Journal of Chemical Theory and Computation最新文献

Electronic excited states are central to a vast arrayof physicaland chemical phenomena, yet accurate and efficient methods for preparingthem on quantum devices remain challenging and are comparatively underexplored.We introduce a general dissipative algorithm for selectively preparing ab initio electronic excited states. The key idea is torecast excited-state preparation as an effective ground-state problemby suitably modifying the underlying Lindblad dynamics so that thetarget excited state becomes the unique steady state of a designedquantum channel. We develop three complementary strategies tailoredto different types of prior information about the excited state, suchas symmetry and approximate energy. We demonstrate the effectivenessand versatility of these schemes through numerical simulations ofatomic and molecular spectra, including valence excitations in prototypicalplanar-conjugated molecules and transition-metal complexes. Takentogether, these results provide a new pathway for advancing quantumsimulation methods for realistic, strongly correlated electronic systems.

Accurate modeling of electronically excited states insolutionis essential for the rational design of photoactive molecular systems,but requires a reliable treatment of environmental effects. We presenta mean-field approach that combines density functional theory witha polarizable continuum model (PCM) to account for environmental dielectricscreening, providing an efficient mean-field alternative to higher-costwavefunction- or many-body-based approaches. The framework employsdielectric-screened range-separated hybrid (SRSH) functionals withinPCM to obtain excited states of solvated azobenzene derivatives, aprototypical molecular photoswitch. SRSH–PCM excitation energiesbetter reproduce experiment when the larger static dielectric constant,which also accounts for nuclear polarization, is used to define long-rangescreening. Using the smaller optical dielectric constant, which representspurely electronic polarization, SRSH–PCM energies align betterwith previously reported high-quality GW–Bethe–Salpeterequation (GW/BSE) energies [Kshirsagar et al., JCTC 2020]. Our findingsdemonstrate that mean-field SRSH–PCM provides an effectivedescription of excited-state properties while offering insights intothe role of environmental screening and, in particular, the importanceof accounting for nuclear contributions to dielectric screening inmany-body calculations of solution-phase excitation energies.

The calculation of accurate thermodynamic properties,such as activationand reaction free energies, is an essential tool for elucidating reactionmechanisms. The method most commonly employed, density functionaltheory (DFT), generally falls short of chemical accuracy (1 kcal/molerror), resulting in order-of-magnitude errors in the calculationof equilibrium constants and rate constants. The inaccuracy of commonexchange-correlation functionals is compounded by the use of smalland medium-sized basis sets, which are mandatory for molecular systemswith realistic sizes. The basis set incompleteness error incurredby the use of these basis sets further diminishes the accuracy ofthermochemical calculations. In previous works, we proposed energycorrections based on the use of atom-centered potentials (ACP) asa way of mitigating functional and basis set incompleteness error.ACPs are one-electron potentials that, when applied in combinationwith a functional and basis set, yield results comparable to a higherlevel of theory with little additional cost. In this work, and basedon previous studies, we develop ACPs for thermochemical calculationsbased on the M06-2X density functional and the 6-31+G(d,p) basis.The resulting M06-2X/6-31+G(d,p)-ACP method is tested on a varietyof cases, demonstrating excellent performance for thermochemistryand kinetics calculations well outside the parametrization set, particularlyfor reaction energies and barrier heights of moderately large systems.At the same time, M06-2X/6-31+G(d,p)-ACP retains the low cost of theuncorrected functional and basis set combination.

We introduce quantumMonte Carlo (QMC)-trained multistate machine-learned(ML) force fields for nonadiabatic excited-state dynamics, targetingphotochemical processes in which the electronic character changesalong the reaction path and a consistent correlated description isrequired. In this framework, variational Monte Carlo (VMC) wave functionscombine compact selected configuration-interaction (CIPSI) expansionswith a Jastrow factor that explicitly accounts for dynamical correlation,while neural networks convert the stochastic VMC/CIPSI data into smoothpotential energy surfaces for large surface-hopping ensembles. Weapply this approach to azomethane, a demanding test case involvingtorsional relaxation through conical-intersection regions and C–Nbond dissociation on the hot ground state. Benchmark calculationssupport the accuracy of the QMC reference data and show robust forceconvergence across isomerization and dissociation geometries. TheQMC-trained dynamics preserves the expected photoisomerization mechanism,strongly reduces the excessive C–N breaking obtained with completeactive space self-consistent field, and predicts a small but non-negligibleprompt dissociation component after internal conversion, with a timescale consistent with femtosecond-resolved mass-spectrometry experiments.These results establish ML-QMC as a practical route to nonadiabaticphotochemical dynamics with accurate wave function reference data.

