System and method for predicting chemical reaction outcomes and conditions with high confidence based on diverse and accurate data sets
A machine learning model trained on diverse chemical reaction datasets addresses the inefficiencies in predicting chemical reaction outcomes, providing accurate and reliable predictions for a broad range of compounds, thereby reducing costs and time in drug discovery.
Patent Information
- Application Number
- JP2024524387
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-06-10
- Filing Date
- 2022-10-24
- Publication Date
- 2024-11-13
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Current methods for predicting the outcomes of chemical reactions are inaccurate and inefficient, particularly due to limitations in data availability and the inability to generalize beyond small molecular spaces, leading to high failure rates and increased costs in drug discovery and other industries.
A machine learning model trained on large-scale, high-throughput chemical reaction datasets, utilizing semi-automated laboratories and diverse data sources, to predict reaction outcomes with high accuracy and reliability, including the use of DNA-encoded libraries, MALDI-TOF mass spectrometry, and proprietary data mining algorithms.
The model achieves high accuracy in predicting chemical reaction outcomes for a wide range of commercially valuable compounds, reducing experimental costs and time, and enabling reliable predictions for complex reactions, such as Suzuki and Heck couplings.
Smart Images

Figure 2024541898000001_ABST
Abstract
Description
[Technical field]
[0001] Cross-references to related cases This application claims priority to U.S. Provisional Patent Application No. 63 / 270,932, entitled "Trust-Worthy Systems And Methods For Discovering Novel Chemical Reactions Or Classes Of Reactions," filed October 22, 2021, and U.S. Provisional Patent Application No. 63 / 351,295, entitled "Tool For Recommending Chemical Reaction Conditions," filed June 10, 2022, both of which are incorporated by reference herein. This application is a related application to U.S. Patent Application No. 17 / 060,765, entitled "Systems And Method For Designing Organic Synthesis Pathways For Desired Organic Molecules," filed January 14, 2021. [Background technology]
[0002] Predicting the outcome of chemical reactions is a central challenge for many industries that use chemistry, including pharmaceuticals, agriculture, and cosmetics. Consider drug discovery: for every drug on the market, typically thousands of different drugs need to be synthesized and tested in the laboratory. Inefficiencies in the chemical processes used, for example not yielding the desired product in the laboratory, can have significant downstream effects and increase the price of goods and services.
[0003] Unfortunately, predicting the outcome of many types of chemical reactions is difficult for humans and computers alike. For example, industry data has shown that the Buchwald-Hartwig coupling reaction has an internal failure rate of 55%. This challenge is difficult even for computers. Although computational chemistry has made great advances, it will be some time before simulations can replace experiments. It may be possible to accurately predict the outcome of a particular chemical reaction from first principles if (a) the reaction mechanism is well understood, (b) the reaction mechanism is simple (e.g., involving only one transition state), and (c) the overall chemical complexity of the reactants and reagents involved is low.
[0004] Other approaches to predicting reaction outcomes are based on machine learning, specifically deep neural networks (DNNs). However, machine learning methods are fundamentally limited by the available data. Deep neural networks trained on publicly available data do not generalize due to inherent biases in the publicly available data. In particular, nearly all data sources omit failed experiments entirely.
[0005] However, current approaches are limited in terms of accurately predicting reaction outcomes for broad chemical spaces. One limiting factor is that most previous studies have considered only a very small range of molecular space. For example, Shields et al. mainly considered reaction spaces with only a few or even a single reaction product.
[0006] Unpredictable reaction outcomes, among other things, increase the costs and lengthen the timelines for introducing new drugs to market. Thus, there is a need for systems and methods for predicting the outcomes and conditions of chemical reactions with high accuracy and reliability.
[0007] These embodiments are illustrated by way of example, and not by way of limitation, in the accompanying drawings in which like reference symbols indicate similar elements and in which: [Brief description of the drawings]
[0008] [Figure 1] 1 is a screen shot of an embodiment of a graphical user interface (GUI) for an embodiment of a model for predicting outcomes and conditions of chemical reactions. [Diagram 2] 13 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting chemical reaction outcomes and conditions illustrating the creation of new functional groups. [Diagram 3] 1 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting the outcomes and conditions of chemical reactions, showing a functionalization search overview. [Figure 4] 13 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting outcomes and conditions of chemical reactions, with marked locations showing a search overview in hover state. [Diagram 5] 1 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting outcomes and conditions of chemical reactions, showing search summary results filtered by functionalization type. [Figure 6] 1 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting chemical reaction outcomes and conditions, showing a detailed view of functionalization. [Figure 7] 13 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting chemical reaction outcomes and conditions, showing functionalization details of an expanded reference reaction. [Figure 8A] 1 is a flow chart illustrating an embodiment of a method for predicting outcomes and conditions of a chemical reaction. [Figure 8B] 1 is a flow chart illustrating an embodiment of a method for predicting outcomes and conditions of a chemical reaction. [Figure 9] 1 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting outcomes and conditions of a chemical reaction. [Figure 10]1 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting outcomes and conditions of a chemical reaction. [Figure 11] 1 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting outcomes and conditions of a chemical reaction. [Figure 12] FIG. 13 illustrates input and output formats in a GUI embodiment for an embodiment that is a model for predicting chemical reaction outcomes and conditions. [Figure 13] 1 is a flow chart illustrating an embodiment of a data collection method for a model for predicting chemical reaction outcomes and conditions. [Figure 14] 1 is a screenshot of an advanced query builder in a GUI embodiment for an embodiment of a model for predicting outcomes and conditions of chemical reactions. [Figure 15] 13 is a screenshot of a depiction of a reference reaction in an embodiment of a GUI for an embodiment of a model for predicting outcomes and conditions of a chemical reaction. [Figure 16] 1 is a screenshot of a reaction editor in a GUI embodiment for an embodiment of a model for predicting the outcomes and conditions of a chemical reaction. [Figure 17] 1 is a flow chart illustrating an embodiment of a method for predicting outcomes and conditions of a chemical reaction. [Figure 18] FIG. 1 is an exemplary block diagram illustrating an embodiment of a system for implementing an embodiment of the method in the present disclosure. [Figure 19] FIG. 1 is an exemplary block diagram illustrating a computing device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0009] The invention aims to obtain machine learning models that can accurately predict the reaction outcomes of a wide range of commercially valuable compounds. The innovation is designed to address difficult problems in chemistry, such as recommending high-yielding conditions for reactions such as Suzuki and Heck couplings.
[0010] There are several embodiments aimed at achieving high accuracy for a wide range of commercially valuable compounds. In particular, large datasets of chemical reactions can be generated using semi-automated high-throughput laboratories. Another innovation is prioritizing reactions (for laboratory execution) using novel methods focused on achieving high accuracy for user-relevant reactions.
[0011] In some embodiments, cost-effective, high-throughput (HT) organic chemistry laboratories may be used in methods to predict the outcomes and conditions of chemical reactions with high accuracy and confidence based on the creation of focused large-scale chemical reaction datasets to train machine learning models. Such embodiments may include a process in which a model (computer program) learns to accurately and reliably predict the outcomes of reactions of a wide range of commercially valuable compounds with high accuracy and appropriate prediction of uncertainty. Such models may be applied to predict challenging problems such as organic chemistry.
[0012] Embodiments may use high-throughput laboratories designed with two key constraints: (a) low cost per reaction (e.g., less than $1 per reaction), and (b) high throughput (e.g., 5000 reactions per week). For example, the cost constraints may be addressed by sourcing building blocks from large-scale providers such as MolPort.
[0013] In one embodiment, to ensure that the model can generalize to a wide range of drug-like molecules, the experimental reactions are selected to include reactions that are specifically related to the target set that includes drug-like molecules, as described later in this specification. Because drug-like molecules have a highly biased structure, selecting such a focused set of experimental reactions can significantly improve performance on drug-like molecules and can be covered with a relatively small number of focused experiments. In some embodiments, the target set may be specified to consist of any molecule type that is most relevant for a given application. In some embodiments, experimental reactions in a chemical space that represent the breadth of the chemical space of the target set are selected to better train the model so that it can generalize to a wide range of molecules in a particular chemical space.
[0014] In some embodiments, methods are disclosed for training machine learning models to generate results within specific reaction classes that are robust enough to directly inform experimental design.
[0015] Some embodiments include the deliberate creation of a dataset of chemical reactions that is used to train a highly specialized model for a particular class of chemical reactions. Additionally, embodiments may include the use of medium or high throughput chemistry experiments to do so, the use of DNA encoded libraries to do so, the use of MALDI-TOF mass spectrometry to do so, the use of MISER chromatography to do so, the use of an automated chemistry lab to do so, the use of proprietary data mining algorithms to do so.
[0016] Some embodiments may include an intuitive graphical user interface that allows users to query the system for desired chemical reactions, intuitively display the system's recommendations, and reference extensive supporting information that gives users new confidence in the results, which may include ML model outputs, proprietary experimental datasets (discussed above), among others.
[0017] Some embodiments may combine all of the above in a methodology / computer system that allows for user-machine and machine-machine interaction with the goal of identifying ways to perform specific chemical reactions, including methods that are better in some respect, such as having higher yields, or identifying chemical reactions that can be performed under specific user-defined constraints, such as a limited range of conditions, or that can be sent directly to an automated synthesis system for execution.
[0018] In some contexts, "machine-to-machine interaction" means outputting the results of a system directly to another computer system, such as the computer system that runs an automated laboratory.
[0019] Some embodiments may include the use of one or more of the above-mentioned methodologies in the areas of late-stage functionalization of compounds in the drug discovery pipeline (such embodiments are described in more detail later in this disclosure) and prediction of conditions for chemical reactions (such embodiments are described in more detail later in this disclosure).
[0020] Some embodiments may include the use of one or more of the methodologies described above in the field of DNA-encoded library (DEL) synthesis. Such embodiments include applying the strategy of steps 1-3, as described in more detail below, to identify reaction conditions that allow for more diverse or more efficient generation of DELs. This may be useful to rule out certain chemical reactions or certain conditions of chemical reactions that are widely known but are not acceptable in the context of DEL synthesis. In such embodiments, the methodology of steps 1-3 (described below) may be used to increase the efficiency of DEL synthesis by better selecting substrates that will successfully undergo certain reactions. This is important because in the context of DEL synthesis, large scale and high purity standards are required. In other such embodiments, a given chemical reaction may be optimized for applicability in conditions suitable for DEL synthesis by applying the methodology of steps 1-3. Running chemical reactions under mild conditions that do not adversely affect the DNA tags (e.g., degrading the DNA) is helpful for DEL synthesis.
[0021] Some embodiments may include the use of one or more of the above-mentioned methodologies in the field of automating the execution of chemical reactions using robots. While typical approaches would focus on the automated synthesis of hardware or the automated synthesis of a specific application, the approach of the embodiments presented in this disclosure focuses on comprehensive coverage of a portion of the chemical space using large reaction databases and / or the development of models that use these databases to make accurate predictions. This is unique in that it addresses a key problem that other procedures for automatically executing chemical reactions have, namely, the need for a user to program the robot with the substrates, conditions, and other parameters required for execution. In one such embodiment, the method of steps 1-3 focuses on establishing conditions for executing a particular chemical reaction that maximize the range, i.e., the yield of a given reaction for a large number of different substrates under the same conditions.
[0022] In embodiments, the method allows chemists to predict new chemical properties in a reliable manner. In this context, "new chemical properties" refers to one or more of the following: discovering a new class of chemical reactions, expanding the molecular scope of a known class of reactions, enabling the synthesis of new compounds, increasing the yield of a known reaction, or discovering new conditions for a known chemical reaction.
[0023] In some embodiments, the method comprises the following steps 1 to 3:
[0024] Step 1) Creating a detailed dataset of chemical reactions that is broad enough to predict with a high degree of confidence the outcome of selected new chemical reactions, focusing on one or more classes of chemical reactions that are of interest to the user of the methodology. Such a dataset can be created based on a combination of one or more of the following methodologies (a-g):
[0025] a. Automatic or manual analysis of reaction text data (e.g., literature, textbooks, patents, laboratory notebooks, internet websites, etc.) to extract chemical reactions (which may be part of the dataset). Such analysis may involve i) modifying commonly known techniques to perform chemical structures of interest. Specific new components are included in the description of the embodiments. Such analysis may involve ii) enriching the dataset with manually specified reaction rules that can be used to automatically generate artificial reactions. Such rules may include information such as the type of transformation, its extent, and the necessary reaction conditions. Such analysis may also involve iii) enriching the dataset with external datasets of molecules or chemical reactions. In particular, chemical reactions taken from patents filed with the U.S. Patent and Trademark Office may be used for that purpose.
[0026] b. Use of medium and high throughput analytical techniques such as MALDI-TOF and MISER. MALDI-TOF is a commonly known technique in mass spectrometry and is a set of techniques used to identify chemical compounds. One feature of MALDI-TOF that is important for the purposes of this methodology is that it can achieve high throughput (less than 1 second per sample). In embodiments, this technique may require adaptation to create data sets of interest. Applying this approach to the creation of data sets of interest lends itself to the use of appropriate analytical procedures, including but not limited to the selection of matrix composition and ionization setup. MISER chromatography with MS detection is a commonly known technique for the separation and identification of compounds. One feature that is useful for the purposes of this methodology is that it can achieve medium to high throughput (less than 30 seconds per sample). This technique may also require adaptation to create data sets of interest. In embodiments, these techniques can be used to create data sets of reaction results that are large and of sufficient quality to drive systems such as those described in this disclosure. In particular, this can be facilitated by automatically feeding these data sets to machine learning models that predict the outcomes of other chemical reactions. In embodiments, deep neural networks can be used to remove noise or improve the fidelity of analytical results generated using high throughput analytical techniques such as MALDI-TOF or MISER. In particular, neural networks can be trained to predict the full mass spectrogram of a molecule based on the output from MALDI-TOF or MISER. For example, in one embodiment, machine learning methods can be used to predict the ionization of compounds analyzed by MALDI-TOF or other MS techniques based on data on the ionization properties of other compounds with known ionization properties, thereby improving the accuracy of the "quantitativeness" of these analytical techniques.
[0027] c. Use of medium and high throughput chemistry experiments. Commonly known techniques may be modified to perform reactions of interest. These techniques may be used to generate data sets of sufficient size and quality to drive systems such as those described in this disclosure. Additionally, in embodiments, users may run chemical reactions themselves to validate system results. Additionally, in some embodiments, experiments may be performed utilizing automated liquid and / or solid dispensers.
[0028] d. Use of DNA-encoded libraries. In some embodiments, DNA-encoded libraries (DELs) can be used as a means to generate experimental data on the reactivity of DNA-tagged reagents, making them suitable for training machine learning models (especially models). In one embodiment of this kind, a library of reagents is used, each tagged with a different DNA tag, having a common functional group. A mixture of such tagged library components (A) can undergo a chemical reaction with a specific reagent(s) (B), so that covalent bonds are formed between some elements of A and some elements of B. Appropriate construction (tagging or immobilization) of reagent B allows subsequent inexpensive identification of the formed AB adducts. In embodiments, by using widely known techniques such as polymerase chain reaction (PCR) or next generation sequencing (NGS), reagent B can be attached to large molecules such as DNA strands, proteins (polypeptides), nanoparticles or polymer resin beads, making it possible to wash out non-reactive library components A and then identify the DNA tags of library components A that undergo reaction with reagent B.
