Machine-learned potentials for next-generation matter simulations.

The choice of simulation methods in computational materials science is driven by a fundamental trade-off: bridging large time- and length-scales with highly accurate simulations at an affordable computational cost. Venturing the investigation of complex phenomena on large scales requires fast yet accurate computational methods. We review the emerging field of machine-learned potentials, which promises to reach the accuracy of quantum mechanical computations at a substantially reduced computational cost. This Review will summarize the basic principles of the underlying machine learning methods, the data acquisition process and active learning procedures. We highlight multiple recent applications of machine-learned potentials in various fields, ranging from organic chemistry and biomolecules to inorganic crystal structure predictions and surface science. We furthermore discuss the developments required to promote a broader use of ML potentials, and the possibility of using them to help solve open questions in materials science and facilitate fully computational materials design.

Free energy analysis and mechanism of base pair stacking in nicked DNA.

The equilibrium of stacked and unstacked base pairs is of central importance for all nucleic acid structure formation processes. The stacking equilibrium is influenced by intramolecular interactions between nucleosides but also by interactions with the solvent. Realistic simulations on nucleic acid structure formation and flexibility require an accurate description of the stacking geometry and stability and its sequence dependence. Free energy simulations have been conducted on a series of double stranded DNA molecules with a central strand break (nick) in one strand. The change in free energy upon unstacking was calculated for all ten possible base pair steps using umbrella sampling along a center-of-mass separation coordinate and including a comparison of different water models. Comparison to experimental studies indicates qualitative agreement of the stability order but a general overestimation of base pair stacking interactions in the simulations. A significant dependence of calculated nucleobase stacking free energies on the employed water model was observed with the tendency of stacking free energies being more accurately reproduced by more complex water models. The simulation studies also suggest a mechanism of stacking/unstacking that involves significant motions perpendicular to the reaction coordinate and indicate that the equilibrium nicked base pair step may slightly differ from regular B-DNA geometry in a sequence-dependent manner.

A compact native 24-residue supersecondary structure derived from the villin headpiece subdomain.

Many small proteins fold highly cooperatively in an all-or-none fashion and thus their native states are well protected from thermal fluctuations by an extensive network of interactions across the folded structure. Because protein structures are stabilized by local and nonlocal interactions among distal residues, dissecting individual substructures from the context of folded proteins results in large destabilization and loss of unique three-dimensional structure. Thus, mini-protein substructures can only rarely be derived from natural templates. Here, we describe a compact native 24-residues-long supersecondary structure derived from the hyperstable villin headpiece subdomain consisting of helices 2 and 3 (HP24). Using a combination of experimental techniques, including NMR and small-angle x-ray scattering, as well as all-atom replica exchange molecular-dynamics simulations, we show that a variant with stabilizing substitutions (HP24stab) forms a densely packed and compact conformation. In HP24stab, interactions between helices 2 and 3 are similar to those observed in native folded HP35, and the two helices cooperatively stabilize each other by completing the hydrophobic core lining the central part of HP35. Interestingly, even though the HP24wt fragment shows a more expanded and less structured conformation, NMR and simulations demonstrate a preference for a native-like topology. Thus, the two stabilizing residues in HP24stab shift the energy balance toward the native state, leading to a minimal folding motif.

Machine learning models to accelerate the design of polymeric long-acting injectables.

Long-acting injectables are considered one of the most promising therapeutic strategies for the treatment of chronic diseases as they can afford improved therapeutic efficacy, safety, and patient compliance. The use of polymer materials in such a drug formulation strategy can offer unparalleled diversity owing to the ability to synthesize materials with a wide range of properties. However, the interplay between multiple parameters, including the physicochemical properties of the drug and polymer, make it very difficult to intuitively predict the performance of these systems. This necessitates the development and characterization of a wide array of formulation candidates through extensive and time-consuming in vitro experimentation. Machine learning is enabling leap-step advances in a number of fields including drug discovery and materials science. The current study takes a critical step towards data-driven drug formulation development with an emphasis on long-acting injectables. Here we show that machine learning algorithms can be used to predict experimental drug release from these advanced drug delivery systems. We also demonstrate that these trained models can be used to guide the design of new long acting injectables. The implementation of the described data-driven approach has the potential to reduce the time and cost associated with drug formulation development.