Chemical potentials are among the most important propertiesthatcan be obtained from a molecular simulation since they define manytechnologically relevant collective properties such as solubilitiesand activity coefficients. The chemical potential of a species insolution is typically obtained by computing the free energy changeof adding that species into a bulk system, a calculation typicallyvery expensive for systems of technological interest such as electrolytes,due to the lack of phase space overlap between “not-inserted″and “inserted″ states. Recently, normalizing flows havebeen introduced as a way to accelerate free energy computations bylearning a bijective function to map the configuration space of oneBoltzmann distribution onto another. Currently, these trainable mappingsare constructed to be as expressive as possible, theoretically havingthe ability to create a perfect mapping between states. This expressivitymakes them difficult to train, limiting their ability to be generated“on-the-fly″ for any new free energy estimation challenge,and in practice these mappings have shown only modest performanceimprovements for liquid systems. We address these issues by introducingthe concept of a “minimal″ normalizing flow (MNF). Thisis a trainable bijective mapping that is intentionally limited inexpressivity and instead applies low-dimensional, physically informedtransformations. Useful MNFs can be trained in approximately 1 minof GPU time due to their simplicity and our introduction of a novelloss function based on the Bhattacharyya distance as an alternativeto the Kullback–Leibler divergence. We show how calculationsof chemical potentials of pure and binary Lennard-Jones particle systemscan be accelerated by at least an order of magnitude with a simpleradial mapping. For a more complex illustration of the approach, weapply a two-dimensional (radial and orientational) mapping to thesolvation of sodium and chloride ions in water, showing that MNFscan increase the effective sample size by as much as 3 times for chargingfree energy calculations and as much as 8 times for calculating freeenergy changes due to force field perturbations. This provides thefoundation for the development of physically informed normalizingflows that can accelerate complex free energy calculations while retaininglow training costs.

Molecular conformations play a critical role in determiningmolecularproperties, such as membrane permeability, binding affinity, and ultimatelytherapeutic efficacy. Experimental and computational approaches cancharacterize conformations and provide insight into why certain conformationsare thermodynamically preferred over others. However, examining staticconformations alone may not fully explain why subtle differences,such as a single LEU-to-ILE mutation in a bicyclic peptide, can producemarkedly distinct conformational ensembles. Furthermore, analyzingtransition pathways between conformations reveals the mechanisms thatshape these ensembles. Here, we introduce a deep learning model, termedICoN-v1, trained on molecular dynamics (MD) simulation data to learnthe underlying physics that governs cyclic peptide conformationaldynamics. We examined cyclic hexapeptides with Nuclear Magnetic Resonance(NMR)-determined structures, and MYC-targeting bicyclic peptides thatare stereo-diversified or have a single LEU-to-ILE mutation. By followingminimum-energy pathways in the latent space constructed by ICoN-v1,we efficiently generated smooth conformational transition paths. Thesepathways reveal sequential sets of concerted backbone and side-chaintorsional rotations moving between energy minima. Notably, smoothtransition pathways that are absent from MD output were observed usingICoN-v1. Our results identify various sets of concerted torsionalmotions that are nonlinearly combined during conformational transitionsand reveal the key residues governing each stage of the transition,thereby elucidating how the observed conformations are generated andinforming molecular design.