[0029] e. Use of Simulation Software. In embodiments, molecular simulation software can be used to predict the outcome of chemical reactions to enrich the dataset. In embodiments, simulation software can be used to predict the outcome of chemical reactions of simple molecules with sufficiently accurate results to bootstrap learning of machine learning models that predict the outcome of more complex reactions.
[0030] f. Experimenting with existing literature data and existing machine learning models to discover which experiments are most informative in the context of the chemistry of interest and testing whether the created dataset is already detailed enough. For example, in embodiments, i. a machine learning model can be used to discover the most useful reactions to run to enable more reliable predictions. In particular, machine learning certainty estimates can be used to select reactions with the highest uncertainty. Reactions can be run and analyzed in the laboratory to strengthen the model by adding reactions from models with the highest prediction uncertainty to the training set. Further, in embodiments, ii. existing literature data or other sources can be used to determine whether the dataset is already detailed enough for the chemistry target of interest. In one such embodiment, the existing sources can be considered in terms of the number of data points included in the chemical structure of interest. Such analysis can be performed by chemists, statistical models, or a combination of both. In other such embodiments, these sources can be used as training datasets for machine learning models such as those referenced in step 2a below. Further, in embodiments, iii. the space of chemical reactions can be grouped using broad chemical features (calculated using commonly known chemistry software and machine learning models). For example, in embodiments, a smaller number of groups of chemical reactions can be selected from these groups such that these selected groups provide the most useful training data for machine learning algorithms to increase robustness in predicting the outcome of a pool of chemical reactions from other groups when these selected groups are performed and analyzed in the laboratory. In another example, embodiments may select groups of reactions that are relatively the most difficult to predict the outcome of such reactions without using a robust machine learning system. Moreover, the selected groups can be used to design experiments such that chemical reactions are densely sampled from each group of interest. Such experimental design allows for the training of more robust machine learning models.In other embodiments, iv. synthetic planning software can also be used to explore which reactions can reach a given molecule of interest.
[0031] g. Manual labeling of reactions generated by machine learning software by a chemist. In embodiments, i. reaction candidates may be generated by a computer system using the methods described below in subsection 3a "Late-stage functionalization" and given to one or more chemists who are instructed to assign reaction labels.
[0032] In embodiments, the various techniques described above (1.a-1.g) can be combined in the context of creating a dataset of chemical reactions (step 2) that can be used to train a machine learning model focused on identifying reliable chemical structures, and continuously updating the dataset with new reaction data carefully selected to maximize robustness while minimizing laboratory experimental costs.
[0033] Step 2) Training a machine learning model focused on identifying reliable chemical structures using the detailed dataset created and other relevant sources of information.
[0034] In embodiments, the machine learning model is not limited to the dataset created in step 1, but may be any model that can be trained using the same and predict a chemical structure of interest. In particular, detailed conditions (reagents, solvents, temperature, etc.) for carrying out a reaction with the chemical structure of interest, if required, are included in the output from the model.
[0035] In an embodiment, a machine learning model is created that focuses on making reliable predictions for novel (undiscovered) chemical structures (e.g., new molecules). The system is designed to show reliable predictions in exchange for the number of predictions shown.
[0036] c. In an embodiment, the machine learning model can be trained in a way that allows it to make sufficiently reliable predictions, which may be provided by using one or more of: i. the detailed dataset created in step 1; ii. training on additional datasets of molecules or reactions to expose the model to a broader knowledge of what molecules exist; iii. ensembles or other techniques used to increase generalization across domains (causal learning in one embodiment).
[0037] d. In embodiments, these techniques can be used to train machine learning models that are confident in their predictions regarding reaction outcomes (Step 2).
[0038] Step 3) Find a way to combine a or b or both to perform the chemical reaction of interest. a. The easy-to-use interface is adapted to the specific chemical structure of interest and focuses on enabling efficient interaction between the user and the machine. For example: i. A reliable model can be combined with a detailed experimental dataset that chemists can trust. ii. The user can explore the most chemically relevant chemical reactions from both sources, so that the results can be reviewed and compared with the literature and the created dataset (step 1). In an embodiment of ii, these relevant chemical reactions are selected by the similarity in the fingerprints of the potentially relevant reaction products and the reaction product of interest. In another embodiment of ii, these relevant chemical reactions are selected based on the similarity of the molecular structure of the potentially relevant reaction products and the reaction product of interest. Examples of such molecular structure include the presence of functional groups, the order of atoms in which the reaction is taking place, and the intramolecular or intermolecular nature of the reaction.
[0039] b. Automated Synthesis Platform. In embodiments, users can propose and execute chemical reactions that experimentally validate new predictions.
[0040] In an embodiment, the procedure of steps 1-3 may be applied iteratively with step 1 using the results obtained from step 3 to improve the results of each step. Late functionalization
[0041] In an embodiment of the proposed methodology, a robust machine learning model, a literature-based dataset (obtained using a custom multi-step extraction pipeline), and an intuitive GUI are combined and applied to late-stage functionalization. Late-stage functionalization is a methodology in the drug discovery process where promising drug candidates are optimized by small modifications to their structure. Every new structure enjoys the activity data available on close analogues, making it highly valuable. However, performing late-stage functionalization comes with a major challenge: finding reaction conditions that allow transformations to be performed chemoselectively.
[0042] The embodiments provide the user with highly reliable predictions of how a molecule will be modified under what conditions, which in turn provides the user with access to a wider range of analogs than would be possible without the use of the embodiments.
[0043] In one such embodiment, 1. the machine learning model is a neural network based on the Transformer architecture (in other embodiments the model may be based on other model architectures as well) and is trained using some of the techniques described above in step 2 to enable reliable predictions for chemical structures of interest.
[0044] 2. In other such embodiments, a. models can be trained using a technique called self-supervised learning, which uses data extracted from publicly available documents (such as patents) that contain information about previously successful chemical reactions. Details of these chemical reactions can be extracted using machine learning methods that create extensive, detailed, and robust datasets. i. Chemical data extraction is performed using a pipeline of machine learning models that analyze the data in multiple stages: 1. The first stage uses a model that predicts whether a piece of text describes a chemical reaction. 2. The second stage uses a model that labels within a paragraph of a chemical reaction description as a header, description, or footer. 3. The third stage uses a model that predicts which parts of the reaction description are names of entities such as reaction products, substrates, solvents, catalysts, and other conditions required to perform the reaction. ii. In embodiments, all stages of the reaction extraction pipeline can be performed by neural networks based on transformer architectures that are task-specific and trained using manually labeled datasets. These are a different type of models than those used for chemistry-related tasks (such as the models described in embodiment 1 above at the beginning of this paragraph). The model described in section i.1-3 - embodiment 2 is trained specifically for natural language processing tasks. iii. The detailed reaction properties predicted by the pipeline improve the efficiency of using the self-supervised learning techniques mentioned above. The model can be trained using b. supervised learning objectives based on the same data as in the previous point, or c. a large number of automatically generated artificial chemical reactions that are likely to be wrong (i.e., probably would not work in a lab).
[0045] 3. For embodiments predicting new chemical properties, only highly reliable predictions are selected from the model output. To achieve this, the model predicts its reliability as a separate output, and only chemical reactions for which the value of this output exceeds a predefined threshold are selected. For example, chemical reactions can be generated by generative models (models that generate substrates based on products, or models that generate products based on substrates) or using so-called reaction templates. In the latter case, the final confidence level can be calculated using a discriminator model (a classification model that outputs a single value indicating the probability that the chemical reaction will be successful) trained on positive and negative examples.
[0046] FIG. 1 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting the outcome and conditions of a chemical reaction. In FIG. 1, searches 10, 20 are each functionalization searches previously performed in the system. Molecules 12, 14 represent molecules for which searches have previously been performed. And schematics 16, 18 are presented for each search. In the GUI, selecting an item in either of the searches 10, 20 redirects the user to the search exploration page (FIG. 3) for the related item. The "New Functionalization" button provides the ability to create a new functionalization (FIG. 2). In this screen, similar atoms may be colored similarly. For example, the F atom in molecule 12 may be light blue and the "O" atom may be red.
[0047] FIG. 2 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting the outcomes and conditions of chemical reactions, showing the creation of a new functionalization. Molecules 22 (already identified as molecule 12), 24, 26 show molecules that can be selected to perform a search. A new molecule can be created by selecting new compound 28. After selecting a compound, a search (prediction) can be started by selecting "Start Prediction". In this screen, similar atoms and similar functional groups can be colored similarly. For example, the F atom in molecule 22 can be colored light blue and the "NH" group in molecule 24 can be colored royal blue.
[0048] FIG. 3 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting the outcomes and conditions of chemical reactions, showing a functionalization search overview. A list 30 includes currently displayed functionalizations 30a-30e, each of which, when selected, redirects to a functionalization detail view (FIG. 6) and is displayed next to the selected input molecule 32 (molecule 12 in FIG. 1). A numerical value indicating the confidence level of the model prediction (e.g., shown as a percentage) for each functionalization 30a-30e may be included. By default, a selection of the top predictions is displayed. The specific positions corresponding to each prediction are marked on the graph with pie charts (34a-34d) that code the confidence score (value) and type (color) of the top prediction at each position. As shown, the functionalizations correspond to the pie charts as follows: 30a vs. 34a, 30b vs. 34b, 30c vs. 34c, and 30d vs. 34d. The pie chart corresponding to functionalization 30e is not shown. By maintaining an indicator over an item, a location can be "hovered" and information can be displayed, as shown in Figure 4. Predictions can be filtered by functionalization type, as shown in Figure 5. In this view, similar functionalizations can be colored similarly by the associated atoms of the functionalization. For example, the F functionalization found in molecule 32 may be colored light blue, while the Br functionalization found in molecule 32 may be colored beige.
[0049] FIG. 4 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting chemical reaction outcomes and conditions, showing a search summary that appears when hovering over a position 40. Hovering displays a list 42 of functionalizations found at the position 40 in a floating menu next to the position. Each functionalization 44a, 44b can be selected to display details about the functionalization, such as the percent confidence level. In this view, similar functionalizations may be colored similarly depending on the associated atom of the functionalization. For example, an F functionalization found on molecule 32 may be colored light blue, while a Br functionalization found on molecule 32 may be colored beige.
[0050] 5 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting outcomes and conditions of chemical reactions, showing search summary results filtered by functionalization type F38. Selecting functionalization type F38 reduces the functionalization list displayed when hovering over position 40 to those that meet the filter criteria.
[0051] FIG. 6 is a screenshot of an embodiment of a GUI for an embodiment of a model for predicting the outcome and conditions of a chemical reaction, showing a detailed view of the functionalization. The screenshot shows that the GUI offers navigation to an overview 44 of the selected position, as well as other functionalizations in the list 30. A predicted reaction graph 46 is displayed, including the substrate (on the right, also associated with the reaction's directional arrow) and the reaction conditions (associated with the reaction's directional arrow). Below the predicted reaction, a list of reference reactions 48, 50 (partially displayed) is displayed. Each reference reaction, e.g. reaction 48, can be further expanded (52) to show detailed information as shown in FIG. 7. In this view, similar functionalizations can be colored similarly according to the relevant atoms of that functionalization. For example, F functionalizations can be light blue, Br functionalizations beige.
[0052] 7 is a screenshot of a GUI embodiment for an embodiment of a model for predicting chemical reaction outcomes and conditions showing functionalization details expanded from reference reaction 48. Expanding link 52 (FIG. 6) results in step 54 being displayed.
[0053] Section 1
[0054] 8 is a flow chart illustrating an embodiment of a method 800 for predicting chemical reaction outcomes and conditions, including an embodiment of a data collection method. In some embodiments, a data collection method or data collection can be understood as a method that includes multiple steps (such as selecting a target set or purchasing reagents) used to design and execute experiments that are used to train machine learning models used in the disclosed system. One goal of the data collection method is to gather data suitable for making accurate predictions about high-yield reaction conditions for a particular class of chemical transformations in a particular subset of chemical space.
[0055] Method 800 includes several steps that may (but need not) be performed in the order shown. FIG. 8 illustrates actions 801 interacting with computer system(s) 851 via GUI 100. The first step 852 is selection of a target set (see text for examples). User input 808 may be received at GUI 100 via paths 838, 876. The target set may be selected from published sources 804 or from external sources 802 that may include products and other structures of reactions of interest to the user 806. The selected target set is provided 888 to step 854, which focuses on prioritizing reactions for execution and may receive user input 810 via paths 840, 878. The prioritization input 810 may be selected by the user based on considerations 812 that may include model output 814, available resources 816, commercially available reagents 818, and reagents available in user inventory 820. In step 894, the selected reactions are sent to step 856. The selected reactions are sent via step 872 (890, 908). In step 872, a computer program can monitor the execution of the experiment by a user and hardware in an automated laboratory. In step 856, the reactions selected in step 854 are executed and the results are analyzed. The analysis of the results in step 856 can, in some embodiments, be assisted by step 874, where a computer program and / or ML model and / or simulation is executed to support the interpretation of the analytical data (910). The results of step 874 are provided to steps 858 (via path 896), 870 (via path 896), 868 (via path 886), 836 (via paths 886, 850). In step 858, a dataset for model training is assembled. In step 870, a dataset is assembled for model evaluation 898. The dataset for model training from step 858 is provided via path 900 to step 860, where an ML model is trained. The trained model is available via path 902 to step 862 where the model is evaluated.In evaluating the model, step 862 may further receive input from step 870 via path 912 of a data set for model evaluation. Step 862 may also receive user selection and prioritization input 832 via paths 844, 882. In step 866, a decision is made whether to repeat any steps. With input from decision 866 via path 906 and possible user selection and prioritization input 834 provided via paths 846, 884, a set of next reactions is prioritized in step 864. The prioritized next reactions may be provided to step 854 via path 892 for inclusion in the list of reactions selected for execution. In step 836, the user interacts with the model to display the model results from step 868 for reactions of interest and for exploration of those reactions. Both steps 836 and 868 can receive user input 822 via path 848, which can include additional data 824 including, for example, data from computer simulations (826), data from other records of chemical reactions (e.g., ELNs) (828), and data from texts (such as scientific literature and patents).
[0056] Basically, as shown in FIG. 8, a user can interact with the model and the data set and can monitor the method through a dedicated user interface, such as a graphical user interface (GUI) 100. In some embodiments, the model or the current model (meaning that the model has been retrained) can be understood as a machine learning model trained with all or part of the chemical data available at a given step of the method. For example, the user can use the obtained model (e.g., at step 836) and the user interface can access the model (at step 868) to predict high-yield conditions for a chemical reaction. The objective of one embodiment of the method is to train a model to provide a prediction of the reaction outcome and predict optimal reaction conditions for a fully specified reaction (see definition) with a satisfactory performance (see text for examples of how performance can be measured), which can be obtained after one or multiple iterations of data collection and model training cycles, as described above.
[0057] In some embodiments, the user may be a computer system that interfaces with the system, for example, by an API. In some embodiments, there may be both human users and computer systems / users. That is, in this description, steps performed by a user may be performed by a human user or a user of a computer system, or both.
[0058] In some embodiments, a target set may be understood to include a set of (fully or partially specified) chemical reactions that are input in one of the steps of the data collection method. The set is typically defined by a user. The set can be explicitly (e.g., a particular class of reactions that form known drug-like compounds that are approved or in clinical trials) or implicitly (e.g., a particular class of reactions that form as products chemical compounds that satisfy the so-called Lipinski rule of five, which imposes constraints on properties such as molecular weight).
