System and method for query-based random access to virtual chemical combination synthesis library

Abstract

The present disclosure relates to systems and methods for querying a combinatorial synthesis library comprising a plurality of compounds and representing a plurality of reaction types, where each reaction type is mapped to a plurality of reactants and each reactant is mapped to a plurality of synthons, accepting a query of a molecular encoder model in the form of a single graph, thereby obtaining a query vector. The query vector is input into a reaction query generator model to obtain a first reaction type and a first plurality of reactants. A synthon for each reactant is determined by inputting the reactants into a synthon query generator model. A set of synthons is thus determined, each synthon corresponding to a reactant of the first plurality of reactants. A molecular structure in the combinatorial synthesis library is identified, the combinatorial synthesis library comprising the set of components arranged according to a synthesis rule associated with the first reaction type.

Description

[0001] Cross-references to related patent applications

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 342,574, filed on May 16, 2022, entitled “GRAPH GENERATIVE MODEL FOR COMBINATORIAL SYNTHESIS LIBRARIES,” which is hereby incorporated by reference herein. Technical Field

[0003] The present disclosure generally relates to screening combinatorial synthetic libraries for analogs of a target. Background Art

[0004] Virtual high-throughput screening (vHTS)

[47] has gained tremendous attention in early drug discovery, in large part due to the availability of on-demand chemical libraries constructed using combinatorial synthesis. These combinatorial synthetic libraries (CSLs) enable access to an extremely large amount of chemical space from a relatively small set of chemically accessible building blocks that can be combined according to known synthetic routines. In recent years, these libraries have grown from millions to billions and now trillions of compounds [22, 33, 44, 58]. As a result, the size of virtual chemical libraries is rapidly approaching a size that would allow explicit enumeration, posing new challenges for virtual screening. For example, the Enamine READily Accessible (REAL) library

[20] leverages readily available molecular building blocks and parallel synthesis, allowing lead times of weeks and ushering in an era of ever-decreasing delays between in silico and in vitro high-throughput screening.

[0005] Due to the combinatorial explosion enabled by these constructs, early drug discovery has now "crossed the Rubicon" into the non-enumerative realm. This poses new challenges for in silico hit discovery and optimization, which rely on screening explicitly enumerated compounds. These approaches are not well suited to the non-enumerative setting, scaling linearly with the number of compounds.

[0006] Virtual high-throughput screening and enumeration. Typically, the first step in a vHTS campaign is to prepare a compound library for subsequent use [1, 17]. Although compound sampling and scoring techniques have been developed

[18] , these approaches still rely on an exhaustive accessible library. An exception is the virtual synthon hierarchical enumeration screening (V-SYNTHES) approach

[44] , which exploits the modular nature of parallel synthetic libraries. However, by design, V-SYNTHES does not allow query-based random access. On the other hand, SpaceMACS

[45] and SpaceLight [4] can provide query-based access to modular libraries by decomposing queries into fragments and matching them to synthons in the library through similarity searches. In parallel with these efforts, machine learning has also received considerable attention in vHTS: for predicting activity scores given docked conformations [15, 43, 57], predicting activity scores given ligands and proteins (undocked) respectively [38, 56], and improving or completely replacing classical molecular docking with machine learning methods [39, 51, 52].

[0007] Deep Learning Methods for Molecular Generation. De novo drug design plays an increasingly important role in identifying novel chemical entities in drug discovery activities [10, 36, 50, 54]. The two main neural network-based paradigms for molecular generation are text-based generative models and graph-based generative models. Early work on text-based generative models (also known as chemical language models) applied recurrent neural networks to SMILES strings [16, 46]. Although these methods have shown great promise and stimulated interest in molecular generation within the ML community, they are not guaranteed to produce valid SMILES strings. Methods that exploit the grammatical constraints of SMILES notation have been proposed to improve validity [12, 29]; in addition, the more recently proposed SELFIES notation [28, 37] guarantees validity and has been increasingly adopted. However, in both cases, modeling using such text-based chemical representations still suffers from known drawbacks (e.g., subjectivity, similar molecular structures may have large edit distances).

[0008] For some applications, it makes sense to exploit generative models that can be fitted to molecular databases, allowing navigation of these databases via the fitted model. The ability of language models to adapt to molecular databases was investigated in a previous study [2], which applied a deep language model to GDB-13 [7], a database of 975,000,000 compounds formed by fully enumerating molecules of up to 13 atoms of elemental types C, N, O, S, and Cl, subject to simple chemical stability and synthetic feasibility rules. The authors trained the model on 0.1% of the entire library and found that the model was able to cover approximately 70% of the compounds in the GDB-13 library. Furthermore, the language model they trained generated compounds that did not satisfy the GDB-13 construct in approximately 15% of the cases.

[0009] In recent years, graph generation models have received significant attention as an alternative to their text-based counterparts. The earliest of these models focused on generating graphs of constant size in a single pass

[48] or generating graphs of arbitrary size autoregressively, one atom or bond at a time [32, 42, 60, 35]. These approaches also struggle to reliably generate chemically valid molecules and encounter difficulties when dealing with large molecular graphs.

[0010] To address these two points, fragment-based graph generation models have been proposed and are becoming increasingly popular [23,24,25,27]. The advantage of these models is that they guarantee chemical validity by decomposing molecules into valid subcomponents and explicitly prohibiting operations that produce invalid fragment combinations. Such explicit validity checks can be performed for each action, but at the cost of additional computation. While other text- and graph-based generation models often have difficulty handling large molecular graphs due to the long autoregressive chains required to generate them, fragment-based graph generation models require an autoregressive length comparable to the number of fragments comprising the molecule. This can be important when the fragments themselves contain many atoms.

[0011] However, some issues remain due to the general difficulty of autoregressive graph generation. Unlike text-based models, where there is less ambiguity in the order of autoregressions (e.g., tokens are usually decoded in a left-to-right order), graphs have no such canonical node order, which poses a challenge for graph-based autoencoders [32, 59]. Furthermore, although they require shorter autoregressive chains than their counterparts, existing fragment-based graph generation models still require autoregressive lengths that grow in terms of the overall size of the molecule, as autoregressive decoding cannot be efficiently parallelized.

[0012] While fragment-based graph generative models and SELFIES-based language models have each addressed the issue of chemical validity, synthetic accessibility presents a separate challenge. Prior work has questioned the synthetic feasibility of many compounds proposed by existing generative models

[13] , which, if not properly addressed, could limit the practical utility of these models in drug discovery applications. Subsequent work has attempted to improve on these shortcomings by (i) including explicit penalties for synthetic inaccessibility through scoring functions

[19] , (ii) restricting models to fragments of known compounds [34, 40, 53], or (iii) inducing biases towards simple and known synthetic pathways [8, 9, 21].

[0013]

[0006] In view of the foregoing background, what is needed in the art is a scalable method for navigating a CSL using query-based random access. Summary of the invention

[0014] The present disclosure solves the above needs in the art. A system and method for querying a combinatorial synthesis library that includes multiple compounds and represents multiple reaction types, wherein each reaction type is mapped to multiple reactants, and each reactant is mapped to multiple synthons. The system accepts a query in the form of a single graph into a molecular encoder model to obtain a query vector. The query vector is input into a reaction query generator model to obtain a first reaction type and a first plurality of reactants. The synthons of each reactant are determined by inputting the reactants into a synthon query generator model. Thus, a set of synthons is determined, each synthon corresponding to a reactant in the first plurality of reactants. Molecular structures in a combinatorial synthesis library are identified, and the combinatorial synthesis library includes a set of synthons arranged according to synthesis rules associated with the first reaction type.

[0015] In one embodiment, the molecular encoder model is a graph-based generative model that exploits the structure of CSL to provide efficient navigation of related chemical space. The model learns a key hierarchy on library components and uses these keys to process queries for retrieval. The encoder processes the molecular graph and returns a query vector as output, which the decoder uses to retrieve molecules from the CSL through a sequence of efficient query key comparisons, which utilizes the hierarchical structure of CSL, requires minimal autoregression and allows efficient parallelization. In such embodiments, the graph-based generative model acts as a "neural database" that provides random access to extremely large, non-enumerable compound libraries. Therefore, the model provides efficient and cost-effective accessible molecules. In addition, the model overcomes the challenges of long autoregressive chains in compound generation and improves the scalability of large molecular graphs. In addition, relative to comparable methods, the model reduces the number of parameters by ten times and provides considerable improvements in the computational complexity of searching through CSL.

[0016] A system and method for querying a combinatorial synthesis library is provided. The combinatorial synthesis library includes a plurality of compounds and represents a plurality of reaction types. Each corresponding reaction type in the plurality of reaction types has a corresponding mapping to a corresponding plurality of reactants. Each corresponding reactant in each corresponding plurality of reactants has a corresponding mapping to a corresponding plurality of synthons.

[0017] In some such embodiments, the plurality of reaction types includes one or more reaction types, and the combinatorial synthesis library includes one or more compounds for each reaction type in the plurality of reaction types.

[0018] A molecular query is input into the molecular encoder model. The query is graph-based. The molecular encoder model includes a message passing neural network including a plurality of message passing layers that collectively include a first plurality of parameters. Inputting the graph-based query into the molecular encoder model generates a query vector as an output of the molecular encoder model by applying the first plurality of parameters to the graph-based query.

[0019] In some such embodiments, the graph-based query is a single graph comprising a plurality of nodes and a plurality of edges, wherein each node of the plurality of nodes is connected to another node of the plurality of nodes by at least one edge of the plurality of edges.

[0020] In some such embodiments, each respective node in the plurality of nodes is associated with: (i) a corresponding element type in a plurality of element types, (ii) a node degree in a plurality of node degrees, (iii) a hybridization in a plurality of hybridizations, (iv) a plurality of bonded hydrogens, (v) a formal charge in a set of formal charges, and (vi) a binary indication of aromaticity.

[0021] In some such embodiments, each respective bond in the plurality of bonds is associated with: (i) a bond type, (ii) a binary indication of conjugation, (iii) a binary indication of whether the respective bond is within a ring, and (iv) an indication of stereochemistry.

[0022] In some such embodiments, the graph can be any graph representing single molecule compounds, where such single molecule compounds are present in a combinatorial synthesis library.

[0023] In some such embodiments, the graph represents a weighted composite of a first graph for a first molecular compound and a second graph for a second molecular compound.

[0024] In some such embodiments, the single graph represents a weighted composite of multiple graphs of a second plurality of compounds, where the second plurality of compounds have a common property.

[0025] In some such embodiments, the common property is a Tanimoto distance to each other compound in the second plurality of compounds that is less than a threshold value.

[0026] In some such embodiments, the common property is a binding coefficient for a macromolecular target that is less than a threshold value.

[0027] The query vector is input into a reactive query generator model comprising the second plurality of parameters. This results in an output from the reactive query generator model 116 of a first reaction type of the plurality of reaction types by applying the second plurality of parameters to the query vector.

[0028] In some such embodiments, the reaction query generator model is a two-layer perceptron with intermediate ReLU activations.

[0029] In some such embodiments, output from a reaction query generator model is used to identify a first reaction key type from a plurality of reaction key types via a first query key lookup, wherein each reaction key type from the plurality of reaction key types represents a synthesis reaction that can be used to synthesize one or more compounds in a combinatorial synthesis library.

[0030] By inputting the corresponding reactant 86 into a synthon query generator model including the third plurality of parameters, thereby obtaining the corresponding synthon as an output of the synthon query generator model by applying the third plurality of parameters to the corresponding reactant, a corresponding synthon for each corresponding reactant in the first plurality of reactants of the first reaction type is determined from the corresponding plurality of synthons mapped to the corresponding reactant. This generates a set of synthons, each synthon in the set of synthons corresponding to a reactant in the first plurality of reactants.

[0031] In some such embodiments, the synthetic subquery generator model is a two-layer perceptron with intermediate ReLU activations.

[0032] In some such embodiments, output from the synthesizer query generator model is used to identify a synthesizer key for a corresponding synthesizer via a second query key lookup.

[0033] In some such embodiments, the first plurality of parameters includes 100,000 parameters, the second plurality of parameters includes 5,000 parameters, and the third plurality of parameters includes 5,000 parameters.

[0034] In some such embodiments, the first plurality of reactants includes three or more reactants, and the corresponding mappings of the corresponding plurality of synthons for reactants of the three or more reactants include ten or more synthons.

[0035] Molecular structures (compounds) are identified in a combinatorial synthesis library, for example, the combinatorial synthesis library comprising a set of synthons arranged according to synthesis rules associated with a first reaction type.

[0036] In some such embodiments, the plurality of compounds in the combinatorial synthesis library includes one billion or more compounds, and the identified molecular structure is any one of the one billion or more compounds that satisfies the query.

[0037] In some such embodiments, the plurality of compounds in the combinatorial synthesis library includes a trillion or more compounds, and the identified molecular structure is any one of the trillion or more compounds that satisfies the query. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In the accompanying drawings, embodiments of the disclosed system and method are shown by way of example. It should be clearly understood that the description and drawings are only for illustrative purposes and as an aid to understanding, and are not intended as a definition of the limits of the disclosed system and method.

[0039] Figure 1A and 1B A computer system according to some embodiments of the present disclosure is presented.

[0040] Figure 2A , 2B 2C and 2D show methods for querying a combinatorial synthesis library according to some embodiments of the present disclosure, wherein the dashed boxes represent optional elements.

[0041] Figure 3A , 3B , 3C, 3D, 3E, 3F, 3G and 3H collectively present an overview of systems and methods for screening compound libraries according to some embodiments of the present disclosure. Figure 3A : Combinatorial synthesis libraries consist of potentially large numbers of tables describing multi-component reactions. Example two-component reactions are illustrated with four synthons per group, resulting in sixteen product chemical spaces. Such constructed libraries can greatly accelerate early drug discovery by describing large amounts of accessible chemical space through combinatorial explosion. Figure 3B : We develop an instance-efficient graph generation model and use it for navigating a very large combinatorial synthesis library. At a high level, the instance model first encodes a molecule (represented as a graph) through a GNN encoder. Given a molecule representation, a decoder generates queries for synthon-based libraries for retrieval. The decoder is constructed in such a way that retrieval is fast and efficient. Figure 3C : The library encoder starts with learning a representation of each synthon (e.g., 88-1, 88-2, 88-3, 88-4) represented as a molecular graph through a synthon encoder (GNN). The synthon representation is used to generate a key for each synthon through a multi-layer perceptron (MLP). Figure 3D : Moving up the hierarchy, group representations are learned by a set of neural networks operating on the representations of the synthons belonging to each group. Figure 3E: Similarly, the reaction representations are learned by a set of neural networks operating on the representations of the constituent groups. The reaction keys are also generated from the reaction representations via an MLP. Figure 3F : Molecular encoders are graph neural networks that take a molecular graph as input and produce molecular keys as output for querying synthon-based libraries. Figure 3G : To reconstruct the input molecule, a reaction query is first generated from the molecule key and then compared with the reaction key to sample the reaction types. Figure 3H : Next, synthon queries are generated independently for each group using the molecular key, reaction representation, and group representation (e.g., no autoregression), which are compared with the synthon key to sample the synthons (e.g., 98-1, 98-2, 98-3, 98-4, 98-5, 98-6, 98-7, and 98-8).

