Nonadditive Conformational Effects of tRNA Core Modifications

Abstract

Post-transcriptional modifications are known to regulate the structure and function of tRNA molecules. Compared to modifications in the anticodon region, the mechanistic effects of modifications in the core region of tRNA remain poorly understood. Using all-atom molecular dynamics (MD) simulations on yeast tRNA^Phe, we investigated the conformational effects of incorporating chemical modifications in the core region (termed hereafter “core modifications”) of tRNA. Specifically, we focused on three core modifications, namely, 7-methylguanosine (m⁷G), 5-methylcytosine (m⁵C), and 1-methyladenosine (m¹A). We discovered that individual modifications increased the structural stability of tRNA, but their combinations produced multimodal free energy landscapes with multiple distinct basins. We identified m¹A₅₈ as a bifunctional regulator that locally disrupts structure while globally coordinating anticorrelated arm movement. The combination of m¹A₅₈ and m⁵C₄₉ exhibited the lowest structural deviation from the crystal structure, while the triply modified variant incorporating all three modifications produced the most integrated residue community network, reflecting nonadditive allosteric coupling that cannot be predicted from any pairwise combination. Critically, core modifications established extensive allosteric pathways that prevent the consistent inward displacement of the anticodon stem loop otherwise observed in all simulations of chemically unmodified tRNA. Our results demonstrate that core modifications function as regulators of conformational ensembles rather than simple stabilizers.

Introduction

Transfer RNA (tRNA) is an essential adaptor molecule that decodes genetic information in cells by delivering cognate amino acids to the ribosome for protein synthesis. , Beyond its canonical role in translation, tRNA has emerged as a key regulator of gene expression, cell signaling, and cellular stress responses. Importantly, tRNA dysregulation contributes to the development of oncogenic diseases, neurodegeneration, and viral infections. ,

One of the most important determinants of tRNA structure and function is the ability of tRNA to undergo a diverse set of post-transcriptional modifications, usually between 10 and 20 modifications per tRNA molecule. , Eukaryotic tRNAs carry approximately 13 modifications per molecule. These modifications have been identified across all structural domains of the canonical L-shaped tRNA structure, which is composed of an aminoacyl acceptor stem (AAS), anticodon stem loop (ASL), and a conserved structural core formed by the D-arm or D-stem loop (DSL) and T-arm or the T-stem loop (TSL), and a variable loop (VL) (Figure ). Collectively, these modifications impact tRNA stability, decoding accuracy, aminoacylation fidelity, and ribosome binding. −

1.

(A) Secondary structure of yeast tRNA^Phe highlighting core modifications: m⁷G₄₆ or G* (red), m⁵C₄₉ or C* (blue), and m¹A₅₈ or A* (green). The architecture is composed of five structural domains color-coded as follows: AAS (purple), DSL (tan), ASL (magenta), VL (gray), and TSL (orange). Three nucleotides at positions 34, 35, and 36 that form the anticodon are highlighted in light blue. (B) Canonical L-shaped tertiary structure (PDB ID: 1EHZ) of tRNA is shown with a distinct color for each of the five structural domains. Three modified nucleotides are depicted in red (m⁷G₄₆), blue (m⁵C₄₉), and green (m¹A₅₈) space-filling representations.

tRNA modifications can be broadly classified based on their locations in the structure. The first group is composed of ASL modifications, such as wybutosine, methylcytosine, and pseudouridine, which have been extensively studied using both experimental and computational methods. ,− MD simulations of the human tRNA_UUU showed that ASL modifications alter hydrogen-bonding patterns and loop conformation within the ASL. Additionally, the ASL modifications impact tRNA-mRNA recognition, with the fully modified wild-type molecule exhibiting greater structural stability than variants lacking individual modifications. Subsequent studies of hypermodified nucleosides at position 37, including hydroxywybutosine (OHyW) in tRNA^Phe, wybutosine and its interaction with Mg²⁺ ions in the same system, and the lysidine and t⁶A pair required for AUA codon recognition in tRNA^Ile, collectively established that anticodon-loop modifications act locally by remodeling loop geometry and base stacking to drive the ASL toward a functionally open conformation. More recently, Nuclear Magnetic Resonance (NMR) and optical melting experiments have shown that m³C32 modification in the ASL domain of human tRNA-Arg-UCU simultaneously disrupts local base pairing and electrostatically stabilizes the ASL region. Together, these studies established that ASL modifications regulate tRNA conformation as well as decoding through direct structural effects on the ASL domain. MD simulation studies have further shown that ASL modifications modulate the conformational ensemble of tRNA, influence codon–anticodon interactions, and thereby affect decoding efficiency and adaptation during cellular stress. ,

The second group, which is less characterized, is composed of core modifications located in the D-arm and T-arm regions. , Although experimental studies have demonstrated that core modifications enhance structural integrity and biological function, their detailed mechanistic effects remain poorly understood, largely because high-resolution structural information capturing tRNA dynamics under solution conditions is limited. NMR studies on Escherichia coli initiator tRNA (tRNA^fMet) demonstrated that unmodified tRNA exhibits enhanced conformational exchange on the microsecond-to-millisecond time scale relative to naturally modified species. The results indicated that post-transcriptional modifications suppress slow-time scale conformational exchange while simultaneously increasing fast (ps–ns) local fluctuations, producing a more stabilized yet dynamically redistributed ensemble. Another study used single-molecule fluorescence resonance energy transfer (smFRET) to show that human mitochondrial tRNA^Lys requires cooperative modification patterns for proper folding, with m²G₁₀ alone failing to facilitate the correct structure. These findings highlight that individual modifications can produce suboptimal structural outcomes when not accompanied by their cooperative partners, underscoring the importance of modification networks.

Computational approaches have provided mechanistic insights into how core region perturbations propagate through the tRNA structure. MD simulations of m²G and m₂ ²G modifications at the D-stem hinge region (position 26) showed that these methylguanosine modifications prevent tRNA from adopting aberrant mitochondrial-like conformations and preserve the canonical L-shaped tertiary fold. , MD simulations of E. coli tRNA^Met D-loop variants further demonstrated that single-nucleotide substitutions perturb allosteric communication networks while preserving the global tertiary fold, with dynamic network analysis identifying conserved core nucleotides as critical hubs for allosteric signal propagation. For example, a U₈C substitution in human mitochondrial tRNA^Met was shown to cause mitochondrial myopathy by impairing aminoacylation efficiency despite preserving the overall tertiary structure, supporting the interpretation that core perturbations can compromise function through allosteric network disruption without global structural destabilization.

