Post-Docking Refinement of Peptide or Protein-RNA Complexes Using Thermal Titration Molecular Dynamics (TTMD): A Stability Insight

Abstract

RNA–protein interactions drive and regulate fundamental cellular processes like transcription and translation. Despite being still limited, the growing body of structural data significantly contributes to the characterization of these interactions. However, RNA complexes involving proteins or peptides are not always available due to the structural determination challenges that this biopolymer entails. Consequently, modeling approaches like molecular docking are exploited to generate complexes relevant to structural and pharmaceutical purposes, including analysis of putative drug targets. Docking methods, despite their widespread adoption, are often hindered by limitations in scoring accuracy, which affects the ranking of the generated poses. Postdocking refining methods, including molecular dynamics (MD) approaches, have been developed to tackle this issue. Thermal Titration Molecular Dynamics (TTMD) is an enhanced sampling molecular dynamics technique that has been previously effectively applied to refine protein or RNA-small-molecule docking poses. This study presents the first application of TTMD to RNA-peptide complexes, validating this method on more complex systems and extending its applicability domain. Our findings showcase the capability of this technique to refine peptide-RNA docking poses, correctly identifying native binding modes among decoys for different pharmaceutically relevant targets.

Introduction

Ribonucleic acid (RNA) is a crucial molecule within the cell, commonly considered a bridging element between genetic information and protein expression. Beyond its canonical “Central Dogma” role, RNA has emerged as a versatile biopolymer involved in several cellular processes like RNA transcription, translation, and splicing.¹⁻³

Remarkably, only a fraction of the genome is translated into proteins,⁴ which have been the main pharmaceutical target for decades, leaving a large part of the genome unexplored. For this reason, noncoding RNAs are now receiving significant attention as their dysregulation is related to conditions like cancer and neurodegenerative diseases,⁵ and understanding these processes can lead to the identification of new pharmaceutically relevant targets. In particular, flaws in protein-RNA interactions are found in these diseases,^6,7 and the proteins that can bind RNA, defined as RNA Binding Proteins (RBP), constitute a not neglectable portion of all the proteins (7.5%).⁸ Consequently, studying the interactions between RNA and proteins could give insights into cellular processes and could be strategic for identifying new therapeutics.

An increasing, yet limited, number of high-resolution structures of RNA-protein complexes (ribonucleoproteins, RNPs) has been solved and deposited in public databases in the past decade; however, the experimental determination of these complexes is still a time-consuming and complex task due to the peculiar nature of RNA. The high structural flexibility, the negatively charged phosphate backbone, and the transient nature of the interaction involving RNA-protein complexes hamper the resolution of these structures.³ Given the lack of solved structures, modeling approaches could help to predict and study these interactions, with molecular docking being one of the most adopted methods.³

The docking procedure is composed of two stages: ranking and scoring. The first one is the generation of possible binding modes (poses) of a molecule at a given target by sampling its orientation and conformation, while the latter consists of ranking the generated models using a scoring function.⁹ One of the main issues associated with sampling is that the recognition process and the peptide-RNA complex formation often involve conformational changes. While many tools perform rigid body docking, neglecting this critical aspect in favor of speed and reducing computing costs, some semiflexible programs can address this conformational issue by considering local rearrangements. Still, conformational changes involving backbone and loop rearrangements are considered a critical challenge in modeling macromolecular complexes.⁹

On the other hand, pose scoring can suffer from inaccuracies.¹⁰ Many docking tools were initially conceived or adapted to generate protein–protein complexes, and specific protein-RNA scoring potentials are not always available.^2,11 A consensus-scoring approach could improve the results,³ as no protein-RNA scoring function can now reach optimal performances. Limited knowledge of protein-RNA interactions restricts the refinement process of these functions. In contrast, the situation is different for protein complexes not involving RNA, as they have been extensively characterized and studied. Many structures are available in the Protein Data Bank (PDB), and the scoring functions used to evaluate these complexes are better tuned.¹²

Given the peculiar charge distribution of RNA, another essential aspect neglected by common docking approaches is the role of water and ions,¹³ which can be crucial for maintaining structural stability and facilitating RNA-protein interactions.¹⁴

An alternative approach based on Molecular Dynamics (MD) could solve these limitations, considering both RNA and protein flexibility, simulating the system in explicit solvent and ions, thus providing a better model of a biological environment.

However, classical MD is poorly efficient in observing infrequent events such as binding and unbinding. Although reaching biochemical relevant time scales (microsecond-to-millisecond simulations) is indeed possible with current computer architectures, it still requires prohibitive computational time, at least for virtual screening purposes. Moreover, RNAs are often characterized by a rugged conformational landscape, where the system may become trapped in a metastable state. High energy barriers usually block transitions between states, leading to the preferential sampling of a state that differs from the desired one.

For these reasons, enhanced sampling protocols are routinely adopted to offset the time scale limitation of classical MD, allowing the observation of such phenomena in reduced computational time.

Many of these methods rely on the definition of the so-called collective variables (CVs), along which an energy bias is added during the simulation.¹⁵ Some relevant examples are metadynamics, steered molecular dynamics (STD), and umbrella sampling. These techniques are widely employed in diverse contexts, from folding and conformational studies to binding and unbinding simulations¹⁶

Some notable applicative examples in the context of nucleic acids include the use of Bias-exchange metadynamics to characterize the binding and unbinding of Mg²⁺ ions on an RNA duplex,¹⁷ steered Molecular dynamics (STD) to characterize the binding between TAR and a cyclic peptide¹⁵ and umbrella sampling coupled with STD to simulate the unbinding of two ligands (BMVC and BMVC0) from a G-quadruplex.¹⁸

Despite the success of these methods, selecting appropriate reaction coordinates can be challenging and system-dependent. CV-free enhanced sampling methods, instead, introduce a bias independently from collective variables. A few examples included in this category are replica-exchange molecular dynamics (REMD), accelerated molecular dynamics (aMD), temperature-accelerated molecular dynamics (TAMD), Gaussian accelerated Molecular Dynamics (GaMD). For example, in relation to nucleic acids, Gaussian accelerated Molecular Dynamics (GaMD) has been recently exploited to simulate the binding between M2-1 protein, from Human respiratory syncytial virus (HRSV), and SH7 RNA¹⁹; while replica exchange has been used in the study of a polycation-siRNA complex to identify possible binding sites.²⁰

