Characterization of interactions between compounds and polymers using attitude ensemble

Abstract

Systems and methods of characterizing interactions between a compound and a polymer include obtaining a plurality of sets of atomic coordinates. Each set of atomic coordinates includes the compound bound to the polymer in a corresponding pose of a plurality of poses. And sequentially inputting each corresponding group of atomic coordinates or codes thereof into the neural network to obtain corresponding initial embedding as output, thereby obtaining a plurality of initial embedding. Each initial embedding corresponds to a set of atomic coordinates of the plurality of sets of atomic coordinates. An attention mechanism is applied to the plurality of initial inserts in a cascade form to obtain an attention insert. A pooled function is applied to the attention embedding to derive a pooled embedding. The pooling is embedded into a model to obtain an interaction score for the interaction between the compound and the polymer.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 336,841, filed on April 29, 2022, entitled “CHARACTERIZATION OF INTERACTIONSBETWEEN COMPOUNDS AND POLYMERS USING POSE ENSEMBLES,” which is hereby incorporated by reference into this document. Technical Field

[0003] This application aims to use the model to characterize the interaction between the test compound and the target polymer. Background Art

[0004] Fundamentally, biological systems operate through the physical interaction of molecules, such as compounds, with target polymers. Structure-based virtual high-throughput screening (vHTS) methods have been used to characterize the interactions between candidate (test) compounds and target polymers through machine learning methods. For example, such characterizations can report continuous or categorical activity labels, pKa, or any other suitable metric to characterize the interaction between candidate compounds and target polymers.

[0005] One drawback of machine learning approaches to vHTS is that they do not adequately consider the enthalpic and entropic components of receptor-ligand complex formation. Structure-based deep learning approaches typically predict bioactivity from docked static ligand poses. However, these approaches ignore the entropic contribution to the free energy change. Predicting bioactivity from an ensemble of docked poses can overcome this limitation, but it requires the model to be able to recognize and be sensitive to different poses. However, traditional machine learning approaches are often insensitive to different poses. Figure 19 illustrates this insensitivity, where machine learning models such as convolutional neural networks incorrectly favor poses that have all the correct components but are not correct overall at all. Figure 18 shows a situation where the left and right poses have the same parts (two eyes, two eyebrows, nose, lips, and the overall shape of the head). Therefore, it may be difficult to teach the machine learning model that the pose on the left is the correct pose. Therefore, there is an inherent pose insensitivity in traditional vHTS machine learning approaches. This pose insensitivity can lead to incorrect or inaccurate characterization of the interaction between the test compound and the target polymer. For example, this pose insensitivity could cause a vHTS machine learning approach (which provides a categorical activity label for each compound in a screening library) to incorrectly label a certain percentage of compounds in a screening library.

[0006] In view of the above background, what is needed in the art is a method to impose posture sensitivity on vHTS machine learning methods, so that such vHTS machine learning methods are sensitive to undesirable postures. Summary of the invention

[0007] The present disclosure solves the problem of identification in context by utilizing a vHTS machine learning model that predicts biological activities from multiple poses simultaneously. The disclosed vHTS machine learning model conditional multi-task architecture enhances sensitivity to different ligand poses, and its architecture includes an attention mechanism that exploits hidden correlations in pose distributions. The disclosed vHTS machine learning model improves biological activity predictions compared to a baseline model that predicts from static ligand poses alone.

[0008] Therefore, one aspect of the present disclosure provides a computer system for characterizing the interaction between a test compound and a target polymer. The computer system includes one or more processors and a memory addressable by the one or more processors. The memory stores at least one program for being executed by one or more processors. The at least one program includes instructions for obtaining multiple sets of atomic coordinates. Each corresponding set of atomic coordinates in the multiple sets of atomic coordinates includes a test compound bound to the target polymer in a corresponding posture in multiple postures. In some embodiments, each corresponding set of atomic coordinates in the multiple sets of atomic coordinates includes atomic coordinates of at least 30 atoms.

[0009] For each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates, the at least one program includes instructions for inputting the corresponding set of atomic coordinates or an encoding of the corresponding set of atomic coordinates into the first neural network to obtain a corresponding initial embedding as an output of the first neural network, thereby obtaining a plurality of initial embeddings. Each initial embedding in the plurality of initial embeddings corresponds to a set of atomic coordinates in the plurality of sets of atomic coordinates. The first neural network includes more than 400 parameters.

[0010] The at least one program further includes instructions for applying an attention mechanism to the plurality of initial embeddings in a cascaded fashion to obtain an attentive embedding.

[0011] The at least one program further includes instructions for applying a pooling function to the attention embedding to derive a pooled embedding.

[0012] The at least one program further includes instructions for inputting the pooled embedding into the first model to obtain a first interaction score for the interaction between the test compound and the target polymer. The first model includes more than 400 parameters.

[0013] In some embodiments, the first interaction score represents the binding coefficient of the test compound to the target polymer. In some such embodiments, the binding coefficient is the IC of the test compound relative to the target polymer.50 ,EC 50、 Kd, KI or pKI.

[0014] In some embodiments, the first interaction score represents a quality score of an in silico pose of the test compound to the target polymer.

[0015] In some embodiments, the first model is a fully connected second neural network.

[0016] In some embodiments, the at least one program further includes an instruction for inputting the pooling embedding into the second model to obtain a second interaction score for the interaction between the test compound and the target polymer. In such embodiments, the first model is a first fully connected neural network, the second model is a second fully connected neural network, the first interaction score represents the computer-simulated gesture quality score of the test compound to the target polymer, and the second interaction score represents the computer-simulated gesture quality score of the test compound to the target polymer. In some such embodiments, at least one program further includes an instruction for inputting the first interaction score and the second interaction score into a third model to obtain a third interaction score, wherein the third model is a third fully connected neural network. In some such embodiments, the third interaction score is a discrete binary activity score having a first value when the test compound is determined to be inactive by the third model, and having a second value when the test compound is determined to be active by the third model.

[0017] In some embodiments, the polymer of interest is an assembly of a protein, a polypeptide, a polynucleic acid, a polyribonucleic acid, a polysaccharide, or any combination thereof.

[0018] In some embodiments, each set of atomic coordinates in the plurality of sets of atomic coordinates comprises three-dimensional coordinates {x1, ..., x2, ..., x3, ..., x4, ..., x5, ..., x6, ..., x7, ..., x8, ..., x9, ..., ..., ...9, ..., ...10, ..., ...11, ..., ...12, ..., ...13, ..., ...14, ..., ...15, ..., ...16, ..., ...17, ..., ... N}, the crystal structure of the target polymer is or better resolution or or better resolution.

[0019] In some embodiments, each set of atomic coordinates in the plurality of sets of atomic coordinates comprises an ensemble of three-dimensional coordinates of at least a portion of the target polymer determined by nuclear magnetic resonance, neutron diffraction, or cryo-electron microscopy.

[0020] In some embodiments, the first interaction score is a binary score, wherein a first value of the binary score represents an IC of the test compound relative to the target polymer that is above a first threshold. 50 ,EC 50、 Kd, KI or pKI, and the second value of the binary score represents the IC of the test compound relative to the target polymer below the first threshold50 ,EC 50 , Kd, ​​KI or pKI.

[0021] In some embodiments, the test compound satisfies two or more, three or more, or all four of the following Lipinski five-fold rule: (i) no more than five hydrogen bond donors, (ii) no more than ten hydrogen bond acceptors, (iii) molecular weight less than 500 Daltons, and (iv) LogP less than 5.

[0022] 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.

[0023] In some embodiments, the test compound is an organic compound having a molecular weight between 400 Daltons and 10,000 Daltons.

[0024] In some embodiments, the multiple sets of atomic coordinates consist of between 3 and 64 poses. In some embodiments, the multiple sets of atomic coordinates consist of between 2 and 64 poses.

[0025] In some embodiments, the first neural network is a convolutional neural network. In some such embodiments, each respective set of atomic coordinates in the plurality of sets of atomic coordinates includes atomic coordinates of at least 80 atoms.

[0026] In some embodiments, the first neural network is a graph neural network. In some such embodiments, the graph neural network is characterized by an initial embedding layer and a plurality of interaction layers, the plurality of interaction layers each contributing an interaction data structure in a plurality of interaction data results for each atom in a respective set of groups of atoms of atomic coordinates of a corresponding pose in a plurality of poses, the plurality of poses being pooled to form a corresponding initial embedding for the corresponding pose.

[0027] In some embodiments, the first neural network is an equivariant neural network or a message passing neural network.

[0028] In some embodiments, the first neural network includes a plurality of graph convolution blocks, and each block uses a plurality of radial graphs to consider connectivity within a corresponding set of atomic coordinates.

[0029] In some embodiments, the first neural network includes 1x10 6 parameters.

[0030] In some embodiments, the corresponding initial embedding includes a data structure having between 128 and 768 values. In some embodiments, the corresponding initial embedding includes a data structure having more than 100 values. In some embodiments, the corresponding initial embedding includes a data structure having more than 80 values, more than 100 values, more than 120 values, more than 140 values, or more than 160 values. In some embodiments, the corresponding initial embedding includes a data structure consisting of between 100 values ​​and 2000 values.

[0031] In some embodiments, the plurality of initial embeddings includes a first plurality of values ​​and applying the attention mechanism includes: (i) inputting the first plurality of values ​​into an attention neural network to obtain a first plurality of weights, wherein each weight in the first plurality of weights corresponds to a corresponding value in the first plurality of values, and (ii) weighting each corresponding value in the first plurality of values ​​by a corresponding weight in the plurality of weights to obtain the attention embedding. In some such embodiments, the first plurality of weights sums to one, and each weight in the first plurality of weights is a scalar value between zero and one.

[0032] In some embodiments, the pooling function folds the attention embedding into the pooled embedding by applying a statistical function to combine each portion of the attention embedding representing a different pose in the plurality of poses to form the pooled embedding. In some such embodiments, the attention embedding includes a corresponding plurality of values ​​for a corresponding plurality of elements for each respective pose in the plurality of poses, and the statistical function is a maximum function that takes the maximum value across the corresponding elements of each respective pose represented in the attention embedding to form the pooled embedding. In some embodiments, the attention embedding includes a corresponding plurality of values ​​for a corresponding plurality of elements for each respective pose in the plurality of poses, and the statistical function is an average function that averages the corresponding elements of each respective pose represented in the attention embedding to form the pooled embedding.

[0033] In some embodiments, the first model is a regression task and the first interaction score quantifies the interaction between the test compound and the target polymer.

[0034] In some embodiments, the first model is a classification task and the first interaction score classifies the interaction between the test compound and the target polymer.

[0035] Another aspect of the present disclosure provides a method for characterizing the interaction between a test compound and a target polymer. The method includes obtaining multiple sets of atomic coordinates. Each corresponding set of atomic coordinates in the multiple sets of atomic coordinates includes a test compound bound to a target polymer in a corresponding posture in multiple postures. In some embodiments, each corresponding set of atomic coordinates in the multiple sets of atomic coordinates includes atomic coordinates of at least 30 atoms. Further, in the method, for each corresponding set of atomic coordinates in the multiple sets of atomic coordinates, the encoding of the corresponding set of atomic coordinates or the corresponding set of atomic coordinates is input into a first neural network to obtain a corresponding initial embedding as the output of the first neural network. In this way, multiple initial embeddings are obtained, wherein each initial embedding in the multiple initial embeddings corresponds to a set of atomic coordinates in the multiple sets of atomic coordinates. In some such embodiments, the first neural network includes more than 400 parameters. The method further includes applying an attention mechanism to multiple initial embeddings in a cascaded form to obtain an attention embedding. The method further includes applying a pooling function to the attention embedding to derive a pooling embedding. The method further includes inputting the pooling embedding into a first model to obtain a first interaction score for the interaction between the test compound and the target polymer. In some embodiments, the first model includes more than 400 parameters.

[0036] Another aspect of the present disclosure provides a non-transitory computer-readable storage medium. A non-transitory computer-readable storage medium stores instructions that, when executed by a computer system, cause the computer system to execute a method for characterizing an interaction between a test compound and a target polymer. The method includes obtaining multiple sets of atomic coordinates. Each corresponding set of atomic coordinates in the multiple sets of atomic coordinates includes a test compound bound to a target polymer in a corresponding posture in a plurality of postures. In some embodiments, each corresponding set of atomic coordinates in the multiple sets of atomic coordinates includes atomic coordinates of at least 30 atoms. The method further includes, for each set of atomic coordinates in the multiple sets of atomic coordinates, inputting a corresponding set of atomic coordinates or an encoding of a corresponding set of atomic coordinates into a first neural network to obtain a corresponding initial embedding as an output of the first neural network. In this way, multiple initial embeddings are obtained, wherein each initial embedding in the multiple initial embeddings corresponds to a set of atomic coordinates in the multiple sets of atomic coordinates. In some embodiments, the first neural network includes more than 400 parameters. The method further includes applying an attention mechanism to multiple initial embeddings in a cascaded form to obtain an attention embedding. The method further includes applying a pooling function to the attention embedding to derive a pooling embedding. The method further includes inputting the pooled embedding into the first model to obtain a first interaction score for the interaction between the test compound and the target polymer. In some embodiments, the first model includes more than 400 parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In the accompanying drawings, embodiments of the disclosed system and method are shown by way of example. It should be clearly understood that the description and drawings are only for illustrative purposes and as an aid to understanding, and are not intended to be a definition of the limits of the disclosed system and method.

[0038] Figure 1A and 1B A computer system according to some embodiments of the present disclosure is shown.

[0039] Figure 2A , 2B , 2C, 2D, 2E and 2F show methods for characterizing the interaction between a test compound and a target polymer according to some embodiments of the present disclosure.

[0040] Figure 3 Schematic diagram of example training compounds showing their pose relative to a target polymer according to some embodiments of the present disclosure.

[0041] Figure 4 is a schematic diagram of a geometric representation of input features in the form of a three-dimensional voxel grid according to some embodiments of the present disclosure.

[0042] Figure 5 and 6 Schematic diagram of a compound encoded onto a two-dimensional voxel grid according to some embodiments of the present invention.

[0043] Figure 7 According to some embodiments of the present invention Figure 6 A visualization of the image where the voxels are numbered.

[0044] Figure 8 is a schematic diagram of a geometric representation of an input feature in the form of coordinate positions of atomic centers according to some embodiments of the present disclosure.

[0045] Fig. 9A A system for characterizing an interaction between a test compound and a target polymer according to an embodiment of the present disclosure is presented.

[0046] Fig. 9B A system for characterizing the interaction between a test compound and a target polymer according to another embodiment of the present disclosure is presented.

[0047] Fig. 9C A system for characterizing the interaction between a test compound and a target polymer according to another embodiment of the present disclosure is presented, wherein the first neural network is a graph-based neural network.

[0048] Fig.10A system for characterizing the interaction between a test compound and a target polymer, wherein the characterization is (i) a binary discrete activity and (ii) a pKi, and wherein the system is trained using the poses of training compounds according to one embodiment of the present disclosure.

[0049] Fig.11 A system for characterizing an interaction between a test compound and a target polymer, wherein the characterization is pKi, and wherein the pKi is conditioned in part on activity, and wherein the system is trained using the poses of training compounds according to an embodiment of the present disclosure.

[0050] Fig.12 A system for characterizing an interaction between a test compound and a target polymer, wherein the characterization is activity, and wherein the activity is conditioned in part on a pKi and a pose quality score, and wherein the system is trained using poses of training compounds according to an embodiment of the present disclosure.

[0051] Fig.13 A system for characterizing an interaction between a test compound and a target polymer, wherein the characterization is activity, and wherein the activity is conditioned in part on a pKi and a binding mode score, and wherein the system is trained using the poses of training compounds according to an embodiment of the present disclosure.

[0052] Fig.14 A system for characterizing the interaction between a test compound and a target polymer, wherein the characterization is activity and two different compound binding mode scores, and wherein the system is trained using the poses of training compounds according to one embodiment of the present disclosure.

[0053] Fig.15 A system for characterizing the interaction between a test compound and a target polymer, wherein the characterization is activity, two different compound binding mode scores and pKi, and wherein the system is trained using the poses of training compounds according to an embodiment of the present disclosure.

[0054] Fig.16A A system for characterizing an interaction between a test compound and a target polymer, wherein the characterization is activity, and wherein the activity is conditioned in part on a pKi and a binding mode score, and wherein the system is trained using the poses of training compounds according to an embodiment of the present disclosure.

[0055] Fig. 16B A system for characterizing an interaction between a test compound and a target polymer, wherein the characterization is activity, and wherein the activity is conditioned in part on pKi and two different binding mode scores, and wherein the system is trained using the poses of training compounds according to an embodiment of the present disclosure.

[0056] Fig.17 In order to apply a plurality of function computation elements (g1, g2, ...) to a voxel input (x1, x2, ..., x 100 ) and use g() to combine the function's element-wise outputs.

[0057] FIG. 18 illustrates the insensitivity faced by prior art machine learning models when characterizing the pose of a compound relative to a target polymer.

[0058] Figure 19 demonstrates the insensitivity of conventional machine learning models according to the prior art to the quality of compound-polymer poses, wherein, as shown, the best possible pose receives the same score by the machine learning model as a worse pose, and wherein an untrustworthy pose receives the same score by the machine learning model as the best possible pose.

[0059] Fig. 20 An activity task conditioned on PoseRanker and Vina scores according to an embodiment of the present disclosure is demonstrated.

[0060] Fig.21A and 21B Performance statistics of the disclosed architectures (o3-2.8.0 and o4-2.8.0) relative to other architectures (n8b-long and n8b-maxlong) are provided.

[0061] Like reference numerals refer to corresponding parts throughout the several views of the drawings. DETAILED DESCRIPTION

[0062] Reference will now be made in detail to the embodiments, examples of which are shown in the accompanying drawings. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it will be apparent that one of ordinary skill in the art may practice the present disclosure without these specific details. In other cases, well-known methods, processes, components, circuits, and networks are not described in detail in order to avoid unnecessarily obscuring aspects of the embodiments.

