Accelerated Chemical Reaction Optimization Using Multi-Task Learning

Enhancing LC×LC separations through multi-task Bayesian optimization

Method development in comprehensive two-dimensional liquid chromatography (LC×LC) is a challenging process. The interdependencies between the two dimensions and the possibility of incorporating complex gradient profiles, such as multi-segmented gradients or shifting gradients, make trial-and-error method development time-consuming and highly dependent on user experience. Retention modeling and Bayesian optimization (BO) have been proposed as solutions to mitigate these issues. However, both approaches have their strengths and weaknesses. On the one hand, retention modeling, which approximates true retention behavior, depends on effective peak tracking and accurate retention time and width predictions, which are increasingly challenging for complex samples and advanced gradient assemblies. On the other hand, Bayesian optimization may require many experiments when dealing with many adjustable parameters, as in LC×LC. Therefore, in this work, we investigate the use of multi-task Bayesian optimization (MTBO), a method that can combine information from both retention modeling and experimental measurements. The algorithm was first tested and compared with BO using a synthetic retention modeling test case, where it was shown that MTBO finds better optima with fewer method-development iterations than conventional BO. Next, the algorithm was tested on the optimization of a method for a pesticide sample and we found that the algorithm was able to improve upon the initial scanning experiments. Multi-task Bayesian optimization is a promising technique in situations where modeling retention is challenging, and the high number of adjustable parameters and/or limited optimization budget makes traditional Bayesian optimization impractical.

Utopia Point Bayesian Optimization Finds Condition-Dependent Selectivity for N-Methyl Pyrazole Condensation

Utopia Point Bayesian Optimization (UPBO) was used to identify reaction conditions that are highly selective for the formation of N1 and N2-methyl-3-aryl pyrazole constitutional isomers. UPBO was used to explore a wide chemical space and identify basic reaction conditions for a typically acid-catalyzed Knorr pyrazole condensation. These studies revealed that selectivity in the reaction stems from a condition-dependent equilibrium of intermediates prior to dehydration. For the N2-methyl isomer reaction pathway, a hemiaminal intermediate was found to form reversibly under the reaction conditions, enabling a highly selective synthesis of the N2 isomer upon dehydrative workup. UPBO was able to successfully optimize conversion and selectivity simultaneously with search spaces of >1 million potential variable combinations without the need for high-performance computational resources.

Combining Bayesian optimization and automation to simultaneously optimize reaction conditions and routes

Reaching optimal reaction conditions is crucial to achieve high yields, minimal by-products, and environmentally sustainable chemical reactions. With the recent rise of artificial intelligence, there has been a shift from traditional Edisonian trial-and-error optimization to data-driven and automated approaches, which offer significant advantages. Here, we showcase the capabilities of an integrated platform; we conducted simultaneous optimizations of four different terminal alkynes and two reaction routes using an automation platform combined with a Bayesian optimization platform. Remarkably, we achieved a conversion rate of over 80% for all four substrates in 23 experiments, covering ca. 0.2% of the combinatorial space. Further analysis allowed us to identify the influence of different reaction parameters on the reaction outcomes, demonstrating the potential for expedited reaction condition optimization and the prospect of more efficient chemical processes in the future.

ORDerly: Data Sets and Benchmarks for Chemical Reaction Data

Machine learning has the potential to provide tremendous value to life sciences by providing models that aid in the discovery of new molecules and reduce the time for new products to come to market. Chemical reactions play a significant role in these fields, but there is a lack of high-quality open-source chemical reaction data sets for training machine learning models. Herein, we present ORDerly, an open-source Python package for the customizable and reproducible preparation of reaction data stored in accordance with the increasingly popular Open Reaction Database (ORD) schema. We use ORDerly to clean United States patent data stored in ORD and generate data sets for forward prediction, retrosynthesis, as well as the first benchmark for reaction condition prediction. We train neural networks on data sets generated with ORDerly for condition prediction and show that data sets missing key cleaning steps can lead to silently overinflated performance metrics. Additionally, we train transformers for forward and retrosynthesis prediction and demonstrate how non-patent data can be used to evaluate model generalization. By providing a customizable open-source solution for cleaning and preparing large chemical reaction data, ORDerly is poised to push forward the boundaries of machine learning applications in chemistry.