On scientific understanding with artificial intelligence.

An oracle that correctly predicts the outcome of every particle physics experiment, the products of every possible chemical reaction or the function of every protein would revolutionize science and technology. However, scientists would not be entirely satisfied because they would want to comprehend how the oracle made these predictions. This is scientific understanding, one of the main aims of science. With the increase in the available computational power and advances in artificial intelligence, a natural question arises: how can advanced computational systems, and specifically artificial intelligence, contribute to new scientific understanding or gain it autonomously? Trying to answer this question, we adopted a definition of 'scientific understanding' from the philosophy of science that enabled us to overview the scattered literature on the topic and, combined with dozens of anecdotes from scientists, map out three dimensions of computer-assisted scientific understanding. For each dimension, we review the existing state of the art and discuss future developments. We hope that this Perspective will inspire and focus research directions in this multidisciplinary emerging field.

Golem: an algorithm for robust experiment and process optimization.

Numerous challenges in science and engineering can be framed as optimization tasks, including the maximization of reaction yields, the optimization of molecular and materials properties, and the fine-tuning of automated hardware protocols. Design of experiment and optimization algorithms are often adopted to solve these tasks efficiently. Increasingly, these experiment planning strategies are coupled with automated hardware to enable autonomous experimental platforms. The vast majority of the strategies used, however, do not consider robustness against the variability of experiment and process conditions. In fact, it is generally assumed that these parameters are exact and reproducible. Yet some experiments may have considerable noise associated with some of their conditions, and process parameters optimized under precise control may be applied in the future under variable operating conditions. In either scenario, the optimal solutions found might not be robust against input variability, affecting the reproducibility of results and returning suboptimal performance in practice. Here, we introduce Golem, an algorithm that is agnostic to the choice of experiment planning strategy and that enables robust experiment and process optimization. Golem identifies optimal solutions that are robust to input uncertainty, thus ensuring the reproducible performance of optimized experimental protocols and processes. It can be used to analyze the robustness of past experiments, or to guide experiment planning algorithms toward robust solutions on the fly. We assess the performance and domain of applicability of Golem through extensive benchmark studies and demonstrate its practical relevance by optimizing an analytical chemistry protocol under the presence of significant noise in its experimental conditions.

Machine learning directed drug formulation development.

Machine learning (ML) has enabled ground-breaking advances in the healthcare and pharmaceutical sectors, from improvements in cancer diagnosis, to the identification of novel drugs and drug targets as well as protein structure prediction. Drug formulation is an essential stage in the discovery and development of new medicines. Through the design of drug formulations, pharmaceutical scientists can engineer important properties of new medicines, such as improved bioavailability and targeted delivery. The traditional approach to drug formulation development relies on iterative trial-and-error, requiring a large number of resource-intensive and time-consuming in vitro and in vivo experiments. This review introduces the basic concepts of ML-directed workflows and discusses how these tools can be used to aid in the development of various types of drug formulations. ML-directed drug formulation development offers unparalleled opportunities to fast-track development efforts, uncover new materials, innovative formulations, and generate new knowledge in drug formulation science. The review also highlights the latest artificial intelligence (AI) technologies, such as generative models, Bayesian deep learning, reinforcement learning, and self-driving laboratories, which have been gaining momentum in drug discovery and chemistry and have potential in drug formulation development.

Data-science driven autonomous process optimization.