[0063] The present disclosure provides a system and method for characterizing the interaction between a compound and a polymer. In the method, multiple groups of atomic coordinates are obtained. Each group of these groups of atomic coordinates includes a compound bound to a polymer in a corresponding posture in a plurality of postures. Each corresponding group of atomic coordinates or its encoding is sequentially input into a neural network to obtain a corresponding initial embedding as an output. Multiple initial embeddings are calculated in this way. Each initial embedding corresponds to a group of atomic coordinates in a plurality of groups of atomic coordinates. An attention mechanism is applied to multiple initial embeddings in a cascade to obtain an attention embedding. In some embodiments, the attention mechanism is a neural network that is trained on test data to emphasize some parts of multiple initial embeddings while weakening some parts of multiple initial embeddings. A pooling function is applied to the attention embedding to derive a pooled embedding. Therefore, the pooling function collapses all initial embeddings representing multiple postures into a single composite embedding representing all postures in multiple postures. The pooled embedding is input into the model to obtain an interaction score for the interaction between the compound and the polymer.

[0064] It should also be understood that although the terms "first", "second", etc. can be used to describe various elements in this article, these elements should not be limited by these terms. These terms are only used to distinguish one element from another element. For example, without departing from the scope of the present disclosure, the first subject can be referred to as the second subject, and similarly, the second subject can be referred to as the first subject. Although the first subject and the second subject are both subjects, these subjects are not the same subject.

[0065] The terms used in this disclosure are only used for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the description of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "one", "a kind of" and "said" are intended to include plural forms as well. It should also be understood that, as used herein, the term "and / or" refers to and encompasses any and all possible combinations of one or more of the associated listed items. It should be further understood that, when used in this specification, the terms "comprises" and / or "comprising" specify the presence of stated features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or groups thereof.

[0066] As used herein, the term "if" may be interpreted to mean "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if [a stated condition or event] is detected" may be interpreted to mean "upon determining" or "in response to determining" or "upon detecting [a stated condition or event]" or "in response to detecting [a stated condition or event]," depending on the context.

[0067] 1 shows a computer system 100 for characterizing the interaction between a test compound and a target polymer. For example, it can be used as a binding affinity prediction system to generate accurate predictions about the binding affinity of one or more test compounds to a target polymer.

[0068] Referring to FIG. 1 , in a typical embodiment, computer system 100 includes one or more computers. For the purpose of illustration in FIG. 1 , computer system 100 is represented as a single computer including all functions of disclosed computer system 100. However, the present disclosure is not limited thereto. The functions of computer system 100 may be distributed across any number of networked computers and / or reside on each of several networked computers and / or virtual machines. Those skilled in the art will appreciate that various different computer topologies are possible for computer system 100, and all such topologies are within the scope of the present disclosure.

[0069] With the foregoing in mind, turning to FIG. 1 , the computer system 100 includes one or more processors (CPUs) 59, a network or other communication interface 84, a user interface 78 (e.g., including 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 disk storage and / or persistent devices 90 optionally accessed by one or more controllers 88, one or more communication buses 12 for interconnecting the above components, and a power supply 79 for powering the above components. To the extent that components of the memory 92 are not persistent, data in the memory 92 can be seamlessly shared with the non-volatile memory 90 or non-volatile / persistent portions of the memory 92 using known computing techniques such as caching. The memory 92 and / or storage 90 may include mass storage located remotely relative to the central processor 59. In other words, some of the data stored in the memory 92 and / or storage 90 may actually be hosted on a computer external to the computer system 100, but may be electronically accessed by the computer system 100 using the network interface 84 over the Internet, an intranet, or other form of network or electronic cable. In some embodiments, computer system 100 utilizes models running from memory associated with one or more graphics processors to improve system speed and performance. In some alternative embodiments, computer system 100 utilizes models running from memory 92 rather than memory associated with a graphics processor.

[0070] The memory 92 of the computer system 100 stores:

[0071] An optional operating system 34, including programs that handle various basic system services;

[0072] A spatial data evaluation module 36 for characterizing the interaction between the test compound (or training compound) and the target polymer;

[0073] Data of a target polymer 38, including structural data of the target polymer (multiple atomic spatial coordinates 40 of the target polymer) and optional active site information 42 of the target polymer;

[0074] a training data set 44 comprising a plurality of electronic descriptions, each electronic description 46 of the plurality of electronic descriptions corresponding to a training compound of a plurality of training compounds (and / or test compounds) and comprising (i) a plurality of poses of the corresponding compound, each corresponding pose 48 of the corresponding compound being represented by: (a) a set of corresponding atomic spatial coordinates 49 detailing the atomic coordinates of the corresponding compound in the corresponding pose relative to the spatial coordinates 40 of the target polymer 38, (b) an optional corresponding voxel map 52 detailing the atomic interactions of the corresponding compound in the corresponding pose relative to the target polymer based on the corresponding set of atomic coordinates, and (c) an optional corresponding vector 54 encoding the interactions of the corresponding compound with respect to the target polymer in the corresponding pose based on the corresponding set of atomic coordinates 49 and / or the corresponding voxel map 52, (ii) a (first) interaction score 50 between the corresponding compound and the target polymer 38, (iii) an optional activity score 56 between the corresponding compound and the target polymer 38, and (iv) an optional (second) interaction score 58 between the corresponding compound and the target polymer 38;

[0075] a first neural network 72 comprising a plurality of parameters, wherein each respective output of the first neural network provides an initial embedding 74 corresponding to a set of atomic coordinates 49;

[0076] An attention mechanism 77, which is jointly applied in a cascade to the initial embedding 74 of each pose 48 of the corresponding compound (a specific training or test compound) to derive an attention embedding 79;

[0077] a pooling function 81 having a plurality of parameters 83, wherein the pooling function is applied to the attention embedding 79 to derive a pooled embedding 85 having a plurality of embedding elements 87;

[0078] a first model 89 having a plurality of parameters 91, which is applied to the pooled embedding 85 to (i) obtain a first interaction score for the interaction between the corresponding compound (a particular training or test compound) and the target polymer and / or (ii) condition any other single model and / or group of models;

[0079] an optional second model 93 comprising a plurality of parameters 95, wherein the output of the second model is used to (i) provide a second interaction score for the interaction between the corresponding compound (specific training or test compound) and the target polymer and / or (ii) adjust any other single model and / or group of models;

[0080] an optional third model 97 comprising a plurality of parameters 99, wherein the output of the third model is used to (i) provide a third interaction score for the interaction between the corresponding compound (a particular training or test compound) and the target polymer and / or (ii) adjust any other single model and / or group of models; and

[0081] Optionally, any number of additional x-th models, each such additional x-th model comprising a corresponding plurality of parameters, wherein the output of the additional x-th model is used, at least in part, to (i) provide a characterization of the interaction between the corresponding compound (a particular training or test compound) and the target polymer and / or (ii) adjust any other single model and / or group of models.

[0082] In some implementations, one or more of the above-identified data elements or modules of the computer system 100 are stored in one or more of the previously mentioned memory devices and correspond to a set of instructions for performing the above-mentioned functions. The above-identified data, modules or programs (e.g., instruction sets) need not be implemented as separate software programs, processes or modules, and thus various subsets of these modules may be combined or otherwise rearranged in various implementations. In some implementations, the memory 92 and / or 90 (and optionally 52) optionally stores a subset of the above-identified modules and data structures. Moreover, in some embodiments, the memory 92 and / or 90 (optionally 52) stores additional modules and data structures not described above. In some embodiments, the first neural network 72 is replaced with another form of model.

[0083] Now that a system for characterizing the interaction between a test compound and a polymer of interest has been disclosed, methods for performing such characterization are described in detail with reference to FIG. 2 and discussed below.

[0084] Box 200. See Figure 2A Frame 200, a computer system 100 for characterizing the interaction between a test compound and a target polymer is provided. As discussed above in conjunction with FIG. 1, the computer system 100 includes one or more processors 59 and a memory 90 / 92 addressable by the one or more processors. The memory stores at least one program for being executed by the one or more processors. The at least one program includes instructions described in detail below.

[0085] Boxes 202 to 218. See Figure 2A In block 202, a plurality of sets of atomic coordinates are obtained. Each corresponding set of atomic coordinates 49 in the plurality of sets of atomic coordinates includes the test compound bound to the target polymer in a corresponding pose 48 in the plurality of poses. In other words, see Figure 1A, each corresponding set of atomic coordinates 49 includes the atomic coordinates of the test compound and at least a subset of the spatial coordinates of the target polymer considered by the first neural network. For example, in some embodiments, the set of atomic coordinates 49 consists of the atomic coordinates of the test compound and the atomic coordinates of the portion of the target polymer that constitutes the active site to which the test compound has docked. In some embodiments, the target polymer includes multiple active sites and the test compound has docked to one of the active sites. Figure 3 The posture 48 of the test compound in the active site of the target polymer 38 is shown. Each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates includes the atomic coordinates of at least 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, 100, 150, 200, 250 or 300 atoms. In some embodiments, in addition to the atomic coordinates of the test compound, a set of atomic coordinates in the plurality of sets of atomic coordinates includes the atomic coordinates of at least 400 atoms of the target polymer. In some embodiments, a set of atomic coordinates in the plurality of sets of atomic coordinates includes, in addition to the atomic coordinates of the test compound, the atomic coordinates of at least 25 atoms, at least 50 atoms, at least 100 atoms, at least 200 atoms, at least 300 atoms, at least 400 atoms, at least 1000 atoms, at least 2000 atoms or at least 5000 atoms of the target polymer. In some embodiments, in addition to the coordinates corresponding to the test compound, in each corresponding set of the plurality of sets of atomic coordinates, there are only coordinates of active sites of the target polymer 38 where the ligand is expected to bind to the target polymer.

[0086] See also Figure 2A In block 204, in some embodiments, the target polymer is an assembly of a protein, a polypeptide, a polynucleic acid, a polyribonucleic acid, a polysaccharide, or any combination thereof.

[0087] In some embodiments, the target polymer 38 is a macromolecule composed of repeating residues. In some embodiments, the target polymer 38 is a natural material. In some embodiments, the target polymer 38 is a synthetic material. In some embodiments, the target polymer 38 is an elastomer, shellac, amber, natural or synthetic rubber, cellulose, bakelite, nylon, polystyrene, polyethylene, polypropylene, polyacrylonitrile, polyethylene glycol, or a polysaccharide.

[0088] In some embodiments, the target polymer 38 is a heteropolymer (copolymer). A copolymer is a polymer derived from two (or more) monomer species, as opposed to a homopolymer using only one monomer. Copolymerization refers to a method for chemically synthesizing a copolymer. Examples of copolymers include, but are not limited to, ABS plastics, SBR, nitrile rubber, styrene-acrylonitrile, styrene-isoprene-styrene (SIS), and ethylene-vinyl acetate. Since copolymers include at least two types of constituent units (also referred to as structural units or particles), copolymers can be classified based on how these units are arranged along the chain. These include alternating copolymers with regular alternating A and B units. See, for example, Jenkins, 1996, "Glossary of Basic Terms in Polymer Science", Pure Appl. Chem. 68 (12): 2287–2311, which is hereby incorporated by reference in its entirety. Another example of a copolymer is a periodic copolymer with A and B units arranged in a repeating sequence (e.g., (ABABBAAAABBB) n ). Another example of a copolymer is a statistical copolymer, in which the sequence of monomer residues in the copolymer follows statistical rules. See, for example, Painter, 1997, Fundamentals of Polymer Science, CRC Press, 1997, page 14, which is hereby incorporated by reference in its entirety. Still other examples of copolymers that can be evaluated using the disclosed system and method are block copolymers, which include two or more homopolymer subunits connected by covalent bonds. The combination of homopolymer subunits may require a non-repeating subunit in the middle called a connecting block. Block copolymers with two or three different blocks are respectively referred to as diblock copolymers and triblock copolymers.

[0089] In some embodiments, the target polymer 38 includes 50 or more, 100 or more, 150 or more, 200 or more, 300 or more, 400 or more, 500 or more, 600 or more, 700 or more, 800 or more, 900 or more, or 1000 or more atoms.

[0090] In some embodiments, the target polymer 38 is actually a plurality of polymers (e.g., 2 or more, 3 or more, 10 or more, 100 or more, 1000 or more, or 5000 or more polymers), wherein the corresponding polymers in the plurality of polymers do not all have the same molecular weight. In some such embodiments, the target polymer 38 in the plurality of polymers shares at least 50%, at least 60%, at least 70%, at least 80%, or at least 90% sequence identity and falls within a weight range having a corresponding chain length distribution. In some embodiments, the target polymer 38 is a branched polymer molecule comprising a main chain having one or more substituent side chains or branches. Types of branched polymers include, but are not limited to, star polymers, comb polymers, brush polymers, dendritic polymers, ladders, and dendrimers. See, for example, Rubinstein et al., 2003, Polymer physics, Oxford; New York: Oxford University Press. Page 6, which is hereby incorporated herein by reference in its entirety.

[0091] In some embodiments, the target polymer is a polypeptide. As used herein, the term "polypeptide" means two or more amino acids or residues connected by a peptide bond. The terms "polypeptide" and "protein" are used interchangeably herein and include oligopeptides and peptides. "Amino acid", "residue" or "peptide" refers to any of the twenty standard structural units of proteins known in the art, including imino acids, such as proline and hydroxyproline. The names of amino acid isomers may include D, L, R and S. The definition of amino acids includes non-natural amino acids. Therefore, selenocysteine, pyrrolysine, lanthionine, 2-aminoisobutyric acid, γ-aminobutyric acid, dehydroalanine, ornithine, citrulline and homocysteine, as non-limiting examples, are all considered amino acids. Other variants or analogs of amino acids are known in the art. Therefore, a polypeptide may include a synthetic peptidomimetic structure, such as a peptoid. See Simon et al., 1992, Proceedings of the National Academy of Sciences USA, 89, 9367, which is hereby incorporated by reference in its entirety. See also Chin et al., 2003, Science 301, 964; and Chin et al., 2003, Chemistry & Biology 10, 511, each of which is incorporated herein by reference in its entirety.

[0092] The target polymer 38 evaluated according to some embodiments of the disclosed system and method can also have any number of post-translational modifications. Therefore, the target polymer 38 includes those polymers modified by acylation, alkylation, amidation, biotinylation, formylation, γ-carboxylation, glutamylation, glycosylation, glycylation, hydroxylation, iodination, isoprenylation, fatty acylation, cofactor addition (e.g., heme, flavin, metal, etc.), addition of nucleosides and derivatives thereof, oxidation, reduction, pegylation, phosphatidylinositol addition, phosphopantetheinization, phosphorylation, pyroglutamate formation, racemization, addition of amino acids (e.g., arginylation) by tRNA, sulfation, selenization, ISGization, SUMOylation, ubiquitination, chemical modification (e.g., citrullination and deamidation) and treatment with other enzymes (e.g., proteases, phosphatases and kinases). Other types of post-translational modifications are known in the art and within the scope of the target polymer 38 disclosed herein.

[0093] In some embodiments, the target polymer 38 is a surfactant. Surfactant is a compound that reduces the surface tension of a liquid, the interfacial tension between two liquids or between a liquid and a solid. Surfactants can act as detergents, wetting agents, emulsifiers, foaming agents and dispersants. Surfactants are generally amphiphilic organic compounds, which means that they contain hydrophobic groups (their tails) and hydrophilic groups (their heads) at the same time. Therefore, surfactant molecules contain water-insoluble (or oil-soluble) components and water-soluble components. In the case of water and oil mixing, surfactant molecules will diffuse in water and adsorb at the interface between air and water or at the interface between oil and water. Insoluble hydrophobic groups can extend out of the bulk water phase, into the air or into the oil phase, while the water-soluble head groups remain in the water phase. This arrangement of surfactant molecules at the surface changes the surface properties of water at the water / air or water / oil interface.

[0094] Examples of ionic surfactants include ionic surfactants such as anionic, cationic or zwitterionic (amphoteric) surfactants. In some embodiments, target object 58 is a reverse micelle or a liposome.

[0095] In some embodiments, the target polymer 38 is a fullerene. A fullerene is any molecule composed entirely of carbon in the form of a hollow sphere, ellipsoid, or tube. Spherical fullerenes are also called buckyballs, and they resemble the ball used in soccer. Cylindrical fullerenes are called carbon nanotubes or buckytubes. Fullerenes are structurally similar to graphite, which consists of stacked graphene sheets of connected hexagonal rings; but they can also contain pentagonal (or sometimes heptagonal) rings.

[0096] See also Figure 2AIn block 206, in some embodiments, each set of atomic coordinates 49 in the plurality of sets of atomic coordinates includes three-dimensional coordinates {x1, . . . , x N} (e.g., by X-ray crystallography), the crystal structure of the target polymer is or better yet, or better yet, or better yet, or better yet, or better yet, or better yet, or better yet, or better yet, or better yet, or better yet, or better, or In certain embodiments, the polymer part from the target polymer crystal structure is composed of the atomic coordinates of less than 25,30,35,40,45,50,55,60,65,70,75,80,85,90,95,100,150,200,250 or 300 atoms. In certain embodiments, the polymer part from the target polymer crystal structure is also composed of the atomic coordinates of less than 400 atoms of the target polymer except the atomic coordinates of the test compound. In certain embodiments, the polymer part from the target polymer crystal structure is composed of the atomic coordinates of less than 25 atoms, less than 50 atoms, less than 100 atoms, less than 200 atoms, less than 300 atoms, less than 400 atoms, less than 1000 atoms, less than 2000 atoms or less than 5000 atoms of the target polymer.

[0097] In some embodiments, each set of atomic coordinates 49 in the plurality of sets of atomic coordinates comprises three-dimensional coordinates {x1, ..., x2, ..., x3, ..., x4, ..., x5, ..., x6, ..., x7, ..., x8, ..., x9, ..., ..., ...9, ..., ...10, ..., ...11, ..., ...12, ..., ...13, ..., ...14, ...,15, ...,16, ...,17, ...,1 N Jumper et al., 2021, “Highly accurate protein structure prediction with AlphaFold”, Nature 596, pp. 583-589, which is hereby incorporated by reference.

