Machine learning addresses the question of how to build computers that improve automatically through experience. It is one of today's most rapidly growing technical fields, lying at the intersection of computer science and statistics, and at the core of artificial intelligence and data science. Recent progress in machine learning has been driven both by the development of new learning algorithms and theory and by the ongoing explosion in the availability of online data and low-cost computation. The adoption of data-intensive machine-learning methods can be found throughout science, technology and commerce, leading to more evidence-based decision-making across many walks of life, including health care, manufacturing, education, financial modeling, policing, and marketing. Copyright © 2015, American Association for the Advancement of Science.

Organic synthesis is one of the key stumbling blocks in medicinal chemistry. A necessary yet unsolved step in planning synthesis is solving the forward problem: Given reactants and reagents, predict the products. Similar to other work, we treat reaction prediction as a machine translation problem between simplified molecular-input line-entry system (SMILES) strings (a text-based representation) of reactants, reagents, and the products. We show that a multihead attention Molecular Transformer model outperforms all algorithms in the literature, achieving a top-1 accuracy above 90% on a common benchmark data set. Molecular Transformer makes predictions by inferring the correlations between the presence and absence of chemical motifs in the reactant, reagent, and product present in the data set. Our model requires no handcrafted rules and accurately predicts subtle chemical transformations. Crucially, our model can accurately estimate its own uncertainty, with an uncertainty score that is 89% accurate in terms of classifying whether a prediction is correct. Furthermore, we show that the model is able to handle inputs without a reactant-reagent split and including stereochemistry, which makes our method universally applicable.Copyright © 2019 American Chemical Society.

Reaction prediction remains one of the major challenges for organic chemistry and is a prerequisite for efficient synthetic planning. It is desirable to develop algorithms that, like humans, "learn" from being exposed to examples of the application of the rules of organic chemistry. We explore the use of neural networks for predicting reaction types, using a new reaction fingerprinting method. We combine this predictor with SMARTS transformations to build a system which, given a set of reagents and reactants, predicts the likely products. We test this method on problems from a popular organic chemistry textbook.

We describe a fully data driven model that learns to perform a retrosynthetic reaction prediction task, which is treated as a sequence-to-sequence mapping problem. The end-to-end trained model has an encoder-decoder architecture that consists of two recurrent neural networks, which has previously shown great success in solving other sequence-to-sequence prediction tasks such as machine translation. The model is trained on 50,000 experimental reaction examples from the United States patent literature, which span 10 broad reaction types that are commonly used by medicinal chemists. We find that our model performs comparably with a rule-based expert system baseline model, and also overcomes certain limitations associated with rule-based expert systems and with any machine learning approach that contains a rule-based expert system component. Our model provides an important first step toward solving the challenging problem of computational retrosynthetic analysis.

In a computing context, cybersecurity is undergoing massive shifts in technology and its operations in recent days, and data science is driving the change. Extractingsecurity incident patternsor insights from cybersecurity data and building correspondingdata-driven model, is the key to make a security system automated and intelligent. To understand and analyze the actual phenomena with data, various scientific methods, machine learning techniques, processes, and systems are used, which is commonly known as data science. In this paper, we focus and briefly discuss oncybersecurity data science, where the data is being gathered from relevant cybersecurity sources, and the analytics complement thelatest data-driven patternsfor providing more effective security solutions. The concept of cybersecurity data science allows making the computing process more actionable and intelligent as compared to traditional ones in the domain of cybersecurity. We then discuss and summarize a number of associatedresearch issues and future directions. Furthermore, we provide amachine learningbasedmulti-layered frameworkfor the purpose of cybersecurity modeling. Overall, our goal is not only to discuss cybersecurity data science and relevant methods but also to focus the applicability towards data-driven intelligent decision making for protecting the systems from cyber-attacks.