Autonomous process optimization involves the human intervention-free exploration of a range process parameters to improve responses such as product yield and selectivity. Utilizing off-the-shelf components, we develop a closed-loop system for carrying out parallel autonomous process optimization experiments in batch. Upon implementation of our system in the optimization of a stereoselective Suzuki-Miyaura coupling, we find that the definition of a set of meaningful, broad, and unbiased process parameters is the most critical aspect of successful optimization. Importantly, we discern that phosphine ligand, a categorical parameter, is vital to determination of the reaction outcome. To date, categorical parameter selection has relied on chemical intuition, potentially introducing bias into the experimental design. In seeking a systematic method for selecting a diverse set of phosphine ligands, we develop a strategy that leverages computed molecular feature clustering. The resulting optimization uncovers conditions to selectively access the desired product isomer in high yield.

Designing and understanding light-harvesting devices with machine learning.

Understanding the fundamental processes of light-harvesting is crucial to the development of clean energy materials and devices. Biological organisms have evolved complex metabolic mechanisms to efficiently convert sunlight into chemical energy. Unraveling the secrets of this conversion has inspired the design of clean energy technologies, including solar cells and photocatalytic water splitting. Describing the emergence of macroscopic properties from microscopic processes poses the challenge to bridge length and time scales of several orders of magnitude. Machine learning experiences increased popularity as a tool to bridge the gap between multi-level theoretical models and Edisonian trial-and-error approaches. Machine learning offers opportunities to gain detailed scientific insights into the underlying principles governing light-harvesting phenomena and can accelerate the fabrication of light-harvesting devices.

From Absorption Spectra to Charge Transfer in Nanoaggregates of Oligomers with Machine Learning.

Fast and inexpensive characterization of materials properties is a key element to discover novel functional materials. In this work, we suggest an approach employing three classes of Bayesian machine learning (ML) models to correlate electronic absorption spectra of nanoaggregates with the strength of intermolecular electronic couplings in organic conducting and semiconducting materials. As a specific model system, we consider poly(3,4-ethylenedioxythiophene) (PEDOT) polystyrene sulfonate, a cornerstone material for organic electronic applications, and so analyze the couplings between charged dimers of closely packed PEDOT oligomers that are at the heart of the material's unrivaled conductivity. We demonstrate that ML algorithms can identify correlations between the coupling strengths and the electronic absorption spectra. We also show that ML models can be trained to be transferable across a broad range of spectral resolutions and that the electronic couplings can be predicted from the simulated spectra with an 88% accuracy when ML models are used as classifiers. Although the ML models employed in this study were trained on data generated by a multiscale computational workflow, they were able to leverage experimental data.

ChemOS: An orchestration software to democratize autonomous discovery.

The current Edisonian approach to discovery requires up to two decades of fundamental and applied research for materials technologies to reach the market. Such a slow and capital-intensive turnaround calls for disruptive strategies to expedite innovation. Self-driving laboratories have the potential to provide the means to revolutionize experimentation by empowering automation with artificial intelligence to enable autonomous discovery. However, the lack of adequate software solutions significantly impedes the development of self-driving laboratories. In this paper, we make progress towards addressing this challenge, and we propose and develop an implementation of ChemOS; a portable, modular and versatile software package which supplies the structured layers necessary for the deployment and operation of self-driving laboratories. ChemOS facilitates the integration of automated equipment, and it enables remote control of automated laboratories. ChemOS can operate at various degrees of autonomy; from fully unsupervised experimentation to actively including inputs and feedbacks from researchers into the experimentation loop. The flexibility of ChemOS provides a broad range of functionality as demonstrated on five applications, which were executed on different automated equipment, highlighting various aspects of the software package.

Beyond Ternary OPV: High-Throughput Experimentation and Self-Driving Laboratories Optimize Multicomponent Systems.

Fundamental advances to increase the efficiency as well as stability of organic photovoltaics (OPVs) are achieved by designing ternary blends, which represents a clear trend toward multicomponent active layer blends. The development of high-throughput and autonomous experimentation methods is reported for the effective optimization of multicomponent polymer blends for OPVs. A method for automated film formation enabling the fabrication of up to 6048 films per day is introduced. Equipping this automated experimentation platform with a Bayesian optimization, a self-driving laboratory is constructed that autonomously evaluates measurements to design and execute the next experiments. To demonstrate the potential of these methods, a 4D parameter space of quaternary OPV blends is mapped and optimized for photostability. While with conventional approaches, roughly 100 mg of material would be necessary, the robot-based platform can screen 2000 combinations with less than 10 mg, and machine-learning-enabled autonomous experimentation identifies stable compositions with less than 1 mg.