[0042] Figure 4A and 4B Interpolation in latent space according to an embodiment of the present disclosure is collectively demonstrated. Figure 4A : Move from left to right in raster order, and perform linear interpolation from the starting molecule to the target molecule. Below each molecule is its Tanimoto similarity with the starting molecule and the target molecule. Figure 4B : Sample two random directions in the latent space around the query compound and visualize the decoded molecules as evenly spaced on the resulting 2D plane.

[0043] Figure 5A and 5B Together, simulated autoencoding according to an embodiment of the present disclosure is shown. Each query molecule (indicated by a box in the upper left corner) is autoencoded and randomly decoded fifteen times according to the system and method of the present disclosure. Below each molecule is its Tanimoto similarity to the query molecule. Figure 5A : Automatically encode the molecules contained in the library. Figure 5B : Automatically encode molecules not included in the library.

[0044] FIG. 6 shows a synthon according to the prior art.

[0045] Figure 7 We show how the fraction of samples in the repertoire generated by RationaleRL decreases rapidly as the number of different reactions present in training increases, dropping below 50% after including only 50 reactions.

[0046] Figure 8 Test time distribution skew according to library size according to some embodiments of the present disclosure is shown.

[0047] Fig. 9 A post hoc density estimation process according to some embodiments of the present disclosure is presented.

[0048] Fig.10 A list of 2021 U.S. Federal Drug Administration novel drug approvals used in simulation comparisons according to some embodiments of the present disclosure is shown.

[0049] Fig.11 A comparison of CSLVAE and Arthor in a top 1 simulation search according to some embodiments of the present disclosure is shown.

[0050] Like reference numerals refer to corresponding parts throughout the several views of the drawings. DETAILED DESCRIPTION

[0051] Reference will now be made in detail to the embodiments, examples of which are shown in the accompanying drawings. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it will be apparent that one of ordinary skill in the art can practice the present disclosure without these specific details. In other cases, well-known methods, processes, components, circuits, and networks are not described in detail in order to avoid unnecessarily obscuring aspects of the embodiments.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs. All patents and publications mentioned herein are incorporated by reference in their entirety.

[0053] definition.

[0054] As used herein, the terms "administer", "administration", and "administering" refer to (1) providing, giving, administering, and / or prescribing by a health practitioner or his or her authorized agent or under his or her direction in accordance with the present disclosure; and / or (2) administering, ingesting, or consuming by a mammal in accordance with the present disclosure.

[0055] As used herein, the terms "co-administration / co-administering", "administered in combination with / administering in combination with", "simultaneous / concurrent" encompass the administration of two or more active pharmaceutical ingredients to a subject such that the active pharmaceutical ingredients and / or their metabolites are present in the subject at the same time. Co-administration includes simultaneous administration in separate compositions, administration at different times in separate compositions, or administration in a composition in which two or more active pharmaceutical ingredients are present. Simultaneous administration in separate compositions and administration in a composition in which both drugs are present are preferred.

[0056] The terms "active pharmaceutical ingredient" and "drug" include compounds described herein and any pharmaceutically acceptable analogs, derivatives, salts, solvates, hydrates, co-crystals or prodrugs thereof. The terms "active pharmaceutical ingredient" and "drug" may also include those compounds described herein and any pharmaceutically acceptable analogs, derivatives, salts, solvates, hydrates, co-crystals or prodrugs thereof that bind to a target molecule.

[0057] The term "in vivo" refers to events occurring within the body of a subject.

[0058] The term "in vitro" refers to events that occur outside the body of a subject. In vitro assays encompass cell-based assays in which live or dead cells are used, and may also encompass cell-free assays in which intact cells are not used.

[0059] As used herein, the term "if" may be interpreted to mean "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if [a stated condition or event] is detected" may be interpreted to mean "upon determining" or "in response to determining" or "upon detecting [a stated condition or event]" or "in response to detecting [a stated condition or event]," depending on the context.

[0060] The term "effective amount" or "therapeutically effective amount" refers to an amount of a compound or combination of compounds described herein that is sufficient to achieve the intended application (including but not limited to disease treatment). The therapeutically effective amount can vary depending on the intended application (in vitro or in vivo) or the subject to be treated and the condition (e.g., the subject's weight, age and gender), the severity of the condition, the mode of administration, etc., which can be easily determined by a person of ordinary skill in the art. The term also applies to doses that will induce a specific response in target cells. The specific dose will vary depending on the specific compound selected, the administration regimen followed, whether the compound is administered in combination with other compounds, the timing of administration, the tissue to which it is administered, and the physical delivery system carried by the compound.

[0061] As used herein, the term "therapeutic effect" encompasses therapeutic benefits and / or prophylactic benefits. A prophylactic effect includes delaying or eliminating the appearance of a disease or condition, delaying or eliminating the onset of symptoms of a disease or condition, slowing, stopping or reversing the progression of a disease or condition, or any combination thereof.

[0062] The term "pharmaceutically acceptable salt" refers to salts derived from various organic and inorganic counterions well known in the art. Pharmaceutically acceptable acid addition salts can be formed with inorganic and organic acids. Preferred inorganic acids from which salts can be derived include, for example, hydrochloric acid, hydrobromic acid, sulfuric acid, nitric acid, and phosphoric acid. Preferred organic acids from which salts can be derived include, for example, acetic acid, propionic acid, glycolic acid, pyruvic acid, oxalic acid, maleic acid, malonic acid, succinic acid, fumaric acid, tartaric acid, citric acid, benzoic acid, cinnamic acid, mandelic acid, methanesulfonic acid, ethanesulfonic acid, p-toluenesulfonic acid, and salicylic acid. Pharmaceutically acceptable base addition salts can be formed with inorganic and organic bases. Inorganic bases from which salts can be derived include, for example, sodium, potassium, lithium, ammonium, calcium, magnesium, iron, zinc, copper, manganese, and aluminum. Organic bases from which salts can be derived include, for example, primary, secondary, and tertiary amines, substituted amines, including naturally occurring substituted amines, cyclic amines, and basic ion exchange resins. Specific examples include isopropylamine, trimethylamine, diethylamine, triethylamine, tripropylamine and ethanolamine. In some embodiments, pharmaceutically acceptable base addition salts are selected from ammonium salts, potassium salts, sodium salts, calcium salts and magnesium salts. The term "cocrystal" refers to a molecular complex derived from a variety of cocrystal formers known in the art. Unlike salts, cocrystals generally do not involve hydrogen transfer between the cocrystal and the drug, but rather involve intermolecular interactions between the cocrystal former and the drug in the crystal structure, such as hydrogen bonding, aromatic ring stacking or dispersion forces.

[0063] "Pharmaceutically acceptable carrier" or "pharmaceutically acceptable excipient" is intended to include any and all solvents, dispersion media, coatings, antibacterial and antifungal agents, isotonic and absorption delaying agents, and inert ingredients. The use of such pharmaceutically acceptable carriers or pharmaceutically acceptable excipients for active pharmaceutical ingredients is well known in the art. Unless any conventional pharmaceutically acceptable carrier or pharmaceutically acceptable excipient is incompatible with the active pharmaceutical ingredient, its use in the therapeutic compositions of the present disclosure is contemplated. Additional active pharmaceutical ingredients, such as other drugs disclosed herein, may also be incorporated into the described compositions and methods.

[0064] When using ranges in this article to describe, for example, physical or chemical properties such as molecular weight or chemical formula, all combinations and subcombinations of ranges and specific embodiments therein are intended to be included. When referring to a number or a numerical range, the use of the term "about" means that the number or numerical range mentioned is an approximation within the experimental variability (or within the statistical experimental error), and therefore the number or numerical range can vary. The variation is usually 0% to 15%, or 0% to 10%, or 0% to 5% of the quantity or numerical range. The term "comprising" (and related terms, such as "comprise / comprises" or "having" or "including") includes those embodiments, such as "consisting of" or "essentially consisting of" the embodiment of any composition of the material, method or process of the feature composition.

[0065] As used interchangeably herein, the terms "classifier" or "model" refer to a machine learning model or algorithm.

[0066] In some embodiments, the "model" is supervised machine learning. Non-limiting examples of supervised learning algorithms include, but are not limited to, logistic regression, neural networks, support vector machines, naive Bayesian algorithms, nearest neighbor algorithms, random forest algorithms, decision tree algorithms, boosted tree algorithms, multinomial logistic regression algorithms, linear models, linear regression, gradient boosting, hybrid models, hidden Markov models, Gaussian NB algorithms, linear discriminant analysis, or any combination thereof. In some embodiments, the classifier is a multinomial classifier algorithm. In some embodiments, the model is a 2-stage stochastic gradient descent (SGD) model. In some embodiments, the model is a deep neural network (e.g., a deep and wide sample-level classifier).

[0067] Neural Network. In some embodiments, the model is a neural network (e.g., a convolutional neural network and / or a residual neural network). Neural network algorithms, also known as artificial neural networks (ANNs), include convolutional and / or residual neural network algorithms (deep learning algorithms). A neural network can be a machine learning algorithm that can be trained to map an input data set to an output data set, wherein the neural network includes an interconnected node group organized into multiple layers of nodes. For example, a neural network architecture can include at least an input layer, one or more hidden layers, and an output layer. A neural network can include any total number of layers and any number of hidden layers, wherein the hidden layers serve as trainable feature extractors that allow a set of input data to be mapped to one or a set of output values. As used herein, a deep learning algorithm (DNN) can be a neural network that includes multiple hidden layers, for example, two or more hidden layers. Each layer of a neural network can contain multiple nodes (or "neurons"). A node can receive input directly from the input data or the output of a node in a previous layer, and perform a specific operation, for example, a sum operation. In some embodiments, the connection from the input to the node is associated with a parameter (e.g., a weight and / or a weighting factor). In some embodiments, a node can perform a summation operation on all input pairs x. iand the product of its associated parameters. In some embodiments, the weighted sum is offset by a deviation b. In some embodiments, a threshold or activation function f can be used to gate the output of a node or neuron, and f can be a linear or nonlinear function. The activation function can be, for example, a rectified linear unit (ReLU) activation function, a leaky ReLU activation function, or other functions such as a saturated hyperbolic tangent function, an identity function, a binary step function, a logistic function, an arcTan function, a softsign function, a parametric rectified linear unit function, an exponential linear unit function, a softPlus function, a curved identity function, a softExponential function, a Sinusoid function, a Sine function, a Gaussian function, or a sigmoid function, or any combination thereof.

[0068] Weight factors, bias values, and thresholds or other computational parameters of a neural network can be "taught" or "learned" during a training phase using one or more sets of training data. For example, the parameters can be trained using input data from a training data set and a gradient descent or back-propagation method so that the output values ​​computed by the ANN are consistent with the examples included in the training data set. The parameters can be obtained from a back-propagation neural network training process.

[0069] According to the present disclosure, any of a variety of neural networks may be applicable. Examples may include, but are not limited to, feedforward neural networks, radial basis function networks, recurrent neural networks, residual neural networks, convolutional neural networks, residual convolutional neural networks, etc., or any combination thereof. In some embodiments, machine learning utilizes pre-trained and / or transfer learning ANNs or deep learning architectures. Convolutional and / or residual neural networks may be used according to the present disclosure.

[0070] For example, a deep neural network classifier includes an input layer, a plurality of individually parameterized (e.g., weighted) convolutional layers, and an output scorer. Each convolutional layer and the parameters (e.g., weights) of the input layer contribute to a plurality of parameters (e.g., weights) associated with the deep neural network classifier. In some embodiments, at least 100 parameters, at least 1000 parameters, at least 2000 parameters, or at least 5000 parameters are associated with the deep neural network classifier. Therefore, deep neural network classifiers require the use of computers because they cannot be solved mentally. In other words, in such embodiments, given the input of the classifier, the classifier output needs to be determined using a computer, rather than by thinking. See, e.g., Krizhevsky et al., 2012, “Imagenet classification with deep convolutional neural networks”, in Advances in Neural Information Processing Systems 2, Pereira, Burges, Bottou, Weinberger, eds., pp. 1097-1105, Curran Associates, Inc.; Zeiler, 2012 “ADADELTA: an adaptive learning rate method”, CoRR, vol. abs / 1212.5701; and Rumelhart et al., 1988, “Neurocomputing: Foundations of research”, chapter Learning Representations by Back-propagating Errors, pp. 696-699, Cambridge, MA, USA: MIT Press, each of which is hereby incorporated by reference.

[0071] For example, neural network algorithms suitable for use as classifiers, including convolutional neural network algorithms, are disclosed in Vincent et al., 2010, “Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion”, J Mach Learn Res 11, pp. 3371-3408; Larochelle et al., 2009, “Exploring strategies for training deep neural networks”, J Mach Learn Res 10, pp. 1-40; and Hassoun, 1995, Fundamentals of Artificial Neural Networks, Massachusetts Institute of Technology, each of which is hereby incorporated by reference. Additional example neural networks suitable for use as classifiers are disclosed in Duda et al., 2001, Pattern Classification, 2nd Edition, John Wiley & Sons, Inc., New York; and Hastie et al., 2001, The Elements of Statistical Learning, Springer-Verlag, New York, each of which is hereby incorporated by reference. Additional example neural networks suitable for use as classifiers are also described in Draghici, 2003, Data Analysis Tools for DNA Microarrays, Chapman & Hall / CRC; and Mount, 2001, Bioinformatics: sequence and genome analysis, Cold Spring Harbor Laboratory Press, Cold Spring Harbor, New York, each of which is hereby incorporated by reference.

[0072] As used herein, the term "parameter" refers to any coefficient or similar value of an internal or external element (e.g., weight and / or hyperparameter) in an algorithm, model, regressor, and / or classifier that can affect (e.g., modify, customize, and / or adjust) one or more inputs, outputs, and / or functions in the algorithm, model, regressor, and / or classifier. For example, in some embodiments, a parameter refers to any coefficient, weight, and / or hyperparameter that can be used to control, modify, customize, and / or adjust the behavior, learning, and / or performance of an algorithm, model, regressor, and / or classifier. In some cases, a parameter is used to increase or decrease the impact of an input (e.g., feature) on an algorithm, model, regressor, and / or classifier. As a non-limiting example, in some embodiments, a parameter is used to increase or decrease the impact of a node (e.g., of a neural network), wherein the node includes one or more activation functions. The parameter assignment to a specific input, output, and / or function is not limited to any one paradigm of a given algorithm, model, regressor, and / or classifier, but can be used for any suitable algorithm, model, regressor, and / or classifier architecture to obtain the desired performance. In some embodiments, the parameter has a fixed value. In some embodiments, the value of the parameter is adjusted manually and / or automatically. In some embodiments, the value of the parameter is modified by a validation and / or training process (e.g., by error minimization and / or back propagation methods) of the algorithm, model, regressor, and / or classifier. In some embodiments, the algorithm, model, regressor, and / or classifier of the present disclosure includes multiple parameters. In some embodiments, the plurality of parameters is n parameters, wherein: n≥2; n≥5; n≥10; n≥25; n≥40; n≥50; n≥75; n≥100; n≥125; n≥150; n≥200; n≥225; n≥250; n≥350; n≥500; n≥600; n≥750; n≥1,000; n≥2,000; n≥4,000; n≥5,000; n≥7,500; n≥10,000; n≥20,000; n≥40,000; n≥75,000; n≥100,000; n≥200,000; n≥500,000, n≥1 x 106, n≥5 x 10 6 , or n ≥ 1x 10 7 . Therefore, the algorithms, models, regressors and / or classifiers of the present disclosure cannot be performed psychologically. In some embodiments, n is between 10,000 and 1x 107, between 100,000 and 5x 106, or between 500,000 and 1x106. In some embodiments, the algorithms, models, regressors and / or classifiers of the present disclosure operate in a k-dimensional space, where k is a positive integer of 5 or greater (e.g., 5, 6, 7, 8, 9, 10, etc.). Therefore, the algorithms, models, regressors and / or classifiers of the present disclosure cannot be performed psychologically.