[0059] In some embodiments, a model or a current model (meaning a model has been retrained) may be understood as a machine learning model trained on all or part of the chemical data available at a given step of data collection. A model takes as input a partially specified chemical reaction (a reaction can be partially defined in the sense that the input may exclude some of the other information such as necessary conditions and outcomes). If the model receives as input a reaction that only lacks outcome information, the model outputs a reaction outcome. If the model receives as input a partially specified reaction, the model outputs one or more predicted fully specified reactions. The model may provide any number of additional outputs, such as an explanation for the prediction or the model's certainty regarding the prediction.
[0060] In some embodiments, a current dataset is a dataset available at a given step of data collection and can be understood as a dataset suitable for training a current model, which may be assembled from experiments already performed and potentially coupled with other available data sources (e.g., from the published literature or based on quantum computation).
[0061] In some embodiments, a chemical reaction or a reaction can be understood as a fully specified chemical reaction, including all information about reactants and conditions, as well as information about the reaction outcome. In some embodiments, the reaction outcome can be understood as which products are formed and in what yields (percentages), with the exception of a binary indication of whether a given product was formed in a yield above a given value (depending on the context, e.g., provided by the user). In particular, a skilled chemist should be able to carry out a chemical reaction based on this information in the laboratory.
[0062] In some embodiments, an outcomeless chemical reaction may be understood as a fully specified chemical reaction that is only lacking information regarding the outcome.
[0063] In some embodiments, reaction conditions may be understood to be part of a comprehensive description of how to carry out a fully specified chemical reaction, excluding the structure of the reactants whose atoms contribute to the structure of the expected products (called reactants). For example, reaction conditions include all physical variables that affect the reaction, i.e., temperature, pressure, reaction time, stirring strength, order of addition of reagents, rate of addition of reagents, all components of the reaction mixture whose atoms do not contribute to the expected target product (solvent(s), catalyst, base, acid, coupling reagent, etc.), quality of reagents used, and ratios of reaction mixture components.
[0064] In some embodiments, a partially specified chemical reaction or a partially specified reaction may be understood as a partially specified reaction in which some information is omitted, which may be any information from a fully specified chemical reaction, in particular, conditions such as solvents or catalysts may be missing.
[0065] In some embodiments, the result of a chemical reaction can be seen as the products that are formed, as well as what yield each product is formed in for the chemical reaction.
[0066] In some embodiments, a computer system is disclosed for predicting chemical reaction conditions that meet a user's requirements (e.g., are high-yielding based on a partially specified chemical reaction) and / or for predicting the outcome of a given chemical reaction designed to obtain with high accuracy a chemical reaction specified (implicitly or explicitly, as illustrative examples) by a user (hereinafter referred to as a target set) by a combination of 1) and, optionally, 2).
[0067] 1) A machine learning model trained on a diverse set of experimental results for chemical reactions combining:
[0068] (1)(i) Experiments designed and performed with a methodology that combines one or more of the following characteristics: chemical reactions are selected with the end goal of enabling a trained model to perform with high accuracy on a target set (see Sections 2.2 and 5 for details on the method for prioritizing chemical reactions); are performed in a high-throughput manner utilizing automated liquid and / or solid dispensers (e.g., by selecting appropriate methods or experimental hardware; see Sections 3 and 5 for details); the reaction results (which products are formed and in what yields) are analyzed (e.g., performed by an LCMS instrument with a UV / Vis detector) and quantified (e.g., processed by off-the-shelf software to determine the yield of the reaction); or involve decisions by one or more users (see Sections 2.4 and 5 for details on what decisions users make in particular); or
[0069] (1)(ii) Where appropriate, other sources of chemical information containing chemical reactions extracted by any method from textual information such as disclosed patents or scientific literature, or other datasets of chemical reactions, including, but not limited to, datasets obtained from written records of experiments after they have been performed (such as electronic lab notebooks); or
[0070] (1)(iii) Where appropriate, the results of computer programs intended to simulate the outcome of chemical reactions using molecular modeling at various levels of theory (level of theory is a term used in computational chemistry, where a higher level of theory means that the computer program achieves greater accuracy of the simulation, at the expense of longer processing times).
[0071] Additionally, machine learning models can be trained using techniques that enable more accurate predictions on the target set, as described below in Section 2.4.
[0072] 2) Optionally, a computer program with a graphical user interface that allows a user to accomplish a desired objective. Embodiments of such a computer program include (i) and (ii):
[0073] (2)(i) A computer program that uses any known algorithm to plan a synthetic route consistent with user-specified constraints (in embodiments, these may be, but are not limited to, available chemistry hardware, preferred conditions or limitations on price or availability of reactants to purchase, etc.). In embodiments, the computer program may use machine learning models to predict and suggest reaction conditions (see also Section 5.3). In one embodiment, the synthetic steps along the computer-designed route may be performed autonomously as part of one or more laboratory instruments. In embodiments, the system may communicate with the laboratory instruments via an application programming interface. In other embodiments, the results of the synthetic route planning may be summarized into a single number indicating the expected cost of synthesizing the compound that is shown to the user.
[0074] (2)(ii) A computer program (see FIG. 5 and FIG. 16) that allows a user to find reaction conditions that satisfy a user's constraints (e.g., high yield) for a user-specified chemical reaction via a graphical interface where the user can receive predicted optima created using a machine learning model. In one such embodiment, the user can effectively interact with the model predictions (e.g., explore or modify the model predictions). In one such embodiment, the program allows the user to input only one of the substrates (rather than the complete reaction) and the program uses the machine learning model to suggest reaction conditions for the synthesis of multiple potential products. For more information, see section "5.2 Prediction of Reaction Results and Optima for Selected Reactions." With regard to the user's ability to select constraints for the predictions, this functionality allows the user to define what is believed to be the optimum through the selection of constraints.
[0075] Additionally, some embodiments provide synthesis plans and potential syntheses of compound collections. As a specific example, a DNA-encoded library can be synthesized based on the predictions made by the system. Also, a virtual catalog of compounds can be created based on the system's predictions.
[0076] Section 2. Data Collection Methods
[0077] Section 2.1
[0078] In an embodiment, the data collection method involves one or more iterations, each iteration including the execution of one or more of the following steps: 1) selecting a batch of chemical reactions to be run in the lab based in part on their similarity to a target set (see Section 2.2 and Section 5 for a description of how to select the batch), 2) running the selected batch of chemical reactions in the lab and analyzing the reaction mixture by an appropriate analytical method with the aim of quantifying the reaction yield (proportion of product formed) (this step may require purchasing chemicals on the market), 3) estimating the reaction outcome after the run using software processing of the analytical data, 4) obtaining a new model by training on the current dataset (which may include all or a subset of the reactions run in the previous step), 5) analyzing the performance of the model according to a number of metrics (which may be displayed in a GUI, as defined later in the specification) and deciding to continue data collection (see Section 2.4 for details), 6) interacting with the model and the dataset in a user interface (graphical or non-graphical). In particular, this interaction involves inputting partially specified chemical reactions into the model and reading / displaying the outputs (predicted results of the chemical reactions and additional information such as reaction conditions) (see Section 5 for this step).
[0079] Section 2.2
[0080] In one embodiment, the target set consists of any set of responses that are relevant to a given application of the final model.
[0081] In one embodiment, the reactions in the target set are any reactions whose products are or have been in clinical trials, or have been identified as potent binders / inhibitors of relevant biological targets. In one such embodiment, only reactions of a given type (e.g., amide coupling) are included in the target set.
[0082] In one embodiment, machine learning models and / or heuristic algorithms and / or user testing are used to reduce the size of the target set by prioritizing certain chemical reactions to achieve one or more objectives, such as reducing the chemical similarity of reactions in the target set to other reactions in the target set, adequate coverage of a particular user-defined chemical space with a minimum size target set, etc.
[0083] With regard to the selection of chemical reactions to run, in one embodiment, the reactions to run, analyze, and add to the current data set at a given step of data collection are selected from the space of all possible chemical reactions as the set of best scoring reactions according to the mathematical formula of Equation (1) below: S=argmax_{S,|S|=N}f(S), equation (1) where f(S) is a scoring function that assigns a score to a set of reactions, where S is the set of N reactions that are part of the batch. In the remainder of the specification, f will be referred to as the reaction prioritization function.
[0084] In embodiments, the function f(S) may include a combination of one or more factors defined below: a) the number of reactions that are part of the target set that are chemically similar (defined below) to reactions in set S; b) the number of reactions that are (1) chemically similar to reactions in set S and (2) that are assigned a low certainty by the current model; c) the number of reactions that are part of the target set that are (1) similar to reaction set S and (2) that are considered unlikely (below a predefined score according to a predefined scale) by one or more users (in some embodiments, the user may indicate each reaction in a GUI (see Section 2.4)); (in some embodiments, a machine learning model may be used to approximate expert opinion). ) the price of the reagents needed to perform reaction S; e) the time the reagents arrived at the lab from the compound provider from the order date (in particular, whether the reagents have already been purchased); f) the certainty of the reactions in set S as assigned by the model; g) the similarity of the reactions to the dataset (used to train the model) (e.g., in the form of Euclidean distance or variance of the model predictions); h) the chemical similarity (defined below) of reaction set S to itself (in one embodiment, the formula takes as input the distribution of chemical similarities between the reactions in set S. In general, lower similarity is desired for the chemical reactions in set S, as it often indicates that the chemical reactions cover a wider subset of the chemical space (and therefore may cover a wider subset of the target set).in other words, they should be similar to the target, yet different from each other); i) an estimate of the model's uncertainty (in one embodiment, the uncertainty is based on the average of different copies of the model, each retrained on the same dataset); j) the type of chemical reaction (in one embodiment, only reactions of a given chemical type are selected (e.g., only amide coupling reactions)); k) a score assigned by one or more users reflecting their opinion of whether the reaction is relevant for improved performance of the model on the target set (in some embodiments, the user(s) are shown the chemical reactions in a GUI (see section 2.4)).
[0085] In one embodiment, the chemical similarity between reactions used in determining the set S is based on numerical representations of the reactants (substrates, products) and reagents in each reaction. In one embodiment, the numerical representations are calculated using any published method of representing chemical compounds, such as MACCS or Morgan Fingerprints (collectively referred to as chemical fingerprints). In one embodiment, the chemical similarity function is based on a chemical fingerprint calculated for a molecule in which atoms farther than a certain distance from the center of the reaction have been removed, where the center of the reaction is defined as the atom affected during the chemical reaction. In one embodiment, the numerical representations are calculated by inputting the chemical reaction into a model and saving a hidden representation of the chemical reaction. In some embodiments, the numerical representations can be used to calculate the chemical similarity using a measure of similarity between two sets of numerical representations, such as Euclidean distance or Jaccard coefficient.
[0086] With respect to any step of collecting data, in one embodiment, the data collection can include purchasing a number of reagents to be readily available for carrying out reactions thereon, and these reactions will naturally be prioritized for execution, particularly if the reaction prioritization function f(S) includes a time availability factor (f).
[0087] In one embodiment, the set of reagents R can be prioritized by finding the set of reagents R that maximizes the mathematical expression in Equation (2) below: R=argmax_{R,|S|=N}g(R), equation (2) where g(R) is a scoring function that assigns a score to the set of reagents R, where R is the set of N reactions that are part of the batch. In one embodiment, the function f(R) includes one or more factors including: a) a factor of any of the following forms: V denotes the set of reactions that can potentially be performed given the set of reagents R (the factor is f(V), where f is a reaction prioritization function); b) the price of the reagents from the chemical compound provider; c) the time the reagents arrive at the lab from the chemical compound provider from the order date (in particular, whether the reagents have already been purchased, in which case time t=0); d) the chemical similarity of the set of reagents R to itself.
[0088] In one embodiment, the set of reagents R or the set of reactions S are selected according to the following iterative optimization algorithm, which aims to find an approximate solution to the optimization problem posed in (2): In a first step, each reagent or reaction is scored individually for f(S) (where g(R) can be decomposed as f(S)=\sum f(s_i)(g(R)=\sum g(r_i)). In a second step, the reagent or reactions with the highest scores are selected. The first and second steps are repeated until a desired number (N) of reagents or reactions is selected. The desired number N is a parameter set by the user of the method and can be different in different steps of the method.
[0089] In one embodiment, the solution to equations (1) and (2) can be solved by using off-the-shelf software for discrete optimization.
[0090] In one embodiment, one or more users can display different sets of reagents R for reaction S in a graphical user interface and prompt for additional input that can be used as part of the scoring function f(S) or g(R).
[0091] Section 2.3
[0092] Some embodiments include using data from other sources. In one embodiment, the current dataset includes chemical reactions extracted from textual information such as journal articles and patents. In one such embodiment, the extraction can be done automatically using a machine learning model trained to automatically extract chemical information from textual data. In one such embodiment, the machine learning model is first trained in a self-supervised manner (using a common pseudo task such as predicting the next word in a sequence) on the text to be used to extract the chemical information. In one such embodiment, a Transformer architecture is used to perform the extraction. In one such embodiment, the following computational pipeline is used to extract information from the textual data: (i) predict (using a Transformer architecture) whether a fragment of text describes a chemical reaction; (ii) label paragraphs in the description of the chemical reaction as headers, descriptions, or footers (using a Transformer architecture); and (iii) predict entities (using a Transformer architecture) from the reaction description, such as reaction products, substrates, solvents, catalysts, and other conditions required to carry out the reaction.
[0093] In one embodiment, the current dataset includes auxiliary information sources that are not directly related to predicting the outcome of the chemical reaction. In one such embodiment, a dataset of molecular properties (of any type and calculated using any means, such as quantum chemical calculations) is combined with the dataset. In one such embodiment, the dataset including the auxiliary information sources is selected based on both similarity to the target set and any type of chemical similarity to the reactions in the current dataset.
[0094] In one embodiment, a quantum computer or quantum chemical calculation program can be run instead of or in parallel with high-throughput experiments to predict (approximate) outcomes of any chemical reaction in order to enrich the dataset with additional data. In one such embodiment, the same procedure used to select reactions to be run in the laboratory can be used to select reactions to be predicted using a quantum computer or quantum chemical calculation program.
[0095] In an embodiment, a method is disclosed for creating more accurate machine learning models for predicting properties of compounds and reactions based on quantum computation. The embodiment is based on the premise that (a) accurate simulations of simplified quantum systems (the term "simple" can refer to both simplification of the reaction mechanism and / or reagents involved in the chemical reaction) can be obtained with a reasonable amount of computation, and (b) machine learning models may greatly benefit from access to simplified quantum system data when predicting more complex data, such as experimental reaction result data or simulation results of more complex quantum systems. In some embodiments of this kind, one or more computational pipelines (as described in the next paragraph) are established to calculate the results of chemical reactions of a wide range of different substrates and products, which are then used (as one of the elements) to train the machine learning model. In one embodiment of this kind, multiple quantum chemistry computational pipelines are established, each intended to contain specific information about one aspect of the chemical reaction. By using the pipelines to calculate the results of the chemical reactions, the dataset can be enriched with relatively accurately simulated reaction results. This technique is particularly useful for chemical compounds that are not well covered in experimental datasets.
[0096] The quantum computing pipeline can be parameterized in a variety of ways, including (i) the algorithm used to calculate the energy of the molecule (e.g., GFN-xTB), (ii) the transition states of the reactions being simulated, and (iii) the parameters of the algorithm used to calculate the energy of the molecule (e.g., the number of molecular orbitals to use, the error tolerance of the algorithm, etc.).