Recent in vivo evidence demonstrates that core modifications regulate intrinsic decoding efficiency at the ribosomal A-site through mechanisms independent of aminoacylation and subcellular localization. Genetic interactions between modifications targeting proximal sites further indicate that individual modification effects on decoding accuracy are not additive, establishing that combinatorial modification states must be characterized to understand their functional consequences. Despite these findings, the combinatorial and cooperative effects of naturally occurring core modifications have not yet been systematically examined through computational approaches.

In this study, we employed all-atom MD simulations (48 μs total) to systematically dissect the individual and combinatorial effects of three evolutionarily conserved core modifications found in yeast tRNA^Phe: 7-methylguanosine at position 46 (m⁷G₄₆, termed G*), 5-methylcytosine at position 49 (m⁵C₄₉ or C*), and 1-methyladenosine at position 58 (m¹A₅₈ or A*) (Figure and Figure S1). These modifications were selected based on their distinct structural locations and established functional roles. For example, G* is located in the VL-TSL region and is suggested to enhance thermal stability through strengthened electrostatic interactions and base stacking. , C* is located in the TSL region and is suggested to enhance base stacking through increased hydrophobicity, maintaining TSL structural integrity and stabilizing tertiary interactions between the TSL and DSL regions. , A* is positioned at the critical TSL/DSL interface that stabilizes the tRNA elbow region. Additionally, A* introduces a positive charge on the Watson–Crick face, preventing canonical base pairing and creating unique structural constraints at the tertiary core. , Through comprehensive analysis of structural dynamics, energy landscapes, residue–residue correlations, rotational dynamics, and allosteric communication pathways, we provide mechanistic insights into the nonadditive conformational effects of core modifications in tRNA.

Computational Methods

System Preparation

The initial atomic coordinates for tRNA^Phe were obtained from a crystal structure (PDB ID: 1EHZ). Based on the crystal structure, we created eight tRNA systems, one unmodified system and seven chemically modified systems. Specifically, we generated three singly modified tRNA^Phe systems by retaining only one of the modified nucleotides (m⁷G₄₆, m⁵C₄₉, or m¹A₅₈) in the crystal structure. Next, we generated three doubly modified tRNA^Phe systems containing combinations of two of the three chemical modifications, and one triply modified structure with all three modifications simultaneously present. We also generated an unmodified tRNA^Phe structure, ensuring that any chemically modified nucleotides in the crystal structure were converted to the corresponding unmodified nucleotide using the psfgen tool in the Visual Molecular Dynamics (VMD) software (Table S1). To ionize each system, we used the cionize plugin in VMD to place Mg²⁺ ions at the minima in the electrostatic potential calculated around the tRNA backbone. Then, we solvated each structure using the CHARMM-modified TIP3P water molecules. We further maintained the bulk ionic concentration by adding Mg²⁺ and Cl^– (150 mM MgCl₂). This concentration was selected to ensure adequate sampling of Mg²⁺-RNA binding and desolvation events, given that the first-shell water exchange for Mg²⁺ occurs on the microsecond-to-millisecond time scale. A typical solvated and ionized tRNA system contained ∼67000 atoms.

Details of MD Simulations

Each of the 8 systems was first energy-minimized using a conjugate-gradient scheme (1000 steps) to remove any steric overlaps. To stabilize the system density and the size of the simulation domain, an initial equilibration step (with a simulation time-step of 2 fs) was performed in the NPT ensemble for 500 ps. The temperature (at 300 K) and the pressure (at 1 atm) were maintained using a Langevin thermostat and a Nosé-Hoover barostat, respectively. Periodic boundary conditions were applied in all simulations, and the electrostatic interactions were computed using the Particle Mesh Ewald (PME) method. A cutoff of 12 Å was used to truncate van der Waals interactions. For the production step, three independent all-atom MD simulations (each 2 μs long) with a 2 fs time-step in the NVT ensemble were conducted for all 8 systems, resulting in a total of 48 μs data set. All MD simulations were performed using the NAMD software, and the trajectories were visualized using VMD. The CHARMM36 force-field for RNA as well as the CHARMM36 all-atom additive force-field for post-transcriptional modifications in RNA were used. Additionally, the improved CHARMM36 force-field parameters for Mg²⁺ ions in the nucleic acid systems, the Roux force-field parameters for the Cl^– ions, and the modified TIP3P water model were used in conducting simulations. ,

Conformational Analyses

Rotational Analysis

To analyze the impact of chemical modifications on intramolecular domain rotations in tRNA, we partitioned tRNA into three structural domains. The upper arm (TSL and AAS, residues 1–7, 49–65, and 66–76), the lower arm (ASL and DSL, residues 10 to 44), and an elbow region (residues 8, 9, 45, 46, 47, and 48), which served as a pivot point. These partitions align with the canonical tRNA L-shape, following structural definitions used to model macromolecular communication and translocation trajectories. We calculated the center of mass (COM) for each domain using mass-weighted atomic coordinates excluding hydrogen atoms, and we constructed vectors from the elbow region’s COM to each arm’s COM. Additionally, we calculated signed rotational angles using a projection-based method analogous to approaches employed in ribosomal intersubunit rotation studies. , Each arm was sequentially treated as a fixed reference axis while the opposing arm vector was projected onto a plane perpendicular to this axis, and rotation angles were calculated from the arccosine of normalized dot products with signs determined using cross products. The sign convention follows established nomenclature in ribosomal translocation studies, where positive angles indicate counterclockwise rotation (CCWR) and negative angles indicate clockwise rotation (CWR) when viewing down the reference axis. This approach has been used to characterize rotations in ribosomal complexes using smFRET and cryogenic electron microscopy (cryo-EM) methods. ,

Principal Component Analysis (PCA)

We performed principal component analysis (PCA) using CPPTRAJ , to identify the dominant collective motions in the tRNA structure. PCA reduces the high-dimensional conformational space to a lower-dimensional representation by identifying orthogonal modes of motion that capture the largest variance in atomic positions. Following the alignment of all trajectory frames to remove overall translation and rotation, a covariance matrix was constructed from the positional deviations of the phosphorus atoms (P) along the sugar–phosphate backbone. The diagonalization of this matrix yielded eigenvectors (principal components) and eigenvalues representing the variance of the motion along each mode. Free energy surface (FES) analysis was performed in the subspace defined by the first two principal components (PC1 and PC2), capturing the dominant conformational motions.