Thermal Titration Molecular Dynamics (TTMD)²¹ also belongs to the CV-free class and, differently from the previously mentioned programs, one of the perks of this method is represented by its simplicity. TTMD aims to qualitatively compare ligand-target complexes and discriminate strong from weak binders, without aiming to reconstruct the free energy landscape of the binding-unbinding process nor to estimate the free energy of the process. TTMD’s workflow is pretty straightforward as it performs a series of MD simulations at increasing temperatures while monitoring the persistence of the original binding mode through an interaction fingerprint-based scoring function²²: the persistency of the pose discriminates strong binders from weak ones. Moreover, TTMD is a user-friendly program, and simulations can be easily carried out without a strong modeling background as they are automatically performed.

This tool has been recently applied to study the stability of RNA-small molecule complexes.²³ As a result of the positive outcome of applying TTMD to RNA-small molecule complexes, this work aims to assess the applicability domain of the technique to study the stability of RNA-peptide/protein complexes. In particular, we evaluated the application of TTMD as a postdocking filter, capable of discriminating poses with native-like binding modes from decoys based on interaction stability.

Materials and Methods

Hardware Overview

The molecular modeling studies, such as the preparation of RNA-peptide/protein complex structures, the setup for MD simulations, and trajectory analyses, were conducted on a 20 CPU Linux workstation equipped with an Intel Core i9-9820X 3.3 GHz processor. Molecular dynamics simulations were carried out on a cluster composed of 30 NVIDIA GPUs ranging from GTX980 to RTX4090.

Structure Preparation

The three-dimensional structures of the RNA-peptide/protein complexes selected for the analysis were retrieved from the Protein Data Bank (PDB).²⁴ The structures of these complexes are reported in Tables S1–S5 (Supporting Information) with their PDB IDs reported as well in Table 1. In the case of NMR structures, the first conformer was selected according to the deposited selection criteria (lowest energy for 1HJI and 5UZZ, and lowest NMR restraint violation energy for 1ETG). All structures were prepared using tools provided by the Molecular Operating Environment (MOE) suite 2022.02.²⁵

Table 1. Summary of the Investigated Complexes.

PDB accession code	peptide/protein	RNA	technique	resolution
1HJI48	nun protein	BoxB RNA	NMR
6XH049	TAR binding protein	TAR trans-activation response element	X-ray	3.1 Å
4PDB50	30S ribosomal protein S8	RNA aptamer	X-ray	2.6 Å
1ETG51	REV peptide	RRE rev responsive element RNA	NMR
5UZZ52	macrocyclic peptide (L50)	pre-miR21	NMR

The “Structure Preparation” tool of the MOE environment was utilized to address issues within the chosen protein and RNA chains. This preparation process included the addition of missing atoms, capping of N/C-termini with acetyl and N-metylamide respectively, and handling of tautomeric forms and alternate conformations of specific residues and nucleotides. Missing loops were reconstructed using a homology-based approach.

The “Protonate3D” tool was used to add hydrogen atoms and assign ionization states (pH = 7.4, T = 310 K, ionic strength = 0.154 M), determining the protonation states. This tool also calculated the most probable tautomer based on hydrogen-bonding contributions, ensuring the stability of the structure. Partial charges were computed using the AMBER14:EHT force field, and hydrogen atoms added during reconstruction underwent energy minimization using the same force field.

The prepared structures were then used for subsequent molecular dynamics and docking calculations, ensuring comprehensive and accurate preparation for computational analysis.

RNA-Protein Docking

Docking calculations, performed to generate RNA-peptide/protein complexes, were carried out using two protocols: HDOCKlite 1.1 for rigid docking and HADDOCK²⁶ 2.5 for flexible docking.

Rigid Docking

HDOCKlite1.1 is a docking program derived from the HDOCK server.²⁷ Unlike the web-based version, the local software does not use a template-based approach. It employs a hierarchical rigid sampling of the rotational and translational space between proteins and/or nucleic acids, with an angle interval of 15°, and a spacing of 1.2 Å for a Fast Fourier Transform (FFT)-based translational search. The best scoring and diverse binding modes are retained.

Flexible Docking

To incorporate the flexibility of RNA and peptide or protein counterparts, flexible docking was performed using a locally installed version of HADDOCK2.5.

HADDOCK2.5 is an integrative modeling docking program derived from its web server interface.²⁶ It can combine experimental and predicted information to guide the structure prediction of biomolecular complexes. The docking calculation is based on a triple consecutive run of docking using protocols of first rigid-body calculation, semiflexible refinement, and, in the end, an explicit solvent analysis.

The docking protocol follows the method developed by Charitou and co-workers,²⁸ designating the entire peptide or protein as passive while defining the RNA interface based on nucleotides within a 4.5 Å distance cutoff from the target in the reference PDB structures. The three-stage docking process generates 5000 poses in a initial rigid body minimization stage, followed by 400 poses during the semiflexible step, terminating with the final refinement in explicit solvent. For the subsequent molecular dynamics (MD) simulations, we employed complexes generated during the semiflexible refinement stage, since solvation and refinement were performed following our standard internal protocol for MD simulation.

Surface of Contact Calculations

The surface of contact of each generated pose and the respective crystal or NMR deposited structure has been computed as interatomic contact surface areas (CSA) through the program dr_sasa (method 1) which exploits the Shrake-Rupley algorithm.²⁹

System Setup for MD Simulations

The RNA-peptide or protein complexes for molecular dynamics (MD) simulations were prepared utilizing Visual Molecular Dynamics (VMD)³⁰ 1.9.3 and AmberTools22.^31,32 Parameters for RNA atoms were assigned using the ff14SB force field with χ modification tuned for RNA (χOL3),^33,34 while protein and peptide ligands were parametrized employing the ff14SB³⁵ force field. Each system was solvated in an orthorhombic box with a padding of 15 Å between the complex and the box boundary, using the TIP3P water model.³⁶ Sodium (Na⁺) and chloride (Cl^–) ions were added to neutralize the system and achieve a physiological salt concentration of 0.154 M. A 500-step energy minimization using the conjugate gradient algorithm was performed to eliminate clashes and bad contacts.