[0097] In one such embodiment, a quantum chemistry computational pipeline is established by searching through various possible parameterizations to achieve significant correlation with experimental data for a given subset of chemical compounds (e.g., smaller compounds). In one embodiment, the data set includes chemical reactions with results that are calculated according to the quantum computational methodology described in the previous paragraph.
[0098] Section 2.4
[0099] Section 2.4.1: Prompt user(s) for input during data collection
[0100] In various embodiments, data collection may include questioning a user(s) with the goal of using the answers to drive the process. In some embodiments, multiple users may be asked the same question and the answers are pooled from all users using a method appropriate for the context (e.g., using maximum voting in the context to decide whether to stop data collection).
[0101] In some embodiments, the user(s) may answer one or more of the following questions (only briefly discussed here and described in detail elsewhere in the specification): a) whether to continue data collection (see Section 2.4.3 for details); b) evaluate on some scale the relevance of a chemical reaction or set of chemical reactions that will improve the performance of the target set (see Section 2.2 for details); c) specify various parameters for data collection, which may include, but are not limited to, (i) how many reactions to prioritize, (ii) what the target set is, and (iii) what function to use during reaction prioritization. (See Section 2.2). Specific details of the parameters are provided in the relevant sections.
[0102] In some embodiments, a machine learning model can be used to predict a user's answer by training an answer prediction model on a dataset of answers collected in previous iterations or runs of data collection. In some embodiments, the answer prediction model can be based on a Transformer architecture with input consisting of a token string representing the relevant context of the question (e.g., the history of the data collection methodology) and output being a token string representing an answer (e.g., a "0" or a "1" indicating whether to stop or continue data collection).
[0103] Section 2.4.2: Supporting the user(s) in answering questions
[0104] In some embodiments, the user(s) may use a dedicated user interface (such as a graphical user interface (GUI)) to help them more accurately answer questions posed during data collection, such as deciding whether to continue data collection. See section 2.4.1 for details on what questions may arise. The following text describes the functionality of the GUI.
[0105] In some embodiments, the GUI has functionality that can be used for any use case, as described in Section 5. In particular, although less formally, the GUI can support querying the underlying dataset, asking the model for predictions, and displaying user-interpretable explanations for model predictions (e.g., scientific justification, see Section 5 for more details).
[0106] In some embodiments, the performance of each of the current models is summarized (according to one or more metrics described in Section 2.4.3) and displayed in a GUI.
[0107] In some embodiments, a user can use the GUI to run queries against the database to help answer questions at any step of data collection. Queries can be specified by a variety of means, such as (a) the presence or absence of a given chemical structure, (b) similarity to a given chemical reaction, (c) the presence or absence or value of a given chemical property (e.g., lipophilicity or acidity). Running queries can provide a deeper understanding of chemical reactivity, which can lead to better selections. In one such embodiment, the GUI can be used to help answer questions regarding the likelihood that a model will accurately predict the yield of a chemical reaction. In one embodiment, the graphical interface used to explore the dataset disclosed as part of the use case (see Section 5) can also be used in the data collection process for purposes of querying the dataset.
[0108] Section 2.4.3: Evaluation and Decision to Stop Data Collection
[0109] In some embodiments, the data collection process includes a step of deciding whether or not to continue the process. In some embodiments, the decision is made partially or completely by the user. In some embodiments, the decision is made completely autonomously.
[0110] In some embodiments, the decision is based on an evaluation of the model's performance (e.g., its ability to predict the outcome of a chemical reaction) by summarizing it into one or more metrics. In some embodiments, the metrics may include (i) the model's accuracy in distinguishing between high-yielding and low-yielding reactions (the user can determine a percentage threshold that distinguishes between high-yielding and low-yielding reactions); (ii) the model's accuracy in predicting a fully-specified reaction based on partially-specified reactions; and (iii) a correlation between the predicted yield (yield is part of the reaction outcome) and the actual yield of a selected product (e.g., the product with the highest yield). In this type of embodiment, where the decision is entirely autonomous, the user specifies a set of logical constraints based on the user metrics that will terminate the process of data collection when desired (either the user is asked to stop data collection or the computer system driving the data collection ensures that the constraints are met and stops data collection).
[0111] In some embodiments, the metrics can be calculated on reactions (hereafter referred to as an "evaluation set of chemical reactions") from one or more of the following sources: (a) the target set; (b) reactions after the run that were not used to train the model; (c) another set of reactions that are specifically selected and run for the purpose of evaluating performance by the user or independently using any method (e.g., this set can include a set of particularly difficult reactions that were not included in the target set). In one such embodiment, the reagents that are part of the evaluation set are more chemically complex (e.g., measured by the number of different chemical structures from a predefined list) than the reagents included in the experiment. In another such embodiment, the reagents in the evaluation set can be selected in a manner that selects the most similar reactions to a known drug or compound in clinical trial in terms of chemical similarity (as described above) between the reactants of the reactions to the known drug or compound in clinical trial.
[0112] Section 2.5
[0113] Section 2.5.1: Model Characteristics
[0114] In some embodiments, the model can be trained to have one or more of the following properties (how the model is trained depends on the exact type of model used; see section 2.5.2 for details): a) it can predict a reaction outcome based on input consisting of a fully specified chemical reaction (in which case the model masks the specified outcome) or a partially specified chemical reaction; b) it can predict one or more chemical reactions based on a fully specified or partially specified chemical reaction, which may contain additional information (e.g., information about the conditions used). A fully specified chemical reaction is one that contains all the information about the reactants and conditions, as well as information about the reaction outcome (what products are produced and in what yields (percentages)). In particular, a skilled chemist should be able to carry out a chemical reaction based on this information in the laboratory. Partially specified chemical reaction or partially specified reaction: A partially specified reaction omits some of the information from a fully specified chemical reaction. In particular, information about conditions such as solvents and catalysts may be missing. A specific example of (b) is predicting high-yield reaction conditions based on input consisting of substrates and products.
[0115] In some embodiments, the model output can include uncertainty about its predictions (details on how to calculate them are in the next paragraph). The model uncertainty can be used to improve the accuracy of predictions at each stage of data collection where the model is used. For example, the model can be used for: (a) reaction prioritization during data collection (see Section 2.2), or (b) additional output that is displayed in the GUI in every use case (such as displaying predicted reaction conditions).
[0116] In some embodiments, the model input may include a set of user requirements for the generated chemical reactions along with additional information such as conditions, etc. In some embodiments, the user requirements include one or more of the following types: (a) The reaction has the highest possible yield. (b) The reaction conditions meet certain constraints, such as the use of low temperatures. (c) The substrates and products satisfy certain logical constraints, such as being below a given purchase price.
[0117] In some embodiments, an ensemble of models can be trained, where a model can be understood as being composed of multiple variants of a model. In some embodiments, individual variants of a model can be obtained by repeating the training procedure, but changing the configuration in a less or more significant way, such as changing the order of training examples displayed during training, or using different parameters of the training procedure (such as the length of training). In these embodiments, when the ensemble model is asked to produce an output, it is input to each variant and the outputs are pooled according to any method, such as averaging the outputs (if the averaging is well-defined) or voting (if the output is categorical).
[0118] In some embodiments, the model output may further include outputs aimed at increasing the interpretability of the model by the user. In some embodiments, the model may have a list of chemical reactions (fully or partially specified) from the dataset, with the ability to display the results of reactions already performed that are related to the predicted reactions, thus helping the user form an opinion on whether the model output is correct (whether the model output matches the experiment). In some embodiments, the model may have a user-interpretable explanation of why it made the prediction, such as: (a) outputting a prediction on physicochemical properties related to the prediction (e.g., solubility of the product in water); (b) including a prediction of the reaction mechanism (e.g., showing the critical transition state along with the predicted energy of the transition state). In some embodiments, the user may be asked questions (e.g., via a GUI; see section 2.5) regarding how convincing or useful the provided explanation is. In some embodiments of this kind, the model may be trained on a dataset enriched with information on the provided information on how convincing or useful the provided explanation is.
[0119] In some embodiments, the uncertainty can be calculated as a confidence interval indicating the likelihood of maximum and minimum values of the predicted quantity. In one embodiment, the mean (for classification output, such as predicting whether the yield will exceed some user-defined threshold) or the variance of the predictions (for regression output, such as predicting the yield) made by the members of the ensemble are used as factors influencing the uncertainty of the model. In another embodiment of this kind, a form of distance between the input and examples from the training set (e.g., the Euclidean distance between the hidden representations obtained from the model, if the model is a neural network) is used to calculate the uncertainty of the model, in the following formula:
[0120] In some embodiments, the model outputs a scalar that quantifies whether the model was trained on similar data, separate from the uncertainty estimate. In one such embodiment, the scalar is calculated by training an ensemble of different models and calculating the variance of the predictions of each member of the ensemble. In some embodiments, the scalar is used to modify the uncertainty estimate, such that if the scalar is low, the uncertainty is increased accordingly. The scalar can be used at any step of data collection where the uncertainty of the model output is used (as specified in the text where appropriate).
[0121] In some embodiments, the model output may include a scalar that estimates how the user judges the associated predicted reaction in terms of how likely the user thinks the reaction is to have a yield above a certain threshold (e.g., on a scale of 1 to 5 where the reaction achieves greater than 5% yield of the desired product). In some such embodiments, the dataset is expanded to include reactions with such assigned scalars, which allows for training a model to predict the scalar.
[0122] Multi-task learning is a broad technique that allows a given machine learning model to be trained for many tasks, such as both predicting what objects are in an image and predicting where objects are in an image. Task-based training can be understood as configuring the training so that the model achieves a given function (task) for a given set of examples. In one embodiment, the model is trained on a dataset using any form of multi-task learning. In one such embodiment, individual weights are assigned to subsets of the dataset. In another such embodiment, the model is first trained on the full dataset and then trained again on subsets of the dataset.
[0123] In one embodiment, the model or training procedure can be modified so that the model predicts reactions that satisfy certain logical conditions, such as temperatures within a certain range. In one embodiment, the model can be trained or fine-tuned (after training on all reactions) on a subset of the dataset consisting of reactions that satisfy the constraints. This property is useful in certain use cases, such as use in the context of predicting and performing synthesis using an automated lab. In an automated lab, it may be possible to use only certain reaction conditions (such as only certain conditions, or only certain temperature ranges). In some embodiments, this can be achieved by adding an additional filtering step added after generating the output from the model that filters out model outputs that do not follow the given logical constraints.
[0124] Achieving high accuracy for reactions with low chemical similarity to reactions in the dataset may be important in some use cases. In particular, drug-like compounds are often chemically complex (e.g., have a high number of atoms) and therefore expensive to purchase. Thus, they may be prioritized less frequently during data collection (see section 2.2). In some embodiments, a model or training procedure may be configured to increase the accuracy of such reactions in the following ways: In some embodiments, an ensemble (as described in the previous paragraph) may be used to create a model. In some embodiments, the training procedure may include contrastive, self-supervised, or semi-supervised learning tasks (which are well-known broad categories of tasks that can be used within a multitask learning framework; see previous paragraph). For example, predicting a removed or masked input portion (e.g., predicting a substrate based on a reaction in which the substrate has been removed or masked). In some embodiments, the training may include predicting one or more properties of a molecule, such as the solubility or boiling point of the molecule (which may be part of one reaction of the dataset or may come from a different data source, such as the CheMBL database). In some embodiments, the training procedure may include tasks or methods aimed at learning representations (in models with internal representations such as neural networks) that causally predict the response outcome. In some embodiments of this kind, any method of causal discovery may be applied to find parts of the representations that predict the response outcome in a more causal way.
[0125] In one embodiment, the model can be trained using a training method that uses meta-information about a given chemical reaction, such as whether the reaction outcome was measured in the lab or simulated using a quantum computing pipeline.
[0126] In some embodiments, the model input may further consist of auxiliary information related to the chemical reaction, such as values of different molecular properties (e.g., the electronegativity of each atom). In some embodiments, these molecular properties can be calculated using quantum simulation software such as ORCA or Schrodinger. In some embodiments, these molecular properties can be predicted using machine learning models trained on datasets that include the molecular properties.
[0127] Section 2.5.2: Implementations based on Transformer or Graph Neural Network architectures
[0128] In one embodiment, the model is based on a sequence-to-sequence deep neural network, such as the Transformer architecture, where both the input and the output comprise a string of tokens, with each token having an assigned chemical meaning. In one such embodiment, the substrates and products of the reaction are encoded in the form of a string (e.g., following the SMILES notation), and the output is encoded in the form of tokens representing the predicted yield and / or reaction conditions. In another such embodiment, the input consists of a reaction where one or more pieces of information of the chemical reaction are missing (e.g., the missing products), and the output consists of the yield and a prediction of the missing information (e.g., what the products should be, etc.). In some embodiments, the model may include as an output a token indicating whether the reaction yield exceeds a certain (user-specified) threshold. A visualization of the input and output representation of the model (for an amide coupling reaction) is shown in Figure 12.
[0129] In one embodiment, the model is based on a graph neural network, a type of neural network that takes as input a graph with vertices (atoms) and edges (chemical bonds), where each vertex and edge has additional properties (e.g., type of atom). The output may be the same as described in the previous paragraph. In some embodiments, the reaction conditions are encoded as properties of additional vertices of the input graph. In some embodiments, each reaction condition is treated as an additional vertex of the input graph.
[0130] FIG. 12 illustrates input and output formats in an embodiment of a GUI 100 for an embodiment that is a model for predicting the outcome and conditions of a chemical reaction. In FIG. 12, the model receives as input 116 the substrates 112, specifically 112a, 112b, the products 114, specifically 114a, and optionally encoded reaction conditions 120 in the form of one-hot encodings 124, 126, 128, 130 or other textual encodings of the molecules. The model outputs 118 the encoded conditions 120 with the predicted class 122 (high and low yields relative to a user-defined yield threshold) as the first token along with the conditions 124, 126, 128, 130 (if not sent as input) as a four-token sequence. In some embodiments, during training, portions of the input 116 may be masked or removed for a portion of the training time.
[0131] Section 3. High-throughput laboratories
[0132] In some embodiments, the model is trained on a dataset that includes reactions performed in a laboratory specifically designed to improve the model's performance on a desired target set.
[0133] In one embodiment, the high throughput laboratory involves the use of medium and high throughput analytical techniques such as MALDI-MS, Echo-MS, MISER chromatography, etc. applied to the analysis of the composition of post-reaction mixtures to determine the amount of product(s) formed in the reaction and the consumption level of starting materials.
[0134] In one embodiment, machine learning models can be used to predict the yield of chemical reactions from raw analytical data. In particular, machine learning models can be used to predict reaction yields based on the output of high-throughput but noisy analytical techniques such as MALDI-MS, Echo-MS, and chromatography in MISER mode. In one embodiment, models can be trained on a dataset of chemical reactions with quantified yields (potentially using lower throughput techniques).
[0135] In another embodiment of this type, machine learning models can be trained and used to determine reaction yields (amount of product) directly from raw analytical data obtained from any analytical device (such as an LCMS instrument), in particular allowing quantification without knowing or measuring the level of analytical signal (i.e., without knowing the molar extinction coefficient) of a known amount of pure compound being analyzed.
[0136] In one embodiment, data obtained from LCMS analysis of the post-reaction mixture can be used to estimate (and recalculate into the yield of the carried out reaction) the amounts of selected components of the post-reaction mixture.
[0137] In one embodiment, an automated solution can be used to generate the reaction mixture and transfer the reaction mixture between different devices. In one such embodiment, laboratory hardware such as an automated liquid handler (e.g., OpentronOT-2) or a 96-channel pipette (e.g., Integra Mini) can be used to automate the generation of the reaction mixture (e.g., by automating pipetting).
