AI-integrated network for RNA complex structure and dynamic prediction

Abstract

RNA complexes are essential components in many cellular processes. The functions of these complexes are linked to their tertiary structures, which are shaped by detailed interface information, such as binding sites, interface contact, and dynamic conformational changes. Network-based approaches have been widely used to analyze RNA complex structures. With their roots in the graph theory, these methods have a long history of providing insight into the static and dynamic properties of RNA molecules. These approaches have been effective in identifying functional binding sites and analyzing the dynamic behavior of RNA complexes. Recently, the advent of artificial intelligence (AI) has brought transformative changes to the field. These technologies have been increasingly applied to studying RNA complex structures, providing new avenues for understanding the complex interactions within RNA complexes. By integrating AI with traditional network analysis methods, researchers can build more accurate models of RNA complex structures, predict their dynamic behaviors, and even design RNA-based inhibitors. In this review, we introduce the integration of network-based methodologies with AI techniques to enhance the understanding of RNA complex structures. We examine how these advanced computational tools can be used to model and analyze the detailed interface information and dynamic behaviors of RNA molecules. Additionally, we explore the potential future directions of how AI-integrated networks can aid in the modeling and analyzing RNA complex structures.

I. INTRODUCTION AND MOTIVATION

RNA complexes are fundamental biomolecules that bind proteins, DNA, or small molecules. They play essential roles in critical biological processes, such as transcription and translation.^1–3 For example, RNA polymerase II (Pol II) recruits the U1 small nuclear ribonucleoprotein particle to initiate co-transcriptional splicing.⁴ RNA complexes are also closely linked to the occurrence of diseases, such as AIDS and COVID-19.^5–7 It is widely known that the function of biomolecules is intricately connected to their tertiary structure. Understanding their structures helps scientists uncover the mechanisms underlying various biological processes, leading to advancements in numerous natural disciplines. This structural insight can inform the development of new therapeutic strategies, enhance our understanding of disease mechanisms, and drive innovations in drug design, ultimately contributing to improved health outcomes and the advancement of biomedical science.^8–11

Experimental methods, such as x-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy, and cryo-electron microscopy (cryo-EM), have been developed to determine the structure of biomolecules. However, these techniques are expensive and time-consuming, limiting their widespread use.^12–14 Additionally, each method faces technical challenges. For example, samples for x-ray crystallography must be crystallizable, which limits the types of samples with high flexibility.¹⁵ NMR spectroscopy presents challenges when analyzing large biomolecules. Cryo-EM density maps have low resolution, making structure determination challenging and expert-intensive.¹⁶ The existing constraints are causing a slowdown in the modeling and analysis of biomolecular structures. This impedes the speed at which scientific and medical advancements can be made. Hence, there is an urgent need to develop computational approaches to speed up research on RNA complexes.

Several computational prediction approaches have recently been developed to overcome the experimental limitations. Monomeric structure prediction methods, such as Alphafold3, RoseTTAFold, and 3dRNA, have been proposed and have achieved accuracy comparable to experimental methods. However, predicting RNA complex structures remains challenging.^17–20 Current RNA complex prediction methods predominantly rely on molecular docking. These methods face challenges due to RNA's flexible conformation and the complicated features of its interfaces. Information on RNA binding sites contributes to determining the locations where molecules interact with RNA, which is particularly useful for identifying targets for small molecules. However, larger molecules involve more extensive binding interfaces, necessitating more detailed information on interfacial contacts. Over the past decade, advancements in network-based approaches have significantly improved the ability to obtain information on binding sites and interfacial contacts. These methods, known for their robustness and simplicity, have greatly enhanced the molecular docking process. In addition, structure prediction methods usually focus on static structures. The functional RNA complexes are often related to their dynamic behavior, particularly the dynamic changes of their interfaces. Dynamic network analysis methods have been extensively utilized in this area of research. Artificial intelligence (AI) and network-based methods have recently been increasingly proposed and developed to overcome the limitations of experimental techniques in RNA complex modeling and dynamical analysis.