A two-step equilibration protocol was carried out before the production phase. The first step involved a 1 ns simulation in the canonical ensemble (NVT), with a 5 kcal mol^–1 Å^–2 harmonic positional restraint applied to each protein/peptide and RNA atom. The second equilibration step consisted of a simulation of equal length in the isothermal–isobaric ensemble (NPT), where the same restraint was applied only to the RNA and peptide/protein backbone. During the NPT phase, the pressure was maintained at 1 atm using a Monte Carlo barostat,³⁷ while the temperature, for all the equilibration phases, was kept constant at the initial (i.e., lowest) TTMD temperature value through a Langevin thermostat.³⁸

All MD simulations were conducted using the ACEMD 3.5³⁹ engine based on the open-source Python library OpenMM.⁴⁰ An integration time step of 2 fs was employed, and the M-SHAKE algorithm⁴¹ was utilized to constrain bonds involving hydrogen atoms. Long-range electrostatic interactions were calculated using the particle-mesh Ewald method,⁴² applying a 9.0 Å cutoff for Lennard–Jones and real-space electrostatic interactions, and a switching distance of 7.5 Å for Lennard–Jones interactions.

TTMD

Thermal Titration Molecular Dynamics (TTMD)^21,43 is an enhanced sampling molecular dynamics protocol, developed for evaluating complex unbinding kinetics,⁴³ initially used for protein–ligand complexes and also applied to RNA-small molecule structures. TTMD operates through a series of short classical MD simulations, performed at progressively increasing temperatures, defined as “TTMD-steps”. Thus, the kinetic energy of the system is progressively increased, thereby reducing the simulation time required to observe unbinding events compared to conventional MD simulations.

In this study, TTMD simulations were carried out using Python 3.10, employing libraries such as NumPy, MDAnalysis,^44,45 and Scikit-learn. The starting temperature was set at 300 K, and the simulation proceeded in 10 K increments up to the final temperature of 450 K. Each simulation window, or TTMD-step, lasted for 10 ns.

During the simulation, the preservation of the native binding mode was assessed using an interaction fingerprint-based scoring function (IFP_cs).²²

This metric, described extensively in the work of Pavan et al.,²¹ evaluates the conservation of the starting binding mode throughout the MD simulation. In particular, for nucleic acids, Prolif fingerprint package is used to determine the presence of nine types of interaction (Hydrophobic, HBDonor, HBAcceptor, PiStacking, Anionic, Cationic, CationPi, PiCation, VdWContact), based on geometric thresholds, between each nucleotide of the nucleic acid and the peptide/protein.

This information is then organized into a matrix (i.e., interaction fingerprint), which is further processed and compared through cosine similarity to the reference one (i.e., computed from the last frame of equilibration), to evaluate changes in the binding mode.

TTMD Trajectory Analysis

During a TTMD simulation, each frame generated is analyzed using the ProLIF⁴⁶ Python package to compute the interaction fingerprint between the two molecular entities of the complex. The fingerprint is then compared for each frame to the last frame of the second equilibration phase using the cosine similarity metric, which helps assess the maintenance of the original binding mode during the simulation. The similarity value is multiplied by −1 to provide a negative score range, where 0 represents a complete deviation from the reference pose, i.e., loss of the refence binding mode. The interaction fingerprint cosine similarity (IFP_cs) score thus ranges from −1 to 0, where −1 indicates complete retention of the original binding mode, and 0 signifies a total loss of native interactions. If the average IFP_cs score, calculated on the last 10% of each TTMD-step, exceeds −0.05, indicating that all native interactions have been lost, the simulation is stopped.

At the end of the simulation, each average IFP_cs score for each TTMD-step is plotted in the “Titration Profile” graph, as well as the respective temperature. In this graph the “MS coefficient” is also reported, which is defined by the slope of the line connecting the origin of the axes and the last point of the graph, and it is an indicator of the complex stability. This coefficient can range from 0 to 1, with 0 representing the conservation of the binding mode and 1 indicating the loss of the initial interactions in the first step of the simulation. The MS score is calculated using the following equation.

1

Five independent TTMD simulations were executed for each examined complex, and the average MS coefficient was computed over three replicates after discarding the ones with the highest and the lowest values.

In addition, a “Titration Timeline” plot is generated by the automatic TTMD procedure, illustrating the time-dependent evolution of IFP_CS scores throughout the simulation. This plot also includes the root-mean-square deviation (RMSD) of protein backbone, of RNA backbone and of the backbone of nucleotides belonging to the binding site (i.e., in contact with the peptide/protein; defined as Bsite in the graph), calculated using the MDAnalysis Python package, providing a comprehensive view of the dynamics and stability of the complex during the simulation.

eRMSD Calculations

All the eRMSD calculations were performed using Barnaba⁴⁷ software and the eRMSD plots obtained are available in Zenodo at the link provided in the Data Availability section.

Results

To assess the applicability of TTMD to the study of RNA-peptide complexes, we applied this technique to five test cases of pharmaceutical relevance, selected based on the availability of an experimental structure in the Protein Data Bank (PDB) and spanning over different “ligand” classes, ranging from cyclic peptides to proteins, mainly focusing on peptides.

In particular, the test cases included the following complexes: Nun protein-BoxB RNA (PDB ID: 1HJI), TAR binding protein-HIV-1 TAR RNA (PDB ID: 6XH0), 30S ribosomal protein S8-RNA aptamer (PDB ID: 4PDB), REV peptide-RRE RNA (PDB ID: 1ETG) and macrocyclic peptide-pre-miR21 (PDB ID: 5UZZ).

1HJI, 1ETG, and 5UZZ were chosen to assess the technique’s performance with RNA-peptide complexes with the latter characterized by a cyclic peptide, while 4PDB and 6XH0 were used to explore the limits of TTMD application with RNA-protein complexes.