Time-Lagged Independent Component Analysis (tICA)

To identify slow conformational motions and kinetically relevant states, we performed the time-lagged independent component analysis (tICA) on the trajectory data. − For selecting the optimal features for characterizing tRNA conformational dynamics, we evaluated three feature types using the variational approach for Markov processes (VAMP2) scoring method, , which is a commonly used metric to determine the optimal input feature for tICA. , The feature types used were torsion angles, atom positions (Cartesian coordinates), and interatomic distances between the backbone atoms. The backbone was defined as the sugar–phosphate chain including the P atom and the ribose atoms (O5′, C5′, C4′, C3′, O3′). The VAMP2 scores were calculated for six different lag times (0.5, 1, 5, 10, 20, and 40 ns), and the highest VAMP2 scores were observed for backbone interatomic distances. Furthermore, to determine an appropriate lag time for tICA, we evaluated the convergence of implied time scale plots across a range of lag times. , We selected the shortest lag time at which the implied time scale plots reached plateau values to satisfy the Markovian assumption while maximizing temporal resolution. Additionally, we performed tICA using multiple different lag times to assess the sensitivity of the resulting free energy surfaces to the choice of lag time. Molecular features were computed using MDTraj. VAMP2 scoring, tICA, and implied time scale plots were implemented using Python with NumPy and SciPy libraries, following established algorithms. ,,

Residue–Residue Correlation Analysis

Past studies have indicated that correlated residue fluctuations can reveal communication pathways and allosteric networks in biomolecules. − Therefore, we used the Carma 3.8 package to compute the atomic covariance matrices from the MD trajectories. The covariance data were subsequently used to generate residue–residue correlation maps. Using these data, we further computed the difference dynamic cross-correlation (ΔDCC) for the differences in correlated motions between the unmodified tRNA and modified tRNA structures to enable the identification of regions that display anticorrelated motions. Such patterns may perturb the conformational flexibility of helical stems and loops, and provide mechanistic insights into potential long-range communication within the tRNA structure. ,

Dynamic Network Analysis

We further investigated allosteric communication networks within tRNA by employing the community network analysis method developed by Sethi et al., , which has proven effective in mapping allosteric pathways in proteins. Using MD simulation trajectories of tRNA, we performed cross-correlation analysis with Carma, treating the P atoms of each nucleotide as network nodes, analogous to the Cα atoms used for protein residues. This approach directly captures the conformational dynamics relevant to tRNA function. Pairwise correlations between nucleotides were calculated according to eq , generating correlation maps that revealed coordinated motions throughout the tRNA structure. The dynamical cross-correlation coefficient C _ij between residues i and j is calculated as

$$C_{\mathit{ij}}=\frac{⟨\mathrm{\Delta }{\mathrm{r⃗}}_i(t)·\mathrm{\Delta }{\mathrm{r⃗}}_j(t)⟩}{\sqrt{⟨\mathrm{\Delta }{\mathrm{r⃗}}_i{(t)}^2⟩⟨\mathrm{\Delta }{\mathrm{r⃗}}_j{(t)}^2⟩}}$$ 1

where Δr⃗ _i (t) = r⃗ _i (t) – ⟨r⃗ _i (t)⟩, r⃗ _i (t) is the position vector of the P atom of residue i at time t, ⟨r⃗ _i (t)⟩ is the mean position over the trajectory, and ⟨⟩ denotes the ensemble average over all frames.

These maps provide the foundation for identifying dynamically coupled regions that may participate in allosteric communication between functionally important sites, such as the ASL, AAS, and regions that interact with aminoacyl-tRNA synthetases or the ribosome. The resulting community structure could potentially highlight how post-transcriptional modifications propagate through the tRNA molecule to affect distant functional sites. We also used this method to identify optimal (shortest pathways) and suboptimal pathways (closest alternative pathways) between residues, particularly, residue 34, which is the wobble position located in the ASL region, and residue 46 in the core region of tRNA. These optimal pathways are crucial in conveying allosteric signals from one residue to another, while suboptimal pathways reveal alternative routes for communication. Identification of the suboptimal pathways depends on the edge length offset parameter, which defines the tolerance threshold for calculating alternative routes between nodes. Although the specific numerical value of this offset is an operational choice, it strictly dictates the breadth of the calculated communication network. To capture the most direct communication routes, edge length offsets of 0 and 5 were initially tested, but these cutoffs did not produce any significant difference in pathway counts for different tRNA variants. We expanded the analysis by increasing the edge length offset to 10, 20, and 30 (Table S4).

Quantitative Epistasis Analysis

To quantitatively assess whether the effects of the three core modifications combine additively, we applied a quantitative epistasis framework. The epistasis residual ε is defined as the difference between the observed value of a structural feature F in a combined-modification system and the value predicted by summing the effects of each modification individually

$$F_{\mathrm{pred}}=\sum F(\mathrm{single\; mods})-(\mathrm{n}-1)F(\mathrm{unmod})$$ 2

for n modifications combined. This residual-based formalism, originating with Fisher’s decomposition of genetic effects and subsequently formalized for biochemical and biophysical systems, has been applied across a broad range of molecular systems − to detect nonadditive coupling between mutations. We applied an analogous mutant-cycle framework, with the three core modifications (G*, C*, A*) as combinatorial perturbations to the tRNA and the unmodified structure as the reference, using two independent structural metrics. Using the data from the community network count and coordinated arm rotation, we computed the epistasis residual ε for each feature. The residual ε represents the deviation of the observed combined effect from the additive prediction, where a value of zero indicates that the two modifications combine additively, while a nonzero value indicates nonadditivity.