[0138] In an embodiment, DNA-encoded libraries (DELs) can be used as a means to generate experimental data (results of chemical reactions) on the reactivity of DNA-tagged reagents, making them suitable for training machine learning models (especially models). In one embodiment of this kind, a library of reagents is used, each tagged with a different DNA tag, having a common functional group. A mixture of such tagged library components (A) can undergo a chemical reaction with a specific reagent(s) (B), so that covalent bonds are formed between some elements of A and some elements of B. The appropriate construction (tagging or immobilization) of reagent B allows for the subsequent cheap and reliable separation and identification of the formed AB adducts. In an embodiment, by using widely known techniques such as polymerase chain reaction (PCR) or next generation sequencing (NGS), reagent B can be attached to large molecules such as DNA strands, proteins (polypeptides), nanoparticles or polymer resin beads, making it possible to wash out the non-reactive library components A and then identify the DNA tags of library components A that undergo a reaction with reagent B.
[0139] In another embodiment of the DEL used as a means of generating data, the DEL can be exposed to a molecule that is a binding target for the fragment of interest attached to some molecules in the DEL. The target molecule can be a protein or a small molecule and can be covalently attached to a solid support material. Molecules that do not bind to the target are washed out. In one embodiment, the remaining molecules that are part of the DEL can be identified using commonly known techniques such as polymerase chain reaction (PCR) or next generation sequencing (NGS).
[0140] Section 4. Data Collection Method Implementation
[0141] FIG. 13 is a flow chart illustrating an embodiment of a data collection method 140 for a model to predict the outcome and conditions of a chemical reaction. In FIG. 13, the process begins with an initial stage 142, which includes steps 144-150. Step 144 is the selection of a target set. In an embodiment, the target set may be configured to consist of reactions that include multiple published drug-like compounds whose targets have been identified as strong binders or inhibitors of a recognized biological target, or that are in or after clinical trials. In an embodiment, the target set may be based on any subset of such compounds. In step 146, a large single batch of reagents is purchased or accessed based on their similarity to the reagents in the target set and their chemical similarity to other reagents in the target set. In step 148, a large number of randomly selected reactions are run that include the reagent ordered in the previous step (because there are so many potential reactions, it is usually not practical to run all reactions that include the purchased reagent). In one embodiment, the initial stage ends with step 150, where the user inspects the performance of the current model on a different set of compounds (including, but not limited to, drug-like compounds that are not part of the reagents mentioned in 1.b above). Based on these tests, the user or users decide whether the data collection method should continue with another iteration 152 of the initial phase 142, rather than 154. If not, the next phase 156 is entered. At each (e.g., biweekly) iteration of phase 156, steps 158-164 are repeated. At step 158, the number of outputs of the current model (with inputs that are reactions for which reagents were purchased in the previous step) that can be used to prioritize reactions is calculated (see later in the text). At step 160, if desired, the user is queried in the GUI as to which reactions (of the reactions for which reagents were purchased in the previous step) should be prioritized. At step 162, after determining the final set of prioritized reactions, the prioritized reactions are run in the laboratory and quantified (i.e., reaction yields are calculated based on reaction mixture analysis).In step 164, the current model is retrained with a current data set that at least partially includes the data generated in step 162, and the user inspects the performance of the current model to determine whether to continue data collection (160 instead of 166). The inspection may include using the GUI to inspect the accuracy of the model against various sets of compounds (see text for evaluation sets of reactions). If not (166), the model may continue to be used (170) as described in any of the various embodiments.
[0142] In one embodiment, the compounds of step 146 can be purchased from an external provider of chemicals, such as MolPort, according to a function g(R), where R is a set of reagents with the following properties:
[0143] g(R) can be decomposed as g(R) = \sum_{i=1}^Ng(R_i), where R_i is a single reagent to be purchased and N is the number of reagents desired.
[0144] If the compound (R_i) price exceeds a predefined threshold or the time to arrival exceeds a predefined threshold, g(R_i) is set to minus infinity. Otherwise, g(R_i) is set to the number of reactions from the target set such that the similarity between one of the substrates and the reagent exceeds a user-defined threshold.
[0145] In another such embodiment, g(R_i) may further include a term indicating one or more user answers regarding how much a response with respect to reagent R_i would improve the performance of the model resulting in the target set (e.g., on a scale of 1 to 10). As described in Section 2.4, in some embodiments, the user may have access to a GUI when answering the questions.
[0146] In some such embodiments, the function g(R) is optimized using the iterative optimization algorithm previously disclosed (Section "2. Data Collection Methods").
[0147] In some embodiments of this type, the decision as to whether to continue data collection is based on the performance of the model predicting response outcomes from the target set, which can be displayed in the GUI (see Section 2.4).
[0148] In the above embodiments, the reactions may be carried out in a high-throughput chemistry laboratory optimized to achieve high throughput (number of reactions performed and analyzed per unit time) and low operating costs.
[0149] In one such embodiment, the reaction mixtures are prepared in separate wells of a standard size multi-well plate.
[0150] In one such embodiment, all solutions of all reactants are prepared and stored in separate wells of a standard size multi-well plate to serve as stock solutions for preparing reaction mixtures.
[0151] In one such embodiment, an automated liquid handler equipped with a single or eight channel pipette, such as an Opentrons OT-2, can be used for one or more steps of preparing the reaction mixture.
[0152] In one such embodiment, a 96-channel pipette (eg, Integra Mini-96) or a 384-channel pipette can be used for one or more steps of preparing the reaction mixture or post-reaction analytical sample.
[0153] In one embodiment of this kind, the multiwell plate containing the reaction mixture is sealed with an adhesive polymer or metal cover, or with a silicone or rubber mat. The sealing mat can be held in place by placing the plate with the sealing mat between two rigid panels (one under the plate and one above the mat) and compressing the panels with screws or the like. The reaction mixtures in the wells of the plate can be stirred by shaking the plate in an orbital (thermo)shaker, or by magnetic stir bars placed in each well and forced to move by varying a magnetic field generated by an external device. The reaction mixtures in the wells of the plate can be heated or cooled by placing the multiwell plate in a thermoshaker or a heating / cooling block.
[0154] In one such embodiment, known amounts of one or more chemically inert compounds are added to selected or all of the post-reaction mixtures to act as internal standards to aid in quantification of the post-reaction mixtures, hi one embodiment, different internal standards or mixtures thereof are added to selected subsets of the post-reaction mixtures.
[0155] In one such embodiment, multi-well filtration plates containing either an inert membrane or a stationary phase capable of selectively absorbing selected components of the post-reaction mixture can be used in one or more steps of preparing the post-reaction analytical sample.
[0156] In embodiments, the post-reaction mixture can be analyzed and quantified using off-the-shelf equipment such as high pressure liquid chromatography (HPLC) coupled with one or more detectors including a single or multi-wavelength UV-Vis detector, a fluorescence detector, an evaporative light scattering detector (ELSD), a charged aerosol detector (CAD), a radiation detector, an electrochemical detector, a chemiluminescence nitrogen detector, or a mass spectrometer.
[0157] In one such embodiment, pre- or post-column derivatization is applied to the analysis of all or selected analytical samples, and different methods of pre- or post-column derivatization can be applied to different subsets of analytical samples.
[0158] In one such embodiment, the reacted analytical sample is analyzed by MALDI-MS or Echo-MS analysis.
[0159] In one such embodiment, an aliquot of the post-reaction mixture is subjected to liquid chromatography and fractions containing the separated product of satisfactory purity are collected using a manual or automated fraction collector. The amount of product in the collected fractions is measured by weighing the solid residue after evaporation of the eluent(s). In one such embodiment, the eluent stream coming out of the column is split with a known split ratio between a fraction collector and a sample destruction detector such as an MS or ELSD. In one such embodiment, a quartz crystal microbalance is used to assess the mass of the solid residue.
[0160] In one such embodiment, the post-reaction analytical sample is analyzed by nuclear magnetic resonance (NMR) spectroscopy. In one such embodiment, the reaction is carried out in a deuterated solvent or mixtures thereof, and the products are quantified by NMR in the untreated or treated reaction mixture.
[0161] In one such embodiment, the running of a selected batch of chemical reactions (including preparation of analytical samples) is supported by dedicated software. The software can use as input such a batch of chemical reactions to be run and perform any combination of the actions listed below: a) To optimize the process of dispensing reagents into wells (containers), the reaction batch is divided into subsets, with each subset performed in wells on one plate (or in a group of containers in one rack). b) To optimize the process of dispensing reagents into the wells (containers), each reaction is assigned a specific location of a well on the plate (or a container in a rack). c) Providing the human laboratory operator(s) with a detailed list of the steps required to run a batch of reactions. d) Monitoring the execution of an experimental protocol by monitoring in an interactive manner the successive steps performed by human(s) and / or laboratory hardware. e) Creating a command set for one or more automated laboratory instruments. f) Prepare a detailed report on the procedures used to carry out the reaction, based inter alia on the monitoring and recording of temperature, pressure, humidity, oxygen (or other relevant gas) concentration(s), stirring time and intensity, and start and end times of each procedure. g) Generating output for use by the analytical device(s) and / or software used to process the raw analytical data.
[0162] In one such embodiment, during step 164 of deciding whether to continue data collection, the accuracy of predictions made by the current model is evaluated based on a combination of one or more of the following sets of reactions with known outcomes: (i) a random subset of reactions for reagents purchased in the previous step, (ii) a random subset of reactions for reagents from a smaller set predetermined at the start of data collection, and (iii) a large number of reactions that form drug-like compounds. The results of such evaluation can be displayed in the GUI. Furthermore, since the evaluation process ensures that the model is meaningfully tested in a setting that feeds previously unseen reactions as input, no reactions from these three sets can be used to train the model during data collection. In the above embodiment, the user is shown the calculated accuracy in the GUI and is asked to decide whether to continue the data collection process at the end of each stage.
[0163] In such an embodiment, the current dataset includes some reactions that have been performed so far during the data collection stage. In another such embodiment, the current dataset can be combined with reactions extracted from patent publications and published patent applications.
[0164] In such an embodiment, the current model may be based on a Transformer architecture (as disclosed in the previous part of this disclosure).
[0165] In this type of embodiment, reaction recommendations may be prioritized during data collection according to three variants of a prioritization function f(S): (i) f(S) is a random number resulting in a random selection of viable reactions using purchased reagents, (ii) f(S) is a weighted sum of a measure of model uncertainty for set S and a measure of chemical similarity for set S, such that the most chemically diverse and uncertain reactions are selected, and (iii) in addition to the construction of f(S) as described in (ii), the function includes an additional factor that measures the chemical similarity between the products of reaction set S and the products of the target set. In embodiments, reaction recommendations can be inspected by one or more users using the GUI to narrow down set S to fewer reactions.
[0166] Section 5. Use Case Examples
[0167] In some use cases, a GUI or UI greatly enhances a user's ability to access a model and thereby achieve a desired goal. However, while this disclosure refers to a particular GUI as illustrated by certain screenshots, it should be understood that other user interfaces (UIs) can be described with reference to a GUI and used in the disclosed embodiments for the interface between a user and a model or between multiple models. Furthermore, since a user may be a computer system, it should be understood that actions described in terms of a GUI or other UI result from an interface of the computer system, such as an API.
[0168] Section 5.1 Use Case Characteristics
[0169] In some embodiments, the GUI or UI of the use case may include an option to run a given chemical reaction in an automated or semi-automated lab. In one such embodiment, this option allows a user to verify or test the model predictions. In an embodiment, the use case may require an application programming interface (API) to communicate with the lab hardware. In some such embodiments, the results of the run experiment are shown to the user in the UI. In some such embodiments, the results of the run experiment may be added to the dataset, for example, in step 160.
[0170] In some embodiments, the reactants and products of the reaction, along with the reaction conditions predicted by the model, are sent via the API to another software that uses the input to generate a user-readable protocol of the synthesis and / or an automated laboratory hardware executable protocol. In some embodiments, such a protocol includes one or more series of steps, each step capable of being performed using a laboratory device. In some embodiments, instructions for performing such steps are provided to the associated laboratory device via the API. Examples of laboratory hardware that can be directed by a protocol include automated liquid dispensers, automated solid dispensers, multi-channel pipettes, reagent dispensers, robotic arms with grippers to move plates, vessels, or racks of vessels, plate sealers, vessel capping / decapping devices, gas / vacuum valves, magnetic stirrers, orbital shakers, cooling / heating devices, centrifuges, evaporators, filtration devices, buffer exchange devices, magnetic modules (for magnetic bead-based chemistry), peristaltic pumps, syringe pumps, vacuum pumps, gas generators, gas compressors, conveyor belts, rail-based plate (or vessel, or rack) movers, car-like plate (or vessel, or rack) movers (including partially autonomous instruments, e.g., Formulatrix's ROVER). The user can monitor and influence the generated protocol(s), for example, by excluding certain reactions from the generated protocol.
[0171] In some embodiments, model predictions in a use case are only displayed if the confidence of such predictions exceeds a certain threshold (per model output, see section 2.5 for models), with the end goal of selecting only the most confident predictions. The invention is particularly useful in the context of the entire system and its potential applications. While the user goal may already be met by narrowing the model to only a subset of the chemical space, narrowing this subspace to only those reactions above a confidence threshold can further significantly increase the confidence of the model.
[0172] In some embodiments, the use case graphical interface may include a GUI that displays reactions from the dataset and allows a user to perform potentially complex queries on them to find reactions that are relevant to the user. In one such embodiment, the user may interact through the GUI and construct a query that reflects what reactions should be retrieved from the dataset. In some embodiments, the query may be (potentially) defined as a nested structure of logical constraints, such as whether a given chemical structure is present in any of the substrates. Figure 14 illustrates one such possible embodiment.
[0173] In some embodiments, the model predictions can be shown along with a selected scientific rationale (see text below for scientific rationale). In one embodiment, the model predictions can be shown along with responses from the dataset (called reference responses) along with a short text explanation explaining why a given example is relevant to a given model prediction. An example of how this embodiment can be implemented in a GUI is shown in FIG. 15.
[0174] In some embodiments, model predictions are presented along with additional outputs designed to enhance the interpretability of the model. See Section 2.5 for more details on the format of such explanations. Some examples include displaying scientific justification or a list of relevant reference reactions from the dataset. In some embodiments of this kind, users may be asked to provide feedback on how useful or persuasive the explanations provided are to them.
[0175] In an embodiment, any graphical interface may provide a user viewing the output of a machine learning model for a given chemical reaction with a user-readable explanation (called the scientific rationale), such as a chemist explaining why a given chemical reaction is or is not possible.
[0176] In one such embodiment, if one or more examples from a dataset (e.g., a training dataset used to train a given model) meet given criteria (defined manually or automatically), they are displayed to the user along with a textual description of such criteria. In the context of a machine learning model predicting the outcome of a chemical reaction, examples of such criteria include: (a) chemical reactions with similarity higher than a defined threshold, (b) chemical reactions with the same user-interpretable chemical features, such as steric hindrance or electron density distribution, and (c) chemical reactions with similar estimated or measured energy barrier magnitudes (similar activation energies).
[0177] In one such embodiment, the scientific justification can be based on a summary of the model's performance on a given set of compounds (e.g., a target set). For example, an expert can be presented with a description of the type: "This model achieves 80% accuracy in predicting high-yield reactions of heterocycles of the type shown in the diagram." In some embodiments, the GUI includes a display of scientific evidence (as defined in the previous paragraph) in the manner described in the previous paragraph.