A concise outline of the selected cases, summarized in Table 1, is reported hereafter.

Nun Protein

Bacteriophages rely on the bacteria machinery to replicate their genomes, synthesize viral proteins, and generate offspring. They can also compete for the same host, and in this scenario, some molecular mechanisms enable a lysogenic phage to exclude superinfecting phages. In particular, HK022 and superinfecting λ phage are involved in a well-studied exclusion mechanism, where HK022-encoded Nun protein acts as a transcription terminator for λ early genes, competing with the λ phage-encoded N protein for the major groove-binding of the stem-loop RNA boxB.⁵³ These complexes have been extensively studied as models to elucidate protein-RNA recognition,⁵⁴ as they both share on the protein side an arginine-rich motif (ARM) like other RNA-binding proteins such as Tat and Rev.⁵⁵ In this paper, a nun/boxB complex, deposited under the PDB ID: 1HJI,⁴⁸ has been considered.

BoxB RNA is characterized by a hairpin conformation closed by an apical tetraloop (GNRA), and in the deposited structure, the peptide assumes a bent α-helix conformation upon binding to the major groove, contacting even portions of the apical loop,⁴⁸ with a predicted surface of contact of 561.76 Ų.

TAR and REV

Human immunodeficiency virus of type 1 (HIV-1) is the causative agent of the acquired immunodeficiency syndrome (AIDS). Despite the significant efforts directed to treat and prevent this retrovirus, its eradication is still hampered by the absence of an effective vaccine and by the limited capability of Antiretroviral Therapy (ART) to completely remove the virus from the host,⁵⁶ as the presence of latently infected cells could be a source of persistent residual viremia⁵⁷ and long-term related comorbidities can occur.⁵⁸ In this scenario, transcription inhibitors could play a key role in controlling viral reactivation, thus reducing viral reservoir. Among the targets involved in the transcription phase, the TAT-TAR complex is one of the most attractive for drug development.⁵⁹ The elongation of HIV-1 transcription is promoted by host CDK9/cyclin T1, recruited by HIV-1 transactivation response (TAR). TAR is activated by the Tat protein, which is essential in this process as, in its absence, the transcription results in short and abortive transcripts.⁶⁰ Not surprisingly, the TAR complexes deposited in the PDB mainly include peptidomimetics and cyclic peptides specifically designed to inhibit the formation of the TAR-Tat complex. In this work, a complex that involves a TAR binding protein, deposited under the PDB ID: 6XH0,⁴⁹ has been selected. This case study involves a protein-RNA complex, and in this structure, TAR is characterized by an A-form helical stem-loop with a central trinucleotide (UCU) bulge and an apical hexaloop, with the major groove and the central bulge being involved in binding with the β2-β3 connecting loop of the protein,⁴⁹ with a predicted surface of contact of 771.90 Ų.

TAR-TAT complex is not the only target that emerged as a potential candidate against the latent form of HIV-1: REV-RRE complex has also gathered significant attention as it is involved in the process that leads to the translation of structural proteins.⁶¹ HIV-1 REV is a regulatory protein that recognizes the Rev response element (RRE), a ∼350 nucleotide sequence located in unspliced and incompletely spliced viral mRNA transcripts. The REV-RRE oligomeric recruits host Crm1/Ran-GTP nuclear export machinery to facilitate the export of these mRNAs from the nucleus to the cytoplasm.^62,63 Subsequently, after the complex dissociation in the cytoplasm, viral RNAs are packaged for new virions or translated into essential viral proteins.^62,64

The interaction between a purine-rich bulge stem-loop IIb of RRE and the arginine-rich α-helix binding domain of REV drives the recognition between Rev and RRE. In the present study, a complex available under PDB code 1ETG(51) and presenting RRE complexed with a REV peptide portion that includes the arginine-rich motif (ARM) has been selected, prepared, and subjected to docking and TTMD pose refinement.

RRE is characterized by a branched structure with multiple stem-loops and bulges. In the deposited structure, the REV peptide, presenting an α-helical conformation, interacts with the stem-loop IIB of RRE, binding the major groove near a purine-rich internal loop,⁵¹ with a predicted surface of contact of 964.49 Ų.

Bacillus anthracis Ribosomal Protein S8

The bacterial S8 ribosomal protein has been studied and taken as a model for the characterization of RNA-protein interaction. This RNA-binding protein contributes to the correct folding of the central domain of 16S rRNA.⁶⁵ Apart from this function in the ribosome assembly, it has a role in the repression of the translation of the genes encoded by the spc operon mRNA, including the ones of S8 itself and ten other ribosomal proteins.⁵⁰ Although greater affinity is commonly attributed to the 16S rRNA site, the region targeted on the mRNA is very similar,⁶⁶ and many interactions between the S8 protein and the two binding sites are the same.⁵⁰ The study of the interaction between RNA and S8 has also been addressed through aptamers: Davlieva et al.,⁵⁰ through a SELEX experiment, discovered an RNA aptamer able to bind the Bacillus anthracis S8 protein with high affinity. They aimed to identify RNA sequences that, while retaining the ability to bind S8 protein, do not present the conserved features of helix 21 (16S rRNA),⁵⁰ and resulted in a valuable insight into the ability of the selected aptamer to adopt a different secondary structure upon binding, stressing the importance of RNA flexibility into RNA-protein recognition. To investigate the impact of RNA flexibility on TTMD performances, we decided to include this complex, registered with the PDB ID: 4PDB, performing both rigid and semiflexible docking. In this structure, the aptamer assumes a hairpin conformation, with the A-form helical stem interrupted by an internal loop, while the S8 protein presents an N-terminal domain with an αβαββ fold and a C-terminal comprising an α-helix followed by an antiparallel four-strand β-sheet. The latter portion is involved in most of the contacts with the RNA aptamer through residues located in the turns at the ends of the β-strands.⁵⁰ The overall predicted surface area is 879.87 Ų.