This review article focuses on RNA complex structure modeling and dynamic analysis. We start by summarizing how AI and network-based methods are integrated into RNA complex structure modeling, emphasizing advances in binding site and interface contact prediction [Figs. 1(a) and 1(b)]. These AI-network-driven approaches have provided more accurate and efficient information to model RNA complexes. Following this, we present an overview of AI-integrated network methods for analyzing complex dynamic motions [Fig. 1(c)]. This includes techniques for understanding how these complexes behave and interact over time, which is crucial for comprehending their biological functions and mechanisms. We then discuss potential future directions in this field and the remaining challenges and opportunities for innovation. We also explore emerging technologies and methodologies that could further enhance our understanding of RNA complexes. We aim to provide a comprehensive outlook on the evolving landscape of RNA complex structure prediction and dynamic behavior analysis in the context of AI and network-based approaches.

FIG. 1.

The AI-integrated network methods are applied in RNA complex structure modeling and dynamic analysis. (a) Prediction of the binding site of the receptor (R) concerning the ligand (L). (b) Complex contact prediction. (c) Dynamic motions analysis for RNA complex.

II. COMPLEX STRUCTURE MODELING

A. Overview

Classical complex structure prediction algorithms are usually based on molecular docking, which predicts the RNA complex structures using individual components. The idea of docking algorithms originates from the lock-and-key model and the induced fit theory.²¹ The docking process aims to obtain the optimal binding mode that satisfies the spatial shape complementarity and energy minimization simultaneously.²² Traditional docking methods typically explore conformational space using techniques, such as Fast Fourier Transform (FFT) and Monte Carlo, to generate numerous candidate complex configurations.²³ Then, these conformations are ranked by using scoring functions to predict the near-native structures according to energy scores.^24,25 Many classical molecular docking algorithms exist, such as HADDOCK and HSYMDOCK for protein–protein docking, HDOCK and 3dRPC for protein-NA docking, and NLDock for NA-ligand docking.^24,26–32 RNA complex docking accuracy is limited by two main factors: (1) the conformational flexibility of RNA and (2) the complex interface interaction characteristics between RNA and protein. Therefore, researchers are interested in identifying binding site and interface contact information to guide the complex modeling.

Binding site information can help quickly identify potential RNA binding hotspots, which is valuable for guiding RNA complex modeling. It has also been shown to improve the prediction of RNA–small molecule interactions, thereby accelerating downstream drug discovery tasks.³³ For larger proteins with extensive binding interfaces, it is important to consider not only the RNA binding area but also the protein binding pose. Therefore, contact information between the two monomers can also aid in predicting the complex structure. In structure prediction, it has been proven that contact information between residues or nucleotides enhances the accuracy of predictions.³⁴ Recently, AI-integrated network methods are gradually beginning to emerge. In these approaches, biomolecules are viewed as networks, with nodes representing residues (proteins) or nucleotides (RNA) and edges typically defined by a distance threshold. This threshold is primarily based on the length of physical interactions. Two residues/nucleotides are considered connected if the distance between any pair of their heavy atoms is less than the defined distance threshold.³⁵

B. Binding sites prediction

The nucleotides in RNA that interact with other molecules are known as binding sites or binding pockets. Experimental RNA binding sites are determined by the nucleotides located within a specific atomic distance cutoff, such as 4 Å, from other molecules.^35,36 RNA binding sites are critical in determining complex structures and provide information on biomolecular interaction and assembly. Accurate knowledge of binding sites can significantly enhance our understanding of molecular regulation, leading to the development of targeted RNA inhibitors.