Results

Core Region Modifications Exhibit Variable Impact on tRNA Stability

We assessed the global structural effects of core modifications on tRNA by calculating the root mean squared deviation (RMSD) for each of the three independent simulations for all eight tRNA variants (Figure S2) along with the cumulative distributions of RMSD relative to the initial crystal structure (PDB ID: 1EHZ) for all eight tRNA variants (Figures A and S3). The unmodified tRNA exhibited the broadest distribution (black trace in Figures A and S3), reflecting the highest structural deviation among all tRNA variants, while the doubly modified C*A* variant produced the narrowest distribution at the lowest RMSD values (purple trace in Figure A), indicating that the double modification (C*A*) most effectively results in a narrower conformational distribution with a single, well-defined state. The remaining singly modified and doubly modified variants occupy intermediate positions, with G* and A* individually producing sharper distributions than C* alone, suggesting that C* exerts its conformational effect primarily in combination with A*. Notably, the triply modified G*C*A* variant does not reproduce the narrow distribution of C*A*, demonstrating that the introduction of G* into the C*A* partially relieves the conformational restriction imposed by C* and A* together. Furthermore, the time-evolution traces of RMSD for each independent simulation showed that RMSD values for all tRNA variants fluctuated within the range between ∼3.0 Å and ∼4.5 Å over the last 1000 ns of most of the MD simulations (Figure S2). Specifically, the lowest RMSD value that we observed in this range was 3.10 ± 0.31 Å for the C*A* variant (red trace in Figure S2) and the highest was 4.55 ± 0.57 Å for the G*A* variant (red trace in Figure S2). However, we also observed two simulations to exceed this RMSD range: one simulation of the unmodified tRNA variant and one simulation of the G*A* variant (both shown as green traces in Figure S2), which exhibited RMSD values of 4.90 ± 0.80 Å and 6.17 ± 0.38 Å, respectively. These larger RMSD values were associated with the increased conformational flexibility of the ASL residues in their respective simulations. Nonetheless, the overall RMSD values remained generally consistent across independent MD simulations for all tRNA variants.

2.

(A) Distributions of RMSD for each tRNA variant. The RMSD was computed relative to the initial crystal structure (PDB ID: 1EHZ). RMSD distributions are based on data from three independent MD simulations. The peak position in each distribution indicates the mean RMSD corresponding to a preferred conformational state, while the width of each distribution quantifies structural variability. (B) Per-residue ΔRMSF analysis for selected variants (singly, doubly, and triply modified) with anticodon domain residues highlighted (magenta).

We also calculated the radius of gyration (R _g) for each tRNA variant across all MD simulations to evaluate changes in the overall shape of each tRNA structure (Figure S4). We observed that the R _g values for all tRNA variants fluctuated in the range between ∼20.0 Å and ∼22.5 Å around the initial R _g value of ∼21.5 Å (Figure S4). This behavior was consistent across all independent simulations, indicating that the overall structure of tRNA was maintained throughout the MD simulations. Together with the RMSD analysis, these results indicate that the overall tRNA structure was maintained despite local conformational fluctuations.

Per-residue root mean squared fluctuations (RMSF) were computed for each simulation of all eight tRNA variants (Figure S5) to assess the residue-level fluctuations across independent trajectories. Overall, the RMSF profiles were comparable among three independent simulations for each tRNA variant. Specifically, larger fluctuations were commonly observed in the terminal residues of the AAS domain, whereas the ASL domain also exhibited elevated fluctuations across several simulations (e.g., simulation 3 of the unmodified tRNA; Figure S5). In contrast, we observed the DSL, VL, and TSL domains to be consistently more stable relative to the AAS domain across all simulations for each variant (Figure S5). Additionally, per-residue difference in root mean squared fluctuation (ΔRMSF) analysis, defined as the difference in RMSF between the modified and unmodified tRNA, was used to resolve whether these globally divergent states retained residue-level stability. Positive ΔRMSF values indicate increased fluctuations of residues in the structure, whereas negative values signify decreased fluctuations (Figure B). In singly modified variants, negative ΔRMSF values were concentrated in the anticodon domain (residues 27–44, magenta), confirming that even a single core modification reduces fluctuations in this distal region. Additionally, the analysis revealed that the increased stability observed in C*A* originates from localized flexibility suppression in the ASL region, which exhibited the largest negative ΔRMSF values among all variants examined (Figure B). Notably, the triply modified G*C*A* variant displayed comparably negative ΔRMSF in the ASL but showed positive ΔRMSF in regions toward the 3′ end, indicating that the introduction of G* along with C*A* redistributes, rather than uniformly amplifies, flexibility changes across the structure.

To determine the effect of core modifications on the intramolecular hydrogen bonding, we calculated the average number of hydrogen bonds (N _HB) formed between all tRNA bases for each tRNA variant (Figure S6). The N _HB values varied between 112 ± 2 for the triply modified tRNA variant and 116 ± 2 for the doubly modified tRNA variants. Overall, the N _HB values were similar across different tRNA variants, with the observed differences falling within the error bars of each other (Figure S6). This indicates that the overall hydrogen-bonding patterns between nucleotides in the tRNA are largely preserved across all modification states, even as RMSD and free energy landscape analyses reveal substantial differences in the conformational ensembles sampled by each system. These results suggest that core modifications alter the tRNA conformational landscape without substantial changes in the intramolecular hydrogen bonding within the tRNA structure.

m¹A₅₈ Coordinates Anticorrelated tRNA Arm Rotation

Rotational dynamics of the upper and lower tRNA arms were quantified across three independent MD simulations for all eight variants (Table S2). The counterclockwise rotation (CCWR) of the upper arm ranged between 11.8% and 100%, and the clockwise rotation (CWR) of the lower arm ranged between 0% and 93.2% across variants. The unmodified tRNA served as a reference at 62.7% upper CCWR and 49.6% lower CWR, on average. The presence of A* emerged as the principal determinant of the coordinated arm rotation. All variants containing A* consistently produced higher simultaneous upper-arm CCWR and lower-arm CWR percentages across all three independent simulations. Variants with the modification A* consistently exhibited higher and more coordinated rotational occupancies across both arms. Among these, G*A* produced the strongest coupling, with a mean upper-arm CCWR of 97.2% and a mean lower-arm CWR of 88.1%, followed by A* alone (84.2% and 79.2%, respectively) and C*A* (73.6% and 66.9%, respectively). In contrast, variants lacking A* showed diminished and less coordinated values: C* produced a mean upper-arm CCWR of 77.6% but only 62.3% lower-arm CWR, while G* and G*C* both showed diminished mean occupancies relative to A*-containing variants, with G* averaging 66.7% upper CCWR and 65.6% lower CWR, and G*C* showing further reduced values of 40.9% and 28.1%, respectively. The triply modified G*C*A* variant constituted a marked exception. The consistency that characterized all A*-containing variants (A*, G*A*, and C*A*) was not conserved in G*C*A*, indicating that neither the introduction of G* into the C*A* variant nor the introduction of C* into the G*A* variant preserves the coordinated rotational behavior otherwise conferred by A*. This loss of directional consistency in G*C*A* regardless of the pairwise context from which it is derived constitutes further evidence of nonadditive interactions among the three modifications.

