Here, we present an update of the CHARMM27 all-atom additive force field for nucleic acids that improves the treatment of RNA molecules. The original CHARMM27 force field parameters exhibit enhanced Watson-Crick (WC) base pair opening which is not consistent with experiment while analysis of MD simulations show the 2′-hydroxyl moiety to almost exclusively sample the O3′ orientation. Quantum mechanical studies of RNA related model compounds indicate the energy minimum associated with the O3′ orientation to be too favorable, consistent with the MD results. Optimization of the dihedral parameters dictating the energy of the 2′-hydroxyl proton targeting the QM data yielded several parameter sets, which sample both the base and O3′ orientations of the 2′-hydroxyl to varying degrees. Selection of the final dihedral parameters was based on reproduction of hydration behavior as related to a survey of crystallographic data and better agreement with experimental NMR J-coupling values. Application of the model, designated CHARMM36, to a collection of canonical and non-canonical RNA molecules reveals overall improved agreement with a range of experimental observables as compared to CHARMM27. The results also indicate the sensitivity of the conformational heterogeneity of RNA to the orientation of the 2′-hydroxyl moiety and support a model whereby the 2′-hydroxyl can enhance the probability of conformational transitions in RNA.

Application of deep learning techniques for de novo generation of molecules, termed as inverse molecular design, has been gaining enormous traction in drug design. The representation of molecules in SMILES notation as a string of characters enables the usage of state of the art models in natural language processing, such as Transformers, for molecular design in general. Inspired by generative pre-training (GPT) models that have been shown to be successful in generating meaningful text, we train a transformer-decoder on the next token prediction task using masked self-attention for the generation of druglike molecules in this study. We show that our model, MolGPT, performs on par with other previously proposed modern machine learning frameworks for molecular generation in terms of generating valid, unique, and novel molecules. Furthermore, we demonstrate that the model can be trained conditionally to control multiple properties of the generated molecules. We also show that the model can be used to generate molecules with desired scaffolds as well as desired molecular properties by conditioning the generation on scaffold SMILES strings of desired scaffolds and property values. Using saliency maps, we highlight the interpretability of the generative process of the model.

No abstract

Urea has long been used to probe the stability and folding kinetics of proteins. 1 In contrast only recently it was shown that the RNA molecules that have a high propensity to misfold can be resolved using moderate amounts of urea. 2 Urea titrations can also be used to probe the interactions that stabilize the folded states of RNA. 2c Although the mechanism by which urea denatures proteins is now fairly well understood 3 the nature of interactions by which urea destabilizes RNA is not known. In order to provide a microscopic basis for the action of urea on RNA we have carried out extensive all atom molecular dynamics (MD) simulations on two RNA constructs using two urea force fields. Destabilization of RNA is due to disruption of base-pair interactions by direct hydrogen bonding of urea with the bases. The simulations also reveal a novel mechanism in which urea molecules engage in stacking interactions with the purine bases. 4 Analyses of 20 ns trajectories generated using MD simulations with a urea force field that was created as a part of the present work (see SI for simulation details, SI Figs. 1 and 2 and Tables 1-3 for urea parameter development, and for assessing the validity of the force field) of the 22-nucleotide RNA hairpin P5GA 5 (Fig. 1A) in various urea concentrations ([C]s) reveal that at high [C] the solvent-exposed stem regions lead to disruption of base pairing. The fraction of intact hydrogen bonds associated with the bases in the stem decreases from about 0.71 in the absence of urea to 0.46 in 8M urea. The loss of the Watson-Crick (WC) hydrogen bonds is accompanied by opening of the base pairs, which is reflected in the distribution of the hydrogen bond donor-acceptor distances (R HB ) in the hairpin stem (Fig. 1B). The base-paired state is indicated by a sharp peak at R HB = 3Å, whose height decreases as [C] increases to 6M. The probability of sampling R HB distances that are greater than 10Å (Fig. 1B) increases greatly in high [C], which results in a rotation of the bases of the helix leading to N1-N3 distances of about 16Å. 6 Examination of opening at the individual base pair level reveals considerable heterogeneity 7 with the largest fluctuations occurring at the GA and GU mismatches. We also show that urea-induced disruption of the base opening due to the loss of WC hydrogen bonds is nonspecific in the sense that urea does not preferentially interact with a specific base pair. These finding suggests that denaturation of RNA is due to favorable non-specific interactions with amide-like surfaces of the nucleic acids. The average base-base interaction energies (GC, thirum@umd.edu. amackere@rx.umaryland Table 4). When averaged over all base pair interactions in the stem the interactions become less favorable by about 2.7 kcal/mol at 6M relative to [C] = 0 (SI Table 4). The average interaction energies for certain base pairs (for example A6G17 and U8A15) are substantially less at high [C] relative to their values in water (see SI Table 4).In contrast, the backbone conformational proper...