Recently, traditional methods for identifying RNA binding sites have been developed. These prediction approaches can be broadly categorized into sequence-based and tertiary structure-based. Rsite2 is a sequence-based method for RNA binding site prediction.³⁷ Zeng et al. found a close relationship between the distance of nucleotides and the locations of RNA binding sites. They developed Rsite2, which calculates the Euclidean distances between each nucleotide and other nucleotides using the predicted secondary structures derived from the sequence. It then identifies the nucleotides that are the extreme points on the distance curve as functional sites. This approach has demonstrated that the connectivity between nucleotides holds the potential for discovering the functional sites by testing three ncRNAs [tRNA (Lys), Diels-Alder ribozyme, and RNase P]. Tertiary structure-based approaches include Rsite and RBind. Rsite replaces the secondary structure Euclidean distance with a tertiary structure Euclidean distance, refining the identification process of these functional sites.³⁸ RBind transforms RNA tertiary structures into networks and utilizes network properties (degree and closeness centrality) to identify the binding connectivity.³⁵ By analyzing these network properties, RBind effectively identifies RNA binding sites, offering a comprehensive understanding of the structural and functional relationships within RNA molecules. These methods have shown the effectiveness of network properties in binding site prediction. Recently, AI-integrated network methods have been developed to further explore the role of network features in identifying binding sites. RNetsite considers multiple local and global network features and utilizes an ensemble machine learning strategy to identify RNA binding sites, effectively demonstrating the effectiveness of AI-integrated networks strategy in RNA binding site prediction.³⁹ RNAsite uses a random forest model that combines structural features derived from network properties, such as degree and closeness centrality, with additional structural features such as accessible surface area and Laplacian norm. It also incorporates multiple sequence alignment (MSA) to consider the evolutionary conservation of each position in the RNA sequence.³⁶ Similarly, RLBind employs a deep learning-based pipeline using a convolutional neural network that integrates both structural and sequence features for binding site prediction.⁴⁰ The sequence properties include nucleotide types and evolutionary conservation, while the structure properties encompass network topological characteristics, biochemical properties, and accessible surface areas. Table I provides a brief description of the methods described above.

TABLE I.

List of RNA binding sites prediction methods.

Name	Proposed year	Type	Technique	Features	References
Rsite	2015	Structure-based	Distance	RNA tertiary structures	38
Rsite2	2016	Sequence-based	Distance	RNA sequence	37
RBind	2018	Structure-based	Network	Network properties	35
RNAsite	2021	Structure-based	AI + Network	Network properties	36
				RNA tertiary structures
				RNA sequence
RLBind	2022	Structure-based	AI + Network	Network properties	40
				RNA tertiary structures
				RNA sequence
RNetsite	2023	Structure-based	AI + Network	Network properties	39

To evaluate the performance of binding site prediction methods, we compare Rsite2 with network-based RBind and AI-integrated network-based RNetsite. The comparison was performed using a combination of the RNA-molecule binding site benchmark datasets TE18 and RB9. After removing three overlapping structures from TE18 and RB9, 24 RNA test cases were included (called TE24). The test results showed that RNetsite achieved the highest performance on TE24 in terms of Precision, Recall, and Matthews correlation coefficient (MCC), with RBind showing the second-best performance [Figs. 2(a)–2(c)]. This demonstrates that network features significantly aid in identifying binding sites, and the integration of AI technologies further enhances the effectiveness of RNA binding site recognition.

FIG. 2.

The performance of RNetsite (purple), RBind (green), and Rsite2 (blue) models on the TE24 set. (a) Precision. (b) Recall. (c) MCC. RNetsite produced the best results, followed by RBind and Rsite2.