[0098] See also Figure 2AIn block 208, in some embodiments, each set of atomic coordinates 49 in the plurality of sets of atomic coordinates comprises an ensemble of three-dimensional coordinates of at least a portion of a target polymer determined by nuclear magnetic resonance, neutron diffraction, or cryo-electron microscopy. In some embodiments, the portion of the polymer consists of atomic coordinates of less than 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 95, 100, 150, 200, 250, or 300 atoms. In some embodiments, in addition to the atomic coordinates of the test compound, the portion of the polymer consists of atomic coordinates of less than 400 atoms of the target polymer. In some embodiments, a portion of the polymer consists of atomic coordinates of less than 25 atoms, less than 50 atoms, less than 100 atoms, less than 200 atoms, less than 300 atoms, less than 400 atoms, less than 1000 atoms, less than 2000 atoms, or less than 5000 atoms of the target polymer. In some embodiments, the three-dimensional coordinate ensemble includes ten or more, twenty or more, or thirty or more atomic structures of at least a portion of the target polymer having or better yet, or better yet, or better yet, or better yet, or better yet, or better yet, or better yet, or better or Or better yet backbone RMSD.

[0099] In some embodiments, the target polymer 38 comprises two different types of polymers, such as a nucleic acid bound to a polypeptide. In some embodiments, the natural target polymer comprises two polypeptides bound to each other. In some embodiments, the natural target polymer under study comprises one or more metal ions (e.g., a metalloproteinase with one or more zinc atoms). In such cases, metal ions and or small organic molecules may be included in the atomic coordinates 40 of the target polymer.

[0100] In some embodiments, the target polymer 38 is a polymer and there are ten or more, twenty or more, thirty or more, fifty or more, one hundred or more, between one hundred and one thousand, or less than 500 residues in the polymer.

[0101] In some embodiments, the atomic coordinates of the target polymer 38 are determined using modeling methods, such as ab initio methods, density functional methods, semi-empirical and empirical methods, molecular mechanics, chemical kinetics, or molecular dynamics.

[0102] In some embodiments, each corresponding set of atomic coordinates 49 is represented by Cartesian coordinates including the centers of atoms of the target polymer 38. In some alternative embodiments, each corresponding set of atomic coordinates 49 is represented by the electron density of the target polymer, such as measured by X-ray crystallography. For example, in some embodiments, the atomic coordinates 40 include 2F calculated using the calculated atomic coordinates of the target polymer 38. 观察到的 -F 计算的 Electron density map, where F 观察到的 is the observed structure factor amplitude of the target polymer, and F 计算的 is the structure factor amplitude calculated from the calculated atomic coordinates of the target polymer 38.

[0103] In various other embodiments, each corresponding set of atomic coordinates 49 is obtained from a variety of sources, including but not limited to a collection of structures generated by solution NMR, co-complexes interpreted from X-ray crystallography, neutron diffraction, cryo-electron microscopy, sampling from computer simulations, homology modeling, rotational isomer library sampling, or any combination thereof.

[0104] See also Figure 2A In block 210, in some embodiments, the test compound satisfies two or more, three or more, or all four of the following Lipinski five-fold rule: (i) no more than five hydrogen bond donors, (ii) no more than ten 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 hereby incorporated by reference in its entirety. In some embodiments, in addition to the Lipinski five-fold rule, the test compound satisfies one or more criteria. 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.

[0105] See also Figure 2A In block 214, 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. However, some embodiments of the disclosed systems and methods do not limit the size of the test compound. For example, in some embodiments, the test compound itself is a large polymer, such as an antibody.

[0106] See also Figure 2AAt block 216, in some embodiments, the test compound is an organic compound having a molecular weight between 400 Daltons and 10,000 Daltons.

[0107] See also Figure 2A In the frame 218 of, in some embodiments, multiple groups of atomic coordinates are composed of 3 to 64 postures. In some embodiments, the target polymer 38 is a polymer with an active site, and each posture is obtained by docking the test compound to the active site of the target polymer. In some embodiments, the test compound is docked to the target polymer 38 multiple times to form multiple postures. In some embodiments, the test compound is docked to the target polymer 38 twice, three times, four times, five times or more times, ten times or more times, fifty times or more times, 100 times or more times, or 1000 times or more times. Each such docking represents a different posture of the test compound docked to the target polymer 38. In some embodiments, the target polymer 38 is a polymer with an active site, and the test compound is docked to the active site in each of a variety of different ways, and each such way represents a different posture. In some embodiments, the target polymer includes a plurality of active sites, and the test compound is docked in one of the active sites in each of a variety of different ways, and each way represents a different posture. In some such embodiments, the system and method of the present disclosure are also used to separately study one or more other active sites of the target polymer.

[0108] In some embodiments, each pose of the test compound is determined by AutoDock Vina. See, Trott and Olson, "AutoDock Vina: improving the speed and accuracy of docking with a newscoring function, efficient optimization and multithreading", Journal of Computational Chemistry 31 (2010) 455-461. In some embodiments, one docking program is used to determine some poses of the test compound, and another docking program is used to determine other poses of the test compound. In some embodiments, Quick Vina 2 (Alhossary et al., 2015, "Fast, accurate, and reliable molecular docking with Quick Vina", Bioinformatics 31:13, pp. 2214-2216), VinaLC (Zhang et al., 2013, "Message Passing Interface and Multithreading Hybrid for Parallel Molecular Docking of Large Databases on Petascale High Performance Computing Machines", J. Comput. Chem. DOI: 10.1002 / jcc.23214), Smina (Koes et al., 2013, "Lessons learned inempirical scoring with smina from the CSAR 2011 benchmarking exercise", Journal of chemical information and modeling 53:8, pp. 1893-1904), or CUina (Morrison et al., "Efficient GPU Implementation of AutoDock Vina", COMP poster 3432389), each of which is incorporated herein by reference in its entirety, was used to determine the pose of the test compound.

[0109] In some embodiments, the plurality of sets of atomic coordinates are ensembles from ensemble docking algorithms, such as Stafford et al., 2022, "AtomNet PoseRanker: Enriching Ligand Pose Quality for Dynamic Proteins in Virtual High-Throughput Screens", Journal of Chemical Information and Modeling 62, pp. 1178-1189, which is hereby incorporated by reference in its entirety. In some such embodiments, the ensemble consists of between 3 and 64, between 4 and 128, between 5 and 32, more than 5, or between 8 and 25 structurally similar poses.

[0110] In some embodiments, each pose of a docked test compound is scored against several different conformations of the target protein (e.g., between 2 and 100). In some embodiments, each pose (e.g., in an ensemble of poses) is scored against a fixed conformation of the target protein.

[0111] In some embodiments, the test compound is docked to the target polymer 38 by random pose generation technology or by biased pose generation. In some embodiments, the test compound is docked to the target polymer 38 by Markov Monte Carlo sampling. In some embodiments, such sampling allows the test compound to be fully flexible in docking calculation and scoring function, and the scoring function is the sum of the interaction energy between the test compound and the target polymer 38 and the conformational energy of the test compound. See, for example, Liu and Wang, 1999, "MCDOCK: A Monte Carlo simulation approach to the molecular docking problem ", Journal of Computer-Aided Molecular Design 13, pp. 435-451, which is hereby incorporated by reference as a whole. In some such embodiments, the pose represented by multiple sets of atomic coordinates is the pose that receives the highest score relative to the poses of all other tests (e.g., the highest 256 points therefore have 256 poses, the highest 128 points therefore have 128 poses, the highest 64 points therefore have 64 poses, the highest 32 points therefore have 32 poses, etc.).

[0112] In some embodiments, algorithms such as DOCK (Shoichet, Bodian and Kuntz, 1992, "Molecular docking using shape descriptors", Journal of Computational Chemistry 13 (3), pp. 380-397; and Knegtel et al., 1997 "Molecular docking to ensembles of protein structures", Journal of Molecular Biology 266, pp. 424-440, each of which is hereby incorporated by reference in its entirety) are used to find multiple poses of the test compound relative to the target polymer 38. This algorithm models the target polymer 38 and the test compound as rigid bodies. The docked conformation is searched using surface complementarity to find pose 48.

[0113] In some embodiments, an algorithm such as AutoDOCK (Morris et al., 2009, "AutoDock4 and AutoDock Tools 4: Automated Docking with Selective Receptor Flexibility", J. Comput. Chem. 30(16), pp. 2785-2791; Sotriffer et al., 2000, "Automated docking of ligands to antibodies: methods and applications", Methods: A Companion to Methods in Enzymology 20, pp. 280-291; and Morris et al., 1998, "Automated Docking Using a Lamarckian Genetic Algorithm and Empirical Binding Free Energy Function", Journal of Computational Chemistry 19: pp. 1639-1662, each of which is hereby incorporated by reference) is used to find multiple poses of the test compound relative to the target polymer 38. AutoDOCK uses a kinematic model of the ligand and supports Monte Carlo, simulated annealing, Lamarckian genetic algorithm, and genetic algorithm. Thus, in some embodiments, for a test compound, a docking score function is used to obtain multiple different poses by Markov chain Monte Carlo sampling, simulated annealing, Lamarckian genetic algorithm, or genetic algorithm.

[0114] In some embodiments, an algorithm such as FlexX (Rarey et al., 1996, "A Fast Flexible Docking Method Using an Incremental Construction Algorithm", Journal of Molecular Biology 261, pp. 470-489, which is hereby incorporated by reference) is used to find multiple poses of the test compound relative to the target polymer. FlexX uses a greedy algorithm to perform incremental construction of the test compound at the active site of the target polymer 38. Therefore, in some embodiments, multiple different poses of the test compound are obtained by the greedy algorithm.

[0115] In some embodiments, algorithms such as GOLD (Jones et al., 1997, "Development and Validation of a Genetic Algorithm for flexible Docking", Journal Molecular Biology 267, pp. 727-748, which is hereby incorporated by reference) are used to find multiple poses of the test compound relative to the target polymer 38. GOLD stands for genetic optimization of ligand docking. GOLD establishes a genetically optimized hydrogen bond network between the test compound and the target polymer 38.

[0116] In some embodiments, molecular dynamics is performed on the target polymer (or a part thereof, such as the active site of the target polymer) and the test compound to identify multiple postures. During the molecular dynamics run, the atoms of the target polymer and the test compound are interacted in a fixed time period, so as to give a view of the dynamic evolution of the system. The trajectory of the atoms in the target polymer and the test compound is determined by numerically solving the Newtonian equation of motion of the interacting particle system, wherein the force between the particles and its potential energy are calculated using the interatomic interaction potential or the molecular mechanics force field. See Alder and Wainwright, 1959, "Studies in Molecular Dynamics. I. General Method", J. Chem. Phys. 31 (2): 459; and Bibcode, 1959, J. Ch. Ph. 31, 459A, doi: 10.1063 / 1.1730376, each document in the document is hereby incorporated by reference. Therefore, in this way, the molecular dynamics run generates the trajectory of the target polymer and the corresponding test compound over time. This trajectory includes the trajectory of atoms in the target polymer and the test compound. In some embodiments, a subset of the plurality of different poses is obtained by taking snapshots of this trajectory over a period of time. In some embodiments, poses are obtained from snapshots of several different trajectories, each of which includes a different molecular dynamics run of the target polymer interacting with the test compound. In some embodiments, prior to the molecular dynamics run, the test compound is first docked into an active site of the target polymer using a docking technique.

[0117] Boxes 220 to 238. See Figure 2BIn block 220, for each corresponding set of atomic coordinates 49 in the plurality of sets of atomic coordinates, the corresponding set of atomic coordinates or the encoding of the corresponding set of atomic coordinates is input into the first neural network 72 to obtain a corresponding initial embedding 74 as an output of the first neural network, thereby obtaining a plurality of initial embeddings 74-1, ..., 74-N. Each initial embedding 74 in the plurality of initial embeddings 74-1, ..., 74-N corresponds to a set of atomic coordinates 49 in the plurality of sets of atomic coordinates 49-1-1, ..., 49-1-N. In some embodiments, the first neural network 72 includes more than 400 parameters. In some embodiments, the corresponding set of atomic coordinates or the encoding of the corresponding set of atomic coordinates input into the first neural network 72 consists of information between 20 bits and 20,000 bits. In some embodiments, the corresponding set of atomic coordinates or the encoding of the corresponding set of atomic coordinates input to the first neural network 72 comprises 20, 40, 60, 80, 100, 200, 300, 400, 500, 600, 700, 800, 900, or 1000 bits of information. In some embodiments, the corresponding set of atomic coordinates or the encoding of the corresponding set of atomic coordinates input to the first neural network 72 comprises 2000, 4000, 6000, 8000, or 10,000 bits of information. In some embodiments, the first neural network 72 comprises 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 more than 1 x 10 6 In some embodiments, the amount of information in the corresponding set of atomic coordinates input into the first neural network combined with the number of parameters of the neural network results in more than 10,000 calculations, more than 100,000 calculations, more than 1 x 10 6 Calculations, more than 5x 10 6 Calculate, or more than 1x 10 7 The performance of the calculation is to calculate the initial embedding 74 using the first neural network 74.

[0118] See also Figure 2B In block 222, in some embodiments, the first neural network 72 is a convolutional neural network. Figure 2B In block 224, in some such embodiments, each respective set of atomic coordinates 49 in the plurality of sets of atomic coordinates includes atomic coordinates of at least 80 atoms (e.g., atoms of the target compound in addition to atoms of the target protein to be considered by the first neural network).

[0119] In some embodiments, the corresponding set of atomic coordinates 49 is converted into a corresponding voxel map 52. Thus, the corresponding voxel map 52 represents the test compound of the target polymer 38 relative to the corresponding pose 48. In some such embodiments, the voxel map 52 is unfolded into a corresponding vector 54 and input into the first neural network 72. In some such embodiments, the first neural network 72 is a convolutional neural network, which in turn provides a corresponding initial embedding 74 of the corresponding pose 48 as an output. In some such embodiments, the corresponding vector 54 mentioned above is a one-dimensional vector. In some embodiments, the corresponding vector 54 includes 10 or more elements, 20 or more elements, 100 or more elements, 500 or more elements, 1000 or more elements, or 10,000 or more elements. In some such embodiments, each such element is represented by a different bit in the data structure input to the first neural network.

[0120] In some embodiments, the first neural network 72 is any of the convolutional neural networks disclosed in Wallach et al., 2015, "AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery", arXiv: 1510.02855v1, or U.S. Patent Nos. 11,080,570; 10,546,237; 10,482,355; 10,002,312; or 9,373,059, each of which is hereby incorporated by reference. More details about using a convolutional neural network as the first neural network 72 to obtain a corresponding initial embedding 74 of a corresponding pose 48 of a test compound relative to a target polymer 38 are provided below under the heading "Using a convolutional neural network as the first neural network 72".

[0121] In some embodiments, the first neural network 72 is an equivariant neural network. Non-limiting examples of equivariant convolutional neural networks are disclosed in Thomas et al., 2018, "Tensor field networks: Rotation-and translation-equivariant neural networks for 3D point clouds", arXiv:1802.08219; Anderson et al., 2019, "Cormorant: Covariant Molecular Neural Networks", Neural Information Processing Systems; Johannes et al., 2020, "Directional Message Passing ForMolecular Graphs", International Conference on Learning Representations; Townshend et al., 2021, "ATOM3D: Tasks On Molecules in Three Dimensions", International Conference on Learning Representations; Jing et al., 2009, "Learning from Protein Structure with Geometric Vector Perceptrons", arXiv:2009.01411; and Satorras et al., 2021, "E(n)Equivariant Graph Neural Networks", Networks", arXiv:2102.09844, each of which is hereby incorporated by reference.

[0122] See also Figure 2C In block 228, in some embodiments, the first neural network 72 is a graph neural network. Figure 2C Box 230, in some such embodiments, the graph neural network is characterized by an initial embedding layer and multiple interaction layers, each of the multiple interaction layers contributing an interaction data structure in multiple interaction data results for each atom in a corresponding group of atomic groups of atomic coordinates 49 for a corresponding pose 48 in a plurality of poses, and the plurality of poses are pooled to form a corresponding initial embedding 74 for the corresponding pose 48.

[0123] Fig. 9BEmbodiments of the present disclosure are shown where the first neural network 72 is a graph neural network (GCN). In some embodiments, the GCN employs three-dimensional coordinates of the protein-ligand pose 48, and a single hot atom encoding that simultaneously identifies element type, target protein / compound, and hybridization state. In some embodiments, connectivity is defined purely by radial functions, without using chemical bonds. In some embodiments, a radial cutoff is used at layer l. To define the radial graph, the neighborhood of atom i is defined as where dij = ‖r i -r j ‖ is the pairwise distance between atoms i and j. At each layer l, atoms There are eigenvectors at in is the number of features at layer l, and are the groups of atoms of a protein, a compound (target compound or training compound), or a protein-compound complex, respectively. In some embodiments, the GCN is configured so that P, L, or P+L atoms can be used as either source atoms (i) or target atoms (j) layer by layer. In some embodiments, the graph convolution block is based on continuous filter convolution.