[0073] For the avoidance of doubt, it is intended that a particular feature (e.g., integer, feature, value, use, disease, formula, compound or group) described in conjunction with a particular aspect, embodiment or example of the present disclosure be understood to be applicable to any other aspect, embodiment or example described herein, unless incompatible therewith. Therefore, such features may be used in conjunction with any definition, claim or embodiment defined herein, where appropriate. All features disclosed in this specification (including any accompanying claims, abstract and drawings) and / or all steps of any method or process so disclosed may be combined in any combination, except combinations in which at least some of the features and / or steps are mutually exclusive. The present disclosure is not limited to any details of any disclosed embodiment. The present disclosure extends to any novel feature or novel combination of features disclosed in this specification (including any accompanying claims, abstract and drawings) or to any novel step or any novel combination of steps in the steps of any method or process so disclosed.

[0074] In addition, as used herein, the term "about" means that dimensions, sizes, formulations, parameters, shapes, and other quantities and characteristics are not and need not be exact, but may be approximate and / or larger or smaller as desired, thereby reflecting tolerances, conversion factors, rounding, measurement errors, etc. and other factors known to those skilled in the art. In general, dimensions, sizes, formulations, parameters, shapes, or other quantities or characteristics are "about" or "approximate", whether or not explicitly stated as such. Note that embodiments of very different sizes, shapes, and dimensions may employ the described arrangements.

[0075] In addition, the transition terms "comprising," "consisting essentially of," and "consisting of," when used in the appended claims in both their original and amended forms, limit the scope of the claims relative to the exclusion of unrecited additional claim elements or steps, if any, from the scope of the claims. The term "comprising" is intended to be inclusive or open-ended, and does not exclude any additional, unrecited elements, methods, steps, or materials. The term "consisting of" does not include any elements, steps, or materials other than those specified in the claim, and in the latter case, does not include impurities normally associated with the specified material. The term "consisting essentially of" limits the scope of the claim to the specified elements, steps, or materials, and those elements, steps, or materials that do not materially affect the basic and novel characteristics of the claimed invention. Alternatively, all embodiments of the present invention may be more specifically defined by any of the transition terms "comprising," "consisting essentially of," and "consisting of."

[0076] Figure 1A and 1BA computer system 100 is shown for querying a library comprising combinatorial synthesis. Figure 1A and 1B In a typical embodiment, computer system 100 includes one or more computers. Figure 1A and 1B For purposes of illustration, computer system 100 is represented as a single computer that includes all of the functionality of the disclosed computer system 100. However, the present disclosure is not limited thereto. The functionality of computer system 100 may be distributed across any number of networked computers and / or reside on each of a number of networked computers and / or virtual machines. Those skilled in the art will appreciate that a variety of different computer topologies are possible for computer system 100, and all such topologies are within the scope of the present disclosure.

[0077] Considering the above content, Figure 1A and 1B , the computer system 100 includes one or more central processing units (CPUs) 64, one or more graphics processing units (GPUs) 74, a network or other communication interface 76, a user interface 68 (e.g., including an optional display 70 and an optional keyboard 72 or other form of input device), memory 58 (e.g., random access memory, persistent memory, or a combination thereof), one or more magnetic memory disk storage and / or permanent devices 60 optionally accessed by one or more controllers 62, one or more communication buses 12 for interconnecting the above components, and a power supply 66 for powering the above components. To the extent that the components of the memory 58 are not persistent, the data in the memory 58 can be seamlessly shared with the non-volatile memory 60 using known computing techniques such as caching. The memory 60 can include a large capacity memory remotely located relative to the central processing unit 64. In other words, some of the data stored in the memory 58 and / or storage 60 may actually be hosted on a computer external to the computer system 100, but can be accessed electronically by the computer system 100 using the network interface 76 over the Internet, an intranet, or other form of network or electronic cable. In some embodiments, computer system 100 utilizes models running from memory associated with one or more graphics processors 74 to increase system speed and performance. In some alternative embodiments, computer system 100 utilizes models running from memory 58 rather than memory associated with a graphics processor.

[0078] The memory 58 of the computer system 100 stores:

[0079] ● an optional operating system 78, including programs that handle various basic system services;

[0080] ●Combinatorial synthesis library query model 80, used to query the combinatorial synthesis library;

[0081] a combinatorial synthesis library 82 comprising a plurality of compounds 92, the combinatorial synthesis library representing a plurality of reaction types 82, each such reaction type being indexed by a reaction type key 84 and having a corresponding mapping to a corresponding plurality of reactants, each respective reactant 86 in each corresponding plurality of reactants having a corresponding mapping to a corresponding plurality of synthons 88, each such synthon being indexed by a synthon key 90, such that each respective compound 92 is indexed by a reaction type key 94, which in turn can be used to determine the reactant 96 and synthon key 98 of the respective compound;

[0082] a molecular query 101 for querying the combinatorial synthesis library 82, wherein the molecular query is in the form of a graph 102 comprising a plurality of nodes 102 and a plurality of edges 104, wherein each node of the plurality of nodes represents an atom, and each node is connected to another node of the plurality of nodes by at least one edge of the plurality of edges, the edge representing a covalent bond between the two nodes;

[0083] The molecular encoder model 110 optionally comprises a message passing neural network comprising a plurality of message passing layers 112 that collectively comprise a first plurality of parameters 114, the molecular encoder model responsive to receiving a molecular query 101 in the form of a graph 102, outputting a query vector by applying the plurality of parameters 114 to the graph 102;

[0084] the reaction query generator model 116 includes a second plurality of parameters 118, the reaction query generator model 116 providing, in response to receiving the query vector, a first reaction type 82-1 of the plurality of reaction types 82 as an output from the reaction query generator model 116 by applying the second plurality of parameters 118 to the query vector; and

[0085] ●The synthesized sub-query generator model 120 includes a third plurality of parameters 122. In response to receiving the identification of the reactant, the synthesized sub-query generator model 120 provides the corresponding synthon 88 as an output from the synthesized sub-query generator model 120 by applying the third plurality of parameters 122 to the corresponding reactant.

[0086] In some embodiments, one or more of the above-identified data elements or modules of the computer system 100 are stored in one or more of the previously mentioned memory devices and correspond to a set of instructions for performing the above-mentioned functions. The above-identified data, modules or programs (e.g., instruction sets) do not need to be implemented as separate software programs, processes or modules, and therefore various subsets of these modules can be combined or otherwise rearranged in various implementations. In some embodiments, the memory 60 (and optional memory 58) optionally stores a subset of the above-identified modules and data structures. In addition, in some embodiments, the memory 60 (and optional memory 58) stores additional modules and data structures not described above.

[0087] Now that a system for querying a combinatorial synthesis library has been disclosed, a method for performing such a query is described in detail with reference to FIG. 2 and discussed below.

[0088] Referring to block 200, in some embodiments, systems and methods are provided for querying a combinatorial synthesis library 82. The combinatorial synthesis library 82 includes a plurality of compounds 92 and represents a plurality of reaction types 82. Each corresponding reaction type 82 in the plurality of reaction types has a corresponding mapping to a corresponding plurality of reactants 86. Each corresponding reactant 86 in each corresponding plurality of reactants has a corresponding mapping to a corresponding plurality of synthons 88.

[0089] One aspect of the present disclosure focuses on combinatorial synthetic chemical libraries (CSLs), which can be used to represent a large amount of accessible chemical space (e.g., more than 10 8 , 10 9 , 10 10 or 10 11 compounds). Figure 3A An illustration of such a construction is provided in . In some embodiments, the CSL consists of a set of multidimensional synthesis tables. In some such embodiments, each synthesis table describes a multi-component chemical reaction, which can be indicated by an index t∈T (e.g., a natural number).

[0090] Following

[44] , the term “R-group”, or simply group, is used to refer to a placeholder group in a chemical reaction, and is denoted by r∈R. Multiple groups can be combined in a (chemically efficient) reaction, denoted by t∈T. We let ψ:T→P(R) be the set of R-groups required to return reaction t function.

[0091] Each R-group is spanned by a large number of molecular building blocks, called synthons, which can be used in the corresponding reaction. Synthons are denoted by s∈S and can be represented by the molecular diagram It is worth noting that a synthon can belong to multiple R-groups. A target molecule is represented by an R-group, which represents a potential starting reagent in the synthesis of that target molecule. For example, in Figure 6, compound 602 includes two R-groups, the electrophilic "PhCH 2+ "R-group 604 and nucleophilic "COOH - ” R-group 606. The two R-groups do not exist by themselves; the synthetic equivalents (synthons) corresponding to each R-group react to produce the desired R-group, and thus the entire compound. For example, in the illustration provided in FIG. 6 , cyanide anion 610 is “COOH - "R-group 606 is a possible synthon, while benzyl bromide 608 is "PhCH 2+ A possible synthon of R-group 604.

[0092] For convenience, in some embodiments, the symbol Used to return the set of synthetic subgroups belonging to a specific R-group r function.

[0093] The product x∈X is a molecule synthesized from a multicomponent reaction where the R-group and the corresponding synthesizer tuple where for all i=1,...,k t , (In some embodiments, k t between 2 and 4). Let f be defined as a synthesis rule that generates a compound x from a pair of reaction and synthon tuples (t,u), e.g., x:=f(t,u). In short, as a simple analogy, a synthesis rule t can be viewed as specifying k t terms, where each term is an R-group and their associated synthons correspond to the allowed values ​​of the corresponding terms.

[0094] Therefore, synthon-based libraries A molecule is characterized entirely by its reaction T, R-group R, and synthon S, as well as a synthesis rule f, a mapping of reactions to R-groups ψ, and a mapping of R-groups to synthons σ. Due to the combinatorial nature of such constructs, they can be used to construct libraries spanning large areas of readily accessible chemical space from which compounds can be obtained (i) with high probability, (ii) at reasonable cost, and (iii) with short lead times.

[0095] Using probability language, the distribution on X caused by D can be described by the following factorization:

[0096]

[0097] in is the set of all eligible synthetic subchains of reaction t. This factorization describes the generation process, where first the reaction Sampling is performed, and then valid synthetic sub-tuples are sampled Include synthons from corresponding groups in t and combine these synthons together by synthesis to form the product x (via deterministic rule f).

[0098] As written, All valid (t,u) pairs in have the same probability under p. Note that if Each product in can be obtained according to only a single synthetic route, It is available through A uniform distribution over a portion of X is achieved.

[0099] Referring to block 202, in some embodiments, the plurality of reaction types 82 includes 20 or more reaction types, and the combinatorial synthesis library 82 includes 100 or more compounds 92 for each reaction type 82 in the plurality of reaction types. In some embodiments, the plurality of reaction types 82 includes 20, 30, 40, 50, 60, 70, 80, 90, 100, 500, 1000, 2000, 3000, 5000, 10,000, or more reaction types, and the combinatorial synthesis library 82 includes 100, 200, 300, 400, 500, 1000, or 10,000 or more compounds 92 for each reaction type 82 in the plurality of reaction types.

[0100] Referring to block 204, a molecular query 101 is input into a molecular encoder model 110. The query is graph-based. In some embodiments, the molecular encoder model 101 includes a message passing neural network including a plurality of message passing layers 112 that collectively include a first plurality of parameters 114. Inputting the graph-based query into the molecular encoder model 110 generates a query vector as an output of the molecular encoder model by applying the first plurality of parameters 114 to the graph-based query.

[0101] In some embodiments, the molecular encoder model 101 (MolecularEncoder) The molecular graph Takes as input and returns d X Dimensional features represent:

[0102]

[0103] In some embodiments, the Molecular Encoder (molecular encoder model 101) is a graph neural network with a variational linear layer stacked on top of a graph readout, which is a graph given an input graph Generate samples z~q(z|x). Figure 3A -H means that z is interpreted as being introduced into the library by x Query in .

[0104] Note that in some embodiments, the molecular encoder model 110 takes a molecular graph as input and is therefore able to generate molecular graphs that are not in the library. In some embodiments, this is useful for finding analogs through a catalog—that is, compounds that can be purchased from a catalog and are chemical analogs of the query molecule.

[0105] Referring to block 206, in some embodiments, the graph-based query is a single graph 102 comprising a plurality of nodes 102 and a plurality of edges 104, wherein each node 102 of the plurality of nodes is connected to another node of the plurality of nodes by at least one edge 104 of the plurality of edges. In some embodiments, the single graph 102 represents a single molecule compound, and each node of the plurality of nodes represents an atom of the single molecule compound, and each edge of the plurality of edges represents a covalent bond between two atoms in the single molecule compound. In some embodiments, the plurality of nodes represents each non-hydrogen atom of the molecular compound. In some embodiments, the plurality of nodes represents each atom of the molecular compound. In some embodiments, the plurality of edges represents each covalent bond of the molecular compound.

[0106] In some such embodiments, each respective node in the plurality of nodes is associated with any 1, 2, 3, 4 or more of the following features: (i) a corresponding element type in a plurality of element types, (ii) a node degree in a plurality of node degrees, (iii) a hybridization in a plurality of hybridizations, (iv) a plurality of bonded hydrogens, (v) a formal charge in a set of formal charges, and (vi) a binary indication of aromaticity.

[0107] Referring to block 208, in some embodiments, each respective node 102 in the plurality of nodes 102 is associated with: (i) a corresponding element type in a plurality of element types, (ii) a node degree in a plurality of node degrees, (iii) a hybridization in a plurality of hybridizations, (iv) a plurality of bonded hydrogens, (v) a formal charge in a set of formal charges, and (vi) a binary indication of aromaticity.

[0108] Referring to block 210, in some embodiments, each respective bond 104 in the plurality of bonds is associated with: (i) a bond type, (ii) a binary indication of conjugation, (iii) a binary indication of whether the respective bond is within a ring, and (iv) an indication of stereochemistry.

[0109] In some embodiments, the atoms and bonds of the molecular compound are represented in the graph as a set of binary features. In some embodiments, these binary features are described in Tables 1 and 2, respectively. In an example embodiment, the dimensions of the nodes and edges are each set to 64. Therefore, in the example, the atom embedding consists of 50×64=3,200 parameters, and the bond embedding consists of 12×64=768 parameters.

[0110] Table 1: Atom feature types.

[0111]

[0112] Table 2: Key feature types.

