Universal Chemical Synthesis and Discovery with ‘The Chemputer’

Recent advances in chemical programming enable adoption and universal automation of chemical discovery and synthesis, combined with artificial intelligence, to efficiently perform laboratory tasks, including the closed-loop data exploration for new reactivity.Robots can perform chemical reactions and analysis much faster than can be done manually, utilizing trial and error, as well as feedback to make autonomous decisions.Chemists can actively seek out to explore chemical space, aiming for discovery of novel or new molecules and reactions using closed-loop robotic chemical search engines. There is a growing drive in the chemistry community to exploit rapidly growing robotic technologies along with artificial intelligence-based approaches. Applying this to chemistry requires a holistic approach to chemical synthesis design and execution. Here, we outline a universal approach to this problem beginning with an abstract representation of the practice of chemical synthesis that then informs the programming and automation required for its practical realization. Using this foundation to construct closed-loop robotic chemical search engines, we can generate new discoveries that may be verified, optimized, and repeated entirely automatically. These robots can perform chemical reactions and analyses much faster than can be done manually. As such, this leads to a road map whereby molecules can be discovered, optimized, and made on demand from a digital code. There is a growing drive in the chemistry community to exploit rapidly growing robotic technologies along with artificial intelligence-based approaches. Applying this to chemistry requires a holistic approach to chemical synthesis design and execution. Here, we outline a universal approach to this problem beginning with an abstract representation of the practice of chemical synthesis that then informs the programming and automation required for its practical realization. Using this foundation to construct closed-loop robotic chemical search engines, we can generate new discoveries that may be verified, optimized, and repeated entirely automatically. These robots can perform chemical reactions and analyses much faster than can be done manually. As such, this leads to a road map whereby molecules can be discovered, optimized, and made on demand from a digital code. Methodologies for the automation of chemical synthesis, optimization, and discovery have not generally been designed for the realities of laboratory-based research, tending instead to focus on engineering solutions to practical problems. We argue that the potential of rapidly developing technologies (e.g., machine learning and robotics) are more fully realized by operating seamlessly with the way that synthetic chemists currently work (Figure 1) [1Steiner S. et al.Organic synthesis in a modular robotic system driven by a chemical programming language.Science. 2019; 363eaav2211Crossref PubMed Scopus (194) Google Scholar]. This is because the organic chemist often works by thinking backwards as much as they do forwards when planning a synthetic procedure. To reproduce this fundamental mode of operation, a new universal approach to the automated exploration of chemical space is needed that combines an abstraction of chemical synthesis with robotic hardware and closed-loop programming [2Sans V. et al.A self optimizing synthetic organic reactor system using real-time in-line NMR spectroscopy.Chem. Sci. 2015; 6: 1258-1264Crossref PubMed Google Scholar, 3Kitson P.J. et al.Digitization of multistep organic synthesis in reactionware for on-demand pharmaceuticals.Science. 2018; 359: 314-319Crossref PubMed Scopus (112) Google Scholar]. However, this leads chemists to constantly test the reactions with different synthetic parameters and conditions. The alternative to this problem, as shown in this opinion article, is the development of an approach to universal chemistry using a programming language with automation in combination with machine learning and artificial intelligence (AI). Chemists already benefit from algorithms in the field of chemometrics and, therefore, automation is one step forward that might help chemists to navigate and search chemical space more quickly, efficiently, and importantly, without bias. Chemometrics is a field that employs a broad range of algorithms to solve chemistry-related problems and has been well established over the past 50 years [4Brereton R.G. A short history of chemometrics: a personal view.J. Chemom. 