C. RNA complex contacts

Intermolecular contacts between complexes are crucial for regulating and executing various cellular processes. Typically, two residues or nucleotides from interacting biomolecules are considered to be in contact if the distance between their atoms is within a defined threshold, usually 6 or 8 Å.⁴¹ The prediction of complex inter-monomer contacts aims to determine whether two positions from different monomers within an interface are spatially close to each other in a tertiary structure. Understanding these contacts offers valuable insight into the mechanisms underlying cellular communication, regulation, and the organization of biological activities. In addition, accurately predicting contacting monomer pairs between interacting biomolecules is crucial for the structural characterization of complex interactions. This prediction can serve as an intermediate step in complex structure prediction,^42,43 as the identified contacts can be incorporated into docking algorithms to enhance the accuracy of complex structure models.^26,44,45 Moreover, these predicted contacts are invaluable for guiding the design of interfaces and can be extended to predict novel inter-complex interactions.⁴⁶

Biomolecules will likely be highly conserved or coevolutionary to maintain fundamental or functional interactions. Due to evolutionary pressure, coevolving monomer pairs are often found to be spatially proximal in complex structures.⁴⁷ Coevolutionary analysis methods, such as mfDCA and plmDCA, have been employed in previous studies on inter-protein contacts.^48,49 However, these methods have certain limitations, such as the need for a large number of homology sequences and relatively low accuracy.

Although AI-based approaches have also been applied to predict inter-contacts,^50–54 RNA complex inter-contact prediction is still in its early stages, hindered by challenges, such as scarce data, the flexible RNA structure, and the complexity of RNA complex binding interfaces. Nonetheless, AI-based network methods are being explored to predict intra-RNA contacts.^55–57 For example, we have developed DIRECT, a hybrid approach that utilizes a restricted Boltzmann machine (RBM) to extract contact pattern information from the structural distance graph.⁵⁸ The results demonstrate that the AI-based network approach enhances the accuracy of contact inference from sequence co-variations. Given that protein datasets are much larger than those available for RNA contact prediction, we found that knowledge from a protein-coevolution transformer-based language model can be effectively transferred to the task. To achieve this, we developed CoT-RNA-Transfer, which significantly enhances RNA contact prediction through transfer learning using a publicly available protein language model.⁵⁹ Our findings suggest that structural patterns learned from proteins can be successfully transferred to RNAs, paving the way for new research opportunities.

III. COMPLEX DYNAMIC ANALYSIS

A. Overview

Numerous biological functions are intricately regulated by complex RNA dynamical processes.^60,61 The binding interface is predominantly governed by atomic and residue-level interactions, which play a crucial role in driving complex conformational changes and motions. Understanding these motions is essential for gaining insights into RNA complex binding interfaces' stability and regulating their functions. Molecular dynamics (MD) simulations offer a practical approach for studying these dynamic processes. MD simulations provide a direct perspective on the dynamic behavior of biomolecular structures. However, as system sizes increase and enhanced sampling techniques become more popular, there is a growing need for advanced analysis tools capable of extracting information from vast amounts of data and providing new insights.⁶² Dynamic network methods have been developed and have become effective tools for extracting interface features from MD simulations, such as key nucleotides/residues and communication pathways within molecular complexes.^63–65 The dynamical network method transforms the atomic representation of MD simulations into a graph representation with nodes and edges, facilitating the analysis of dynamic structures in biomolecular systems through graph analysis methods.^66,67 Building on the development of dynamic network analysis methods, various AI techniques have recently been introduced to the dynamic analysis community.^68–70

B. Dynamic network analysis methods

In the context of MD simulations, residues and nucleotides within an RNA complex are represented as individual network nodes. Edges are established between pairs of nodes when the corresponding residues or nucleotides are in contact during the MD simulation. Two nonconsecutive residues or nucleotides are defined to be in contact if any of their heavy atoms (non-hydrogen atoms) are within a defined distance threshold (such as 4.5 Å) for the majority of frames throughout the MD simulations.⁶³ The binding interface is the residue-nucleotide pairs if the distance between residue and nucleotide is less than 4 Å.^39,63

In the dynamic network analysis, the edges are weighted by the correlation values derived from the simulation. The distance between two nodes connected by an edge decreases as the correlation (or interaction energy) between the residues/nucleotides increases. The correlation, denoted as $C_{ij}$, between the motions of nodes $i$ and $j$ provides a quantitative measure of the information transfer between these nodes. In essence, the motion of residue or nucleotide $i$ can be used to predict the directional motion of residue or nucleotide $j$, indicating a level of coordinated movement or interaction within the molecular system. The weighted edges can be easily employed to investigate the behavior within the binding interface.