[0124]

[0125] In some embodiments, the convolution kernel W l (r), where

[0126]

[0127] It is constructed using the linear combination of the zero-order spherical Bessel function j0(x) = sin x / x. 0n is the nth zero of j0(x), and are learnable weights, and the number of basis functions is chosen to be

[0128]

[0129] As an example, in some embodiments, However, other values ​​may be used. This parameterization is used to ensure that the convolution kernel vanishes at the edges of the inner neighborhood. In some embodiments, no forces are computed and hence no smoothing cutoff is required here. After the graph convolution step, a Linear-LeakyReLU-Linear layer is included in some embodiments. In some embodiments, no residual connections are applied at the output of the convolution block. Instead, each convolution layer uses a bottleneck layer to take all layers as input. It has been empirically observed that this performs better than skip connections from the previous layer for the same set of source atoms. Fig. 9CIn the illustrated embodiment, the network 72 has five graph convolution blocks. In the first two graph convolution blocks (902 / 904), all ligand and receptor atoms are considered, so And in some embodiments, R c is set to and For the third graph convolution block 906, in some embodiments, the cutoff radius and filter are increased to and Filters, however, in some embodiments, only atoms in a compound (test compound or training compound) are considered as atoms of interest (e.g., But only ). Finally, in the last two layers, 908 and 910, only the polymerized ligand atoms are embedded In some embodiments, the cutoff value and To construct the initial embedding 74 therefrom, readout operations are then applied independently at each layer to the target protein 38 and compound (test compound or training compound) features: These embeddings are then concatenated together to form the initial embedding for pose 74: in In some embodiments, z is applied to z 读取 Dropout 0.2 -LeakyReLU - Two iterative constructions of linear.

[0130] Fig. 20 is a system for characterizing the interaction between a test compound and a target polymer, wherein the characterization is (i) posture ranking and (ii) Vina score and activity. In this embodiment, the shared embedding z 读取 =g(f 0…L ), PoseRanker score Score with Vina By converting z 读取 (initial embedding 74) is passed to a separate multilayer perceptron y 姿态 ( Fig. 20 The first model in 89) and y vina ( Fig. 20 The conditional embedding z′ is then formed by passing z to the conditional mapping:

[0131]

[0132] in And passed to the final MLPp 活动性 ( Fig. 20 The third model in 97) is used to calculate the activity score. Alternatively, z′ can be mapped by the second condition,

[0133]

[0134] to obtain the embedding z″ conditioned on the PoseRanker score and the Vina score.

[0135] Non-limiting additional examples of graph convolutional neural networks are disclosed in 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; Schütt 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., “AtomNet PoseRanker: Enriching Ligand Pose Quality for Dynamic Proteins in Virtual High Throughput Screens”, https: / / chemrxiv.org / engage / chemrxiv / article-details / 614b905e39ef6a1c36268003, each of which is hereby incorporated by reference.

[0136] See also Figure 2CIn block 232, in some embodiments, the first neural network 72 is an equivariant neural network or a message passing neural network. See Bao and Song, 2020, “Equivariant Neural Networks and Equivarification”, arXiv:1906.07172v4, and Gilmer et al., 2020, “Message Passing Neural Networks”, in Schütt et al. (eds.), Machine Learning Meets Quantum Physics, Lecture Notes in Physics 968, Springer, Cham, each of which is hereby incorporated by reference.

[0137] See also Figure 2C In block 234, in some embodiments, the first neural network 72 includes a plurality of graph convolution blocks, and each block uses a plurality of radial graphs to consider connectivity within a corresponding set of atomic coordinates. For example, in some such embodiments, the first neural network 72 includes 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, or 20 or more graph convolution blocks, and each block uses a plurality of radial graphs to consider connectivity within a corresponding set of atomic coordinates.

[0138] See also Figure 2C In block 236, in some embodiments, the first neural network 72 includes 1x 10 6 In some embodiments, the first neural network 72 includes more than 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 10,000, 50,000, 100,000, or 1 x 10 6 parameters.

[0139] See also Figure 2C At block 238 , in some embodiments, the corresponding initial embedding 74 includes a data structure having values ​​between 128 and 768.

[0140] Boxes 240 to 246. See Figure 2CIn block 240, the attention mechanism 77 is applied to the multiple initial embeddings (74-1 to 74-P) in a cascaded form, thereby obtaining an attention embedding 79. In some embodiments, the multiple initial embeddings consist of 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 or more than 20 initial embeddings. In some embodiments, the attention mechanism is a mapping of a query (multiple initial embeddings in a cascaded form) and a set of key-value pairs to an output (attention embedding 79), where the query, key, value, and output are all vectors. In some such embodiments, the output (attention embedding 79) is calculated as a weighted sum of values, where the weight assigned to each value is calculated by a compatibility function of the query (multiple initial embeddings in a cascaded form) with the corresponding key.

[0141] Thus, according to block 240, each initial embedding 74 for each pose of the compound is concatenated together and applied to the attention mechanism. For example, if the compound has five poses 48, resulting in five initial embeddings 74, then the five initial embeddings are concatenated together to form Fig. 9A The z shown 计算 , and this z 计算 Applied to the attention mechanism 77 to obtain the attention embedding 79. The example attention mechanism is described in Chaudhari et al., "An Attentive Survey of Attention Models", arXiv: 1904-02874v3, July 12, 2021, and Vaswani et al., "Attention is All You Need", 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, California, USA, each of which is hereby incorporated by reference herein. The attention mechanism 77 draws such an inference that some parts of the posture 48 are more important than other parts, and therefore some parts (elements or groups of elements) within the initial embedding 74 are more important than other parts. For example, in an instance where each initial embedding consists of twenty elements, elements 1-4 and 9-15 may contain more information about the characterization of the interaction between the compound and the target polymer than elements 5-9 and 16-20. The attention mechanism is trained to discover such observations using training compounds, and then the learned (trained) observations are applied to the initial embedding 74 of the test compound to form an attention embedding. Therefore, the attention mechanism can improve the accuracy of the proposed algorithm by allowing models downstream of the attention mechanism (e.g., the first model 89) to dynamically pay more attention to the input embedding (e.g., z 注意力 、z 池 ), which helps to effectively perform the task at hand (characterize the interaction between the test compound and the target polymer).

[0142] See also Figure 2D In block 244, in some embodiments, the plurality of initial embeddings 74 (e.g., in cascade form) include a first plurality of values ​​and applying the attention mechanism 77 includes (i) inputting the first plurality of values ​​into an attention neural network to obtain a first plurality of weights, wherein each weight in the first plurality of weights corresponds to a corresponding value in the first plurality of values, and (ii) weighting each corresponding value in the first plurality of values ​​by a corresponding weight in the plurality of weights to obtain an attention embedding. Thus, for example, consider a case where the compound has five poses 48, resulting in five initial embeddings 74, wherein each of the five initial embeddings includes 20 values. Thus, the concatenation of the five initial embeddings 74 produces a merged vector having 100 elements ( Fig. 9A The z shown in 计算 ), each element has a value. 计算 The result of inputting to the attention neural network is that the attention neural network returns 100 weights, each weight corresponding to z 计算 Each element of is then used to adjust z 计算 The corresponding values ​​in Fig. 9A z 注意力 Therefore, if z 计算 has 100 elements, then in some embodiments z 注意力 will also have 100 elements. 注意力 The value of element 1 of is (a)z 计算 The value of element 1 of (b) z returned by the attention neural network 计算 The product of the weights of the elements 1, z 注意力 The value of element 2 of is (a)z 计算 The value of element 2 of (b) and z returned by the attention neural network 计算 The product of the weights of 2 elements of , and so on.

[0143] See also Figure 2D In some such embodiments, the first plurality of weights sums to one (or some other constant value), and each weight in the first plurality of weights is a scalar value between zero and one (or some other constant value). Thus, in the example given above, where z 计算has 100 elements, and therefore the attention neural network returns 100 weights, which sum to 1 (or some other constant value) according to block 246, and the values ​​are scalar values ​​between 0 and 1 (or some other constant value). In embodiments having an attention neural network, the attention neural network is jointly trained with at least the first neural network 72 and the first model 89 for known labels (e.g., pKa, activity, binding score) of a plurality of training compounds. In some embodiments, the plurality of training compounds includes 25 or more training compounds, 100 or more training compounds, 500 or more training compounds, 1000 or more training compounds, 10,000 or more training compounds, 100,000 or more training compounds, or 1 x 10 6 Training compounds.

[0144] Boxes 248 to 256. See Figure 2D At block 248, a pooling function 81 is applied to the attention embedding 79 to derive a pooled embedding 85. Examples of pooling functions include, but are not limited to, average, sum, or max pooling. Thus, consider the example above, where the compound has five poses 48, resulting in five initial embeddings 74, where each of the five initial embeddings includes 20 values, and thus z 计算 And further 注意力 has 100 elements. Apply pooling function 81 to z 注意力 To generate z 池 Consider the case where the pooling function is a mean function. In this example, z 池 There will be 20 elements, and z 池 The first element of will be z 注意力 The average value of elements 1, 21, 41, 61, and 81, z 池 The second element of will be z 注意力 The average of elements 2, 22, 42, 62, and 82 of 池 The twentieth element of will be z 注意力 The average of the 20th, 40th, 60th, 80th, and 100th elements of . That is, in other words, and see Figure 2D In some such embodiments, the pooling function 81 combines the attention embeddings 79z by applying a statistical function 注意力 For each corresponding part of 注意力 The attention embedding 79 is collapsed into the pooling embedding 85z 池 , different poses 48 among the plurality of poses are represented to form a pooled embedding 85 .

[0145] Now consider again the above example, where the compound has five poses 48, resulting in five initial embeddings 74, where each of the five initial embeddings includes 20 values, and thus z 计算 And further 注意力 has 100 elements and is applied to z 注意力 To generate z 池 The pooling function 81 of is a summation function. In this example, z 池 will have 20 elements, and z 池 The first element of will be z 注意力 The sum of elements 1, 21, 41, 61, and 81 of z 池 The second element of will be z 注意力 The sum of the elements 2, 22, 42, 62, and 82, ..., and z 池 The twentieth element of will be z 注意力 The sum of the 20th, 40th, 60th, 80th, and 100th elements of . In other words, see Figure 2D In some such embodiments, the pooling function 81 combines the attention embeddings 79z by applying a statistical function 注意力 Each corresponding part of , embeds the attention into 79z 注意力 Folded into a pooled embedding 85, different poses 48 (now weighted by the attention mechanism) among the multiple poses are represented to form a pooled embedding 85.

[0146] See also Figure 2D In some embodiments, the attention embedding 79z 注意力 includes corresponding values ​​of corresponding elements for each corresponding pose 48 in the plurality of poses, and the statistical pooling function is a maximum function across the attention embedding 79z 注意力 The corresponding elements of each corresponding pose 48 represented in are maximized to form a pooled embedding 85. To illustrate, consider again the above example where the compound has five poses 48, resulting in five initial embeddings 74, where each of the five initial embeddings includes 20 values, and thus z 计算 And further 注意力 has 100 elements and is applied to z 注意力 To generate z 池 The pooling function 81 of is a maximum function. In this example, z 池 will have 20 elements, and z 池 The first element of will be z 注意力 The maximum value among elements 1, 21, 41, 61 and 81 of z 池 The second element of will be z 注意力 The maximum value among the 2nd, 22nd, 42nd, 62nd and 82nd elements of池 The twentieth element of will be z 注意力 The maximum value among elements 20, 40, 60, 80, and 100 of .

[0147] See also Figure 2D In block 256, in some embodiments, the attention embedding 79 includes a corresponding plurality of values ​​for a corresponding plurality of elements for each corresponding pose 48 in the plurality of poses, and the statistical pooling function is an average function that averages the corresponding elements for each corresponding pose 48 represented in the attention embedding 79 to form the pooled embedding 85. Such embodiments are similar to the maximum pooling function described above, except that an average function is applied instead of a maximum function.

[0148] Boxes 258 to 280. See Figure 2E In block 258, the pooled embedding 85 is input into a first model 89 to obtain a first interaction score for the interaction between the test compound and the target polymer. In some embodiments, the first model 89 includes more than 400 parameters 91. In some embodiments, the first model 89 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 more than 1x10 6 In some embodiments, the amount of information in the pooled embeddings input into the first model 89 combined with the number of parameters of the first model 89 results in more than 10,000 computations, more than 100,000 computations, more than 1 x 10 6 Calculations, more than 5x10 6 Calculate or exceed 1x 10 7 Calculate the performance to calculate the first interaction score.

[0149] In some embodiments, the system 100 or one or more programs hosted, stored, or addressable by the system 100 is capable of characterizing the interaction between the test compound and the target polymer 38. In some such embodiments, the characterization is a discrete (e.g., discrete binary) activity score. In other words, the characterization is absolute. For example, see Figure 2E In block 262, in some embodiments, the first model 89 is a classification task and the first interaction score classifies the interaction between the test compound and the target polymer. In some such embodiments, the characterization (e.g., the first interaction score) is discrete binary, and the computer system provides a value, such as "1", when the test compound is determined to be active on the target polymer by the computer simulation method disclosed herein, and provides another value, such as "0", when the test compound is determined to be inactive on the target polymer.

[0150] In some embodiments, the characterization (e.g., a first interaction score) is on a non-binary discrete scale. For example, in some embodiments, when the activity of the test compound is determined to be below a first threshold by the computer simulation method disclosed herein, the characterization provides a first value, such as "0", when the activity of the test compound is determined to be between the first threshold and the second threshold, the characterization provides a second value, such as "1", and when the activity of the test compound is determined to be above the second threshold, the characterization provides a third value, such as "2". In such embodiments, the first and second thresholds are predetermined and constant for a specific experiment (e.g., a specific evaluation of a specific database of test compounds relative to a specific target polymer), and are selected to have values ​​that prove useful for identifying suitable test compounds for activity against the test polymer from a database of test compounds. For example, in some embodiments, any of the threshold values ​​disclosed herein are designed to identify 0.1% or less, 0.5% or less, 1% or less, 2% or less, 5% or less, 10% or less, 20% or less, or 50% or less of a database of test compounds as being active against a target polymer, wherein the database of test compounds comprises 100 or more compounds, 1000 or more compounds, 10,000 or more compounds, 100,000 or more compounds, 1x10 6 Compounds, 10x 10 6 compounds or more compounds.

[0151] In an alternative embodiment, the system 100 or one or more programs hosted, stored, or addressable by the system 100 is capable of characterizing the interaction between the test compound and the target polymer 38 as an activity on a continuous scale. That is, the system 100 or one or more programs hosted, stored, or addressable by the system 100 provides a number on a continuous scale that indicates the activity of the test compound against the target polymer. The activity value on the continuous scale is useful, for example, when comparing the activity of each test compound in the test compound database with the target polymer specified by the trained spatial data evaluation module 36. As an example, see Figure 2E At block 260 , in some embodiments, the first model 89 is a regression task and the first interaction score quantifies the interaction between the test compound and the target polymer.

[0152] The disclosed systems and methods are not limited to characterizing the interaction between the test compound and the target polymer 38 as activity on a continuous scale or a discrete scale. In alternative embodiments, the system 100 or one or more programs hosted, stored, or addressable by the system 100 characterizes the interaction between the test compound and the target polymer as an IC of the test compound on the target polymer on a continuous scalar or a discrete (categorical) scalar. 50 ,EC50 , Kd, ​​KI or pKI.

[0153] Although a binary discrete scalar and a discrete scalar with three possible outcomes have been identified, the present disclosure is not limited to these two examples of discrete scalars that characterize the interaction between a test compound and a target polymer 38. In fact, any discrete scalar can be used to characterize the interaction between a test compound and a target polymer 38, including, as non-limiting examples, discrete scalars with 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, or 20 different outcomes.

[0154] See also Figure 2E At block 264, in some embodiments, the first interaction score represents a binding coefficient of the test compound to the target polymer.

[0155] See also Figure 2E In block 266, in some embodiments, the first interaction score is the IC 50 ,EC 50 , Kd, ​​KI or pKI. IC 50 ,EC 50 , Kd, ​​KI, and pKI 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, each of which is hereby incorporated by reference into this document.

[0156] See also Figure 2F In block 270, in some embodiments, the first interaction score is a binary score, wherein a first value of the binary score represents an IC of the test compound relative to the target polymer that is above a first threshold. 50 ,EC 50 , Kd, ​​KI or pKI, and the second value of the binary score represents the IC of the test compound relative to the target polymer below the first threshold 50 ,EC 50 , Kd, ​​KI or pKI.

[0157] See also Figure 2FAt block 272, in some embodiments, the first interaction score represents a computer simulated pose quality score of the test compound to the target polymer.

[0158] See also Figure 2F In box 274, in some embodiments, the first model 89 is a fully connected second neural network, also known as a multilayer perceptron (MLP). In some embodiments, the MLP is a type of feedforward artificial neural network (ANN) comprising at least three layers of nodes: an input layer, a hidden layer, and an output layer. In such embodiments, except for the input node, each node is a neuron using a nonlinear activation function. More disclosure about suitable MLPs used as the first model 89 in some embodiments of the present disclosure can be found in 2020, Multilayer Perceptrons: Theory and Applications, Nova Science Publishers, Hauppauge, New York, edited by Vang-mata, which is hereby incorporated by reference into this document.

[0159] See also Figure 2F In block 276, in some embodiments, the pooled embedding 85 is also input into the second model 93 to obtain a second interaction score for the interaction between the test compound and the target polymer. In such embodiments, the first model 89 is a first fully connected neural network, the second model 93 is a second fully connected neural network, the first interaction score represents a computer simulated pose quality score of the test compound to the target polymer, and the second interaction score represents a computer simulated pose quality score of the test compound to the target polymer.

[0160] Fig.10 A system for characterizing the interaction between a test compound and a target polymer, wherein the characterization is (i) a binary discrete activity and (ii) a pKi, and wherein the system is trained using training compounds according to one embodiment of the present disclosure, whose activity and pKi are known. In the training, a multi-task loss function is calculated and the loss is calculated in all tasks (in Fig.10 binary discrete activity and pKi) and N b Minimize over a mini-batch of training compounds: where w t and l t Represent the weight and loss of task t respectively, is the output of each task t given a training compound i, and y it is the pooled embedding 81 passed to each task t. This means that the pooled embedding y itThis is because it is not required that each task (e.g., first model 89, second model 93, third model 97, etc.) receive the same pooled embedding 85. In some embodiments, although Fig.10 Not shown in , but each task (e.g., first model 89, second model 93, third model 97, etc.) receives a different pooled embedding 85 from a different pooling function 81. Thus, an architecture is envisioned in which the attention embedding 79 is passed to multiple pooling functions to arrive at multiple pooled embeddings 85, and each of the multiple pooled embeddings is passed to a different task t (e.g., first model 89, second model 93, third model 97, etc.). In other embodiments, there is a single pooling function 81 that produces a single pooled embedding that is then passed to one or more tasks (e.g., first model 89, second model 93, third model 97, etc.).