Chemical Modifications Reshape tRNA Conformational Landscapes

To understand the collective impact of chemical modifications on tRNA conformational dynamics, we performed PCA on our MD trajectories and projected the FES along the first two principal components. These FES plots provide a comprehensive characterization of the conformational space sampled by tRNA in its chemically modified forms, revealing the energetic distribution of structural ensembles and metastable conformational states. The unmodified tRNA exhibited a broad, diffuse energy landscape with a single basin (Figure A), reflecting extensive conformational sampling and high structural flexibility. The low-energy barriers indicate readily accessible transitions between microstates. These observations align with experimental NMR studies on E. coli tRNA^fMet, which demonstrated that unmodified tRNA exhibits greater conformational exchange between substrates compared to modified variants.

3.

Free energy surface (FES) projected along the first two principal components (PC1 and PC2) for (A) the unmodified tRNA variant, (B–D) the singly modified tRNA variants, (E–G) the doubly modified tRNA variants, and (H) the triply modified tRNA variant. Blue regions indicate low-energy, thermodynamically favorable conformations, while red indicates higher-energy regions. The global free energy minimum of each system is indicated with an asterisk symbol, which is projected to the corresponding representative conformation of tRNA, with regions of high deviation from the crystal structure (>4.0 Å) highlighted in red.

Incorporation of single chemical modifications resulted in a significant reduction of the accessible conformational landscape (Figure B–D). All singly modified systems converged to single, well-defined energy basins, indicating that individual modifications suppress the conformational heterogeneity observed in unmodified tRNA. The localization of these basins along both principal components demonstrates that individual modifications reduce the conformational variability and impose defined geometric constraints on the tRNA structure. In contrast, doubly modified tRNA systems showed transitions toward conformational heterogeneity (Figure E–G). This behavior is most pronounced in C*A* and G*A*, which display clear multimodality via distinct split low-energy basins. The G*C* system retains a single primary minimum, though its free energy surface appears more diffuse than those of the singly modified G* and C* systems. This indicates that the introduction of a second modification begins to relax the constraints imposed by single modifications, either by splitting the landscape into discrete basins or by increasing the spread of the accessible microstates.

Furthermore, the progression from unmodified to triply modified tRNA demonstrated a non-monotonic relationship between modification number and conformational landscape complexity. Unmodified tRNA samples a broad, continuous region of conformational space, and the introduction of single modifications narrows the sampling to more localized regions. The addition of a second modification relaxes these constraints either by increasing the spread of accessible conformations or fragmenting the landscape into discrete basins, while the incorporation of the third modification drives the system into a multimodal character rather than consolidating the conformational ensemble.

To further assess conformational sampling across independent simulations, we resolved additional PCA free energy landscapes for each MD simulation across all tRNA variants (Figures S7 and S8). We observed the dominant free energy minima in similar regions of the PCA free energy landscapes across three simulations for each tRNA variant (Figures S7 and S8), suggesting that the independent trajectories sampled similar low-energy tRNA conformations. However, we additionally conducted clustering analysis on each trajectory consisting of 25,000 tRNA configurations to directly compare the averaged structures from the most populated clusters (Figure S9). The centroid structures of the most populated clusters, collectively representing at least 50% of the sampled configurations for each trajectory, showed substantial overlap across the independent simulations (Figure S9). Structural differences were primarily observed in the AAS and ASL domains, consistent with the greater flexibility of these regions observed in the RMSF analysis (Figure S5), while the other tRNA domains (DSL, VL, and TSL) exhibited greater conformational similarity across the independent simulations. Overall, the PCA and clustering analyses demonstrated that similar tRNA conformations were sampled across independent MD simulations, while local structural variability was observed in the more flexible tRNA regions (e.g., AAS domain).

tICA Reveals Kinetic Organization of Conformational Landscapes

To characterize slow conformational dynamics, we applied tICA with a lag time (τ) of 10 ns to data from MD trajectories. We initially computed VAMP2 scores for backbone torsion angles, atomic coordinates, and interatomic distances at different lag times (Figure S10). This analysis identified interatomic distances as the optimal feature for capturing slow conformational motions. We therefore performed tICA using backbone interatomic distances as input features.

To determine an appropriate tICA lag time, we evaluated implied time scale convergence across a range of lag time values, following established validation procedures. The implied time scales approached plateau values between 10 and 25 ns across different systems, with most systems showing plateauing at ∼10 ns (Figure S11). Therefore, we selected τ = 10 ns as a common lag time for subsequent tICA at which the implied time scales plateaued (Figure S11) to satisfy the Markovian assumption while maximizing temporal resolution. Additionally, we probed the sensitivity of the resulting free energy landscapes to the choice of lag time (Figure S12). We observed that shorter lag times (0.5–1 ns) resulted in highly fragmented free energy surfaces characterized by numerous localized minima separated by high energetic barriers (Figure S12). As the lag time increased to 5 ns, the landscapes progressively became less fragmented, with kinetically adjacent basins merging into broader low-energy basins (Figure S12). For example, for the G* system, increasing the lag time to 10 ns resulted in the merging of previously separated regions and a reduction of free energy barriers to ∼2.5–4 kcal/mol (Figure S12). At larger lag times, additional smaller free energy minima began merging into broader low-energy basins, further reducing energetic barriers that previously separated them (Figure S12).

The resulting free energy landscapes projected onto the first two tICA components (tIC1 and tIC2) revealed markedly different conformational dynamics compared to PCA (Figure S13). Unlike PCA, which identifies directions of maximum structural variance irrespective of time scale, tICA extracts the slowest dynamical modes by maximizing autocorrelation at a fixed lag time, thereby isolating kinetically relevant conformational transitions from fast, high-variance fluctuations. Analyzing the cumulative free energy landscapes of the different variants at a lag time of 10 ns (Figure S13), the unmodified tRNA^Phe exhibited a connected landscape with multiple accessible basins linked through a shallow, low-barrier region, reflecting high kinetic accessibility between states. The G* modified variant retained a similarly connected topology but showed greater separation between two principal basins, indicating increased conformational heterogeneity relative to the unmodified system while preserving overall connectivity (Figure S13). The C* modification resulted in a different landscape, dominated by a single deep basin accompanied by less populated conformational regions, indicating that this modification reduces conformational heterogeneity toward one dominant conformation (Figure S13). In contrast, the A* variant exhibited several basins, separated by moderate barriers (∼2–4 kcal/mol), indicating increased conformational heterogeneity relative to the unmodified system while maintaining the overall connected landscape (Figure S13).

