of circular RNAs predicted using the rsRNASP scoring function for 3dRNA and isRNAcirc. Figure 10 (A) shows that the scoring results for helix-circular RNAs are similar to those using the DFIRE-RNA scoring function. With the exception of TBCD, isRNAcirc scores for this type of circular RNA outperformed those for 3dRNA. Figure 10 (B) shows that the scoring results for helix-circular RNAs show that isRNAcirc scores outperformed those for 3dRNAs, particularly EZH2. Figure 10 (C) shows that the scoring results for internal-circular RNAs show that isRNAcirc scores overall outperformed those for 3dRNA. However, isRNAcirc scores for all three-dimensional structures of MBOTA2 and SDHAF2 were lower than those for 3dRNA. Compared to the scores using the DFIRE-RNA scoring function, the scoring results using rsRNASP are more discrete. Part (D) of Figure 10 shows the scoring results of junction-circular RNAs. Except for SNHG4, the scoring results of this type of circular RNA, IsRNAcirc, are better than those of 3dRNA.

[0036] Figure 11. Runtime for predicting the three-dimensional structures of 34 circular RNAs using IsRNAcirc. Best Mode for Carrying Out the Invention

[0037] The first step is to obtain the sequence information of the circular RNA for which 3D structure prediction is required. Second, cRNAsp12 is used to predict the secondary structure of the circular RNA, represented by a dotted bracket (DBB) format, as shown in Figure 2 (A). The cRNAsp12 server provides a user-friendly web interface for predicting the secondary structure and folding stability of circular RNA sequences. After entering the circular RNA sequence information, parameters such as the folding temperature and the number of output structures are set and submitted. After the prediction is completed, a number of secondary structures are generated based on the number of output structures set. These secondary structures are represented by a dotted bracket (DBB) format, each followed by a free energy value. The predicted structures are sorted from top to bottom based on their free energy value. The secondary structure with the lowest free energy DBB is used for the next prediction step. Third, the RNAcomposer tool is used to predict the initial 3D structure of the circular RNA, using the circular RNA sequence from the first step and the secondary structure with the lowest free energy value predicted by cRNAsp12 in the second step as input files. RNAComposer is a user-friendly and freely available server for 3D structure prediction of RNAs up to 500 nucleotides. During prediction, the circRNA sequence information and previously obtained secondary structure information are input, and the temperature and number of output 3D structure models are set before submission. After the task completes, the 3D structure information output by RNAcomposer is obtained, as shown in Figure 2 (B). In the fourth step, the circRNA sequence from the first step, the bracketed secondary structure information from the second step, and the 3D structure model from the third step are input as files into IsRNAcirc. In the fifth step, the two ends of the initial 3D structure input from the fourth step are connected, as shown in Figure 2 (C). Using simulated annealing, the temperature is raised to initially eliminate unreasonable structures predicted by the template method. During the simulated annealing process, the connection between the 3' and 5' ends is gradually strengthened. The conformational ensembles generated during simulated annealing were then clustered. In the sixth step, replica exchange molecular dynamics (REMD) simulations were performed on the clustered results from the fourth step using 10 temperatures between 260 and 480 K to improve sampling efficiency in the three-dimensional conformational space. Each replica was simulated for 100 nanoseconds, as shown in Figure 2 (D). To obtain a predicted structure, the top 10% of structures with the lowest potential energy in the collected snapshots were clustered based on their pairwise root mean square deviation (RMSD). At the same time, all-atom reconstructions of the top five coarse-grained circular RNA structures were performed for user visualization. In the seventh step, QRNAS was used to refine the structure, as shown in Figure 2 (E). The above steps complete the prediction of the three-dimensional structure of circular RNA using IsRNAcirc.Finally, five predicted circular RNA PDB files will be obtained. The number of circular RNA three-dimensional structures generated by the final prediction can be adjusted by modifying the parameters in the config file. Modes for Carrying Out the Invention