[0161] exist Fig.10 In some embodiments, the pKi model is trained as a regression task using a loss function such as mean squared error for pKi values ​​of training compounds, while the activity model is trained as a classification task using a loss function such as binary cost entropy against known binary discrete activity values ​​of training compounds.

[0162] like Fig.10 As shown, in some embodiments, the pooling embedding 85 is input into a first model 89 (providing a characterization of the interaction between the test compound and the target polymer in the form of a calculated pKi value) and a second model 93 (providing a characterization of the interaction between the test compound and the target polymer in the form of the activity of the test compound relative to the target polymer 38). Fig.10 In the illustrated embodiment, the interaction between the test compound and the target polymer is characterized by both a pkI score (e.g., a discrete binary score or a scalar score) and an activity score (e.g., a classification as "good binder", "bad binder", etc.). Fig.10 pKi is calculated in the embodiment shown, but it should be understood that Fig.10 In other embodiments of the topology, the second model calculates IC 50 ,EC 50、 Kd or KI instead of pKI.

[0163] Fig.11 A system for characterizing an interaction between a test compound and a target polymer, wherein the characterization is pKi, and wherein the pKi is modulated in part by activity, and wherein the system is trained using known pKi and activity of training compounds according to one embodiment of the present disclosure. Fig.11In a system of the invention, the pKi model is conditioned on the activity model. In some embodiments, the pKi model is trained as a regression task using a loss function such as mean squared error against the pKi values ​​of the training compounds, while the activity model is trained as a classification task against the activity values ​​of the training compounds using a loss function such as binary cost entropy.

[0164] like Fig.11 As shown, in some embodiments, the pooling embedding 85 is input to the first model 89 (via edge 1102) and the second model 93 (via edge 1104). In addition, the output of the second model 93 (which is a calculation of the activity of the test compound relative to the target polymer) is input to the first model 89 via edge 1106. In some embodiments, the characterization provided by the second model 93 is an activity score for the test compound. In some embodiments, the activity score is a discrete binary score, for example, where "1" indicates that the test compound is active on the target polymer, and "0" indicates that the test compound is inactive on the target polymer. In some embodiments, the activity score provided by the second model 93 is a scalar. Therefore, the first model 89 receives the output of the second model 93 and the pooling embedding 85. The first model 89 uses these two inputs to determine a characterization of the interaction between the test compound and the target polymer (e.g., in the form of a pKi of the test compound relative to the target polymer 37). The first model 89 adjusts the pKi calculations of both the pooling embedding 85 and the second model 93 to improve the performance of the first model 89 in characterizing the test compound. Although the second model 89 is Fig.11 pKi is calculated in the embodiment shown, but it should be understood that Fig.11 In other embodiments of the topology, the second model calculates IC 50 ,EC 50、 Kd or KI instead of pKI.

[0165] like Fig.12As shown, in some embodiments, the pooled (shared) embedding 85 is input into both the first model 89 (via edge 1202) and the second model 93 (via edge 1204). In addition, the output of the first model 89 (which is a calculation of the pKi of the test compound relative to the target polymer) is input into the second model 93 via edge 1206. Thus, the second model 93 receives the output of the first model 89 and the pooled embedding 85. The second model 93 uses these two inputs to determine a characterization of the interaction between the test compound and the target polymer (e.g., in the form of an activity score test compound). In some embodiments, the activity score is a discrete binary score, for example, where "1" indicates that the test compound is active on the target polymer, and "0" indicates that the test compound is inactive on the target polymer. In some embodiments, the activity score provided by the second model 93 is a scalar. Adjustments to the activity score of the second model 93 of both the pooled embedding 85 and the output of the first model 93 are used to improve the performance of the second model 83 in characterizing the test compound. Although the first model 89 is Fig.12 pKi is calculated in the embodiment shown, but it should be understood that Fig.12 In other embodiments of the topology, the first model 89 calculates IC 50 ,EC 50、 Kd or KI instead of pKI.

[0166] See also Figure 2F In block 278, in some such embodiments, the first interaction score and the second interaction score are input into a third model 97 to obtain a third interaction score. In some such embodiments, the third model 97 is a third fully connected neural network. Figure 2F In some such embodiments, the third interaction score is a discrete binary activity score having a first value when the test compound is determined by the third model 97 to be inactive and having a second value when the test compound is determined by the third model 97 to be active. For example, Fig.13 A system for characterizing an interaction between a test compound and a target polymer is presented, wherein the characterization is activity (via activity model 97), and wherein the activity depends in part on pKi (via pKi model 89) and binding mode score (via PoseRanker model 93), and wherein according to one embodiment of the present disclosure, the pKi model is trained using known pKi values ​​of training compounds, and the PoseRanker model is trained using binding mode scores of training compounds. Fig.13For the purpose of the activity model 97, in some embodiments, if the pKi (or IC50) of each training compound is less than 10 μM, it is marked as active; otherwise it is marked as inactive. To train the PoseRanker model 93, in some embodiments, the binding mode scores of the training compounds are obtained by docking the training compounds with the following docking program, such as CUina [Morrison et al., 2020, “CUina: An Efficient GPU Implementation of AutoDock Vina. August 2020. URL https: / / blog.atomwise.com / efficient-gpu-implementation-of-autodock-vina.] and then ranking each pose of each docked compound using the PoseRanker model [Stafford et al., 2021, “AtomNet PoseRanker: Enriching Ligand Pose Quality for Dynamic Proteins in Virtual High Throughput Screens”, doi:10.33774 / chemrxiv-2021-t6xkj. URL https: / / chemrxiv.org / engage / chemrxiv / article-details / 614b905e39ef6a1c36268003. Therefore, in such embodiments, the combined pattern score used to train the PoseRanker model 93 is the PoseRanker ranking.

[0167] exist Fig.13 In the system, the activity model 97 is conditioned on the pKi model 89 and the PoseRanker model 93. In some embodiments, the pKi model and the PoseRanker model (Stafford et al., 2022, "AtomNet PoseRanker: Enriching Ligand Pose Quality for Dynamic Proteins in Virtual High-Throughput Screens", Journal of Chemical Information and Modeling 62, pp. 1178-1189, which is hereby incorporated by reference herein) are trained as regression tasks using a loss function such as mean squared error, while the activity model is trained as a classification task using a loss function such as binary cost entropy.

[0168] Therefore, if Fig.13 As shown, in some embodiments, the pooled (shared) embedding 85 is input into both the first model 89 (via edge 1302) and the second model 93 (via edge 1304). In addition, the output of the first model 89 (which is the calculation of the pKi of the test compound relative to the target polymer) is input into the third model 97 via edge 1306. In addition, the output of the second model 93 (which is the calculation of the pose quality of the test compound relative to the target polymer represented by the pooled embedding 85) is also input into the third model 97 via edge 1308. Fig.13 In the present disclosure, the second model is called the PoseRanker model. In the present disclosure, the terms "PoseNet" and "PoseRanker" are used interchangeably. The PoleRanker (PoseNet) model is described in more detail in Stafford et al., 2022, "AtomNet PoseRanker: Enriching Lig and Pose Quality for Dynamic Proteins in Virtual High-Throughput Screens", Journal of Chemical Information and Modeling, Vol. 62, pp. 1178-1189, which is hereby incorporated by reference herein. In addition, pooling embedding 85 is input into the third model. Therefore, the third model 97 receives the output of the first model 89, the output of the second model 93, and the pooling embedding 85. The third model 97 uses all these inputs to determine the characterization of the interaction between the test compound and the target polymer (for example, in the form of an activity score test compound). In some embodiments, the activity score is a discrete binary score, for example, wherein "1" indicates that the test compound has activity to the target polymer, and "0" indicates that the test compound has no activity to the target polymer. Although the first model 89 Fig.13 pKi is calculated in the embodiment shown, but it should be understood that Fig.13 In other embodiments of the topology, the first model 89 calculates IC 50 ,EC 50、 Kd or KI instead of pKI.

[0169] Fig.14 A system for characterizing an interaction between a test compound and a target polymer, wherein the characterization is activity and two different compound binding mode scores, and wherein the system is trained using training compounds with known activity scores according to one embodiment of the present disclosure. Fig.14In a system of , the activity model is conditioned on the pose quality score model. In some embodiments, the pose quality model is trained as a regression task using a loss function such as mean squared error, while the activity model is trained as a classification task using a loss function such as binary cost entropy.

[0170] Fig.15 A system for characterizing the interaction between a test compound and a target polymer, wherein the characterization is activity, two different compound binding mode scores, and pKi, and wherein according to one embodiment of the present disclosure, the system is trained using positive and negative poses for coupling of the training compounds. Fig.15 In a system of , the activity model is conditioned on the pose quality score model. In some embodiments, the pose quality model is trained as a regression task using a loss function such as mean squared error, while the activity model is trained as a classification task using a loss function such as binary cost entropy.

[0171] exist Fig.16A In the illustrated embodiment, the pose of the test compound, for example in the form of atomic coordinates 49 or vectors 54, is introduced into the first neural network 72 to ultimately generate a Fig. 9A 85. The pooling embedding 85 is input to the first model 89 (via edge 1630), the second model 93 (via edge 1610) and the third model 97 (via edge 1620). Further, the output of the second model 93 of the test compound (which is the calculation of the interaction score, such as the posture quality score, etc.) is input to the first model 89 via edge 1640. Further, the output of the third model 97 of the test compound (which is the calculation of the interaction score, such as pKi, etc.) is input to the first model 89 via edge 1650. Therefore, the first model receives the output of the third model, the second model and the pooling embedding 84. The first model 89 uses each of these inputs together to determine the characterization of the interaction between the test compound and the target polymer. In some embodiments, the characterization is the activity score of the test compound. In some embodiments, the activity score is a discrete binary score, for example, where "1" indicates that the test compound is active to the target polymer, and "0" indicates that the test compound is inactive to the target polymer. In some embodiments, the activity score provided by the first model 89 is a scalar.

[0172] In some embodiments, see Fig.16A, the pooled embedding 85 is used to predict three outputs: activity (via the first model 89), CUina pose quality score (via the second model 93), and pKi score (via the third model 97). For a description of the fully connected CUina pose quality model, see Morrison et al. "Efficient GPU Implementation of AutoDock Vina", COMP poster 3432389, which is hereby incorporated by reference. Fig.16A In the embodiment shown, it is performed in two stages. First, by transferring the pooling embedding 85 to the second model 93 and the third model 97, CUina and pKi score predictions are calculated. Secondly, conditional embedding 1690 is formed by connecting (i) pooling embedding 85, (ii) the second model 93 score predictions from the first stage, and (iii) the third model 97 score predictions from the first stage. Then the embedding 1690 is transferred to the first model 89, which is in the form of a multilayer perceptron, to calculate the activity prediction of the test compound. In certain embodiments, embedding 1690 is not simply connected (i) pooling embedding 85, (ii) the second model 93 score predictions from the first stage, and (iii) the third model 97 score predictions from the first stage, but represents that these three components are multiplied by each other, or some other mathematical combination of these three components. In such embodiments, the product of three components or some other mathematical combination of three components are input into the third model as embedding 1690. In certain embodiments, embedding 1690 is not connected to three components, but each of the three sources is transformed, and the transformation is used as the input of the first model 89. More generally, embedding 1690 is capable of performing any mathematical function, including but not limited to multiplication, concatenation, linear or non-linear transformations, on all or any portion of any input to embedding 1690 to form a conditional embedding that is passed to the first model 72. Although the third model 97 is Fig.16A In the embodiment shown, pKi is estimated, but it should be understood that in Fig.16A In other embodiments of the topology, the third model 97 estimates the IC of the test compound relative to the target polymer. 50 ,EC 50、 Kd or KI, not pKI.

[0173] See also Fig. 16B , the first model 89 can also be adjusted on the additional model. Fig. 16BIn addition to the pooling embedding 85, the first model 89 is conditioned on the output of the second model 93 that provides the CUina score of the test compound relative to the target polymer, the third model 97 that provides the pKi score of the test compound relative to the target polymer, and the fourth model 990 that provides the PoseRanker score of the test compound relative to the target polymer. Fig. 16B In the embodiment shown, pKi is estimated, but it should be understood that in Fig. 16B In other embodiments of the topology, the third model 97 estimates the IC of the test compound relative to the target polymer. 50 ,EC 50、 Kd or KI, not pKI. Fig. 16B In some such embodiments, the first model, the second model, the third model, and the fourth model 990 are each fully connected neural networks. Such fully connected neural networks are also referred to as multilayer perceptrons (MLPs). In some embodiments, the MLP is a type of feedforward artificial neural network (ANN) comprising at least three layers of nodes: an input layer, a hidden layer, and an output layer. In such embodiments, except for the input nodes, each node is a neuron using a nonlinear activation function. Further disclosure on suitable MLPs for use as the first model 72 in some embodiments may be found in Multilayer Perceptrons: Theory and Applications, Nova Science Publishers, Hauppauge, New York, 2020, edited by Vang-mata, which is hereby incorporated by reference into this document. See Fig. 16B In some embodiments, the corresponding activity score provided by the first model 89 is a binary activity score. For example, in some embodiments, an activity score having a value of "1" means that the test compound is "active" in inhibiting the activity or function of the target polymer (e.g., enzyme activity) and an activity score of "0" means that the test compound does not inhibit the activity or function of the target polymer.

[0174] Representative test compounds and training compounds. Several different architectures or systems have been described for characterizing the interaction between a test compound and a target polymer. Before each such architecture or system can be used to characterize the interaction between a test compound and a target polymer, it needs to be trained for the training compound. A significant difference between a test compound and a training compound is that the training compound is labeled (e.g., using complementary binding data for the target polymer obtained from a wet lab binding assay, etc.), and such labels are used to train the first compound neural network 72, the attention mechanism 77, the first model 89, the second model 93, and the optional third and subsequent models, while each test compound is either unlabeled or does not use a label, and the first neural network 72, the attention mechanism 77, the first model 89, the second model 93, and the optional third and subsequent models of the present disclosure are used to characterize the interaction between each test compound and the target polymer. In other words, the training compound has been characterized by a label (characterization of the interaction between the training compound and the target polymer), and such characterization is used to train the models of the present disclosure, so that they can characterize the interaction between the test compound and the target polymer. The interaction between the test compound and the target polymer is usually not characterized before applying the first neural network 72 and other models of the present disclosure. In a typical embodiment, the characterization of the interaction between the available training compounds and the target polymer is binding data for the target polymer 38 obtained by a wet lab binding assay.

[0175] In some embodiments, a prediction model (such as Fig. 9A ) is trained to receive geometric data input of a compound (e.g., a pose 48 of the compound) and output a representation of the interaction between the compound and the target polymer. For example, in some embodiments, each of the several poses of each of a plurality of training compounds (e.g., 50 or more training compounds, 100 or more training compounds, 1000 or more training compounds, 100,000 or more training compounds) has known binding data for the target polymer, which are sequentially trained by Fig. 9A The model shown is run and provides a single value for each corresponding training compound.

[0176] In some such embodiments, the systems of the present disclosure (e.g., Fig. 9AThe system shown in the figure outputs one of two possible activity categories for a training compound of a given target polymer. For example, a single value provided by the system of the present disclosure for each corresponding training compound is in a first activity category (e.g., a binder) when it is below a predetermined threshold, and in a second activity category (e.g., a non-binder) when its amount is above a predetermined threshold. The activity category specified by the system of the present disclosure is compared with the actual activity category represented by the training compound binding data. In a typical non-limiting embodiment, such training compound binding data is determined from independent network laboratory binding. Then, the errors in the activity category assignment performed by the system of the present disclosure as verified for the binding data (e.g., the first neural network 72, the attention mechanism 77, and / or the first model 89, etc.) are back-propagated through the parameters of each model of the system of the present disclosure to train the system. In an exemplary embodiment, the model of the present disclosure is trained by stochastic gradient descent utilizing the AdaDelta adaptive learning method to combat errors in the activity class assignments made by the model, given the binding data (Zeiler, 2012 “ADADELTA: an adaptive learning rate method”, CoRR, volume abs / 1212.5701, which is hereby incorporated by reference herein), and in Rumelhart et al., 1988, “Neurocomputing: Foundations of research”, chapter Learning Representations by Back-propagating Errors, pages 696-699, Cambridge, MA, USA: MIT Press, which is hereby incorporated by reference herein. In some such embodiments, the two possible activity classes are respectively a binding constant greater than a given threshold amount (e.g., an IC of the training compound relative to the target polymer) and a 50 ,EC 50 or KI greater than 1 nanomolar, 10 nanomolar, 100 nanomolar, 1 micromolar, 10 micromolar, 100 micromolar, or 1 millimolar) and a binding constant less than a given threshold amount (e.g., IC of a training compound relative to a target compound). 50 ,EC 50 or KI is less than 1 nanomolar, 10 nanomolar, 100 nanomolar, 1 micromolar, 10 micromolar, 100 micromolar or 1 millimolar).

[0177] In some embodiments, for each training compound for a given target polymer, the system of the present disclosure outputs one of a plurality of possible activity categories (e.g., three or more activity categories, four or more activity categories, five or more activity categories). For example, when the quantity falls within the first range, the single value provided by the system and method of the present disclosure for each corresponding training compound is in the first activity category, when the quantity falls within the second range, the single value is in the second activity category, when the quantity falls within the third range, the single value is in the third activity category, and so on. The activity categories assigned by the system of the present disclosure are compared with the actual activity categories represented by the training compound combination data of other forms of training data. Then, errors in the activity category assignments made by the system of the present disclosure, as verified against the combination data (or other forms of measured or independently calculated data), are used to train the system of the present disclosure using the above-described techniques. In some embodiments, each corresponding classification in the plurality of classifications is the IC value of the training compound relative to the target polymer. 50 ,EC 50 , pkA or KI range.