We present new developments in the AMBIT open source software package for efficient searching of chemical structures and structural fragments. AMBIT-SMARTS is a Java based software built on top of The Chemistry Development Kit. The AMBIT-SMARTS parser implements the entire SMARTS language specification with several syntax extensions that enable support for custom modifications introduced by third party software packages such as OpenEye, MOE and OpenBabel. The goal of yet another open-source SMARTS parser implementation is to achieve better performance and compatibility with multiple existing flavours of the SMARTS language, as well as to provide utilities for running efficient SMARTS queries in large structural databases. We describe a combination of approaches towards lowering the computational cost and improving the response time of substructure queries. An exhaustive comparison of the AMBIT algorithm with several subgraph isomorphism implementations is performed. To demonstrate the performance of the entire system from an end-user point of view, response time statistics for Web service substructure search queries against a database of 4.5 M structures are also reported. The package has wide applicability in the implementation of various chemoinformatics tasks. It has already been used in several projects dealing with descriptor calculation and predictive algorithms, database queries, web applications and web services. Copyright © 2011 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim.

We investigate practical selection of hyper-parameters for support vector machines (SVM) regression (that is, epsilon-insensitive zone and regularization parameter C). The proposed methodology advocates analytic parameter selection directly from the training data, rather than re-sampling approaches commonly used in SVM applications. In particular, we describe a new analytical prescription for setting the value of insensitive zone epsilon, as a function of training sample size. Good generalization performance of the proposed parameter selection is demonstrated empirically using several low- and high-dimensional regression problems. Further, we point out the importance of Vapnik's epsilon-insensitive loss for regression problems with finite samples. To this end, we compare generalization performance of SVM regression (using proposed selection of epsilon-values) with regression using 'least-modulus' loss (epsilon=0) and standard squared loss. These comparisons indicate superior generalization performance of SVM regression under sparse sample settings, for various types of additive noise.

近年来，由于计算能力、大数据和算法的不断进步，人工智能（Artificial intelligence，AI）重新兴起，已成为诸多研究领域变革性发展背后的重要推动力.机器学习（Machine learning，ML）是人工智能一个重要的研究领域.随着化学信息学的发展，机器学习在化学领域展现出巨大的发展潜力，也为有机化学的发展带来了新的机遇.为帮助有机化学家了解这一新兴领域，对如何将机器学习策略应用于有机化学研究做简单介绍，同时，概括总结了机器学习在化合物性质预测、分子从头设计、化学反应预测、逆合成分析和智能合成机器方面的应用实例，分析讨论了当前机器学习在有机化学领域面临的挑战和难题.

Driven by nowadays’ computing power, big data technology as well as learning algorithm, artificial intelligence (AI) has gained trenmendous attentions and become a transformative approach in many research areas. One of the most extensively explored AI approaches in chemistry is (deep) machine learning, which provides new twists in the fields of organic chemistry. The workflow of machine learning (ML) study in organic chemistry is briefly introduced. Meanwhile, the application of ML in the accurate prediction of chemical properties, molecular <i>de novo</i> design, chemical reaction prediction, retrosynthetic analysis and artificial intelligence synthetic machine are also summarized. In the end, the current challenges in this field are analyzed and discussed.

Design of asymmetric catalysts generally involves time- and resource-intensive heuristic endeavors. In view of the steady increase in interest toward efficient catalytic asymmetric reactions and the rapid growth in the field of machine learning (ML) in recent years, we envisaged dovetailing these two important domains. We selected a set of quantum chemically derived molecular descriptors from five different asymmetric binaphthyl-derived catalyst families with the propensity to impact the enantioselectivity of asymmetric hydrogenation of alkenes and imines. The predictive power of the random forest (RF) built using the molecular parameters of a set of 368 substrate-catalyst combinations is found to be impressive, with a root-mean-square error (rmse) in the predicted enantiomeric excess (%) of about 8.4 ± 1.8 compared to the experimentally known values. The accuracy of RF is found to be superior to other ML methods such as convolutional neural network, decision tree, and eXtreme gradient boosting as well as stepwise linear regression. The proposed method is expected to provide a leap forward in the design of catalysts for asymmetric transformations.