How machine learning can assist the interpretation of ab initio molecular dynamics simulations and conceptual understanding of chemistry.

Molecular dynamics simulations are often key to the understanding of the mechanism, rate and yield of chemical reactions. One current challenge is the in-depth analysis of the large amount of data produced by the simulations, in order to produce valuable insight and general trends. In the present study, we propose to employ recent machine learning analysis tools to extract relevant information from simulation data without a priori knowledge on chemical reactions. This is demonstrated by training machine learning models to predict directly a specific outcome quantity of ab initio molecular dynamics simulations - the timescale of the decomposition of 1,2-dioxetane. The machine learning models accurately reproduce the dissociation time of the compound. Keeping the aim of gaining physical insight, it is demonstrated that, in order to make accurate predictions, the models evidence empirical rules that are, today, part of the common chemical knowledge. This opens the way for conceptual breakthroughs in chemistry where machine analysis would provide a source of inspiration to humans.

Oscillatory Active-site Motions Correlate with Kinetic Isotope Effects in Formate Dehydrogenase.

Thermal motions of enzymes have been invoked to explain the temperature dependence of kinetic isotope effects (KIE) in enzyme-catalyzed hydride transfers. Formate dehydrogenase (FDH) from Candida boidinii exhibits a temperature independent KIE that becomes temperature dependent upon mutation of hydrophobic residues in the active site. Ternary complexes of FDH that mimic the transition state structure allow investigation of how these mutations influence active-site dynamics. A combination of X-ray crystallography, two-dimensional infrared (2D IR) spectroscopy, and molecular dynamic simulations characterize the structure and dynamics of the active site. FDH exhibits oscillatory frequency fluctuations on the picosecond timescale, and the amplitude of these fluctuations correlates with the temperature dependence of the KIE. Both the kinetic and dynamic phenomena can be reproduced computationally. These results provide experimental evidence for a connection between the temperature dependence of KIEs and motions of the active site in an enzyme-catalyzed reaction consistent with activated tunneling models of the hydride transfer reaction.

Chimera: enabling hierarchy based multi-objective optimization for self-driving laboratories.

Finding the ideal conditions satisfying multiple pre-defined targets simultaneously is a challenging decision-making process, which impacts science, engineering, and economics. Additional complexity arises for tasks involving experimentation or expensive computations, as the number of evaluated conditions must be kept low. We propose Chimera as a general purpose achievement scalarizing function for multi-target optimization where evaluations are the limiting factor. Chimera combines concepts of a priori scalarizing with lexicographic approaches and is applicable to any set of n unknown objectives. Importantly, it does not require detailed prior knowledge about individual objectives. The performance of Chimera is demonstrated on several well-established analytic multi-objective benchmark sets using different single-objective optimization algorithms. We further illustrate the applicability and performance of Chimera with two practical examples: (i) the auto-calibration of a virtual robotic sampling sequence for direct-injection, and (ii) the inverse-design of a four-pigment excitonic system for an efficient energy transport. The results indicate that Chimera enables a wide class of optimization algorithms to rapidly find ideal conditions. Additionally, the presented applications highlight the interpretability of Chimera to corroborate design choices for tailoring system parameters.

Phoenics: A Bayesian Optimizer for Chemistry.