[0178] In some embodiments, the classification of multiple training compounds by the system of the present disclosure is compared with the training data using non-parametric techniques (e.g., binding data or other independently measured data for training compounds). For example, the system of the present disclosure is used to rank multiple training compounds relative to a given property (e.g., binding to a given target polymer) and compare the ranking with the ranking provided by the training data obtained by the wet laboratory binding assay of multiple training compounds. This produces the ability to train the system of the present disclosure for errors in the calculated ranking order using the systematic error correction techniques discussed above. In some embodiments, the error (difference) between the ranking of training compounds by the system of the present disclosure and the ranking of training compounds determined by the binding data (or other independently measured data for training compounds) is calculated using the Wilcoxon Mann Whitney function (Wilcoxon signed rank test) or other non-parametric tests, and the error is used to further train the system of the present disclosure (e.g., the first neural network 72, the attention mechanism 77 and / or the first model 89, etc.).

[0179] In some embodiments, model training may involve modifying parameters of one or more component models. Parameters may be further constrained using various forms of regularization, such as L1, L2, weight decay, and dropout.

[0180] In an embodiment, any model disclosed herein may optionally have their parameters (e.g., weights) adjusted (adjusted to minimize the error between the binding affinity and / or classification predicted by the system and the binding affinity and / or classification reported by the training data) with the training data labeled (e.g., with binding data). Various methods may be used to minimize the error function, such as gradient descent, which may include but are not limited to logarithmic loss, sum of squared errors, hinge loss methods. These methods may include second-order methods or approximations such as momentum, Hessian-free estimation methods, Nesterov accelerated gradient methods, Adagrad, etc. Unlabeled generative pre-training and labeled discriminative training can also be combined.

[0181] A convolutional neural network is used as the first neural network 72. In order to characterize the interaction between the test compound and the target polymer, a voxel map 52 is created for the corresponding pose 48 of the compound in some embodiments. In some embodiments, the voxel map 52 is created by: (i) sampling the compound and the target polymer 38 in the pose 48 on the basis of a three-dimensional grid, thereby forming a corresponding three-dimensional uniform space-filling honeycomb, the corresponding three-dimensional uniform space-filling honeycomb comprising a corresponding plurality of space-filling (three-dimensional) polyhedral units; and (ii) for each corresponding three-dimensional polyhedral unit in the corresponding plurality of three-dimensional units, filling the voxel (a discrete set of regularly spaced polyhedral units) in the corresponding voxel map based on the properties (e.g., chemical properties) of the corresponding three-dimensional polyhedral unit. Thus, for a specific pose 48 of a specific compound, a corresponding voxel map 52 is created. Examples of space-filling honeycombs include cubic honeycombs with parallelepiped units, hexagonal honeycombs with hexagonal units, rhombic dodecahedrons with rhombic dodecahedron units, elongated dodecahedrons with elongated dodecahedron units, and truncated octahedrons with truncated octahedron units.

[0182] In some embodiments, the space-filling cells are cubic cells with cubical cells, and the size of such voxels determines their resolution. For example, one may select This means that in such embodiments, each voxel represents a The corresponding cube of the geometric data of the size (for example, the corresponding height, width and depth of the corresponding cell are However, in some embodiments, a finer grid spacing is used (e.g., or even ) or a coarser grid spacing (e.g. ), where the spacing produces an integer number of voxels to cover the input geometry data. In some embodiments, sampling occurs to As an illustration, for a resolution between Resolution For an input cube, such an arrangement would yield 40*40*40=64,000 input voxels.

[0183] In some embodiments, the properties of the atoms generated in sampling (i) are placed in a single voxel in the corresponding voxel map, and each voxel in the plurality of voxels represents the properties of a maximum of one atom. In some embodiments, the properties of the atoms consist of an enumeration of the types of atoms. As an example, some embodiments of the disclosed systems and methods are configured to represent the presence of each atom in a given voxel of the voxel map 52 as a different number of entries, e.g., if carbon is in the voxel, then because the atomic number of carbon is 6, the value 6 is assigned to the voxel. However, such encoding may mean that atoms with close atomic numbers will behave similarly, which may not be particularly useful, depending on the application. Further, element behavior may be more similar within a family (column on the periodic table), and therefore such encoding postures bring additional work to the convolutional neural network 24 decoding.

[0184] In some embodiments, the properties of atoms are encoded in voxels as binary categorical variables. In such embodiments, the atom types are encoded in a so-called "one-hot" encoding: each atom type has a separate channel. Therefore, in such embodiments, each voxel has multiple channels, and at least a subset of the multiple channels represents atom types. For example, one channel within each voxel can represent carbon, while another channel within each voxel can represent oxygen. When a given atom type is found in the three-dimensional grid element corresponding to a given voxel, a first value of the binary categorical variable, such as "1", is assigned to the channel of the atom type within the given voxel, and when the atom type is not found in the three-dimensional grid element corresponding to the given voxel, a second value of the binary categorical variable, such as "0" within the given voxel, is assigned to the channel of the atom type.

[0185] Although there are more than 100 elements, most of them are not encountered in biology. However, even if the most common biological elements (e.g., H, C, N, O, F, P, S, Cl, Br, I, Li, Na, Mg, K, Ca, Mn, Fe, Co, Zn) are represented, each voxel may produce 18 channels or 10,483*18=188,694 inputs to the receptor field. Thus, in some embodiments, each corresponding voxel in the voxel map includes multiple channels, and each of the multiple channels represents a different property that may appear in the three-dimensional space-filling polyhedral unit corresponding to the corresponding voxel. In those embodiments where additional properties of atoms (e.g., partial charge, presence in ligands and protein targets, electronegativity, or SYBYL atom type) are additionally presented as independent channels for each voxel, the number of possible channels for a given voxel is even higher, thereby requiring more input channels to distinguish other equivalent atoms.

[0186] In some embodiments, each voxel has five or more input channels. In some embodiments, each voxel has fifteen or more input channels. In some embodiments, each voxel has twenty or more input channels, twenty-five or more input channels, thirty or more input channels, fifty or more input channels, or one hundred or more input channels. In some embodiments, each voxel has five or more input channels selected from the descriptors found in Table 1 below. For example, in some embodiments, each voxel has five or more channels, each of which is encoded as a binary categorical variable, wherein each such channel represents a SYBYL atom type selected from Table 1 below. For example, in some embodiments, each corresponding voxel in the voxel map includes a channel for C.3 (sp3 carbon) atom type, which means that if the grid in the space of a given test object-target object (or training compound-target object) complex represented by the corresponding voxel covers sp3 carbon, the channel adopts a first value (e.g., "1"), and otherwise a second value (e.g., "0").

[0187] Table 1 – SYBYL atom types

[0188]

[0189]

[0190]

[0191] In some embodiments, each voxel includes ten or more input channels, fifteen or more input channels, or twenty or more input channels selected from the descriptors found above in Table 1. In some embodiments, each voxel contains a channel for a halogen.

[0192] In some embodiments, a structural protein-ligand interaction fingerprint (SPLIF) score is generated for the pose 48 of the corresponding compound. In such embodiments, the SPLIF score is used as an additional input to the underlying first neural network 72, or is encoded separately in the voxel map. For a description of SPLIF, see Da and Kireev, 2014, J. Chem. Inf. Model. 54, pp. 2555-2561, "Structural Protein-Ligand Interaction Fingerprints (SPLIF) for Structure-Based Virtual Screening: Method and Benchmark Study", which is hereby incorporated by reference herein. SPLIF implicitly encodes all possible types of interactions that may occur between the interaction fragments of the compound (test compound or training compound) and the target polymer 38 (e.g., π-π, CH-π, etc.). In the first step, the intermolecular contacts of the compound (test compound or training compound)-target polymer 38 are checked. If the distance between the two atoms is within a specified threshold (e.g., within ), they are considered to be in contact. For each such intermolecular atom pair, the corresponding atoms of the compound and the target polymer are expanded into circular fragments, for example, fragments including the atom in question and its continuous neighborhood up to a certain distance. Each type of circular fragment is assigned an identifier. In some embodiments, such identifiers are encoded in separate channels in the corresponding voxels. In some embodiments, the extended connection fingerprint (ECFP2) up to the first nearest neighbor defined in the Pipeline Pilot software can be used. See, Pipeline Pilot, Version 8.5, Accelrys Software Inc., 2009, which is hereby incorporated by reference. ECFP retains information about all atom / bond types and uses a unique integer identifier to represent a substructure (e.g., a circular fragment). SPLIF fingerprints encode all circular fragment identifiers found. In some embodiments, the SPLIF fingerprint is not a single voxel encoded, but is used as a separate independent input in the neural network discussed below.

[0193] In some embodiments, instead of or in addition to SPLIF, a structural interaction fingerprint (SIFt) is calculated for each pose of a given compound (test compound or training compound) relative to the target polymer and is independently provided as input to the first neural network 72 or encoded in the voxel map 52. For the calculation of SIFt, see Deng et al., 2003, "Structural Interaction Fingerprint (SIFt): A Novel Method for Analyzing Three-Dimensional Protein-Ligand Binding Interactions", J. Med. Chem. 47 (2), pp. 337-344, which is hereby incorporated by reference.

[0194] In some embodiments, instead of or in addition to SPLIF and SIFT, for each pose of a given compound (test compound or training compound) relative to the target polymer 38, an atom pair based interaction fragment (APIF) is calculated and provided independently as input to the first neural network, or encoded separately in the voxel map. For the calculation of APIF, see Perez-Nueno et al., 2009, "APIF: a new interaction fingerprint based on atoms and its application to virtual screening", J. Chem. Inf. Model. 49 (5), pp. 1245-1260, which is hereby incorporated by reference.

[0195] The data representation may be encoded in a manner that enables the expression of various structural relationships associated with, for example, molecules / proteins. According to various embodiments, the geometric representation may be implemented in various manners and topologies. The geometric representation is used for visualization and analysis of data. For example, in an embodiment, geometric shapes may be represented using voxels arranged on various topologies, such as 2-D, 3-D Cartesian / Euclidean space, 3-D non-Euclidean space, manifolds, etc. For example, Figure 4An example three-dimensional grid structure 400 including a series of sub-containers according to an embodiment is shown. Each sub-container 402 can correspond to a voxel. A coordinate system can be defined for the grid so that each sub-container has an identifier. In some embodiments of the disclosed systems and methods, the coordinate system is a Cartesian coordinate system in 3-D space, but in other embodiments of the system, the coordinate system can be any other type of coordinate system, such as an oblate spherical, cylindrical or spherical coordinate system, a polar coordinate system, other coordinate systems designed for various manifolds and vector spaces, etc. In some embodiments, voxels can have specific values ​​associated with them, for example, the values ​​can be represented by applying labels and / or determining their positioning, etc.

[0196] In some embodiments, the first neural network 72 is a convolutional neural network that requires a fixed input size. In some such embodiments of the disclosed systems and methods, the geometric data (e.g., the voxel map 52 and / or the atomic coordinate set 49) is clipped to fit within an appropriate bounding box. For example, a bounding box with a side length of Cube. In some embodiments where the compound (test compound or training compound) has been docked at the active site of the target polymer 38, the center of the active site is used as the center of the cube. Although in some such embodiments, a square cube of fixed size centered on the active site of the target polymer 38 is used to divide the space into a voxel grid, the disclosed system is not limited thereto. In some embodiments, any of a variety of shapes is used to divide the space into a voxel grid. In some embodiments, polyhedra such as rectangular prisms, polyhedron shapes, etc. are used to divide the space. In one embodiment, the grid structure can be configured to be similar to the arrangement of voxels. For example, each substructure can be associated with the channel of each analyzed atom. In addition, a coding method can be provided to represent each atom digitally.

[0197] In some embodiments, the voxel map takes into account time factors (eg, running along the molecular dynamics of the compound docked to the target polymer) and can therefore be four dimensions (X, Y, Z, and time).

[0198] In some embodiments, other embodiments such as pixels, points, polygonal shapes, polyhedrons, or any other type of shape in multiple dimensions (e.g., shapes in 3D, 4D, etc.) may be used instead of voxels.

[0199] In some embodiments, the geometric data is normalized by selecting the origin of the X, Y, and Z coordinates as the center of mass of the binding site of the target polymer 38 determined by the cavity flooding algorithm. For representative details of such algorithms, see Ho and Marshall, 1990, "Cavity search: An algorithm for the isolation and display of cavity-like binding regions", Journal of Computer-Aided Molecular Design 4, pages 337-354; and Hendlich et al., 1997, "Ligsite: automatic and efficient detection of potential small molecule-binding sites in proteins", J. Mol. Graph. Model 15: 6, each of which is hereby incorporated by reference. Alternatively, in some embodiments, the origin of the voxel map is centered on the center of mass of the entire co-complex (compounds docked with corresponding postures relative to the target polymer). In some embodiments, the origin of the voxel map is centered on the center of mass of the compound (test compound or training compound). In some embodiments, the origin of the voxel map is centered on the center of mass of the target polymer 38. The basis vectors may optionally be selected as the principal moments of inertia of the entire co-complex, only the target polymer, or only the compound (test compound or training compound). In some embodiments, the target polymer 38 has active sites, and the compound (test compound or training compound) is sampled by sampling the poses in the active sites of the target polymer based on a three-dimensional grid, wherein the mass of the active site is the origin, and the corresponding sampled three-dimensional uniform cell represents a portion of the polymer and compound (test compound or training compound) centered at the center of mass. In some embodiments, the uniform cell is a regular cubic cell, and the portion of the target polymer and compound (test compound or training compound) is a cube of a predetermined fixed size. In such embodiments, the use of a cube of a predetermined fixed size ensures that the relevant portion of the geometric data is used and that each voxel map is the same size. In some embodiments, the predetermined fixed size of the cube is Wherein N is an integer or real value between 5 and 100, an integer between 8 and 50, or an integer between 15 and 40. In some embodiments, the uniform cell is a rectangular prism cell and a portion of the target polymer and the compound (test compound or training compound) are rectangular prisms, and the predetermined fixed size is Where Q is a first integer between 5 and 100, R is a second integer between 5 and 100, S is a third integer or real value between 5 and 100, and at least one number in the set {Q, R, S} is not equal to another value in the set {Q, R, S}.

[0200] In an embodiment, each voxel has one or more input channels that may have respective values ​​associated therewith, which in a simple implementation may be on / off and may be configured to encode an atom type. The atom type may represent the element of the atom, or the atom type may be further refined to distinguish other atomic properties. The atoms present may then be encoded in each voxel. Various techniques and / or methods may be used to utilize various types of encoding. As an example encoding method, the atomic number of the atom may be utilized, with each voxel producing a value ranging from one for hydrogen to 118 for ununoctium (or any other element).

[0201] However, as discussed above, other encoding methods can be used, such as "one-hot encoding", in which each voxel has many parallel input channels, each channel is either on or off and encodes one type of atom. The atom type can represent the element of the atom, or the atom type can be further refined to distinguish other atomic properties. For example, SYBYL atom types distinguish single-bonded carbon from double-bonded, triple-bonded, or aromatic carbon. For SYBYL atom types, see Clark et al., 1989, "Validation of the General Purpose Tripos Force Field, 1989, J. Comput. Chem. 10, pp. 982-1012, which is hereby incorporated by reference.

[0202] In some embodiments, each voxel further includes one or more channels to distinguish between atoms that are part of the target polymer 38 or cofactor and atoms that are part of the compound (test compound or training compound). For example, in one embodiment, each voxel further includes a first channel for the target polymer 38 and a second channel for the compound (test compound or training compound). When the atoms in the portion of space represented by the voxel are from the target polymer 38, the first channel is set to a value, such as "1", otherwise it is zero (for example, because the portion of space represented by the voxel does not include atoms or includes one or more atoms from the compound). Further, when the atoms in the portion of space represented by the voxel are from the compound (test compound or training compound), the second channel is set to a value, such as "1", otherwise it is zero (for example, because the portion of space represented by the voxel does not include atoms or includes one or more atoms from the target polymer 38). Similarly, other channels can additionally (or alternatively) specify additional information, such as partial charge, polarizability, electronegativity, solvent accessible space, and electron density. For example, in some embodiments, the electron density map of the target polymer covers a three-dimensional coordinate set, and the creation of the voxel map further samples the electron density map. Examples of suitable electron density maps include, but are not limited to, multiple isomorphous substitution maps, single isomorphous substitution with anomalous signal maps, single wavelength anomalous dispersion maps, multi-wavelength anomalous dispersion maps, and 2Fo-Fc maps (260). See McRee, 1993, Practical Protein Crystallography, Academic Press, which is hereby incorporated by reference.

[0203] In some embodiments, voxel encoding according to the disclosed systems and methods may include additional optional encoding refinements. Two examples are provided below.

[0204] In a first encoding refinement, based on the fact that most elements rarely appear in biological systems, the required memory can be reduced by reducing the set of atoms represented by a voxel (e.g., by reducing the number of channels represented by a voxel). Atoms can be mapped to share the same channel in a voxel by combining rare atoms (so they may rarely affect the performance of the system) or by combining atoms with similar properties (so they can minimize the inaccuracy of the combination). In some embodiments, two, three, four, five, six, seven, eight, nine, or ten different atoms share the same channel in a voxel.

[0205] The encoding refinement is done by partially activating neighboring voxels to make the voxels representing the atomic position. This leads to partial activation of neighboring neurons in the subsequent neural network and changes from one-hot encoding to "multi-hot" encoding. For example, to illustrate, consider the chlorine atom, which has The van der Waals diameter, and hence the volume, is When placed When the grid is formed, the voxels inside the chlorine atom will be completely filled, while the voxels on the edge of the atom will only be partially filled. Therefore, the channel representing chlorine in the partially filled voxels will be opened in proportion to the amount of such voxels falling into the chlorine atom. For example, if 50% of the voxel volume falls within the chlorine atom, the channel in the voxel representing chlorine will be activated 50%. Relative to discrete one-hot encoding, this can produce a "smooth" and more accurate representation. Therefore, in some embodiments, the characteristics of the atoms generated in the sampling are distributed across a subset of voxels in the voxel map, and this subset of voxels includes two or more voxels, three or more voxels, five or more voxels, ten or more voxels, or twenty-five or more voxels. In some embodiments, the characteristics of the atoms are composed of an enumeration of atom types (e.g., one of the SYBYL atom types).

