Rolf David

L’obtention, dans des conditions douces, sélectives et de chimie durable de molécules ciblées est aujourd’hui un enjeu majeur. Les métalloenzymesartificielles représentent une voie d’investigation importante, car en jouant, par exemple, sur la seconde sphère de coordination,il est possible de modifier fortement la réactivité de ces systèmes bio-inspirés. Le développement de cette chimie suppose une connaissanceapprofondie des différentes étapes du mécanisme de la réaction envisagée. Pour cela, la chimie théorique est essentielle à la rationalisation dela réactivité chimique mais elle souffre encore de nombreuses insuffisances pour les systèmes que nous nous proposons d’étudier.Dans ce travail, nous avons choisi d’étudier la superoxyde réductase, enzyme détoxifiante du radical superoxyde. Si de nombreuses expérimentalessont disponibles détaillant certains intermédiaires, le mécanisme précis est peu documenté. Le but a été de mettre en place uneméthodologie complète allant du développement de paramètres MM spécifiques à l’étude de la réactivité par métadynamiques QM/MM.Le développement de paramètres MM pour le site actif à fer a permis son étude en dynamique MM donnant des informations sur la conformationsdu squelette peptidique ainsi que l’interaction avec les molécules de solvant. De part la nature du fer, une description QM du site actifà été nécessaire via l’utilisation de DFT hybride. Les métadynamiques QM/MM ont permis quant à elles d’explorer les chemins réactionnelset de caractériser les espèces ainsi formées et les énergies d’activations.Cette méthodologie a permis la compréhension en premier lieu de la réactivité native de la forme sauvage et elle a aussi permis d’explorer lesréactivités nouvelles des mutations de la SOR permettant ainsi de définir le rôle crucial de la seconde sphère de coordination.

Enzyme catalysis can be simulated accurately and efficiently by coupling machine-learned potentials trained on gas-phase data to the environment (ML/MM) using electrostatic machine-learning embedding (EMLE).

To simulate enzyme reactions, multiscale quantum mechanics/molecular mechanics (QM/MM) approaches are well established and popular. However, accurately and efficiently estimating enzyme activity is a challenge, because in general, precise methods are too computationally expensive. Here, we demonstrate that enzyme catalysis can be captured by coupling efficient machine-learned potentials (MLPs) for a reaction to the wider enzyme environment using electrostatic machine-learning embedding (EMLE). Our EMLE model is first applied to the natural Diels-Alderase AbyU, showing that it correctly differentiates the catalytic action on different enzyme-substrate conformations. Then, we show that training a reaction-specific EMLE model allows us to accurately capture the enzyme catalytic effects of the conversion of chorismate to prephenate, a reaction with a highly polarizable and charged transition state. In both cases, in contrast to mechanical embedding approaches, EMLE embedding allows accurate and efficient predictions of enzyme activity, agreeing with high-level QM/MM reference calculations. This approach facilitates the use of gas phase-trained MLPs in MLP/molecular mechanics (ML/MM) simulations and should thus be highly beneficial for computational activity screening of enzyme biocatalysts.

Phosphoester bonds are fundamental components of both small molecules and polymeric biomolecules, most notably nucleic acids. Their formation from alcohol groups and phosphate is a significant challenge for living systems, as it involves substantial free-energy barriers. These barriers, while hindering bond formation, also ensure the kinetic stability of the resulting biomolecular structures. This duality presents a major obstacle in understanding the emergence of RNA-based systems at the origins of life, where protein enzymes capable of catalyzing such reactions were absent. Understanding the mechanisms underlying phosphoester formation is therefore essential for developing non-biological catalytic strategies. In a recent proof-of-concept study (Benayad et al. Proc. Natl. Acad. Sci. USA 121, e2322040121 (2024)), we demonstrated that machine-learned interatomic potentials (MLIPs) can be effectively trained and employed to sample the reactive pathways of the uncatalyzed reaction in water on microsecond timescales. In this work, we significantly expand upon our previous analysis to offer a detailed molecular characterization of all possible reaction pathways. We focus on the two protonation states of phosphate prevalent under near-neutral conditions (H2PO4- and HPO42-), using a range of enhanced sampling techniques. Special attention is given to the role of the solvent, elucidating its involvement in the reaction process, as well as the contributions of proton transfer events and oxygen substitution at the phosphorus center. We conclude by proposing molecular-level explanations for the observed differences in reactivity between the two phosphate protonation states.