The doubly modified variants exhibited distinct landscapes rather than a uniform increase in fragmentation. The G*C* variant contained a dominant basin connected to a secondary basin, with a shallow ridge extending toward an essentially disconnected minor basin (Figure S13). The G*A* variant exhibited three free energy basins connected through a shared central transition state, remaining kinetically linked rather than isolated (Figure S13). The C*A* conformational landscape is characterized by three basins that converge upon a shared central region, forming a centralized network topology analogous to that of G*A*. Thus, rather than existing as isolated populations, the individual arms retain kinetic connectivity via mutual intersection. Relative to G*A*, however, this central junction sits at a visibly higher free energy, indicating a more heterogeneous, although still connected, pathway between arms.

The triply modified G*C*A* system displayed a three-lobed converging topology similar to G*A*, but occupying a smaller region of tICA space connected by narrower channels, indicating that combining all three modifications constrains rather than expands the accessible conformational space. Overall, the modifications reshape the landscape topology distinctly: for example, the unmodified, G*, and G*C* variants occupy elongated, continuous free energy surfaces; A*, G*A*, and G*C*A* adopt multibasin topologies; and C*A* alone shows clear separation into several disconnected regions.

Core Modifications Exhibit Localized Effects on Inter-Residue Correlations

We computed dynamic cross-correlation matrices (DCCMs) for all modified tRNA systems to assess how chemical modifications alter the network of correlated motions within tRNA (Figure ). We further generated difference maps (ΔDCCM) by subtracting the unmodified correlation matrix from each modified state. These ΔDCCM maps (Figure S14) reveal regions where modifications enhance (positive ΔDCCM values) or diminish (negative ΔDCCM values) correlated motions between residue pairs relative to the unmodified reference.

4.

DCCMs showing correlated (red) and anticorrelated (blue) motions for (A) the unmodified tRNA variant, (B) the singly modified G* tRNA variant, (C) the doubly modified C*A* tRNA variant, and (D) the triply modified G*C*A* tRNA variant. The maps shown here indicate a positive correlation between the TSL, AAS, and DSL domains, highlighted by the red-dashed box showing the brown region on the map. Negative correlations between the anticodon arm and tRNA upper-arm regions (TSL and AAS) are marked with the blue box showing blue regions on the map.

In unmodified tRNA, interdomain cross-correlations were weak to moderate across all domain pairs. Two interdomain features were identified: an anticorrelated relationship between the ASL residues (27–43) and the AAS and the DSL regions (blue inset box; Figure ), and a positive correlation between the TSL and the AAS regions (red inset box; Figure ). Correlations between sequentially adjacent residues were consistently strong across all modification states, while off-diagonal elements, reflecting interdomain coupling, showed modification-dependent variation.

The DCCMs reveal distinct interdomain coupling patterns across modification states. In G*, the cross-correlation blocks involving the ASL and its coupling partners shift to more negative values relative to unmodified tRNA, indicating that core and anticodon arm motions become more anticorrelated upon introduction of G*. In C*A*, negative cross-correlation develops distinctly within the ASL coupling region (the blue inset box), alongside a subtle region of positive correlation appearing at lower residue pairs. In G*C*A*, the off-diagonal coupling within the blue box deepens into the strongest anticorrelation of all states. This demonstrates that the G*C*A* correlation pattern is not a linear combination of the G* and C*A* patterns.

Core Modifications Restructure tRNA Communication Networks

The communities of correlated residues were identified by applying the Girvan-Newman algorithm , to our correlation network data. The algorithm operates by sequentially removing high “betweenness” edges which serve as intercommunity bridges until distinct residue communities emerge. The “betweenness” centrality of an edge is defined as the sum of the fractions of shortest paths between all pairs of nodes that traverse a given edge. Edges with high betweenness centrality identify critical bridges linking distinct network communities. ,

We performed community network analysis on all trajectory data for each of the eight tRNA variants to identify correlated communities and intercommunity communication pathways. Despite variation in total community count across variants, a core set of 6–9 communities was consistently identified in all variants (Figure ). Differences in the total community count reflected changes in the number of peripheral single-node communities rather than reorganization of this core structure. Single-node peripheral communities were reassigned to the neighboring community with the highest intercommunity edge weight. The reduction in community count upon introduction of chemical modifications reflects modification-dependent integration of peripheral nodes into the core network structure, with the degree of integration showing nonadditive dependence on modification identity and combination. Overall, the unmodified tRNA produced 22 communities (Table S3), the highest among all variants, distributed across all structural domains, including a peripheral community at the ASL connected by a single low-weight intercommunity edge.

5.

Dynamic network analysis of tRNA conformational dynamics using the Girvan-Newman community detection algorithm for (A) the unmodified tRNA variant, (B-D) the singly modified tRNA variants, (E-G) the doubly modified tRNA variants, and (H) the triply modified tRNA variant. Core communities are represented by colored circles connected by intercommunity edges that represent the sum of all pairwise correlations. Thicker edges indicate stronger intercommunity communication. Modifications progressively reduce network modularity, indicating enhanced long-range communication between residues. Shown here are different modified states of tRNA with the number of core communities formed.

Among singly modified variants, C* produced the greatest reduction in community count (10 communities), followed by G* (13 communities) and A* (14 communities). G* and C* showed the highest number of intercommunity edges at the tRNA core region, while A* retained the peripheral ASL community observed in unmodified tRNA, absent in C* and G*.

Among doubly modified variants, G*A* produced the fewest communities (12), followed by C*A* (11), both similar to their respective singly modified forms. In contrast, G*C* produced 14 communities, equal to those of A* alone and greater than either G* (13) or C* (10) individually, indicating that the combination of G* and C* does not produce additive integration of peripheral nodes. The triply modified G*C*A* variant produced the fewest communities of all eight variants (9 communities), fewer than any singly modified or doubly modified state. This reduction is not predictable from any pairwise combination, as G*C* alone shows a greater community count than either of its constituent modifications, yet the introduction of A* into this context produces a degree of network integration that exceeds all doubly modified variants, constituting evidence of nonadditive interactions among the three modifications at the level of the dynamic correlation network.

