Determination of graph edit distance in drug-like chemical spaces

Abstract

A system and method for estimating the graph edit distance (GED) between compounds are provided. A first graph representing a first compound contains multiple nodes and multiple edges. The atoms of the first compound are represented by nodes, and the bonds of the first compound are represented by edges in the first graph. A second graph representing a second compound also contains multiple nodes and multiple edges. The atoms of the second compound are represented by nodes, and the bonds of the second compound are represented by edges in the second graph. The first graph is input into the model to generate a first latent embedding. The second graph is input into the model to generate a second latent embedding. The estimate (GED) between the two compounds is determined as the difference between the two latent embeddings.

Description

Technical Field

[0001] Cross - reference to Related Applications This application claims the priority of U.S. Provisional Patent Application No. 63 / 492,200, entitled "GRAPH EDIT DISTANCE DETERMINATION IN DRUG-LIKE CHEMICAL SPACES", filed on March 24, 2023, which is incorporated herein by reference.

[0002] This application claims the priority of U.S. Provisional Patent Application No. 63 / 492,208, entitled "IMPROVED BINDING SITE MODELS", filed on March 24, 2023, which is incorporated herein by reference.

[0003] This application is directed to determining an estimated value of the graph edit distance between compounds.

Background Art

[0004] Figure 3A illustrates graphic edit distance and how it is a measure of chemical similarity. A molecule can be represented as a graph, where the atoms of the molecule are represented as nodes in the graph and the bonds of the molecule are represented by edges in the graph. Graph editing involves the following six basic operations: (i) adding nodes, (ii) deleting nodes, (iii) substituting or relabeling nodes, (iv) adding edges, (v) deleting edges, and (vi) substituting or relabeling edges. Each of these six basic operations can be associated with an operation cost. A path defines a set of basic operations that transform one graph into another. The sum of the costs of each basic operation in a path defines the cost of the path. Graph edit distance is the path with the lowest cost. Therefore, if there are two or more (sets of basic operations) that can transform one graph into another, the graph edit distance is defined as the path with the lowest cost. An example of pyridine being transformed into aniline is shown in Figure 3A. Multiple paths may exist to perform the transformation, but the path with the lowest cost is the graph edit distance. For the conversion from pyridine to aniline, the minimum-cost path is the substitution of N to C, the addition of an N atom, and the addition of a CN bond. This converts the graph for pyridine to the graph for aniline. If the cost of each of these three basic operations is 1, the total cost for the transition from pyridine to aniline, the graph edit distance, is 3.

[0005] Graph Edit Distance (GED) can be compared to other molecular similarity metrics such as fingerprinting and maximum common substructures (MCS) based methods. See Zhang et al., 2015, “Design of chemical space networks using a Tanimoto similarity variant based upon maximum common substructures,” J.Comput.Aided Des. 29, 937, which is incorporated herein by reference in its entirety. Fingerprinting-based methods can be hypersensitive to minor molecular changes, especially when the change is central to the molecule. Thus, a minor change in the center of the molecule will result in a large value for the fingerprinting metric, while the same change around the molecule will not. In contrast, moving a minor change from the periphery to the center of the molecule can reduce the MCS metric by as much as 50 percent. GED is sensitive to molecular changes regardless of where the change occurs. Therefore, GED produces similar values ​​for minor changes both centrally and peripherally. For example, if carbon is changed to nitrogen, the GED value is always 1, regardless of where in the molecule the change occurred.

[0006] GED and MCS metrics are chemically intuitive in that they adequately track the actual chemical changes required to transform one molecule into another, whereas fingerprinting methods are typically not. GED is chemically intuitive in the sense that it defines a transformation that makes one molecule appear as another. MCS is chemically intuitive in a different sense; it shows how much of two molecules are in common.

[0007] A drawback of GED is its extremely high computational cost. In fact, accurate GED calculations are characterized as NP-complete tasks. The computational cost increases or decreases unfavorably with the molecular weight of the compounds being compared. For example, for compounds with molecular weights between 350 and 500 Daltons, GED calculations become prohibitively expensive for large-scale use in most drug discovery applications, such as QSAR modeling of compounds and identification of chemically related analogs. Due to the computational cost, methods have been developed to approximate or estimate accurate GEDs instead of performing accurate GED calculations. One such GED approximation algorithm is called the Hungarian method. See Bougleux et al., 2017, “A Hungarian Algorithm for Error-Correcting Graph Matching,” Lecture Notes in Computer Science, Conference: International Workshop on Graph-Based Representations in Pattern Recognition (GbR), Anacapri, Capri, Italy, LNCS 10310, which is incorporated herein by reference. The Hungarian method scales well with respect to molecular weight. However, the GED approximation produced by the Hungarian method has an insufficient mean squared error compared to the true GED.

[0008] Another GED approximation approach is SimGNN. See Bai et al. 2019, ACM International Conference on Web Search and Data Mining, which is incorporated herein by reference. SimGNN is a general program for computing graph similarity, including similarity between composite graphs. SimGNN was trained on the AIDS, LINUX program-dependent graphs, and IMDB ego-network datasets. Of these three datasets, only the AIDS dataset is related to drug-like systems. Thus, SimGNN can address general graph editing problems and can be used with graphs much larger than molecules. Furthermore, many of what SimGNN was trained on do not represent drug-like chemical spaces. As will be discussed in more detail below, SimGNN GED approximations, like those produced by the Hungarian method, do not have a satisfactory mean squared error when applied to drug-like applications.

[0009] Therefore, considering the above background, what is needed in this field is a computer system and method for performing an improved GED approximation. [Overview of the Initiative]

[0010] This disclosure addresses the issues raised in the background. A system and method are provided for estimating the graph edit distance (GED) between compounds. A first graph representing a first compound includes a plurality of nodes and a plurality of edges. The atoms of the first compound are represented by the nodes, and the bonds of the first compound are represented by the edges of the first graph. A second graph representing a second compound also includes a plurality of nodes and a plurality of edges. The atoms of the second compound are represented by the nodes, and the bonds of the second compound are represented by the edges of the second graph. The first graph is input into a model to generate a first isometrically preserved latent embedding. The second graph is input into a second model to generate a second isometrically preserved latent embedding. The estimate (GED) between the two compounds is provided using the difference between the first isometrically preserved latent embedding and the second isometrically preserved latent embedding.

[0011] A. Graph-based embodiment. Accordingly, one aspect of the present disclosure provides a system and method for estimating the graph edit distance between a first compound and a second compound.

[0012] A first graph representing the first compound is obtained. The first graph includes a first set of nodes and a first set of edges. Each atom of the first compound is represented by a node within the first set of nodes. Each bond of the first compound is represented by an edge within the first set of edges.

[0013] In some embodiments, a second graph representing the second compound is obtained. The second graph includes a second set of nodes and a second set of edges. Each atom of the second compound is represented by a node within the second set of nodes. Each bond of the second compound is represented by an edge within the second set of edges.

[0014] In some such embodiments, the first and second compounds have precise graph editing distances of 1 to 10.

[0015] In some such embodiments, the first and second compounds have precise graph editing distances of 1 to 15.

[0016] In some embodiments, the first compound and the second compound are organic compounds having molecular weights of less than 500 daltons, less than 1000 daltons, less than 2000 daltons, less than 4000 daltons, less than 6000 daltons, less than 8000 daltons, less than 10000 daltons, or less than 20000 daltons, respectively.

[0017] In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of 400 daltons to 10,000 daltons.

[0018] The first graph is input to a model containing multiple parameters. The model applies the multiple parameters of the model to the first graph (through at least 10,000 calculations, e.g., processor instructions) to generate a first isometrically retained latent embedding representing the first compound as the first output from the model.

[0019] The second graph is input to the model. The second model applies multiple parameters to the second graph (through at least 10,000 calculations, e.g., processor instructions) to generate a second isometrically retained latent embedding representing the second compound as a second output from the model.

[0020] In some embodiments, the model is a message passing network.

[0021] In some embodiments, the model is a graph neural network.

[0022] In some embodiments, the model is a message-passing graph neural network having 3 to 8 layers.

[0023] In some embodiments, the model is a message-passing graph neural network consisting of five layers.

[0024] In some embodiments, the model has more than 500,000 parameters, or 1 × 10⁻¹⁶ parameters. 6 It contains more than one parameter.

[0025] In some embodiments, the model performs at least 100,000 calculations to compute a first isometrically preserved latent implant, and the model performs at least 100,000 calculations to compute a second isometrically preserved latent implant.

[0026] An estimate of the graph edit distance between the first compound and the second compound is provided using the difference between the first isometrically retained latent embedding and the second isometrically retained latent embedding.

[0027] In some embodiments, the difference is the L2 norm difference between a first isometrically preserved latent embedding and a second isometrically preserved latent embedding. In some embodiments, the first latent space is a first vector containing a first set of vector values, and the second latent space is a second vector containing a second set of vector values, and the L2 norm difference between the first and second vectors is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first and second vectors.

[0028] In some embodiments, it is determined whether a second compound is a chemically related analog of a first compound when the estimated graph edit distance falls below a threshold (e.g., less than 5, less than 4, less than 3, or less than 2).

[0029] In some embodiments, for each compound-specific pair in the compound database, the process of obtaining a first graph, obtaining a second graph, inputting the first graph, inputting the second graph, and providing an approximate GED is repeated, where the first compound in each compound-specific pair is the first compound, the second compound in each compound-specific pair is the second compound, thereby forming multiple estimated graph edit distances. The multiple estimated graph edit distances are then clustered into multiple clusters by evaluating the multiple distances using a criterion function. Compounds are then selected from one or more clusters within the multiple clusters for evaluation in a drug discovery program.

[0030] In some embodiments, a quantitative structure-activity relationship model is constructed using an estimate of the graph edit distance, including the first and second compounds.

[0031] In some embodiments, an estimate of the graph edit distance is used to determine the activity cliff for the drug target.

[0032] Another aspect of the present disclosure provides a method for estimating the graph edit distance between a first compound and a second compound. The method comprises (A) obtaining a first graph representing the first compound, the first graph comprising a first plurality of nodes and a first plurality of edges. Each atom of the first compound is represented by a node in the first plurality of nodes, and each bond of the first compound is represented by an edge in the first plurality of edges. The method further comprises (B) obtaining a second graph representing the second compound, the second graph comprising a second plurality of nodes and a second plurality of edges. Each atom of the second compound is represented by a node in the second plurality of nodes, and each bond of the second compound is represented by an edge in the second plurality of edges. The method further comprises (C) inputting the first graph into a model comprising a plurality of parameters. The model applies multiple parameters to a first graph (e.g., through at least 10,000 calculations such as processor instructions) to generate a first isometrically preserved latent embedding representing a first compound as a first output from the model. The method further includes (D) inputting a second graph into the model. The model applies multiple parameters to the second graph (e.g., through at least 10,000 calculations) to generate a second isometrically preserved latent embedding representing a second compound as a second output from the model. The method further includes (E) using the difference between the first isometrically preserved latent embedding and the second isometrically preserved latent embedding to provide an estimate of the graph edit distance between the first compound and the second compound.

[0033] Another aspect of the present disclosure provides a non-temporary computer-readable storage medium. The non-temporary computer-readable storage medium stores instructions that, when executed by a computer system, cause the computer system to perform a method for estimating the graph edit distance between a first compound and a second compound. The method comprises (A) obtaining a first graph representing the first compound. The first graph comprises a first plurality of nodes and a first plurality of edges. Each atom of the first compound is represented by a node in the first plurality of nodes, and each bond of the first compound is represented by an edge in the first plurality of edges. The method further comprises (B) obtaining a second graph representing the second compound. The second graph comprises a second plurality of nodes and a second plurality of edges. Each atom of the second compound is represented by a node in the second plurality of nodes, and each bond of the second compound is represented by an edge in the second plurality of edges. The method further comprises (C) inputting the first graph into a model comprising a plurality of parameters. The model applies multiple parameters to a first graph (e.g., through at least 10,000 calculations) to generate a first isometrically preserved latent embedding representing a first compound as a first output from the model. The method further includes (D) inputting a second graph into the model. The model applies multiple parameters to the second graph (e.g., through at least 10,000 calculations) to generate a second isometrically preserved latent embedding representing a second compound as a second output from the model. The method further includes (E) using the difference between the first isometrically preserved latent embedding and the second isometrically preserved latent embedding to provide an estimate of the graph edit distance between the first compound and the second compound.

[0034] B. Embodiments of Compound Representations. Another aspect of the present disclosure provides a computer system for estimating a graph edit distance between a first compound and a second compound. The computer system comprises one or more processors and a memory addressable by one or more processors. The memory stores at least one program for execution by one or more processors. The at least one program includes (A) instructions for inputting a first representation of the first compound into a model comprising a plurality of parameters. The model applies the plurality of parameters to the first representation of the first compound to produce a first isometrically preserved latent embedding representing the first compound as a first output from the model. The at least one program further includes (B) instructions for inputting a second representation of the second compound into the model. The model applies the plurality of parameters to the second representation of the second compound to produce a second isometrically preserved latent embedding representing the second compound as a second output from the model. At least one program further includes instructions for providing an estimate of the graph edit distance between a first compound and a second compound using the difference between a first isometrically preserved latent embedding and a second isometrically preserved latent embedding.

[0035] In some embodiments, the difference is the L2 norm difference between a first isometrically preserved latent embedding and a second isometrically preserved latent embedding. In some such embodiments, the first isometrically preserved latent embedding is a first vector containing a first set of vector values, and the second isometrically preserved latent embedding is a second vector containing a second set of vector values, and the L2 norm difference between the first vector and the second vector is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first vector and the second vector, and the model is trained such that the L2 norm of the difference between the first vector and the second vector is an estimate of the GED between the first numerator and the second numerator.

[0036] In some embodiments, at least one program further includes instructions for determining that a second compound is a chemically related analog of a first compound when the estimated graph edit distance falls below a threshold.

[0037] In some embodiments, at least one program further includes instructions for repeating input (A), input (B), and provide (C) for each compound-specific pair in a compound database, where the first compound in each compound-specific pair is the first compound, the second compound in each compound-specific pair is the second compound, and so on, thereby forming multiple graph edit distances. In some such embodiments, the multiple graph edit distances are clustered into multiple clusters by evaluating the multiple distances using a criterion function, and compounds from one or more of the multiple clusters are selected for evaluation in a drug discovery program.

[0038] In some embodiments, at least one program further includes instructions for constructing a quantitative structure-activity relationship model comprising a first compound and a second compound, using an estimate of the graph edit distance.

[0039] In some embodiments, at least one program further includes instructions for determining the activity cliff for a drug target using an estimate of the graph edit distance.

[0040] In some embodiments, the model is a graph convolutional neural network.

[0041] In some embodiments, the model is a graph neural network (message passing network).

[0042] In some embodiments, the model is a message-passing graph neural network consisting of 3 to 8 layers.

[0043] In some embodiments, the model is a message-passing graph neural network consisting of five layers.

[0044] In some embodiments, the model is an autoencoder. In some such embodiments, the autoencoder is a variational autoencoder, an adversarial autoencoder, a normalized autoencoder, a sparse autoencoder, or a metric learning autoencoder.

[0045] In some embodiments, the first representation is a first fingerprint of the first compound, and the second representation is a second fingerprint of the second compound.

[0046] In some embodiments, the first expression is a plurality of first descriptors of a first compound, and the second expression is a plurality of second descriptors of a second compound.

[0047] In some embodiments, the first representation is a first three-dimensional structural representation of the first compound, and the second representation is a second three-dimensional structural representation of the second compound.

[0048] In some embodiments, the first representation is a first tensor-based representation of the first compound, and the second representation is a second tensor-based representation of the second compound.

[0049] In some embodiments, the first representation is a first string representation of the first compound, and the second representation is a second string representation of the second compound.

[0050] In some embodiments, the model includes more than 500,000 parameters. In some embodiments, the model is 1 × 10 6 It contains more than one parameter.

[0051] In some embodiments, the first and second compounds each satisfy two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 Daltons, and (iv) a LogP of less than 5.

[0052] In some embodiments, the first and second compounds have accurate graph editing distances of 1 to 10. In some embodiments, the first and second compounds have accurate graph editing distances of 1 to 15.

[0053] In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than 500 daltons. In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than 1000 daltons. In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than 2000 daltons. In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than 4000 daltons, less than 6000 daltons, less than 8000 daltons, less than 10000 daltons, or less than 20000 daltons. In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of 300 daltons to 1500 daltons.

[0054] In some embodiments, the model performs at least 100,000 calculations to compute a first isometrically preserved latent implant, and the model performs at least 100,000 calculations to compute a second isometrically preserved latent implant.

[0055] Another aspect of the present disclosure provides a method for estimating a graph edit distance between a first compound and a second compound. The method comprises (A) inputting a first representation of the first compound into a model comprising a plurality of parameters (for example, on a computer system having one or more processors and memory addressable by one or more processors), the model applying the plurality of parameters to the first representation of the first compound to produce a first isometrically preserved latent embedding representing the first compound as a first output from the model. The method further comprises (B) inputting a second representation of a second compound into a model (for example, on a computer system having one or more processors and memory addressable by one or more processors), the model applying the plurality of parameters to the second representation of the second compound to produce a second isometrically preserved latent embedding representing the second compound as a second output from the model. The method further includes (C) providing an estimate of the graph edit distance between a first compound and a second compound using the difference between a first isometrically preserved latent embedding and a second isometrically preserved latent embedding (for example, in a computer system having one or more processors and memory addressable by one or more processors).

[0056] Another aspect of the present disclosure provides a non-temporary computer-readable storage medium. The non-temporary computer-readable storage medium stores instructions that, when executed by a computer system, cause the computer system to perform a method for estimating the graph edit distance between a first compound and a second compound. The method includes (A) inputting a first representation of the first compound into a model that includes a plurality of parameters. The model applies the plurality of parameters to the first representation of the first compound to produce a first isometrically preserved latent embedding representing the first compound as a first output from the model. The method further includes (B) inputting a second representation of the second compound into the model. The model applies the plurality of parameters to the second representation of the second compound to produce a second isometrically preserved latent embedding representing the second compound as a second output from the model. The method further includes (C) using the difference between the first isometrically preserved latent embedding and the second isometrically preserved latent embedding to provide an estimate of the graph edit distance between the first compound and the second compound. [Brief explanation of the drawing]

[0057] The drawings illustrate examples of the systems and methods of this disclosure. It should be clearly understood that the descriptions and drawings are for illustrative purposes only and to aid understanding, and are not intended to define any limitation of the systems and methods of this disclosure.

[0058] [Figure 1] Computer systems according to several embodiments of this disclosure are illustrated. [Figure 2-2D] The following illustrations show a method for estimating the graph edit distance between a first compound and a second compound, according to some embodiments of the present disclosure, where optional elements are indicated by dashed boxes. [Figure 3A] This diagram illustrates the graphic edit distance and how it is a measure of chemical similarity that follows prior art. [Figure 3B] Desired characteristics of a training set for training a model to approximate the graph edit distance, according to some embodiments of this disclosure, are illustrated. [Figure 4] Statistics for training and test drug-like datasets and training and test fragment datasets for training a model to approximate graph edit distance according to some embodiments of this disclosure are illustrated. [Figure 5] The performance of the trained prior art model SimGNN on the fragment dataset and drug-like dataset of this disclosure, according to several embodiments of this disclosure, is shown. [Figure 6] This paper describes the desired formal properties of a model for approximating graph edit distance, and how the prior art model SimGNN lacks some of these formal properties. [Figure 7] A schematic diagram of a system for approximating graph editing distance according to some embodiments of the present disclosure. [Figure 8] Figure 4 illustrates the performance of the system according to this disclosure, MolGEDNet, in terms of mean squared error, compared to the prior art model SimGNN, when the graph edit distance is approximated for the test dataset shown in Figure 4. [Figure 9] The system according to this disclosure, MolGEDNet, compares the ranking of graph edit distances using the approximate graph edit distance of the test dataset in Figure 4 with the ranking of graph edit distances using the true graph edit distance of the test dataset in Figure 4, compared with the prior art model SimGNN, using Spearman's ρ and Kendall's ρ.

number

number

number

[0059] Similar reference numbers refer to corresponding parts across multiple drawings. [Modes for carrying out the invention]

[0060] Introduction. This disclosure provides a system and method for estimating the graph edit distance (GED) between compounds. A first graph representing a first compound includes a plurality of nodes and a plurality of edges. The atoms of the first compound are represented by the nodes, and the bonds of the first compound are represented by the edges of the first graph. A second graph representing a second compound also includes a plurality of nodes and a plurality of edges. The atoms of the second compound are represented by the nodes, and the bonds of the second compound are represented by the edges of the second graph. The first graph is input into a model to generate a first isometrically preserved latent embedding. The second graph is input into a model to generate a second isometrically preserved latent embedding. The estimate (GED) between the two compounds is provided using the difference between the first isometrically preserved latent embedding and the second isometrically preserved latent embedding.