Fluorescent molecules, fluorophores or dyes, play essential roles in bioimaging. Effective bioimaging requires fluorophores with diverse colors and high quantum yields for better resolution. An essential computational component to design novel dye molecules is an accurate model that predicts the electronic properties of molecules. Here, we present statistical machines that predict the excitation energies and associated oscillator strengths of a given molecule using the random forest algorithm. The excitation energies and oscillator strengths of a molecule are closely related to the emission spectrum and the quantum yields of fluorophores, respectively. In this study, we identified specific molecular substructures that induce high oscillator strengths of molecules. The results of our study are expected to serve as new design principles for designing novel fluorophores.

\n Chemists often discover reactions by applying catalysts to a series of simple compounds. Tweaking those reactions to tolerate more structural complexity in pharmaceutical research is time-consuming. Ahneman\n et al.\n report that machine learning can help. Using a high-throughput data set, they trained a random forest algorithm to predict which specific palladium catalysts would best tolerate isoxazoles (cyclic structures with an N–O bond) during C–N bond formation. The predictions also helped to guide analysis of the catalyst inhibition mechanism.\n

Drug toxicity is frequently caused by electrophilic reactive metabolites that covalently bind to proteins. Epoxides comprise a large class of three-membered cyclic ethers. These molecules are electrophilic and typically highly reactive due to ring tension and polarized carbon-oxygen bonds. Epoxides are metabolites often formed by cytochromes P450 acting on aromatic or double bonds. The specific location on a molecule that undergoes epoxidation is its site of epoxidation (SOE). Identifying a molecule's SOE can aid in interpreting adverse events related to reactive metabolites and direct modification to prevent epoxidation for safer drugs. This study utilized a database of 702 epoxidation reactions to build a model that accurately predicted sites of epoxidation. The foundation for this model was an algorithm originally designed to model sites of cytochromes P450 metabolism (called XenoSite) that was recently applied to model the intrinsic reactivity of diverse molecules with glutathione. This modeling algorithm systematically and quantitatively summarizes the knowledge from hundreds of epoxidation reactions with a deep convolution network. This network makes predictions at both an atom and molecule level. The final epoxidation model constructed with this approach identified SOEs with 94.9% area under the curve (AUC) performance and separated epoxidized and non-epoxidized molecules with 79.3% AUC. Moreover, within epoxidized molecules, the model separated aromatic or double bond SOEs from all other aromatic or double bonds with AUCs of 92.5% and 95.1%, respectively. Finally, the model separated SOEs from sites of sp(2) hydroxylation with 83.2% AUC. Our model is the first of its kind and may be useful for the development of safer drugs. The epoxidation model is available at http://swami.wustl.edu/xenosite.

Drug toxicity is often caused by electrophilic reactive metabolites that covalently bind to proteins. Consequently, the quantitative strength of a molecule's reactivity with glutathione (GSH) is a frequently used indicator of its toxicity. Through cysteine, GSH (and proteins) scavenges reactive molecules to form conjugates in the body. GSH conjugates to specific atoms in reactive molecules: their sites of reactivity. The value of knowing a molecule's sites of reactivity is unexplored in the literature. This study tests the value of site of reactivity data that identifies the atoms within 1213 reactive molecules that conjugate to GSH and builds models to predict molecular reactivity with glutathione. An algorithm originally written to model sites of cytochrome P450 metabolism (called XenoSite) finds clear patterns in molecular structure that identify sites of reactivity within reactive molecules with 90.8% accuracy and separate reactive and unreactive molecules with 80.6% accuracy. Furthermore, the model output strongly correlates with quantitative GSH reactivity data in chemically diverse, external data sets. Site of reactivity data is nearly unstudied in the literature prior to our efforts, yet it contains a strong signal for reactivity that can be utilized to more accurately predict molecule reactivity and, eventually, toxicity.