Augmenting DMTA using predictive AI modelling at AstraZeneca

Design-Make-Test-Analyse (DMTA) is the discovery cycle through which molecules are designed, synthesised, and assayed to produce data that in turn are analysed to inform the next iteration. The process is repeated until viable drug candidates are identified, often requiring many cycles before reaching a sweet spot. The advent of artificial intelligence (AI) and cloud computing presents an opportunity to innovate drug discovery to reduce the number of cycles needed to yield a candidate. Here, we present the Predictive Insight Platform (PIP), a cloud-native modelling platform developed at AstraZeneca. The impact of PIP in each step of DMTA, as well as its architecture, integration, and usage, are discussed and used to provide insights into the future of drug discovery.

A Slug Flow Platform with Multiple Process Analytics Facilitates Flexible Reaction Optimization

Flow processing offers many opportunities to optimize reactions in a rapid and automated manner, yet often requires relatively large quantities of input materials. To combat this, the use of a flexible slug flow reactor, equipped with two analytical instruments, for low-volume optimization experiments are reported. A Buchwald-Hartwig amination toward the drug olanzapine, with 6 independent optimizable variables, is optimized using three different automated approaches: self-optimization, design of experiments, and kinetic modeling. These approaches are complementary and provide differing information on the reaction: pareto optimal operating points, response surface models, and mechanistic models, respectively. The results are achieved using <10% of the material that would be required for standard flow operation. Finally, a chemometric model is built utilizing automated data handling and three subsequent validation experiments demonstrate good agreement between the slug flow reactor and a standard (larger scale) flow reactor.

Transfer Learning Bayesian Optimization to Design Competitor DNA Molecules for Use in Diagnostic Assays

With the rise in engineered biomolecular devices, there is an increased need for tailor-made biological sequences. Often, many similar biological sequences need to be made for a specific application meaning numerous, sometimes prohibitively expensive, lab experiments are necessary for their optimization. This paper presents a transfer learning design of experiments workflow to make this development feasible. By combining a transfer learning surrogate model with Bayesian optimization, we show how the total number of experiments can be reduced by sharing information between optimization tasks. We demonstrate the reduction in the number of experiments using data from the development of DNA competitors for use in an amplification-based diagnostic assay. We use cross-validation to compare the predictive accuracy of different transfer learning models, and then compare the performance of the models for both single objective and penalized optimization tasks.

On-Demand Optimization of Colorimetric Gas Sensors Using a Knowledge-Aware Algorithm-Driven Robotic Experimental Platform

Synthesizing the best material globally is challenging; it needs to know what and how much the best ingredient composition should be for satisfying multiple figures of merit simultaneously. Traditional one-variable-at-a-time methods are inefficient; the design-build-test-learn (DBTL) method could achieve the optimal composition from only a handful of ingredients. A vast design space needs to be explored to discover the possible global optimal composition for on-demand materials synthesis. This research developed a hypothesis-guided DBTL (H-DBTL) method combined with robots to expand the dimensions of the search space, thereby achieving a better global optimal performance. First, this study engineered the search space with knowledge-aware chemical descriptors and customized multiobjective functions to fulfill on-demand research objectives. To verify this concept, this novel method was used to optimize colorimetric ammonia sensors across a vast design space of as high as 19 variables, achieving two remarkable optimization goals within 1 week: first, a sensing array was developed for ammonia quantification of a wide dynamic range, from 0.5 to 500 ppm; second, a new state-of-the-art detection limit of 50 ppb was reached. This work demonstrates that the H-DBTL approach, combined with a robot, develops a novel paradigm for the on-demand optimization of functional materials.