The results of prior work from our group also guided the selection of this test case, as a binding simulation between the aptamer and the protein was successfully performed through Supervised Molecular Dynamics (SuMD), with the last frame of the trajectory converging to the crystal structure.⁶⁷

Pre-miR21

The overexpression of miR21 is implicated in the development of cancer. In many solid tumors, it is linked with cell proliferation, differentiation, and apoptosis, contributing to tumor growth, invasion, and metastasis.⁶⁸ After the first cleavage of the primary miRNA transcripts inside the nucleus, the pre-miRNA is exported into the cytoplasm, where it is further cleaved at the apical portion of the stem-loop by another RNase III enzyme named Dicer. While the passenger strand of the obtained duplex is degraded, the guide strand forms the miRNA-induced silencing complex,⁶⁹ which targets complementary RNAs, leading to mRNA cleavage, degradation, and translational repression.⁷⁰

For these reasons, miR21 is investigated as a potential tumor biomarker for diagnostic purposes. Moreover, targeting miRNA-21 could provide adjuvant therapy for tumors like glioblastoma, in which this RNA acts like an oncogene capable of blocking apoptosis.⁶⁸

However, inhibition of miR-21 using small molecules has had limited success. A cyclic peptides library screening led to the identification of a cyclic peptide (L50) that can bind pre-miR21 at the Dicer cleavage site,⁵² and the same authors also deposited the structure of the complex under the PDB ID: 5UZZ.⁵² In this complex, the cyclic peptide binds the major groove at the junction between the single-stranded RNA loop and the double-stranded RNA,⁵² with a predicted surface of contact of 636.99 Ų.

In the present work, starting from different docking poses generated through HDOCKlite, the TTMD protocol has been applied to a pharmaceutically relevant target like miR21, evaluating the behavior of the technique in the challenging field of cyclic peptides.

Simulation Workflow

For each RNA-peptide complex, different docking poses were generated by exploiting HDOCKlite and HADDOCK programs, and the obtained structures are reported in Tables S1–S5 (Supporting Information). These two different algorithms were considered to compare a fully rigid docking procedure (HDOCKlite) to one including some extent of flexibility (HADDOCK). The only exception was represented by structure 5UZZ, characterized by the presence of a cyclic peptide with a d-Proline, that was not easily managed with HADDOCK, which was thus excluded from the comparison.

The five top-scoring docking poses were selected for each protocol and subjected to TTMD simulations. When the pose with the lowest RMSD to the experimental reference (i.e., the experimental-like pose) was not included in the top five structures, a lower-ranked experimental-like pose was retrieved, regardless of its ranking position, and subjected to TTMD as well. The results were compared to reference TTMD simulations performed on the experimental structure, to assess the complexes stabilities and evaluate the capability of TTMD to discriminate experimental and experimental-like poses from decoys.

As an example, the poses for 6XH0 are reported in Table 2, while the docking poses for the remaining test cases are reported in Tables S1–S5 (Supporting Information). In this case, as regards HDOCKlite, the 5 top scoring poses have been selected, with the crystal-like pose being at the top of the ranking. In the case of HADDOCK, the 5 top scoring poses were selected as well, but they did not include the crystal-like pose, which was instead at the 212th position of the ranking, and additionally submitted to TTMD simulations.

Table 2. Docking Results of 6XH0.

^aThe docking pose of 6XH0, obtained exploiting HDOCKlite and HADDOCK, and the structure deposited in the PDB along with its assertion code, are reported. The RNA is colored in Blue while the protein is pictured orange.

Scoring Results

The Prolif fingerprint package, according to the experience gained in the previous application of TTMD to RNA-small molecule complexes,²³ was adopted to compute the RNA-peptide interaction fingerprints. In TTMD, the cosine similarity (IFP_cs) between the interaction fingerprints during the simulation and the reference position (end of equilibration) is used to monitor the persistence of the binding mode during time. In this work, the starting temperature of the simulation was set to 300 K, the end temperature to 450 K, and the duration of each TTMD-step to 10 ns with a temperature increase of 10 K between each step. The TTMD MS coefficient, i.e., the slope of the line connecting the IFP_cs at the beginning and end of the simulation (and deeply explained in the method section), was adopted to rank the poses and characterize the conservation of the binding mode between the start and the end of the simulation, with lower MS values representing more stable poses.

Tables 3 and 4 summarize the TTMD results, for each test case, indicating the average MS values calculated from the TTMD simulations, while the whole set of results is reported in Tables S6–S10 (Supporting Information).

Table 3. TTMD Simulation Results for Each Significant Replica Generated with HADDOCK^a.

HADDOCK
6XH0		1ETG		1HJI		4PDB
pose	mean MS	pose	mean MS	pose	mean MS	pose	mean MS
crystal	0.0038	NMR	0.0041	NMR	0.0041	crystal	0.0032
1	0.0048	1	0.0060	1	0.0082	1	0.0036
2	0.00516	2	0.00596	2	0.00806	2	0.00388
3	0.00529	3	0.00573	3	0.00665	3	0.00417
4	0.00488	4	0.00552	4	0.67135	4	0.00456
5	0.00580	5	0.00465	5	0.00549	5	0.00444
C-like (Pose212)	0.0041	NMR-like (Pose247)	0.00535	NMR-like (Pose73)	0.0043	C-like (Pose317)	0.0038

^aThe PDB accession code, the docking program used, and the average MS score calculated, are reported for each test case simulated. The mean MS value for the experimental structure and experimental-like pose, the docking protocol, and the PDB code of each case are highlighted in bold.

Table 4. TTMD Simulation Results for Each Significant Replica Generated with HDOCKLite^a.

HDOCKlite
6XH0		1ETG		1HJI		4PDB		5UZZ
pose	mean MS	pose	mean MS	pose	mean MS	pose	mean MS	pose	mean MS
crystal	0.0038	NMR	0.0041	NMR	0.0041	crystal	0.0032	NMR	0.0052
1	0.0038	1	0.0054	1	0.0042	1	0.0041	1	0.0055
2	0.00503	2	0.00614	2	0.00710	2	0.01154	2	0.00563
3	0.00615	3	0.0053	3	0.00517	3	0.066865	3	0.00561
4	0.00532	4	0.00529	4	0.00464	4	0.00524	4	0.00911
5	0.02103	5	0.00484	5	0.0040	5	0.01168	5	0.00551
C-like (Pose1)	0.0038	NMR-like (Pose52)	0.0045	NMR-like (Pose1)	0.0042	C-like (Pose1)	0.0041	NMR-like (Pose1)	0.0055

^aThe PDB accession code, the docking program used, and the average MS score calculated, are reported for each test case simulated. The mean MS value for the experimental structure and experimental-like pose, the docking protocol, and the PDB code of each case are highlighted in bold.