[0178] Section 5.2. Prediction of reaction outcomes and optimal conditions for selected reactions
[0179] In one embodiment, the model can be used to predict conditions such as solvents, temperatures, catalysts, etc. that will achieve high yields for a chemical reaction or satisfy other user-specified constraints inserted by a user in a graphical user interface or user interface. Figure 16 shows one instantiation of the GUI for this use case.
[0180] In one embodiment, the model can be used to predict product yields in user-specified chemical reactions by inputting the chemical reaction with additional information about the conditions into the model (see Section 2.5).
[0181] In one embodiment, by inputting a fully specified chemical reaction, including yield information, into the model, the model can be used to estimate the probability that a reaction carried out under given conditions will result in a yield of a user-provided product higher than a selected threshold (see Section 2.5).
[0182] Section 5.3. Using the model to plan synthesis with optional control of an autonomous chemistry laboratory
[0183] With regard to designing synthetic pathways and controlling an autonomous chemistry laboratory, in some embodiments, the model can be used to design synthetic pathways that end in a user-specified target molecule. In one such embodiment, the synthetic planning algorithm (such as Retro* or AiZynthFinder) can be modified in a number of ways (examples are given in the next paragraph) so that the model output influences the final designed synthetic plan.
[0184] In one such embodiment, the synthetic planning algorithm may be modified such that the predicted yields and associated confidence of reactions included in the synthetic plan influence the prioritization of the synthetic plan relative to other synthetic plans. In one such embodiment, the Retro* or AiZynthFinder synthetic planning algorithms are used. In one such embodiment, the score assigned to a reaction by the algorithm (during operation) includes one or more factors including the predicted yield of the reaction and the confidence associated with the prediction (see also section 2.5). The output of the synthetic planning can be displayed to the user in the form of a GUI or can be read programmatically via an API.
[0185] In one such embodiment, a synthetic plan can be used to control automated or semi-automated laboratory operations to suggest precise sequences of chemical reactions along with high-yield conditions.
[0186] In one embodiment, the GUI allows an end user to submit a given chemical reaction to be run in an automated or semi-automated laboratory.
[0187] In one embodiment, a separate machine learning model can be used to predict the results of the composite planning system based on the model in order to speed up the composite planning algorithm. In one embodiment, a neural network based on the Transformer architecture is used to predict the final depth of the composite tree predicted by the composite planning software, or to predict other quantities extracted from the output of the composite planning software.
[0188] Section 5.4
[0189] In some embodiments, the model can be used to predict which late functionalization of a given molecule is likely to be successful under what reaction conditions, and the output can be displayed in a GUI or accessed using a non-graphical interface. Late functionalization is a stage in the drug discovery process where a potential drug candidate is optimized by (usually) small changes to its structure. In this type of embodiment, the input to the model is one of substrates, and the output additionally includes (in addition to yield and / or conditions) predicted missing substrate(s) and predicted product(s). In an embodiment, the model is adapted to output likely functionalization chemistry along with certainty estimates and condition information by specifying these requirements as constraints as part of the input to the model (see Section 2.5). In one embodiment, the model predictions can be shown to the user in a GUI, as shown in screenshots within this disclosure. In one embodiment, the model is trained to predict parts of the masked reaction (e.g., during training, the input is the reaction with a masked substrate and the output is the identity of the masked substrate), which allows the model to be used for late functionalization.
[0190] 6. Applications for creating libraries of compounds (physical or virtual)
[0191] In one embodiment, a synthetic plan for the synthesis of a (large) collection of compounds is designed using a model that predicts the reaction conditions for the reactions included in the plan using the methods described in Section 5.3. In one embodiment, the synthetic plan is designed according to the following steps: A user inputs how to synthesize an ensemble, excluding some information, such as reaction conditions for some or all of the steps involved in carrying out a chemical reaction. b. The model is used by the user to predict missing pieces of information, such as a set of step-by-step conditions that additionally satisfy user-specified constraints (as described in Section 2.5 "You can specify constraints on models"), such as that the reaction must be performed at room temperature or that the yield of a desired product must be above a certain threshold.
[0192] A DNA-encoded library (DEL) is a mixture of a large number (up to millions) of compounds in a solution, where each compound is tagged (usually with a DNA strand) and can be identified using inexpensive analytical methods such as DNA sequencing. The envisioned benefit of applying the model in this context is to create DELs that are more diverse (e.g., containing new chemical structures or chemical reactions for a wider range of molecules) or of higher quality (lower proportion of unexpected / unidentified compounds in the mixture).
[0193] DNA-encoded libraries are typically created by performing a series of steps that involve running chemical reactions. In each step, a mixture of hundreds to millions of tagged compounds is reacted with a single reagent under selected conditions. A key challenge in creating DELs is that the chemical reactions must give the desired products in very high yields for all compounds in the mixture, and they must be performed under conditions that are compatible with the DNA tags, typically under relatively mild conditions such as using room temperature for the temperature at which the reactions are run, and those conditions do not destroy the DNA strands that are attached to the compounds in the mixture.
[0194] In one embodiment of the above type, this procedure is used to design a synthesis plan and possibly execute the synthesis of a DEL library. In this type of embodiment, the user may provide constraints such that the recommended conditions satisfy certain constraints related to the synthesis of the DEL library, such as using a temperature that is low enough to maintain the integrity of the DNA tags.
[0195] In any of these embodiments, the synthetic plan can be performed in any laboratory and the collection of compounds is physically available. In some of these embodiments, a human user can modify any part of the plan, for example, by examining the model predictions using a user interface having one or more functions of any of the use case applications.
[0196] In embodiments, synthetic plans for synthesizing large collections of compounds can be generated more automatically using the following steps: (a) The user specifies a list of "starting" compounds. (b) The user specifies a list of constraints that each reaction in the final synthetic plan must satisfy. (c) The user specifies a maximum number of synthesis steps. (d) Every compound is listed such that there is a series of reactions produced by the model that ends with this compound as a product.
[0197] In another embodiment of this type, the predicted response is shown in a GUI with one or more functions of any use case application.
[0198] In one embodiment, a large number of virtual chemical structures can be generated and potentially synthesized by enumerating and applying chemical reactions to commercially available compounds that are predicted to be likely according to the model (predicted to achieve sufficiently high yields with high confidence). In one such embodiment, the model predictions are filtered using the uncertainty-related output of the model to include only the most confident predictions.
[0199] In one embodiment, a user can access a GUI or programmatic API to explore which compounds are part of a collection of compounds.
[0200] FIG. 9 is a screenshot of a portion of the GUI 100 that can be used to enhance user decision making (answering questions, see section 2.4 for details) during the data collection process. The screenshot in FIG. 9 is a basic view of the GUI 100 showing a loaded dataset 101 of target reactions that can be run in the next step "runnable set", a current ML model 102, a loaded dataset 103 of reactions: a graphical interactive view 104 of all reactions from the target set, e.g. 104a-104g, that can be color coded to indicate the certainty of the model predictions, a graphical interactive view 105 of all reactions from the runnable set (the set of chemical reactions that contain reactions that can be run in the laboratory used for data collection), e.g. 104a and 104d coded in red r "low" certainty, 104b coded in green t "high" certainty, and 104g coded in orange o "medium" certainty, with an explanation of the color coding 110. The "Upload" button allows uploading either a target reaction dataset, an actionable reaction dataset, or a model, and after user interaction with the system the "Download" button becomes active.
[0201] From the screenshot of FIG. 9, as shown in the screenshot of FIG. 10, the user can select one or more reactions 104a-104g from the target set 104. Once a reaction is selected, the graphic symbol corresponding to the selected reaction is highlighted (circles around 104a, 104d, 104e, 104g), the selected reactions are displayed as a list 108 (reaction 108a corresponds to 104a, reaction 108b corresponds to 104d, reaction 108c corresponds to 104e, and reaction 108d corresponds to 104g), and the ML model identifies reaction(s) 105a-105j from the feasible set 105. It is known that this may improve the certainty of the prediction of the selected reaction 104a, 104d, 104e, 104g in the scenario where the outcome of the indicated reaction is obtained from the "feasible set" (a set of chemical reactions including reactions that are feasible in the laboratory used for data collection). The number of reactions of the selected subset of the target set 104 that are supported by a given reaction from the feasible set is indicated by numbers 105a-105j and corresponds to the number of reactions from the selected subset of the target set that could improve the prediction of the ML model if the outcome of this reaction from the "feasible set" were known. The list of reactions 109 from the feasible set 105 is represented as reactions 109a-109d with their assigned numbers (circles 105a-105j correspond to one or none of the reactions 109a-109d). The information thus obtained can be exported to a file of a suitable format via download 111b.
[0202] From the screenshot of FIG. 10, as shown in the screenshot of FIG. 11, the user can select one or more reactions 105j-105m from the executable set 105, the selected reactions are highlighted as shown and listed at 109 as reactions 109a-109d. Reactions 104h-104k from the target set 104 that can be supported by the selected reactions 104a, 104d, 104e, and 104g are highlighted and listed at 108 (reaction 108a corresponds to 104h, reaction 108b corresponds to 104i, reaction 108c corresponds to 104j, and reaction 108d corresponds to 104k). The current certainty of the model for the prediction of the highlighted reactions of the target set 104 is displayed by a color code 110. The information thus obtained can be exported to a file of a suitable format using download 111b.
[0203] FIG. 14 is a screenshot of an advanced query builder that can be used by a user in embodiments involving prediction of chemical reaction outcomes and conditions, GUIs used to enhance human decision making in data collection, and other embodiments. The advanced query builder allows the creation of "parent" filters 202, 204, each specifying a logical operator for combining the parent filter with its "children". Child filters 204, 206c, 206d are children of parent filter 202. Child filters 206a, 206b are children of filter 204. The query builder allows the creation of new child filters within the query builder using buttons 208a, 208b. A filter can be both a parent and a child at the same time. Filter 204 is a child of filter 202, but a parent of filters 206a, 206b. All filters and all filter elements, except for the "root" filter 202, can be freely rearranged via drag and drop functionality 201. Each filter can be specified as one or more functional groups selected from a predefined set 206a-206c or custom SMARTS 206d. A filter consists of: I. the functional group name or custom SMARTS 212, II. if applicable, a button 214 that opens a separate graphic and / or text list of functional groups arranged into subsets of similar chemical properties, III. three selection fields that specify the logic (in / out) 216 and position 218, 220 at which a given filter is applied to the reaction, IV. a delete button 222.
[0204] FIG. 15 is a screenshot 300 of an embodiment of the GUI 100 for an embodiment of a model for predicting the outcome and conditions of a chemical reaction, a GUI used to enhance human decision making in data collection, and a depiction of reference reactions that can be used by a user in other embodiments. FIG. 15 shows reference reactions 302a, 302b for a single prediction 116a. Each reference reaction 302a, 302b displays a reaction graph with a list of conditions 304a, 304b, a button 306 to toggle source patent information or a mark 308 if the reaction was performed in a laboratory, and clues 310a, 310b that explain why it was chosen for this particular prediction. The reference reactions can be filtered in two ways: by clues 312 or by custom filters 314, and the filtered result is the intersection of the two.
[0205] FIG. 16 is a screenshot 400 of a reaction editor that can be used by a user in an embodiment of the GUI 100 for an embodiment of a model for predicting the outcome and conditions of a chemical reaction, a GUI used to enhance human decision making in data collection, and other embodiments. In FIG. 16, the reaction editor is in an empty state. The editor allows the user to draw a reaction graph consisting of substrates 112, specifically 112e, 112f, products 114, specifically 114c, and reaction conditions 402. The editor button 120a allows adding atoms and entire structures. A popup 404 includes buttons 406a, 406b for starting a reaction graph from a predefined template. The plausibility of the reaction (plausibility here refers to logical plausibility, not what is predicted by the model, e.g., no carbon is bonded to more than four atoms, which is physically impossible) is checked live during editing, and the status 120a is displayed to the user. The predicted conditions button 408 to the result view (FIG. 15) shows a set of predicted conditions for the specified reaction, along with other outputs of the model, such as yield. The PREDICTION CONDITIONS button 408 is disabled until a valid response is specified.
[0206] FIG. 17 is a flow chart illustrating an embodiment of a method 1700 for predicting the outcome and conditions of a chemical reaction. In the method 1700, in step 1702, a target set of chemical reactions is defined. In step 1704, a first set of chemical reactions is selected based in part on measurements related to the target set. In step 1706, the first set of chemical reactions is executed. In step 1708, an outcome is determined for each executed chemical reaction from the first set. In step 1710, a training data set is assembled that includes at least one determined outcome. In step 1712, a model is built and trained using a computer system, machine learning, and the training data set to predict properties of the chemical reaction and / or suggest reagents or products for completing the partially specified chemical reaction. Additionally, the method 1700 can include steps 1714 to 1718. In step 1714, inputs are provided to the model, the inputs including one or more products, substrates, or conditions. In step 1716, using the inputs and the computer system executing the model, one or more of the following are generated: a predicted outcome of the chemical reaction, a predicted optimal set of reaction conditions, or a suggested reagent or product that will complete a partial chemical reaction. In step 1718, the generated prediction or suggestion is provided to a user.
[0207] Generally, characteristics of a chemical reaction can be understood to include any property or outcome of the reaction, such as reactants, products, reaction conditions, yields, etc.
[0208] In general, in any embodiment, one or more separate GUIs may be implemented, each corresponding to a different function, may be used by different users, and may not communicate with each other. In particular, in embodiments involving use for recommending reaction conditions or operating an automated laboratory, the GUI used for these uses by some users may be separate from the GUI used by users to operate the data collection methods.
[0209] Figure 18 is an exemplary block diagram illustrating one embodiment of a system for implementing embodiments of the methods of the present disclosure, e.g., as described with reference to previous figures. In Figure 18, a computer network 1800 includes a number of computing devices 1810a-b and one or more server systems 1820 connected to a communication network 1860 via a number of communication links 1830. The communication network 1860 provides mechanisms that allow various components of the distributed network 1800 to communicate with each other and exchange information.
[0210] The communications network 1860 itself is comprised of one or more interconnected computer systems and communications links. The communications links 1830 may include wired links, optical links, satellite or other wireless communications links, wave propagation links, or any other mechanism for communicating information. Various communications protocols may be used to facilitate communications between the various systems depicted in FIG. 18. These communications protocols may include TCP / IP, UDP, HTTP protocol, Wireless Application Protocol (WAP), BLUETOOTH, Zigbee, 802.11, 802.15, 6LoWPAN, LiFi, Google Weave, NFC, GSM, CDMA, other mobile data communications protocols, wireless telephony protocols, Internet telephony, IP telephony, digital voice, Voice over Broadband (VoBB), Broadband Telephony, Voice over IP (VoIP), vendor-specific protocols, customized protocols, and the like. In one embodiment, communications network 1860 is the Internet, while in other embodiments, communications network 1860 may be any suitable communications network including a local area network (LAN), a wide area network (WAN), a wireless network, a cellular network, a personal area network, an intranet, a private network, a near field communication (NFC) network, a public network, a switched network, a peer-to-peer network, and combinations thereof.
[0211] In an embodiment, the server 1820 is not located near the user of the computing device and communicates through a network. In another embodiment, the server 1820 is a device that the user can wear or keep nearby. In an embodiment, the server 1820 has a large battery to power a long-range communication network such as a cell network or Wi-Fi. The server 1820 communicates with other components of the system via a wired link or a low-power short-range wireless communication such as BLUETOOTH. In an embodiment, one of the other components of the system, for example the PC 1810b, plays the role of the server.