[0061] Herein, embodiments are given in detail, and examples of these embodiments are illustrated in the accompanying drawings. In the following detailed description, many specific details are given in order to provide a full understanding of the present disclosure. However, it will be apparent to those skilled in the art that the present disclosure can be carried out without these specific details. In other cases, well-known methods, procedures, components, circuits, and networks are not described in detail so as not to unnecessarily obscure the aspects of the embodiments.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which this disclosure pertains. All patents and publications referenced herein are incorporated herein by reference in their entirety.

[0063] Multiple examples may be provided for any component, operation, or structure described herein as a single example. Ultimately, the boundaries between various components, operations, and data stores are somewhat arbitrary, and a particular operation is illustrated in the context of a particular exemplary configuration. Other forms of functionality are conceivable and may be within the scope of implementations. In general, structures and functionalities presented as separate components of an exemplary configuration may be implemented as a combined structure or component. Similarly, structures and functionalities presented as single components may be implemented as separate components. These and other variations, modifications, additions, and improvements are within the scope of implementations.

[0064] Furthermore, while terms such as "first," "second," etc., may be used herein to describe various elements, it should be understood that these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, without departing from the scope of this disclosure, a first subject may be referred to as a second subject, and similarly, a second subject may be referred to as a first subject. Both the first and second subjects are subjects, but they are not the same subject.

[0065] The terms used in this disclosure are for the sole purpose of describing specific embodiments and are not intended to limit the invention. Where used in the description of the invention and the accompanying claims, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. Where used herein, the terms “and / or” will also be understood to mean and include any and all possible combinations of one or more of the enumerated items relating to the invention. Where used herein, the terms “comprises” and / or “comprising” specify the presence of the described features, integers, steps, actions, elements, and / or components, but will not be understood to exclude the presence or addition of one or more other features, integers, steps, actions, elements, components, and / or groups thereof.

[0066] As used herein, the term "if" may, depending on the context, be interpreted as meaning "when," "upon," or "in response to a determination," "according to a determination," or "in response to detection" that the preceding described condition is true. Similarly, the phrases "if (the preceding described condition is) determined to be true," "if (the preceding described condition is) true," or "when (the preceding described condition is) true" may, depending on the context, be interpreted as meaning "after (the preceding described condition is) determined to be true," "in response to a determination," "according to a determination," "after (the preceding described condition is) detected," or "in response to detection."

[0067] Furthermore, when a reference number is denoted as "i," this reference number refers to a general component, set, or embodiment. For example, a cell component referred to as "cell component i" refers to the i-th cell component in a group of cell components.

[0068] The preceding description included exemplary systems, methods, techniques, instruction sequences, and computing machine program products that embody exemplary implementations. For illustrative purposes, many specific details are described to provide an understanding of various implementations of the subject matter of the present invention. However, it will be apparent to those skilled in the art that implementations of the subject matter of the present invention can be carried out without these specific details. Generally, well-known instruction instances, protocols, structures, and techniques are not described in detail.

[0069] For illustrative purposes, the foregoing description is written with reference to specific implementations. However, the following exemplary considerations are not intended to be exhaustive, nor are they intended to limit implementations to the exact forms disclosed. Many modifications and variations are possible in light of the above teachings. The implementations are selected and described to best illustrate the principles and their practical applications, thereby enabling those skilled in the art to best utilize implementations suitable for specific intended uses and various implementations with various modifications.

[0070] For clarity, not all routine features of the implementations described herein are shown or explained. It will be understood that in the development of such actual implementations, many implementation-specific decisions are made to achieve specific designer objectives, such as compliance with use cases and business-related constraints, and these specific objectives will vary from implementation to implementation and from designer to designer. Furthermore, while such design efforts can be complex and time-consuming, it will be understood that they are nevertheless routine engineering endeavors for those skilled in the art who are interested in this disclosure.

[0071] Several parts of this description describe embodiments of the present invention in terms of algorithms and symbolic representations of operations on information. These algorithmic descriptions and representations are commonly used by those skilled in data processing technology to effectively communicate the nature of their work to others skilled in the art. These operations are described functionally, computationally, or logically, but are understood to be implemented by computer programs or equivalent electrical circuits, microcode, etc.

[0072] The language used herein has been selected primarily for readability and guidance purposes and may not have been selected to describe or limit the subject matter of the invention. Accordingly, the scope of the invention is intended to be limited not by this detailed description but by any claims relating to an application filed herein. Accordingly, the disclosure of embodiments of the invention is intended to be illustrative but not limiting of the scope of the invention.

[0073] In general, terms used in the claims and specification are intended to be interpreted as having an obvious meaning that can be understood by those skilled in the art. Specific terms are defined below for further clarity. In the event of any conflict between the plain meaning and the provided definition, the provided definition shall prevail.

[0074] Any terms not directly defined herein should be understood to have meanings generally associated with them as understood within the art of the invention. Certain terms are provided herein to provide additional guidance to practitioners when describing compositions, devices, methods, etc., of embodiments of the invention, and methods of making or using them. It should be understood that the same thing may be said in two or more ways. As a result, alternative languages ​​and synonyms may be used for one or more of the terms considered herein. It is not important whether a term is elaborated on or discussed herein. Several synonyms or alternative methods, materials, etc. are provided. The enumeration of one or more synonyms or equivalents does not preclude the use of other synonyms or equivalents unless expressly stated otherwise. The use of examples, including examples of terms, is for illustrative purposes only and does not limit the scope and meaning of embodiments of the invention herein.

[0075] To avoid any doubt, it is intended that any specific features (e.g., integers, properties, values, uses, diseases, formulas, compounds, or groups) described herein in relation to any particular aspect, embodiment, or example of the Disclosure may be understood to be applicable to any other aspect, embodiment, or example described herein, provided that they do not conflict. Thus, such features may be used in conjunction with any of the definitions, claims, or embodiments defined herein, as necessary. All features disclosed herein (including any appended claims, abstracts, and drawings) and / or all steps of any method or process so as disclosed may be combined in any combination, except for any combination in which at least some of the features and / or steps are mutually exclusive. This Disclosure is not limited to any details of any disclosed embodiment. This Disclosure extends to any novel features or combinations of features disclosed herein (including any appended claims, abstracts, and drawings), or any novel steps or combinations of any method or process so as disclosed.

[0076] definition As used herein, the terms “compound” and “molecule” are to be used interchangeably.

[0077] Where used herein, the term “about” means that dimensions, sizes, formulations, parameters, shapes, and other quantities and characteristics are not, and do not need to be, exact, but may be approximate and / or greater or less, as desired, taking into account tolerances, conversion factors, rounding, measurement errors, and equivalents, as well as 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 explicitly stated as such. It should be noted that very different embodiments of sizes, shapes, and dimensions may adopt the arrangement described.

[0078] As used herein, the term “fingerprint” refers to a digital digest of a compound, as well as a fingerprint of the compound. Non-exclusive examples of such digital digests include Daylight fingerprints, BCI fingerprints, ECFC4 fingerprints, ECFP4 fingerprints, EcFC fingerprints, MDL fingerprints, atomic pair fingerprints (APFP fingerprints), topological twist fingerprints (TTFP) fingerprints, UNITY 2D fingerprints, RNNS2S fingerprints, and GraphConv fingerprints. See Franco, 2014, “The Use of 2D fingerprint methods to support the assessment of structural similarity in orphan drug legislation,” J.Cheminform 6, p.5, and Rensi and Altman, 2017, “Flexible Analog Search with Kernel PCA Embedded Molecule Vectors,” Computational and Structural Biotechnology Journal, doi:10.1016 / j.csbj.2017.03.003. Each of these is incorporated herein by reference.See also Raymond and Willett, 2002, “Effectiveness of graph-based and fingerprint-based similarity measures for virtual screening of 2D chemical structure databases,” Journal of Computer-Aided Molecular Design 16, 59-71, and Franco et al., 2014, “The use of 2D fingerprint methods to support the assessment of structural similarity in orphan drug legislation,” Journal of Chemoinformatics 6(5). Each of these is incorporated herein by reference.

[0079] Where used interchangeably herein, the terms “classifier,” “model,” “algorithm,” “regressor,” and / or “classifier” refer to a machine learning model or algorithm. In some embodiments, the model is unsupervised learning. In some embodiments, the model is supervised machine learning. Non-exclusive examples of supervised learning include, but are not limited to, logistic regression, neural networks, support vector machines, naive Bayes algorithms, nearest neighbor algorithms, random forest algorithms, decision trees, boosted trees, polynomial logistic regression, linear models, linear regression, gradient boosting, mixture models, hidden Markov models, Gaussian NB algorithms, linear discriminant analysis, or any combination thereof. In some embodiments, the model is a multinomial classifier algorithm. In some embodiments, the model is a two-stage stochastic gradient descent (SGD) model. In some embodiments, the model is a deep neural network (e.g., a deep and wide sample-level model). In some embodiments, the classifiers or models of the Disclosure may number 25 or more, 100 or more, 1,000 or more, 10,000 or more, 100,000 or more, or 1 × 106 It has more than one parameter, and therefore, the model's calculations cannot be performed mentally.

[0080] Furthermore, as used herein, the term “parameter” refers to any coefficient or value of an internal or external element (e.g., weights and / or hyperparameters) in an algorithm, model, regressionr, and / or classifier that can affect (e.g., modify, adapt, and / or tune) one or more inputs, outputs, and / or features in the algorithm, model, regressionr, 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, adapt, and / or tune the behavior, learning, and / or performance of the algorithm, model, regressionr, and / or classifier. In some cases, a parameter is used to increase or decrease the influence of an input (e.g., a feature) to the algorithm, model, regressionr, and / or classifier. In a non-limiting example, in some embodiments, a parameter is used to increase or decrease the influence of a node (e.g., a neural network), which includes one or more activation functions. The assignment of parameters to specific inputs, outputs, and / or functions is not limited to any single paradigm for a given algorithm, model, regressionr, and / or classifier, but can be used in any suitable algorithm, model, regressionr, and / or classifier architecture for desired performance. In some embodiments, the parameters have fixed values. In some embodiments, the parameter values ​​are manually and / or automatically adjustable. In some embodiments, the parameter values ​​are modified by a validation and / or training process for the algorithm, model, regressionr, and / or classifier (e.g., by error minimization and / or backpropagation). In some embodiments, the algorithms, models, regressionrs, and / or classifiers of this disclosure include a plurality of parameters.In some embodiments, the multiple parameters are n parameters, where 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≧1×10⁶, n≧5×10. 6 , or n≧1×10 7 Therefore, the algorithms, models, regressors, and / or classifiers of this disclosure cannot be implemented in the mind. In some embodiments, n is 10,000 to 1 × 10 7 , 100,000~5×10 6 , or 500,000~1×10 6 In some embodiments, the algorithms, models, regressors, and / or classifiers of the Disclosure operate in a k-dimensional space, where k is a positive integer greater than or equal to 5 (e.g., 5, 6, 7, 8, 9, 10, etc.). Therefore, the algorithms, models, regressors, and / or classifiers of the Disclosure cannot be implemented in the mind.

[0081] Neural network. In some embodiments, the model is a neural network (e.g., a convolutional neural network and / or a residual neural network). A neural network, also known as an artificial neural network (ANN), includes convolutional and / or residual neural networks (deep learning). In some embodiments, the neural network is a machine learning model trained to map an input data set to an output data set, and the neural network includes a group of interconnected nodes organized into multiple layers of nodes. For example, in some embodiments, the neural network architecture includes at least an input layer, one or more hidden layers, and an output layer. In some embodiments, the neural network includes any total number of layers and any number of hidden layers, and the hidden layers function as trainable feature extractors that enable mapping a set of input data to an output value or a set of output values. In some embodiments, the deep learning model is a neural network that includes multiple hidden layers, e.g., two or more hidden layers. In some embodiments, each layer of the neural network includes a number of nodes (or "neurons"). In some embodiments, a node receives an input coming directly from either the input data or the output of the nodes in the previous layer and performs a specific operation, e.g., a summation operation. In some embodiments, the connection from the input to the node is associated with parameters (e.g., weights and / or weight coefficients). In some embodiments, the node is an input, x i, and the products of all pairs of the associated parameters are summed. In some embodiments, the weighted sum is offset by a bias b. In some embodiments, the output of a node or neuron is gated using a threshold function or activation function f, which is a linear or nonlinear function. In some embodiments, the activation function is, for example, a rectified linear unit (ReLU) activation function, a leaky ReLU activation function, or other functions such as saturated hyperbolic tangent, identity, binary step, logistic, arcTan, soft sine, parametric rectified linear unit, exponential linear unit, softPlus, bent identity, softExponential, sine curve, sine, Gaussian, or sigmoid function, or any combination thereof.

[0082] In some embodiments, the parameters of a neural network (weighting coefficients, bias values, and thresholds, or other computational parameters) are "instructed" or "learned" during the training phase using one or more sets of training data. For example, in some implementations, the parameters are trained using gradient descent or backpropagation with input data from the training dataset so that the output values ​​computed by the ANN match examples contained in the training dataset. In some embodiments, the parameters are obtained from a backpropagation neural network training process.

[0083] Any of the various neural networks are suitable for use in accordance with this disclosure. Examples 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, or any combination thereof. In some embodiments, machine learning utilizes a pre-trained and / or transfer-trained ANN or deep learning architecture. In some implementations, convolutional and / or residual neural networks are used in accordance with this disclosure.

[0084] For example, a deep neural network model includes an input layer, several individually parameterized (e.g., weighted) convolutional layers, and an output scorer. Each parameter (e.g., weight) of the convolutional layer, as well as the input layer, contributes to several parameters (e.g., weights) associated with the deep neural network model. In some embodiments, at least 50 parameters, at least 100 parameters, at least 1000 parameters, at least 2000 parameters, or at least 5000 parameters are associated with the deep neural network model. Thus, deep neural network models require the use of a computer because they cannot be solved mentally. In other words, given an input to the model, in such embodiments, the model output must be determined using a computer, not mentally. For example, see 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,” ch. Learning Representations by Back-propagating Errors, pp. 696-699, Cambridge, MA, USA: MIT Press. Each of these is incorporated herein by reference.

[0085] Suitable convolutional neural networks for use as models are disclosed, for example, 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 these is incorporated herein by reference. Further exemplary neural networks suitable for use as models are disclosed in Duda et al., 2001, Pattern Classification, Second Edition, John Wiley & Sons, Inc., New York, and Hastie et al., 2001, The Elements of Statistical Learning, Springer-Verlag, New York. Each of these is incorporated herein by reference in its entirety. Further exemplary neural networks suitable for use as models are 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 these is incorporated herein by reference in its entirety.

[0086] Support Vector Machines. In some embodiments, the model is a Support Vector Machine (SVM). Suitable SVMs for use as models include, for example, Cristianini and Shawe-Taylor, 2000, An Introduction to Support Vector Machines, Cambridge University Press, Cambridge; Boser et al., 1992, “A training algorithm for optimal margin classifiers,” in Proceedings of the 5th Annual ACM Workshop on Computational Learning Theory, ACM Press, Pittsburgh, Pa., pp. 142-152; Vapnik, 1998, Statistical Learning Theory, Wiley, New York; Mount, 2001, Bioinformatics: sequence and genome analysis, Cold Spring Harbor Laboratory Press, Cold Spring Harbor, NY; Duda, Pattern Classification, Second Edition, 2001, John Wiley & Sons, Inc., pp. 259, 262-265; and Hastie, 2001, The Elements of Statistical Learning, Springer, New York. This is described in York and Furey et al., 2000, Bioinformatics 16, 906-914. Each of these is incorporated herein by reference in its entirety. When used for classification, SVM separates given binary-labeled data using a hyperplane as far away as possible from the labeled data. In certain cases where linear separation is not possible, SVM works in combination with the technique of a “kernel” that automatically provides a nonlinear mapping to the feature space. The hyperplane found by SVM in the feature space corresponds in some cases to a nonlinear decision boundary in the input space.In some embodiments, multiple parameters associated with the SVM (e.g., weights) define a hyperplane. In some embodiments, the hyperplane is defined by at least 10, at least 20, at least 50, or at least 100 parameters, and the SVM model requires a computer to compute it because it cannot be solved mentally.

[0087] Naive Bayes models. In some embodiments, the model is a naive Bayes model. A suitable naive Bayes model for use as a model is disclosed, for example, in Ng et al., 2002, “On discriminative vs. generative classifiers: A comparison of logistic regression and naive Bayes,” Advances in Neural Information Processing Systems, 14, which is incorporated herein by reference. Naive Bayes models are a family of “stochastic modes” based on applying Bayes’ theorem with strong (naive) independence assumptions between features. In some embodiments, they are combined with kernel density estimation. For example, see Hastie et al., 2001, The elements of statistical learning: data mining, inference, and prediction, eds. Tibshirani and Friedman, Springer, New York, which is incorporated herein by reference.

[0088] Nearest Neighbor Method. In some embodiments, the model is the nearest neighbor method. In some implementations, the nearest neighbor model is memory-based and does not include the model being fitted. In the nearest neighbor method, given a query point x0 (under test), k training points x that are closest to x0 are selected. (r) ,r,...,k (here the training targets) are identified, and then the point x0 is classified using the k-nearest neighbor method. In some embodiments, the Euclidean distance in the feature space is

number

[0089] A k-neighborhood model is a non-parametric machine learning method in which the input consists of k nearest training examples in a feature space. The output is class membership. An object is classified by multiple votes from its nearest neighbors, and the object is assigned to the most common class among its k nearest neighbors (k is typically a small positive integer). When k=1, the object is simply assigned to the class of its single nearest neighbor. See Duda et al., 2001, Pattern Classification, Second Edition, John Wiley & Sons. This is similar to using a computer to solve a model for a given input, because in some embodiments incorporated herein by reference, the number of distance calculations required to solve a k-neighborhood model is too many to perform mentally.

[0090] Random forest, decision tree, and boost tree algorithms. In some embodiments, the model is a decision tree. Decision trees are generally described by Duda, 2001, Pattern Classification, John Wiley & Sons, Inc., New York, pp. 395–396, which is incorporated herein by reference. Tree-based methods divide the feature space into a set of rectangles and fit a model (such as a constant) into each rectangle. In some embodiments, the decision tree is a random forest regression. For example, one particular model is a classification and regression tree (CART). Other particular decision tree models, but not limited to, include ID3, C4.5, MART, and random forests. CART, ID3, and C4.5 are described by Duda, 2001, Pattern Classification, John Wiley & Sons, Inc., New York, pp. 396–408 and pp. 411–412, which is incorporated herein by reference. CART, MART, and C4.5 are described in Hastie et al., 2001, The Elements of Statistical Learning, Springer-Verlag, New York, Chapter 9. This is incorporated herein by reference in its entirety. Random forests are described in Breiman, 1999, “Random Forests--Random Features,” Technical Report 567, Statistics Department, UC Berkeley, September 1999. This is incorporated herein by reference in its entirety. In some embodiments, decision tree models include at least 10, at least 20, at least 50, or at least 100 parameters (e.g., weights and / or decisions) and require a computer to compute them because they cannot be solved mentally.

[0091] Regression. In some embodiments, the model uses a regression equation. In some embodiments, the regression equation is determined using any type of regression. For example, in some embodiments, the regression is logistic regression. In some embodiments, the regression is logistic regression with lasso, L2, or elastic net normalization. In some embodiments, extracted features with corresponding regression coefficients that cannot satisfy a threshold are excluded (removed) from consideration. In some embodiments, a generalization of the logistic regression model for handling multi-category responses is used as the model. Logistic regression is disclosed in Agresti, An Introduction to Categorical Data Analysis, 1996, Chapter 5, pp. 103-144, John Wiley & Son, New York, which is incorporated herein by reference. In some embodiments, the model uses the regression model disclosed in Hastie et al., 2001, The Elements of Statistical Learning, Springer-Verlag, New York. In some embodiments, the logistic regression model includes at least 10, at least 20, at least 50, at least 100, or at least 1000 parameters (e.g., weights), and it cannot be solved mentally, thus requiring a computer to compute it.

[0092] Linear Discriminant Analysis Models. In some embodiments, the model is linear discriminant analysis (LDA), normal discriminant analysis (NDA), or discriminant function analysis. In some embodiments, such a model is a generalization of Fisher's linear discriminant, a method used in statistics, pattern recognition, and machine learning to find a linear combination of features that characterize or separate two or more classes of objects or events. In some embodiments, the resulting combination is used as a model (linear model) in some embodiments of this disclosure.

[0093] Mixed models and hidden Markov models. In some embodiments, the model is a mixed model, such as that described in McLachlan et al., Bioinformatics 18(3):413-422, 2002. In some embodiments, particularly those embodiments including a time component, the model is a hidden Markov model, such as that described in Schliep et al., 2003, Bioinformatics 19(1):i255-i263.