[0113]

[0114] In some embodiments, a single graph represents the query molecular compound as a set of atomic features and a set of bond features. For example, in some embodiments, a single graph represents the query molecular compound as a set of atomic features and a set of bond features of Table 1 and Table 2, respectively. In some such embodiments, the set of atomic features includes element type, node degree, hybridization, chirality, bonded hydrogen, formal charge, aromaticity, and the set of bond features includes bonding time, conjugation, ring endo and stereochemistry, as described in Table 1 and Table 2. In some such embodiments, each non-hydrogen atom in the query molecular compound is represented by 500, 1000, 1500, 2000, 2500, 3000 or more than 3000 parameters in a set of atomic features, and each covalent bond in the molecular compound is represented by 100, 200, 300, 400, 500, 600, 700 or more than 700 parameters in a set of bond features.

[0115] Referring to block 212, in some embodiments, the individual graphs 102 represent single molecule compounds present in the combinatorial synthesis library. In some embodiments, the individual graphs 102 represent single molecule compounds not present in the combinatorial synthesis library.

[0116] Referring to block 214, in some embodiments, the single graph 102 represents a weighted composite of a first graph of a first molecular compound and a second graph of a second molecular compound. In some embodiments, the single graph 102 represents a composite of two or more molecular compounds.

[0117] Referring to block 216, in some embodiments, the single graph 102 represents a weighted composite of multiple graphs of a second plurality of compounds, wherein the second plurality of compounds have a common property. Referring to block 218, in some embodiments, the common property is a Tanimoto distance to each other compound in the second plurality of compounds that is less than a threshold value. Referring to block 220, in some embodiments, the common property is a binding coefficient to a macromolecular target that is less than a threshold value.

[0118] Referring to block 220, the query vector z (e.g., represented by the above equation) is input into the reaction query generator model 116 including the second plurality of parameters 118. This results in an output from the reaction query generator model 116 of a first reaction type from the plurality of reaction types by applying the second plurality of parameters 118 to the query vector.

[0119] In some embodiments according to block 220, given the query MolecularEncoder of block 204 and Library The task of the decoder is to extract The molecules that respond to the query are retrieved from the molecule - for example, in other words, the reactions and synthon tuples that will produce the molecule as a product are identified. Based on the factorization discussed previously, in some embodiments, decoding is performed by first generating the reactions.

[0120] In some embodiments, the reactive query generator model 116 is utilized. In some embodiments, the reactive query generator model 116 may be directed by an operation

[0121] q T =ReactionQueryGenerator(z). (11)

[0122] In some such embodiments, the ReactionQueryGenerator samples It generates a response query q based on the query z T With CSL In some embodiments, the operation q performed by some embodiments of the reactive query generator model 116 is T =ReactionQueryGenerator(z) is mathematically equivalent to:

[0123]

[0124] This defines the probability distribution of the reaction types in T. Sampling can be performed according to this probability distribution to achieve the reaction

[0125] Thus, referring to block 224, in some embodiments, the output from the reaction query generator model 116 is used to identify a first reaction key type 84 of a plurality of reaction key types via a first query key lookup, wherein each reaction key type of the plurality of reaction key types represents a synthesis reaction that can be used to synthesize one or more compounds 92 in the combinatorial synthesis library 82. In some such embodiments, the output from the reaction query generator model 116 is a reaction key type t. In other embodiments, the output from the reaction query generator model 116 is a key type A probability distribution of , such as the probability distribution derived from equation (12), from which a particular reaction key type t is obtained by sampling the probability distribution.

[0126] Referring to box 222, in some embodiments, the reactive query generator model 116 is a two-layer perceptron with an intermediate ReLU activation (e.g., not counting the input layer and the output layer). In some embodiments, the reactive query generator model 116 is an N-layer perceptron (MLP), where N is a positive integer of 2 or greater (e.g., 2, 3, 4, 5, 6, 7, 8, 9, or 10) and represents the number of hidden layers in the model. In some embodiments, the MLP is a type of feed-forward artificial neural network (ANN) that includes at least three layers of nodes: an input layer, one or more hidden layers, and an output layer. In such embodiments, except for the input node, each node is a neuron that uses a nonlinear activation function. In some embodiments, the activation function is a rectified linear unit ReLU activation. In some embodiments, the activation function is a rectified linear unit (ReLU) activation function, a leaky ReLU activation function, or other functions such as a saturated hyperbolic tangent function, an identity function, a binary step function, a logistic function, an arcTan function, a softsign function, a parametric rectified linear unit function, an exponential linear unit function, a softPlus function, a curved identity function, a softExponential function, a Sinusoid function, a Sine function, a Gaussian function, or a sigmoid function, or any combination thereof. Further disclosure of suitable MLPs for use as the reaction query generator model 116 in some embodiments of the present disclosure can be found in Multilayer Perceptrons: Theory and Applications, NovaScience Publishers, Hauppauge, New York, ed., 2020, which is hereby incorporated by reference herein. In some embodiments, the reaction query generator model 116 has the architecture of any model disclosed in the definition section above.

[0127] Given a sample reaction t from block 220, the required R-groups (via ψ) are known, and furthermore the synthons suitable for each group (via σ) are known. Thus, referring to block 226, in some embodiments, a corresponding synthon 88 for each corresponding reactant t in a first plurality of reactants of a first reaction type is determined from the corresponding plurality of synthons mapped to the corresponding reactant 86 by inputting the corresponding reactant 86 into a synthon query generator model 120 including a third plurality of parameters 122, thereby obtaining the corresponding synthon as an output of the synthon query generator model 120 by applying the third plurality of parameters 122 to the corresponding reactant. This produces a set of synthons, each synthon in the set of synthons corresponding to a reactant in the first plurality of reactants. In other words, to decode a synthon tuple for reaction t, the synthon query generator model 120 is implemented. In some embodiments, the synthon query generator model 120 implements the operation SynthonQueryGenerator: It is used in some embodiments of the present disclosure to query each R-group as follows Synthon:

[0128]

[0129]

[0130] Referring to box 228, in some embodiments, the synthetic sub-query generator model 120 is a two-layer perceptron with an intermediate ReLU activation. In some embodiments, the synthetic sub-query generator model 120 is an N-layer perceptron (MLP), where N is a positive integer of 2 or greater (e.g., 2, 3, 4, 5, 6, 7, 8, 9, or 10) and represents the number of hidden layers in the model. In some embodiments, the MLP is a type of feed-forward artificial neural network (ANN) that includes at least three layers of nodes: an input layer, one or more hidden layers, and an output layer. In such embodiments, except for the input node, each node is a neuron that uses a nonlinear activation function. In some embodiments, the activation function is a rectified linear unit ReLU activation. In some embodiments, the activation function is a rectified linear unit (ReLU) activation function, a leaky ReLU activation function, or other functions such as a saturated hyperbolic tangent function, an identity function, a binary step function, a logistic function, an arcTan function, a softsign function, a parametric rectified linear unit function, an exponential linear unit function, a softPlus function, a curved identity function, a softExponential function, a Sinusoid function, a Sine function, a Gaussian function, or a sigmoid function, or any combination thereof. Further disclosure of suitable MLPs for use as synthetic subquery generator models 120 in some embodiments of the present disclosure can be found in Multilayer Perceptrons: Theory and Applications, Nova Science Publishers, Hauppauge, New York, 2020, edited by Vang-mata, which is hereby incorporated by reference herein. In some embodiments, the synthetic subquery generator model 120 has the architecture of any model disclosed in the definition section above.

[0131] Referring to block 230, in some embodiments, the output of the synthesized sub-query generator model 120 is used to identify the synthesized sub-key 80 of the corresponding synthesizer via a second query key lookup.

[0132] Referring to block 232, in some embodiments, the first plurality of parameters 114 (of the molecular encoder model 110) includes 100,000 parameters, the second plurality of parameters 118 (of the reaction query generator model 116) includes 5,000 parameters, and the third plurality of parameters 122 (of the synthetic sub-query generator model 120) includes 5,000 parameters. In some embodiments, the first plurality of parameters 114 (of the molecular encoder model 110) includes 10,000, 50,000, 100,000, or 1 x 10 6parameters, the second plurality of parameters 118 (of the reactive query builder model 116) includes 1,000, 2,000, 3,000, 4,000, 5,000, or 10,000 parameters, and the third plurality of parameters 122 (of the synthetic sub-query builder model 120) includes 1,000, 2,000, 3,000, 4,000, 5,000, or 10,000 parameters.

[0133] Referring to block 234, in some embodiments, first plurality of reactants 86 includes three or more reactants, and corresponding mappings of corresponding plurality of synthons for reactants in the three or more reactants include ten or more synthons. In some embodiments, first plurality of reactants 86 includes 2, 3, 4, 5, 6, 7, 8, 9, or 10 or more reactants. In some embodiments, corresponding mappings of corresponding plurality of synthons for reactants in first plurality of reactants 86 include 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 25, 30, 35, 40, 45, 50, or 100 or more synthons.

[0134] Referring to block 236, a molecular structure (compound) 92 is identified in a combinatorial synthesis library 82 that includes a set of synthons 88 arranged according to a synthesis rule f associated with a first reaction type 82. As used herein, a synthesis rule f generates a compound x from a reaction and synthon tuple pair (t;u), e.g., x:=f(t;u).

[0135] Referring to block 238, in some embodiments, the plurality of compounds in combinatorial synthesis library 82 includes one billion or more compounds, and the identified molecular structure is any one of the one billion or more compounds that satisfy the query. Referring to block 240, in some embodiments, the plurality of compounds in combinatorial synthesis library 82 includes one trillion or more compounds, and the identified molecular structure is any one of the one trillion or more compounds that satisfy the query. In some embodiments, combinatorial synthesis library 82 includes 1 x 10 8 , 1x 10 9 , 1x 10 10 , 1x 10 11 , 1x 10 12 or 1x 10 13 Compounds.

[0136] Combinatorial synthesis library variational autoencoders.

[0137] Next, consider The task of finding the product x in , which is equivalent to finding the reaction t and the synthesizer tuple u that satisfy x = f(t,u). This can be seen as an inference problem. A variational formula is obtained for the latent variable model of x:

[0138]

[0139] where p(z) denotes the prior distribution of the latent variable. By first selecting the reaction t, and then selecting the synthon We can further simplify the joint conditional distribution of t and u to be independently conditional on t:

[0140]

[0141]

[0142] This leads to a strategy where the molecule x is encoded into a latent space z~q(z|x) and the reaction type is first decoded And then decode a synthon for each group based on the sampled reactions i=1,...,k t (where t is a positive integer of 2 or greater) to form a composite sub-tuple

[0143] The resulting latent variable model can be called a combinatorial synthesis library variational autoencoder. Figure 3B -H gives a step-by-step description of the method. The three main modules that make up this autoencoder are (i) the library encoder, (ii) the molecular encoder, and (iii) the molecular decoder. Module (ii) forms the basis of q(z|x), while modules (ii) and (iii) form basis.

[0144] Library encoder.

[0145] In some embodiments, the combinatorial synthesis library Organized hierarchically, with synthons S at the bottom of the hierarchy, R-groups R in the middle, and reactions T at the top. Here we describe a strategy for learning an associated hierarchy that describes libraries at these three levels of resolution in an end-to-end manner. These representations can then be used to retrieve query results into the library. The library encoder is described in Figure 3C -E is shown.

[0146] Starting from the bottom of the hierarchy with synthons S, in order to learn the representation of each synthon in a fully inductive way, a graph neural network is used to parameterize the SynthonEncoder: It applies a series of message passing steps followed by a read to arrive at d S - A dimensional representation of each synthon. In some embodiments, dS The dimensions in is 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, or more.

[0147] Moving up the hierarchy, a deep neural network is used to represent R-groups R by aggregating the representations of synthons belonging to a particular group. Formally, RGroupEncoder: From the composition of the synthon d S -dimensional representation set to learn the d R -dimensional representation. Following

[61] , an R-group encoder has the form RGroupEncoder In some embodiments, φ and ρ are parameterized as multilayer perceptrons, and is a permutation-invariant aggregation operator.

[0148] In some embodiments, φ and ρ are parameterized as models other than multilayer perceptrons. For example, in some embodiments, φ and ρ are parameterized as any of the models disclosed in the definition of "Model" disclosed above. In some embodiments, a higher performance set-to-vector neural network is used, such as, for example, a set transformer

[31] . However, the advantage of using a multilayer perceptron is that it allows for fast querying of a subset of the library at test time, because All s∈S can be cached, reducing the computation required to query partitions of the library. In some embodiments, mean pooling is used as an aggregation operator to focus on features of the distribution of synthons in the group (as opposed to sum pooling, which typically expresses features of multisets), which improves performance when dealing with R-groups of varying cardinality.

[0149] In some embodiments, to represent reactions, another deep neural network is used as a ReactionEncoder: The model takes as input a set of variable-sized group representations corresponding to the reactants in t and produces d T -dimensional representation. In some embodiments, summation is used as an aggregation operator to learn multiset properties of R-groups in a reaction.

[0150] Putting this together, the cascade is as follows:

[0151]

[0152] In some embodiments, factorization of likelihood is considered such that given a molecule representation z, the molecule decoder first decodes the reaction type, and then decodes one synthon per group, respectively, according to the reaction type. Thus, in some embodiments, a key vector is used for each reaction and each synthon to compare with the associated query vector. Thus, in some embodiments, a query vector having the function The ReactionKeyGenerator is used to generate a key vector for a given reaction representation, and similarly, the SynthonKeyGenerator is used: Generate a key for each synthesizer. In some embodiments, each of these key generators is parameterized as a multilayer perceptron:

[0153]

[0154] In some embodiments, each of these key generators is parameterized as some other form of model. For example, in some embodiments, each of these key generators is parameterized as any model disclosed under the definition "model" in the definition section above.

[0155] Example training.

[0156] For large CSL Encoding the entire library in each iteration of the training loop may require too much GPU memory. To overcome this problem, in some embodiments, a mini-batch strategy is used, where the number of Draw a random subset from the full library from In some embodiments, a subsampler is used to form a synthon, an R-group, and a reaction representation and a key. It (i) uniformly randomly samples a subset of reactions contained in the entire library, retaining only the R-groups contained in the sampled reactions, and (ii) for each reaction, samples a random number of products, retaining only the synthons contained in the sampled products. In some embodiments, teacher forcing is further used to feed the ground truth reactions when generating synthon queries for corresponding R-groups. Algorithm 1 below describes one embodiment of such a training procedure used in some embodiments of the present disclosure.

[0157] Algorithm 1 - Training procedure:

[0158] Import the complete library Library Sampler Batch size N, model parameters θ, KL divergence weight β ≥ 0, optimization selection

[0159] However, it did not meet the highest standards

[0160] #Prepare small batches

[0161] Library Take samples

[0162] Reaction to n = 1 ... N - Synthetic subchain pairs A small batch of samples are taken to form the corresponding products through synthesis, for n = 1...N, x n =f(t n ,u n )

[0163] #Encode the library

[0164] For each s∈S, we obtain the synthesizer representation according to equation (5):

[0165] For each r∈R, we obtain the group representation according to equation (6):

[0166] For each t∈T, the reaction expression is obtained according to equation (7):

[0167] For each t∈T, the reaction key is obtained according to equation (8):

[0168] For each s∈S, the synthetic subkey is obtained according to equation (9):

[0169] # Encode the product

[0170] For each n=1...N, sample the product query z according to equation (10) n

[0171] For each n=1…N, keep the KL divergence contribution KLD n =D KL (q(z|x n ))||P(z))

[0172] #Decode library-related artifacts

[0173] For n=1…N, the reaction query [q T ] n

[0174] For each t∈T and n=1…N, the reaction probability is obtained according to equation (12):

[0175] For every n=1…N and r∈σ(t n ), according to equation (13) we get the synthetic subquery

[0176] For each n=1…N, and According to equation (14), the synthon probability is obtained

[0177] For each n=1…N Calculating LL contribution

[0178] #Calculate loss and update parameters

[0179] Compute the loss as the negative of the weighted ELBO,

[0180] Compute gradients using the optimizer of choice And update θ, for example, θ←optimizer

[0181] End Time

[0182] Returns the fitted model parameters θ * ←θ

[0183] Post hoc density estimation.