Oxidosqualenecyclases (OSCs) convert linear 2,3-oxidosqualeneinto diverse polycyclic triterpenoids. The precise control over highlyreactive carbocation intermediates, particularly the ultimate quenchingmechanism (deprotonation vs hydroxylation), remains a fundamentalchallenge in mechanism-driven enzyme reshaping. Here, we investigateAaOSC-20433, a rare OSC from Artemisia argyi exhibitingabsolute specificity for pure deprotonation to produce dammara-20,24-dien-3β-ol.This starkly contrasts with the homologous dammarenediol synthasefrom Panax ginseng (PgDDS), which favors hydroxylationto yield dammarenediol II. Integrating multiscale QM/MM simulationsand site-directed mutagenesis, we modeled the impact of crucial activesite mutations on the reaction cascade. Our simulations revealed thatS411 acts as a general base governing the terminal deprotonation.More importantly, we demonstrate that mechanism-guided active sitesubstitutions can dramatically alter the chemical trajectory. Relaxingsteric constraints (F728A) induces catalytic promiscuity to yieldtricyclic and tetracyclic products. Furthermore, the L259Y mutationsignificantly alters the quenching network. It unlocks the hydroxylationpathway to form dammarenediol II and acts as an engineered base extendingcyclization depth to yield the pentacyclic lupeol. Notably, combiningthese functional loci (L259Y+F728S) reveals notable synergistic epistasisbetween the steric cavity and the proton-transfer network, drivingcompetitive hydration events to diversify the hydroxylated productprofile. This study provides a comprehensive mechanistic frameworkfor understanding how OSC active site mutations govern product specificity,offering a robust theoretical basis for the tailored, mechanism-drivenfunctional reprogramming of triterpene synthases in synthetic biology.

Root-mean-square deviation (RMSD) is the standard metricof structuralcomparison in molecular dynamics (MD) simulations. In its conventionalform, RMSD assigns equal weight to all atoms regardless of mobility.Hence, flexible loops and disordered regions can dominate a globalRMSD, while the rigid functional core contributes negligibly to theoverall metric. To address this issue, we introduce the Bayes-optimalRMSD (BRMSD), which optimizes per-atom weights jointlywith structural averages by maximizing a Bayesian posterior. In atrade-off between low RMSD and weight uniformity, a position-fluctuationparameter σ controls the transition from classical RMSD (σ→ ∞) to a progressive focus on a rigid core (σ→ 0). The BRMSD framework supports analysis modules for structuralalignment, focused alignment onto a user-specified domain, trajectorysmoothing, soft K-means conformational clustering,and rigid-domain identification. These modules are implemented inthe open-source Python package BRMSD and benchmarkedon two MD systems, the endoplasmic reticulum translocon-associatedprotein SND3 and the phosphotransferase adenylate kinase.

Bootstrap Embedding (BE) has recentlydemonstrated animpressiveability to produce highly accurate single-point energies at a lowcomputational cost, utilizing systematically determined fragmentsthat overlap with each other. In many chemical applications, suchas scanning the potential energy surface, one performs numerous single-pointcalculations that differ by local geometric changes. This often leadsto a frustratingly time-consuming process, even with computationallyaffordable methods like BE. We present here a locally updated versionof BE that dramatically accelerates these types of calculations witha minimal loss in accuracy. Specifically, we propose a method thatupdates only the fragments that involve the change in geometry, whilereusing information from a single correlated calculation at a referencegeometry for all other fragments. We illustrate the key characteristicsof this method and compare it with conventional BE. This novel schemecan potentially lead to a correlation method that scales sub-linearlyin time with respect to system size. It exhibits chemical accuracy(<1 kcal/mol energy error) for bond stretches of about 1 Åor rotations of tens of degrees relative to the reference geometry,while recomputing only a small number of fragments, independent ofsystem size. We discuss potential applications in simulating realisticchemical phenomena, such as scanning the potential energy surface,reaction pathway studies, molecular dynamics simulations, and geometryoptimizations.

Machine-learned interatomicpotentials hold the promise to enablethe modeling of highly concentrated liquids over meaningful time scales,far from reach for current ab initio electronic structuremethods. Here we evaluate the performances of various MACE potentialsin modeling a 21 m water-in-salt electrolyte based on lithium bis(trifluoromethanesulfonyl)imide.We test out-of-the-box foundation models, as well as both fine-tuningand from-scratch training strategies. Our simulations demonstratethat surrogate models allow us to overcome sampling limitations of ab initio molecular dynamics, reaching an excellent agreementwith experimental observables such as the structure factor. We alsodemonstrate the benefit of fine-tuning a foundation model over trainingfrom scratch, in terms of data efficiency, but most importantly asa means to provide information regarding configurations hard to sample,such as short Li⁺–Li⁺ distances. Finally,we show that depending on the reference exchange-correlation functional,empirical dispersion correction schemes can be detrimental. All inall, our work shows that machine-learned interatomic potentials area good fit for the modeling of highly concentrated electrolytes overlong time scales.