[0094] Ensembles of models and boosts. In some embodiments, an ensemble of models (two or more) is used. In some embodiments, boosting techniques such as AdaBoost are used in conjunction with many other types of learning algorithms to improve the performance of the models. In this approach, any of the outputs of the models disclosed herein, or their equivalents, are combined into a weighted sum representing the final output of the boosted models. In some embodiments, multiple outputs from the models are combined using any measure of central tendency known in the art, including but not limited to the mean, median, mode, weighted mean, weighted median, weighted mode, etc. In some embodiments, multiple outputs are combined using a voting method. In some embodiments, each model in the ensemble of models is weighted or unweighted.

[0095] As used herein, the term “latent space” refers to a multidimensional space in which data points are represented after being encoded by the model. Such data points are multidimensional, but have a lower dimension than the input to the model. As used herein, the latent space has a lower dimension compared to the original input space of the mode, enabling more efficient storage and computation.

[0096] As used herein, the term “latent embedding” refers to the embedding of data points into a lower-dimensional space by a model.

[0097] As used herein, the term “vector” refers to an enumerated list of elements, such as an array of elements, where each element has a meaning to which it is assigned. Therefore, as used in this disclosure, the term “vector” is interchangeable with the term “tensor.” For example, if a vector contains a population count, then for each of the multiple cells, there exists a predetermined element in the vector for each of the multiple cells. For ease of presentation, in some cases, a vector may be described as one-dimensional; however, this disclosure is not limited thereto. A vector of any dimension may be used in this disclosure, provided that a description of what each element of the vector represents is defined (for example, element 1 may represent the population count of cell 1 of multiple cells).

[0098] Exemplary computer system Figure 1 illustrates a computer system 100 for estimating the graph edit distance between a first compound and a second compound.

[0099] Referring to Figure 1, in a typical embodiment, the computer system 100 comprises one or more computers. For illustrative purposes in Figure 1, the computer system 100 is represented as a single computer containing all of the disclosed computer system 100 functions. However, the disclosure is not limited thereto. The functions of the computer system 100 may be distributed across any number of networked computers and / or reside in each of multiple networked computers and / or virtual machines. Those skilled in the art will understand that a variety of different computer topologies are possible for the computer system 100, and that all such topologies are within the scope of the disclosure.

[0100] Referring to Figure 1 with the above in mind, the computer system 100 comprises one or more processing units (CPU and / or GPU) 59, a network or other communication interface 84, a user interface 78 (including, for example, an optional display 82 and an optional keyboard 80 or other form of input device), memory 92 (e.g., random access memory, persistent memory, or a combination thereof), one or more magnetic disk storage and / or persistent devices 90 optionally accessed by one or more controllers 88, one or more communication buses 12 for interconnecting the aforementioned components, and a power supply 79 for supplying power to the aforementioned components. Unless the components of memory 92 are persistent, data in memory 92 can be seamlessly shared with non-volatile memory 90 or portions of memory 92 that are non-volatile / persistent using known computing techniques such as caching. Memory 92 and / or memory 90 may include mass storage located remotely from the central processing unit 59. In other words, some data stored in memory 92 and / or memory 90 may actually be outside the computer system 100 but be hosted on a computer that can be electronically accessed by the computer system 100 via the internet, intranet, or other form of network or electronic cable using the network interface 84. In some embodiments, the computer system 100 utilizes a model in which the data is executed from memory associated with one or more graphical processing units in order to improve the speed and performance of the system. In some alternative embodiments, the computer system 100 utilizes a model in which the data is executed from memory 92 rather than from memory associated with the graphical processing units.

[0101] In some embodiments, the computer system 100 presents media to the user through the display 82. Examples of media presented by the display 82 include one or more images, videos, audio (e.g., waveforms of audio samples), or a combination thereof. In a typical embodiment, one or more images, videos, audio, or a combination thereof are presented by the display 82 through a client application (e.g., a client application hosted by the computer system 100, a client application accessible through the computer system 100, etc.). In some embodiments, audio is presented through an external device (e.g., a speaker, headphones, an input / output (I / O) subsystem, etc.) that receives audio information from the computer system 100 and presents audio data based on this audio information. In some embodiments, the user interface 78 also includes an audio output device, such as a speaker or audio output, for connection to a speaker, earphone, or headphones.

[0102] Memory 92 includes high-speed random-access memory such as DRAM, SRAM, DDR RAM, or other random-access solid-state memory devices, and optionally also includes non-volatile memory such as one or more magnetic disk storage devices, optical disk storage devices, flash memory devices, or other non-volatile solid-state storage devices. Memory 92 may optionally include one or more storage devices located remotely from the CPU 59. Memory 92, or alternatively non-volatile memory devices within memory 92, includes non-temporary computer-readable storage media. Access to memory 92 by other components of the computer system 100, such as the CPU 59, is optionally controlled by the controller 88. In some embodiments, memory 92 includes mass storage located remotely from the CPU 59. In other words, some data stored in memory 93 may actually be outside the computer system 100 but can be electronically accessed by the computer system 100 via the internet, intranet, or other forms of network or electronic cable using the network interface 84.

[0103] The memory 92 of computer system 100 is • A selection of 30 operating systems (e.g., embedded operating systems such as ANDROID, iOS, DARWIN, RTXC, LINUX, UNIX, OS X, WINDOWS, or VxWorks) that include procedures for handling various basic system services. • Approximate graph edit distance (GED) calculation module 32 for spatial data evaluation module 36 for estimating the graph edit distance between the first compound and the second compound. • Identity of the first compound 34, A graph 36 of a first compound 34, the graph comprising a plurality of nodes (38-1, ..., 38-N, where N is a positive integer greater than or equal to 2) and a plurality of edges (40-1, ..., 40-M, where M is a positive integer greater than or equal to 2), where each edge 40 identifies a source node (e.g., 42-1-1) and a destination node (42-1-2) within the plurality of nodes, where each atom of the first compound 36 is represented by a node 38 within the plurality of nodes, and each bond of the first compound is represented by an edge 40 within the plurality of edges, graph 36. • The identity of the second compound 44, A graph 46 of the second compound 44, the graph 46 comprising a plurality of nodes (48-1, ..., 48-X, where X is a positive integer greater than or equal to 2) and a plurality of edges (50-1, ..., 50-Y, where Y is a positive integer greater than or equal to 2), where each edge 50 identifies a source node (e.g., 52-1-1) and a destination node 52-1-2 within the plurality of nodes, where each atom of the second compound 46 is represented by node 48 within the plurality of nodes, and each bond of the second compound is represented by edge 50 within the plurality of edges, graph 46. The training library 56 stores pairs of training compounds 58-1, ..., 58-Q, where Q is a positive integer greater than or equal to 2, and each pair of training compounds is associated with a true graph edit distance 60.

[0104] In some implementations, one or more of the identified data elements or modules of the computer system 100 are stored in one or more of the aforementioned memory devices and correspond to a set of instructions for performing the functions described above. The identified data, modules, or programs (e.g., sets of instructions) do not need to be implemented as separate software programs, procedures, or modules, and therefore various subsets of these modules can be combined or rearranged in various implementations. In some implementations, memories 92 and / or 90 (and optionally 52) optionally store a subset of the modules and data structures identified above. Furthermore, in some embodiments, memories 92 and / or 90 (and optionally 52) store additional modules and data structures not described above.

[0105] Exemplary Method A system for estimating the graph edit distance between the first and second compounds is disclosed here, and a method for performing such characterization is described in detail with reference to Figure 2 and discussed below.

[0106] Block 200. Referring to Block 200, a system and method for estimating the graph edit distance between a first compound 34 and a second compound 44 are provided.

[0107] Block 202. Referring to Block 202, we optionally obtain a first graph 36 representing a first compound. The first graph includes a first set of nodes and a first set of edges. Each atom of the first compound is represented by a node 38 in the first set of nodes. Each bond of the first compound is represented by an edge 40 in the first set of edges.

[0108] Block 204. Referring to Block 204, we optionally obtain a second graph 46 representing a second compound 44. The second compound 46 includes a second set of nodes and a second set of edges. Each atom of the second compound is represented by a node 48 in the second set of nodes. Each bond of the second compound is represented by an edge 50 in the second set of edges.

[0109] Block 206. Referring to Block 206, in some embodiments, the first and second compounds each satisfy two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) molecular weight less than 500 Daltons, and (iv) LogP less than 5. See Lipinski, 1997, Adv. Drug Del. Rev. 23, 3, which is incorporated herein by reference in its entirety. In some embodiments, the first and second compounds each satisfy one or more criteria in addition to Lipinski's Rule of Five. For example, in some embodiments, the first and second compounds each have five or fewer aromatic rings, four or fewer aromatic rings, three or fewer aromatic rings, or two or fewer aromatic rings.

[0110] Blocks 208-210. Referring to block 208, in some embodiments, the first and second compounds have an exact graph edit distance of 1 to 10. Referring to block 210, in some embodiments, the first and second compounds have an exact graph edit distance of 1 to 15. In some embodiments, the exact graph edit distance between the first and second compounds is unknown.

[0111] Blocks 212-214. Referring to Block 212, in some embodiments, the first and second compounds are organic compounds having molecular weights of less than 500 daltons, less than 1000 daltons, less than 2000 daltons, less than 4000 daltons, less than 6000 daltons, less than 8000 daltons, less than 10000 daltons, or less than 20000 daltons, respectively. Referring to Block 214, in some embodiments, the first and second compounds are organic compounds having molecular weights of 400 daltons to 10000 daltons, respectively.

[0112] In some embodiments, the first compound and the second compound each have a molecular weight of at least 10 Da, at least 20 Da, at least 50 Da, at least 100 Da, at least 200 Da, at least 500 Da, at least 1 kDa, at least 2 kDa, at least 3 kDa, at least 5 kDa, at least 10 kDa, at least 20 kDa, at least 30 kDa, at least 50 kDa, at least 100 kDa, or at least 500 kDa. In some embodiments, the first compound and the second compound each have a molecular weight of 1000 kDa or less, 500 kDa or less, 100 kDa or less, 50 kDa or less, 10 kDa or less, 5 kDa or less, 2 kDa or less, 1 kDa or less, 500 Da or less, 300 Da or less, 100 Da or less, or 50 Da or less. In some embodiments, the first and second compounds each have molecular weights of 10 Da to 900 Da, 50 Da to 1000 Da, 100 Da to 2000 Da, 1 kDa to 10 kDa, 5 kDa to 500 kDa, or 100 kDa to 1000 kDa. In some embodiments, the first and second compounds each have molecular weights in a different range, starting at 10 Daltons or less and ending at 1000 kDa or less.

[0113] In some embodiments, the first compound and the second compound are each low molecular weight compounds. For example, in some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than approximately 1000 daltons (e.g., less than 900 daltons).

[0114] Block 216. Referring to Block 216, a first graph, or a representation of several other forms of the first compound, is input to a model containing multiple parameters. The model applies the multiple parameters to the first graph (for example, through at least 10,000 calculations) to produce a first isometrically retained latent embedding representing the first compound as a first output from the model.

[0115] As shown in Figure 10, in some embodiments, the first isometrically retained latent implant has dimensions of 64, 96, 128, 192, or 256. In some embodiments, the first isometrically retained latent implant has dimensions of 64 to 1024. In some embodiments, the first isometrically retained latent implant has dimensions of 64 to 512.

[0116] In some embodiments, the model (for example, model 702 in Figure 7) has more than 400 parameters, more than 1000 parameters, more than 2000 parameters, more than 5000 parameters, more than 10,000 parameters, more than 100,000 parameters, or 1 × 10 6 It contains more than one parameter.

[0117] In some embodiments, the model has more than 10,000 parameters, more than 100,000 parameters, or 1 × 10 6It includes more than 1000 parameters. In some embodiments, the model includes 1000 parameters. In some embodiments, the model consists of 10 to 10 million parameters. In some embodiments, the model includes at least 10, at least 50, at least 100, at least 500, at least 1000, at least 2000, at least 5000, at least 10,000, at least 20,000, at least 50,000, at least 100,000, at least 200,000, at least 500,000, at least 1 million, at least 2 million, at least 3 million, at least 4 million, or at least 5 million parameters. In some embodiments, the model consists of 8 million or fewer parameters, 5 million or fewer parameters, 4 million or fewer parameters, 1 million or fewer parameters, 500,000 or fewer parameters, 100,000 or fewer parameters, 50,000 or fewer parameters, 10,000 or fewer parameters, 5,000 or fewer parameters, 1,000 or fewer parameters, or 500 or fewer parameters. In some embodiments, the first model consists of 10 to 5,000 parameters, 500 to 10,000 parameters, 10,000 to 500,000 parameters, 20,000 to 1 million parameters, or 1 million to 5 million parameters. In some embodiments, the number of parameters in the model is in another range, starting with 10 or more parameters and ending with 8 million or fewer parameters.

[0118] In some embodiments, the number of model parameters and the amount of information in the first graph, or other representations of the first compound, input into Model 702, result in more than 10,000 calculations, more than 100,000 calculations, or 1 × 10⁻¹⁶ calculations to compute the first isometrically retained latent embedding. 6 Calculations exceeding 5 × 10 6 Calculations exceeding the number of times, or 1 × 10 7 It delivers performance exceeding 10

[0119] In some embodiments, compounds are represented in forms other than graphs for input into the models of this disclosure. For example, in some embodiments, compounds are represented as fingerprints, and the compound fingerprints are input into the model to generate isometrically retained latent embeddings of the molecule. Non-limiting examples of compound fingerprint representations are shown in the definitions section above.

[0120] In some embodiments, compounds are represented as descriptors, and these descriptors are input into a model to generate an isometrically retained latent embedding of the compound. Descriptors include, but are not limited to, physicochemical descriptors, topological descriptors, electronic descriptors, geometric descriptors, functional group numbers, and toxicological and environmental descriptors, as well as any combination thereof.

[0121] Examples of physicochemical descriptors, though not limited to them, include molecular weight (the sum of the atomic weights of all atoms in a compound), LogP, penalized logP, compound acid dissociation constant (pKA), compound acid dissociation constant (kA), compound topological polar surface area (TPSA), polar surface area (PSA) (the surface area of ​​a molecule occupied by polar atoms and functional groups, often used to predict membrane permeability and bioavailability), solubility (a prediction of the solubility of a compound in different solvents or aqueous solutions), and drug similarity. See, for example, Prasanna and Doerksen et al., 2009, “Topological Polar Surface Area: A Useful Descriptor in 2D-QSAR,” Current Medicinal Chemistry 16(1), pp.21-41, which is incorporated herein by reference.

[0122] LogD is the distribution constant of a compound's lipophilicity. In some embodiments, LogD is calculated from the compound's chemical structure (cLogD). Generally, cLogD refers to the calculated distribution constant of a molecule's lipophilicity. This is typically determined for a compound in an aqueous phase that can be adjusted to a specific pH using a buffer. This parameter can serve as an estimate of the compound's overall lipophilicity, such as solubility, permeability across biological membranes, hepatic clearance, lack of selectivity, and / or nonspecific toxicity, which affect the compound's behavior in a range of biological processes related to drug discovery. In some cases, lipophilicity can be used as an estimate of drug similarity. See Cambridge MedChem Consulting, 2019, “Lipophilicity,” available at cambridgemedchemconsulting.com / resources / physiochem / logD.html, which is incorporated herein by reference in its entirety.

[0123] logP is a measure of the hydrophilicity of a compound. The fractionation coefficient (abbreviated as P) is defined specifically for unionized solutes as a particular ratio of the solute concentrations between two solvents (two phases of the liquid phase), and the logarithm of this ratio is log P. As used herein, the numerator of this ratio is the concentration of the compound in octanol, and the denominator is the concentration of the compound in water.

[0124] Penalized logP is a logP score that also takes ring size and synthetic accessibility into account. See, for example, Ertl, 2009, “Estimation of synthetic accessibility score of drug-like molecules,” J. Cheminform. This is incorporated herein by reference.

[0125] In some embodiments, pKA is estimated using the structure of the compound. See, for example, Seybold and Shields, 2015, “Computational estimation of pKa values,” WIREs Comput Mol Sci 5, pp. 290–297, which is incorporated herein by reference.

[0126] The topological polar surface area (TPSA) of a compound is a measure of its polar surface area that utilizes the contribution of functional groups based on a large database of structures, avoiding the need to calculate the three-dimensional structure of the compound or to determine which biological conformation(s) are relevant. See, for example, Prasanna and Doerksen et al., 2009, “Topological Polar Surface Area: A Useful Descriptor in 2D-QSAR,” Current Medicinal Chemistry 16(1), pp.21-41, which is incorporated herein by reference.

[0127] Topological descriptors, though not limited to them, include the number of atoms and bonds (the number of atoms and bonds in a compound), molecular bonding indexes (graph-based indices representing structural bonding, such as the Wiener index, Balaban index, and Zagreb index), and molecular diameter (a measure of molecular size or compactness, often calculated as the maximum distance between any two atoms in the molecule). Balaban and Ivanciuc, 1999, Historical Development of Topological Indices, CRC press, 1 st Please refer to the Edition, which is incorporated into this specification by reference.

[0128] Examples of electron descriptors, though not limited to them, include electronegativity (a measure of an atom's tendency to attract electrons in chemical bonds), ionization potential (the energy required to remove an electron from an atom or molecule), and electron density (the distribution of electron charge within a molecule, often represented using molecular orbitals or electron density maps).

[0129] Geometric descriptors include, but are not limited to, bond length and angle (measurements of the length and angle between atoms within a molecule), torsion angle (measurements of the rotation around a single bond within a molecule), and surface area and volume (geometric properties that characterize the shape and size of a molecule).

[0130] Functional group descriptors are not limited to, but include the number of specific functional groups or substructures within a compound, such as hydroxyl groups, aromatic rings, or carboxylic acid groups.

[0131] Toxicological and environmental descriptors include, but are not limited to, toxic hors (structural features related to the toxicity or adverse effects of a compound) and environmental trend descriptors (predictions of the compound's behavior in the environment, including biodegradability, persistence, and potential for bioaccumulation).

[0132] The QED score is described in Bickerton et al., 2012, “Quantifying the chemical beauty of drugs,” Nature Chemistry 4(2), pp. 90–98. This is incorporated herein by reference.

[0133] Non-restrictive examples of compound descriptors are disclosed in Seko, 2017, “Representation of compounds for machine-learning prediction of physical properties,” Phys. Rev. B 95, 144110, which is incorporated herein by reference.

[0134] In some embodiments, compounds are represented as three-dimensional structural representations, which are input into a model to generate isometrically retained latent embeddings of the molecule. Non-limiting examples of three-dimensional structural representations of compounds include, but are not limited to, drug-Hore models (hypothetical representations of the spatial arrangement of chemical features of a compound, such as hydrogen bond donors, acceptors, aromatic rings, etc.), conformational ensembles (representations of multiple conformations or energetically accessible states of a compound), and quantum mechanical calculations, such as those generated by abu initio and density function theory (DFT) methods, for calculating the electronic structure and energy of a molecule, which can inform the three-dimensional geometry.

[0135] In some embodiments, compounds are represented as tensor-based representations, which are input into a model to generate isometrically retained latent embeddings of the molecules. Non-limiting examples of tensor-based representations of compounds are disclosed in Huo and Rupp, 2022, “Unified representation of molecules and crystals for machine learning,” Machine Learning 3,045017, which is incorporated herein by reference.

[0136] In some embodiments, the compound is represented as a string representation, and this string representation of the compound is input into the model to generate an isometrically retained latent embedding of the molecule. Non-exclusive examples of string representations of compounds include, but are not limited to, Simplified Molecular Input Line Entry System (SMILES; see, e.g., Honda et al., 2019, “SMILES Transformer: Pre-trained Molecular Fingerprint for Low Data Drug Discovery,” arXiv:1911.04738), SMiles ARbitrary Target Specification (SMARTS; see, e.g., SMARTS - A Language for Describing Molecular Patterns,” 2022 on the Internet at daylight.com / dayhtml / doc / theory / theory.smarts.html (accessed December 2020)), and DeepSMILES (see, e.g., O'Boyle and Dalke, 2018, “DeepSMILES: an adaptation of SMILES for use in machine-learning of chemical structures,” Preprint at ChemRxiv. Examples include the self-referencing embedded string (SELFIES) format (see https: / / doi.org / 10.26434 / chemrxiv.7097960.v1.) and the self-referencing embedded string (SELFIES) format (see Krenn et al., 2022, “SELFIES and the future of molecular string representations,” Patterns 3(10), pp.1–27). These are incorporated herein by reference.