[0038] To make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0039] The present embodiment provides a method for predicting the three-dimensional structure of a circular RNA, comprising the following steps: a first step, obtaining sequence information of a circular RNA for which three-dimensional structure prediction is required; a second step, using cRNAsp12 to predict the secondary structure information of the circular RNA, including free energy expressed in the form of dotted brackets, and folding stability; a third step, using the sequence of the circular RNA in the first step and the secondary structure expressed in the form of dotted brackets with the lowest free energy obtained in the second step as input files, and predicting the three-dimensional initial structural model of the circular RNA using the RNAcomposer tool; a fourth step, inputting the sequence of the circular RNA in the first step, the information of the secondary structure represented in the form of brackets obtained in the second step, and the three-dimensional initial structural model obtained in the third step as input files into IsRNAcirc; a fifth step, connecting the two ends of the initial three-dimensional structural model of the circular RNA input into IsRNAcirc in the fourth step, and preliminarily eliminating unreasonable structures predicted by the template method by increasing the simulation temperature, while gradually strengthening the connection between the 3' end and the 5' end during the simulated annealing process, and then clustering the conformational ensemble in the simulated annealing process; a sixth step, using 10 PCR products at 260-480 ℃ to generate the final product. K, a replica exchange molecular dynamics simulation is performed on the results obtained by clustering in the fifth step. The top 10% structures with the lowest potential energy in the collected snapshots are clustered according to the root mean square deviation (RMSD) between each pair. At the same time, the top five coarse-grained circular RNA three-dimensional structures are reconstructed by all atoms to facilitate user visualization. In the seventh step, QRNAS is used to refine the circular RNA three-dimensional structure obtained in the sixth step. In the eighth step, two knowledge-based scoring functions, rsRNASP and DFIRE-RNA, are used to score the circular RNA three-dimensional structure model refined in the seventh step, and the prediction accuracy is compared with the results predicted by the template-based 3dRNA prediction method.

[0040] Example 1 Coarse-grained model construction

[0041] Coarse-graining methods combine several atoms into a single bead and use only a few beads to model nucleotides, or use an even coarser-grained representation by treating helices and loops as vertices and edges. Due to the smaller number of atoms involved in coarse-graining, the energy landscape in coarse-grained models is smoother, reaching the global energy minimum more quickly, and resulting in faster computation. Coarse-grained models offer advantages in RNA structure prediction and folding simulations because they significantly improve the efficiency of sampling the conformational space, especially for large RNA molecules. In recent years, many different coarse-grained models of nucleotide representation have been developed to predict RNA three-dimensional structure and study RNA folding behavior. Both the YUP model and the NAST model use a single bead to represent nucleotides. The former uses a single bead to represent the nucleotide at the P atom and uses a harmonic potential function to describe bond stretching, bond angle bending, and torsion angle distortion. The NAST model simplifies nucleotides by using only the C3' atom. The iFoldRNA model uses three beads to represent nucleotides. Two beads are located at the center of mass of the phosphate group and the five-atom sugar ring. The third bead is located at the center of the six-atom ring of the nucleobase. The TOPRNA model uses three beads to describe the phosphate, sugar, and base portions of a nucleotide. In the oxRNA model, nucleotides are treated as rigid bodies with five interaction sites, including backbone, hydrogen bonding, cross-stacking, and 3' and 5' stacking interactions. The SimRNA model reduces the 20-23 heavy atoms of a nucleotide to five: atoms P, C4, C2, N1, and C4' for pyrimidines, and N9 and C6 for purines. In the HiRE-RNA model, the coarse-grained representation of a nucleotide consists of six or seven beads: five beads representing the backbone (atoms P, O5', C5', C4', and C1'), and one bead (for pyrimidines) or two beads (for purines) placed at the center of mass of the heavy atoms in the base ring. Ernwin is a helix-based coarse-grained model that maps helices as reduced cylinders and loops as edges connecting helices or two helices. The relative orientation and position of two helices is determined by the loop connecting them.

