Fig 1.
The framework for predicting circular RNA 3D structures using IsRNAcirc.
(A) Converting an all-atom model (PDB ID:1Z5C, left) into the coarse-grained representation (right) through predefined mapping relationships in IsRNAcirc. IsRNAcirc predicts circular RNA 3D structure using four main steps: (B) Input preparation. From sequence information, cRNAsp12 [31] was used to predict the 2D structure of circular RNAs. Based on the sequence and predicted 2D structure, RNAComposer [18] was used to generate initial 3D structures of the corresponding linear counterpart. Sequence information, predicted 2D structure, and those obtained initial 3D structures are used as inputs for IsRNAcirc for circular RNA 3D structure predictions; (C) End closure. The 5’- and 3’-end of circular RNA are circularized using a harmonic restraint between the two ends of linear structure and the simulated annealing technique. A clustering procedure was performed based on the last 10% MD trajectories to obtain starting structures for the subsequent step; (D) Structure prediction. Possible 3D structures of circular RNA were predicted following a similar process to IsRNA2, including conformational sampling through REMD simulations, clustering the lowest energy conformations to generate the most probable structures, and all-atom reconstruction. By default, five predicted 3D models are provided; (E) Model refinement. QRNAS [32] was employed to refine the predicted 3D models by eliminating problems such as spatial steric resistance and bond breakage.
Fig 2.
IsRNAcirc can significantly reduce the locally irrational topologies contained in initial 3D structures (generated by RNAComposer).
(A-D) Four types of locally irrational topology were identified in initial 3D structures through view inspection: (A) template-free loop due to a lack of related templates in the library, (B) backbone tangle caused by improper template assembly, (C) base collision, in which a backbone segment passes through the middle of two nucleobases, and (D) sharp twist, in which an apparently unreasonable sharp turn is observed in the backbone. The names of circular RNAs for which the relevant irrational topology was observed in the initial 3D structures were also given in top left. (E) The total numbers of locally irrational topologies observed in all initial 3D structures (three structures for each RNA) and the predicted 3D models by IsRNAcirc (five models for each RNA, if available) were displayed in a type-dependent manner. All 34 circular RNAs were considered.
Fig 3.
A representive prediction for helix-circular RNAs.
(A) Predicted 2D structure of POLR2A consisting of 336 nucleotides (circRNA ID: hsa_circ_0000741), in which the BSJ is located in the helical region (marked by red arrow). (B) Superposition of the initial linear 3D structure (colored by light brown) and the predicted 3D structure of POLR2A circular RNA by IsRNAcirc (colored by light blue). (C-D) Zoom-in show the local details of 3D models: (C) the presence of template-free loop in the initial 3D structure and (D) the corresponding segment in the predicted 3D model by IsRNAcirc. The atoms carbon, oxygen, and nitrogen are colored by silver, red, and blue, respectively.
Fig 4.
A representative prediction for hairpin-circular RNAs.
(A) Predicted 2D structure of FKBP8 consisting of 259 nucleotides (circRNA ID: hsa_circ_0000915), in which the BSJ is located in the hairpin loop (marked by red arrow). (B) Superposition of the initial linear 3D structure (colored by light brown) and the predicted 3D structure of FKBP8 by IsRNAcirc (colored by light blue). (C-D) Zoom-in show the local details of 3D models: (C) the presence of a backbone tangle in the initial 3D structure and (D) the optimized segments in the 3D model predicted by IsRNAcirc. The atoms carbon, oxygen, and nitrogen are colored by silver, red, and blue, respectively.
Fig 5.
A representative prediction for internal-circular RNAs.
(A) Predicted 2D structure of MBOTA2 consisting of 224 nucleotides (circRNA ID: hsa_circ_0007334), in which the BSJ is located in the 1x1 internal loop (marked by red arrow). (B) Superposition of the initial linear 3D structure (colored by light brown) and the predicted 3D structure of MBOTA2 by IsRNAcirc (colored by light blue). (C-D) Zoom-in show the local details of 3D models: (C) the presence of sharp twist of backbone in the initial 3D structure and (D) the optimized topology in the 3D model predicted by IsRNAcirc. The atoms carbon, oxygen, and nitrogen are colored by silver, red, and blue, respectively.
Fig 6.
A representative prediction for junction-circular RNAs.
(A) Predicted 2D structure of SLC22A23 consisting of 259 nucleotides (circRNA ID: hsa_circ_0075504), in which the BSJ is located in the 3-way junction loop (marked by red arrow). (B) Superposition of the initial linear 3D structure (colored by light brown) and the predicted 3D structure of SLC22A23 by IsRNAcirc (colored by light blue). (C-D) Zoom-in show the local details of 3D models: (C) the presence of base collision in the initial 3D structure and (D) the corresponding segments in the 3D model predicted by IsRNAcirc. The atoms carbon, oxygen, and nitrogen are colored by silver, red, and blue, respectively.
Fig 7.
Comparison of the performance of circular RNA 3D structure predictions between IsRNAcirc and 3dRNA based on the (A) DFIRE-RNA and (B) rsRNASP scoring function. Each 3D model predicted by IsRNAcirc or 3dRNA method was scored individually and displayed according to the circularized types. e.u. (energy unit).
Table 1.
Summary of the performance of IsRNAcirc relative to 3dRNA in circular RNA 3D structure prediction.
Lists are the number of cases where IsRNAcirc prediction has a lower energy score in different range based on the DFIRE-RNA and rsRNSP scoring schemes.
Fig 8.
Runtime (total CPU time) in IsRNAcirc as a function of length of circular RNAs.
Circles represent 34 circular RNAs and the line indicates the fitting result. Predictions were performed on intel(R) Xeon(X) CPU E5-2620 v4 @ 2.10GHz.