[0137] Block 218. Referring to Block 218, the second graph, or other form of representation of the first compound, is input into the model. Figure 7 shows two models, 702 and 704, but in a typical embodiment, models 702 and 704 are actually the same model. That is, the first graph (or other form of representation) of the first compound is input into model 702 to obtain the first isometrically retained latent embedding, and then the second graph (or other form of representation) of the second compound is input into model 702 for the second isometrically retained latent embedding. Two models 702 / 704 are depicted in Figure 7 to better illustrate the comparison of model outputs. In some embodiments, the number of model parameters and the amount of information in the second graph, or other form of representation of the second compound, input into the model, result in more than 10,000 calculations, more than 100,000 calculations, or 1 × 10⁻¹⁶ calculations to compute the second isometrically retained latent embedding. 6 Calculations exceeding 5 × 10 6 Calculations exceeding the number of times, or 1 × 10 7 It delivers performance exceeding 10

[0138] As used herein, in some embodiments, the first isometrically preserved latent embedding and the second isometrically preserved latent embedding refer to latent embeddings generated by a model that, within a threshold, adheres to the symmetry and self-distance requirements illustrated in Figure 6.

[0139] As used herein, in some embodiments, the first isometrically retained latent embedding and the second isometrically retained latent embedding are used to calculate the symmetry properties illustrated in Figure 6 within a threshold across a test dataset of compounds, such as those illustrated in Figure 6, which are used to calculate the statistics illustrated in Figure 6.

number

[0140] When used herein, in some embodiments, the first isometrically retained latent embedding and the second isometrically retained latent embedding are used to calculate the self-inter

number

[0141] Blocks 220-222. Referring to block 220, in some embodiments, the model is a message-passing network. Referring to block 222, in some embodiments, the model is a graph neural network.

[0142] In some embodiments, the model is any model that takes a representation of a compound and outputs an isometrically retained latent embedding that represents the compound.

[0143] Isometry-preserving latent embeddings maintain the geometric properties of the input space within the latent space. In some embodiments, the model used to generate such isometry-preserving latent embeddings is an autoencoder.

[0144] In some embodiments, the model is a variational autoencoder (VAE). A VAE is a generative model that learns to probabilistically encode and decode data points. They are trained to maximize the likelihood of generating input data while simultaneously minimizing the divergence between a learned latent distribution and a prior distribution (typically a Gaussian distribution). By encouraging the latent space to follow a specific distribution (e.g., a Gaussian distribution), VAEs aim to implicitly preserve local structure and distance relationships in the input space, which can result in some degree of isometric preservation. See, for example, Kingma and Max, 2019, Foundations and Trends in Machine Learning, 12(4), ISSN 1935-8237, which is incorporated herein by reference.

[0145] In some embodiments, the model is an adversarial encoder (AAE). An AAE combines the principle of variational autoencoders with adversarial training. In addition to accurately reconstructing the input data, they also train a discriminator to distinguish between samples from the true latent distribution and samples from the encoder's distribution. This adversarial training component encourages the encoder to generate latent embeddings that are indistinguishable from samples drawn from the prior distribution, which can result in improved isometric retention.

[0146] In some embodiments, the model is a normalized encoder. Various normalization techniques can be applied to a standard autoencoder to facilitate isometric preservation in the latent space. For example, a penalty term is added to the loss function to punish deviations from a desired structure or distribution in the latent space. In some embodiments, as described in the Examples section, L2 norm normalization is applied to the autoencoder to preserve isometrics.

[0147] In some embodiments, the model is a sparse autoencoder, a denoising autoencoder (see, for example, Zheng, 2019, Stacked Denoising Auto-Encoder for Short-Term Load Forecasting: Deep Learning with Stacked Denoising Auto-Encoder Algorithm, Lambert Academic Publishing), or a contractive autoencoder. In some embodiments, such autoencoders indirectly promote isometric preservation by facilitating robust representations that capture prominent features of the input data.

[0148] In some embodiments, the model is a metric-learning autoencoder. As used herein, a metric-learning autoencoder is explicitly designed to learn embeddings that preserve a specific distance metric or relationship in the input space. These models incorporate loss functions that directly facilitate isometric preservation, such as triplet loss or contrastive loss. By optimizing these loss functions during training, the autoencoder learns to map similar input samples to nearby points in the latent space, thereby preserving local distance and geometric structure.

[0149] In some embodiments, the model is a graph neural network (GNN), also known as a message-passing network. See Baskin et al., 1997, “A neural device for searching direct correlations between structures and properties of chemical compounds,” Journal of Chemical Information and Computer Sciences, 37(4):715-721; Sperduti & Starita, 1997, “Supervised neural networks for the classification of structures,” IEEE Transactions on Neural Networks 8(3), pp.714-735; Scarselli et al., 2008, “The graph neural network model,” IEEE Transactions on Neural Networks 20(1), pp.61-80; and Gilmer et al., 2017, “Message Passing Neural Networks,” Schutt et al. (eds), Machine Learning Meets Quantum Physics, Lecture Notes in Physics 968, Springer, Cham. Each of these is incorporated herein by reference.

[0150] In some embodiments, the model is a graph convolutional neural network (GCN). Generally, message passing in GNNs is computed in a nonlinear neural network, which can result in enormous computational costs. In contrast, GCN (Kipf and Welling, 2017) is a considerably simplified version of GNN and uses a linear message passing scheme expressed as follows:

number

number

number

number

number

number

number

[0151] Non-limiting examples of GCNs include Behler Parrinello, 2007, “Generalized Neural-Network Representation of High Dimensional Potential-Energy Surfaces,” Physical Review Letters 98,146401, Chmiela et al., 2017, “Machine learning of accurate energy-conserving molecular force fields,” Science Advances 3(5):e1603015, Schutt et al. al., 2017, “SchNet: A continuous-filter convolutional neural network for modeling quantum interactions,” Advances in Neural Information Processing Systems 30, pp. 992-1002, Feinberg et al., 2018, “PotentialNet for Molecular Property Prediction,” ACS Cent.Sci.4,11,1520-1530, and Stafford et al., 2022, “AtomNet PoseRanker:Enriching Ligand Pose Quality for Dynamic Proteins This is disclosed in "Virtual High Throughput Screens," J. Chem Inf Model 62, pp. 1178–1189. Each of these is incorporated herein by reference.

[0152] In some embodiments, as shown in Figure 18, the message passing network is used for the model rather than the GCN. In some embodiments, as shown in Figure 19, the message passing network is used for the model rather than the GCN. In some embodiments, as shown in Figure 20, the message passing network is used for the model rather than the GCN.

[0153] In some embodiments, the model is an arbitrary encoder that takes a graph of a compound, or another form of representation of a compound, and outputs a vector representing the compound.

[0154] Blocks 224-228. Referring to block 224, in some embodiments, the model is a message-passing graph neural network with 3 to 8 layers. Referring to block 226, in some embodiments, the model is a message-passing graph neural network with 5 layers. Referring to block 228, in some embodiments, the model has more than 500,000 parameters, or 1 × 10⁶ 6 It contains more than one parameter.

[0155] Block 230. Referring to Block 230, in some embodiments, the model performs at least 100,000 calculations to compute a first isometrically preserved latent embedding, and the model performs at least 100,000 calculations to compute a second isometrically preserved latent embedding.

[0156] Blocks 232-234. Referring to block 232, in some embodiments, an estimate of the graph edit distance between a first compound and a second compound is provided as the L2 norm difference between a first isometrically retained latent embedding and a second isometrically retained latent embedding. Referring to block 234, in some embodiments, the first isometrically retained latent embedding is a first vector containing a first set of vector values, and the second isometrically retained latent embedding is a second vector containing a second set of vector values, and the L2 norm difference between the first vector and the second vector is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first vector and the second vector.

[0157] Advantageously, by using the L2 norm difference between the outputs of the first and second outputs of the model as the GED estimate, all properties of the distance metric outlined at the top of Figure 6, except for positive valuation, are satisfied.

[0158] Block 236. One practical application / technical effect of GED is to find chemically related analogues. For example, consider a drug discovery effort where the first compound is a known lead compound. In this case, one or more databases of compounds can be used to search for compounds that are chemically related analogues by calculating the GED distance between the first compound and each compound in the one or more databases of compounds. An example of such a database of compounds is the PubChem database (see Kim et al., 2021, “PubChem in 2021: new data content and improved web interfaces,” Nucleic Acids Research 8(49)(D1), pp. D1388–D1395, which is incorporated herein by reference). Another example of such a database is the zinc database (see Irwin et al., 2020, “ZINC20-A Free Ultralarge-Scale Chemical Database for Ligand Discovery,” J. Chem. Inf. Model 60(12), 6065-6073, which is incorporated herein by reference). Another example of such a database is the ChEMBL database, a manually curated database of bioactive molecules with drug-like properties. See Gaulton et al., 2019, “ChEMBL: towards direct deposition of bioassay data,” Nucleic Acids Res. 47(D1), D930-D940, which is incorporated herein by reference.For further examples of such databases, see MCULE (Kiss et al., 2012, “Http: / / Mcule.Com: A Public Web Service for Drug Discovery,” J. Cheminformatics 4 (1), p. 17.) and ENAMINE (Irwin et al., 2016, “Docking Screens for Novel Ligands Conferring New Biology,” J. Med. Chem. 59(9), pp. 4103-4120). Each of these is incorporated herein by reference.

[0159] In some embodiments, the GED distance between a first compound and each of 10 or more compounds in one or more databases of compounds is calculated in such an effort.

[0160] In some embodiments, the GED distance between a first compound and each of 100 or more compounds in one or more databases of compounds is calculated in such an effort.

[0161] In some embodiments, the GED distance between a first compound and each of 1000 or more compounds in one or more databases of compounds is calculated in such an effort.

[0162] In some embodiments, the GED distance between a first compound and each of 10,000 or more compounds in one or more databases of compounds is calculated in such an effort.

[0163] In some embodiments, the GED distance between a first compound and each of 100,000 or more compounds in one or more databases of compounds is calculated in such an effort.

[0164] In some embodiments, a first compound and 1 × 10 are included in one or more databases of compounds. 6 pieces, 2×10 6 pieces, 3×106 pieces, 4×10 6 pieces, 5×10 6 pieces, 6×10 6 pieces, 7×10 6 pieces, 8×10 6 individual pieces, or 9 x 10 6 The GED distance between each of the individual or more compounds is calculated in such an effort.

[0165] In some embodiments, a first compound and 1 × 10 are included in one or more databases of compounds. 7 pieces, 2×10 7 pieces, 3×10 7 pieces, 4×10 7 pieces, 5×10 7 pieces, 6×10 7 pieces, 7×10 7 pieces, 8×10 7 individual pieces, or 9 x 10 7 The GED distance between each of the individual or more compounds is calculated in such an effort.

[0166] In the drug discovery process, lead compounds are biologically active molecules that are likely to be useful but may require modification to better suit the target or to have better pharmacological properties. These compounds, having a threshold estimated GED below (e.g., <10, <9, <8, <7, <6, <5, <4, <3, or <2), can then be evaluated as chemically related analogs of the first compound and thus can be evaluated to determine whether they are better suited to the target or have better pharmacological properties. Referring to block 236, in some embodiments, it is determined whether a second compound is a chemically related analog of the first compound when the estimated GED is below a threshold. In some embodiments, this threshold is any real number between 2 and 10, including 2 and 10. In some embodiments, the threshold is 2, 3, 4, 5, 6, 7, 8, 9, or 10. In some embodiments, the threshold is 2.5, 3.5, 4.5, 5.5, 6.5, 7.5, 8.5, or 9.5.

[0167] Block 238. Another practical application / technical effect of GED is to ensure diversity within a chemical library or among chemical substances selected for a drug discovery program. For example, one or more databases of compounds can be clustered using GED as a distance metric to form clusters, and then compounds from each of the resulting clusters can be selected for evaluation in a drug discovery program.

[0168] One example of such a database of compounds is the PubChem database (see Kim et al., 2021, “PubChem in 2021: new data content and improved web interfaces,” Nucleic Acids Research 8(49)(D1), pp.D1388-D1395, which is incorporated herein by reference). Another example of such a database is the zinc database (see Irwin et al., 2020, “ZINC20-A Free Ultralarge-Scale Chemical Database for Ligand Discovery,” J.Chem.Inf.Model 60(12), 6065-6073, which is incorporated herein by reference). Yet another example of such a database is the ChEMBL database, a manually curated database of bioactive molecules with drug-like properties. See Gaulton et al., 2019, “ChEMBL: towards direct deposition of bioassay data,” Nucleic Acids Res. 47(D1), D930-D940, which is incorporated herein by reference. Further examples of such databases are MCULE (Kiss et al., 2012, “Http: / / Mcule.Com: A Public Web Service for Drug Discovery,” J. Cheminformatics 4 (1), p. 17) and ENAMINE (Irwin et al., 2016, “Docking Screens for Novel Ligands Conferring New Biology,” J. Med. Chem. 59 (9), pp. 4103-4120), each of which is incorporated herein by reference.

[0169] Accordingly, referring to block 238, in some embodiments, for each compound-specific pair in the compound database, the process is repeated: optionally obtain a first graph (block 202), optionally obtain a second graph (block 204), input the first graph or another representation of the first compound (block 216), input the second graph or another representation of the second compound (block 218), and provide an approximate GED, where the first compound in each compound-specific pair is the first compound, the second compound in each compound-specific pair is the second compound, thereby forming a plurality of estimated graph edit distances. The plurality of graph edit distances are then clustered into a plurality of clusters by evaluating the plurality of distances using a criterion function. Compounds are then selected from one or more clusters within the plurality of clusters for evaluation in the drug discovery program. In some embodiments, compounds are selected from two, three, four, five, six, seven, eight, nine, or ten or more different clusters for the drug discovery program. Some embodiments include 10 or more, 100 or more, 1,000 or more, 10,000 or more, 100,000 or more, or 1 × 10 6 One or more compounds are selected from each cluster within one or more clusters.

[0170] In some embodiments, more than 100 compound-specific pairs are evaluated for block 238. In some embodiments, more than 1,000 compound-specific pairs are evaluated for block 238. In some embodiments, more than 10,000 compound-specific pairs are evaluated for block 238. In some embodiments, more than 100,000 compound-specific pairs are evaluated for block 238. In some embodiments, 1 × 10⁶ 6 One or more compound-specific pairs are evaluated.

[0171] In some embodiments, block 238 is 1 × 10 6 pieces, 2×10 6 pieces, 3×10 6pieces, 4×10 6 pieces, 5×10 6 pieces, 6×10 6 pieces, 7×10 6 pieces, 8×10 6 individual pieces, or 9 x 10 6 One or more compound-specific pairs are evaluated. In some embodiments, for block 238, 1 × 10 7 pieces, 2×10 7 pieces, 3×10 7 pieces, 4×10 7 pieces, 5×10 7 pieces, 6×10 7 pieces, 7×10 7 pieces, 8×10 7 individual pieces, or 9 x 10 7 One or more compound-specific pairs are evaluated. In some embodiments, for block 238, 1 × 10 8 pieces, 2×10 8 pieces, 3×10 8 pieces, 4×10 8 pieces, 5×10 8 pieces, 6×10 8 pieces, 7×10 8 pieces, 8×10 8 individual pieces, or 9 x 10 8 One or more compound-specific pairs are evaluated.

[0172] In some embodiments, the clustering is unsupervised clustering. In some embodiments, the clustering is supervised clustering. A clustering algorithm suitable for use according to block 348 is described, for example, on pages 211–256 of Duda and Hart, Pattern Classification and Scene Analysis, 1973, John Wiley & Sons, Inc., New York (hereinafter, "Duda 1973"). This is incorporated herein by reference in its entirety. As an illustrative example, in some embodiments, the clustering problem is described as one of the headings for natural grouping in a dataset. To identify natural groupings, two problems are addressed. First, determine how to measure the similarity (or dissimilarity) between two samples. This metric (e.g., a similarity measure) is used to ensure that samples in one cluster are more similar to each other than samples in other clusters. Second, using the similarity measure, determine a mechanism for dividing the data into clusters. One way to begin a clustering investigation is to define a distance function and compute a matrix of distances between all pairs of samples in the training set. If distance is a good measure of similarity, the distance between reference entities within the same cluster is significantly smaller than the distance between reference entities in different clusters. However, in some implementations, clustering does not use a distance metric. For example, in some embodiments, a non-metric similarity function s(x, x') is used to compare two latent embeddings x and x'. In some such embodiments, s(x, x') is a symmetric function with large values ​​when x and x' are "similar" in some way. Once a method is chosen for measuring the "similarity" or "dissimilarity" between latent embeddings, clustering uses a criterion function to measure the clustering quality of any partition of the data. Partitions of multiple compound pairs that make the criterion function extreme are used to cluster the data.Specific exemplary clustering techniques intended for use in this disclosure include, but are not limited to, hierarchical clustering (agglomerative clustering using nearest neighbor algorithms, farthest neighbor algorithms, mean-connected algorithms, centroid algorithms, or sum-of-squares algorithms), k-means clustering, fuzzy k-means clustering algorithms, and Jarvis-Patrick clustering. In some embodiments, the clustering includes unsupervised clustering (e.g., without prior determination of a predetermined number of clusters and / or cluster assignments).

[0173] Block 240. Another practical application / technical effect of GED is the construction of quantitative structure-activity relationship (QSAR) models. A QSAR model is a computational or mathematical model that attempts to find a significant correlation between molecular structure and molecular activity. The vast amount of available data on chemical compounds and their reactivity necessitates computational tools to reduce the execution time of drug synthesis and testing cycles. These tools are crucial when analyzing activity data and creating new models for virtual screening techniques. In the case of QSAR models, the GED distance forms the basis for quantifying differences in molecular structure. See, for example, Garcia-Hernandez, 2020, “Learning the Edit Costs of Graph Edit Distance Applied to Ligand-Based Virtual Screening,” Current Topics in Medicinal Chemistry 20, 1582-1592, which is incorporated herein by reference. Thus, referring to Block 240, in some embodiments, estimates of the GED are used to construct quantitative structure-activity relationship models involving first and second compounds.

[0174] In some embodiments, the QSAR is constructed in relation to a first compound, which is a known lead compound in a drug discovery effort. In this example, one or more databases of compounds can be searched for compounds having a predetermined GED of the first compound, and then the QSAR is constructed using these compounds.

[0175] One example of such a database of compounds is the PubChem database (see Kim et al., 2021, “PubChem in 2021: new data content and improved web interfaces,” Nucleic Acids Research 8(49)(D1), pp.D1388-D1395, which is incorporated herein by reference). Another example of such a database is the zinc database (see Irwin et al., 2020, “ZINC20-A Free Ultralarge-Scale Chemical Database for Ligand Discovery,” J.Chem.Inf.Model 60(12), 6065-6073, which is incorporated herein by reference). Yet another example of such a database is the ChEMBL database, a manually curated database of bioactive molecules with drug-like properties. See Gaulton et al., 2019, “ChEMBL: towards direct deposition of bioassay data,” Nucleic Acids Res. 47(D1), D930-D940, which is incorporated herein by reference. Further examples of such databases are MCULE (Kiss et al., 2012, “Http: / / Mcule.Com: A Public Web Service for Drug Discovery,” J. Cheminformatics 4 (1), p. 17) and ENAMINE (Irwin et al., 2016, “Docking Screens for Novel Ligands Conferring New Biology,” J. Med. Chem. 59(9), pp. 4103-4120), each of which is incorporated herein by reference.

[0176] Therefore, in some embodiments, the first compound and 10, 100, 1,000, 10,000, 100,000, and 1 × 10¹⁶ compounds from one or more databases of compounds. 6pieces, or 1 x 10 7 The GED distance between each of the first compound and the first compound is calculated, and those compounds that fall within the threshold-estimated GED of the first compound are used to determine the QSAR. In some embodiments, the threshold-estimated GED is less than 10, less than 9, less than 8, less than 7, less than 6, less than 5, less than 4, less than 3, or less than 2. In some embodiments, the threshold-estimated GED is any real number between 2 and 10, including 2 and 10. In some embodiments, the threshold-estimated GED is 2, 3, 4, 5, 6, 7, 8, 9, or 10. In some embodiments, the threshold-estimated GED is 2.5, 3.5, 4.5, 5.5, 6.5, 7.5, 8.5, or 9.5. In some embodiments, 10 to 90 percent of the compounds evaluated for the estimated GED with respect to the first compound in one or more databases fall within the threshold-estimated GED. In some embodiments, 20 to 80 percent of compounds evaluated for the estimated GED with respect to a first compound in one or more databases are within the threshold estimated GED. In some embodiments, at least 20, 30, 40, 50, or 60 percent of compounds evaluated for the estimated GED with respect to a first compound in one or more databases are within the threshold estimated GED. In some embodiments, less than 90, 80, 70, 60, or 50 percent of compounds evaluated for the estimated GED with respect to a first compound in one or more databases are within the threshold estimated GED.

[0177] Block 242. Another practical application / technical effect of GED is defining the activity cliff. An activity cliff occurs when a limited number of modifications are made to a molecule, such as one or two small changes, and the activity of the derived molecule drops orders of magnitude compared to the original molecule. Knowledge of the activity cliff is an important aspect of structure-activity analysis. See, for example, Wassermann, et al., 2010, “Activity Landscape Representations for Structure-Activity Relationship Analysis,” J.Med.Chem. 53, 8209-8233, which is incorporated herein by reference. Thus, referring to Block 242, in some embodiments, estimates of the GED are used to determine the activity cliff for a drug discovery program.