Although the complete unbinding event was never sampled during the simulations, applying TTMD as a postdocking refinement tool with this set of test cases has brought positive results since the stability of the simulated complexes, evaluated as the retention of the binding mode throughout the simulation, has been assessed. In detail, all the experimental structures have lower MS values when compared to the derived docking poses, regardless of the protocol used to generate them. The only exception is represented by pose number 5 (RMSD 5.02 Å) of 1HJI, generated through the HDOCKlite rigid docking approach, which has an MS score slightly lower than the NMR and the NMR-like structure (pose 1, RMSD 1.63 Å), but still comparable to them.

For most of the selected test cases, the technique can assign the lowest MS value, among the different poses, to the experimental-like predicted structures. Detailed information about the docking scores for each simulation and RMSD values for each docking pose can be found in Tables S1–S5 (Supporting Information).

Remarkably, even when exploiting a semiflexible docking protocol, like HADDOCK, TTMD successfully distinguished experimental-like poses from decoys in all selected cases except for 4PDB and 1ETG.

All HADDOCK poses obtained for 30S ribosomal protein S8 (4PDB) have reasonably low and close RMSD values as shown in Table S4 (Supporting Information), and TTMD failed to discriminate their subtle differences. In this situation, TTMD identified the crystal pose as the most stable, but, among the docking poses, the lowest MS was assigned to pose 1 (RMSD of 2.52 Å) rather than pose 317 (RMSD of 1.82 Å). Considering that both RMSD values are below the crystallographic resolution (2.60 Å), the two poses can be described as crystal-like, and TTMD still discriminated them from decoys.

In the case of REV peptide (1ETG), TTMD assigned the lowest MS score to the NMR pose, while the second MS score was not assigned to the NMR-like pose (pose 247), RMSD of 5.62 Å, but to pose 5 (RMSD of 7.28 Å). However, both poses are topologically similar to the NMR structure and are assigned the lowest MS scores among the docking poses.

Despite these situations, the technique was able to identify experimental-like poses for the remaining cases. 6XH0 and 1HJI are particularly significant since, following HADDOCK’s ranking, most of the top poses are characterized by high RMSD. Therefore, the application of TTMD and the evaluation of the MS score are valuable in rescoring HADDOCK’s poses, identifying the more experimental-like poses regardless of the ranking provided by the HADDOCK score.

On the other hand, when used starting from the HDOCKlite-generated poses, TTMD obtained positive results, isolating the experimental-like pose in almost all the cases apart from 1HJI. As mentioned above, the NMR structure, pose 1 (NMR-like pose), and pose 5, share low and close MS score values, 0.0042, 0.00423, and 0.00414 respectively. The technique still manages to identify a stable pose among the others, but it also describes pose number 5 as the most stable. Intriguingly, pose 1 and 5 have the lowest RMSD values among the generated ones. Moreover, as observable in the Titration Timeline graphs Figure S1 (Supporting Information), the fluctuation of the IFP_cs is higher for pose 5, whereas, for pose number 1 and the NMR reference, it is steadier over time. The IFF score²³ a previously reported additional TTMD scoring metric evaluating the fingerprint fluctuation around the mean IFP_cs value, provides further support, suggesting a slightly stabler fingerprint fluctuation for the crystal-like pose, with an IFF value of 0,8 compared to 0,9 attributed to pose 5. The MS score is calculated on the last TTMD-step, as explained in the Materials and Methods section. For this reason, at the end of the simulation, rearrangements can lead the complex to assume native interactions previously lost during the initial part of the simulation, leading to a slightly lower MS score.

Although this docking program was indeed able to identify the NMR-like pose among the others, for 5UZZ the MS scores are high and close, indicating a lower resolution of TTMD arguably determined by a suboptimal temperature ramp for this test case.

Differently from the HADDOCK-originated poses of 1ETG discussed above, in the case of HDOCKlite, TTMD managed to identify the NMR-like pose among the others even if it was out of the five top scoring poses.

Notably, TTMD was able to seamlessly identify the crystal-like pose among the ones generated for 4PDB, as, in contrast with HADDOCK, they are much more diversified, as can be observed in Table S4 (Supporting Information).

This specific test case was also selected because an enhanced sampling MD technique called Supervised Molecular Dynamics (SuMD) has already been used to simulate the binding event with this test case reaching a near-crystal complex.⁶⁷ This achievement opens the possibility of using SuMD as an alternative tool for the complex generation, coupling it with TTMD to score the obtained structures, an approach already tested for a Berenil-dodecanucleotide complex.⁷¹

Starting from the last frame of the SuMD trajectory retrieved from a previous study, we also performed a TTMD simulation as a proof of concept. As shown in Table 5, despite being higher than the crystal one, the average MS score, calculated from the TTMD simulations carried on the last frame of the SuMD trajectory, was lower than those computed from the HDOCKlite and HADDOCK-generated crystal-like poses. Given the encouraging result, SuMD seems able to generate stable crystal-like complexes. Despite the need for further validation and the long computational times required, this technique could represent a viable alternative to standard docking approaches for complex generation.

Table 5. TTMD Simulation Results of the Last Frame from the SuMD Trajectory.

^aThe PDB accession code, the MD protocol or docking program used, the MS score for each simulation, and the average MS coefficient are reported for 4PDB. The replica with the nearest MS value to the average is highlighted in bold.

TTMD Simulation Analysis

For each test case, an in-depth description of the TTMD simulation is presented hereafter, comparing experimental structures with experimental-like ones and the MS worst-performing pose.