We present a supervised learning approach to predict the products of organic reactions given their reactants, reagents, and solvent(s). The prediction task is factored into two stages comparable to manual expert approaches: considering possible sites of reactivity and evaluating their relative likelihoods. By training on hundreds of thousands of reaction precedents covering a broad range of reaction types from the patent literature, the neural model makes informed predictions of chemical reactivity. The model predicts the major product correctly over 85% of the time requiring around 100 ms per example, a significantly higher accuracy than achieved by previous machine learning approaches, and performs on par with expert chemists with years of formal training. We gain additional insight into predictions the design of the neural model, revealing an understanding of chemistry qualitatively consistent with manual approaches.

Machine learning enables computers to address problems by learning from data. Deep learning is a type of machine learning that uses a hierarchical recombination of features to extract pertinent information and then learn the patterns represented in the data. Over the last eight years, its abilities have increasingly been applied to a wide variety of chemical challenges, from improving computational chemistry to drug and materials design and even synthesis planning. This review aims to explain the concepts of deep learning to chemists from any background and follows this with an overview of the diverse applications demonstrated in the literature. We hope that this will empower the broader chemical community to engage with this burgeoning field and foster the growing movement of deep learning accelerated chemistry.

Prediction of bond dissociation energies for charged molecules with a graph neural network enabled by global molecular features and reaction difference features between products and reactants.

\n Asymmetric catalysis is widely used in chemical research and manufacturing to access just one of two possible mirror-image products. Nonetheless, the process of tuning catalyst structure to optimize selectivity is still largely empirical. Zahrt\n et al.\n present a framework for more efficient, predictive optimization. As a proof of principle, they focused on a known coupling reaction of imines and thiols catalyzed by chiral phosphoric acid compounds. By modeling multiple conformations of more than 800 prospective catalysts, and then training machine-learning algorithms on a subset of experimental results, they achieved highly accurate predictions of enantioselectivities.\n

Drug molecules consist of a few tens of atoms connected by covalent bonds. How many such molecules are possible in total and what is their structure? This question is of pressing interest in medicinal chemistry to help solve the problems of drug potency, selectivity, and toxicity and reduce attrition rates by pointing to new molecular series. To better define the unknown chemical space, we have enumerated 166.4 billion molecules of up to 17 atoms of C, N, O, S, and halogens forming the chemical universe database GDB-17, covering a size range containing many drugs and typical for lead compounds. GDB-17 contains millions of isomers of known drugs, including analogs with high shape similarity to the parent drug. Compared to known molecules in PubChem, GDB-17 molecules are much richer in nonaromatic heterocycles, quaternary centers, and stereoisomers, densely populate the third dimension in shape space, and represent many more scaffold types.

SchNetPack is a toolbox for the development and application of deep neural networks that predict potential energy surfaces and other quantum-chemical properties of molecules and materials. It contains basic building blocks of atomistic neural networks, manages their training, and provides simple access to common benchmark datasets. This allows for an easy implementation and evaluation of new models. For now, SchNetPack includes implementations of (weighted) atom-centered symmetry functions and the deep tensor neural network SchNet, as well as ready-to-use scripts that allow one to train these models on molecule and material datasets. Based on the PyTorch deep learning framework, SchNetPack allows one to efficiently apply the neural networks to large datasets with millions of reference calculations, as well as parallelize the model across multiple GPUs. Finally, SchNetPack provides an interface to the Atomic Simulation Environment in order to make trained models easily accessible to researchers that are not yet familiar with neural networks.