[0178] Additional Embodiments Another aspect of the present disclosure provides a computer system for estimating a graph edit distance between a first compound and a second compound. The computer system comprises one or more processors and a memory addressable by one or more processors. The memory stores at least one program for execution by one or more processors. The at least one program includes (A) instructions for inputting a first representation of the first compound into a model having a plurality of parameters. The model applies the plurality of parameters to the first representation of the first compound to produce a first latent embedding representing the first compound as a first output from the model. The at least one program further includes (B) instructions for inputting a second representation of the second compound into the model. The model applies the plurality of parameters to the second representation of the second compound to produce a second latent embedding representing the second compound as a second output from the model. The at least one program further includes (C) instructions for providing an estimate of the graph edit distance between the first compound and the second compound using the difference between the first latent embedding and the second latent embedding.

[0179] In some embodiments, the difference is the L2 norm difference between a first latent embedding and a second latent embedding. In some embodiments, the first latent embedding is a first vector containing a first set of vector values, and the second latent embedding is a second vector containing a second set of vector values, and the L2 norm difference between the first and second vectors is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first and second vectors. In some such embodiments, the model is trained such that the L2 norm of the difference between the first and second vectors is an estimate of the GED between the first and second numerators.

[0180] In some embodiments, at least one program further includes instructions for determining that a second compound is a chemically related analog of a first compound when the estimated graph edit distance falls below a threshold.

[0181] In some embodiments, at least one program further includes instructions for repeating input (A), input (B), and provide (C) for each compound-specific pair in a compound database, where the first compound in each compound-specific pair is the first compound, the second compound in each compound-specific pair is the second compound, and so on, thereby forming multiple graph edit distances. In some such embodiments, the multiple graph edit distances are clustered into multiple clusters by evaluating the multiple distances using a criterion function, and compounds from one or more of the multiple clusters are selected for evaluation in a drug discovery program.

[0182] In some embodiments, at least one program further includes instructions for constructing a quantitative structure-activity relationship model comprising a first compound and a second compound, using an estimate of the graph edit distance.

[0183] In some embodiments, at least one program further includes instructions for determining the activity cliff for a drug target using an estimate of the graph edit distance.

[0184] In some embodiments, the model is a graph convolutional neural network.

[0185] In some embodiments, the model is a graph neural network (message passing network).

[0186] In some embodiments, the model is a message-passing graph neural network consisting of 3 to 8 layers.

[0187] In some embodiments, the model is a message-passing graph neural network consisting of five layers.

[0188] In some embodiments, the model is an autoencoder. In some such embodiments, the autoencoder is a variational autoencoder, an adversarial autoencoder, a normalized autoencoder, a sparse autoencoder, or a metric learning autoencoder.

[0189] In some embodiments, the first representation is a first fingerprint of the first compound, and the second representation is a second fingerprint of the second compound.

[0190] In some embodiments, the first expression is a plurality of first descriptors of a first compound, and the second expression is a plurality of second descriptors of a second compound.

[0191] In some embodiments, the first representation is a first three-dimensional structural representation of the first compound, and the second representation is a second three-dimensional structural representation of the second compound.

[0192] In some embodiments, the first representation is a first tensor-based representation of the first compound, and the second representation is a second tensor-based representation of the second compound.

[0193] In some embodiments, the first representation is a first string representation of the first compound, and the second representation is a second string representation of the second compound.

[0194] In some embodiments, the model includes more than 500,000 parameters. In some embodiments, the model is 1 × 10 6 It contains more than one parameter.

[0195] In some embodiments, the first and second compounds each satisfy two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 Daltons, and (iv) a LogP of less than 5.

[0196] In some embodiments, the first and second compounds have accurate graph editing distances of 1 to 10. In some embodiments, the first and second compounds have accurate graph editing distances of 1 to 15.

[0197] In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than 500 daltons. In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than 1000 daltons. In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than 2000 daltons. In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of less than 4000 daltons, less than 6000 daltons, less than 8000 daltons, less than 10000 daltons, or less than 20000 daltons. In some embodiments, the first compound and the second compound are each organic compounds having a molecular weight of 300 daltons to 1500 daltons.

[0198] In some embodiments, the model performs at least 100,000 calculations to compute a first latent embedding, and the model performs at least 100,000 calculations to compute a second latent embedding.

[0199] In some embodiments, the first latent implant is a first isometrically retained latent implant. In some embodiments, the second latent implant is a second isometrically retained latent implant.

[0200] Another aspect of the present disclosure provides a method for estimating a graph edit distance between a first compound and a second compound. The method includes (A) inputting a first representation of the first compound into a model comprising a plurality of parameters (e.g., on a computer system having one or more processors and memory addressable by one or more processors). The model applies the plurality of parameters to the first representation of the first compound to produce a first latent embedding representing the first compound as a first output from the model. The method further includes (B) inputting a second representation of a second compound into the model (e.g., on a computer system having one or more processors and memory addressable by one or more processors). The model applies the plurality of parameters to the second representation of the second compound to produce a second latent embedding representing the second compound as a second output from the model. The method further includes (C) using the difference between the first latent embedding and the second latent embedding to provide an estimate of the graph edit distance between the first compound and the second compound (e.g., on a computer system having one or more processors and memory addressable by one or more processors). In some embodiments, the first latent implant is a first isometrically retained latent implant. In some embodiments, the second latent implant is a second isometrically retained latent implant. In some embodiments, the difference is the L2 norm difference between the first latent implant and the second latent implant.

[0201] Another aspect of the present disclosure provides a non-temporary computer-readable storage medium. The non-temporary computer-readable storage medium stores instructions that, when executed by a computer system, cause the computer system to perform a method for estimating the graph edit distance between a first compound and a second compound. The method includes (A) inputting a first representation of the first compound into a model that includes a plurality of parameters. The model applies the plurality of parameters to the first representation of the first compound to produce a first latent embedding representing the first compound as a first output from the model. The method further includes (B) inputting a second representation of the second compound into the model. The model applies the plurality of parameters to the second representation of the second compound to produce a second latent embedding representing the second compound as a second output from the model. The method further includes (C) using the difference between the first latent embedding and the second latent embedding to provide an estimate of the graph edit distance between the first compound and the second compound. In some embodiments, the first latent embedding is a first isometrically preserved latent embedding. In some embodiments, the second latent implant is a second isometrically retained latent implant. In some embodiments, the difference is the L2 norm difference between the first latent implant and the second latent implant.

[0202] Training a model to predict the efficacy of a test compound against a target protein. Successful drug discovery campaigns rely on identifying biologically active lead molecules that are chemically distinct from known compounds for disease targets. This is particularly challenging when there is little or no readily available ligand data nearby, as is the case with novel targets, or when the novel scaffold is far removed in chemical space. Structure-based virtual high-throughput screening (vHTS) methods are designed to overcome this challenge by identifying novel compounds with predicted activity from extensive chemical libraries, e.g., MCULE (Kiss et al., http: / / Mcule.Com: A Public Web Service for Drug Discovery. J. Cheminformatics 2012, 4(1), P17) or ENAMINE (Irwin and Shoichet, 2016, Docking Screens for Novel Ligands Conferring New Biology, J. Med. Chem. 59(9), 4103-4120). Therefore, virtual screening is a tool for identifying novel inhibitors of proteins with known structures. See, for example, Wallach et al., 2015, “AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery,” doi:29 10.48550 / arxiv.1510.02855, Lyu et al., 2019, “Ultra large library docking for discovering new chemotypes,” Nature 566:7743 566,20 224-229. doi:10.1038 / s41586-019-0917-9, and Bender et al., 2021, “A practical guide to large-scale docking,” Nature Protocols 16:10 16,39 4799-4832. Each of these is incorporated by reference. Conventional structure-based virtual high-throughput screening approaches rank compounds by docking ligands primarily to rigid receptors using empirical or force-field-based scoring functions. (See Trott and Olson, 2010, “AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading,” J Comput Chem 31, 455-461, which is incorporated herein by reference.) Docking to structures that deviate from the ligand-binding state can lead to inaccurate predictions of the binding complex and inadequate compound ranking. For example, recovering active compounds is often difficult when docking to experimental structures that do not contain the ligand (e.g., the apo state) or when the homologous ligand is small.See, for example, Abagyan et al., 2010, “Recipes for the selection of experimental protein conformations for virtual screening,” J Chem Inf Model 50, 32 186-193. This is incorporated herein by reference. Another obstacle is that experimentally derived structures are not available for many protein targets that have disordered or flexible domains. AlphaFold (AF) has the potential to accelerate drug discovery thanks to its accurate structural prediction of such proteins. See Jumper et al., 2024, “Highly accurate protein structure prediction with AlphaFold,” Nature, doi:41 10.1038 / s41586-021-03819-2. This is incorporated herein by reference. However, these are still merely rigid structures, and their usefulness will be limited if they do not represent binding-like structures. For example, see Vijayan et al., 2015, “Conformational analysis of the DFG-out kinase motif and biochemical profiling of structurally validated type II inhibitors,” J Med Chem 58, 466-479, and Wankowicz et al., 2022, “Ligand binding remodels protein side chain conformational heterogeneity,” Elife 11. doi:10.7554 / ELIFE.74114. Each of these is incorporated herein by reference.

[0203] Applying deep learning to structure-based drug discovery using vHTS methods provides pharmaceutical companies with a path to identifying promising new compounds. These methods are likely to have a significant impact on research areas that have been challenging for pharmaceutical companies, such as studies focused on identifying compounds of target proteins with multiple binding sites. However, using these computational methods requires a large amount of well-annotated data on compounds and the protein sites to which they bind. This problem is made more difficult because many protein targets have allosteric sites. Allosteric sites are alternating positions on a protein outside of the orthosteric (active) site that can modulate protein function in some way. Compounds and proteins are typically matched using measurements of binding activity. This is suitable when most of the compounds being tested bind to a single protein site, but there are many protein targets to which various compounds can bind to one of several possible sites. In such cases, incorrect binding site information used to train predictive models leads to decreased model performance. Unfortunately, because most compound activity measurements are not directly mapped to the precise protein binding sites, there is insufficient site-level annotated data available in public databases such as the Protein Data Bank (PDB) or ChEMBL to train models directed to alternative sites.

[0204] Considering the above, an improved method is needed for training models to predict the potency of target compounds. One aspect of this disclosure addresses this drawback by assigning a binding site to a training compound known to bind to a target protein having orthosteric and one or more allosteric sites, by comparing such a compound with a reference compound whose binding site is known. In this way, a more accurate pose of such a training compound is achieved because the training compound is positioned at the binding site based on the chemical intuition provided by the reference compound. It has been found that by training a binding model using this type of training data, such a binding model trained on such multi-site data provides improved potency predictions.

[0205] Figure 22 illustrates a computer system 2200 for predicting the efficacy of a test compound against a target protein. Referring to Figure 22, in a typical embodiment, the computer system 2200 comprises one or more computers. For illustrative purposes in Figure 22, the computer system 2200 is represented as a single computer containing all of the disclosed functions of the computer system 2200. However, the disclosure is not limited thereto. The functions of the computer system 2200 may be distributed across any number of networked computers and / or reside in each of multiple networked computers and / or virtual machines. Those skilled in the art will understand that a variety of different computer topologies are possible for the computer system 2200, and that all such topologies are within the scope of the disclosure.

[0206] Referring to Figure 22 with the above in mind, the computer system 2200 comprises one or more processing units (CPUs) 2259, a network or other communication interface 2284, a user interface 2278 (including, for example, an optional display 2282 and an optional keyboard 2280 or other form of input device), memory 2292 (e.g., random access memory, persistent memory, or a combination thereof), one or more magnetic disk storage and / or persistent devices 2290 optionally accessed by one or more controllers 2288, one or more communication buses 2212 for interconnecting the aforementioned components, and a power supply 2279 for supplying power to the aforementioned components. Unless the components of memory 2292 are persistent, data in memory 2292 can be seamlessly shared with non-volatile memory 2290 or portions of memory 2292 that are non-volatile / persistent using known computing techniques such as caching. Memory 2292 and / or memory 2290 may include mass storage located remotely from the central processing unit 2259. In other words, some data stored in memory 2292 and / or memory 2290 may actually be outside the computer system 2200 but may be hosted on a computer that can be electronically accessed by the computer system 2200 via the internet, intranet, or other form of network or electronic cable using the network interface 2284. In some embodiments, the computer system 2200 utilizes Model 2248, which runs from memory associated with one or more graphical processing units, in order to improve the speed and performance of the system. In some alternative embodiments, the computer system 2200 utilizes Model 2248, which runs from memory 2292 rather than from memory associated with the graphical processing units.

[0207] The memory 2292 of computer system 2200 is • An optional operating system 2230, including procedures for handling various basic system services, • A potency prediction computation module 2232 for training model 2248 to predict the potency of the test compound against the target protein, Training data 2234, wherein the training data includes data on multiple training compounds, and each training compound 2236 in the multiple training compounds is labeled with (i) the respective protein target 2240 in the multiple protein targets, (ii) the binding site to each protein target 2242, and (iii) the potency to the protein target 2238, and training data 2234, • The atomic structure of each of the 2246 target proteins within multiple proteins, • Stores a binding site model 2248 for predicting the efficacy of the test compound against the target protein.

[0208] In some embodiments, one or more of the identified data elements or modules of the computer system 2200 are stored in one or more of the aforementioned memory devices and correspond to a set of instructions for performing the functions described above. The identified data, modules, or programs (e.g., sets of instructions) do not need to be implemented as separate software programs, procedures, or modules, and therefore various subsets of these modules can be combined or rearranged in various implementations. In some implementations, memories 2292 and / or 2290 optionally store a subset of the modules and data structures identified above. Furthermore, in some embodiments, memories 2292 and / or 2290 store additional modules and data structures not described above.

[0209] Here, a system for training Model 2248 to predict the efficacy of a test compound against a target protein is disclosed, and the method for performing such characterization is described in detail with reference to Figure 23 and discussed below.

[0210] Block 2300. Referring to Block 2300, in some embodiments, systems and methods are provided for training a model to predict the potency of a test compound against a target protein. In some embodiments, this model is part of a compound selection workflow. As a non-limiting example, in some embodiments, the systems and methods of the present disclosure are used to train model 2402 of the compound selection workflow in Figure 24.

[0211] Block 2302. Referring to Block 2302, a dataset containing data for multiple training compounds is obtained. Each training compound is labeled with its respective protein target within multiple protein targets, its binding site for each protein target, and its potency against the protein target. In some embodiments, potency is represented by the model as a calculated pKi. In some embodiments, potency is represented by the model as a calculated IC50, EC50, Kd, ​​KI, or pKI of the corresponding compound against the target protein. Such measures of potency are called binding coefficients and are generally described in Huser ed., 2006, “High-Throughput-Screening in Drug Discovery,” Methods and Principles in Medicinal Chemistry 35, and Chen ed., 2019, A Practical Guide to Assay Development and High-Throughput Screening in Drug Discovery, CRC Press, Boca Raton. Each of these is incorporated herein by reference.

[0212] One non-exclusive source of such compounds is ChEMBL, a manually curated database of bioactive molecules with drug-like properties. See Gaulton et al., 2019, “ChEMBL: towards direct deposition of bioassay data,” Nucleic Acids Res. 47(D1), D930-D940, which is incorporated herein by reference.

[0213] Block 2304. Referring to Block 2304, in some embodiments, each training compound satisfies two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) molecular weight less than 500 Daltons, and (iv) LogP less than 5. See Lipinski, 1997, Adv. Drug Del. Rev. 23, 3, which is incorporated herein by reference in its entirety. In some embodiments, the training compound satisfies one or more criteria in addition to Lipinski's Rule of Five. For example, in some embodiments, the training compound has five or fewer aromatic rings, four or fewer aromatic rings, three or fewer aromatic rings, or two or fewer aromatic rings.

[0214] Blocks 2306-2308. Referring to Block 2306, in some embodiments, each training compound is an organic compound having a molecular weight of less than 500 daltons, less than 1000 daltons, less than 2000 daltons, less than 4000 daltons, less than 6000 daltons, less than 8000 daltons, less than 10000 daltons, or less than 20000 daltons. Referring to Block 208, in some embodiments, each training compound is an organic compound having a molecular weight of 400 daltons to 10000 daltons.

[0215] In some embodiments, each training compound has a molecular weight of at least 10 Da, at least 20 Da, at least 50 Da, at least 100 Da, at least 200 Da, at least 500 Da, at least 1 kDa, at least 2 kDa, at least 3 kDa, at least 5 kDa, at least 10 kDa, at least 20 kDa, at least 30 kDa, at least 50 kDa, at least 100 kDa, or at least 500 kDa. In some embodiments, each training compound has a molecular weight of 1000 kDa or less, 500 kDa or less, 100 kDa or less, 50 kDa or less, 10 kDa or less, 5 kDa or less, 2 kDa or less, 1 kDa or less, 500 Da or less, 300 Da or less, 100 Da or less, or 50 Da or less. In some embodiments, each training compound has a molecular weight in the range of 10 Da to 900 Da, 50 Da to 1000 Da, 100 Da to 2000 Da, 1 kDa to 10 kDa, 5 kDa to 500 kDa, or 100 kDa to 1000 kDa. In some embodiments, each training compound has a molecular weight that begins below 10 daltons and ends within another range that ends at 1000 kDa or less.

[0216] In some embodiments, each training compound is a small molecule. For example, in some embodiments, each training compound is an organic compound having a molecular weight of approximately less than 1000 daltons (e.g., less than 900 daltons).

[0217] Block 2310. Referring to Block 2310, in some embodiments, the plurality of training compounds includes 100 or more, 1000 or more, 2000 or more, 10,000 or more, 100,000 or more, or 1×10 6 or more training compounds. In some embodiments, the plurality of training compounds is from 500 training compounds to 1×10 7 training compounds. In some embodiments, the plurality of training compounds is from 1000 training compounds to 1×10 6 training compounds. In some embodiments, the plurality of training compounds is from 10,000 training compounds to 100,000 training compounds.

[0218] Block 2312. Referring to Block 2312, in some embodiments, each of the protein targets within the plurality of protein targets is represented by one or more of the training compounds within the plurality of training compounds, and the plurality of protein targets includes 10 or more protein targets or 100 or more protein targets. In other words, in some embodiments, for each respective protein target within the plurality of protein targets, there is at least one training compound docked to the protein target for the purpose of generating a pose for training the model.

[0219] In some embodiments, each of the protein targets within the plurality of protein targets is represented by 5 or more of the compounds within the plurality of compounds, and the plurality of protein targets includes 10 or more protein targets or 100 or more protein targets. In other words, in some embodiments, for each respective protein target within the plurality of protein targets, there is at least a given training compound docked to each protein target for the purpose of generating 5 or more poses for training the model.

[0220] In some embodiments, each of the protein targets within the plurality of protein targets is represented by 50 or more of the compounds within the plurality of compounds, and the plurality of protein targets includes 10 or more protein targets or 100 or more protein targets.

[0221] In some embodiments, each of the protein targets within the plurality of protein targets is represented by 100 or more of the compounds within the plurality of compounds, and the plurality of protein targets includes 10 or more protein targets or 100 or more protein targets.

[0222] In some embodiments, each of the protein targets within the plurality of protein targets is represented by 500 or more of the compounds within the plurality of compounds, and the plurality of protein targets includes 10 or more protein targets or 100 or more protein targets.

[0223] In some embodiments, each protein target within a plurality of protein targets is represented by 1,000, 10,000, or 100,000 or more compounds within a plurality of compounds, and the plurality of protein targets comprises 10 or more protein targets or 100 or more protein targets.

[0224] Block 2314. Referring to Block 2314, in some embodiments, at least 5 percent, 10 percent, or 15 percent of multiple protein targets contain multiple binding sites. Identifying the binding sites for training compounds that bind to the protein is crucial to obtaining correct pose information for such training compounds docked to target proteins. For example, referring to Figure 29A, if a training compound is docked to the binding site depicted on the left side of the figure, then if the training compound actually binds to the binding site depicted on the right side of the figure, the pose data of the training compound provided to the model for the purpose of training the model will be inaccurate. One way to determine the true binding site of such training compounds is to cocluster them with compounds that have known binding sites based on a compound fingerprint, such as ECFP4, as illustrated in Figure 29B. As illustrated in Figures 29B and 30, when each compound with a known binding site in each cluster maps to the same binding site, there is confidence that the remaining compounds in the cluster also bind to the same binding site, even if no structural data indicating this exists.