1ETG

The NMR structure remains stable throughout the simulation, obtaining the lowest MS value, and the nucleic acid globally retains the native structure even at high temperatures as can be evinced by low to medium eRMSD⁷² values, which rise with temperature, just like the following poses. Interactions mediated by key RNA residues U66, G67, U45, G46, G47, and A73, as identified in the original NMR structural analysis, are consistently preserved during the simulation. Despite minor conformational changes in the RNA, the key residues within the binding site remain geometrically stable, likely due to the stabilizing effect of the peptide, which helps maintain the overall integrity of the complex.

Similarly, the NMR-like structures generated by both HDOCKlite and HADDOCK exhibit stable interactions between the peptide and RNA, with key interactions involving U66, G67, U45, G46, G47, and A73 remaining even as the temperature increases.

However, for the MS worst-performing poses from both HDOCKlite and HADDOCK, the interactions deviate from the NMR-like configurations. These poses show a significant reduction in the complex stability throughout the simulation, with an increased peptide flexibility and mobility, and a partial loss of the initial interactions.

1HJI

The NMR structure remains stable throughout the simulation, achieving a low MS value. Key interactions mediated by C2, C3, C4, and A7, as previously described by Faber et al.,⁴⁸ are largely preserved. Specifically, the stacking interaction between A7 and Tyr39 on the peptide is consistently maintained, supporting binding mode stability.

The HDOCKlite NMR-like structure similarly maintains a low MS score and stable interactions mediated by C2, C3, C4, and A7, including the critical A7-tyrosine stacking interaction, which undergoes minor fluctuations at higher temperatures. The HADDOCK NMR-like structure also preserves a stable binding mode, though it partially loses the A7 stacking interaction.

In contrast, HDOCKlites’s least stable pose quickly loses peptide-RNA interactions due to poor shape complementarity, altered by the increasing temperatures. Similarly, HADDOCK’s least stable pose loses stability early in the simulation, leading to partial peptide dissociation, which impacts the RNA stability as, for this test case, it is the only simulation reported characterized by a rapid drift toward medium eRMSD values.

5UZZ

In the NMR structure, the peptide ligand (L50) binds to the major groove of the RNA helix, where it interacts primarily with residues such as U27, G28, and A29, also interacting with residues of the apical loop as C39 and U40.

In the study of Shortridge⁵² Arg12 was defined as essential for recognition and binding within the core of the binding cavity, and during the simulation of the NMR structure, these interactions are maintained despite the high dynamic behavior of the complex, including Arg12 starting interaction with G28 and A29.

NMR structure and NMR-like pose share the same RNA-peptide conformation, which guarantees common interactions during the simulations like Arg12-G28 and Arg12-A29, while the MS worst-pose presents the L50 peptide flipped inside the major groove. The starting position falls within the same binding site region of the NMR and NMR-like pose and ensures common starting contacts, but the interactions are not mediated by the same amino acid residues. During the simulation, the peptide drifts and consequently loses part of the starting interactions with a consequent worsening of IFP_cs. These interactions are mediated by Arg12, which interacts with the apical loop at C39 and U40 and then moves toward U6 and G7, while Arg8, at the beginning of the simulation, is solvent exposed rapidly heading to the apical loop.

Globally the eRMSD values for these complexes increase during the simulations reaching medium/high values, instability likely due to a suboptimal temperature ramp.

4PDB

The crystal structure is the most stable one, obtaining the lowest MS value. Interactions mediated by residues C16, C17, A24, U25, A26, U27, and C28, described by Davlieva et al.,⁵⁰ are visibly retained during the simulation.

The crystal-like HADDOCK-generated structure is also stable overall, but toward the end of the simulation, where temperatures are highest, the C28-Tyr88 interaction is lost, contributing to a higher MS score. The U27-Glu126 hydrogen bond also experiences fluctuations, becoming less stable in response to the dynamic changes.

The crystal-like structure generated through HDOCKlite displays a similarly low MS value, comparable to that achieved with the other docking protocol. The simulation reveals some translational movement of the protein in relation to the initial bound state, with coordinated movements in the protein followed by a slight reduction in IFPcs, as shown in Figure S2 (Supporting Information). In this case, the U27-Glu126 interaction is largely preserved, while the hydrogen bond between the hydroxyl group of Ser107 and A26 fluctuates following the geometrical stability.

The worst-performing pose of HDOCKlite (pose 4) is highly unstable from the beginning of the simulation, leading to an early loss of the initial protein-nucleic interactions that are indeed highly divergent from the crystal ones.

The worst-performing HADDOCK-generated pose is stabler than the last described one, as it is still similar to the experimental structure (i.e., low RMSD values, as all the poses generated with this technique), and therefore it presents crystal-like interactions that are retained almost until the end of the simulation. Lastly, the high temperatures favor the movement of the protein leading to the partial loss of interactions and, as a consequence, worsening of the IFPcs, hinting at the sensitivity of the technique.

Both RMSD stable and unstable complexes have a temperature-dependent eRMSD trend starting from low values and reaching medium/high ones.

6XH0

The crystal structure exhibits stability throughout the simulation, achieving a low MS value. The integrity of the binding mode is primarily governed by key interactions mediated by RNA residues G26, G28, and G36. Arginines, in particular Arg47, Arg49, and Arg52, form essential interactions within the TAR major groove. Furthermore, Gln48, located in the β2-β3 loop, contributes to the stabilization of the RNA phosphate backbone, while Thr50 and Gln54 support the structural conformation of the loop by forming intrapeptide hydrogen bonds. Collectively, these interactions form a robust network that ensures the sustained stability of the binding mode throughout the simulation.

The crystal-like structures generated by both HDOCKlite and HADDOCK maintain low MS scores and exhibit similar binding modes, with preservation of the critical interactions with G26, G28, and G36. The persistence of these key interactions contributes to the stability of the binding mode during the simulations.

The least stable pose generated by HADDOCK shows only partial interactions with key RNA residues. This results in an unstable binding mode with structural fluctuations throughout the simulation. Similarly, the lowest-performing pose generated by HDOCKlite lacks the key native interactions with the RNA, resulting in the destabilization of the binding configuration starting from the initial stages of the simulation. All the complexes have low eRMSD values at the start of the simulation, rising to medium values, accordingly to the temperature increase.