[0184] Given a trained generative model p θ (x|z), or through (x,z)~p θ (x|z)p(z) samples the products (discarding z) (Ref. 3). However, this does not generally correspond well to the product due to the bias introduced by the batch sampling strategy outlined in this paper (first uniformly sampling the reactions, then uniformly sampling the products of a given reaction). Uniform distribution of products.

[0185] While this can be corrected for during the training phase by importance weighting, in some embodiments an alternative approach is performed by utilizing the following post hoc density estimation strategy

[14] . A large number of products are sampled and passed through the molecular encoder z~q φ (z|x) encodes these products. Then, in some embodiments, a density estimator is fitted to the clustered samples, writing q λ (z). In some embodiments, for simplicity, a multivariate normal distribution is used, but in some examples, more expressive density estimators are used herein (e.g., a mixture of multivariate normals).

[0186] In some embodiments, it can then be obtained by (x, z) ~ p θ (x|z)q λ (z) Sampling of the product, which will more closely match the This helps to correct for biases in the spatial distribution of the products caused by the choice of a batch sampling strategy.

[0187] Computational complexity, scalability, and efficiency.

[0188] In this section, the computational complexity, scalability, and efficiency of some embodiments of the presently disclosed systems and methods are described.

[0189] First, please note that the library can be encoded with O(|S|+|R|+|T|) complexity, where the constants depend on the complexity of the synthon, the R-group, and the reaction encoder. Nevertheless, This is logarithmic compared to the naive encoding of each product in Complexity.

[0190] Of more note is the computational complexity of the molecular decoder. For clarity, consider a simplified Let M i indicates the number of synthons for groups i=1,…,k. Naively speaking, The nearest neighbor lookup in will have On the other hand, in some embodiments, the currently disclosed systems and methods independently perform searches with synthons in each R-group, which achieves Complexity—logarithmic growth. Therefore, the structure of the currently disclosed molecular decoders is well suited for very large combinatorial libraries of interest in early drug discovery.

[0191] Another advantage of the disclosed decoding strategy is that it relies only minimally on autoregression. In fact, in some embodiments, only a single autoregression step is required, regardless of the size of the generated graph (the autoregression length is exactly two). Therefore, the disclosed system and method can be extended to large and variable-sized molecular graphs that follow combinatorial synthesis construction.

[0192] Furthermore, the disclosed systems and methods guarantee the generation of chemically valid, synthetically accessible molecular graphs without having to perform explicit validity checks. This compares favorably to prior work, where the validity of each candidate action is verified at each step of the autoregression and invalid actions are excluded from the selection set. While cheminformatics libraries such as RDKit

[30] have efficient C++ implementations for these checks, they still increase runtime quite significantly. Further, in the absence of explicit validity checks, these models have been shown to be able to generate invalid molecular graphs at significantly higher rates

[23] .

[0193] Conclusion.

[0194] The disclosed combinatorial synthesis library variational autoencoder is a new graph-based generative model for combinatorial synthesis library navigation. The disclosed model utilizes minimum autoregression, allowing efficient generation of large molecular graphs and improving scalability. The compounds generated by the disclosed model are chemically effective and can be obtained cost-effectively. In some embodiments, the disclosed model is a neural database that provides random access to a non-enumerable library. In the following experiments, the ability of the disclosed model in modeling very large and realistic on-demand manufacturing libraries is demonstrated, paving the way for exploring more scalable strategies for non-enumerable chemical libraries for early drug discovery.

[0195] In some embodiments, synthon lookup in the disclosed decoder scales linearly with the number of synthons in the R-group, which may present a challenge as the library continues to add many synthons per R-group. In some embodiments, this can be mitigated by more scalable query key designs [11, 26]. Furthermore, in some embodiments, due to the presence of significant R-group symmetries (e.g., as in polymers), some embodiments of the decoder include modifications that break parity, and some such modifications may not allow for equally convenient parallelization. Finally, due to its strict substitution pattern, softmax has limitations in mapping from real-valued potentials to selected probabilities

[55] . Therefore, in some embodiments, instead of using softmax, sparse or substitution-aware softmax variants are used.

[0196] SynthonEncoder. In some embodiments, the SynthonEncoder is a message passing graph neural network that takes node, edge, and graph features as input and produces updates to these features in a residual manner after each round of message passing. The following single message passing layer.

[0197] Preliminary round. A set of node features Edge Features And the graphic feature g 0 The zeros indicate that these are initial or input features; Marking instructions The result after a round of message passing.

[0198] Edge model. The edge model updates edge features as follows. For edge (i, j) ∈ ε, the current iteration The edge features, graph features, and two corresponding node features are concatenated. Then, the concatenated features are layer-normalized and processed through two layers of MLP to produce The output dimension of the final linear layer is set equal to the dimension of the side features, which allows the side features to be updated in a residual manner, e.g., Note that for edge (i, j) ∈ ε, the edge features are maintained for updates i→j and i←j respectively; these features are is the same, but is usually different for subsequent layers.

[0199] Node model. Given a node The message is formed from all its neighboring nodes by summing the associated incoming edge features, Similar to the edge model, node features, messages, and graph features are concatenated, layer normalized, and passed through two layers of MLP, where the output dimension of the final linear layer is equal to the dimension of the node features. This produces a residual update of the node features, given by Therefore, the node characteristics are given by to update.

[0200] Graphical model. Graphical features are accumulated from The message of each node in is updated as shown below. Using the now familiar layer norm + MLP design to form the graph The final linear layer has an output dimension equal to the dimension of the graph features. These messages,

[0201] Then, it is aggregated by summing and pooling to form the residual of the graph features. Thus updating the graph features accordingly, In terms of implementation, the node and graph models use a shared MLP whose output dimension is equal to the sum of the node and graph feature dimensions, and then the output is split into two parts: one part is routed to node updates, while the other part is routed to graph updates.

[0202] Putting this together. The message passing neural network applies a series of message passing layers as described above. In one embodiment according to the present disclosure, the node, edge and graph feature dimensions are set to 64 and four message passing layers are utilized. For the input graph features, a zero vector is used in some embodiments. The final graph features are used as synthon representations for generating synthon keys and cascade representations of R-groups and reactions. In total, one embodiment of a SynthonEncoder according to the present disclosure is described by 152,832 parameters.

[0203] R-group encoder. In some embodiments according to the present disclosure, the R-group encoder (RGroupEncoder) follows the design outlined in DeepSets, where each synthon representation is processed separately by an MLP, merged together (here, average merging is used to allow the network to focus on the distribution characteristics of the synthons belonging to the R-group), and the result is processed by another MLP. In some embodiments according to the present disclosure, the two MLPs are set as a two-layer network with ReLU activation between two linear operations. In some embodiments, all dimensions are 64, and therefore the RGroupEncoder utilizes a total of 16,640 parameters. In some alternative embodiments, the R-group encoder is constructed using any model disclosed in the definition section above.

[0204] Reaction Encoder. In some embodiments, the reaction encoder (ReactionEncoder) follows the same design as RGroupEncoder, except that sum pooling is utilized instead of mean pooling to allow the network to focus on multiple groups of R-groups in the reaction. In some embodiments, all dimensions are 64, and thus ReactionEncoder also utilizes 16,640 parameters in total. In some alternative embodiments, the reaction encoder is constructed using any of the models disclosed in the definition section above.

[0205] SynthonKeyGenerator. In some embodiments, the synthonkey encoder (SynthonKeyGenerator) is an MLP that generates synthonkeys from synthon representations. In some embodiments, a linear layer is used. In some embodiments, both the input and output dimensions are set to 64, and thus, in such embodiments, SynthonKeyGenerator utilizes 4,160 parameters in total. In some alternative embodiments, the synthonkey encoder is constructed using any of the models disclosed in the definitions section above.

[0206] ReactionKeyGenerator. In some embodiments, the reaction key generator (ReactionKeyGenerator) is an MLP that generates reaction keys from reaction representations. In some embodiments, a linear layer is used. In some embodiments, both the input and output dimensions are set to 64, and in some embodiments, the ReactionKeyGenerator utilizes 4,160 parameters in total. In some alternative embodiments, the reaction key generator is constructed using any of the models disclosed in the definitions section above.

[0207] Molecular encoder. In some embodiments, the molecular encoder (MolecularEncoder) generates molecular queries and utilizes the same message passing design as SynthonEncoder. In some such embodiments, nodes, edges and graph features are all set to 64, and four layers of message passing are used. However, in some embodiments, unlike SynthonEncoder, MolecularEncoder has an additional variational linear layer, which produces conditional mean and conditional logarithm variance vectors according to the graph features produced by the final message passing round. In some embodiments, the variational linear layer uses 64-dimensional graph features as input and produces 128-dimensional outputs, which are divided into mean and logarithm variance parts. In some embodiments, MolecularEncoder utilizes a total of 152,832+8,320=161,152 parameters. In some alternative embodiments, any model disclosed in the above definition part is used to build a molecular encoder.

[0208] Molecular Query Processing Network. In some embodiments, as a result of the VAE objective, the molecular queries are regularized to be close to the prior p(z). To allow the decoder to better exploit these features, in some embodiments, the molecular queries are processed by an MLP before being passed to the reactive and synthetic subquery generators. For simplicity, in some embodiments, a simple two-layer MLP with intermediate ReLU activations is chosen; all dimensions are 64, so 8,320 parameters are used. In some embodiments, any of the models disclosed in the definitions section above are used in place of such an MLP.

[0209] Reaction Query Generator. In some embodiments, the Reaction Query Generator is an MLP that generates reaction queries from molecular queries. In some embodiments, a two-layer network with intermediate ReLU activations is used. In some embodiments, all dimensions are set to 64, so the Reaction Query Generator utilizes 8,320 parameters in total. In some embodiments, any of the models disclosed in the Definitions section above are used in place of such an MLP for the Reaction Query Generator.

[0210] SynthonQueryGenerator. In some embodiments, the SynthonQueryGenerator is an MLP that generates synthetic subqueries from molecule queries, reaction representations, and R-group representations. In general, in some embodiments, these three feature types are cascaded while always using a common dimension of 64, and to keep the implementation simple, and in some embodiments, summation is used instead of cascading (which can be shown to be equivalent to cascading the weight matrices of subsequent linear layers with additional constraints). Similar to ReactionQueryGenerator, in some embodiments, a two-layer MLP with intermediate ReLU activations results in a total of 8,320 parameters for SynthonQueryGenerator. In some alternative embodiments, any of the models disclosed in the definitions section above are used in place of such an MLP for the SynthonQueryGenerator.

[0211] Examples

[0212] Example 1 - Comparison of the published model with JT-VAE and RationaleRL.

[0213] data.

[0214] To demonstrate the capabilities of the presently disclosed systems and methods on real-world combinatorial synthesis catalogs of the type employed in today's large-scale hit discovery programs, example experiments were conducted using the Enamine REadily Accessible (REAL) library, which consists of 340K synthons and over a thousand reactions. Reactions in REAL range from two to four components, and the number of synthons per R-group ranges from single digits to tens of thousands. In total, the REAL library describes a chemical space of over 16 billion commercially available compounds [4] that can be obtained at low cost in approximately three to four weeks. Advantageously, this library, used in conjunction with the presently disclosed systems and methods, can be used to reproduce example experiments and promote further research on combinatorial synthesis libraries in the machine learning community.

[0215] train.

[0216] During training, subsets of the library were sampled as follows. Of the approximately 1300 reaction types in the REAL database, 20 reactions were first randomly and uniformly sampled, followed by 100 products of each reaction, including the relevant synthons in the library subset. Thus, these library subsets each describe approximately 300K-1.5M compounds, which is significantly less than the full library containing 16 billion compounds. See Algorithm 1 above for details.

[0217] For test-time inference, decoding was performed on the full library of 16 billion compounds. This constitutes a test-time distribution shift relative to training, but the disclosed systems and methods are observed to generalize well to the full library without modification. For completeness, Figure 8 An analysis of the test time distribution shift is provided in .

[0218] As described in Algorithm 1, small subsets of the full library are sampled at each training iteration to achieve tractability. In particular, the library subsampler utilizes the first sample uniformly across reactions and subsequently randomly samples a constant number of products in each reaction; the synthons associated with these sampled products constitute the synthons in the library subset. When training the CSLVAE model used in this experiment, 20 reactions and 100 products per reaction were sampled, resulting in a small batch of 2000 products. As described in the paper, these relevant library subsets each describe chemical space of approximately 300K-1.5M compounds. However, at test time, decoding is performed based on the full library of 16B compounds, which constitutes a rather drastic test-time distribution shift. Figure 8 The magnitude of the test time distribution shift (measured by the average likelihood) of the reconstruction quality is shown. Starting with sampling of five library subsets as described above, 2000 compounds are drawn for each subset; each of the five plots is represented by a different line in the plot. The library is then further increased by first including all synthons in the 20 sampled reactions and calculating the average reconstruction likelihood of the same 2000 products based on the expanded library, and by adding all synthons of 100 randomly sampled additional reactions, and then adding all synthons of another 100 randomly sampled additional reactions, and finally including all synthons and all reactions. At each step, the library subset describes an increasingly larger chemical space. It is observed that the average likelihood of the library size decreases in a roughly linear-logarithmic manner due to this distribution shift.

[0219] Molecular reconstruction and generation.

[0220] An embodiment of the disclosed system and method (CSLVAE) is compared with two existing molecular graph generation models: JT-VAE

[23] and RationaleRL

[25] . All three models are trained from scratch on the Enamine REAL library.

[0221] In JT-VAE, the molecular graph is represented by a connection tree over chemical fragments. The decoding process first generates a connection tree in a depth-first manner, placing the fragments in each node, and then subsequently orienting the fragments to match the attachment points. On the other hand, RationaleRL takes the starting rationale as input. The goal of the decoder is to complete the molecule in an autoregressive manner (one graph edit per step). In this example, a product is taken from the library and all synthons except one are deleted, and the resulting graph is used as the starting rationale. Therefore, in this example, the task of RationaleRL is to generate the missing synthons.