[0206] Therefore, voxelization (rasterization) of the encoded geometric data (test or training compounds docked onto the target polymer) is performed based on various rules applied to the input data.

[0207] According to some embodiments, Figure 5 and 6 A view of two compounds 502 encoded onto a two-dimensional grid 500 of voxels is provided. Figure 5 Two compounds superimposed on a two-dimensional grid are provided. Figure 6 A one-hot encoding is provided that uses different shading patterns to encode the presence of oxygen, nitrogen, carbon, and empty space, respectively. As described above, such encoding may be referred to as a "one-hot" encoding. Figure 6 Shows Figure 5 Grid 500 with compound 502 omitted. Figure 7 Provided Figure 6 A view of a 2D voxel grid where the voxels are numbered.

[0208] In some embodiments, feature geometry is represented in a form other than voxels. Figure 8 A view of various representations is provided, where features (e.g., atomic centers) are represented as 0-D points (representation 802), 1-D points (representation 804), 2-D points (representation 806), or 3-D points (representation 808). Initially, the spacing between points can be randomly selected. However, as the predictive model is trained, the points may move closer or farther apart.

[0209] In some embodiments, the input representation of the first neural network 72 may be in the form of a 1D array of features, including but not limited to three-dimensional coordinates.

[0210] Expand the voxel map into corresponding vectors. In some embodiments, each voxel map 52 is optionally expanded into corresponding vectors 54. In some embodiments, each such vector is a one-dimensional vector. For example, in some embodiments, each side The cube is centered on the active site of the target polymer 38, and the compound (target compound or training compound) is docked in one pose and The three-dimensional fixed grid spacing is sampled to form the corresponding voxel of the voxel map, which holds the basis of the voxel structural characteristics in the corresponding channel, such as atom type, and optionally, more complex compound-target polymer descriptors, as described above. In some embodiments, the voxels of this three-dimensional voxel map are unfolded into a one-dimensional floating point vector.

[0211] In some embodiments, the vectorized representation of the voxel map is input into the first neural network 72. In some embodiments, the vectorized representation of the voxel map is stored in GPU memory along with the first neural network 72. This provides the advantage of processing the vectorized representation of the voxel map by the first neural network 72 at a faster rate. However, in other embodiments, any or all of the vectorized representation of the voxel map and the first neural network 72 are located in the memory 92 of the system 100 or are simply addressable across the network by the system 92. In some embodiments, any or all of the vectorized representation of the voxel map and the first neural network 72 are in a cloud computing environment.

[0212] In some embodiments, each vector 54 is provided to a graphics processor memory, wherein the graphics processor memory includes a network architecture, the network architecture including a first neural network 72, the first neural network being in the form of a convolutional neural network including an input layer for sequentially receiving the vector, a plurality of convolutional layers, and an optional scorer. The plurality of convolutional layers include an initial convolutional layer and a final convolutional layer. In some embodiments, the convolutional neural network is not in GPU memory, but in the memory 92 of the system 100. In some embodiments, the voxel map 52 is not vectorized before being input to the first neural network 72.

[0213] In some embodiments where the first neural network 72 is a convolutional neural network, a convolutional layer in the plurality of convolutional layers within the network comprises a set of learnable filters (also referred to as kernels). Each filter has a fixed three-dimensional size that is convolved (stepped at a predetermined step rate) across the depth, height, and width of the input volume of the convolutional layer, computing a dot product (or other function) between the entries (weights, or more generally parameters) of the filter and the input, thereby creating a multi-dimensional activation map of the filter. In some embodiments, the filter step rate is one element, two elements, three elements, four elements, five elements, six elements, seven elements, eight elements, nine elements, ten elements, or more than ten elements of the input space. Thus, considering a filter size of 53 In some embodiments, for a total of 125 input space values ​​per voxel channel, this filter will compute the dot product (or other mathematical function) between consecutive cubes of the input space with a depth of five elements, a width of five elements, and a height of five elements.

[0214] The input space of the initial convolutional layer (e.g., the output from the input layer) is formed by a voxel map 52 or a vectorized representation 54 of the voxel map. In some embodiments, the vectorized representation of the voxel map is a one-dimensional vectorized representation of the voxel map used as the input space of the initial convolutional layer. However, when the filter convolves its input space and the input space is a one-dimensional vectorized representation of the voxel map, the filter still obtains those elements of the corresponding continuous cube representing the fixed space in the target polymer 38-compound complex from the one-dimensional vectorized representation. In some embodiments, the filter uses a bookkeeping technique to select those elements of the corresponding continuous cube that form the fixed space in the target polymer 38-compound complex from the one-dimensional vectorized representation. Therefore, in some cases, this necessarily involves obtaining a non-continuous subset of elements in the one-dimensional vectorized representation in order to obtain the element values ​​of the corresponding continuous cube of the fixed space in the target polymer 38-compound complex.

[0215] In some embodiments, the filter is initialized (e.g., for Gaussian noise) or trained to have 125 corresponding weights (per input channel) in which the dot product (or some other form of mathematical operation, such as Fig.17 ) to compute a first single value (or set of values) for the activation layer corresponding to the filter. In some embodiments, the values ​​computed by the filter are summed, weighted, and / or biased. To compute additional values ​​for the activation layer corresponding to the filter, the filter is then stepped (convolved) in one of the three dimensions of the input volume at a step rate (stride) associated with the filter, at which point a dot product (or some other form of mathematical operation, such as ) is taken between the filter weights and the 125 input spatial values ​​(per channel) at the new location in the input volume. Fig.17). This stepping (convolution) is repeated until the filter samples the entire input space at the step rate. In some embodiments, the boundaries of the input space are zero-padded to control the spatial volume of the output space produced by the convolutional layer. In a typical embodiment, each of the filters of the convolutional layer covers the entire three-dimensional input volume in this way, thereby forming a corresponding activation map. The collection of activation maps of the filters from the convolutional layers together form the three-dimensional output volume of one convolutional layer, and are thereby used as the three-dimensional (three spatial dimensions) input of subsequent convolutional layers. Therefore, each entry in the output volume can also be interpreted as the output of a single neuron (or a collection of neurons), which looks at a small area in the input space of the convolutional layer and shares parameters with neurons in the same activation map. Therefore, in some embodiments, a convolutional layer in a plurality of convolutional layers has multiple filters, and each of the multiple filters converts N 3 Convolution of a cubic input space with stride Y (in three spatial dimensions), where N is an integer of two or greater (e.g., 2, 3, 4, 5, 6, 7, 8, 9, 10, or greater than 10) and Y is a positive integer (e.g., 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, or greater than 10).

[0216] Each of the plurality of convolutional layers is associated with a different set of weights, or more generally, a different set of parameters. More specifically, each of the plurality of convolutional layers includes a plurality of filters, and each filter includes an independent plurality of parameters (e.g., weights). In some embodiments, the convolutional layer has 128 filters of size 5. 3 The convolutional layer has 128 x 5 x 5 x 5 or 16,000 parameters (e.g., weights) for each channel in the voxel map. Thus, if there are five channels in the voxel map, the convolutional layer will have 16,000 x 5 parameters (e.g., weights) or 80,000 parameters (e.g., weights). In some embodiments, some or all such parameters (and optionally biases) for each filter in a given convolutional layer can be combined together, e.g., constrained to be the same.

[0217] In response to the input of the corresponding vector, the input layer feeds the first plurality of values ​​to the initial convolution layer as a first function of the values ​​in the corresponding vector, wherein the first function is optionally calculated using a graphics processor. In some embodiments, the computer system 100 has more than one graphics processor, and each such graphics processor is used simultaneously to facilitate the calculation of the first neural network 72.

[0218] Except for the final convolutional layer, each respective convolutional layer feeds the intermediate value to another convolutional layer in the plurality of convolutional layers as a respective second function of (i) a different set of parameters (e.g., weights) associated with the respective convolutional layer and (ii) an input value received by the respective convolutional layer. In some embodiments, a graphics processor is used to calculate the respective second function. For example, in some embodiments, each respective filter of the respective convolutional layer overlays the input volume (in three spatial dimensions) to the convolutional layer according to a characteristic three-dimensional stride of the convolutional layer, and at each respective filter position, a dot product (or some other mathematical function) of the filter parameters (e.g., weights) of the respective filter and the value of the input volume (a continuous cube as a subset of the total input space) at the respective filter position is calculated, thereby producing a calculated point (or set of points) on the activation layer corresponding to the respective filter position. The activation layer of the filters of the respective convolutional layer collectively represents the intermediate value of the respective convolutional layer.

[0219] In some embodiments, the convolutional neural network has one or more activation layers. In some embodiments, the activation layer is a neuron layer that applies a non-saturated activation function f(x)=max(0,x). This increases the nonlinearity of the decision function and the entire network without affecting the receptive field of the convolutional layer. In other embodiments, the activation layer has other functions that increase nonlinearity, for example, a saturated hyperbolic tangent function f(x)=tanh, f(x)=│tanh(x)│ and a sigmoid function f(x)=(1+e -x ) -1 Non-limiting examples of other activation functions found in other activation layers in some embodiments of the neural network may include, but are not limited to, a logistic (or sigmoid) function, a softmax function, a Gaussian function, a Boltzmann weighted average function, an absolute value function, a linear function, a rectified linear function, a bounded rectified linear function, a soft rectified linear function, a parameterized rectified linear function, an average function, a maximum function, a minimum function, a certain vector norm LP (for p=1, 2, 3, ..., ∞) function, a sign function, a square function, a square root function, a multivariate quadratic function, an inverse quadratic function, an inverse multivariate quadratic function, a polyharmonic spline function, and a thin plate spline function.

[0220] Filters in a network's learned convolutional layers activate when they see a certain type of feature at a certain spatial location in the input. In some embodiments, the initial parameters (e.g., weights) for each filter in a convolutional layer are obtained by training a convolutional neural network on a training library of compounds. Thus, the operation of a convolutional neural network can produce more complex features than those historically used to make binding affinity predictions. For example, a filter used as a hydrogen bond detector in a given convolutional layer in the network can not only identify that a hydrogen bond donor and acceptor are at a given distance and angle, but also that the biochemical environment surrounding the donor and acceptor enhances or weakens the bond. Furthermore, filters within the network can be trained to effectively distinguish between binders and non-binders in the underlying data.

[0221] As described above, in some embodiments, the first neural network 72 is configured to develop three-dimensional convolutional layers. The input region of the lowest level convolutional layer can be a cube (or other continuous region) of voxel channels from the receptive field. Higher convolutional layers evaluate the output of lower convolutional layers while still making their output a function of a bounded region of voxels that are close to each other (in 3-D Euclidean distance). In an embodiment, the first neural network 72 is configured to apply regularization techniques to reduce the tendency of the model to overfit the training data.

[0222] Zero or more network layers in the above-mentioned convolutional neural network can be composed of pooling layers. Like convolutional layers, pooling layers are a collection of function calculations that apply the same function to different spatial local input blocks. For pooling layers, the output is given by a pooling operator, such as a vector norm LP on several voxels (for p = 1, 2, 3, ..., ∞). In some embodiments, pooling is performed by channel, while in other embodiments, pooling is performed across channels. In some embodiments, pooling divides the input space into a set of three-dimensional boxes, and for each such sub-region, the maximum value or some other mathematical pooling operation, such as average pooling, is output. The pooling operation provides a form of translation invariance. The function of the pooling layer is to gradually reduce the spatial size of the representation to reduce the amount of parameters and calculations in the network, and therefore also control overfitting. In some embodiments, pooling layers are inserted between continuous convolutional layers of the above-mentioned convolutional neural network. Such pooling layers operate independently on each depth slice of the input and adjust their size spatially. In addition to or instead of max pooling, the pooling unit may also perform other functions such as average pooling or even L2 norm pooling.

[0223] Zero or more layers in the above convolutional neural network can consist of normalization layers, such as local response normalization or local contrast normalization, which can be applied to the same position across channels, or to a specific channel across multiple positions. These normalization layers can promote diversity in the response of several functions to the same input.

[0224] In some embodiments, the scorer is not present. Instead, the convolutional neural network outputs an initial embedding 74 of the input pose 48 instead of a score.

[0225] Use cases.

[0226] The following are example use cases provided for illustrative purposes only, which describe some applications of some embodiments of the present disclosure. Other uses may be contemplated, and the examples provided below are non-limiting and may be varied, omitted, or may include additional elements.

[0227] Hit Discovery. Pharmaceutical companies spend millions of dollars screening compounds to discover new potential drug leads. To find the few compounds that have any interaction with the disease target of interest, large collections of compounds are tested. Unfortunately, wet lab screening is subject to experimental error, and in addition to the cost and time of performing assay experiments, collecting large screening collections poses a significant challenge due to storage limitations, storage stability, or chemical costs. Even the largest pharmaceutical companies have only a few hundred thousand to a few million compounds, while there are tens of millions of molecules available commercially and hundreds of millions, billions, or even trillions of molecules that can be simulated. Examples of commercially available molecular databases include 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).

[0228] A potentially more efficient alternative to physical experiments is virtual high-throughput screening. In the same way that physical simulations can help aerospace engineers evaluate possible wing designs before physically testing the models, computational screening of molecules can focus experimental testing on a small subset of high-likelihood molecules. This can reduce screening costs and time, reduce false negatives, increase success rates, and / or cover a wider range of chemical space.

[0229] In the present application, a protein target can be provided as an input to the system. A large set of compounds can also be provided. For each compound, the binding affinity for the protein target is predicted. The resulting scores can be used to rank the compounds, with the highest scoring compounds being most likely to bind the target protein. Alternatively, the ranked compounds can be analyzed for clusters of similar compounds; large clusters can be used as stronger predictions of compound binding, or compounds can be selected across clusters to ensure diversity in validation experiments.

[0230] Off-target side effect prediction. Many drugs may be found to have side effects. Often, these side effects are caused by interactions with biological pathways other than the one responsible for the drug’s therapeutic effect. These off-target side effects can be uncomfortable or harmful to the person and limit the patient population in which the drug can be safely used. Off-target side effects are therefore an important criterion for evaluating which drug candidates to develop further. While it is important to characterize the interactions of a drug with many alternative biological targets, the development and running of such tests can be expensive and time-consuming. Computational predictions can make this process more efficient.

[0231] In the application of the embodiments of the present disclosure, a group of protein targets associated with significant biological responses and / or side effects can be constructed. The system can then be configured to predict the binding for each protein target in the group in turn. Strong activity against a specific target (i.e., activity as effective as a compound known to activate off-target proteins) may suggest that the molecule produces side effects due to off-target effects.

[0232] Toxicity prediction. Toxicity prediction is a particularly important special case of off-target side effect prediction. Approximately half of drug candidates are eliminated in late-stage clinical trials due to unacceptable toxicity. As part of the novel drug approval process (and before a candidate can be tested in humans), the FDA requires toxicity testing data for a set of targets including cytochrome P450 liver enzymes (whose inhibition may lead to toxicity from drug-drug interactions) or the hERG channel (whose binding may lead to QT prolongation, resulting in ventricular arrhythmias and other adverse cardiac effects).

[0233] In toxicity prediction, the system can be configured to limit off-target proteins to key anti-targets (e.g., CYP450, hERG, or 5-HT 2B Receptors). The binding affinity of drug candidates against these proteins can then be predicted. Alternatively, compounds can be analyzed to predict the set of metabolites (subsequent molecules produced by the body during metabolism / degradation of the original compound), and the binding of molecules to anti-targets can also be analyzed. Problematic compounds can be identified and modified to avoid toxicity, or development of a series of molecules can be stopped to avoid wasting additional resources.

[0234] Potency optimization. One of the key requirements for drug candidates is strong binding to their disease targets. Screening rarely yields compounds that bind strongly enough to produce a clinical effect. Therefore, initial compounds go through a lengthy optimization process, where medicinal chemists repeatedly modify the compounds’ molecular structures to come up with new compounds with higher target binding strength. Each new compound is synthesized and tested to determine if the changes successfully improved binding. The system can be configured to facilitate this process by replacing physical testing with computational predictions.

[0235] In this application, a disease protein target and a set of lead compounds can be input into the system. The system can be configured to generate binding affinity predictions for a set of leads. Optionally, the system can highlight differences between the candidate compounds that help explain the reasons for the differences in predicted binding affinity. A medicinal chemist user can use this information to propose a new set of compounds that will hopefully have better activity against the target. These new, alternative compounds can be analyzed in the same manner.

[0236] Selectivity optimization. As mentioned above, compounds tend to bind to multiple proteins with varying strengths. For example, the binding sites of protein kinases (common chemotherapy and inflammation targets) are very similar, and most kinase inhibitors affect many different kinases. This means that various biological pathways are modified simultaneously, resulting in "dirty" drug signatures and many side effects. Therefore, the key challenge of many drug designs is not activity per se, but specificity: the ability to selectively target one protein (or group of proteins) from a group of potentially closely related proteins.

[0237] The system can reduce the time and cost of optimizing the selectivity of drug candidates. In this application, a user can input two groups of target proteins. One group describes the target proteins against which the compound should be active, while the other group describes the target proteins against which the compound should not be active. The system can be configured to make predictions for the compounds for all proteins in both groups, thereby establishing interaction strength profiles. Optionally, these profiles can be analyzed to propose explanatory patterns in the target proteins. Users can use the information generated by the system to consider structural modifications to the compounds to improve relative binding to different groups of target proteins and design new candidate compounds with better specificity. Optionally, the system can be configured to highlight the differences between the candidate compounds, which helps understand the reasons for the differences in predicted selectivity. The proposed candidates can be iteratively analyzed to further refine the specificity of their activity profiles.