[0225] As illustrated in Figure 31A, tools such as AF2 can extend the mapping assignment to protein targets for which experimental structural data, such as X-ray crystallography data, has not yet been elucidated. AF2 is described in Jumper et al., 2021, “Highly accurate protein structure prediction with AlphaFold,” Nature 596, pp. 583-589, and Tunyasuvunakool et al., 2021, “Highly accurate protein structure prediction for the human proteome,” Nature 596, 590-596. Each of these is incorporated herein by reference. Referring to Figure 31A, a protein target bound to a test or training compound is called ligand-bound or HOLO conformation, and a protein target without a bound test or training compound is called ligand-free or APO conformation. As shown in Figure 31A, in one approach to expand the number of target proteins that can be used to train models beyond those whose structures are experimentally determined, indexed in the Protein Data Bank (PDB) database accessed through means such as X-ray crystallography or nuclear magnetic resonance, and via BLASTp, the AF2 structure is generated from a UniProt protein identifier. In some embodiments, generated structures that have less than 30% sequence identity to the top BLASTp hit from a BLASTp search in the Protein Databank, or that do not have a HOLO homolog or structure, are discarded.Blastp is described in Altschul., et al., 1997, “Gapped BLAST and PSI-BLAST: a new generation of protein database search programs.” Nucleic Acids Res 25(17), pp.3389-402 and Zhang and Freddolino, 2023, “A large-scale assessment of sequence database search tools for homology-based protein function prediction,” https: / / doi.org / 10.1101 / 2023.11.14.567021. Each of these is incorporated herein by reference. UniProt is described in Universal Protein Knowledgebase in 2023, Nucleic Acids Res. 51:D523-D531 (2023). This is incorporated herein by reference. The PDB database is described in Berman et al., 2000, “The Protein Data Bank (2000)” Nucleic Acids Research See 28:235-242, which is incorporated herein by reference. Referring to Figure 31A, in some embodiments, the HOLO template of the target protein generated by AF2 is considered average when it has 30 to 70 percent sequence identity with the top BLASTp hit from a PDB database query using the UniProt sequence of the target protein. In some embodiments, the HOLO template of the target protein generated by AF2 is considered good when it has 70 to 95 percent sequence identity with the top BLASTp hit from a PDB database query using the UniProt sequence of the target protein.In some embodiments, the HOLO template of the target protein generated by AF2 is considered resolved when it has more than 95 percent sequence identity with the top BLASTp hit from a PDB database query using the UniProt sequence of the target protein.

[0226] Block 2315. Referring to Block 2315, the model is trained using the training dataset. For each of the multiple training compounds, one or more poses of each training compound at the target protein binding site identified in the training compound dataset are input into the model and evaluated against the actual known potency of each training compound.

[0227] In some embodiments, this is done using the command parameter line, -exhaustiveness 384 --energy_range 99999 --num_modes 64 --mc_steps 3 --minimize_iters 40 --accurate_line --approximation linear --autobox_add 2.0 --seed 42, these parameters are described in Stafford et al., 2022, “AtomNet PoseRanker: Enriching Ligand Pose Quality for Dynamic Proteins in Virtual High-Throughput Screens,” J.Chem.Inf.Model 62, pp. 1178-1189, which is incorporated herein by reference. In such an approach, a bounding box is drawn around a designated binding site within the target protein, and the training compound is docked to the portion of the target protein contained within this bounding box. Figure 8A illustrates the bounding boxes for two different binding sites of the target protein. In some embodiments, the conformation of the training compound is obtained using a program such as RDKit. See Landrum, G. RDKit: Open-Source Cheminformatics. In some such embodiments, the description of the SMILES of the training compound is the minimized UFF energy using RDKit, and the conformation of the training compound docked to the binding site is obtained.Furthermore, for information on how the compound pose at the protein binding site is constructed in some embodiments, please refer to Meller et al., 2023, “Discovery of a cryptic pocket in the AI-predicted structure of PPM1D phosphatase explains the binding site and potency of its allosteric inhibitors,” bioRxiv [Preprint], 2023 Mar 24:2023.03.22.533829. doi:10.1101 / 2023.03.22.533829.

[0228] In some embodiments, multiple descriptors are generated for each pose of each training compound within the binding site for input to the model. In some embodiments, each heavy atom (non-hydrogen atom) is described by its corresponding SYBYL atom type indicating its chemical environment, which is generated, for example, by converting it to Mol2 format using OpenBabel. See O'Boyle et al., 2011, “Open Babel: An Open Chemical Toolbox.” J. Cheminf. 3, 33, which is incorporated herein by reference. In some embodiments, protons are not explicitly represented in the descriptor set, and all metals are of a single type. In some embodiments, the target protein binding site and the training compound are maintained as separate entities, and their atom types are coded separately.

[0229] The error in the predicted potency of each training compound is backpropagated through the parameters of the system model of this disclosure (e.g., Model 2248 in Figure 22, Model 2402 in Figure 24). In some embodiments, the model is trained against the estimated potency error made by the model, taking into account the actual potency of the training compounds. One way to carry out such training is to use stochastic gradient descent with the AdaDelta adaptive learning method (see Zeiler, 2012 “ADADELTA: an adaptive learning rate method,” 'CoRR, vol. abs / 1212.5701, which is incorporated herein by reference) and the backpropagation algorithm shown in Rumelhart et al., 1988, “Neurocomputing: Foundations of research,” ch. Learning Representations by Back-propagating Errors, pp. 696-699, Cambridge, MA, USA: MIT Press). However, any backpropagation method can be used for training.

[0230] Block 2316. Referring to Block 2316, the test chemical compounds can be obtained.

[0231] Block 2318. Referring to Block 218, in some embodiments, the test compound satisfies two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) molecular weight less than 500 Daltons, and (iv) LogP less than 5. See Lipinski, 1997, Adv. Drug Del. Rev. 23, 3, which is incorporated herein by reference in its entirety. In some embodiments, the test compound satisfies one or more criteria in addition to Lipinski's Rule of Five. For example, in some embodiments, the test compound has five or fewer aromatic rings, four or fewer aromatic rings, three or fewer aromatic rings, or two or fewer aromatic rings.

[0232] Blocks 2320-2322. Referring to Block 2320, in some embodiments, the test compound is an organic compound having a molecular weight of less than 500 daltons, less than 1000 daltons, less than 2000 daltons, less than 4000 daltons, less than 6000 daltons, less than 8000 daltons, less than 10000 daltons, or less than 20000 daltons. Referring to Block 2322, in some embodiments, the test compound is an organic compound having a molecular weight of 400 daltons to 10000 daltons.

[0233] Block 2324. Referring to Block 2324, a test pose is obtained by posing a test chemical compound to the binding site of a target protein. The pose is described above in conjunction with the training compound. Examples of additional poses are described in U.S. Patents 11,080,570, 10,546,237, 10,482,355, 10,002,312, 9,379,059, and U.S. Provisional Patent Application 63 / 336,841, filed on April 29, 2022, entitled "Characterization of Interactions Between Compounds and Polymers Using Pose Ensembles." Each of these is incorporated herein by reference.

[0234] Block 2325. Referring to Block 2325, the test pose is input into the model to predict the potency of the test compound to the binding site. As illustrated in Figures 32, 33A, and 33B, advantageously, the model trained with the multi-site data of this disclosure shows improved performance in predicting the potency of the compound to the target protein. Specifically, Figure 32 shows an improved area under the receiver operating characteristic curve (ROC) of the model trained with the binding site-specific data of this disclosure compared to the static assumption that each training compound binds to an orthosteric site of the target protein. This improvement is also observed for the primary binding sites (orthosteric site, active site, and binding site 1 in Figure 32). Figure 33A shows the improvement in the mean squared error of the model when trained with multiple protein targets enhanced with the AF2-producing protein target described herein (holo on the Y-axis in Figure 33A), referring to Figure 31A (af2_holo on the Y-axis in Figure 33A). Figure 33B shows the improvement in TauB when trained with multiple protein targets enhanced with the AF2-producing protein target described herein, compared to when trained with multiple protein targets not enhanced with the AF2-producing protein target (holo on the Y-axis in Figure 33B), with reference to Figure 31A (af2_holo on the Y-axis in Figure 33B).

[0235] Block 2326. Referring to Block 2326, in some embodiments the model includes a first set of parameters. In some embodiments the model applies the parameters to the pause through at least 10,000 calculations to produce a predicted potency. In some embodiments the model performs at least 100,000 calculations or at least 1 × 10⁶ 6 Through a series of calculations, multiple parameters are applied to the pause to generate the predicted potency.

[0236] Block 2328. Referring to Block 2328, in some embodiments the model is a graph neural network. In some embodiments the model is an SE(3) equivariant graph attention transformer. In some embodiments the model is a message-passing neural network. Message-passing neural networks are described, for example, in Gilmer et al., 2020, “Message Passing Neural Networks” and Schutt et al. (eds), Machine Learning Meets Quantum Physics, Lecture Notes in Physics 968, Springer, Cham. Each of these is incorporated herein by reference.

[0237] In some embodiments, the model is a graph convolutional neural network. Examples of non-restrictive additions to graph convolutional neural networks include Behler Parrinello, 2007, “Generalized Neural-Network Representation of High Dimensional Potential-Energy Surfaces,” Physical Review Letters 98, 146401; Chmiela et al., 2017, “Machine learning of accurate energy-conserving molecular force fields,” Science Advances 3(5):e1603015; Schutt et al., 2017, “SchNet: A continuous-filter convolutional neural network for modeling quantum interactions,” Advances in Neural Information Processing Systems 30, pp.992-1002; Feinberg et al., 2018, “PotentialNet for Molecular Property Prediction,” ACS Cent.Sci.4, 11, 1520-1530; and Stafford et al., 2022, “AtomNet PoseRanker: Enriching Ligand This is disclosed in "Pose Quality for Dynamic Proteins in Virtual High Throughput Screens," J. Chem Inf Model 62, pp. 1178-1189. Each of these is incorporated herein by reference.

[0238] In some embodiments, the final readout layer of the model independently extracts relevant features from the training compound and protein target embeddings and passes them through the final multilayer perceptron.

[0239] In some embodiments, the model is the GRAPHite architecture as described in Stafford et al., 2022, “AtomNet PoseRanker: Enriching Ligand Pose Quality for Dynamic Proteins in Virtual High Throughput Screens,” J. Chem Inf Model 62, pp. 1178 - 1189, or is similar thereto. The GRAPHite architecture is a graph convolutional network (GCN) composed of four graph convolutional layers. In some embodiments, the first two layers include both ligand and receptor features, while the last two layers are ligand-only. In some embodiments, the nodes in the graph represent ligand and receptor atoms. In some embodiments, only receptor atoms within 7 angstroms of any ligand atom are used as part of the graph. In some embodiments, the edges are defined by atoms within 4 angstroms of each other, and the edge weights are distance-dependent. In some embodiments, the final layer is globally pooled into an embedding. In some embodiments, this embedding then passes through two (independent) multi-layer perceptrons to predict two outputs, namely, the ANPR pose quality score and the Vina docking score. In some embodiments, these outputs are then concatenated into the embedding and passed through a third multi-layer perceptron that outputs the predicted potency in the form of pKi.

[0240] Block 2330. Referring to block 2330, in some embodiments, the model includes 500,000 or more parameters, or 1×10 6 or more parameters. In some embodiments, the model includes more than 400 parameters, more than 1000 parameters, more than 2000 parameters, more than 5000 parameters, more than 10,000 parameters, more than 100,000 parameters, or 1×10 6It includes more than 10 parameters. In some embodiments, the number of parameters in the model and the amount of information in the pause input to the combined model result in more than 10,000 calculations, more than 100,000 calculations, or 1 × 10⁻¹⁶ calculations to calculate the predicted potency. 6 Calculations exceeding 5 × 10 6 Calculations exceeding the number of times, or 1 × 10 7 It delivers performance exceeding 10 [Examples]

[0241] Example 1: Constructing training and test datasets. Referring to Figure 3B, training and test datasets were created to identify the methods that provided the GED approximation. One desirable characteristic for the training / test split was to create training / test splits of drug-like molecules / fragments with similar GED distributions. Another desirable characteristic for the training / test split was to require that the training / test pairs be sufficiently different to ensure that performance generalizes into an invisible chemical space (rather than overfitting the model). To create training and test datasets with these characteristics, compounds were randomly selected from ChEMBL30. ChEMBL is a manually curated database of bioactive molecules with drug-like properties. See Gaulton et al., 2019, “ChEMBL: towards direct deposition of bioassay data,” Nucleic Acids Res. 47(D1), D930-D940, which is incorporated herein by reference. The randomly selected compounds were then filtered for drug similarity. One method for calculating drug similarity scores is a quantitative estimate of drug similarity (QED). Bickerton et al., 2012, “Quantifying the chemical beauty of drugs,” Nat. Chem. 4, pp. 90-98, which is incorporated herein by reference. Compounds were screened to have a QED drug similarity score of 0.5 or higher. Furthermore, these compounds were constrained to have a graph edit distance of less than 10. In addition, these compounds were stratified on GED so that all GED values ​​from 0 to 10 were sampled uniformly. Finally, the compounds were split into training / test sets using Taylor-Butina clustering with an extended connectivity fingerprint of up to four bindings (ECFP4) for each compound.As shown in Figure 3B, all pairs within a given cluster are placed in either the training set or the test set, ensuring that there are no pairs of compounds in the test set that are sufficiently similar to the pairs in the training set, so that the model can simply remember the pairs in the training set and extend this memory to the pairs in the test set, rather than learning a true approximation of the GED distance, even for chemical pairs not seen by the model.

[0242] The above describes the construction of drug-like training / test datasets. Referring to Figure 4, the training drug-like dataset had 45,000 compounds, while the test dataset had 4,500 compounds. GED was calculated for the compounds using commercial software based on chemical GED. Figure 21 illustrates the difference between chemical edit distance and accurate graph edit distance. As can be seen in Figure 21, chemical graph edit distance considers chemical changes, while accurate GED considers the number of basic changes mentioned above. Fragment datasets were created by randomly sampling drug-like molecules and cutting them into fragments. The fragments were then filtered using the same criteria used to filter the chemical dataset (stratified by GED, with QED greater than 0.5 and GED less than or equal to 10). However, accurate GED was calculated for the fragments using Depth-first GED, not chemical GED. Depth-first GED is described in Abu-Aisheh et al., 2015, 4. thThis is described in the International Conference on Pattern Recognition Applications and Methods, Lisbon, Portugal, which is incorporated herein by reference. Figure 4 further illustrates the distribution of compounds as a function of the number of heavy atoms in the training and test datasets. Each pair of compounds in the training and test datasets in the fragment dataset summarized in Figure 4 has a training label for the exact GED between each compound pair. Each pair of compounds in the training and test datasets in the drug-like dataset summarized in Figure 4 has a training label for the chemical GED between each compound pair.

[0243] Example 2: Testing SimGNN on test and training datasets. Referring to Figure 5, the performance of the model following prior art (SimGNN, Bai et al. 2019, ACM International Conference on Web Search and Data Mining) after training on the fragment dataset and drug-like dataset shown in Figure 4 is as follows: MSE, Spearman's ρ, and Kendall's ρ, respectively.

number

number

[0244] Example 3: Constructing an improved approximate GED calculation module 32. To address the shortcomings identified in Figure 6, a system summarized in Figure 7 was constructed. Figure 7 shows the first graph G representing the first compound. i The diagram illustrates how 36 is obtained. The first graph includes a first set of nodes and a first set of edges. Each atom of the first compound is represented by a node 38 within the first set of nodes, and each bond of the first compound is represented by an edge 40 within the first set of edges. Furthermore, the second graph G represents the second compound 44. j46 is obtained. The second graph includes a second set of nodes and a second set of edges. Each atom of the second compound is represented by node 48 in the second set of nodes, and each bond of the second compound is represented by edge 50 in the second set of edges. The first graph is input to model 702, which includes multiple parameters. Model 702 applies the multiple parameters to the first graph to obtain a first latent embedding h representing the first compound as a first output from model 702. i This generates the second graph, which is input into Model 704 (actually Model 702, but shown separately for illustrative purposes only). Model 704 applies several parameters to the second graph to generate a second latent embedding h representing the second compound as a second output from Model 704. j This generates the graph edit distance between the first and second compounds, which is then determined as the L2 norm difference between the first and second latent embeddings. If the first latent embedding is the first vector and the second latent embedding is the second vector, the L2 norm difference (between the first and second vectors) is taken as the square root of the sum of the squared differences of the corresponding vector element values ​​between the corresponding elements in the first vector and the second vector. Figure 18 illustrates the mathematical specification of Figure 7.

[0245] Models 702 / 704 in this example are message-passing neural networks. Message-passing neural networks are described, for example, in Gilmer et al., 2020, “Message Passing Neural Networks” and Schutt et al. (eds), Machine Learning Meets Quantum Physics, Lecture Notes in Physics 968, Springer, Cham. Each of these is incorporated herein by reference.

[0246] Example 4: Training the system according to the disclosure. All properties of the distance metric outlined at the top of Figure 6 are satisfied, except for positive valuation, by a trained model 702 / 704 such that the L2 norm difference between the model's first and second outputs is equal to the trained GED estimate. In some embodiments, the axiom of positive valuation is applied when the model 702 / 704 cannot distinguish between two graphs.

[0247] The system 700 illustrated in Figure 7 outputs an estimated graph edit distance (e.g., the L2 norm difference of the outputs of models 702 / 704). The system of this disclosure then compares the estimated GED assigned to each training compound pair in the fragment dataset or drug-like dataset, summarized in Figure 4, with the actual GED for the compound pair. For the fragmented dataset, the actual GED was the exact GED, as summarized in Figure 4. For the drug-like dataset, the actual GED was considered as the chemical GED, as summarized in Figure 4. The difference between the estimated GED generated by model 700 and the actual GED for the training set was considered the error generated by the system of this disclosure (e.g., model 700). To further train model 700, these errors were backpropagated through the parameters of models 702 / 704. In some embodiments, the system 700 was trained against the error of the estimated GED generated by the system 700, taking into account the actual GED for the training set. One way to carry out such training is to use stochastic gradient descent with the AdaDelta adaptive learning method (see Zeiler, 2012 “ADADELTA: an adaptive learning rate method,” CoRR, vol. abs / 1212.5701, which is incorporated herein by reference) and the backpropagation algorithm shown in Rumelhart et al., 1988, “Neurocomputing: Foundations of research,” ch. Learning Representations by Back-propagating Errors, pp. 696-699, Cambridge, MA, USA: MIT Press). However, any backpropagation method can be used for training.

[0248] Figure 8 shows the performance of System 700 in Figure 7, referred to as MolGEDNet. From Figure 8, it can be seen that MolGEDNet is significantly better than SimGNN in mean squared error for both the fragment dataset and the drug-like dataset in Figure 4. Figure 9 shows the performance of System 700 in Figure 7, referred to as MolGEDNet. From Figure 9, it can be seen that MolGEDNet is significantly better than SimGNN in Spearman's ρ and Kendall's ρ for both the fragment dataset and the drug-like dataset in Figure 4.

number

[0249] Example 5: Hyperparameter tuning of the system according to the present disclosure. Using the superior metrics of MolGEDNet for SimGNN established and discussed in Example 4, as shown in Figures 8 and 9, the hyperparameters of MolGEDNet were tuned as shown in Figure 10. For Figure 10, as shown in Figures 8 and 9, Model 702 / 704 was a message-passing graph neural network. As shown in Figure 10, the dimensional determination of Model 702 / 704 with respect to the potential embedding size was found to be optimal when a size of 128 was used, in terms of edge embedding size and node embedding size. To obtain the results shown in Figure 10, System 700 was trained on the drug-like training dataset of Figure 4, and then the mean squared error, Spearman's ρ, and Kendall's ρ were measured across the drug-like test dataset of Figure 4.

number

number

[0250] Figure 11 illustrates the model performance of MolGEDNet when different dropout percentages are considered. Here, dropout refers to the percentage of output from each layer (the connection between layer i and layer i+1) that was dropped during model training. Thus, with a dropout of 0.1, 10 percent of the output from each layer was randomly selected and dropped during training, rather than passing these outputs to subsequent layers in models 702 and 704. As illustrated in Figure 11, the addition of dropout did not improve model performance (measured here by mean squared error). To obtain the results illustrated in Figure 11, system 700 was trained on the drug-like training dataset in Figure 4 with the different dropout values ​​illustrated in Figure 11, and then the mean squared error was calculated across the drug-like test dataset in Figure 4 without dropout (regardless of the amount of dropout used in training).

[0251] Figure 12 illustrates the model performance of MolGEDNet when Model 702 / 704 has a different number of layers. For Figure 12, Model 702 / 704 is a message-passing graph neural network, and what was changed was the number of layers in Model 702 / 704 (4, 5, 6, or 7 layers). As illustrated in Figure 12, Model 702 / 704 with 5 layers was found to be optimal for the test dataset in Figure 4. To obtain the results illustrated in Figure 12, System 700 was trained on the drug-like training dataset in Figure 4, and then the mean squared error, Spearman's RHO, and Kendall's RHO were measured across the drug-like test dataset in Figure 4.

number

number

[0252] Referring to Figure 13, System 700 was trained on the drug-like training dataset in Figure 4 using the optimal hyperparameters: 128 variable dimensions, 5 layers, and no dropout for the messaging-passing graph neural network 702 / 704. Figure 19 illustrates how Model 702 / 704 performs node, edge, and graph updates using these optimal hyperparameters. Figure 20 illustrates the details of the message-passing neural network used to generate the performance data in Figure 13. The trained system was then used to obtain the mean squared error, Spearman's ρ, and Kendall's ρ across the drug-like test dataset in Figure 4.

number

number

[0253] Example 6: Investigating the cause of error in the hyperparameter optimization system of the present disclosure. Figure 14 examines the cause of error in the hyperparameter optimization system 700 of Figure 13 / Example 5. Figure 14 shows that the disclosed models 702 / 704 tend to overestimate the GED when the true graph edit distance is less than 5 or 6, and tend to underestimate the GED when the true graph edit distance is greater than 6.