[0222] Table 3 summarizes the main findings of this exercise. First, it is worth noting that, due to the inductive nature of the library encoder, the implementation of the disclosed systems and methods used has approximately 10 times fewer parameters than the two alternatives considered. All three approaches achieved 100% chemical validity, but the disclosed systems and methods achieved this result without explicit validity checks. The average likelihood is calculated by taking the average of the likelihood of reconstruction of each compound over a large number of products sampled from the library. This is a measure of the model's ability to reconstruct the complete molecular graph (e.g., on average, how likely it is to reproduce the query molecule through the decoder), and can also be loosely interpreted as a measure of coverage / reachability (e.g., what percentage of the library it is able to faithfully cover). Finally, the example experiments highlight the challenges faced by existing graph generation models when applied to very large combinatorial synthesis libraries, namely that they have difficulty reliably generating in-library compounds. For JT-VAE, less than 1 out of 34 compounds was found in REAL. On the other hand, RationaleRL only generates in-library completions in about half of the cases (see Figure 7 ), but its advantage is that it provides a starting rationale in the form of compounds that remove all but one synthon from a REAL. In contrast, the disclosed systems and methods ensure by design that the results come from a library.

[0223] Table 3: Comparison of RationaleRL, JT-VAE, and CSLVAE on synthon-based generative modeling

[0224]

[0225] Dataset preparation.

[0226] The Enamine REAL library is a combinatorial synthesis library of approximately 1300 reaction types, ranging from 2- to 4-component reactions, and approximately 340K synthons. In total, the REAL library describes a chemical space of over 16B compounds made on demand. For the two baseline models currently compared, memory issues arise when using the author-provided code on the full REAL library (at the vocabulary generation step and when writing products to disk). Therefore, the two baselines are trained and evaluated on a subset of the full REAL library. CSLVAE (the model of the present disclosure) does not face memory limitations and is trained and evaluated on the full library. Therefore, RationaleRL and JT-VAE can be compared on a per-project basis (each trained on the same subset of REAL), while the results of CSLVAE reflect training on the larger and more diverse full REAL library. RationaleRL and JT-VAE were not developed for searching combinatorial synthesis libraries, and they are used in this comparison as an attempt to compare the disclosed method (CSLVAE) with the application of a state-of-the-art graphical generation model that is readily available and unmodified.

[0227] To construct the data for RationaleRL and JT-VAE training, 1300 reaction types were ranked according to the number of products contained in each reaction, and 50 medium-sized reactions were selected. These reactions describe a total chemical space of 125M compounds. Products were sampled from each reaction so that all synthons were represented, totaling a training set of ~500K compounds.

[0228] RationaleRL details.

[0229] We exploit the pre-training phase of RationaleRL, which trains a graph-based variational autoencoder that seeks to reconstruct the molecular graph from the starting fundamentals (and does not require RL, as the name implies; this is part of the fine-tuning phase and is not used here). Given the starting fundamentals and the complete molecule, RationaleRL’s decoder completes the autoregression of the molecule in an atom-by-atom, bond-by-bond manner. To form the starting fundamentals, take the product of REAL and remove all but one synthon. Thus, the task of RationaleRL is to complete the missing synthons given the complete molecule (as input to the encoder) and the starting fundamentals (as input to the decoder, along with the latent code). Figure 7It is shown that the probability of a completion produced by RationaleRL producing an artifact from the library decreases dramatically as the number of different reactions in training increases; after including 50 reactions, less than half of the completions produce an artifact from the REAL library. These experiments are based on code provided by the authors and available at https: / / github.com / wengong-jin / multiobj-rationale.

[0230] JT-VAE details.

[0231] JT-VAE is a graph-based variational autoencoder that generates molecules from a tree-like structure of fragments. Unlike RationaleRL, JT-VAE samples chemical fragments in an autoregressive manner, rather than atoms or bonds. Before training, the vocabulary generation step described in the JT-VAE paper was applied to sample products from the REAL subset used in the baseline experiment, which produced a total of 325 fragments. This experiment is based on code provided by the authors and is available at https: / / github.com / wengong-jin / icml18-jtnn Found in.

[0232] Details of the public model (CSLVAE) are available.

[0233] During training, we used an annealing schedule of β

[49] (see Algorithm 1), starting from β = 0 and increasing by 1e-5 every 2000 iterations, to a maximum of β = 1. A total of 200K iterations were trained, during which CSLVAE had observed a total of 2000 200K = 400M compounds (although no 400M unique compounds were observed due to the batch sampling strategy). Therefore, when training stopped, CSLVAE had seen no more than 2.5% of the entire REAL library.

[0234] Example 2 - Latent space visualization.

[0235] The latent space learned by the disclosed systems and methods was qualitatively checked. It is interesting to verify whether the proposed model has learned a latent space that varies relatively smoothly over the covered chemical space (e.g., small perturbations to the query cause only minor edits to the resulting molecular graph). Two checks were performed: latent space interpolation and local neighborhood visualization.

[0236] Figure 4A An example of interpolation contained in latent space. In particular, the molecular queries for the starting compound (top left) and the target compound (bottom right) are interpolated in raster order. The molecules immediately adjacent to the starting and target compounds are the related reconstructions. The products are decoded against the full REAL library of 16 billion compounds. Figure 4ABelow each molecule is its Tanimoto similarity to the starting and target molecules, respectively (5). It is observed that the interpolation passes through regions of chemical space that have decreasing (increasing) similarity relative to the starting (vs. target) compound.

[0237] Figure 4B Visualization of the latent space around randomly sampled products from the REAL library. Following previous work

[29] , random 2D planes are formed in the high-dimensional latent space by sampling in two random directions around a molecular query (central compound) and decoding the resulting products using the argmax decision rule. It is observed that the latent space is smooth as the molecules are gradually deformed, only minor edits are made when the motions in the latent space are small (e.g., one synthon at a time, modifications to smaller functional groups), and the molecular scaffolds are generally locally conserved.

[0238] Example 3 - Simulated retrieval via automatic encoding.

[0239] The disclosed system and method are used to find analogs of a query compound in a large CSL. Figure 5A and 5B In the disclosed model of the present disclosure, there are two types of molecules: molecules in the library ( Figure 5A ) and molecules not in the library ( Figure 5B ). Figure 5A Each query molecule of (top left) and Figure 5B Each query molecule (top left) is analyzed using the disclosed model (e.g., Figure 1A , 1B 3) were encoded and randomly decoded fifteen times. Thus, given a molecular query, conditionally random completions were generated from the decoder. In both cases, the autoencoder retrieved highly similar compounds as measured by Tanimoto similarity. For the in-library examples, the model was able to successfully retrieve the query compound ( Figure 5A ). In addition, it was observed that in many cases the compounds returned by the autoencoder had the same reaction type as the query molecule but differed in one or two synthons. For examples outside the library, the disclosed systems and methods found scaffold-related compounds with high Tanimoto similarity, demonstrating the utility of this approach for rapid analog searches in large (non-enumerable) libraries.

[0240] Example 4 - Post hoc density estimation

[0241] The use of post hoc density estimation for the latent code corresponding to the product of uniform sampling from the library is disclosed above as a way to force the random samples from the CSLVAE to more closely track the uniform random sampling from the library. Fig. 9 Algorithm 2 in summarizes this process.

[0242] The experiments performed show that Algorithm 2 achieves the expected results. The experiments were performed as follows. 10,000 molecules were sampled uniformly at random from the library and considered as a training set for the post hoc density estimator. Three density estimators of increasing expressiveness were considered: a multivariate normal distribution (MVNormal), a mixture of five normal distributions (MoG-5), and a mixture of ten normal distributions (MoG-10). Comparison of standard sampling methods for isotropic multivariate normal priors. Therefore, this experiment used four alternatives for sampling latent codes that were subsequently decoded into compounds in the library. Another sample of 10,000 molecules was then sampled from this library, also uniformly at random, as a reference set for comparison.

[0243] Table 4: Divergence between the query sampling strategy and a reference set of compounds uniformly randomly sampled from the library.

[0244]

[0245] Following

[41] , the set of molecules generated above was compared to the reference set of molecules on the following computable molecular properties: synthetic accessibility (SA), quantitative estimate of drug-likeness (QED), molecular weight (MW), and logarithm of the octanol-water partition coefficient (logP). In particular, the Jensen-Shannon distance (square root of the Jensen-Shannon divergence) was calculated between the distributions of these properties for the reference set and each of the discussed sets. Table 4 summarizes the results of this exercise. Three points are worth noting: (a) the reference compounds and the compounds sampled to train the post hoc density estimators have low divergences on various properties, (b) sampling from the prior yields compounds with high divergences relative to uniform sampling in the library, and (c) using more expressive density estimators for the latent codes leads to increasingly lower divergences with the reference set on various properties because they are able to better match the distribution of latent codes in the training set (which is interchangeable with the reference set).

[0246] Example 5 - Comparison with existing simulation enumeration methods

[0247] In Example 5, an attempt was made to compare the analogy capabilities of CSLVAE with those of Arthor, a state-of-the-art commercial similarity search tool for synthetic libraries developed by NextMove Software. Arthor performs analog enumeration using a custom ECFP4 bit vector representation of the molecule, returning compounds with high Tanimoto similarity based on this fingerprint. For a given query compound, the top 100 analogs returned by Arthor from REAL were enumerated. Similarly, each query compound was encoded with the CSLVAE encoder and 100 corresponding random decodings were generated. For each analog, its RDKit ECFP4 Tanimoto similarity was calculated with the query compound, and the top 1 analog was retained. The distribution of Tanimoto similarities of the top 1 analogs returned in this way was compared between Arthor and CSLVAE. As a control, a simple random baseline strategy was used, randomly sampling 100 compounds from REAL, and the top 1 analog was selected again based on RDKit ECFP4 Tanimoto similarity. Since both CSLVAE and random baseline constitute random strategies, this process was repeated 30 times for each query compound and the top-1 Tanimoto similarity was averaged.

[0248] For query compounds, 24 of the 51 novel drugs approved by the FDA in 2021 were used to screen out drugs that do not meet conventional small molecule standards (such as monoclonal antibodies); the compounds used in this experiment are Fig.10 shown. Fig.11 Box plots of top-1 Tanimoto similarities for Arthor, CSLVAE, and random baselines are shown.

[0249] Although CSLVAE found more distant analogs of ECFP4 compared to Arthor (the gold standard for fingerprint similarity searches), it is still able to routinely identify analogs of unknown novel drugs. In this regard, a Tanimoto similarity threshold in the range of 0.3-0.35 is commonly used to indicate whether a pair of molecules can be considered analogs.

[0250] It is particularly noteworthy that CSLVAE is much less resource-intensive than Arthor, which is designed for fast and efficient fingerprint-based simulated enumeration and requires appropriate infrastructure and setup. For these reasons, it is difficult to make a direct apples-to-apples comparison of computational requirements. However, the following relevant information is derived from this experiment. The CSLVAE experiment was run on a machine equipped with an NVIDIA Tesla K80 GPU and an Intel Xeon E5-2686CPU. On average, using this setup (encoding and decoding), 100 simulated samplings of a given query using CSLVAE completed in 11.41 seconds. In addition, all parameters and buffers of the trained CSLVAE model (including representations and keys for synthons, R-groups, and reactions) require only 170MB of memory. In contrast, the Arthor experiment was run in a distributed manner on 100 pods, each equipped with 8 CPUs, and the first 100 simulated enumerations took a total of 32.46 seconds, for a total CPU time of approximately 25,968 seconds. Additionally, Arthor requires a fairly large amount of memory and storage to perform simulated enumeration; our internal setup uses 3GB per shard and issues a 64GB RAM request for each Arthor worker.

[0251] We note that CSLVAE is able to represent large CSLs with significantly fewer resources and perform analog retrieval with significantly improved execution time (potentially three orders of magnitude faster), thanks to its decoding strategy that exploits parallel synthon lookup, requiring multiple keys commensurate with the number of synthons in the library rather than the number of products in the library, and also allows a similarity search that scales in time logarithmically (rather than linearly) with the number of products in the library.

[0252] Example 6 - Encoder conversion to molecular property prediction

[0253] The CSLVAE training objective can be viewed as a comparative pretext task that seeks to align the representation of a given molecule with the representation corresponding to the extracted instructions in CSL. Therefore, it is natural to wonder whether the encoder of a trained CSLVAE model can show good transfer performance on prediction tasks of possible interest. To investigate, an MLP trained on CSLVAE is compared with an MLP trained on molecular fingerprints (ECFP4 and ECFP6) for the task of molecular property prediction. The octanol-water partition coefficient (logP) and the quantitative estimate of drug-likeness (QED) [6] are used as prediction objectives.

[0254] For this experiment, the dataset was constructed by uniformly randomly sampling 100K compounds from REAL, dividing these examples into training, validation, and test folds using an 80-10-10 split. For each compound, its CSLVAE query was extracted as its feature descriptor in addition to its ECFP4 and ECFP6 fingerprints. Using the training fold, the MLP was fit to each feature descriptor separately to predict the logP and QED scores of the molecule. The iteration that achieved the lowest validation RMSE was selected and its test RMSE was recorded. This was repeated five times and the average test RMSE and standard deviation are reported. To demonstrate the extent to which CSLVAE learns molecular features that predict such molecular properties of out-of-domain compounds, the exercise was repeated on a dataset of 250K molecules from ZINC. Table 5 summarizes the results of this exercise.

[0255] Table 5: Encoder transformations based on logP and QED predictions. Cells report the mean RMSE ± 1 standard deviation, calculated from five runs.

[0256]

[0257] The results of this experiment confirm that the latent space learned by CSLVAE can indeed be used to successfully predict quantities like logP and QED, especially in the in-domain case, where it outperforms predictors fitted to chemical fingerprints. However, in the out-of-domain case, predictors fitted on ECFP6 fingerprints perform significantly better than those fitted on CSLVAE queries, suggesting that the features learned by CSLVAE may have lost some relevant predictive information about the input molecules that is significantly different from the CSL on which it was trained.

[0258] Example 7 - Overview of the CSLVAE architecture used in the example.

[0259] Table 6 summarizes the number of parameters used in the example CSLVAE architecture of the examples of the present disclosure.

[0260] Table 6: Overview of the CSLVAE architecture used in the examples

[0261]

[0262]

[0263] References:

[0264] [1] Atanu Acharya, Rupesh Agarwal, Matthew B Baker, Jerome Baudry, Debsindhu Bhowmik, Swen Boehm, Kendall G Byler, SY Chen, Leighton Coates, Connor J Cooper et al. Supercomputer-based ensemble docking drug discovery pipeline with application to COVID-19. Journal of Chemical Information and Modeling, 60(12):5832-5852, 2020.

[0265] [2] Josep Arús-Pous, Thomas Blaschke, Silas Ulander, Jean-Louis Reymond, Hongming Chen and Ola Engkvist. Exploring the GDB-13 chemical space using deep generative models. Journal of Cheminformatics, 11(1):1-14, 2019.

[0266] [3] Dávid Bajusz, Anita Rácz and Károly Héberger. Why is Tanimoto index an appropriate choice for fingerprint-based similarity calculations?Journal of Cheminformatics, 7(1):1-13, 2015.

[0267] [4] Louis Bellmann, Patrick Penner and Matthias Rarey. Topological similarity search in large combinatorial fragment spaces. Journal of Chemical Information and Modeling, 61(1):238-251, 2020.