Machine learning (ML) promises to tackle the grand challenges in chemistry and speed up the generation, improvement and/or ordering of research hypotheses. Despite the overarching applicability of ML workflows, one usually finds diverse evaluation study designs. The current heterogeneity in evaluation techniques and metrics leads to difficulty in (or the impossibility of) comparing and assessing the relevance of new algorithms. Ultimately, this may delay the digitalization of chemistry at scale and confuse method developers, experimentalists, reviewers and journal editors. In this Perspective, we critically discuss a set of method development and evaluation guidelines for different types of ML-based publications, emphasizing supervised learning. We provide a diverse collection of examples from various authors and disciplines in chemistry. While taking into account varying accessibility across research groups, our recommendations focus on reporting completeness and standardizing comparisons between tools. We aim to further contribute to improved ML transparency and credibility by suggesting a checklist of retro-/prospective tests and dissecting their importance. We envisage that the wide adoption and continuous update of best practices will encourage an informed use of ML on real-world problems related to the chemical sciences.© 2022. Springer Nature Limited.

The ice arches that usually develop at the northern and southern ends of Nares Strait play an important role in modulating the export of Arctic Ocean multi-year sea ice. The Arctic Ocean is evolving towards an ice pack that is younger, thinner, and more mobile and the fate of its multi-year ice is becoming of increasing interest. Here, we use sea ice motion retrievals from Sentinel-1 imagery to report on the recent behavior of these ice arches and the associated ice fluxes. We show that the duration of arch formation has decreased over the past 20 years, while the ice area and volume fluxes along Nares Strait have both increased. These results suggest that a transition is underway towards a state where the formation of these arches will become atypical with a concomitant increase in the export of multi-year ice accelerating the transition towards a younger and thinner Arctic ice pack.

An enantioselective, intermolecular dehydrogenative Heck arylation of trisubstituted alkenes to construct remote quaternary stereocenters has been developed. Using a new chiral pyridine oxazoline ligand, good to high enantioselectivity is achieved for various combinations of indole derivatives and trisubstituted alkenes. However, some combinations of substrates led to lower enantioselectivity, which provided the impetus to use structure enantioselectivity correlations to design a better performing ligand.

To improve product yields in synthetic reactions, it is important to use appropriate catalysts. In this study, we used machine learning to design catalysts for a reaction system in which both Buchwald-Hartwig-type and Suzuki-Miyaura-type cross-coupling reactions proceed simultaneously. First, using an existing dataset, yield prediction models were constructed with machine learning between experimental conditions, including the substrate and catalyst and the yields of the two products. Seven methods for calculating both the substrate and catalyst descriptors were proposed, and the predictive ability of the yield prediction models was discussed in terms of the descriptors and machine learning methods. Then, the constructed models were used to predict the compound yields for new combinations of substrates and catalysts, and the predictions were experimentally validated with high reproducibility, confirming that machine learning can predict yields from experimental conditions with high accuracy. In addition, to design catalysts that will improve the yields in our dataset, we added datasets collected from scientific papers and designed catalyst ligands. The proposed catalyst candidates were tested in actual synthetic experiments, and the experimental results exceeded the existing yields.© 2021 The Authors. Published by American Chemical Society.

Computer assistance in synthesis design has existed for over 40 years, yet retrosynthesis planning software has struggled to achieve widespread adoption. One critical challenge in developing high-quality pathway suggestions is that proposed reaction steps often fail when attempted in the laboratory, despite initially seeming viable. The true measure of success for any synthesis program is whether the predicted outcome matches what is observed experimentally. We report a model framework for anticipating reaction outcomes that combines the traditional use of reaction templates with the flexibility in pattern recognition afforded by neural networks. Using 15 000 experimental reaction records from granted United States patents, a model is trained to select the major (recorded) product by ranking a self-generated list of candidates where one candidate is known to be the major product. Candidate reactions are represented using a unique edit-based representation that emphasizes the fundamental transformation from reactants to products, rather than the constituent molecules' overall structures. In a 5-fold cross-validation, the trained model assigns the major product rank 1 in 71.8% of cases, rank ≤3 in 86.7% of cases, and rank ≤5 in 90.8% of cases.