The dynamical cross coefficients (DCC) have been widely used in the analysis of MD simulation data.^71–74 The DCC assesses the consistency of motion direction between two nodes, which can be calculated by the following equation:

$$C_{ij}=\frac{⟨\Delta {\overrightarrow{r}}_i\left(t\right)·\Delta {\overrightarrow{r}}_j\left(t\right)⟩}{{\left(⟨{\Delta {\overrightarrow{r}}_i\left(t\right)}^2⟩⟨{\Delta {\overrightarrow{r}}_j\left(t\right)}^2⟩\right)}^{\frac{1}{2}}}$$ (1)

where ${\overrightarrow{r}}_i\left(t\right)$ is the position of the $C_{\alpha }$ or P of the $i$ th residue or nucleotide of the protein or RNA at time $t$. $\Delta {\overrightarrow{r}}_i\left(t\right)={\overrightarrow{r}}_i\left(t\right)-⟨{\overrightarrow{r}}_i\left(t\right)⟩$, and $⟨·⟩$ refers to the time average of the quantity within the brackets. The DCC between two residues is represented by a value ranging from −1 to 1. A positive $C_{ij}$ value indicates correlated motion, where the residues move in the same direction in most frames. A negative $C_{ij}$ value indicates anticorrelated motion, with residues moving in opposite directions across most frames. If the correlation value between the two residues is close to zero, their motion is considered uncorrelated. The property of DCC has made it a tool to assess the stability of binding interfaces: a positive DCC indicates a tendency for the interface to maintain stability while a negative DCC suggests a tendency to breakup. However, it may not fully capture the impact of varying velocities on the interfaces. The method DCC studies the consistency of motion direction to determine whether the interface is disintegrating. Due to the constraints imposed by the backbone connections in monomeric structures, the structure tends to be rigid, leading to less variation in velocity. As a result, assessing motion correlation based on direction alone is feasible in monomers. However, even when the motion directions are aligned in multimeric complexes, differences in velocity can also cause the interface to breakup. Therefore, relying solely on direction to assess interface stability is inaccurate.

To address this limitation, we have developed distance-based dynamical network correlation (DDNC) analysis, which integrates direction and velocity to more precisely characterize interface binding behavior.³⁹ The DDNC for residue $i$ and nucleotide $j$ at time $t$ can be calculated as follows:

$$M_{ij}\left(t\right)=\frac{\left({\overrightarrow{r}}_i\left(t\right)-{\overrightarrow{r}}_{ij}\right)·\left({\overrightarrow{r}}_j\left(t\right)-{\overrightarrow{r}}_{ij}\right)}{\left|{\overrightarrow{r}}_i\left(t\right)-{\overrightarrow{r}}_{ij}\right|\left|{\overrightarrow{r}}_j\left(t\right)-{\overrightarrow{r}}_{ij}\right|}.$$ (2)

${\overrightarrow{r}}_i\left(t\right)$ is the position of the $i$th residue or nucleotide at time $t$. ${\overrightarrow{r}}_{ij}$ serves as a steady-state reference point derived from the initial structure, obtained by solving an optimization problem as follows:

$$\begin{array}{c}\text{min}\hspace{0.17em}\left|{\overrightarrow{r}}_{ij}-\overrightarrow{r}\right|,\\ {\overrightarrow{r}}_{ij}\hspace{0.17em}\hspace{0.17em}\mathrm{s}.\mathrm{t}.\hspace{0.17em}\{\begin{array}{c}{(\overrightarrow{r}}_{ij}-m)\cdot ({\overrightarrow{r}}_i-{\overrightarrow{r}}_j)=0,\\ \hspace{0.17em}({\overrightarrow{r}}_i-{\overrightarrow{r}}_{ij})\cdot ({\overrightarrow{r}}_j-{\overrightarrow{r}}_{ij})=0.\end{array}\end{array}$$ (3)