To simulate enzyme reactions, multiscale quantum mechanics/molecular mechanics (QM/MM) approaches are well established and popular. However, accurately and efficiently estimating enzyme activity is a challenge, because in general, precise methods are too computationally expensive. Here, we demonstrate that enzyme catalysis can be captured by coupling efficient machine-learned potentials (MLPs) for a reaction to the wider enzyme environment using electrostatic machine-learning embedding (EMLE). Our EMLE model is first applied to the natural Diels-Alderase AbyU, showing that it correctly differentiates the catalytic action on different enzyme-substrate conformations. Then, we show that training a reaction-specific EMLE model allows us to accurately capture the enzyme catalytic effects of the conversion of chorismate to prephenate, a reaction with a highly polarizable and charged transition state. In both cases, in contrast to mechanical embedding approaches, EMLE embedding allows accurate and efficient predictions of enzyme activity, agreeing with high-level QM/MM reference calculations. This approach facilitates the use of gas phase-trained MLPs in MLP/molecular mechanics (ML/MM) simulations and should thus be highly beneficial for computational activity screening of enzyme biocatalysts.

ArcaNN is a comprehensive framework that employs concurrent learning to generate training datasets for reactive MLIPs in the condensed phase.

While RNA appears as a good candidate for the first autocatalytic systems preceding the emergence of modern life, the synthesis of RNA oligonucleotides without enzymes remains challenging. Because the uncatalyzed reaction is extremely slow, experimental studies bring limited and indirect information on the reaction mechanism, the nature of which remains debated. Here, we develop neural network potentials (NNPs) to study the phosphoester bond formation in water. While NNPs are becoming routinely applied to nonreactive systems or simple reactions, we demonstrate how they can systematically be trained to explore the reaction phase space for complex reactions involving several proton transfers and exchanges of heavy atoms. We then propagate at moderate computational cost hundreds of nanoseconds of a variety of enhanced sampling simulations with quantum accuracy in explicit solvent conditions. The thermodynamically preferred reaction pathway is a concerted, dissociative mechanism, with the transient formation of a metaphosphate transition state and direct participation of water solvent molecules that facilitate the exchange of protons through the nonbridging phosphate oxygens. Associative–dissociative pathways, characterized by a much tighter pentacoordinated phosphate, are higher in free energy. Our simulations also suggest that diprotonated phosphate, whose reactivity is never directly assessed in the experiments, is significantly less reactive than the monoprotonated species, suggesting that it is probably never the reactive species in normal pH conditions. These observations rationalize unexplained experimental results and the temperature dependence of the reaction rate, and they pave the way for the design of more efficient abiotic catalysts and activating groups.

The formation of an amide bond is an essential step in the synthesis of materials and drugs, and in the assembly of amino acids to form peptides. The mechanism of this reaction has been studied extensively, in particular to understand how it can be catalyzed, but a representation capable of explaining all the experimental data is still lacking. Numerical simulation should provide the necessary molecular description, but the solvent involvement poses a number of challenges. Here, we combine the efficiency and accuracy of neural network potential-based reactive molecular dynamics with the extensive and unbiased exploration of reaction pathways provided by transition path sampling. Using microsecond-scale simulations at the density functional theory level, we show that this method reveals the presence of two competing distinct mechanisms for peptide bond formation between alanine esters in aqueous solution. We describe how both reaction pathways, via a general base catalysis mechanism and via direct cleavage of the tetrahedral intermediate respectively, change with pH. This result contrasts with the conventional mechanism involving a single pathway in which only the barrier heights are affected by pH. We show that this new proposal involving two competing mechanisms is consistent with the experimental data, and we discuss the implications for peptide bond formation under prebiotic conditions and in the ribosome. Our work shows that integrating deep potential molecular dynamics with path sampling provides a powerful approach for exploring complex chemical mechanisms.

Recent advances in machine learning oer powerful tools for exploring complex reaction mechanisms in condensed phases via reactive simulations. In this tutorial review, we describe the key challenges associated with simulating reactions in condensed phases, we introduce neural network potentials and detail how they can be trained. We emphasize the importance of active learning to construct the training set, and show how these reactive force elds can be integrated with enhanced sampling techniques, including transition path sampling. We illustrate the capabilities of these new methods with a selection of applications to chemical reaction mechanisms in solution and at interfaces.