[0268] [5]Andreas Bender and Robert C Glen. Molecular similarity: a key technique in molecular informatics. Organic and Biomolecular Chemistry, 2(22): 3204 - 3218, 2004。

[0269] [6]G Richard Bickerton, Gaia V Paolini, Jérémy Besnard, Sorel Muresan and Andrew L Hopkins. Quantifying the chemical beauty of drugs. Nature Chemistry, 4(2): 90 - 98, 2012。

[0270] [7]Lorenz C Blum and Jean-Louis Reymond. 970 million druglike small molecules for virtual screening in the chemical universe database GDB - 13. Journal of the American Chemical Society, 131(25): 8732 - 8733, 2009。

[0271] [8]John Bradshaw, Brooks Paige, Matt J Kusner, Marwin Segler and José Miguel Hernández - Lobato. A model to search for synthesizable molecules. Advances in Neural Information Processing Systems, 32, 2019。

[0272] [9] John Bradshaw, Brooks Paige, Matt J Kusner, Marwin Segler, and José Miguel Hernández-Lobato. Barking up the right tree: an approach to search over molecule synthesis DAGs. Advances in Neural Information Processing Systems, 33: 6852-6866, 2020。

[0273]

[10] Hongming Chen. Can generative-model-based drug design become a new normal in drug discovery?Journal of Medicinal Chemistry, 65(1): 100-102, 2021。

[0274]

[11] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020。

[0275]

[12] Hanjun Dai, Yingtao Tian, Bo Dai, Steven Skiena, and Le Song. Syntax-directed variational autoencoder for structured data. arXiv preprint arXiv:1802.08786, 2018。

[0276]

[13] Wenhao Gao and Connor W Coley. The synthesizability of molecules proposed by generative models. Journal of Chemical Information and Modeling, 60(12):5714-5723, 2020。

[0277]

[14] Partha Ghosh, Mehdi SM Sajjadi, Antonio Vergari, Michael Black and Bernhard From variational to deterministic autoencoders. arXiv preprint arXiv:1903.12436, 2019。

[0278]

[15] Paweł Gniewek, Bradley Worley, Kate Stafford, Henry van den Bedem and Brandon Anderson. Learning physics confers pose-sensitivity in structure-based virtual screening. arXiv preprint arXiv:2110.15459, 2021。

[0279]

[16] Rafael Gómez-Bombarelli, Jennifer N Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez-Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D Hirzel, Ryan P Adams and Alán Aspuru-Guzik. Automatic chemical design using a data-driven continuous representation of molecules. ACS Central Science, 4(2):268-276, 2018。

[0280]

[17] Christoph Gorgulla, Andras Boeszoermenyi, Zi-Fu Wang, Patrick D Fischer, Paul W Coote, Krishna M Padmanabha Das, Yehor S Malets, Dmytro S Radchenko, Yurii S Moroz, David A Scott et al. An open-source drug discovery platform enables ultra-large virtual screens. Nature, 580(7805), 2020.

[0281]

[18] David E Graff, Eugene I Shakhnovich and Connor W Coley. Accelerating high-throughput virtual screening through molecular pool-based active learning. Chemical Science, 12(22):7866 - 7881, 2021.

[0282]

[19] Ryan-Rhys Griffiths and José Miguel Hernández-Lobato. Constrained bayesian optimization for automatic chemical design using variational autoencoders. Chemical Science, 11(2):577 - 586, 2020.

[0283]

[20] Oleksandr O Grygorenko, Dmytro S Radchenko, Igor Dziuba, Alexander Chuprina, Kateryna E Gubina and Yurii S Moroz. Generating multibillion chemical space of readily accessible screening compounds. Iscience, 23(11), 2020.

[0284]

[21] Julien Horwood and Emmanuel Noutahi. Molecular design in synthetically accessible chemical space via deep reinforcement learning. ACS Omega, 5(51):32984 - 32994, 2020。

[0285]

[22] John J Irwin and Brian K Shoichet. Docking screens for novel ligands conferring new biology: Miniperspective. Journal of Medicinal Chemistry, 59(9):4103 - 4120, 2016。

[0286]

[23] Wengong Jin, Regina Barzilay and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation. International Conference on Machine Learning, 2018。

[0287]

[24] Wengong Jin, Regina Barzilay and Tommi Jaakkola. Hierarchical generation of molecular graphs using structural motifs. International Conference on Machine Learning, pp. 4839 - 4848, 2020。

[0288]

[25] Wengong Jin, Regina Barzilay and Tommi Jaakkola. Multi - objective molecule generation using interpretable substructures. International Conference on Machine Learning, 2020。

[0289]

[26] Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. Reformer: the efficient transformer. arXiv preprint arXiv:2001.04451, 2020.

[0290]

[27] Xiangzhe Kong, Zhixing Tan, and Yang Liu. Graphpiece: Efficiently generating high-quality molecular graph with substructures. arXiv preprint arXiv:2106.15098, 2021.

[0291]

[28] Mario Krenn, Florian

[28] Akshat Kumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik. Self-referencing embedded strings (selfies): A 100% robust molecular string representation. Machine Learning: Science and Technology, 1(4), 2020.

[0292]

[29] Matt J Kusner, Brooks Paige, and José Miguel Hernández-Lobato. Grammar variational autoencoder. International Conference on Machine Learning, 2017.

[0293]

[30] Greg Landrum. RDKit: Open-source cheminformatics, 2006.

[0294]

[31] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. International Conference on Machine Learning, pp. 3744-3753, 2019.

[0295]

[32] Qi Liu, Miltiadis Allamanis, Marc Brockschmidt, and Alexander Gaunt. Constrained graph variational autoencoders for molecule design. Advances in Neural Information Processing Systems, 31, 2018.

[0296]

[33] Jiankun Lyu, Sheng Wang, Trent E Balius, Isha Singh, Anat Levit, Yurii S Moroz, Matthew J O'Meara, Tao Che, Enkhjargal Algaa, Kateryna Tolmachova, et al. Ultra-large library docking for discovering new chemotypes. Nature, 566(7743): 224-229, 2019.

[0297]

[34] Krzysztof Maziarz, Henry Jackson-Flux, Pashmina Cameron, Finton Sirockin, Nadine Schneider, Nikolaus Stiefl, Marwin Segler, and Marc Brockschmidt. Learning to extend molecular scaffolds with structural motifs. arXiv preprint arXiv:2103.03864, 2021.

[0298]

[35] Rocío Mercado, Tobias Rastemo, Edvard Günter Klambauer, Ola Engkvist, Hongming Chen and Esben Jannik Bjerrum. Graph networks for molecular design. Machine Learning: Science and Technology, 2(2):025023, 2021。

[0299]

[36] Joshua Meyers, Benedek Fabian and Nathan Brown. De novo molecular design and generative models. Drug Discovery Today, 26(11):2707 - 2715, 2021。

[0300]

[37] Akshat Kumar Nigam, Robert Pollice, Mario Krenn, Gabriel dos Passos Gomes and Alan Aspuru - Guzik. Beyond generative models: superfast traversal, optimization, novelty, exploration and discovery (stoned) algorithm for molecules using selfies. Chemical Science, 12(20):7079 - 7090, 2021。

[0301]

[38] Hakime Arzucan and Elif Ozkirimli. DeepDTA: deep drug - target binding affinity prediction. Bioinformatics, 34(17):i821 - i829, 2018。

[0302]

[39] Joseph M Paggi, Julia A Belk, Scott A Hollingsworth, Nicolas Villanueva, Alexander S Powers, Mary J Clark, Augustine G Chemparathy, Jonathan E Tynan, Thomas K Lau, Roger K Sunahara et al. Leveraging nonstructural data to predict structures and affinities of protein-ligand complexes. Proceedings of the National Academy of Sciences, 118(51), 2021.

[0303]

[40] Pavel Polishchuk. Control of synthetic feasibility of compounds generated with CReM. Journal of Chemical Information and Modeling, 60(12):6074 - 6080, 2020.

[0304]

[41] Daniil Polykovskiy, Alexander Zhebrak, Benjamin Sanchez-Lengeling, Sergey Golovanov, Oktai Tatanov, Stanislav Belyaev, Rauf Kurbanov, Aleksey Artamonov, Vladimir Aladinskiy, Mark Veselov et al. Molecular sets(MOSES): a benchmarking platform for molecular generation models. Frontiers in Pharmacology, 11:1931, 2020.

[0305]

[42] Mariya Popova, Mykhailo Shvets, Junier Oliva, and Olexandr Isayev. Molecular rnn: Generating realistic molecular graphs with optimized properties. arXiv preprint arXiv:1905.13372, 2019.

[0306]

[43] Matthew Ragoza, Joshua Hochuli, Elisa Idrobo, Jocelyn Sunseri, and David Ryan Koes. Protein-ligand scoring with convolutional neural networks. Journal of Chemical Information and Modeling, 57(4):942 - 957, 2017.

[0307]

[44] Arman A Sadybekov, Anastasiia V Sadybekov, Yongfeng Liu, Christos Iliopoulos-Tsoutsouvas, Xi-Ping Huang, Julie Pickett, Blake Houser, Nilkanth Patel, Ngan K Tran, Fei Tong, et al. Synthon-based ligand discovery in virtual libraries of over 11 billion compounds. Nature, 601(7893):452 - 459, 2022.

[0308]

[45] Robert Schmidt, Raphael Klein, and Matthias Rarey. Maximum common substructure searching in combinatorial make-on-demand compound spaces. Journal of Chemical Information and Modeling, 2021.

[0309]

[46] Marwin HS Segler, Thierry Kogej, Christian Tyrchan, and Mark PWaller. Generating focused molecule libraries for drug discovery with recurrent neural networks. ACS Central Science, 4(1): 120 - 131, 2018.

[0310]

[47] Brian K Shoichet. Virtual screening of chemical libraries. Nature, 432(7019): 862 - 865, 2004.

[0311]

[48] Martin Simonovsky and Nikos Komodakis. Graphvae: Towards generation of small graphs using variational autoencoders. International Conference on Artificial Neural Networks, 2018.

[0312]

[49] Casper Kaae Tapani Raiko, Lars Kaae and Ole Winther. Ladder variational autoencoders. Advances in Neural Information Processing Systems, 29, 2016.

[0313]

[50] Tiago Sousa, Correia, Vítor Pereira, and Miguel Rocha. Generative deep learning for targeted compound design. Journal of Chemical Information and Modeling, 61(11): 5343 - 5361, 2021.

[0314]

[51] Kate Stafford, Brandon M Anderson, Jon Sorenson, and Henry van den Bedem. AtomNet PoseRanker: Enriching ligand pose quality for dynamic proteins in virtual high-throughput screens. Journal of Chemical Information and Modeling, 62(5):1178 - 1189, 2022.

[0315]

[52] Hannes Octavian-Eugen Ganea, Lagnajit Pattanaik, Regina Barzilay, and Tommi Jaakkola. Equibind: Geometric deep learning for drug binding structure prediction. arXiv preprint arXiv:2202.05146, 2022.

[0316]

[53] Kosuke Takeuchi, Ryo Kunimoto, and Jürgen Bajorath. R-group replacement database for medicinal chemistry. Future Science OA, 7(8):FSO742, 2021.

[0317]

[54] Xiaochu Tong, Xiaohong Liu, Xiaoqin Tan, Xutong Li, Jiaxin Jiang, Zhaoping Xiong, Tingyang Xu, Hualiang Jiang, Nan Qiao, and Mingyue Zheng. Generative models for de novo drug design. Journal of Medicinal Chemistry, 64(19):14011 - 14027, 2021.

[0318]

[55] Kenneth E Train. Discrete choice methods with simulation. Cambridge University Press, 2009.

[0319]

[56] Masashi Tsubaki, Kentaro Tomii, and Jun Sese. Compound-protein interaction prediction with end-to-end learning of neural networks for graphs and sequences. Bioinformatics, 35(2): 309-318, 2019.

[0320]

[57] Izhar Wallach, Michael Dzamba, and Abraham Heifets. AtomNet: a deep convolutional neural network for bioactivity prediction in structure-based drug discovery. arXiv preprint arXiv:1510.02855, 2015.

[0321]

[58] Wendy A Warr, Marc C Nicklaus, Christos A Nicolaou, and Matthias Rarey. Exploration of ultra-large compound collections for drug discovery. Journal of Chemical Information and Modeling, 62(9): 2021-2034, 2022.

[0322]

[59] Robin Winter, Frank Noé, and Djork-Arné Clevert. Permutation-invariant variational autoencoder for graph-level representation learning. Advances in Neural Information Processing Systems, 34, 2021.

[0323]

[60] Jiaxuan You, Bowen Liu, Zhitao Ying, Vijay Pande, and JureLeskovec. Graph convolutional policy network for goal-directed molecular graph generation. Advances in Neural Information Processing Systems, 31, 2018.

[0324]

[61] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep Sets. Advances in NeuralInformation Processing Systems, 30, 2017.

[0325] in conclusion

[0326] For the purpose of explanation, the foregoing description has been described with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the embodiments to the precise forms disclosed. In view of the above teachings, many modifications and variations are possible. These embodiments are selected and described in order to best explain these principles and their practical applications, thereby enabling other persons skilled in the art to best use these embodiments and various embodiments with a variety of modifications suitable for the intended specific use.

Claims

1. A computer system for querying a combinatorial synthesis library comprising a plurality of compounds, wherein The combinatorial synthesis library represents a variety of reaction types, Each respective reaction type in the plurality of reaction types has a corresponding mapping to a corresponding plurality of reactants, and Each respective reactant in each corresponding plurality of reactants has a corresponding mapping to a corresponding plurality of synthons, the computer system comprising: one or more central processing units; one or more graphics processors, wherein each of the one or more graphics processors includes 100 or more cores; as well as a memory addressable by the one or more central processing units, the memory storing at least one program for execution at least in part by the one or more graphics processors, the at least one program comprising instructions for: (A) inputting a query into a molecular encoder model, wherein the query is a single graph, wherein the molecular encoder model comprises a message passing neural network, the message passing neural network comprising a plurality of message passing layers that collectively comprise a first plurality of parameters, thereby obtaining a query vector by applying the first plurality of parameters to the single graph; (B) inputting the query vector into a reaction query generator model including a second plurality of parameters, thereby obtaining a first reaction type from the plurality of reaction types as an output of the reaction query generator model by applying the second plurality of parameters to the query vector; (C) determining a corresponding synthon for each corresponding reactant in a first plurality of reactants of the first reaction type from among the corresponding plurality of synthons mapped to the corresponding reactants by inputting the corresponding reactant into a synthon query generator model including a third plurality of parameters, thereby obtaining a corresponding synthon as an output of the synthon query generator model by applying the third plurality of parameters to the corresponding reactant, thereby determining a set of synthons, each synthon in the set of synthons corresponding to a reactant in the first plurality of reactants; and (D) identifying molecular structures in the combinatorial synthesis library, the combinatorial synthesis library comprising the set of synthons arranged according to the synthesis rules associated with the first reaction type.

2. The computer system of claim 1 , wherein the reactive query generator model is a two-layer perceptron with an intermediate ReLU activation.

3. The computer system of claim 1 , wherein the synthetic subquery generator model is a two-layer perceptron with an intermediate ReLU activation.

4. The computer system of claim 1, wherein The first plurality of parameters comprises 100,000 parameters, The second plurality of parameters includes 5,000 parameters, and The third plurality of parameters includes 5,000 parameters.

5. The computer system of claim 1, wherein The single graph includes a plurality of nodes and a plurality of edges, and Each node of the plurality of nodes is connected to another node of the plurality of nodes through at least one edge of the plurality of edges.