We report Phoenics, a probabilistic global optimization algorithm identifying the set of conditions of an experimental or computational procedure which satisfies desired targets. Phoenics combines ideas from Bayesian optimization with concepts from Bayesian kernel density estimation. As such, Phoenics allows to tackle typical optimization problems in chemistry for which objective evaluations are limited, due to either budgeted resources or time-consuming evaluations of the conditions, including experimentation or enduring computations. Phoenics proposes new conditions based on all previous observations, avoiding, thus, redundant evaluations to locate the optimal conditions. It enables an efficient parallel search based on intuitive sampling strategies implicitly biasing toward exploration or exploitation of the search space. Our benchmarks indicate that Phoenics is less sensitive to the response surface than already established optimization algorithms. We showcase the applicability of Phoenics on the Oregonator, a complex case-study describing a nonlinear chemical reaction network. Despite the large search space, Phoenics quickly identifies the conditions which yield the desired target dynamic behavior. Overall, we recommend Phoenics for rapid optimization of unknown expensive-to-evaluate objective functions, such as experimentation or long-lasting computations.

ChemOS: Orchestrating autonomous experimentation.

ChemOS aims to catalyze the expansion of autonomous laboratories and to disrupt the conventional approach to experimentation.

Machine learning for quantum dynamics: deep learning of excitation energy transfer properties.

Understanding the relationship between the structure of light-harvesting systems and their excitation energy transfer properties is of fundamental importance in many applications including the development of next generation photovoltaics. Natural light harvesting in photosynthesis shows remarkable excitation energy transfer properties, which suggests that pigment-protein complexes could serve as blueprints for the design of nature inspired devices. Mechanistic insights into energy transport dynamics can be gained by leveraging numerically involved propagation schemes such as the hierarchical equations of motion (HEOM). Solving these equations, however, is computationally costly due to the adverse scaling with the number of pigments. Therefore virtual high-throughput screening, which has become a powerful tool in material discovery, is less readily applicable for the search of novel excitonic devices. We propose the use of artificial neural networks to bypass the computational limitations of established techniques for exploring the structure-dynamics relation in excitonic systems. Once trained, our neural networks reduce computational costs by several orders of magnitudes. Our predicted transfer times and transfer efficiencies exhibit similar or even higher accuracies than frequently used approximate methods such as secular Redfield theory.

Absence of Selection for Quantum Coherence in the Fenna-Matthews-Olson Complex: A Combined Evolutionary and Excitonic Study.

We present a study on the evolution of the Fenna-Matthews-Olson bacterial photosynthetic pigment-protein complex. This protein complex functions as an antenna. It transports absorbed photons-excitons-to a reaction center where photosynthetic reactions initiate. The efficiency of exciton transport is therefore fundamental for the photosynthetic bacterium's survival. We have reconstructed an ancestor of the complex to establish whether coherence in the exciton transport was selected for or optimized over time. We have also investigated the role of optimizing free energy variation upon folding in evolution. We studied whether mutations which connect the ancestor to current day species were stabilizing or destabilizing from a thermodynamic viewpoint. From this study, we established that most of these mutations were thermodynamically neutral. Furthermore, we did not see a large change in exciton transport efficiency or coherence, and thus our results predict that exciton coherence was not specifically selected for.

Machine learning exciton dynamics.

Obtaining the exciton dynamics of large photosynthetic complexes by using mixed quantum mechanics/molecular mechanics (QM/MM) is computationally demanding. We propose a machine learning technique, multi-layer perceptrons, as a tool to reduce the time required to compute excited state energies. With this approach we predict time-dependent density functional theory (TDDFT) excited state energies of bacteriochlorophylls in the Fenna-Matthews-Olson (FMO) complex. Additionally we compute spectral densities and exciton populations from the predictions. Different methods to determine multi-layer perceptron training sets are introduced, leading to several initial data selections. In addition, we compute spectral densities and exciton populations. Once multi-layer perceptrons are trained, predicting excited state energies was found to be significantly faster than the corresponding QM/MM calculations. We showed that multi-layer perceptrons can successfully reproduce the energies of QM/MM calculations to a high degree of accuracy with prediction errors contained within 0.01 eV (0.5%). Spectral densities and exciton dynamics are also in agreement with the TDDFT results. The acceleration and accurate prediction of dynamics strongly encourage the combination of machine learning techniques with ab initio methods.