2014; 28: 749-760Crossref Scopus (36) Google Scholar]. Figure 2 presents a standard chemometrics workflow for processing data. The process begins with data that may be of various formats that depend upon the experiment type and/or posed question. The next step is data preprocessing, which covers a variety of procedures depending on the type of data analyzed (e.g., peak detection, input of missing data, and/or normalization). This process is followed by statistical modeling, which is divided into supervised and unsupervised approaches. Probably one of the most well-known unsupervised approaches is principal component analysis (see Glossary), which allows summarizing large data sets into several components that capture most of the information. Support vector machine, along with partial least-square discriminant analysis, are probably the most well-known supervised approaches that allow samples to be classified into distinctive groups based on relevant information. If the produced outcomes are relevant, the next steps incorporate validation to ensure high quality conclusions are formed. Finally, each of the analyses ends with data interpretation. Further details on chemometrics and algorithms that enable exploration of chemical space are found elsewhere [4Brereton R.G. A short history of chemometrics: a personal view.J. Chemom. 2014; 28: 749-760Crossref Scopus (36) Google Scholar, 5Brereton R.G. et al.Chemometrics in analytical chemistry–part I: history, experimental design and data analysis tools.Anal. Bioanal. Chem. 2017; 409: 5891-5899Crossref PubMed Scopus (70) Google Scholar, 6Brereton R.G. et al.Chemometrics in analytical chemistry–part II: modeling, validation, and applications.Anal. Bioanal. Chem. 2018; 410: 6691-6704Crossref PubMed Scopus (72) Google Scholar, 7Henson A.B. et al.Designing algorithms to aid discovery by chemical robots.ACS Cent. Sci. 2018; 4: 793-804Crossref PubMed Scopus (49) Google Scholar]. In the following paragraphs, we show how chemometrics can be synergistically combined with automation. As we will demonstrate through several examples, the process of automation allows for: (i) increased productivity through design of complex experiments that are entirely automated; (ii) increased reliability by reducing human error and increased confidence in the outcomes (i.e., experiments are directly linked with software/algorithms that can produce graphical representation of the results directly); (iii) improved safety when experiments must be performed in a closed environment (e.g., fume hood, glove box, or sealed reactor) equipped with software to indicate when there is a risk or incident; (iv) increased efficiency of processes performed due to increased productivity, reduced waste, and improved outcome quality; and finally (v) valuable walk-away time where the chemist may focus on research [7Henson A.B. et al.Designing algorithms to aid discovery by chemical robots.ACS Cent. Sci. 2018; 4: 793-804Crossref PubMed Scopus (49) Google Scholar, 8Gromski P.S. et al.How to explore chemical space using algorithms and automation.Nat. Rev. Chem. 2019; 3: 119-128Crossref Scopus (117) Google Scholar]. The ability to make small molecules autonomously and automatically will be fundamental to many applications, including searching for new drugs and materials. So far, automation of small molecule synthesis has relied on a single reaction class limiting its overall universality (e.g., in iterative N-methyliminodiacetic (MIDA) boronate synthesis [9Li J. et al.Synthesis of many different types of organic small molecules using one automated process.Science. 2015; 347: 1221-1226Crossref PubMed Scopus (325) Google Scholar] or enzyme-assisted carbohydrate synthesis [10Li T. et al.An automated platform for the enzyme-mediated assembly of complex oligosaccharides.Nat. Chem. 2019; 11: 229-236Crossref PubMed Scopus (87) Google Scholar]). Additionally, automated synthesis requires (in many cases) optimization of reaction yields; following optimization, the best conditions can be fed to the synthesis robot to increase the overall yield. There are many approaches to automated yield optimization, some of which are described below. As optimization of reaction conditions requires live feedback from the robotic system, many different detectors have been introduced to monitor progress of the reactions, including benchtop nuclear magnetic resonance spectroscopy [2Sans V. et al.A self optimizing synthetic organic reactor system using real-time in-line NMR spectroscopy.Chem. Sci. 2015; 6: 1258-1264Crossref PubMed Google Scholar], infrared spectroscopy [11Dragone V. et al.An autonomous organic reaction search engine for chemical reactivity.Nat. Commun. 