[0042] Example 2 IsRNA2 predicts the three-dimensional structure of linear RNA

[0043] IsRNA2 employs a coarse-grained folding model. The RNA backbone is represented by two CG beads, P and S, located at the P and C4' atoms, defining the phosphate group and the ribose sugar ring, respectively. The IsRNA2 model uses three CG beads for purines and pyrimidines (Figure 1, panel (A)). The Watson-Crick edges, Hoogsteen edges, and sugar edges of the bases are fully preserved, better accounting for non-canonical base-pairing interactions in large RNAs. The CG beads for all bases are located at the center of mass of the clustered heavy atoms. Detailed information can be found at: https: / / www.ncbi.nlm.nih.gov / pmc / articles / PMC9731381 / . Using IsRNA2 to predict the three-dimensional structure of linear RNA requires the RNA sequence, the RNA secondary structure represented in dot-bracket form, and an initial three-dimensional structure as initial input files. Molecular dynamics simulations of the IsRNA2 model were performed in the modified LAMMPS software, along with Langevin dynamics simulations with an integration time step of Δt = 1 fs. To predict or refine the three-dimensional structure of linear RNA, 10 replica exchange molecular dynamics (REMD) simulations were performed at temperatures between 200 and 425 K to improve sampling efficiency in the three-dimensional conformational space. Each simulation lasted 50 nanoseconds and was repeated three times using different initial structures. After sufficient relaxation, structural snapshots were collected at 50-picosecond intervals for the final 25 nanoseconds of the simulation. To obtain a predicted structure, the top 10% of the collected snapshots (5000 total) with the lowest potential energy were clustered based on their pairwise root mean square deviation (RMSD). The detailed process and selection of the RMSD threshold for clustering can be found at https: / / www.ncbi.nlm.nih.gov / pmc / articles / PMC9731381 / . The centroid structures of the top clusters (sorted by size) provide the predicted top 3D structures (Figure 1, panel (B)).

[0044] Example 3 IsRNAcirc predicts the three-dimensional structure of circular RNA

[0045] The first step is to obtain the sequence information of the circular RNA for which 3D structure prediction is required. Second, cRNAsp12 is used to predict the secondary structure of the circular RNA, represented by a dotted bracket (DBB) format, as shown in Figure 2 (A). The cRNAsp12 server provides a user-friendly web interface for predicting the secondary structure and folding stability of circular RNA sequences. After entering the circular RNA sequence information, parameters such as the folding temperature and the number of output structures are set and submitted. After the prediction is completed, a number of secondary structures are generated based on the number of output structures set. These secondary structures are represented by a dotted bracket (DBB) format, each followed by a free energy value. The predicted structures are sorted from top to bottom based on their free energy value. The secondary structure with the lowest free energy DBB is used for the next prediction step. Third, the RNAcomposer tool is used to predict the initial 3D structure of the circular RNA, using the circular RNA sequence from the first step and the secondary structure with the lowest free energy value predicted by cRNAsp12 in the second step as input files. RNAComposer is a user-friendly and freely available server for 3D structure prediction of RNAs up to 500 nucleotides. During prediction, the circRNA sequence information and previously obtained secondary structure information are input, and the temperature and number of output 3D structure models are set before submission. After the task completes, the 3D structure information output by RNAcomposer is obtained, as shown in Figure 2 (B). In the fourth step, the circRNA sequence from the first step, the bracketed secondary structure information from the second step, and the 3D structure model from the third step are input as files into IsRNAcirc. In the fifth step, the two ends of the initial 3D structure input from the fourth step are connected, as shown in Figure 2 (C). Using simulated annealing, the temperature is raised to initially eliminate unreasonable structures predicted by the template method. During the simulated annealing process, the connection between the 3' and 5' ends is gradually strengthened. The conformational ensembles generated during simulated annealing were then clustered. In the sixth step, replica exchange molecular dynamics (REMD) simulations were performed on the clustered results from the fourth step using 10 temperatures between 260 and 480 K to improve sampling efficiency in the three-dimensional conformational space. Each replica was simulated for 100 nanoseconds, as shown in Figure 2 (D). To obtain a predicted structure, the top 10% of structures with the lowest potential energy in the collected snapshots were clustered based on their pairwise root mean square deviation (RMSD). At the same time, all-atom reconstructions of the top five coarse-grained circular RNA structures were performed for user visualization. In the seventh step, QRNAS was used to refine the structure, as shown in Figure 2 (E). The above steps complete the prediction of the three-dimensional structure of circular RNA using IsRNAcirc.Finally, five predicted circular RNA PDB files will be obtained. The number of circular RNA three-dimensional structures generated by the final prediction can be adjusted by modifying the parameters in the config file.