“ ${\overrightarrow{r}}_i$” (“ ${\overrightarrow{r}}_j$”) represents the position of the $i$th ( $j$th) residue or nucleotide of the protein or RNA in the initial structure. $m$ is the midpoint of the line connecting ${\overrightarrow{r}}_i$ and ${\overrightarrow{r}}_j$. $\overrightarrow{r}$ is the average position of the binding interface in the initial structure This ensures that ${\overrightarrow{r}}_{ij}$ is one of the closest points to the average position binding interface and guarantees that (1) the correlation between residue $i$ and nucleotide $j$ calculated by Eq. (2) is uncorrelated, indicating a steady state in the initial structure and (2) the stabilization and breaking of the interface both lead to corresponding changes in the correlation coefficients, highlighting the dynamic motions of the interface. Finally, a period from time $T_1$ to time $T_2$ is denoted as $\tau$, the DDNC of this period is defined as follows:

$$M_{ij}\left(\tau \right)={⟨M_{ij}\left(t\right)⟩}_{\tau },$$ (4)

where ${⟨·⟩}_{\tau }$ denotes the time average of the quantity within the brackets during period $\tau$. To compare the performance of DCC and DDNC, we tested both methods in the context of experimentally validated interface disruption between HIV TAR RNA and P-TEFb/Tat complex caused by an RNA inhibitor.^6,75 The HIV viral protein Tat hijacks the cellular positive transcription elongation factor b (P-TEFb) to initiate viral transcription. The P-TEFb/Tat complex binds to the viral transactivation response element (TAR) RNA to overcome the elongation pause. Figure 3(a) shows the dynamic changes of this system. The competitive inhibitor JB181 has been found to disrupt the P-TEFb/Tat and TAR interface based on both experimental data and molecular dynamics simulations. Figures 3(b)–3(d) represent the interface dynamics characterized by DCC, while Figs. 3(e)–3(g) represent the interface dynamics characterized by DDNC. In these figures [Figs. 3(b)–3(g)], positive values indicate that the interface remains stable, whereas negative values suggest that the interface tends to disrupt. The analysis results of DDNC show that the interface breaks up in Fig. 3(e). Subsequently, the interface exhibits fluctuations [Fig. 3(f)] and ultimately breaks up [Fig. 3(g)]. In contrast, the analysis results of DCC depict a different breaking-up process. The interface breaks up in Fig. 3(a), returns to a stable state [Fig. 3(b)], and ultimately breaks up [Fig. 3(c)]. Simulation results indicate that DDNC outperforms the traditional DCC method in capturing the dynamic behavior of the interface [Figs. 3(a)–3(g)].

FIG. 3.

Comparison of DCC and DDNC in molecular dynamics simulations of unstable interfaces. (a) MD simulation of HIV TAR RNA and P-TEFb/Tat complex. (b)–(d) DCC characterization of interface behavior. (e)–(g) DDNC characterization of interface behavior. The interface changes characterized by DDNC are more reasonable than DCC.

Biomolecular systems are represented by networks with connected nodes. Various computational methods, such as Dijkstra's and Floyd-Warshall algorithms, can be applied to determine the essential pathways within the system.⁷⁶ This computational approach has been widely studied, leading to the development of tools that precisely identify information pathways transmitting signals from the allosteric site to the target.^77–79 These tools have significantly advanced the ability to model signal propagation within biomolecular systems, offering deeper insight into allosteric regulation and its impact on molecular interactions.^63,66,80

IV. FUTURE DIRECTIONS

AI-integrated network methods have significantly influenced the field of structural biology, offering advanced approaches to tackle the modeling of complex and analyzing dynamic motions. These methods enable researchers to learn the intricate interactions within biological structures, facilitating a deeper understanding of their structures, behavior, and functions. There are several future directions to explore in this area.