[0254] To determine where the errors in the compounds originate, we examined compound pairs in the test set that had a true GED of exactly 1. Examples of such pairs are shown in Figure 15, along with the estimated GEDs given by the disclosed model. The examples in the left boxes of Figure 15 are pairs with large errors, while the examples in the right boxes of Figure 15 are pairs with small errors. This shows that large errors occur when the edit is at the center of the compound, but high accuracy is achieved when the edit is around the compound. In Figure 16, the mid-centrality of the nodes requiring modification in each pair in Figure 15 is given. Mid-centrality is a measure of graph centrality based on the shortest paths. For each pair of vertices in a connected graph, there is at least one shortest path between the vertices such that either the number of edges the path passes through (for an unweighted graph) or the sum of the edge weights (for a weighted graph) is minimized. The mid-centrality of each vertex is the number of these shortest paths passing through the vertex. See Freeman, 1977, “A set of measures of centrality based on betweenness,” Sociometry. 40(1):35-41, which is incorporated herein by reference. Nodes with high mesocentrality are near the center of the graph, while nodes on the periphery of the graph have low mesocentrality. What we can see is that if the mesocentrality of the node that must be substituted in a given compound pair is high, then for the disclosed model 702 / 704, a higher error occurs in the GED estimation.

[0255] Example 7: Execution time of the hyperparameter optimization system of the present disclosure. Referring to Figure 17, the average runtime as a function of the average molecular weight of the two compounds being compared is given for the disclosed Model 702 / 704 (MolGEDNet, having the optimal hyperparameters described in Example 5), the exact method, and the Hungarian method. MolGEDNet is found to be approximately 50 times faster than the Hungarian approximation algorithm, while also exhibiting significantly superior performance at a much more accurate speed.

[0256] conclusion For explanatory purposes, the foregoing explanation is based on reference to specific implementations. However, the above illustrative considerations are not intended to be exhaustive, nor are they intended to limit the implementation to the exact form disclosed. Many modifications and variations are possible in light of the above teachings. The implementations have been selected and described to best illustrate the principles and their practical applications, thereby enabling those skilled in the art to best utilize implementations suitable for the specific intended use, as well as various implementations with various modifications.

Claims

1. A computer system for estimating the graph edit distance between a first compound and a second compound, One or more processors, The system comprises a memory addressable by one or more processors, wherein the memory stores at least one program for execution by the one or more processors, and the at least one program is (A) To obtain a first graph representing the first compound, comprising a first plurality of nodes and a first plurality of edges, wherein each atom of the first compound is represented by a node within the first plurality of nodes, and each bond of the first compound is represented by an edge within the first plurality of edges, (B) To obtain a second graph representing the second compound, comprising a second plurality of nodes and a second plurality of edges, wherein each atom of the second compound is represented by a node within the second plurality of nodes, and each bond of the second compound is represented by an edge within the second plurality of edges, (C) Inputting the first graph into a model including a plurality of parameters, wherein the model applies the plurality of parameters to the first graph through at least 10,000 calculations to generate a first isometrically retained latent embedding representing the first compound as a first output from the model, (D) Inputting the second graph into the model, wherein the model applies the plurality of parameters to the second graph through at least 10,000 calculations to generate a second isometrically retained latent embedding representing the second compound as a second output from the model, (E) A computer system including instructions for providing an estimate of the graph edit distance between the first compound and the second compound using the difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

2. The computer system according to claim 1, wherein the difference is the L2 norm difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

3. The first isometrically preserved latent embedding is a first vector containing a first plurality of vector values, The second isometrically preserved latent embedding is a second vector containing a second set of vector values, The computer system according to claim 2, wherein the L2 norm difference between the first vector and the second vector is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first vector and the second vector, and the model is trained such that the L2 norm of the difference between the first vector and the second vector is an estimate of the GED between the first numerator and the second numerator.

4. The computer system according to any one of claims 1 to 3, wherein the at least one program further includes an instruction for determining that the second compound is a chemically related analog of the first compound when the estimated value of the graph edit distance falls below a threshold.

5. The aforementioned at least one program, For each unique compound pair in the compound database, the process of obtaining (A), obtaining (B), inputting (C), inputting (D), and providing (E) is repeated, wherein the first compound in each unique compound pair is the first compound, the second compound in each unique compound pair is the second compound, and thereby a plurality of graph edit distances are formed, and this process is repeated. By evaluating the multiple distances using a reference function, the multiple graph editing distances are clustered into multiple clusters, A computer system according to any one of claims 1 to 4, further comprising instructions for selecting a compound from one or more clusters within the plurality of clusters for evaluation in a drug discovery program.

6. The computer system according to any one of claims 1 to 5, wherein the at least one program further includes instructions for constructing a quantitative structure-activity relationship model comprising the first compound and the second compound using the estimated graph edit distance.

7. The computer system according to any one of claims 1 to 6, wherein the at least one program further includes instructions for determining an activity cliff for a drug target using the estimated value of the graph edit distance.

8. The computer system according to any one of claims 1 to 7, wherein the model is a graph convolutional neural network.

9. The computer system according to any one of claims 1 to 7, wherein the model is a graph neural network (message passing network).

10. The computer system according to any one of claims 1 to 7, wherein the model is a message-passing graph neural network consisting of three to eight layers.

11. The computer system according to any one of claims 1 to 7, wherein the model is a message-passing graph neural network consisting of five layers.

12. The computer system according to any one of claims 1 to 7, wherein the model is an autoencoder.

13. The computer system according to claim 12, wherein the autoencoder is a variational autoencoder, an adversarial autoencoder, a normalization autoencoder, a sparse autoencoder, or a metric learning autoencoder.

14. The computer system according to any one of claims 1 to 13, wherein the model includes 500,000 or more parameters.

15. The aforementioned model is 1 × 10 ６ A computer system according to any one of claims 1 to 13, comprising one or more parameters.

16. The computer system according to any one of claims 1 to 15, wherein the first compound and the second compound each satisfy two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 Daltons, and (iv) a LogP of less than 5.

17. The computer system according to any one of claims 1 to 16, wherein the first compound and the second compound have accurate graph editing distances of 1 to 10.

18. The computer system according to any one of claims 1 to 16, wherein the first compound and the second compound have accurate graph editing distances of 1 to 15.

19. The computer system according to any one of claims 1 to 18, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 500 daltons.

20. The computer system according to any one of claims 1 to 18, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 1,000 daltons.

21. The computer system according to any one of claims 1 to 18, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 2,000 daltons.

22. The computer system according to any one of claims 1 to 18, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 4,000 Daltons, less than 6,000 Daltons, less than 8,000 Daltons, less than 10,000 Daltons, or less than 20,000 Daltons.

23. The computer system according to any one of claims 1 to 18, wherein the first compound and the second compound are each organic compounds having a molecular weight of 300 daltons to 1500 daltons.

24. The aforementioned model performs at least 100,000 calculations to calculate the first isometrically preserved latent implantation. The computer system according to any one of claims 1 to 23, wherein the model performs at least 100,000 calculations to calculate the second isometrically preserved latent implant.

25. A method for estimating the graph edit distance between a first compound and a second compound, In a computer system comprising one or more processors and memory addressable by the one or more processors, (A) To obtain a first graph representing the first compound, comprising a first plurality of nodes and a first plurality of edges, wherein each atom of the first compound is represented by a node within the first plurality of nodes, and each bond of the first compound is represented by an edge within the first plurality of edges, (B) To obtain a second graph representing the second compound, comprising a second plurality of nodes and a second plurality of edges, wherein each atom of the second compound is represented by a node within the second plurality of nodes, and each bond of the second compound is represented by an edge within the second plurality of edges, (C) Inputting the first graph into a model including a plurality of parameters, wherein the model applies the plurality of parameters to the first graph through at least 10,000 calculations to generate a first isometrically retained latent embedding representing the first compound as a first output from the model, (D) Inputting the second graph into the model, wherein the model applies the plurality of parameters to the second graph through at least 10,000 calculations to generate a second isometrically retained latent embedding representing the second compound as a second output from the model, (E) A method comprising providing an estimate of the graph edit distance between the first compound and the second compound using the difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

26. The method according to claim 25, wherein the difference is the L2 norm difference between the first latent embedding and the second latent embedding.

27. The first isometrically preserved latent embedding is a first vector containing a first plurality of vector values, The second isometrically preserved latent embedding is a second vector containing a second set of vector values, The method according to claim 26, wherein the L2 norm difference between the first vector and the second vector is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first vector and the second vector, and the model is trained such that the L2 norm of the difference between the first vector and the second vector is an estimate of the GED between the first numerator and the second numerator.

28. The method according to any one of claims 25 to 27, wherein the at least one program further includes an instruction for determining that the second compound is a chemically related analog of the first compound when the estimated value of the graph edit distance falls below a threshold.

29. The method described above is For each unique compound pair in the compound database, the process of obtaining (A), obtaining (B), inputting (C), inputting (D), and providing (E) is repeated, wherein the first compound in each unique compound pair is the first compound, the second compound in each unique compound pair is the second compound, and thereby a plurality of graph edit distances are formed, and this process is repeated. By evaluating the multiple distances using a reference function, the multiple graph editing distances are clustered into multiple clusters, The method according to any one of claims 25 to 28, further comprising selecting a compound from one or more clusters within the plurality of clusters for evaluation in a drug discovery program.

30. The method according to any one of claims 25 to 29, further comprising constructing a quantitative structure-activity relationship model comprising the first compound and the second compound using the estimated value of the graph edit distance.

31. The method according to any one of claims 25 to 30, further comprising determining an activity cliff for a drug target using the estimated value of the graph edit distance.

32. The method according to any one of claims 25 to 31, wherein the model is a graph convolutional neural network.

33. The method according to any one of claims 25 to 31, wherein the model is a graph neural network (message passing network).

34. The method according to any one of claims 25 to 31, wherein the model is a message-passing graph neural network consisting of three to eight layers.

35. The method according to any one of claims 25 to 31, wherein the model is a message-passing graph neural network consisting of five layers.

36. The method according to any one of claims 25 to 31, wherein the model is an autoencoder.

37. The method according to claim 36, wherein the autoencoder is a variational autoencoder, an adversarial autoencoder, a normalized autoencoder, a sparse autoencoder, or a metric learning autoencoder.

38. The method according to any one of claims 25 to 38, wherein the model includes 500,000 or more parameters.

39. The aforementioned model is 1 × 10 ６ The method according to any one of claims 25 to 38, comprising one or more parameters.

40. The method according to any one of claims 25 to 39, wherein the first compound and the second compound each satisfy two or more rules, three or more rules, or all four rules of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 Daltons, and (iv) a LogP of less than 5.

41. The method according to any one of claims 25 to 40, wherein the first compound and the second compound have an accurate graph edit distance of 1 to 10.

42. The method according to any one of claims 25 to 40, wherein the first compound and the second compound have accurate graph edit distances of 1 to 15.

43. The method according to any one of claims 25 to 42, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 500 daltons.

44. The method according to any one of claims 25 to 42, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 1,000 daltons.

45. The method according to any one of claims 25 to 42, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 2,000 daltons.

46. The method according to any one of claims 25 to 42, wherein the first compound and the second compound are organic compounds having a molecular weight of less than 4,000 Daltons, less than 6,000 Daltons, less than 8,000 Daltons, less than 10,000 Daltons, or less than 20,000 Daltons, respectively.

47. The method according to any one of claims 25 to 42, wherein the first compound and the second compound are each organic compounds having a molecular weight of 300 daltons to 1500 daltons.

48. The aforementioned model performs at least 100,000 calculations to calculate the first isometrically preserved latent implantation. The method according to any one of claims 25 to 47, wherein the model performs at least 100,000 calculations to calculate the second isometrically retained latent implant.

49. A non-temporary computer-readable storage medium, which, when executed by a computer system, stores instructions causing the computer system to perform a method for estimating the graph edit distance between a first compound and a second compound, and the method is (A) To obtain a first graph representing the first compound, comprising a first plurality of nodes and a first plurality of edges, wherein each atom of the first compound is represented by a node within the first plurality of nodes, and each bond of the first compound is represented by an edge within the first plurality of edges, (B) To obtain a second graph representing the second compound, comprising a second plurality of nodes and a second plurality of edges, wherein each atom of the second compound is represented by a node within the second plurality of nodes, and each bond of the second compound is represented by an edge within the second plurality of edges, (C) Inputting the first graph into a model including a plurality of parameters, wherein the model applies the plurality of parameters to the first graph through at least 10,000 calculations to generate a first isometrically retained latent embedding representing the first compound as a first output from the model, (D) Inputting the second graph into the model, wherein the model applies the plurality of parameters to the second graph through at least 10,000 calculations to generate a second isometrically retained latent embedding representing the second compound as a second output from the model, (E) A non-temporary computer-readable storage medium comprising providing an estimate of the graph edit distance between the first compound and the second compound using the difference between the first isometrically retained latent embedding and the second isometrically retained latent embedding.

50. The non-temporary computer-readable storage medium according to claim 45, wherein the difference is the L2 norm difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

51. The first isometrically preserved latent embedding is a first vector containing a first plurality of vector values, The second isometrically preserved latent embedding is a second vector containing a second set of vector values, The non-temporary computer-readable storage medium according to claim 50, wherein the L2 norm difference between the first vector and the second vector is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first vector and the second vector, and the model is trained such that the L2 norm of the difference between the first vector and the second vector is an estimate of the GED between the first numerator and the second numerator.

52. A non-temporary computer-readable storage medium according to any one of claims 49 to 51, wherein the at least one program further includes instructions for determining that the second compound is a chemically related analog of the first compound when the estimated value of the graph edit distance falls below a threshold.

53. The method described above is For each unique compound pair in the compound database, the process of obtaining (A), obtaining (B), inputting (C), inputting (D), and providing (E) is repeated, wherein the first compound in each unique compound pair is the first compound, the second compound in each unique compound pair is the second compound, and thereby a plurality of graph edit distances are formed, and this process is repeated. By evaluating the multiple distances using a reference function, the multiple graph editing distances are clustered into multiple clusters, A non-temporary computer-readable storage medium according to any one of claims 49 to 52, further comprising selecting a compound from one or more clusters within the plurality of clusters for evaluation in a drug discovery program.

54. The non-temporary computer-readable storage medium according to any one of claims 49 to 53, further comprising constructing a quantitative structure-activity relationship model comprising the first compound and the second compound using the estimated value of the graph edit distance.

55. A non-temporary computer-readable storage medium according to any one of claims 49 to 54, further comprising determining an activity cliff for a drug target using the estimated value of the graph edit distance.

56. A non-temporary computer-readable storage medium according to any one of claims 49 to 55, wherein the model is a graph convolutional neural network.

57. A non-temporary computer-readable storage medium according to any one of claims 49 to 55, wherein the model is a graph neural network (message passing network).

58. A non-temporary computer-readable storage medium according to any one of claims 49 to 55, wherein the model is a message-passing graph neural network consisting of three to eight layers.

59. A non-temporary computer-readable storage medium according to any one of claims 49 to 55, wherein the model is a message-passing graph neural network consisting of five layers.

60. The non-temporary computer-readable storage medium according to any one of claims 59 to 55, wherein the model is an autoencoder.

61. The non-temporary computer-readable storage medium according to claim 60, wherein the autoencoder is a variational autoencoder, adversarial autoencoder, normalization autoencoder, sparse autoencoder, or metric learning autoencoder.

62. The non-temporary computer-readable storage medium according to any one of claims 49 to 61, wherein the model includes 500,000 or more parameters.

63. The aforementioned model is 1 × 10 ６ A non-temporary computer-readable storage medium according to any one of claims 49 to 61, comprising one or more parameters.

64. A non-temporary computer-readable storage medium according to any one of claims 49 to 63, wherein the first compound and the second compound each satisfy two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 Daltons, and (iv) a LogP of less than 5.

65. A non-temporary computer-readable storage medium according to any one of claims 49 to 64, wherein the first compound and the second compound have an accurate graph editing distance of 1 to 10.

66. A non-temporary computer-readable storage medium according to any one of claims 49 to 64, wherein the first compound and the second compound have accurate graph editing distances of 1 to 15.

67. A non-temporary computer-readable storage medium according to any one of claims 49 to 66, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 500 daltons.

68. A non-temporary computer-readable storage medium according to any one of claims 49 to 66, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 1,000 daltons.

69. A non-temporary computer-readable storage medium according to any one of claims 49 to 66, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 2,000 daltons.

70. A non-temporary computer-readable storage medium according to any one of claims 49 to 66, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 4,000 Daltons, less than 6,000 Daltons, less than 8,000 Daltons, less than 10,000 Daltons, or less than 20,000 Daltons.

71. A non-temporary computer-readable storage medium according to any one of claims 49 to 66, wherein the first compound and the second compound are each organic compounds having a molecular weight of 300 daltons to 1500 daltons.

72. The aforementioned model performs at least 100,000 calculations to calculate the first isometrically preserved latent implantation. The computer system according to any one of claims 49 to 71, wherein the model performs at least 100,000 calculations to calculate the second isometrically preserved latent implant.

73. A computer system for estimating the graph edit distance between a first compound and a second compound, One or more processors, The system comprises a memory addressable by one or more processors, wherein the memory stores at least one program for execution by the one or more processors, and the at least one program is (A) Inputting a first representation of the first compound into a model including a plurality of parameters, wherein the model applies the plurality of parameters to the first representation of the first compound to generate a first isometrically preserved latent embedding representing the first compound as a first output from the model, (B) Inputting a second representation of the second compound into the model, wherein the model applies the plurality of parameters to the second representation of the second compound to generate a second isometrically retained latent embedding representing the second compound as a second output from the model, (C) A computer system including instructions for providing an estimate of the graph edit distance between the first compound and the second compound using the difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

74. The computer system according to claim 73, wherein the difference is the L2 norm difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

75. The first isometrically preserved latent embedding is a first vector containing a first plurality of vector values, The second latent isometric preservation embedding is a second vector containing a second set of vector values, The computer system according to claim 73, wherein the L2 norm difference between the first vector and the second vector is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first vector and the second vector, and the model is trained such that the L2 norm of the difference between the first vector and the second vector is an estimate of the GED between the first numerator and the second numerator.

76. The computer system according to any one of claims 73 to 75, wherein the at least one program further includes an instruction for determining that the second compound is a chemically related analog of the first compound when the estimated value of the graph edit distance falls below a threshold.

77. The aforementioned at least one program, Repeating the input (A), input (B), and provide (C) for each compound-specific pair in the compound database, wherein the first compound in each compound-specific pair is the first compound, the second compound in each compound-specific pair is the second compound, and thereby a plurality of graph edit distances are formed, and repeating this process. By evaluating the multiple distances using a reference function, the multiple graph editing distances are clustered into multiple clusters, A computer system according to any one of claims 73 to 76, further comprising instructions for selecting a compound from one or more clusters within the plurality of clusters for evaluation in a drug discovery program.

78. The computer system according to any one of claims 73 to 77, wherein the at least one program further includes instructions for constructing a quantitative structure-activity relationship model comprising the first compound and the second compound using the estimated graph edit distance.

79. The computer system according to any one of claims 73 to 78, wherein the at least one program further includes instructions for determining an activity cliff for a drug target using the estimated value of the graph edit distance.

80. The computer system according to any one of claims 73 to 79, wherein the aforementioned model is a graph convolutional neural network.

81. The computer system according to any one of claims 73 to 79, wherein the aforementioned model is a graph neural network (message passing network).

82. The computer system according to any one of claims 73 to 79, wherein the model is a message-passing graph neural network consisting of three to eight layers.

83. The computer system according to any one of claims 73 to 79, wherein the model is a message-passing graph neural network consisting of five layers.

84. The computer system according to any one of claims 73 to 79, wherein the model is an autoencoder.

85. The computer system according to claim 84, wherein the autoencoder is a variational autoencoder, an adversarial autoencoder, a normalization autoencoder, a sparse autoencoder, or a metric learning autoencoder.

86. The first representation is the first fingerprint of the first compound, The computer system according to any one of claims 73 to 85, wherein the second representation is a second fingerprint of the second compound.

87. The first expression is a plurality of first descriptors of the first compound, The computer system according to any one of claims 73 to 85, wherein the second expression is a second plurality of descriptors of the second compound.