Identifying general reaction conditions by bandit optimization

Reaction conditions that are generally applicable to a wide variety of substrates are highly desired, especially in the pharmaceutical and chemical industries1-6. Although many approaches are available to evaluate the general applicability of developed conditions, a universal approach to efficiently discover these conditions during optimizations is rare. Here we report the design, implementation and application of reinforcement learning bandit optimization models7-10 to identify generally applicable conditions by efficient condition sampling and evaluation of experimental feedback. Performance benchmarking on existing datasets statistically showed high accuracies for identifying general conditions, with up to 31% improvement over baselines that mimic state-of-the-art optimization approaches. A palladium-catalysed imidazole C-H arylation reaction, an aniline amide coupling reaction and a phenol alkylation reaction were investigated experimentally to evaluate use cases and functionalities of the bandit optimization model in practice. In all three cases, the reaction conditions that were most generally applicable yet not well studied for the respective reaction were identified after surveying less than 15% of the expert-designed reaction space.

Pushing the boundaries of phosphorylase cascade reaction for cellobiose production II: Model-based multiobjective optimization

The inherent complexity of coupled biocatalytic reactions presents a major challenge for process development with one-pot multienzyme cascade transformations. Kinetic models are powerful engineering tools to guide the optimization of cascade reactions towards a performance suitable for scale up to an actual production. Here, we report kinetic model-based window of operation analysis for cellobiose production (≥100 g/L) from sucrose and glucose by indirect transglycosylation via glucose 1-phosphate as intermediate. The two-step cascade transformation is catalyzed by sucrose and cellobiose phosphorylase in the presence of substoichiometric amounts of phosphate (≤27 mol% of substrate). Kinetic modeling was instrumental to uncover the hidden effect of bulk microviscosity due to high sugar concentrations on decreasing the rate of cellobiose phosphorylase specifically. The mechanistic-empirical hybrid model thus developed gives a comprehensive description of the cascade reaction at industrially relevant substrate conditions. Model simulations serve to unravel opposed relationships between efficient utilization of the enzymes and maximized concentration (or yield) of the product within a given process time, in dependence of the initial concentrations of substrate and phosphate used. Optimum balance of these competing key metrics of process performance is suggested from the model-calculated window of operation and is verified experimentally. The evidence shown highlights the important use of kinetic modeling for the characterization and optimization of cascade reactions in ways that appear to be inaccessible to purely data-driven approaches.

Autonomous nanomanufacturing of lead-free metal halide perovskite nanocrystals using a self-driving fluidic lab

Lead-based metal halide perovskite (MHP) nanocrystals (NCs) have emerged as a promising class of semiconducting nanomaterials for a wide range of optoelectronic and photoelectronic applications. However, the intrinsic lead toxicity of MHP NCs has significantly hampered their large-scale device applications. Copper-base MHP NCs with composition-tunable optical properties have emerged as a prominent lead-free MHP NC candidate. However, comprehensive synthesis space exploration, development, and synthesis science studies of copper-based MHP NCs have been limited by the manual nature of flask-based synthesis and characterization methods. In this study, we present an autonomous approach for the development of lead-free MHP NCs via seamless integration of a modular microfluidic platform with machine learning-assisted NC synthesis modeling and experiment selection to establish a self-driving fluidic lab for accelerated NC synthesis science studies. For the first time, a successful and reproducible in-flow synthesis of Cs3Cu2I5 NCs is presented. Autonomous experimentation is then employed for rapid in-flow synthesis science studies of Cs3Cu2I5 NCs. The autonomously generated experimental NC synthesis dataset is then utilized for fast-tracked synthetic route optimization of high-performing Cs3Cu2I5 NCs.

Parallel multi-droplet platform for reaction kinetics and optimization

We present an automated droplet reactor platform possessing parallel reactor channels and a scheduling algorithm that orchestrates all of the parallel hardware operations and ensures droplet integrity as well as overall efficiency. We design and incorporate all of the necessary hardware and software to enable the platform to be used to study both thermal and photochemical reactions. We incorporate a Bayesian optimization algorithm into the control software to enable reaction optimization over both categorical and continuous variables. We demonstrate the capabilities of both the preliminary single-channel and parallelized versions of the platform using a series of model thermal and photochemical reactions. We conduct a series of reaction optimization campaigns and demonstrate rapid acquisition of the data necessary to determine reaction kinetics. The platform is flexible in terms of use case: it can be used either to investigate reaction kinetics or to perform reaction optimization over a wide range of chemical domains.