Predicting binding sites of RNA remains an improvement for further exploration. Unlike proteins, RNA molecules exhibit more flexibility due to the higher number of backbone torsion angles and intricate base pairing, resulting in a more complicated and flexible topology.^81,82 The unique structure and dynamic nature of RNA make it challenging to predict its binding sites accurately. Despite these difficulties, network-based binding site prediction using artificial intelligence techniques shows a promising approach for advancing RNA research. Current efforts are focused on understanding the relationship between RNA structure and functional binding sites through the analysis of the RNA structural network. The ongoing exploration and improvement of AI-based network methods, such as graph neural networks, which have shown potential in predicting protein binding sites, have significant potential to advance RNA binding site research. For example, GraphBind uses an end-to-end graph neural network to identify nucleic acid-binding residues on proteins. Since binding sites often exhibit highly conserved patterns within local tertiary structures, GraphBind constructs graphs based on the structural contexts of target residues and their spatial neighborhoods. To recognize binding residues accurately, GraphBind employs hierarchical graph neural networks (HGNNs) to capture latent local patterns of structural and physicochemical characteristics.⁸³ Similarly, GPSite utilizes predicted protein structures to construct protein graphs. It uses an edge-enhanced graph neural network to extract residual and relational geometric contexts end-to-end, facilitating precise binding site identification.⁸⁴

Predicting interactions within RNA complexes is still in its infancy, presenting challenges that could shape future research directions. One of the primary challenges lies in the structural flexibility of RNA, which differs significantly from proteins. RNA complex interfaces tend to exhibit complicated characteristics. For example, RNA exerts its biological functions by forming complexes with proteins. During the binding process, RNA and proteins recognize and attract each other through long-range electrostatic interactions, followed by short-range interactions, such as van der Waals forces and hydrogen bonds, which optimize their conformations and facilitate tight binding.⁸⁵ RNA regions often exhibit flexible conformational changes upon binding. Determining how these flexible structural regions interact with the complex is critical. Traditional coevolutionary methods rely heavily on multiple sequence alignment (MSA) information, leading to lower prediction accuracy in “orphan” structures whose homologous sequence data are scarce.^86,87 To address this, AI-driven large language models can be employed to overcome the “orphan” effect caused by limited homologous information, thereby improving the accuracy of RNA complex contact predictions. AI-driven large language models have already been applied to predict inter-protein contacts.^50–54 Graph neural networks have been also increasingly applied to protein complexes to predict inter-chain residue-residue contacts. GLINTER and PLMGraph-Inter exemplify this approach, using the structures of interacting monomers and their rotationally invariant graph representations as input features.^52,88 The network model can further help address the challenges posed by the flexible changes in the RNA binding regions.

Static networks are valuable for identifying potential binding sites and complex binding interface contacts. They provide insight into the potential interaction hotspots and post-binding structure. However, function regulation is closely related to conformational dynamics, especially in flexible RNA structures. Understanding the dynamic changes in complex structures can help elucidate their functional mechanisms, which is particularly important for downstream applications such as drug discovery and RNA inhibitor design (Fig. 4). Dynamic networks offer a deeper understanding of the post-binding behavior by analyzing conformational changes and interaction dynamics over time. Then, researchers can comprehensively understand the static and dynamic aspects of regulating RNA function. We recently successfully designed two competitive inhibitors to RNA hairpin motifs, SL3 and RBE RNA, with the help of this strategy.⁸⁹ The SL3 RNA plays a crucial role in recognizing and packaging the HIV-1 genome by interacting with the NC protein. Among potential inhibitors, L22#15 has demonstrated a promising low binding affinity of approximately 5 μM in NMR experiments, outperforming other SL3 RNA inhibitors. Similarly, RBE RNA, which regulates viral gene expression through its interaction with the HIV-1 Rev protein, has been targeted by the inhibitor L22 + 16, achieving a binding affinity of around 11 μM. These findings indicate the effectiveness of network-based models in RNA function regulation and inhibitor design.