[0046] Example 4 Input and output files for predicting the three-dimensional structure of circular RNA using IsRNAcirc

[0047] Using IsRNAcirc to predict the 3D structure of circular RNAs requires several input files, including the circRNA sequence information and the corresponding secondary structure (in dot-bracket format). It should be noted that the circRNA sequence and secondary structure information must be placed in the same file. The starting 3D structure (in standard PDB format) and the configuration files for the basic prediction parameters to be performed (e.g., simulation step size, clustering cutoff, number of predicted models output) are also included. The standard output of IsRNAcirc primarily includes the predicted 3D structural model of the circular RNA (in standard PDB format), trajectory files for replica exchange molecular dynamics simulations, and log files for the circularization and simulation runs.

[0048] Example 5 Selection of test set

[0049] The three-dimensional structure of circular RNA in the dataset of Chen Lingling's research group was selected. This dataset contains 34 circular RNAs with lengths ranging from 161 to 435 nucleotides, as shown in Table 1. The secondary structure of RNA can be divided into two major categories: helix and loop. According to the characteristics of RNA secondary structure, there are many types of circular structures, such as hairpin loops, internal loops, protruding loops and multi-way junction loops. Based on the location of BSJ in the secondary structure of circular RNA predicted by cRNAsp12, these 34 circular RNAs were divided into four categories: helix-circular RNAs, hairpin-circular RNAs, internal-circular RNAs, and junction-circular RNAs. The number of these four categories is 11, 5, 13, and 5 respectively;

[0050] .

[0051] Example 6 Verification of the Predictive Effect of IsRNAcirc

[0052] For the three-dimensional structures of the 34 circular RNAs in Example 5, cRNAsp12 and RNAcomposer were used to predict the secondary structures and initial three-dimensional structures of the 34 circular RNAs, respectively. cRNAsp12 can predict multiple secondary structures based on the circular RNA sequence and rank them according to free energy, selecting the secondary structure with the lowest free energy as input to predict the initial three-dimensional structure. In the RNAcomposer user interface, the number of output predicted structures is set to 3, and the predicted structure file contains prediction models for the three initial three-dimensional structures. The root mean square deviation (RMSD) of the initial three-dimensional structure and the IsRNAcirc predicted three-dimensional structure is calculated simultaneously, as shown in Figure 3. The number of local unreasonable structures in the initial three-dimensional structure predicted by RNAcomposer was counted. Based on the observed results, the local unreasonable structures were roughly divided into four categories: ASXL1, FKBPB, TBCD, and MBOTA2, and they were named: I, II, III, and IV, respectively. The types are shown in sections AD in Figure 4, and the corresponding numbers are shown in section E in Figure 4.

[0053] Example 7 Prediction Effect of IsRNAcirc on Helix-Circular RNAs

[0054] Figure 5 shows the prediction results for POLR2A (336 nt) (circRNA ID: hsa_circ_0000741) from the helix-circular RNA class using IsRNAcirc. Figure 5 (A) shows the initial three-dimensional structure and the predicted three-dimensional structure of IsRNAcirc, indicated in light gray and dark gray, respectively. Figure 5 (B) shows the helix region where the circRNA's BSJ is located within the secondary structure. Figure 5 (C) and (D) compare the local unreasonable structures in the initial three-dimensional model and the local structures of the 3D model predicted using IsRNAcirc. The predicted three-dimensional structures of 11 helix-circular RNAs using IsRNAcirc show significantly fewer unreasonable structures overall than those in the initial three-dimensional model. In particular, the prediction results for FCHO2 (268 nt) and KIAA0368 (435 nt) are better than for similar circRNAs. However, the prediction for DHX34 (218 nt) is poor, as shown in Figure 5 (E). The results show that IsRNAcirc has a good effect on predicting the three-dimensional structure of helix-circular RNAs.