Discussion and Conclusion

This work showcases the first application of Thermal Titration Molecular Dynamics (TTMD) to RNA-peptide complexes.

This Python-based tool was initially developed for the qualitative estimation of protein–ligand unbinding kinetic thanks to the analysis of the conservation of the interaction fingerprint during a series of consecutive MD simulations performed at increasing temperatures. In this paper, we employed this program, previously adapted to the study of RNA-ligand complexes,²³ as a postdocking refinement tool that identifies experimental-like stable poses according to the conservation of the original binding mode, separating them from decoys.

Five test cases have been identified, basing the decision on structural differentiation and pharmaceutical relevance, and state-of-the-art docking tools like HDOCKlite and HADDOCK have been used to generate the poses starting from the selected deposited structures. We decided to carry out both rigid and semiflexible docking approaches to assess whether a fast posing could affect or invalidate the TTMD capability to identify experimental-like poses, rather than a sophisticated yet slower method, like HADDOCK, that keeps into account local flexibility and generates refined docking poses.

In a previous study, RNA-small molecule complexes were characterized through TTMD using different temperature scales and alternative Python packages to compute the interactions, resulting in the definition of a protocol feasible for nucleic acids.²³ In this work, TTMD simulations were performed adopting the same protocol (see Materials and Methods).

Globally, examining Tables S6–S10, TTMD seems capable of refining docking results, being able, in most cases, to discriminate experimental or experimental-like poses, from docking poses that consistently differ from experimental structures. As an example of the discriminatory capability of the technique, Videos S1–S4 (Supporting Information) report the TTMD simulations of the crystal structure of 6XH0 compared respectively to the following poses of the same test-case: HADDOCK-generated crystal-like pose (lowest RMSD) and pose 5 (highest MS), HDOCKlite-generated pose 1 (lowest RMSD) and pose 5 (highest MS). As can be observed in the videos, the crystal structure and the crystal-like poses maintain structural integrity despite the high temperatures, while the structural instability is particularly amplified by the rising temperatures for noncrystal-like poses.

Subtle differences between the poses, such as sharing the same interactions and having low RMSD, can reduce the technique’s resolution, like in the scenario where all the poses under discussion may be described as crystal-like, for instance, poses with RMSD below crystallographic resolution, a situation close to the one reported for 4PDB.

The rigid docking approach seems optimal when no local rearrangements are expected upon binding, and the structural data is available like in the current validation, even considering the short computational times: just a few minutes in the case of HDOCKlite. In fact, with this protocol, for each case a native-like pose with RMSD below 3 Å was obtained, and, except in one case, it coincided with the top-scoring pose. On the other side, in a hypothetical scenario where structure prediction programs are exploited due to the absence of a deposited structure and MD pose refinement is needed,⁷³ it is not trivial to get such outstanding performances with a rigid docking approach, thus following a semiflexible docking procedure, like HADDOCK, could be a better choice. Still, it is important to consider that while HADDOCK’s posing seems robust and preferable in specific situations, aside from the longer computational times required (hours), its scoring capabilities are not equally effective, with native-like poses poorly ranked, sometimes even above the 100th position, as shown in Tables S6–S10 (Supporting Information).

Customizing the parameters used to execute HADDOCK could be beneficial in this scenario, but finding the optimal protocol could be difficult. In this circumstance, TTMD could be a critical resource since, as discussed previously, it can discriminate native-like poses from decoys and refine the docking ranking results, at least for the selected test cases. However, coupling TTMD to a docking protocol like HADDOCK, and implementing it in a high-throughput pipeline is still hampered by the long computational times required to carry out MD simulations. Consequently, further testing is needed to determine a robust docking setup, and additional TTMD code optimization and hardware progress are also necessary to reach the computational power required to execute such workflow rapidly. Hopefully, the rate of recent tech development suggests that near-future applications of enhanced sampling methods like TTMD on a large scale will be feasible.⁷⁴

Alternatively, rescoring or consensus approaches are an option to reduce the number of poses to attention and refine with TTMD, or the top poses generated through rigid docking, followed by MD relaxation, could be used instead. Despite the long simulation time, generating only a few complexes through tools like SuMD (Supervised Molecular Dynamics) could also be a viable solution, although it has not been thoroughly explored. The encouraging results in obtaining stable near-native complexes at the final state of the simulation, as mentioned for 4PDB in the Results section, pave the way for explorative future applications of the technique.

Glossary

Abbreviations

GPU

graphic processing unit

PDB

Protein Data Bank

SuMD

supervised molecular dynamics

TTMD

thermal titration molecular dynamics

Data Availability Statement

All molecular structures utilized in this work have been retrieved from the publicly available PDB. The structure of the SuMD simulation last frame of 4PDB, all the TTMD simulations, and the configuration files are available at 10.5281/zenodo.14056191. TTMD Python code to reproduce the simulations performed in this work is available at https://github.com/molecularmodelingsection/TTMD and released under a permissive MIT license.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.4c01393.

• Docking and TTMD results for all the test cases and titration timeline comparison between HDOCKlite-generated pose 1 and pose 5 of 1HJI (PDF)

• Comparison between the TTMD replica of the crystal and HADDOCK-generated crystal-like pose for 6XH0 (MP4)

• Comparison between the TTMD replica of the crystal and HADDOCK-generated pose 5 for 6XH0 (MP4)

• Comparison between the TTMD replica of the crystal and HDOCKlite-generated crystal-like pose for 6XH0 (MP4)

• Comparison between the TTMD replica of the crystal and HDOCKlite-generated pose 5 for 6XH0 (MP4)

Author Contributions

The manuscript was written through the contributions of all authors. All authors have approved the final version of the manuscript. A.D. and G.N. contributed equally.

The author(s) declare financial support was received for the research, authorship, and/or publication of this article. This work has been funded by the European Union – Next-Generation EU (“PNRR M4C2-Investimento 1.4-CN00000041”).

The authors declare no competing financial interest.

Special Issue

Published as part of Journal of Chemical Information and Modelingspecial issue “Editing DNA and RNA through Computations”.