FIG. 4.

Flowchart illustrating the process of designing RNA inhibitor. (a) Predicting potential binding sites for inhibitor binding. (b) Inferring interface interactions within RNA complexes. (c) Analyzing functional regulation through MD simulations.

Recently, AI-based techniques have opened new possibilities to capture changes in simulations, providing deeper insights into biomolecules' complex motions and interactions. One direction is to utilize AI-based techniques to capture subtle MD dynamic behaviors within complex systems. Karamzadeh et al. used statistical machine learning methods to extract critical structural features that exhibit significant differences in dynamics.⁶⁸ Structural stability and subtle, flexible changes can help us better understand the regulatory mechanisms of biomolecules, facilitating the design of RNA molecules with new functionalities. Another direction is to use short simulation trajectories to predict future simulation states. For example, Zhu et al. employed a neural relational inference model based on a graph neural network to predict latent interactions associated with allostery to understand the allosteric pathway from the ligand binding or mutation site to the active center in a protein.⁹⁰ Thus, comprehending the patterns of biomolecular changes can save substantial time and resources, thereby accelerating the discovery of structural changes induced by drug design.

Integrating AI and physics-based network models enables rapid identification of binding sites and physical interactions, facilitating the prediction of complex docking structures. However, predicting binding sites' robustness remains a challenge due to the flexibility of RNA. Combining AI with coarse-grained complex network models can significantly improve the accuracy and robustness of binding site predictions. Once the complex structure is determined, studying its dynamic behavior becomes equally important. Considering the nodes' direction and velocity in the network allows for a more precise interface analysis. Integrating more dynamic data with AI-integrating network methods makes predicting long-term dynamic structures from short-term dynamic simulations possible. Integrating AI and physics-based network models can reduce interference from irrelevant information, significantly enhancing the efficiency of learning abstract features.

V. CONCLUSION

This review introduces AI-integrated network methods, highlighting their significance in biomolecular complex modeling and dynamic analysis. We have reviewed the evolution of binding site prediction methods, tracing the development from traditional network analysis to the integration of AI technologies. Subsequently, once the binding sites have been identified, the contacts between the binding interfaces can be studied. The progress in predicting contacts for protein complexes has been successful, whereas the prediction of contacts in RNA complexes is still at an early stage. Further development in these methods could contribute to the structural prediction of RNA complexes, marking a promising area for future research. Moreover, we discussed dynamic network analysis techniques for biomolecular MD simulations, which are crucial for understanding allosteric effects and functional regulation in biomolecules. These methods are helpful to enhance our ability to design inhibitors and tackle other downstream tasks more effectively. As AI progresses, its application in this field is expected to offer valuable insight and guidance for future research in complex structure prediction and practical applications, such as inhibitor design and regulation of molecular function.

NOMENCLATURE

AI

artificial intelligence

Cryo-EM

cryo-electron microscopy

DCC

dynamical cross coefficients

DDNC

distance-based dynamical network correlation

MCC

Matthews correlation coefficient

MD

molecular dynamics

MSA

multiple sequence alignment

NMR

nuclear magnetic resonance

AUTHOR DECLARATIONS

Conflict of Interest

The authors have no conflicts to disclose.

Author Contributions

Haoquan Liu: Data curation (equal); Methodology (equal); Writing – original draft (lead); Writing – review & editing (equal). Chen Zhuo: Data curation (supporting); Investigation (supporting). Jiamin Gao: Data curation (supporting); Investigation (supporting). Chengwei Zeng: Data curation (supporting); Investigation (supporting). Yunjie Zhao: Funding acquisition (lead); Investigation (lead); Project administration (lead); Supervision (lead); Writing – review & editing (lead).

DATA AVAILABILITY

Data sharing is not applicable to this article as no new data were created or analyzed in this review.