Correction to Improving the Performance of the Amber RNA Force Field by Tuning the Hydrogen-Bonding Interactions

In our paper¹ we have, among other auxiliary results, presented series of MD simulations using the DESRES² RNA force field (ff). The simulations were carried out using our own implementation of the DESRES ff with the AMBER code on GPU clusters (Supporting Information of the original paper.¹) The simulations were carried out with the TIP4P-D water model³ and Joung-Cheatham (JC) ion parameters originally derived for the TIP4P-EW water model.⁴ However, the simulations testing the DESRES ff in the original study by Tan et al.² were carried out with the CHARMM22⁵ ion parameters. Although the RNA simulations are widely considered to be insensitive to monovalent ion parameters,⁶ we later realized that one of the studied systems revealed ion-sensitive behavior.

To prevent any confusion, we include here the DESRES ff implementation for the AMBER simulation engine, where the JC ions are replaced by the CHARMM22 ones. We have repeated all original DESRES simulations with the CHARMM22 ions to see the effect of ion choice (JC vs CHARMM22). The simulation results with the two sets of ion parameters appear, within the limits of sampling, virtually identical for all studied systems except for the 14-mer of UUCG tetraloop (TL). In particular, we have observed the same disruption of signature interactions, loss of folding and then complete rearrangement of the Kink-turn RNA motif in standard MD (Figure S19 in the Supporting Information of the original paper¹), on the same timescale. Also the entire loss of folded structure of the L1 stalk ribosomal RNA is insensitive to the ion parameters (Figure 6 and Supporting Figure S20 of the original paper¹). We have again observed unfolding of the UUCG and GAGA tetraloops (8-mers, starting from the folded states in all replicas) towards the canonical A-RNA form with the REST2 protocol. Standard MD simulations of the Sarcin-Ricin motif and the hairpin ribozyme still revealed weakening of key base-phosphate interactions. We have again observed unfolding of the gcUUCGgc 8-mer system during a standard MD simulation, where both signature and base-pair interactions were lost after ~500 ns. In contrast, the standard MD simulations with UUCG TL 14-mer (starting from PDB ID 2KOC structure⁷) revealed increased stability of the native structure of the TL with CHARMM22 ions. Originally, we reported five 10 μs-long simulations using the JC ions and the UUCG native structure was lost in all of them.¹ With the CHARMM22 ions, we have noticed slower disruption of the native structure and some tendency for refolding. Thus, we have performed ten standard 20 μs simulations and at the end four copies were in the fully folded state with all signature interactions (Figure A).

Figure A:

20 μs-long unbiased MD simulations of the UUCG TL 14-mer (ggcacUUCGgugcc) with DESRES potential and CHARMM22 ionic parameters. All ten simulations were initiated from the native (PDB ID 2KOC⁷) structure. The panel shows the time evolution of major conformers, i.e., (i) structures with correctly folded A-form stem and loop with all signature interactions formed (native, blue), (ii) states with folded A-form stem and G_L4 repelled from the loop but still maintaining canonical syn conformation of χ dihedral (G_L4 bulge out (syn), cyan), and (iii) other states with folded A-form stem and G_L4 bulged out of the loop sampling mainly anti conformation (G_L4 bulge out (anti), orange).

The results indicate that different ion parameters have generally negligible effect on the performance of DESRES ff for simulations of folded RNAs considering results obtained for the above-listed systems on our computer clusters, using our simulation protocols and installed codes (GPU accelerated AMBER). This is also consistent with observations for many other systems that different monovalent ion parameters (in combination with proper water model, i.e., parameters avoiding major ff artifacts like ion crystallization) do not have any substantial influence on RNA simulations.⁶ However, the UUCG TL appears to be an exception and the CHARMM22 ions lead to visibly increased stability of the native state, though it is still not dominant. The high-resolution NMR data⁷ show the fully folded UUCG tetraloop as the only detectably populated structure suggesting that other possible conformations should have significantly lower populations.

The origin of the sensitivity of the UUCG TL 14-mer simulations to the ion parameters is not yet clear and will be further investigated. The unfolding events are not complete unfolding events but rather unfolding into misfolded states, where G_L4 is repelled from the loop. These noncanonical states (known as G_L4 bulge out states^8–9) are still structurally close to the native one. Another recent study of ours¹⁰ with DESRES ff (using already the correct CHARMM22 ions) using yet different ionic conditions (only neutralizing cations were added) revealed the UUCG tetraloop to be more thermodynamically stable relative to the standard χ_OL3 RNA ff^11–15 (without gHBfix term¹). However, both the ffs resulted in the simulated 8-mer not having the native folded structure as its global minimum, in agreement with the findings reported here. Obviously, longer sequences can promote formation of the hairpin by stabilizing the stem, and this additional stabilization may be ff-dependent.

In summary, although our simulations are certainly not converged they demonstrate higher stability of the folded UUCG TL with the CHARMM22 ions. On the other hand, large-scale instability of some folded RNAs with the DESRES ff is observed with both ions. We suggest that the file introduced here containing the DESRES ff with the originally suggested CHARMM22 ion parameters is utilized in eventual future simulations using our implementation rather than the one with the JC ions.¹ For the sake of completeness we report updated Supporting Figures S1–S5 corresponding to Figures S13-S17 of the original paper;¹ within the sampling limits the data is identical. For the structural Figures visualizing large disruptions of the Kink-turn and the L1 stalk ribosomal RNAs see the original paper;¹ they would be entirely identical. We are grateful to Dr. Stefano Piana-Agostinetti for spotting the difference in used ion parameters. Core results of the original paper¹ (the gHBfix parameters and their testing) are unrelated to this Addition and Correction.