[0212] The distributed computer network 1800 of FIG. 18 is merely illustrative of one embodiment incorporating multiple embodiments and is not intended to limit the scope of the invention as recited in the claims. One of ordinary skill in the art would recognize other variations, modifications, and alternatives. For example, two or more server systems 1820 may be connected to the communications network 1860. As another example, multiple computing devices 1810a-1810b may be connected to the communications network 1860 via an access provider (not shown) or some other server system.
[0213] Computing devices 1810a-1810b typically request information from a server system that provides the information. A server system, by definition, typically has more computing and storage capacity than these computing devices and is often a portable device, mobile communications device, or other computing device that plays the role of a client in a client-server operation. However, a particular computing device can function as both a client and a server depending on whether the computing device is requesting or providing information. Aspects of the embodiments may be implemented using client-server or cloud-cloud computing environments.
[0214] The server 1820 is responsible for receiving information requests from the computing devices 1810a-1810b, performing the necessary processing to fulfill the request, and forwarding the results corresponding to the request to the requesting computing device. The processing required to fulfill the request may be performed by the server system 1820 or may be delegated to other servers connected to the communication network 1860 or other communication networks. The server 1820 may be located close to the computing device 1810 or may be located far from the computing device 1810. The server 1820 may be a hub that controls a local enclave of things in an Internet of Things scenario.
[0215] The computing devices 1810a-1810b enable users to access and query information or applications stored by the server system 1820. Examples of computing devices include portable electronic devices (e.g., mobile communications devices), such as the Apple iPhone, Apple iPad, Palm Pre, or any computing device running Apple iOS, Android OS, Google Chrome OS, Symbian OS, Windows 10, Windows Mobile OS, Palm OS, or Palm Web OS, Internet of Things (IoT) devices, such as RIOT OS, Windows 10 for IoT, WindRiver VxWorks, Google Brillo, ARM Mbed OS, embedded Apple OS and OS X, Nucleus RTOS, Green Hills Integrity, or Contiki, any of the various operating systems used in automobiles or other vehicles, or real-time operating systems (RTOS), such as Microware OS-9, VxWorks, QNX Neutrino, FreeRTOS, Micrium μC / OS-II, Micrium μC / OS-III, Windows Various programmable logic controller (PLC) or programmable automation controller (PAC) operating systems, such as CE, TI-RTOS, RTEMS, etc. Other operating systems may be used. In a particular embodiment, a "web browser" application running on the computing device allows a user to select, access, retrieve, or query information and / or applications stored by the server system 1820.Examples of web browsers include the Android browser provided by Google, the Safari® browser provided by Apple, the Opera web browser provided by Opera Software, the BlackBerry® browser provided by Research In Motion, the Internet Explorer® and Internet Explorer Mobile browsers provided by Microsoft Corporation, and the Firefox® and Firefox for Mobile browsers provided by Mozilla®.
[0216] Figure 19 is an exemplary block diagram illustrating a computing device 1900 of one embodiment. The computing device 1900 may be any of the computing devices 1810a, 1810b, 1820 of Figure 18. The computing device 1900 may include a display, screen, or monitor 1905, a housing 1910, and input devices 1915. The housing 1910 encases familiar computer components such as a processor 1920, memory 1925, a battery 1930, a speaker, a transceiver, an antenna 1935, a microphone, ports, jacks, connectors, a camera, an input / output (I / O) controller, a display adapter, a network interface, a mass storage device 1940, and various sensors, some of which are not shown.
[0217] The input device 1915 may also include a touch screen (e.g., resistive, surface acoustic wave, capacitive sensing, infrared, optical imaging, dispersive signal, or acoustic pulse recognition), a keyboard (e.g., electronic or physical), buttons, switches, a stylus, or combinations thereof.
[0218] Mass storage device 1940 may include flash and other non-volatile solid state storage or solid state drives (SSDs), such as flash drives, flash memory, or USB flash drives. Other examples of mass storage include mass disk drives, floppy disks, magnetic disks, optical disks, magneto-optical disks, fixed disks, hard disks, SD cards, CD-ROMs, writable CDs, DVDs, writable DVDs (e.g., DVD-R, DVD+R, DVD-RW, DVD+RW, HD-DVD, or Blu-ray Discs), battery-backed volatile memory, tape storage, readers, and other similar media, and combinations thereof.
[0219] The embodiments may also be used with computer systems having different configurations, such as computer systems with additional or fewer subsystems. For example, the computer system may include multiple processors (i.e., a multiprocessor system that allows for parallel processing of information), or the system may include a cache memory. The computer system shown in FIG. 19 is only one example of a computer system suitable for use with the embodiments. Other configurations of subsystems suitable for use with the embodiments will be readily apparent to those skilled in the art. For example, in certain implementations, the computing device is a mobile communication device, such as a smartphone or tablet computer. Specific examples of smartphones include the Droid Incredible and Google Nexus One offered by HTC Corporation, the iPhone or iPad offered by Apple, and many others. The computing device may be a laptop or netbook. In another particular implementation, the computing device is a non-portable computing device, such as a desktop computer or a workstation.
[0220] Computer-implemented or computer-executable versions of program instructions useful for practicing the embodiments can be embodied using, stored on, or associated with a computer-readable medium. A computer-readable medium can include any medium involved in transmitting instructions to one or more processors for execution, such as memory 1925 or mass storage device 1940. Such media can take many forms, including but not limited to non-volatile media, volatile media, transmission media, non-print media, and print media. Non-volatile media include, for example, flash memory, optical disks, magnetic disks, and the like. Volatile media include static or dynamic memory, such as cache memory and RAM. Transmission media include coaxial cables, copper wire, fiber optic lines, and wires arranged in a bus. Transmission media can take the form of electromagnetic, radio frequency, acoustic, or light waves, such as those generated during radio wave or infrared data communications.
[0221] For example, a binary, machine-executable version of the software useful for practicing the embodiments may be stored or resident in RAM or cache memory, or in mass storage device 1940. The source code of this software may also be stored or resident in mass storage device 1940 (e.g., a flash drive, hard disk, magnetic disk, tape, or CD-ROM). As a further example, code useful for practicing the embodiments may be transmitted over a network, such as wires, radio waves, or the Internet. Another particular embodiment comprises a computer program product including various software program code for implementing features of the embodiments.
[0222] The computer software product may be written in any of a variety of suitable programming languages, such as C, C++, C#, Pascal, Fortran, Perl, Matlab (from MathWorks, www.mathworks.com), SAS, SPSS, JavaScript, CoffeeScript, Objective-C, Swift, Objective-J, Ruby, Rust, Python, Erlang, Lisp, Scala, Clojure, and Java. The computer software product may be a separate application with a data entry module and a data display module. Alternatively, the computer software product may be a class that can be instantiated as a distributed object. The computer software product may be component software, such as Java Beans (from Oracle) or Enterprise Java Beans (from Oracle EJB).
[0223] The system's operating system may be the Android operating system, iPhone OS (i.e., iOS), Symbian, BlackBerry OS, Palm web OS, Bada, MeeGo, Maemo, Limo, or Brew OS. Other examples of operating systems include any of the various operating systems used for Internet of Things (IoT) devices, automobiles or other vehicles, or Real Time Operating Systems (RTOS), such as the Microsoft Windows family of operating systems (e.g., Windows 95, 98, Me, Windows NT, Windows 2000, Windows XP, Windows XP x64 Edition, Windows Vista, Windows 10, or other Windows versions, Windows CE, Windows Mobile, Windows Phone, Windows 10 Mobile), Linux, HP-UX, UNIX, Sun OS, Solaris, Mac OS X, Alpha OS, AIX, IRIX32, or IRIX64, or RIOT OS, Windows 10 for IoT, WindRiver VxWorks, Google Brillo, ARM Mbed OS, embedded Apple iOS and OS X, Nucleus RTOS, Green Hills Integrity, or Contiki, or Microware OS-9, VxWorks, QNX Neutrino, FreeRTOS, Micrium μC / OS-III, Windows Any one of a variety of programmable logic controller (PLC) or programmable automation controller (PAC) operating systems may be used, such as CE, TI-RTOS, RTEMS, etc. Other operating systems may also be used.
[0224] Further, a computer may be connected to a network, which may be used to connect to other computers. The network may be, among other things, an intranet, the Internet, the Internet, etc. The network may be a wired network (e.g., using copper wires), a telephone network, a packet network, an optical network (e.g., using fiber optics), a wireless network, or any combination thereof. For example, data and other information may be sent between a computer and components (or steps) of a system useful for implementing the embodiments using a wireless network using protocols such as Wi-Fi (IEEE standards 802.11, 802.11a, 802.11b, 802.11e, 802.11g, 802.11i, and 802.11n, to name a few), or other protocols such as BLUETOOTH, NFC, 802.15, or cellular, or communications protocols that may include TCP / IP, UDP, HTTP protocols, Wireless Application Protocol (WAP), BLUETOOTH, Zigbee, 802.11, 802.15, 6LoWPAN, LiFi, Google Weave, NFC, GSM, CDMA, other cellular data communications protocols, wireless telephone protocols, etc. For example, signals from a computer may be transferred at least in part wirelessly to a component or to another computer.
[0225] The following paragraphs contain enumerated embodiments. Embodiment 1. Defining a target set of chemical reactions; selecting a first set of chemical reactions based in part on the measures of relevance to the target set; carrying out said first set of chemical reactions; determining an outcome for each chemical reaction performed in said first set; Assembling a training data set including at least one determined outcome; 1. A method comprising: using a computer system, machine learning, and a training data set to build and train a model that predicts properties or outcomes of a chemical reaction or suggests one or more reactants, reaction conditions, or products for completing an incomplete chemical reaction.
[0226] Embodiment 2. providing inputs to the model, the inputs including one or more products, substrates, or conditions; using said inputs and said computer system executing said model to generate one or more of the following: Prediction of chemical reactions, A predicted set of reaction conditions, or One or more of the reactants, reaction conditions, or products suggested for completing an incomplete chemical reaction; or a predicted result of the incomplete chemical reaction; and 2. The method of claim 1, further comprising providing the generated prediction or suggestion to a user.
[0227] Embodiment 3. 3. The method of embodiment 2, wherein the providing step is performed using a user interface.
[0228] Embodiment 4. After constructing and training the model, determining one or more iterations of selecting a first set of chemical reactions, executing the first set of chemical reactions, determining the determined results, or assembling a training data set; 2. The method of embodiment 1, further comprising repeating said one or more steps.
[0229] Embodiment 5. 5. The method of embodiment 4, wherein the decision to repeat one or more of said steps is performed automatically by said computer system.
[0230] Embodiment 6. said first set of chemical reactions being carried out using an automated or semi-automated laboratory apparatus; 2. The method of embodiment 1, wherein determining the determined result comprises measuring each post-reaction mixture and quantifying using software processing to determine at least one yield.
[0231] Embodiment 7. 2. The method of embodiment 1, wherein defining the target set includes defining the target set by specifying one or more constraints that chemical reactions of the target set must satisfy.
[0232] Embodiment 8. Defining the target set comprises: providing, by a user, a list of chemical compounds, one or more constraints on the chemical compounds, or one or more constraints on a reaction; 2. The method of embodiment 1, comprising defining said target set as hypothetical reactions satisfying constraints having products from said list of chemical compounds or products that satisfy said constraints.
[0233] EMBODIMENT 9. 2. The method of embodiment 1, wherein said first set of chemical reactions is selected based in part on one or more factors including: (a) the chemical similarity between the set of reactions and the target set of reactions; (b) the chemical similarity between the reactions in said set; (c) the cost of a reagent or reactant in a first chemical reaction; (d) the availability of reagents or reactants in said first chemical reaction; (e) one or more predictions of said model when inputted with said chemical reactions; or (f) one or more estimates of uncertainty regarding the predictions of said model when said chemical reactions are input.
[0234] EMBODIMENT 10. providing inputs to said model including one or more products, substrates, or conditions from: the target set, a set of chemical reactions that is more chemically complex than said first set of reactions; or a portion of the reactions run that were not used to train the model; using said inputs and said computer system executing said model to generate one or more of the following: Chemical reaction prediction results, A predicted optimal set of reaction conditions, or Proposed reagents or products for completing a partial chemical reaction; comparing the generated predictions or suggestions with responses from said target set; determining a level of performance of the model based on the comparison; 5. The method of embodiment 4, further comprising determining, wherein the decision to repeat one or more of the steps is based on the level of performance.
[0235] Embodiment 11. 2. The method of embodiment 1, wherein the training data set includes one or more of the following: (i) a result of the chemical reaction determined by carrying out the chemical reaction; (ii) chemical reaction results extracted from the text by the computer system; (iii) the results of a computer program that uses molecular modelling to simulate the outcome of a chemical reaction; or (iv) Results of chemical reactions recorded in an electronic lab notebook.
[0236] Embodiment 12. using said inputs and said computer system to execute said model; Chemical reaction prediction results, A predicted optimal set of reaction conditions, or Producing one or more of the proposed reagents or products to complete a partial chemical reaction; using said inputs and said computer system executing said model to generate a plurality of predicted outcomes of a chemical reaction or a plurality of sets of optimal conditions for carrying out said chemical reaction; 3. The method of claim 2, comprising filtering the plurality of predicted outcomes or the plurality of sets of optimal conditions by the model to remove predicted outcomes with a certainty level below a threshold level of certainty or remove sets of optimal conditions with a performance level below a threshold level of performance.
[0237] Embodiment 13. 2. The method of claim 1, wherein when a person is asked a question that affects the method in some way, a user interface is presented that has one or more of the following characteristics: (a) the performance of said model according to any metric; (b) predictions of the model supplemented by examples taken from the dataset used to train the model; (c) Functionality may also be present in the user interface used by the user to interact with the model.
[0238] Embodiment 14. 14. The method of embodiment 13, wherein the set of reactions is also selected based on factors including a numerical score assigned by a person using the user interface to answer questions about one or more chemical reactions.
[0239] EMBODIMENT 15. A user inputs a partially specified recipe for how to synthesize a collection of compounds that is not yet ready to be run; using the model to generate missing information for the recipe that satisfies user-provided constraints; and Optionally, displaying said recipe and / or said collection of compounds in a user interface; 2. The method of embodiment 1, wherein, if desired, the synthesis of a new collection of compounds is planned and potentially performed by carrying out said process to synthesize said collection of compounds.
[0240] EMBODIMENT 16. 16. The method of embodiment 15, wherein said collection of compounds is dissolved in a single solution and each compound is identified by a DNA strand or another set of atoms that allows its identification.
[0241] EMBODIMENT 17. User-provided constraints include: (a) the reaction yield exceeds a predetermined threshold; and 16. The method of embodiment 15, wherein the conditions include one or more of: (b) satisfying certain logical constraints.
[0242] EMBODIMENT 18. The synthesis plan is (a) the user specifying a list of starting compounds; (a) the user specifying a list of constraints that each reaction in the final synthetic plan must satisfy; (b) the user specifying a maximum number of synthesis steps; 16. The method of embodiment 15, wherein (c) every compound is generated using steps enumerated such that there is a series of reactions generated by said model that ends with this compound as the product.
[0243] EMBODIMENT 19. The user inputs a partially specified chemical reaction (e.g., containing only one specified product and one specified substrate); further training the model to predict missing portions of the reaction, and then completing the reaction using the model; 2. The method of embodiment 1, further comprising: a model generating a prediction regarding the optimal conditions and yield of the completed reaction.