Core Modifications Expand Allosteric Pathway Density between the tRNA Core and the Anticodon Region

Negative ΔRMSF values concentrated in the anticodon domain (residues 32–42, Figure B) indicated that core modifications collectively suppressed conformational flexibility in the ASL region, which is spatially remote from the modification sites. The unmodified tRNA exhibited elevated anticodon arm RMSD across all three simulations, corresponding to continuous inward displacement of the ASL region and disruption of anticodon arm base pairs (Figure ). Notably, this conformational behavior was absent in all seven modified tRNA variants. We observed that the introduction of a single core modification, for example, G* was sufficient to prevent the displacement of the ASL toward the tRNA interior (Figure S14). Based on these observations, we probed for allosteric communication pathways between core region residues and anticodon residues using the community network. Given the combinatorial scale of a comprehensive residue–residue pathway analysis across all eight modification states, the analysis was strategically restricted to G₄₆ as the source residue across the unmodified and G*-containing variants (G*, G*A*, G*C*, and G*C*A*) and G₃₄ in the ASL region as the target residue.

6.

Different conformations of the anticodon domain (ASL and the nucleotides) in both modified and unmodified tRNA. Unmodified tRNAs consistently exhibited an inward displacement of the ASL across all independent simulations, which disrupted base pairs (such as G₃₀–C₄₀, A₃₁–U₃₉, and C₃₂–A₃₈). See also Figure S15.

Using the method of Sethi et al., we identified communication pathways between these two distal residues (G₄₆ and G₃₄). G₄₆ was selected on the basis of RMSD and ΔRMSF results, identifying G* at this position as the most stabilizing core modification among the singly modified variants. G₃₄ was selected as the target because it occupies the wobble position of the anticodon, directly engaging the third codon nucleotide during decoding, making it the most functionally consequential residue in the ASL region. The number of suboptimal pathways identified by the algorithm was sensitive to the chosen edge length offset. At an offset of 0, all systems yielded only one or two pathways, providing little differentiation between the tRNA variants. The identified pathways increased substantially with increasing offset values until a threshold of 20, beyond which the count increased at a higher rate due to the inclusion of kinetically convoluted loops of low physical relevance. Importantly, across all tested offset values, the unmodified tRNA variant consistently exhibited the fewest number of suboptimal pathways, whereas the singly modified G* tRNA variant consistently exhibited the greatest number (Figure S16). An edge length offset of 20 was applied based on the stability of intercommunity pathway counts across modification states relative to offsets of 0, 5, 10, and 30. Thus, although the absolute number of suboptimal pathways depends on the selected edge length offset value, the relative ranking among the tRNA variants was maintained across all tested offsets (Table S4).

The single optimal path (red pathway; Figure ), representing the shortest pathway between G₄₆ and G₃₄ nucleotides, remained identical across all tRNA variants, especially in the ASL domain, where the pathway becomes confined to a narrow region along the anticodon arm backbone. We also demonstrated that the optimal pathway remained unchanged across the different edge length offsets tested (Figure S16), indicating that the optimal pathway connecting the two nucleotides is insensitive to the choice of this parameter. The unmodified tRNA produced 7 suboptimal pathways between G₄₆ and G₃₄ (Table S4). G* produced 271 pathways, a 39-fold increase relative to that of unmodified, distributed across multiple routes spanning the full cross-sectional width of the anticodon arm. For the doubly modified variants, G*A* produced 125 pathways, while G*C* produced 91 pathways with pathway density concentrated in the midsection of the anticodon arm. G*C*A* produced 98 pathways, with pathways concentrated proximal to the G₄₆ source and G₃₄ target nodes.

7.

Optimal pathways between the core region residue G₄₆ and the anticodon residue G₃₄ (wobble position) for different tRNA variants: unmodified (A), G* (B), G*A* (C), G*C* (D), and G*C*A* (E).

The shortest communication distance decreased from 472 (unmodified) to 268 upon the introduction of G*. G*C* produced a comparable shortest distance of 271, similar to that of G* alone, indicating that the addition of C* to G* does not further enhance core-to-anticodon communication efficiency. G*A* (288) and G*C*A* (301) showed modestly increased shortest distances relative to those of G* and G*C*, suggesting that the introduction of A* into G*-containing variants increases the shortest path distance despite the greater number of suboptimal pathways available.

Quantitative Epistasis Confirms Nonadditivity

To quantitatively test whether the effects of the three core modifications combine additively, we adapted an epistasis framework used to characterize interactions among genetic perturbations. − We computed the epistasis residuals (ε) for each doubly modified variant across two structural metrics: community network count and coordinated arm rotation. The magnitude of the ε quantifies the deviation of the observed value for a combined-modification system from that predicted from the corresponding singly modified systems under an additive model. Our findings demonstrated that the combinatorial effects of these chemical modifications are nonadditive, as evidenced by the nonzero values of ε obtained for both metrics. Specifically, for the community network count, all three pairwise combinations showed large residuals (Table S5): |ε­(G*, C*)|= 13, |ε(G*, A*)| = 7, and |ε(C*, A*)| = 9 communities. For the coordinated arm rotation, residuals ranged from 7.1 to 50.2 percentage points across the six upper-CCWR/lower-CWR measurements in G*A*, C*A*, and G*C* (Table S6). G*C* showed the largest deviation from additivity (|ε| = 40.7 for upper CCWR, 50.2 for lower CWR), followed by C*A* (25.5, 25.0). G*A* deviated least from the additive prediction (9.0, 7.1).

Thus, all three doubly modified variants showed nonzero epistasis residuals across both metrics, indicating that the combined effect of two core modifications is not predictable from the sum of their individual effects. These results provide a quantitative measure of the nonadditive effects observed in the combinatorial modification states and further indicate that the influence of one core modification on tRNA conformational behavior depends on the presence of another modification.

Discussion

In this study, we investigated the individual and combinatorial effects of three core modifications (m⁷G₄₆, m⁵C₄₉, and m¹A₅₈) on the conformational dynamics and allosteric communication networks in yeast tRNA^Phe, addressing gaps in our understanding of the cooperative effects of these chemical modifications. Prior computational studies primarily examined the effects on local base configurations and base-pairing interactions of individual modifications such as m²G, which is most prevalent at positions 6 and 10 in the acceptor arm; m^2,2G at position 26, which is the hinge region between the DSL and the ASL regions; and hydroxywybutosine (OHyW) or wybutosine (yW) at position 37 in the ASL region. ,, Other studies focused on ASL modifications and their localized structural consequences. , Additional computational investigations have also provided insights into global tRNA dynamics. ,, However, these studies typically considered single modifications or nucleotide substitutions, without exploring combinatorial states of chemical modifications. , By systematically evaluating combinatorial modification states, we demonstrated that tRNA dynamics are governed by nonadditive mechanisms, where the conformational outcome of a combination of selected modifications cannot be readily inferred from the effects of individual modifications alone.