6. The computer system of claim 5, wherein each of the plurality of nodes is associated with: (i) a corresponding element type among a plurality of element types, (ii) the node degree among multiple node degrees, (iii) Hybridization within multiple hybridizations, (iv) multiple bonded hydrogens, (v) a formal charge in a set of formal charges, and (vi) Binary indication of aromaticity.

7. A computer system according to claim 5 or 6, wherein each respective key of the plurality of keys is associated with: (i) key type, (ii) binary indication of conjugation, (iii) a binary indication of whether the corresponding bond is within a ring, and (iv) Indication of stereochemistry.

8. The computer system of any one of claims 1 to 7, wherein the plurality of reaction types comprises 20 or more reaction types, and the combinatorial synthesis library comprises 100 or more compounds for each of the plurality of reaction types.

9. The computer system of any one of claims 1 to 8, wherein the first plurality of reactants comprises three or more reactants, and the corresponding mapping of the corresponding plurality of synthons for reactants of the three or more reactants comprises ten or more synthons.

10. The computer system according to any one of claims 1 to 9, wherein The output of the reaction query generator model is used to identify a first reaction key of a plurality of reaction keys by a first query key lookup, and Each reaction key in the plurality of reaction keys represents a synthesis reaction that can be used to synthesize one or more compounds in the combinatorial synthesis library.

11. The computer system of any one of claims 1 to 10, wherein the output of the synthesizer query generator model is used to identify a synthesizer key of the corresponding synthon by a second query key lookup.

12. The computer system of any one of claims 1 to 11, wherein the single graph represents single molecule compounds present in the combinatorial synthesis library.

13. The computer system of any one of claims 1 to 11, wherein the single graph represents a weighted composite of a first graph for a first molecular compound and a second graph for a second molecular compound.

14. A computer system according to any one of claims 1 to 11, wherein The single graph represents a weighted composite of the plurality of graphs for the second plurality of compounds, and The second plurality of compounds have a common property.

15. The computer system of claim 14, wherein the common property is a Tanimoto distance to each other compound in the second plurality of compounds that is less than a threshold value.

16. The computer system of claim 14, wherein the common property is a binding coefficient to a macromolecular target that is less than a threshold value.

17. The computer system of any one of claims 1 to 16, wherein the plurality of compounds includes one billion or more compounds, and the molecular structure output by recognition (D) is any one of the one billion or more compounds that satisfies the query.

18. The computer system of any one of claims 1 to 16, wherein the plurality of compounds includes trillions or more compounds, and the molecular structure output by recognition (D) is any one of the trillions or more compounds that satisfies the query.

19. The computer system of any one of claims 1 to 18, wherein the single graph represents a query molecular compound as a set of atomic features and a set of bond features.

20. The computer system of claim 19, wherein The set of atomic characteristics includes element type, nodal degree, hybridization, chirality, bonded hydrogens, formal charge, and aromaticity, and The set of bonding characteristics includes bonding time, conjugation, intra-ring and stereochemistry.

21. The computer system of claim 20, wherein each non-hydrogen atom in the query molecular compound is represented by 2000 or more parameters in the set of atomic features, and each covalent bond in the molecular compound is represented by 500 or more parameters in the set of bond features.

22. A method for querying a combinatorial synthesis library comprising a plurality of compounds, wherein The combinatorial synthesis library represents a variety of reaction types, Each respective reaction type in the plurality of reaction types has a corresponding mapping to a corresponding plurality of reactants, and Each corresponding reactant in each corresponding plurality of reactants has a corresponding mapping to a corresponding plurality of synthons, and The method is performed on a computer system, the computer system comprising: one or more central processing units; one or more graphics processors, wherein each of the one or more graphics processors includes 100 or more cores; as well as a memory addressable by the one or more central processing units, the memory storing at least one program for execution at least in part by the one or more graphics processors, the at least one program comprising instructions for performing the method, the method comprising: (A) inputting a query into a molecular encoder model, wherein the query is a single graph, wherein the molecular encoder model comprises a message passing neural network, the message passing neural network comprising a plurality of message passing layers that collectively comprise a first plurality of parameters, thereby obtaining a query vector by applying the first plurality of parameters to the single graph; (B) inputting the query vector into a reaction query generator model including a second plurality of parameters, thereby obtaining a first reaction type from the plurality of reaction types as an output of the reaction query generator model by applying the second plurality of parameters to the query vector; (C) determining a corresponding synthon for each corresponding reactant in a first plurality of reactants from the corresponding plurality of synthons mapped to the corresponding reactants by inputting the corresponding reactant into a synthon query generator model including a third plurality of parameters, thereby obtaining a corresponding synthon as an output of the synthon query generator model by applying the third plurality of parameters to the corresponding reactant, thereby determining a set of synthons, each synthon in the set of synthons corresponding to a reactant in the first plurality of reactants; and (D) identifying molecular structures in the combinatorial synthesis library, the combinatorial synthesis library comprising the set of synthons arranged according to the synthesis rules associated with the first reaction type.

23. A computer readable storage medium storing one or more programs, the one or more programs comprising instructions that, when executed by a computer system having one or more central processing units and one or more graphics processors, wherein each of the one or more graphics processors comprises 100 or more cores, and a memory that causes the computer system to query a combinatorial synthesis library, wherein The combinatorial synthesis library represents a variety of reaction types, Each respective reaction type in the plurality of reaction types has a corresponding mapping to a corresponding plurality of reactants, and Each corresponding reactant in each corresponding plurality of reactants has a corresponding mapping to a corresponding plurality of synthons, The querying of the combinatorial synthesis library is performed at least in part by the one or more graphics processors, by a method comprising: (A) inputting a query into a molecular encoder model, wherein the query is a single graph, wherein the molecular encoder model comprises a message passing neural network, the message passing neural network comprising a plurality of message passing layers that collectively comprise a first plurality of parameters, thereby obtaining a query vector by applying the first plurality of parameters to the single graph; (B) inputting the query vector into a reaction query generator model including a second plurality of parameters, thereby obtaining a first reaction type from the plurality of reaction types as an output of the reaction query generator model by applying the second plurality of parameters to the query vector; (C) determining a corresponding synthon for each corresponding reactant in a first plurality of reactants from the corresponding plurality of synthons mapped to the corresponding reactants by inputting the corresponding reactant into a synthon query generator model including a third plurality of parameters, thereby obtaining a corresponding synthon as an output of the synthon query generator model by applying the third plurality of parameters to the corresponding reactant, thereby determining a set of synthons, each synthon in the set of synthons corresponding to a reactant in the first plurality of reactants; and (D) identifying molecular structures in the combinatorial synthesis library, the combinatorial synthesis library comprising the set of synthons arranged according to the synthesis rules associated with the first reaction type.

24. A computer system for querying a combinatorial synthesis library comprising a plurality of compounds, wherein The combinatorial synthesis library represents a variety of reaction types, Each respective reaction type in the plurality of reaction types has a corresponding mapping to a corresponding plurality of reactants, and Each respective reactant in each corresponding plurality of reactants has a corresponding mapping to a corresponding plurality of synthons, the computer system comprising: one or more processors; a memory addressable by the one or more processors, the memory storing at least one program for execution by the one or more processors, the at least one program comprising instructions for: (A) inputting a query into a molecular encoder model, wherein the query is an arbitrary graph, wherein the molecular encoder model comprises a first plurality of parameters, thereby obtaining a query vector by applying the first plurality of parameters to the arbitrary graph; (B) inputting the query vector into a reaction query generator model including a second plurality of parameters, thereby obtaining a first reaction type from the plurality of reaction types as an output of the reaction query generator model by applying the second plurality of parameters to the query vector; (C) determining a corresponding synthon for each corresponding reactant in a first plurality of reactants of the first reaction type from among the corresponding plurality of synthons mapped to the corresponding reactants by inputting the corresponding reactant into a synthon query generator model including a third plurality of parameters, thereby obtaining a corresponding synthon as an output of the synthon query generator model by applying the third plurality of parameters to the corresponding reactant, thereby determining a set of synthons, each synthon in the set of synthons corresponding to a reactant in the first plurality of reactants; and (D) identifying molecular structures in the combinatorial synthesis library, the combinatorial synthesis library comprising the set of synthons according to the synthesis rules associated with the first reaction type.

25. The computer system of claim 24, wherein the reactive query generator model is a two-layer perceptron with an intermediate ReLU activation.

26. The computer system of claim 24, wherein the synthetic subquery generator model is a two-layer perceptron with an intermediate ReLU activation.

27. The computer system of claim 24, wherein The first plurality of parameters comprises 100,000 parameters, The second plurality of parameters includes 5,000 parameters, and The third plurality of parameters includes 5,000 parameters.

28. The computer system of claim 24, wherein The arbitrary graph includes a plurality of nodes and a plurality of edges, and Each node of the plurality of nodes is connected to another node of the plurality of nodes through at least one edge of the plurality of edges.

29. The computer system of claim 28, wherein each of the plurality of nodes is associated with: (i) a corresponding element type among a plurality of element types, (ii) the node degree among multiple node degrees, (iii) Hybridization within multiple hybridizations, (iv) multiple bonded hydrogens, (v) a formal charge in a set of formal charges, and (vi) Binary indication of aromaticity.

30. A computer system according to claim 28 or 29, wherein each respective key of the plurality of keys is associated with: (i) key type, (ii) binary indication of conjugation, (iii) a binary indication of whether the corresponding bond is within a ring, and (iv) Indication of stereochemistry.

31. The computer system of any one of claims 24 to 30, wherein the plurality of reaction types comprises 20 or more reaction types, and the combinatorial synthesis library comprises 100 or more compounds for each reaction type in the plurality of reaction types.

32. The computer system of any one of claims 24 to 31, wherein the first plurality of reactants comprises three or more reactants, and the corresponding mapping of the corresponding plurality of synthons for reactants of the three or more reactants comprises ten or more synthons.

33. A computer system according to any one of claims 24 to 32, wherein The output of the reaction query generator model is used to identify a first reaction key of a plurality of reaction keys by a first query key lookup, and Each reaction key in the plurality of reaction keys represents a synthesis reaction that can be used to synthesize one or more compounds in the combinatorial synthesis library.

34. A computer system according to any one of claims 24 to 32, wherein the output of the synthesizer query generator model is used to identify a synthesizer key of the corresponding synthon by a second query key lookup.

35. The computer system of any one of claims 24 to 33, wherein the arbitrary graph represents single molecule compounds present in the combinatorial synthesis library.

36. The computer system of any one of claims 24 to 33, wherein the arbitrary graph represents a weighted composite of a first graph of a first molecular compound and a second graph of a second molecular compound.

37. A computer system according to any one of claims 24 to 33, wherein The arbitrary graph represents a weighted composite of a plurality of graphs for a second plurality of compounds, and Any number of compounds have common properties.

38. The computer system of claim 36, wherein the common property is a Tanimoto distance to each other compound in the second plurality of compounds that is less than a threshold value.

39. The computer system of claim 36, wherein the common property is a binding coefficient to a macromolecular target that is less than a threshold value.

40. The computer system of any one of claims 24 to 38, wherein the plurality of compounds includes one billion or more compounds, and the molecular structure output by identification (D) is any one of the one billion or more compounds that satisfies the query.

41. A computer system according to any one of claims 24 to 38, wherein the plurality of compounds includes a trillion or more compounds, and the molecular structure output by identification (D) is any one of the trillion or more compounds that satisfies the query.

42. The computer system of any one of claims 24 to 40, wherein the arbitrary graph represents a query molecular compound as a set of atomic features and a set of bond features.

43. The computer system of claim 41, wherein The set of atomic characteristics includes element type, nodal degree, hybridization, chirality, bonded hydrogens, formal charge, and aromaticity, and The set of bonding characteristics includes bonding time, conjugation, intra-ring and stereochemistry.

44. The computer system of claim 42, wherein each non-hydrogen atom in the query molecular compound is represented by 2000 or more parameters in the set of atomic features, and each covalent bond in the molecular compound is represented by 500 or more parameters in the set of bond features.

45. The computer system of any one of claims 24 to 43, wherein the molecular encoder model comprises a message passing neural network comprising a plurality of message passing layers that collectively comprise the first plurality of parameters.

46. ​​A method for querying a combinatorial synthesis library comprising a plurality of compounds, wherein The combinatorial synthesis library represents a variety of reaction types, Each respective reaction type in the plurality of reaction types has a corresponding mapping to a corresponding plurality of reactants, and Each corresponding reactant in each corresponding plurality of reactants has a corresponding mapping to a corresponding plurality of synthons, and The method is performed on a computer system, the computer system comprising: one or more processors; a memory addressable by the one or more processors, the memory storing at least one program for execution by the one or more processors, the at least one program comprising instructions for performing a method comprising: (A) inputting a query into a molecular encoder model, wherein the query is an arbitrary graph, and the molecular encoder model comprises a first plurality of parameters, thereby obtaining a query vector by applying the first plurality of parameters to the arbitrary graph; (B) inputting the query vector into a reaction query generator model including a second plurality of parameters, thereby obtaining a first reaction type from the plurality of reaction types as an output of the reaction query generator model by applying the second plurality of parameters to the query vector; (C) determining a corresponding synthon for each corresponding reactant in a first plurality of reactants from the corresponding plurality of synthons mapped to the corresponding reactants by inputting the corresponding reactant into a synthon query generator model including a third plurality of parameters, thereby obtaining a corresponding synthon as an output of the synthon query generator model by applying the third plurality of parameters to the corresponding reactant, thereby determining a set of synthons, each synthon in the set of synthons corresponding to a reactant in the first plurality of reactants; and (D) identifying molecular structures in the combinatorial synthesis library, the combinatorial synthesis library comprising the set of synthons arranged according to the synthesis rules associated with the first reaction type.

47. A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions that, when executed by a computer system having one or more processors and a memory that causes the computer system to query a combinatorial synthesis library, The combinatorial synthesis library represents a variety of reaction types, Each respective reaction type in the plurality of reaction types has a corresponding mapping to a corresponding plurality of reactants, and Each corresponding reactant in each corresponding plurality of reactants has a corresponding mapping to a corresponding plurality of synthons, By methods including: (A) inputting a query into a molecular encoder model, wherein the query is an arbitrary graph, and the molecular encoder model comprises a first plurality of parameters, thereby obtaining a query vector by applying the first plurality of parameters to the arbitrary graph; (B) inputting the query vector into a reaction query generator model including a second plurality of parameters, thereby obtaining a first reaction type from the plurality of reaction types as an output of the reaction query generator model by applying the second plurality of parameters to the query vector; (C) determining a corresponding synthon for each corresponding reactant in a first plurality of reactants from the corresponding plurality of synthons mapped to the corresponding reactants by inputting the corresponding reactant into a synthon query generator model including a third plurality of parameters, thereby obtaining a corresponding synthon as an output of the synthon query generator model by applying the third plurality of parameters to the corresponding reactant, thereby determining a set of synthons, each synthon in the set of synthons corresponding to a reactant in the first plurality of reactants; and (D) identifying molecular structures in the combinatorial synthesis library, the combinatorial synthesis library comprising the set of synthons arranged according to the synthesis rules associated with the first reaction type.

Images