Nucleophilicity provides important information about the chemical reactivity of organic molecules. Experimental determination of the nucleophilicity parameter is a tedious and resource-intensive approach. Herein, we present a novel machine learning protocol that uses key structural descriptors to predict the nucleophilicities of organic molecules, which agree well with the experimental values. A data driven approach was used where quantum mechanical molecular and thermodynamic descriptors from a wide range of structurally diverse nucleophiles and relevant solvents were extracted and modelled using advanced algorithms against the experimentally available nucleophilicity values. Despite the structural diversity of nucleophiles, we are able to achieve statistically robust models with a high predictive power using tree-based and neural network algorithms trained on an in-house developed unique dataset consisting of 752 nucleophilicity values and 27 molecular descriptors.

Random Forest regression (RF), Partial-Least-Squares (PLS) regression, Support Vector Machines (SVM), and Artificial Neural Networks (ANN) were used to develop QSPR models for the prediction of aqueous solubility, based on experimental data for 988 organic molecules. The Random Forest regression model predicted aqueous solubility more accurately than those created by PLS, SVM, and ANN and offered methods for automatic descriptor selection, an assessment of descriptor importance, and an in-parallel measure of predictive ability, all of which serve to recommend its use. The prediction of log molar solubility for an external test set of 330 molecules that are solid at 25 degrees C gave an r2 = 0.89 and RMSE = 0.69 log S units. For a standard data set selected from the literature, the model performed well with respect to other documented methods. Finally, the diversity of the training and test sets are compared to the chemical space occupied by molecules in the MDL drug data report, on the basis of molecular descriptors selected by the regression analysis.

Our previous study demonstrated increased expression of Heat shock protein (Hsp) 90 in the skin of patients with systemic sclerosis (SSc). We aimed to evaluate plasma Hsp90 in SSc and characterize its association with SSc-related features. Ninety-two SSc patients and 92 age-/sex-matched healthy controls were recruited for the cross-sectional analysis. The longitudinal analysis comprised 30 patients with SSc associated interstitial lung disease (ILD) routinely treated with cyclophosphamide. Hsp90 was increased in SSc compared to healthy controls. Hsp90 correlated positively with C-reactive protein and negatively with pulmonary function tests: forced vital capacity and diffusing capacity for carbon monoxide (DLCO). In patients with diffuse cutaneous (dc) SSc, Hsp90 positively correlated with the modified Rodnan skin score. In SSc-ILD patients treated with cyclophosphamide, no differences in Hsp90 were found between baseline and after 1, 6, or 12 months of therapy. However, baseline Hsp90 predicts the 12-month change in DLCO. This study shows that Hsp90 plasma levels are increased in SSc patients compared to age-/sex-matched healthy controls. Elevated Hsp90 in SSc is associated with increased inflammatory activity, worse lung functions, and in dcSSc, with the extent of skin involvement. Baseline plasma Hsp90 predicts the 12-month change in DLCO in SSc-ILD patients treated with cyclophosphamide.

Gaining predictable control over various forms of selectivities, such as enantio- and/or regio-selectivities, has been a long-standing goal in chemical catalysis. Although a number of factors such as the molecular features of the reactants and catalysts, as well as the reaction conditions, can influence the outcome of a reaction, it is not quite conspicuous as to what combinations of these parameters would offer a desired form of selectivity. We use machine learning tools, such as the neural network (NN), decision tree (DT), logistic regression (LR) and Random forest algorithms, to (a) analyze the outcome of an important catalytic regio-selective difluorination reaction of alkenes, and (b) decipher the complex interplay of various molecular parameters and their non-linear dependencies. The connection between what features of alkenes will yield 1,1-difluorination and how subtle changes would steer the reaction to 1,2-difluorination under identical conditions is enunciated. The NN was able to accurately predict whether a given alkene would yield a 1,1- or 1,2-difluorinated product. A combination of DT and the random forest classifier offered important chemical insights, which could be used in making a more rational choice of the reactant alkene for the desired regioisomeric product. The results could have far reaching implications in predicting which regioisomer is likely to be formed under a given set of conditions, and thus this technique is capable of expediting the development of catalytic transformations.