2017; 815733Crossref PubMed Scopus (52) Google Scholar], mass spectrometry [12Yunker L.P.E. et al.Practical approaches to the ESI-MS analysis of catalytic reactions.J. Mass Spectrom. 2014; 49: 1-8Crossref PubMed Scopus (89) Google Scholar], Raman spectroscopy [13Svensson O. et al.Reaction monitoring using Raman spectroscopy and chemometrics.Chemom. Intell. Lab. Syst. 1999; 49: 49-66Crossref Scopus (73) Google Scholar], UV-Vis spectroscopy [14Yue J. et al.Microreactors with integrated UV/Vis spectroscopic detection for online process analysis under segmented flow.Lab Chip. 2013; 13: 4855-4863Crossref PubMed Scopus (55) Google Scholar], and high-performance liquid chromatography [15Malig T.C. et al.Real-time HPLC-MS reaction progress monitoring using an automated analytical platform.React. Chem. Eng. 2017; 2: 309-314Crossref Google Scholar]. Harvested data are then fed to optimization algorithms to explore the often multidimensional parameter space. For example, Bédard and colleagues showed an automated-flow system for the optimization of many different types of chemical reactions, including Buchwald-Hartwig amination, Suzuki-Miyaura cross couplings, nucleophilic aromatic substitution (SNAr), Horner-Wadsworth-Emmons olefination, and photoredox catalysis. The platform could be easily reconfigured to the desired task in a plug-and-play fashion, by attaching different modules (e.g., a photo light-emitting diode or cooled reactor) to the platform core [16Bédard A.-C. et al.Reconfigurable system for automated optimization of diverse chemical reactions.Science. 2018; 361: 1220-1225Crossref PubMed Scopus (237) Google Scholar]. Robotic approaches also promise to speed up chemical space exploration. To this point, high-throughput experimentation (HTE) appears particularly promising because it can perform thousands of nanomole-scale reactions per day. These HTE approaches could deliver the vast amount of information necessary to train machine learning and AI models, yielding chemical ‘big data’. Perera and colleagues demonstrated a flow platform for nanomole screening of Suzuki-Miyaura reactions allowing for screening of greater than 1500 reactions per 24 h [17Perera D. et al.A platform for automated nanomole-scale reaction screening and micromole-scale synthesis in flow.Science. 2018; 359: 429-434Crossref PubMed Scopus (184) Google Scholar]. Despite the high-throughput capability, the search of chemical space is not guided by a specific objective. Therefore, many different machine learning algorithms have been developed to explore chemical space. Machine learning approaches are fundamental to scientific investigation in many disciplines. In chemistry, many of these methods are well-covered within chemometrics. These methods, linked with chemistry and automation, are rapidly changing the face of chemical research and discovery. Here, we explore how chemometrics and robotics/automation are helping to progress discovery through exploring chemical space and beyond. Scientists have begun to embrace the power of machine learning coupled with statistically driven design in their research to predict the performance of synthetic reactions. For example, the yield of a Pd-catalyzed Buchwald-Hartwig reaction was predicted using random forest in the multidimensional chemical space obtained via HTE [18Ahneman D.T. et al.Predicting reaction performance in C-N cross-coupling using machine learning.Science. 2018; 360: 186-190Crossref PubMed Scopus (397) Google Scholar]. Furthermore, Nielsen and colleagues applied random forest to map the yield landscape of intricate deoxyfluorination with sulfonyl fluoride allowing improved prediction of high-yielding conditions for untested substrates [19Nielsen M.K. et al.Deoxyfluorination with sulfonyl fluorides: navigating reaction space with machine learning.J. Am. Chem. Soc. 2018; 140: 5004-5008Crossref PubMed Scopus (124) Google Scholar]. More recently, Phoenics was developed, which combines a concept from Bayesian optimization with ideas from Bayesian kernel density estimation to solve optimization problems and afford efficient exploitation of the search space [20Hase F. et al.Phoenics: a Bayesian optimizer for chemistry.ACS Cent. Sci. 2018; 4: 1134-1145Crossref PubMed Scopus (147) Google Scholar]. Meanwhile, our emphasis is on automation of discovery, which is controlled by robots/computers rather