Two methodological choices underpin the kinetic and network-level conclusions drawn above. First, the tICA lag time was selected as the shortest value at which implied time scales plateaued, a criterion that balances two competing failure modes. The lag times shorter than this plateau violate the Markovian assumption required for tICA and yield free energy surfaces fragmented by transient, nonphysical barriers, while lag times substantially longer than the plateau average over transitions between kinetically distinct substates, artificially smoothing the landscape and obscuring genuine barriers. A uniform lag time of 10 ns was therefore adopted across all eight variants to facilitate a direct comparison of their landscape topologies, as this value consistently marked the onset of the implied time scale plateau, despite modest variations (10–25 ns) observed between individual systems. Second, the community network pathway analysis is inherently sensitive to the edge length offset, which controls how many kinetically indirect, suboptimal routes are retained between a source and target residue; because absolute pathway count scales with this parameter rather than converging to it, we did not treat any single offset as defining a “true” pathway count. Instead, an offset of 20 was selected as a working resolution at which pathway counts across systems were neither collapsed to the one or two trivial routes obtained at an offset of 0 nor inflated by the inclusion of low-relevance loops that emerged beyond an offset of 20. The relative ranking of systems by pathway count was confirmed to be stable across the full range of offsets tested.

We briefly also note that the findings about the three core modifications studied in this work are relevant to other tRNA species. A sequence and structural comparison showed that m⁷G₄₆, m⁵C₄₉, and m¹A₅₈ occupy identical positions in bovine tRNA^Lys (PDB ID: 1FIR) and human tRNA^Phe (PDB ID: 8CTH) and that both of these tRNAs are of nucleotide lengths similar to those of yeast tRNA^Phe, preserving the spacing between the core modification sites and the anticodon domain. This conservation of both sequence position and chemical identity indicates that the core-to-ASL allosteric coupling identified in this study may also be relevant to cytoplasmic tRNAs with comparable lengths and domain architectures. In contrast, mammalian mitochondrial tRNAs frequently lack a canonical D-arm and do not consistently retain the three modifications studied here. Therefore, they are not considered to be a direct point of comparison for the mechanisms described in this study. Thus, although the tRNA investigated in the present study is from yeast, the conservation of these modification sites and the overall tRNA architecture provide a structural basis for considering the potential relevance of our findings to other eukaryotic cytoplasmic tRNAs.

The computational findings of this study could probably be directly tested by using experimental methods such as smFRET and NMR. Specifically, the long-range coupling between the core modification sites and the ASL could be probed by smFRET using dye pairs positioned to report on interdomain rather than local distances, which has been recently used to show that a single anticodon-loop modification (m¹G₃₇) shifts the conformational landscape of tRNA^ProL and suppresses ribosomal frameshifting. The nonadditive effects observed for the combinatorial modification states can be similarly investigated using solution NMR. Chemical-shift perturbation mapping across the eight modification states examined here could be performed following an approach that was previously used to resolve how individual modifications redistribute local and long-range dynamics in tRNA^fMet. This approach could further provide a residue-resolved assessment of whether the correlated motions identified computationally are recapitulated in solution.

Altogether, these observations and findings suggest that yeast tRNA modification effects are inherently combinatorial. Loss of a single modification could potentially shift the molecule toward a distinct conformational and allosteric ensemble rather than reverting to an unmodified state. Loss of m⁷G₄₆, in particular, could also dismantle an allosteric network linking the yeast tRNA core to ASL. These findings suggest that combinatorial modification states must be explicitly considered to accurately describe the tRNA structure, dynamics, and function. Further studies of ribosome-bound tRNA and aminoacyl-tRNA synthetase complexes would be necessary to properly interpret modification-dependent translational regulation and the pathologies that may arise from defects in tRNA modification enzymes. Lastly, although currently available force-field parameters enable characterization of the dynamical and thermophysical properties of tRNAs in the modified and unmodified forms, , further refinement of force-field parameters for chemically modified nucleotides will be important for improving the accuracy of simulations probing the effects of RNA modifications.

Conclusion

Using a large-scale all-atom MD simulation data set across several modification states of yeast tRNA^Phe, we demonstrated that three core modifications (m⁷G₄₆, m⁵C₄₉, and m¹A₅₈) regulate tRNA conformational dynamics through nonadditive mechanisms that cannot be inferred from individual modification effects alone. The presence of any core modification suppresses the anticodon arm collapse observed in all unmodified simulations, while combinations produce multiple free energy basins with distinct kinetic organization. Community network analysis identified modification-specific allosteric roles: m⁷G₄₆ establishes redundant core-to-anticodon communication pathways and m¹A₅₈ coordinates anticorrelated arm rotation. Collectively, these results establish that the conformational state of yeast tRNA^Phe is determined by its specific modification combination and that such combinatorial states must be explicitly characterized to accurately describe its dynamics.

All methodological details for the data generated in this study are available in the Computational Methods section, including PDB codes of input files and simulation conditions. The input files for MD simulations of the unmodified and modified tRNA systems along with the trajectory files are available on Zenodo at: 10.5281/zenodo.22164985. The simulation and visualization software , and the force field for RNA , used for the calculations are openly available.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.6c02045.

• Additional data and analyses for all eight tRNA^Phe systems, including chemical modifications present in the crystal structure, rotational angle analysis, community network analysis, optimal and suboptimal pathway counts, epistasis counts, RMSD and R _g time traces, additional RMSF and RMSD profiles, number of hydrogen bonds analysis, clustering analysis, implied time scale plots, VAMP2 scores, tICA free energy landscapes, difference dynamic cross-correlation matrices (ΔDCCMs), and structural insights into the ASL from MD simulations (PDF)

Conceptualization: T.E.A., L.L., and H.V.; methodology: T.E.A. and L.L.; software: T.E.A. and L.L.; validation: T.E.A. and L.L.; formal analysis: T.E.A.; investigation: T.E.A.; data curation: T.E.A.; writingoriginal draft preparation: T.E.A.; writingreview and editing: L.L. and H.V.; visualization: T.E.A.; supervision: L.L. and H.V.; resources: H.V.; project administration: H.V.; funding acquisition: H.V. All authors have read and agreed to the published version of the manuscript.

The authors declare no competing financial interest.