Deep neural networks have been increasingly used in various chemical fields. In the nature of a data-driven approach, their performance strongly depends on data used in training. Therefore, models developed in data-deficient situations can cause highly uncertain predictions, leading to vulnerable decision making. Here, we show that Bayesian inference enables more reliable prediction with quantitative uncertainty analysis. Decomposition of the predictive uncertainty into model- and data-driven uncertainties allows us to elucidate the source of errors for further improvements. For molecular applications, we devised a Bayesian graph convolutional network (GCN) and evaluated its performance for molecular property predictions. Our study on the classification problem of bio-activity and toxicity shows that the confidence of prediction can be quantified in terms of the predictive uncertainty, leading to more accurate virtual screening of drug candidates than standard GCNs. The result of log  prediction illustrates that data noise affects the data-driven uncertainty more significantly than the model-driven one. Based on this finding, we could identify artefacts that arose from quantum mechanical calculations in the Harvard Clean Energy Project dataset. Consequently, the Bayesian GCN is critical for molecular applications under data-deficient conditions.This journal is © The Royal Society of Chemistry 2019.

Stimulated cells and cancer cells have widespread shortening of mRNA 3’-untranslated regions (3’UTRs) and switches to shorter mRNA isoforms due to usage of more proximal polyadenylation signals (PASs) in introns and last exons. U1 snRNP (U1), vertebrates’ most abundant non-coding (spliceosomal) small nuclear RNA, silences proximal PASs and its inhibition with antisense morpholino oligonucleotides (U1 AMO) triggers widespread premature transcription termination and mRNA shortening. Here we show that low U1 AMO doses increase cancer cells’ migration and invasion in vitro by up to 500%, whereas U1 over-expression has the opposite effect. In addition to 3’UTR length, numerous transcriptome changes that could contribute to this phenotype are observed, including alternative splicing, and mRNA expression levels of proto-oncogenes and tumor suppressors. These findings reveal an unexpected role for U1 homeostasis (available U1 relative to transcription) in oncogenic and activated cell states, and suggest U1 as a potential target for their modulation.

Deep reinforcement learning was employed to optimize chemical reactions. Our model iteratively records the results of a chemical reaction and chooses new experimental conditions to improve the reaction outcome. This model outperformed a state-of-the-art blackbox optimization algorithm by using 71% fewer steps on both simulations and real reactions. Furthermore, we introduced an efficient exploration strategy by drawing the reaction conditions from certain probability distributions, which resulted in an improvement on regret from 0.062 to 0.039 compared with a deterministic policy. Combining the efficient exploration policy with accelerated microdroplet reactions, optimal reaction conditions were determined in 30 min for the four reactions considered, and a better understanding of the factors that control microdroplet reactions was reached. Moreover, our model showed a better performance after training on reactions with similar or even dissimilar underlying mechanisms, which demonstrates its learning ability.

We introduce an artificial intelligence approach to de novo design of molecules with desired physical or biological properties.

Machine learning provides effective computational tools for exploring the chemical space via deep generative models. Here, we propose a new reinforcement learning scheme to fine-tune graph-based deep generative models for molecular design tasks. We show how our computational framework can successfully guide a pretrained generative model toward the generation of molecules with a specific property profile, even when such molecules are not present in the training set and unlikely to be generated by the pretrained model. We explored the following tasks: generating molecules of decreasing/increasing size, increasing drug-likeness, and increasing bioactivity. Using the proposed approach, we achieve a model which generates diverse compounds with predicted DRD2 activity for 95% of sampled molecules, outperforming previously reported methods on this metric.