than by humans. Discovery through automation offers far better efficiency and accuracy, as recently shown by Duros and colleagues, where the authors compared human- and robot-based discovery of gigantic polyoxometalates. Specifically, it was shown that algorithm-based search covered approximately nine times more crystallization space than a random search and approximately six times more than human-based discovery. Perhaps even more importantly is that the rate of successful crystallization also increased by ∼5% [21Duros V. et al.Human versus robots in the discovery and crystallization of gigantic polyoxometalates.Angew. Chem. 2017; 56: 10815-10820Crossref Scopus (83) Google Scholar]. In addition, the algorithm explored a wider range of space that would need to be performed either by human or purely random search. Recently, the same researchers observed that collaboration between smart robotics and humans may be even more efficient than either alone [22Duros V. et al.Intuition-enabled machine learning beats the competition when joint human-robot teams perform inorganic chemical experiments.J. Chem. Inf. Model. 2019; 59: 2664-2671Crossref PubMed Scopus (19) Google Scholar]. Grizou and colleagues described a chemical robotic discovery assistant equipped with a curiosity algorithm that can efficiently explore a complex chemical system in search of complex emergent phenomena exhibited by proto-cell droplets [23Grizou J. et al.A closed loop discovery robot driven by a curiosity algorithm discovers proto-cells that show complex and emergent behaviours.ChemRxiv. 2018; (Published online 13 February, 2019. https://doi.org/10.26434/chemrxiv.6958334.v1)Google Scholar]. This brings the development of automation, optimization, and discovery very close, a topic widely described in the work by Aspuru-Guzik and Henson, where the authors highlight the fact that self-driven laboratories/robots lead the way forward to fast-track discovery by boosting automated experimentation platforms with machine learning to explore chemical space [7Henson A.B. et al.Designing algorithms to aid discovery by chemical robots.ACS Cent. Sci. 2018; 4: 793-804Crossref PubMed Scopus (49) Google Scholar, 24Häse F. et al.Next-generation experimentation with self-driving laboratories.Trends Chem. 2019; 1: 282-291Abstract Full Text Full Text PDF Scopus (122) Google Scholar]. The automated synthesis could make also use of retrosynthetic analysis for planning the synthesis routes to the target molecules. There are many approaches to automated retrosynthesis, and the most recent one by Segler and colleagues seems to be particularly promising [25Segler M.H.S. et al.Planning chemical syntheses with deep neural networks and symbolic AI.Nature. 2018; 555: 604-610Crossref PubMed Scopus (813) Google Scholar]. It used Monte Carlo tree search and symbolic AI to discover retrosynthetic routes. The neural networks were trained on all reactions published in organic chemistry. The system allowed cracking for nearly twice as many molecules, 30 times faster than the traditional computer-aided search method, which is based on extracted rules and hand-designed heuristics. In general, this approach allowed for faster and more efficient retrosynthetic analysis than any other well-known method. Figure 3 shows a workflow for joining automated retrosynthesis with a synthesis robot and reaction optimization. The retrosynthetic module will generate a valid synthesis of the target that can then be transferred into synthesis code that can be executed in a robotic platform. The optimization module can optimize the whole sequence, getting the feedback from the robot. We recently showed a modular platform for automating batch organic synthesis, which embodies our abstraction in ‘the Chemputer’ (Figure 4) [1Steiner S. et al.Organic synthesis in a modular robotic system driven by a chemical programming language.Science. 2019; 363eaav2211Crossref PubMed Scopus (194) Google Scholar]. Our abstraction of organic synthesis (Figure 4A) contains the key four stages of synthetic protocols: reaction, workup, isolation, and purification, that can be linked to the physical operations of an automated robotic platform. Software control over hardware allowed combination of individual unit operations into multistep organic synthesis. A Chempiler was created to program the platform (Figure 4B); the Chempiler creates low-level instructions for the hardware taking graph representation of the platform and abstraction representing organic synthesis (Figure 4C). In this way, it is possible to script and run published syntheses without reconfiguration of the platform, providing that necessary modules are present in the The synthesis of small molecules was and performed automatically with to [1Steiner S. et al.Organic synthesis in a modular robotic system driven by a chemical programming language.Science. 