[0055] Example 8 Prediction Effect of IsRNAcirc on Helix-Circular RNAs

[0056] The prediction results for FKBP8 (259 nt) (circRNA ID: hsa_circ_0000915) in the hairpin-circular RNAs using IsRNAcirc are shown in Figure 6. Figure 6 (A) shows the initial three-dimensional structure and the predicted three-dimensional structure of IsRNAcirc, indicated in light gray and dark gray, respectively. Figure 6 (B) shows the hairpin loop region where the circRNA's BSJ is located within the secondary structure. Figure 6 (C) and (D) compare the local unreasonable structures in the initial three-dimensional model and the local structures of the 3D model predicted using IsRNAcirc. The predicted three-dimensional structures of five circRNAs in the hairpin-circular RNAs using IsRNAcirc show significantly fewer unreasonable structures overall than those in the initial three-dimensional model. In particular, LDLRAD3 (346 nt) and RELL1 (434 nt) show better prediction results than other circRNAs of the same type, as shown in Figure 6 (E). These results demonstrate that the prediction of the three-dimensional structure of hairpin-circular RNAs is effective.

[0057] Example 9 Prediction effect of IsRNAcirc on internal-circular RNAs

[0058] The prediction results for MBOTA2 (224 nt) (circRNA ID: hsa_circ_0007334) from the internal-circular RNA class using IsRNAcirc are shown in Figure 7. Figure 7 (A) shows the initial three-dimensional structure and the predicted three-dimensional structure of IsRNAcirc, represented in light brown and light blue, respectively. Figure 7 (B) shows the internal loop region where the circRNA's BSJ is located within the secondary structure. Figure 7 (C) and (D) compare the local unreasonable structures in the initial three-dimensional model and the local structures of the 3D model predicted using IsRNAcirc. The predicted three-dimensional structures of 13 internal-circular RNAs using IsRNAcirc show significantly fewer unreasonable structures overall than those in the initial three-dimensional model. In particular, CNNB1 (378 nt) and FGFR1-2 (315 nt) show better prediction results than other circRNAs of the same type, as shown in Figure 7 (E). These results demonstrate that the prediction of the three-dimensional structure of internal-circular RNAs is effective.

[0059] Example 10 Prediction Effect of IsRNAcirc on Junction-circular RNAs

[0060] The prediction results for SLC22A23 (259 nt) (circRNA ID: hsa_circ_0075504) in the junction-circular RNAs using IsRNAcirc are shown in Figure 8. Figure 8 (A) shows the initial three-dimensional structure and the predicted three-dimensional structure of IsRNAcirc, indicated in light gray and dark gray, respectively. Figure 8 (B) shows the junction region where the circRNA's BSJ is located within the secondary structure. Figure 8 (C) and (D) compare the localized unreasonable structures in the initial three-dimensional model and the local structures of the 3D model predicted using IsRNAcirc. The predicted three-dimensional structures of five circRNAs in the junction-circular RNAs using IsRNAcirc show significantly fewer unreasonable structures overall than those in the initial three-dimensional model. In particular, EPHB4 (362 nt) shows a significantly better prediction than other circRNAs of the same type, as shown in Figure 8 (E). These results demonstrate that the prediction of the three-dimensional structure of junction-circular RNAs is effective. It's worth noting that the number of 3D structural models predicted by IsRNAcirc and the initial structures used for comparison differed. Using RNAcomposer's three predicted structural models as input, IsRNAcirc outputted a maximum of five structural models for each circular RNA. Furthermore, the dot-bracket secondary structure predicted for the FGFR1-1 3D structure was not the lowest free energy secondary structure, and for some reason, the lowest free energy secondary structure could not be used to predict the initial 3D structure.

[0061] Example 11 Comparison of the prediction effects of IsRNAcirc and 3dRNA on the three-dimensional structure of circular RNA

[0062] Since there is currently no experimental three-dimensional structure of circular RNA, it is impossible to directly verify the performance of IsRNAcirc. Therefore, two knowledge-based scoring functions, rsRNASP and DFIRE-RNA, are used to score the three-dimensional structural model predicted by IsRNAcirc. The present invention uses the three-dimensional structures of the 34 circular RNAs described in Table 1 in Example 5 predicted by the template-based prediction method 3dRNA as a comparison. The three-dimensional structures predicted by the template-based method 3dRNA and the three-dimensional structures predicted by the IsRNAcirc of the present invention were scored separately. From the overall results of the two scoring functions, the prediction effect of IsRNAcirc of the present invention on circular RNA is better than 3dRNA, as shown in Figures 9 and 10.