[0244] EMBODIMENT 20. A user inputs a target molecular structure; generating one or more synthetic pathways for a predicted target molecular structure using said model and a synthetic planning algorithm utilizing said model and any synthetic planning software; 2. The method of embodiment 1, further comprising displaying the predicted synthetic pathway using a user interface.
[0245] EMBODIMENT 21. 21. The method of embodiment 20, further comprising generating a synthetic pathway using a retrosynthetic algorithm that uses the optima predicted by the model as factors influencing the selection of the synthetic pathway.
[0246] EMBODIMENT 22. 2. The method of embodiment 1, wherein any of the following is true: (i) performing the chemical reaction using an automated solid and / or liquid dispenser; (ii) the chemical reactions are carried out in multiwell plates of standardized dimensions, with each well carrying an individual reaction; and (iii) performing chemical reactions in glass, plastic, or metal containers, each reaction being performed in a separate container, and the containers being spatially organized in a rack; (iv) carrying out the chemical reaction is accomplished by stirring the reaction mixture with a magnetic stir bar or an orbital shaker; (v) Carrying out a chemical reaction is accomplished by heating or cooling said reaction mixture in a heating / cooling block or a thermoshaker.
[0247] EMBODIMENT 23. 2. The method of embodiment 1, wherein the analysis of the expected amount of product in the post-reaction mixture is accomplished by any combination of: (i): Liquid chromatography in combination with one or more of the following detectors: Single or multi-wavelength UV-Vis detectors, Fluorescence detector, Evaporative Light Scattering Detector (ELSD), Charged Aerosol Detector (CAD), Radiation detectors, Electrochemical detectors, Nitrogen chemiluminescence detector, or mass spectrometry, (ii) determining the amount of said expected product in the combined fractions containing said product by liquid chromatography with manual or automated collection of fractions, followed by an applicable analytical method; (iii) A method similar to (ii), but involving separation of the product by solid phase extraction (SPE); (iv) A method similar to (i) but in MISER (multiple injections in a single experimental run) mode; (v) MALDI-MS or Echo-MS analysis of a suitably prepared analytical sample of the post-reaction mixture without separation of the components; (vi) Nuclear magnetic resonance (NMR) spectroscopy of the untreated or in some way treated post-reaction mixture.
[0248] EMBODIMENT 24. 2. The method of embodiment 1, wherein any of the following is true: (i) processing signals (data) from an analytical instrument that obtained the analytical sample of the reaction mixture by a dedicated computer program to automatically quantify the expected product; (ii) predicting the level of analytical signal of the expected reaction product using any computational method or ML model; (iii) Any computational method or ML model using said analytical signals of internal analytical standard(s) and reaction products to quantify the amount of product in said analytical sample.
[0249] EMBODIMENT 25. 2. The method of claim 1, wherein the machine learning model has any of the following characteristics: (a) The model architecture is a sequence ordering deep neural network, such as a Transformer architecture; and (b) the model is based on an ensemble; and (c) the model outputs additionally include measures of the uncertainties of other outputs; and (d) the model uncertainties are calculated based on the individual outputs of the ensemble members; and (e) the model accepts as inputs a set of logical constraints that are satisfied by the outputs, and produces outputs that satisfy those constraints.
[0250] 26. 2. The method of embodiment 1, further comprising a user interface that allows a user to explore and view reactions that are part of the dataset used to train the model.
[0251] EMBODIMENT 27. 2. The method of embodiment 1, further comprising a user interface that allows a user to run selected reactions in an external semi-automated or automated laboratory.
[0252] 28. The method of embodiment 27 is however enabled by a mechanism that allows queries to be performed against said database that surfaces reactions that satisfy user-provided constraints, such as the chemical structures present in the reaction.
[0253] 29. 2. The method of embodiment 1, further comprising a user interface that allows a user to use the software programmatically, for example by encoding on a computer medium a set of instructions that instruct the software to perform any actions that a user can perform manually.
[0254] EMBODIMENT 30. The model is trained to provide predictions or suggestions that satisfy any of the following constraints: (a) the confidence level of the prediction exceeds a certain threshold; (b) the logical constraints that are satisfied by the predicted condition or outcome; (c) the logical constraints that are satisfied by the proposed response; providing inputs to the model including one or more products, substrates, or conditions, the inputs further including any of the following: (a) a selected confidence level threshold; (b) selected logical constraints that are satisfied by the predicted condition or outcome; (c) selected logical constraints that are satisfied by the proposed response; using said inputs and said computer system executing said model to generate one or more of the following: a predicted outcome for the chemical reaction, with an associated possible confidence level that matches the input confidence level; A predicted optimal set of reaction conditions with associated possible confidence levels that match the input confidence level, or Suggested reactants, reaction conditions, or products for completing the partial chemical reaction, with associated possible confidence levels that match the input confidence level; and 3. The method of claim 2, further comprising providing the generated prediction or suggestion to a user.
[0255] EMBODIMENT 31. said user inputting a target molecular structure; generating, using said model and a composite planning algorithm utilizing said model and composite planning software; Prediction of one or more synthetic pathways for a target molecular structure; and Optionally, displaying the predicted synthetic route using a user interface; or 2. The method of embodiment 1, optionally further comprising carrying out said synthetic scheme using an automated or semi-automated laboratory.
[0256] EMBODIMENT 32. 2. The method of embodiment 1, wherein any of the following is true: (i) processing signals (data) from an analytical instrument that has acquired the analytical sample of the post-reaction mixture by a dedicated computer program in order to automatically quantify the expected product; (ii) predicting the level of analytical signal of the expected reaction product using any computational method or ML model; (iii) Any computational approach or ML model uses said analytical signals of internal analytical standard(s) and reaction products to quantify the amount of product in said analytical sample.
[0257] EMBODIMENT 33. The synthesis of a compound or a collection of compounds, designing, by a user or by said first computer system or by a second computer system, a partially specified recipe for how to synthesize said compound or said collection of compounds; The method of claim 1 , further comprising planning by using the model to generate missing information for the recipe that satisfies user-provided constraints.
[0258] A system comprising at least one processor and a memory comprising instructions that, when executed by the at least one processor, cause the system to perform operations according to any of the above embodiments 1-33.
[0259] A non-transitory computer-readable medium comprising instructions that, when executed by a processor of a computing device, cause the computing device to perform operations according to any of embodiments 1 to 33 above.
[0260] While the embodiments have been described with respect to specific embodiments, it will be recognized that further variations can be devised without departing from the inventive concept.
[0261] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the subject matter described in the claims. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items. As used herein, the singular forms "a", "an" and "the" are intended to include the plural and singular unless the context clearly indicates otherwise. It will be further understood that the terms "comprise" and / or "comprising", as used herein, specify the presence of stated features, steps, operations, elements and / or components but do not exclude the presence of one or more other features, steps, operations, elements, components and / or groups thereof.
[0262] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which the embodiments belong. Furthermore, it should be understood that terms as defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the relevant art and this disclosure, and should not be interpreted in an idealized or overly formal sense unless expressly defined herein.
[0263] In describing the embodiments, it will be understood that a number of elements, techniques, and steps are disclosed, each of which has separate advantages and each of which can be used in combination with one or more, and in some cases all, of the other disclosed elements or techniques, and the specification and claims should be read with the understanding that such combinations are fully within the scope of the subject matter described in the embodiments and claims.
[0264] In the above and throughout the entire description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure. However, it will be apparent to those skilled in the art that the embodiments may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form to facilitate explanation. The description of the preferred embodiments is not intended to limit the scope of the claims appended hereto. Furthermore, in the methods disclosed herein, various steps are disclosed that illustrate some of the functionality of the embodiments. These procedures are merely examples and are not intended to be limiting in any way. Other steps and functions may be considered without departing from the scope of the disclosure or the embodiments.
Claims
1. Defining a target set of chemical reactions; selecting a first set of chemical reactions based in part on the measures of relevance to the target set; carrying out said first set of chemical reactions; determining an outcome for each chemical reaction performed in the first set; Assembling a training data set including at least one determined outcome; 1. A method comprising: using a first computer system, machine learning, and a training data set to build and train a model that predicts properties or outcomes of a chemical reaction or suggests one or more reactants, reaction conditions, or products for completing an incomplete chemical reaction.
2. providing inputs to the model including one or more products, reactants, or reaction conditions; using said inputs and said computer system executing said model to generate one or more of the following: Predicted properties or results of chemical reactions; A predicted set of reaction conditions, or One or more of the reactants, reaction conditions, or products proposed to drive an incomplete chemical reaction to completion; or a predicted result of the incomplete chemical reaction; and The method of claim 1 , further comprising providing the generated prediction or suggestion to a user.
3. After constructing and training the model, determining one or more iterations of selecting a first set of chemical reactions, executing the first set of chemical reactions, determining the determined results, or assembling a training data set; The method of claim 1 , further comprising repeating the one or more steps.
4. The method of claim 3 , wherein the decision to repeat one or more of the steps is performed automatically by the first computer system or the second computer system.
5. said first set of chemical reactions being carried out using an automated or semi-automated laboratory apparatus; 10. The method of claim 1, wherein determining the determined results comprises measuring each post-reaction mixture and quantifying using software processing to determine at least one yield.
6. The method of claim 1 , wherein defining the target set comprises defining the target set by specifying one or more constraints that chemical reactions of the target set must satisfy.
7. Defining the target set comprises: providing, by a user, a list of chemical compounds, one or more constraints on the chemical compounds, or one or more constraints on a reaction; The method of claim 1 , comprising defining the target set as products from the list of chemical compounds or hypothetical reactions that satisfy the constraints and have products that satisfy the constraints.
8. 10. The method of claim 1 , wherein the first set of chemical reactions is selected based in part on one or more factors including: (a) the chemical similarity between the set of reactions and the target set of reactions; (b) the chemical similarity between the reactions in the set; (c) the cost of a reagent or reactant in the first chemical reaction; (d) the availability of reagents or reactants in said first chemical reaction; (e) one or more predictions of the model when inputted with the chemical reactions; or (f) one or more estimates of uncertainty regarding the predictions of said model when said chemical reactions are input.
9. providing inputs to the model including one or more products, substrates, or conditions from: the target set, a set of chemical reactions that are more chemically complex than said first set of reactions; or a portion of the reactions run that were not used to train the model; using said inputs and said computer system executing said model to generate one or more of the following: Prediction of chemical reactions, A predicted optimal set of reaction conditions, or Proposed reagents or products for completing a partial chemical reaction; comparing the generated predictions or suggestions with responses from said target set; determining a level of performance of the model based on the comparison; The method of claim 3 , further comprising: determining to repeat one or more of the steps based on the level of performance.
10. 10. The method of claim 1, wherein the training data set includes one or more of the following: (i) an outcome of the chemical reaction determined by carrying out the chemical reaction; (ii) the results of chemical reactions extracted from the text by a computer system; (iii) the results of a computer program that uses molecular modeling to simulate the outcome of a chemical reaction; or (iv) Results of chemical reactions recorded in an electronic lab notebook.
11. using said computer system to execute said inputs and said model; Prediction of chemical reactions, A predicted optimal set of reaction conditions, or Producing one or more of the proposed reagents or products to complete a partial chemical reaction; using said inputs and said computer system executing said model to generate a plurality of predicted outcomes of a chemical reaction or a plurality of sets of optimal conditions for carrying out said chemical reaction; 3. The method of claim 2, further comprising filtering the plurality of predicted outcomes or the plurality of sets of optimal conditions by the model to remove predicted outcomes with a certainty level below a threshold level of certainty or remove sets of optimal conditions with a performance level below a threshold level of performance.
12. 10. The method of claim 1, wherein when a user is asked a question that affects the method in some way, a user interface is presented that has one or more of the following characteristics: (a) the performance of the model by any metric; (b) predictions of the model supplemented by examples taken from the dataset used to train the model; or (c) Features that may also be present in the user interface used by the user to interact with the model.
13. 13. The method of claim 12, wherein the set of reactions is selected based also on factors including a numerical score assigned by a person using the user interface to answer questions about one or more chemical reactions.
14. The synthesis of a compound or a collection of compounds, designing, by a user or by said first computer system or by a second computer system, a partially specified recipe for how to synthesize said compound or said collection of compounds; The method of claim 1 , further comprising planning by using the model to generate missing information for the recipe that satisfies user-provided constraints.
15. 1. A system comprising at least one processor and a memory comprising instructions that, when executed by the at least one processor, cause the system to perform operations, the operations being: receiving a target set of chemical reactions; receiving a first set of selected chemical reactions based in part on the measures of relevance to the target set; determining a result for each chemical reaction performed in the first set; and receiving an assembled training data set including at least one outcome from each chemical reaction from the first set, each of the at least one outcome being determined from performance of a different chemical reaction from the first set; and The system further comprises using machine learning and the training dataset to build and train a model to predict properties of a chemical reaction and to suggest reagents and products for completing a partially specified chemical reaction.
16. The operation includes: receiving inputs to the model including one or more products, substrates, or conditions; running the model with the input to generate one or more of the following: Prediction of chemical reactions, A predicted optimal set of reaction conditions, or Proposed reagents or products for completing a partial chemical reaction, and The system of claim 15 , further comprising providing the generated prediction or suggestion to a user.
17. using said computer system to execute said inputs and said model; Prediction of chemical reactions, A predicted optimal set of reaction conditions, or Producing one or more of the proposed reagents or products to complete a partial chemical reaction; running the model with the inputs to generate a plurality of predicted outcomes of a chemical reaction or a plurality of sets of optimal conditions for carrying out the chemical reaction; and 17. The system of claim 16, further comprising using the model to filter the plurality of predicted outcomes or the plurality of sets of optimal conditions to remove predicted outcomes with a certainty level below a threshold level of certainty or remove sets of optimal conditions with a performance level below a threshold level of performance.
18. 1. A non-transitory computer-readable medium comprising instructions that, when executed by a processor of a computing device, cause the computing device to perform operations, the operations including: receiving a target set of chemical reactions; receiving a first set of selected chemical reactions based in part on the measures of relevance to the target set; determining a result for each chemical reaction performed in the first set; and receiving an assembled training data set including at least one outcome from each chemical reaction from the first set, each of the at least one outcome being determined from performance of a different chemical reaction from the first set; and The non-transitory computer-readable medium further comprises: building and training a model using machine learning and the training dataset to predict properties of a chemical reaction and to suggest reagents and products for completing a partially specified chemical reaction.
19. The operation includes: receiving inputs to the model including one or more products, substrates, or conditions; running the model with the input to generate one or more of the following: Chemical reaction prediction results, A predicted optimal set of reaction conditions, or Proposed reagents or products for completing a partial chemical reaction, and 20. The non-transitory computer-readable medium of claim 18, further comprising providing the generated prediction or suggestion to a user.
20. using said computer system to execute said inputs and said model; Chemical reaction prediction results, A predicted optimal set of reaction conditions, or Producing one or more of the proposed reagents or products to complete a partial chemical reaction; running the model with the inputs to generate a plurality of predicted outcomes of a chemical reaction or a plurality of sets of optimal conditions for carrying out the chemical reaction; and 20. The non-transitory computer-readable medium of claim 19, further comprising using the model to filter the plurality of predicted outcomes or the plurality of sets of optimal conditions to remove predicted outcomes with a certainty level below a threshold level of certainty or remove sets of optimal conditions with a performance level below a threshold level of performance.
Citation Information
Cited By
Method and system for predicting and profiling impurities in chemical reaction
WO2026111478A1