88. The first representation is a first three-dimensional structural representation of the first compound, The computer system according to any one of claims 73 to 85, wherein the second representation is a second three-dimensional structural representation of the second compound.

89. The first representation is a first tensor-based representation of the first compound, The computer system according to any one of claims 73 to 85, wherein the second representation is a second tensor-based representation of the second compound.

90. The first representation is a first string representation of the first compound, The computer system according to any one of claims 73 to 85, wherein the second representation is a second string representation of the second compound.

91. The computer system according to any one of claims 73 to 90, wherein the model includes 500,000 or more parameters.

92. The aforementioned model is 1 × 10 ６ A computer system according to any one of claims 73 to 90, comprising one or more parameters.

93. The computer system according to any one of claims 73 to 92, wherein the first compound and the second compound each satisfy two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 Daltons, and (iv) a LogP of less than 5.

94. The computer system according to any one of claims 73 to 93, wherein the first compound and the second compound have accurate graph editing distances of 1 to 10.

95. The computer system according to any one of claims 73 to 93, wherein the first compound and the second compound have accurate graph edit distances of 1 to 15.

96. The computer system according to any one of claims 73 to 95, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 500 daltons.

97. The computer system according to any one of claims 73 to 96, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 1,000 daltons.

98. The computer system according to any one of claims 73 to 96, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 2,000 daltons.

99. The computer system according to any one of claims 73 to 96, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 4,000 Daltons, less than 6,000 Daltons, less than 8,000 Daltons, less than 10,000 Daltons, or less than 20,000 Daltons.

100. The computer system according to any one of claims 73 to 96, wherein the first compound and the second compound are each organic compounds having a molecular weight of 300 daltons to 1500 daltons.

101. The aforementioned model performs at least 100,000 calculations to calculate the first isometrically preserved latent implantation. The computer system according to any one of claims 73 to 100, wherein the model performs at least 100,000 calculations to calculate the second isometrically preserved latent implant.

102. A method for estimating the graph edit distance between a first compound and a second compound, In a computer system comprising one or more processors and memory addressable by the one or more processors, (A) Inputting a first representation of the first compound into a model including a plurality of parameters, wherein the model applies the plurality of parameters to the first representation of the first compound to generate a first isometrically preserved latent embedding representing the first compound as a first output from the model, (B) Inputting a second representation of the second compound into the model, wherein the model applies the plurality of parameters to the second representation of the second compound to generate a second isometrically retained latent embedding representing the second compound as a second output from the model, (C) A method comprising providing an estimate of the graph edit distance between the first compound and the second compound using the difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

103. The method according to claim 102, wherein the difference is the L2 norm difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

104. The first isometrically preserved latent embedding is a first vector containing a first plurality of vector values, The second isometrically preserved latent embedding is a second vector containing a second set of vector values, The method according to claim 103, wherein the L2 norm difference between the first vector and the second vector is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first vector and the second vector, and the model is trained such that the L2 norm of the difference between the first vector and the second vector is an estimate of the GED between the first numerator and the second numerator.

105. The method according to any one of claims 102 to 104, further comprising determining that the second compound is a chemically related analog of the first compound when the estimated value of the graph edit distance falls below a threshold.

106. The method described above is Repeating the input (A), input (B), and provide (C) for each compound-specific pair in the compound database, wherein the first compound in each compound-specific pair is the first compound, the second compound in each compound-specific pair is the second compound, and thereby a plurality of graph edit distances are formed, and repeating this process. By evaluating the multiple distances using a reference function, the multiple graph editing distances are clustered into multiple clusters, The method according to any one of claims 102 to 105, further comprising selecting a compound from one or more clusters within the plurality of clusters for evaluation in a drug discovery program.

107. The method according to any one of claims 102 to 106, further comprising constructing a quantitative structure-activity relationship model comprising the first compound and the second compound using the estimated value of the graph edit distance.

108. The method according to any one of claims 102 to 107, further comprising determining an activity cliff for a drug target using the estimated value of the graph edit distance.

109. The method according to any one of claims 102 to 108, wherein the model is a graph convolutional neural network.

110. The method according to any one of claims 102 to 108, wherein the model is a graph neural network (message passing network).

111. The method according to any one of claims 102 to 108, wherein the model is a message-passing graph neural network consisting of three to eight layers.

112. The method according to any one of claims 102 to 111, wherein the model is a message-passing graph neural network consisting of five layers.

113. The method according to any one of claims 102 to 111, wherein the model is an autoencoder.

114. The method according to claim 113, wherein the autoencoder is a variational autoencoder, an adversarial autoencoder, a normalized autoencoder, a sparse autoencoder, or a metric learning autoencoder.

115. The first representation is the first fingerprint of the first compound, The method according to any one of claims 102 to 114, wherein the second expression is a second fingerprint of the second compound.

116. The first expression is a plurality of first descriptors of the first compound, The method according to any one of claims 102 to 114, wherein the second expression is a second plurality of descriptors of the second compound.

117. The first representation is a first three-dimensional structural representation of the first compound, The method according to any one of claims 102 to 114, wherein the second representation is a second three-dimensional structural representation of the second compound.

118. The first representation is a first tensor-based representation of the first compound, The method according to any one of claims 102 to 114, wherein the second representation is a second tensor-based representation of the second compound.

119. The first representation is a first string representation of the first compound, The method according to any one of claims 102 to 114, wherein the second expression is a second string representation of the second compound.

120. The method according to any one of claims 102 to 119, wherein the model includes 500,000 or more parameters.

121. The aforementioned model is 1 × 10 ６ The method according to any one of claims 102 to 119, comprising one or more parameters.

122. The method according to any one of claims 102 to 121, wherein the first compound and the second compound each satisfy two or more rules, three or more rules, or all four rules of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 Daltons, and (iv) a LogP of less than 5.

123. The method according to any one of claims 102 to 122, wherein the first compound and the second compound have an accurate graph edit distance of 1 to 10.

124. The method according to any one of claims 102 to 123, wherein the first compound and the second compound have accurate graph edit distances of 1 to 15.

125. The method according to any one of claims 102 to 124, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 500 daltons.

126. The method according to any one of claims 102 to 124, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 1,000 daltons.

127. The method according to any one of claims 102 to 124, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 2,000 daltons.

128. The method according to any one of claims 102 to 124, wherein the first compound and the second compound are organic compounds having a molecular weight of less than 4,000 Daltons, less than 6,000 Daltons, less than 8,000 Daltons, less than 10,000 Daltons, or less than 20,000 Daltons, respectively.

129. The method according to any one of claims 102 to 124, wherein the first compound and the second compound are each organic compounds having a molecular weight of 300 daltons to 1500 daltons.

130. The aforementioned model performs at least 100,000 calculations to compute the first latent embedding. The method according to any one of claims 102 to 129, wherein the model performs at least 100,000 calculations to calculate the second latent embedding.

131. A non-temporary computer-readable storage medium, which, when executed by a computer system, stores instructions causing the computer system to perform a method for estimating the graph edit distance between a first compound and a second compound, and the method is (A) Inputting a first representation of the first compound into a model including a plurality of parameters, wherein the model applies the plurality of parameters to the first representation of the first compound to generate a first isometrically preserved latent embedding representing the first compound as a first output from the model, (B) Inputting a second representation of the second compound into the model, wherein the model applies the plurality of parameters to the second representation of the second compound to generate a second isometrically retained latent embedding representing the second compound as a second output from the model, (C) A non-temporary computer-readable storage medium comprising providing an estimate of the graph edit distance between the first compound and the second compound using the difference between the first isometrically retained latent embedding and the second isometrically retained latent embedding.

132. The non-temporary computer-readable storage medium according to claim 131, wherein the difference is the L2 norm difference between the first isometrically retained latent implant and the second isometrically retained latent implant.

133. The first isometrically preserved latent embedding is a first vector containing a first plurality of vector values, The second isometrically preserved latent embedding is a second vector containing a second set of vector values, A non-temporary computer-readable storage medium according to claim 132, wherein the L2 norm difference between the first vector and the second vector is taken as the square root of the sum of the squared differences of the corresponding vector values ​​between the first vector and the second vector, and the model is trained such that the L2 norm of the difference between the first vector and the second vector is an estimate of the GED between the first numerator and the second numerator.

134. A non-temporary computer-readable storage medium according to any one of claims 131 to 133, further comprising determining that the second compound is a chemically related analog of the first compound when the estimated value of the graph edit distance falls below a threshold.

135. The method described above is Repeating the input (A), input (B), and provide (C) for each compound-specific pair in the compound database, wherein the first compound in each compound-specific pair is the first compound, the second compound in each compound-specific pair is the second compound, and thereby a plurality of graph edit distances are formed, and repeating this process. By evaluating the multiple distances using a reference function, the multiple graph editing distances are clustered into multiple clusters, A non-temporary computer-readable storage medium according to any one of claims 131 to 134, further comprising selecting a compound from one or more clusters within the plurality of clusters for evaluation in a drug discovery program.

136. A non-temporary computer-readable storage medium according to any one of claims 131 to 135, further comprising constructing a quantitative structure-activity relationship model comprising the first compound and the second compound using the estimated value of the graph edit distance.

137. A non-temporary computer-readable storage medium according to any one of claims 131 to 136, further comprising determining an activity cliff for a drug target using the estimated value of the graph edit distance.

138. A non-temporary computer-readable storage medium according to any one of claims 131 to 137, wherein the model is a graph convolutional neural network.

139. A non-temporary computer-readable storage medium according to any one of claims 131 to 137, wherein the model is a graph neural network (message passing network).

140. A non-temporary computer-readable storage medium according to any one of claims 131 to 137, wherein the model is a message-passing graph neural network consisting of three to eight layers.

141. A non-temporary computer-readable storage medium according to any one of claims 131 to 137, wherein the model is a message-passing graph neural network consisting of five layers.

142. The non-temporary computer-readable storage medium according to any one of claims 131 to 137, wherein the model is an autoencoder.

143. The non-temporary computer-readable storage medium according to claim 142, wherein the autoencoder is a variational autoencoder, adversarial autoencoder, normalization autoencoder, sparse autoencoder, or metric learning autoencoder.

144. The first representation is the first fingerprint of the first compound, The non-temporary computer-readable storage medium according to any one of claims 131 to 143, wherein the second representation is a second fingerprint of the second compound.

145. The first expression is a plurality of first descriptors of the first compound, A non-temporary computer-readable storage medium according to any one of claims 131 to 143, wherein the second expression is a second plurality of descriptors of the second compound.

146. The first representation is a first three-dimensional structural representation of the first compound, A non-temporary computer-readable storage medium according to any one of claims 131 to 143, wherein the second representation is a second three-dimensional structural representation of the second compound.

147. The first representation is a first tensor-based representation of the first compound, A non-temporary computer-readable storage medium according to any one of claims 131 to 143, wherein the second representation is a second tensor-based representation of the second compound.

148. The first representation is a first string representation of the first compound, A non-temporary computer-readable storage medium according to any one of claims 131 to 143, wherein the second representation is a second string representation of the second compound.

149. A non-temporary computer-readable storage medium according to any one of claims 131 to 148, wherein the model includes 500,000 or more parameters.

150. The aforementioned model is 1 × 10 ６ A non-temporary computer-readable storage medium according to any one of claims 131 to 148, comprising one or more parameters.

151. A non-temporary computer-readable storage medium according to any one of claims 131 to 150, wherein the first compound and the second compound each satisfy two or more, three or more, or all four of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 Daltons, and (iv) a LogP of less than 5.

152. A non-temporary computer-readable storage medium according to any one of claims 131 to 151, wherein the first compound and the second compound have an accurate graph editing distance of 1 to 10.

153. A non-temporary computer-readable storage medium according to any one of claims 131 to 151, wherein the first compound and the second compound have accurate graph editing distances of 1 to 15.

154. A non-temporary computer-readable storage medium according to any one of claims 131 to 153, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 500 daltons.

155. A non-temporary computer-readable storage medium according to any one of claims 131 to 153, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 1,000 daltons.

156. A non-temporary computer-readable storage medium according to any one of claims 131 to 153, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 2,000 daltons.

157. A non-temporary computer-readable storage medium according to any one of claims 131 to 153, wherein the first compound and the second compound are each organic compounds having a molecular weight of less than 4,000 Daltons, less than 6,000 Daltons, less than 8,000 Daltons, less than 10,000 Daltons, or less than 20,000 Daltons.

158. A non-temporary computer-readable storage medium according to any one of claims 131 to 153, wherein the first compound and the second compound are each organic compounds having a molecular weight of 300 daltons to 1500 daltons.

159. The aforementioned model performs at least 100,000 calculations to compute the first latent embedding. The non-temporary computer-readable storage medium according to any one of claims 131 to 153, wherein the model performs at least 100,000 calculations to compute the second latent embedding.

160. A computer system for training a model to predict the efficacy of a test compound against a target protein, One or more processors, The system comprises a memory addressable by one or more processors, wherein the memory stores at least one program for execution by the one or more processors, and the at least one program is Obtaining a dataset containing data for multiple training compounds, wherein each training compound in the multiple training compounds is labeled with (i) a protein target in a plurality of protein targets, (ii) a binding site for the respective protein target, and (iii) efficacy for the protein target, wherein at least 5 percent of the plurality of protein targets contain a plurality of binding sites, and at least 5 percent of the multiple training compounds specify allosteric sites of the corresponding protein targets in the plurality of protein targets. A computer system comprising instructions for training a model using the dataset, wherein for each of the plurality of training compounds, the pose of each training compound within the binding site of the target protein identified in the dataset for the training compound is input to the model and evaluated for the effectiveness of each training compound.

161. The aforementioned at least one program, When examining the test chemical compound, The process involves posing the test chemical compound at the binding site of the target protein to obtain a test pose. The computer system according to claim 160, further comprising instructions for inputting the test pose into the model to obtain a prediction of the efficacy of the test compound against the target protein.

162. The computer system according to claim 161, wherein the target protein is located within the plurality of protein targets.

163. The computer system according to claim 161, wherein the target protein is any protein other than any of the plurality of protein targets.

164. The computer system according to any one of claims 160 to 163, wherein each training compound in the plurality of training compounds satisfies two or more rules of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) molecular weight less than 500 Daltons, and (iv) LogP less than 5.

165. The computer system according to any one of claims 160 to 163, wherein each training compound in the plurality of training compounds satisfies three or more of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) molecular weight less than 500 Daltons, and (iv) LogP less than 5.

166. The computer system according to any one of claims 160 to 163, wherein each training compound in the plurality of training compounds has (i) 5 or fewer hydrogen bond donors, (ii) 10 or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 daltons, and (iv) less than 5 LogP.

167. The computer system according to any one of claims 160 to 166, wherein each training compound in the plurality of training compounds is an organic compound having a molecular weight of less than 500 daltons.

168. The computer system according to any one of claims 160 to 166, wherein each training compound in the plurality of training compounds is an organic compound having a molecular weight of less than 1,000 daltons.

169. The computer system according to any one of claims 160 to 166, wherein each training compound in the plurality of training compounds is an organic compound having a molecular weight of less than 2,000 daltons.

170. The computer system according to any one of claims 160 to 166, wherein each training compound in the plurality of training compounds is an organic compound having a molecular weight of less than 4,000 Daltons, less than 6,000 Daltons, less than 8,000 Daltons, less than 10,000 Daltons, or less than 20,000 Daltons.

171. The computer system according to any one of claims 160 to 166, wherein each training compound in the plurality of training compounds is an organic compound having a molecular weight of 350 daltons to 10,000 daltons.

172. The computer system according to any one of claims 160 to 171, wherein the plurality of training compounds comprises 1,000 or more training compounds.

173. The computer system according to any one of claims 160 to 171, wherein the plurality of training compounds comprises 2,000 or more training compounds.

174. The computer system according to any one of claims 160 to 171, wherein the plurality of training compounds comprises 10,000 or more training compounds.

175. The computer system according to any one of claims 160 to 174, wherein each of the protein targets in the plurality of protein targets is represented by one or more compounds in the plurality of compounds, and the plurality of protein targets comprises 10 or more protein targets.

176. The computer system according to any one of claims 160 to 174, wherein each of the protein targets in the plurality of protein targets is represented by one or more compounds in the plurality of compounds, and the plurality of protein targets comprises 100 or more protein targets.

177. The computer system according to any one of claims 160 to 174, wherein each of the protein targets in the plurality of protein targets is represented by 10 or more compounds in the plurality of compounds, and the plurality of protein targets comprises 10 or more protein targets.

178. The computer system according to any one of claims 160 to 177, wherein at least 10 percent of the plurality of protein targets include a plurality of binding sites.

179. The computer system according to any one of claims 160 to 177, wherein at least 15 percent of the plurality of protein targets include a plurality of binding sites.

180. The computer system according to claim 161, wherein the test compound satisfies two or more of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 daltons, and (iv) a LogP of less than 5.

181. The computer system according to claim 161, wherein the test compound satisfies three or more of Lipinski's Rule of Five: (i) five or fewer hydrogen bond donors, (ii) ten or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 daltons, and (iv) a LogP of less than 5.

182. The computer system according to claim 161, wherein the test compound has (i) 5 or fewer hydrogen bond donors, (ii) 10 or fewer hydrogen bond acceptors, (iii) a molecular weight of less than 500 daltons, and (iv) less than 5 logP.

183. The computer system according to claim 161, wherein the test compound is an organic compound having a molecular weight of less than 500 daltons.

184. The computer system according to claim 161, wherein the test compound is an organic compound having a molecular weight of less than 1,000 daltons.

185. The computer system according to claim 161, wherein the test compound is an organic compound having a molecular weight of less than 2,000 daltons, less than 4,000 daltons, less than 6,000 daltons, less than 8,000 daltons, less than 10,000 daltons, or less than 20,000 daltons.

186. The computer system according to claim 161, wherein the test compound is an organic compound having a molecular weight of 350 daltons to 10,000 daltons.

187. The aforementioned model includes multiple parameters, The computer system according to claim 161, wherein the model applies the plurality of parameters to the test pose through at least 10,000 calculations to generate the predicted potency of the test compound.

188. The computer system according to any one of claims 160 to 187, wherein the model is a graph neural network.

189. The computer system according to any one of claims 160 to 188, wherein the model includes 500,000 or more parameters.

190. The aforementioned model is 1 × 10 ６ A computer system according to any one of claims 160 to 189, comprising one or more parameters.

191. The computer system according to any one of claims 160 to 190, wherein the efficacy against the protein target is pKi.

192. The computer system according to any one of claims 160 to 90, wherein the efficacy against the protein target is IC50, EC50, Kd, ​​or KI.

193. The computer system according to any one of claims 160 to 192, wherein the efficacy of the test compound against the target protein is pKi.

194. The computer system according to any one of claims 160 to 192, wherein the efficacy of the test compound against the target protein is IC50, EC50, Kd, ​​or KI.

195. A method for training a model to predict the efficacy of a test compound against a target protein, In a computer system comprising one or more processors and memory addressable by the one or more processors, Obtaining a dataset containing data for multiple training compounds, wherein each training compound in the multiple training compounds is labeled with (i) a protein target in a plurality of protein targets, (ii) a binding site for the respective protein target, and (iii) efficacy for the protein target, wherein at least 5 percent of the plurality of protein targets contain a plurality of binding sites, and at least 5 percent of the multiple training compounds specify allosteric sites of the corresponding protein targets in the plurality of protein targets. A method for training a model using the dataset, wherein for each of the plurality of training compounds, the pose of each training compound within the binding site of the target protein identified in the dataset for the training compound is input to the model and evaluated for the effectiveness of each training compound.

196. The method described above is When examining the test chemical compound, The process involves posing the test chemical compound at the binding site of the target protein to obtain a test pose. The method according to claim 195, further comprising inputting the test pose into the model to obtain a prediction of the efficacy of the test compound against the target protein.

197. A non-temporary computer-readable storage medium, which, when executed by a computer system, stores instructions causing the computer system to perform a method for training a model to predict the efficacy of a test compound against a target protein, the method being: Obtaining a dataset containing data for multiple training compounds, wherein each training compound in the multiple training compounds is labeled with (i) a protein target in a plurality of protein targets, (ii) a binding site for the respective protein target, and (iii) efficacy for the protein target, wherein at least 5 percent of the plurality of protein targets contain a plurality of binding sites, and at least 5 percent of the multiple training compounds specify allosteric sites of the corresponding protein targets in the plurality of protein targets. A non-temporary computer-readable storage medium, comprising training a model using the dataset, wherein for each of the plurality of training compounds, the pose of each training compound within the binding site of the target protein identified in the dataset for the training compound is input to the model and evaluated for the effectiveness of each training compound.

198. The method described above is When examining the test chemical compound, The process involves posing the test chemical compound at the binding site of the target protein to obtain a test pose. A non-temporary computer-readable storage medium according to claim 197, further comprising inputting the test pose into the model to obtain a prediction of the efficacy of the test compound against the target protein.

Images