[0063] Example 12 IsRNAcirc running time

[0064] To determine the time it takes for IsRNAcirc to predict the three-dimensional structure of circular RNAs, Figure 11 shows the runtime for the 34 circular RNAs listed in Table 1 described in Example 5. These circular RNAs range in length from 161 to 435 nucleotides. Figure 11 shows that the runtime for predicting the three-dimensional structure of circular RNAs increases linearly with circular RNA length. The runtime for the smallest circular RNA is approximately 100 hours. For circular RNAs longer than 400 nt, the runtime is approximately 300 hours. Industrial Applicability

[0065] Since there is currently no experimental three-dimensional structure of circular RNA, it is impossible to directly verify the performance of IsRNAcirc. Therefore, two knowledge-based scoring functions, rsRNASP and DFIRE-RNA, are used to score the three-dimensional structural model predicted by IsRNAcirc. The present invention uses the three-dimensional structures of the 34 circular RNAs described in Table 1 in Example 5 predicted by the template-based prediction method 3dRNA as a comparison. The three-dimensional structures predicted by the template-based method 3dRNA and the three-dimensional structures predicted by the IsRNAcirc of the present invention were scored separately. From the overall results of the two scoring functions, the prediction effect of IsRNAcirc of the present invention on circular RNA is better than 3dRNA, as shown in Figures 9 and 10. Sequence Listing Free Content

[0066] Unordered listing of the present invention.

Claims

1. A method for predicting the three-dimensional structure of circular RNA, characterized in that: The method comprises the following steps: The first step is to obtain the sequence information of the circular RNA whose three-dimensional structure needs to be predicted; In the second step, cRNAsp12 was used to predict the secondary structure information and folding stability of circular RNA, including free energy expressed in dot-bracket form; The third step is to use the sequence of the circular RNA in the first step and the secondary structure represented by the bracket with the lowest free energy point obtained in the second step as input files, and use the RNAcomposer tool to predict the three-dimensional initial structure model of the circular RNA; Step 4: input the circular RNA sequence in the first step, the secondary structure information represented by the brackets obtained in the second step, and the three-dimensional initial structure model obtained in the third step as input files into IsRNAcirc; The fifth step is to connect the two ends of the initial three-dimensional structure model of the circular RNA in IsRNAcirc input in the fourth step, and preliminarily eliminate the unreasonable structure predicted by the template method by increasing the simulation temperature, and gradually strengthen the connection between the 3' end and the 5' end during the simulated annealing process, and then cluster the conformational ensemble in the simulated annealing process; In the sixth step, 10 temperatures between 260 and 480 K were used to perform replica exchange molecular dynamics simulations on the clustered results in the fifth step. The top 10% structures with the lowest potential energy in the collected snapshots were clustered according to the root mean square deviation between each pair. At the same time, the top five coarse-grained circular RNA three-dimensional structures were reconstructed by all atoms to facilitate user visualization. Step 7, using QRNAS to refine the three-dimensional structure of the circular RNA obtained in the sixth step; In the eighth step, two knowledge-based scoring functions, rsRNASP and DFIRE-RNA, are used to score the circular RNA three-dimensional structure model predicted in the seventh step, and the prediction accuracy is compared with the results predicted by the template-based 3dRNA prediction method.

2. The method for predicting the three-dimensional structure of circular RNA according to claim 1, characterized in that: The coarse-grained model is used in the process of constructing the circular RNA three-dimensional structure model.

3. A method for predicting circular RNA IRES according to claim 1, characterized in that The IsRNAcirc was developed based on the IsRNA2 software for predicting the three-dimensional structure of linear RNA.

4. A method for predicting circular RNA IRES according to claim 1, characterized in that In the process of IsRNAcirc predicting the three-dimensional structure of circular RNA, the sequence information and secondary structure information of the circular RNA for which the three-dimensional structure needs to be predicted need to be placed in the same file.

5. A method for predicting circular RNA IRES according to claim 1, characterized in that The standard output of IsRNAcirc includes a three-dimensional structural model of the predicted circular RNA in the standard PDB format, trajectory files for replica exchange molecular dynamics simulation, and log files for circularization and running simulations.

6. A method for predicting circular RNA IRES according to claim 1, characterized in that: The IsRNAcirc predicts the three-dimensional structure of circular RNA within 300 hours.

Images