2019; 363eaav2211Crossref PubMed Scopus (194) Google Scholar]. Finally, by robotic with it is possible to autonomous in closed based on We recently demonstrated a flow system for navigating a of organic reactions utilizing an infrared as the for data The system was to the most autonomously on the of in the infrared between and [11Dragone V. et al.An autonomous organic reaction search engine for chemical reactivity.Nat. Commun. 2017; 815733Crossref PubMed Scopus (52) Google Scholar]. on that we a robotic platform for autonomous searching of chemical space with benchtop analytical nuclear magnetic resonance and mass for The search of chemical space is in Figure The platform in a closed loop with a machine learning the machine learning algorithm the most promising reactions that were then and analyzed automatically within the platform. The results of each experiment were automatically and the data were then used to the machine learning The use of machine learning allowed for autonomous exploration of reaction space allowing for discovery of four new chemical et an organic synthesis robot with machine learning to search for new 2018; PubMed Scopus Google Scholar]. In example, and colleagues developed a for using and successful experiments to synthesis et chemical in synthesis of Commun. 2019; PubMed Scopus Google Scholar]. This has been through of automation and machine learning to capture chemical in the synthesis of The exploration of chemical space by autonomous robots requires to the of the obtained results P.S. et al.How to explore chemical space using algorithms and automation.Nat. Rev. Chem. 2019; 3: 119-128Crossref Scopus (117) Google Scholar]. To we a for and of the experimental results (Figure the experiment must be to be valid and experimental and of the next step is to this has a This can be by a of a If the search that has been the experiment can be classified as not not information to our However, the has not been observed we need to it could be predicted using all the The that this is not novel new to some that obtained is for example, a reaction that be predicted can be classified as a new of In the this will enable of the experimental results by autonomous robots P.S. et al.How to explore chemical space using algorithms and automation.Nat. Rev. Chem. 2019; 3: 119-128Crossref Scopus (117) Google Scholar]. the ability to incorporate hardware and learning to out many smart automation the discovery of new molecules and to chemical synthesis [1Steiner S. et al.Organic synthesis in a modular robotic system driven by a chemical programming language.Science. 2019; 363eaav2211Crossref PubMed Scopus (194) Google Scholar]. In addition, learning coupled with ‘big and already can in some directly for many in the various of chemistry. This is because the fundamental of learning the to be and as new data is to more discoveries that a of chemical space and In our the of the field be on the potential of for chemical discovery with emphasis on automation coupled with machine learning (see the of these approaches shown this opinion Here, we have shown how automation and machine learning can efficiency and and are a universal combination for synthesis, optimization, and discovery in the chemistry can we enable synthetic chemists to without to how to much of chemistry can be done with the we drive adoption of the via development of a new way to synthesis can we enable synthetic chemists to without to how to much of chemistry can be done with the we drive adoption of the via development of a new way to synthesis developed to learning in humans that can an and complex chemical The for the search is designed in a way that the algorithm autonomously the experiments that the of new and a supervised approach used for a between or more different groups of The process is through of between the and a optimization algorithm that to a of conditions of an experimental or which desired unsupervised that data into a space that allows as much of relevant information as a supervised approach that to a of the is based on a approach that allows of many which by using input from the data a machine learning that can be applied for and The approach data into space in to any between