[0238] Fitness functions for automated molecular design. Automated tools that perform the above optimizations are valuable. Successful drug compounds require optimization and balance between potency, selectivity, and toxicity. "Backbone hopping" (when the activity of the lead compound is retained but the chemical structure is significantly changed) can improve pharmacokinetic, pharmacodynamic, toxicity, or intellectual property characteristics. There are algorithms that iteratively suggest new compounds, such as random generation of compounds, growing molecular fragments to fill given binding sites, genetic algorithms that "mutate" and "hybridize" compound populations, and exchange compound fragments with bioisosteric alternatives. Compounds generated by each of these methods can be evaluated for the multiple objectives mentioned above (potency, selectivity, toxicity), and, in the same way that the technology can provide information about each of the previous manual settings (binding predictions, selectivity, side effects, and toxicity predictions), it can be incorporated into an automated compound design system.

[0239] Drug Repurposing. Drugs often have side effects, but sometimes these side effects are beneficial. For example, aspirin, commonly used to treat headaches, is also used for cardiovascular health. Drug repositioning can significantly reduce the cost, time, and risk of drug discovery because these drugs have been shown to be safe in humans and are already optimized for rapid absorption and good stability in patients. Unfortunately, drug repositioning is largely accidental. For example, sildenafil (Viagra) was developed as a hypertension drug and was accidentally observed to be effective in treating erectile dysfunction. Computational predictions of off-target effects can be used in the context of drug repurposing to identify compounds that can be used to treat alternative diseases.

[0240] In the present disclosure, as in off-target side effect prediction, users can assemble a set of possible target proteins, each of which is associated with a disease. That is, inhibiting each target protein can treat (possibly different) diseases; for example, inhibitors of cyclooxygenase-2 can relieve inflammation, while inhibitors of factor Xa can be used as anticoagulants. These target proteins are annotated with the binding affinity of approved drugs (if any). A set of compounds is then assembled, limiting the set of compounds to compounds that have been approved or studied for use in humans. Finally, for each pair of target proteins and compounds, users can use the system to predict binding affinity. If the predicted molecular binding affinity is close to the binding affinity of the effective drug to the protein, candidates for drug repurposing can be identified.

[0241] Drug resistance prediction. Drug resistance is an inevitable consequence of drug use, which places selection pressure on rapidly dividing and mutating pathogen populations. Drug resistance is seen in a variety of disease pathogens such as viruses (HIV), foreign microorganisms (MRSA), and dysregulated host cells (cancer). Over time, a given drug will become ineffective, whether that drug is an antibiotic or a chemotherapeutic agent. At this point, intervention can be switched to another drug that will hopefully still be effective. In HIV, there are several well-known disease progression pathways that are defined by the mutations that the virus accumulates while a patient is treated.

[0242] There is a great deal of interest in predicting how disease pathogens will adapt to medical interventions. One approach is to characterize which mutations a disease pathogen will undergo during treatment. Specifically, the protein target of a drug will need to mutate to avoid binding the drug while continuing to bind its natural substrate.

[0243] In this application, a set of possible mutations in a target protein can be proposed. For each mutation, the resulting protein shape can be predicted. For each of these mutant protein forms, the system can be configured to predict the binding affinity for both the natural substrate and the drug. Mutations that cause the protein to no longer bind to the drug but continue to bind to the natural substrate are candidate mutations that confer drug resistance. These mutated proteins can be used as targets for designing drugs, for example by using these proteins as input for one of these other prediction use cases.

[0244] Personalized medicine. No drug should be taken that doesn’t work. In addition to the cost and hassle, all drugs have side effects. For ethical and economic reasons, drugs must be given only when the benefits outweigh the risks. Being able to predict when a drug will be helpful may be important. People differ from one another by a few mutations. However, small mutations can have profound effects. When these mutations occur at active (orthosteric) or regulatory (allosteric) sites of a disease target, they can prevent the drug from binding, thereby blocking the drug’s activity. When the protein structure for a specific person is known (or predicted), the system can be configured to predict whether a drug will be effective, or the system can be configured to predict when a drug will not work.

[0245] For such applications, the system can be configured to receive as input the chemical structure of a drug and a specific protein expressed by a specific patient. The system can be configured to predict the binding between the drug and the protein, and if the predicted binding affinity of the drug to the protein structure of the specific patient is too weak to be clinically effective, the clinician or practitioner can prevent the drug from being futilely prescribed to the patient.

[0246] Drug trial design. This application generalizes the personalized medicine use case described above to the context of patient populations. When the system can predict whether a drug will be effective for a specific patient phenotype, that information can be used to help design clinical trials. By excluding patients whose specific disease targets will not be adequately affected by the drug, clinical trials can achieve statistical power with fewer patients. Fewer patients directly reduces the cost and complexity of clinical trials.

[0247] For this application, users can segment possible patient populations into subpopulations characterized by expression of different proteins (e.g., due to mutations or subtypes). The system can be configured to predict the binding strength of drug candidates for different protein types. If the predicted binding strength for a particular protein type indicates that the necessary drug concentration is lower than the clinically achievable hospital concentration (e.g., based on test tubes, animal models, or physical characteristics of healthy volunteers), the candidate drug is expected to fail for that protein subpopulation. Patients carrying that protein may then be excluded from drug trials.

[0248] Agrochemical design. In addition to pharmaceutical applications, the agrochemical industry uses binding predictions in the design of new pesticides. For example, one consideration for pesticides is that they deter the species of concern without adversely affecting any other species. For ecological safety, it may be desirable to kill weevils without killing bumblebees.

[0249] For this application, the user can input a set of target protein structures from different species under consideration into the system. One set of target proteins can be designated as targets against which the compound is active, while the remaining proteins will be designated as targets against which the compound should be inactive. As with the previous use case, certain groups of compounds will be considered for each target protein (either in an existing database or ab initio), and the system will designate compounds with the greatest potency against the first set of target proteins, while avoiding the second set of target proteins.

[0250] Materials science. To predict the behavior and properties of new materials, it can be useful to analyze molecular interactions. For example, to study solvation, the user can input a repeating crystal structure of a given small molecule and evaluate the binding affinity of another example of that small molecule on the crystal surface. To study polymer strength, a set of polymer chains similar to the protein target structure can be input, and oligomers of the polymer can be input as small molecules. The system can thus predict the binding affinity between polymer chains.

[0251] Simulations. Simulators often measure the binding affinity of compounds for proteins, because the propensity of a compound to stay in a region of a target protein correlates with its binding affinity there. An accurate description of the features that govern binding can be used to identify regions and gestures with particularly high or low binding energies. Energy descriptions can be folded into molecular dynamics simulations to describe the motion of molecules and the occupancy of protein binding regions. Similarly, stochastic simulators used to study and model systems biology can benefit from accurate predictions of how small changes in compound concentrations affect biological networks.

[0252] Examples

[0253] Fig.21A The area under the curve (AUC) receiver operator characteristic (ROC) as a function of training iterations for two architectures of the present disclosure, o3-2.8.0 (2106) and o4-2.8.0 (2108), are provided, as Fig. 9B As shown, relative to the other two deep learning neural networks, architectures n8b-long (2102) and n8b-max-long (2104). The n8b-long and n8b-max-long architectures are of the type disclosed in U.S. Provisional Patent Application No. 63 / 251,142, filed on October 1, 2021, entitled “Characterization of Interactions Between Compounds and Polymers Using Negative Pose Data and Model Conditioning,” which is hereby incorporated by reference, and require that the pose of each training compound include a “positive pose” and a “negative pose.” Architectures n8b-long and n8b-max-long do not apply the attention mechanism 77 to multiple initial embeddings 74, each representing a different pose 48 of the compound, in order to obtain an attention embedding 79, as is the case with architectures o3-2.8.0 and o4-2.8.0. For Fig.21A According to the calculations summarized in , 3,097 different target proteins were utilized and a library of 5,562,818 compound-protein pairs was available for model training. All 3,097 different target proteins were used in this training. The 3,097 different target proteins were considered non-allosteric. See above for Fig.10 As noted, each of the four models (n8b-long, n8b-max-ling, o3-2.8.0, and o4-2.8.0) is trained using mini-batches of corresponding compound-protein pairs, each mini-batch of compound-protein pairs being selected from available compound-protein pairs using a graphics processing unit (GPU), where the number of compound-protein pairs in a mini-batch is limited by the GPU memory size. Fig.21AFor the training cited in , each mini-batch consisted of 64 different compound-protein pair poses. Each compound used in training had a known binary activity label for at least some of the 3,097 target proteins, either “active” when the compound had a pKa less than 10 μM relative to the target protein, or “inactive” when the compound had a pKa greater than 10 μM relative to the target protein. Thus, Fig.21A Each compound used in the training summarized in consists of at least some of the 3097 target protein activity markers. Fig.21A During the training summarized in , if the activity of a particular training compound against a particular target protein is unknown, the corresponding compound-protein pair is not included in the training. For each corresponding compound-protein pair of the corresponding mini-batch, the error exhibited by the corresponding model in calculating the activity signature of the corresponding training compound for the paired target protein pair of the corresponding compound-protein pair is used to refine the parameters of the corresponding model as part of the interaction. Therefore, each iteration represents the training generated by all compound-protein pairs of the corresponding mini-batch. Fig.21A As shown, going from 100,000 to 5,000,000 such iterations gives the ROC AUC performance of each model in predicting the activity of the training compounds in the compound-protein pairs. Fig.21A It is shown that the ROC AUC statistics of the models of the present disclosure represented by curves 2106 and 2108 continue to improve with increasing iteration numbers, while the ROC AUC statistics of models n8b-long and n8b-max-long represented by curves 2102 and 2104 exhibit overtraining, causing their ROC AUC values ​​to decrease after approximately 500,000 iterations.

[0254] Fig.21B AUCROC statistics of two architectures of the present disclosure, o3-2.8.0 (2106) and o4-2.8.0 (2108) relative to two other deep learning neural network architectures - long (2102) and n8b-max-long (2104) relative to the allosteric benchmark AA103 are provided. Fig.21B It is shown that, relative to the allosteric benchmark AA103, o3-2.8.0 and o4-2.8.0 have improved AUC ROC statistics relative to architecture-long (2102) and n8b-max-long (2104). Fig.21B For the calculations summarized in , 103 different target proteins were used, and a library of 552,011 compound-protein pairs was available for model training. All 103 different target proteins were used in this training. 103 different target proteins were considered allosteric. See above Fig.21AAs noted, each of the four models (n8b-long, n8b-max-long, o3-2.8.0, and o4-2.8.0) is trained using mini-batches of corresponding training compound-protein pairs, each mini-batch of compound-protein pairs being selected from the available training compound-protein pairs using a graphics processing unit (GPU), where the number of compound-protein pairs in a mini-batch is limited by the GPU memory size. Fig.21B For the training cited in , each mini-batch consisted of all poses for 64 different compound-protein pairs. Each training compound used in model training had a known binary activity label for at least some of the 103 target proteins, either “active” when the pKa of the training compound relative to the target protein was less than 10 μM, or “inactive” when the pKa of the training compound relative to the target protein was greater than 10 μM. Thus, Fig.21B Each training compound used in the training summarized in consists of active markers for at least some of the 103 target proteins. Fig.21B During the training summarized in , if the activity of a particular training compound against a particular target protein is unknown, the corresponding compound-protein pair is not included in the training. For each corresponding compound-protein pair of the corresponding mini-batch, the error of the activity signature of the corresponding training compound against the paired target protein pair of the corresponding compound-protein pair is calculated and used to improve the model parameters as part of the interaction. Therefore, each iteration represents the training generated by all compound-protein pairs of the corresponding mini-batch. Fig.21B As shown, the ROC AUC performance is given from 100,000 to 5,000,000 such iterations. Fig.21B It is shown that the ROC AUC statistics of the models of the present disclosure represented by curves 2106 and 2108 continue to improve as the number of iterations increases, while the ROC AUC of models n8b-long and n8b-max-long represented by curves 2102 and 2104 show overtraining as the number of training iterations increases.

[0255] in conclusion

[0256] For the purpose of explanation, the foregoing description has been described with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the embodiments to the precise forms disclosed. In view of the above teachings, many modifications and variations are possible. These embodiments are selected and described in order to best explain these principles and their practical applications, thereby enabling other persons skilled in the art to best use these embodiments and various embodiments with a variety of modifications suitable for the intended specific use.

Claims

1. A computer system for characterizing an interaction between a test compound and a target polymer, comprising: one or more processors; as well as a memory addressable by the one or more processors, the memory storing at least one program for execution by the one or more processors, the at least one program comprising instructions for: (A) obtaining a plurality of sets of atomic coordinates, wherein each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates comprises the test compound bound to the target polymer in a corresponding pose in a plurality of poses, wherein each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates comprises atomic coordinates of at least 5 atoms; (B) for each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates, inputting the corresponding set of atomic coordinates or an encoding of the corresponding set of atomic coordinates into a first neural network to obtain a corresponding initial embedding as an output of the first neural network to obtain a plurality of initial embeddings, wherein each initial embedding in the plurality of initial embeddings corresponds to a set of atomic coordinates in the plurality of sets of atomic coordinates, and wherein the first neural network includes more than 400 parameters and performs at least 10,000 calculations to calculate each initial embedding in the plurality of initial embeddings; (C) applying an attention mechanism to the multiple initial embeddings in a cascaded manner to obtain an attention embedding; (D) applying a pooling function to the attention embedding to derive a pooled embedding; and (E) inputting the pooled embedding into a first model to obtain a first interaction score for the interaction between the test compound and the target polymer. 2 . The computer system of claim 1 , wherein the first interaction score represents a binding coefficient of the test compound to the target polymer.

3. The computer system of claim 2, wherein the binding coefficient is the IC of the test compound relative to the target polymer. 50 ,EC 50 , Kd, ​​KI or pKI.

4. The computer system of claim 1, wherein the first interaction score represents a computer simulated gesture quality score of the test compound to the target polymer.

5. The computer system according to any one of claims 1 to 4, wherein the first model is a fully connected second neural network.

6. The computer system of claim 1, wherein the at least one program further comprises instructions for inputting the pooled embedding into a second model to obtain a second interaction score for the interaction between the test compound and the target polymer, wherein The first model is a first fully connected neural network, The second model is a second fully connected neural network, The first interaction score represents a computer simulated gesture quality score of the test compound to the target polymer, and The second interaction score represents a computer simulated pose quality score of the test compound to the target polymer.

7. The computer system of claim 6, wherein the at least one program further comprises instructions for inputting the first interaction score and the second interaction score into a third model to obtain a third interaction score, wherein the third model is a third fully connected neural network.

8. The computer system of claim 7, wherein the third interaction score is a discrete binary activity score having a first value when the test compound is determined by a third model to be inactive and having a second value when the test compound is determined by the third model to be active.

9. The computer system according to any one of claims 1 to 8, wherein the target polymer is an assembly of a protein, a polypeptide, a polynucleic acid, a polyribonucleic acid, a polysaccharide or any combination thereof.

10. The computer system of any one of claims 1 to 9, wherein each of the plurality of sets of atomic coordinates comprises three-dimensional coordinates {x1, ..., x2, ..., x3, ..., x4, ..., x5, ..., x6, ..., x7, ..., x8, ..., x9, ..., x10, ..., ..., x11, ..., ..., ...12, ..., ...13, ..., ...14, ..., ...15, ..., ...16, ..., ...17, ..., ...18, ..., ...19, ..., ...20, ..., ...21, ..., ...22, ..., ...23, ..., ...24, ..., ...25, ..., ...26, ..., ...27, ..., ...28, ..., ...30, ..., ...31, ..., ...32, ...33, ...34, ...35, ...36, ...37, ...38, ...39, ...40, ...41, ...42, N }, the crystal structure of the target polymer is or better resolution or or better resolution.

11. The computer system of any one of claims 1 to 9, wherein each of the plurality of sets of atomic coordinates comprises an ensemble of three-dimensional coordinates of at least a portion of the target polymer determined by nuclear magnetic resonance, neutron diffraction, or cryo-electron microscopy.

12. The computer system of claim 1, wherein the first interaction score is a binary score, wherein The first value of the binary score indicates that the IC value of the test compound relative to the target polymer is above a first threshold. 50 ,EC 50、 Kd, KI or pKI, and The second value of the binary score represents an IC value of the test compound relative to the target polymer that is below the first threshold. 50 ,EC 50 , Kd, ​​KI or pKI.

13. The computer system of any one of claims 1 to 12, wherein the test compound satisfies two or more, three or more, or all four of the following Lipinski five-fold rule: (i) no more than five hydrogen bond donors, (ii) no more than ten hydrogen bond acceptors, (iii) a molecular weight less than 500 Daltons, and (iv) a LogP less than 5.

14. The computer system of any one of claims 1 to 13, wherein 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.

15. The computer system of any one of claims 1 to 13, wherein the test compound is an organic compound having a molecular weight between 400 Daltons and 10,000 Daltons.

16. The computer system of any one of claims 1 to 15, wherein the plurality of sets of atomic coordinates consists of between 2 and 64 poses.

17. The computer system of any one of claims 1 to 16, wherein the first neural network is a convolutional neural network.

18. The computer system of claim 17, wherein each respective set of atomic coordinates in the plurality of sets of atomic coordinates comprises atomic coordinates of at least 80 atoms.

19. The computer system of any one of claims 1 to 16, wherein the first neural network is a graph neural network.

20. A computer system according to claim 19, wherein the graph neural network is characterized by an initial embedding layer and multiple interaction layers, each of which contributes an interaction data structure in multiple interaction data structures for each atom in a corresponding group of atomic groups of atomic coordinates of the corresponding posture in the multiple postures, and the multiple postures are pooled to form the corresponding initial embedding of the corresponding posture.

21. The computer system of any one of claims 1 to 16, wherein the first neural network is an equivariant neural network or a message passing neural network.

22. The computer system of any one of claims 1 to 16, wherein the first neural network comprises a plurality of graph convolution blocks, and each block uses a plurality of radial graphs to consider connectivity within the corresponding set of atomic coordinates.

23. The computer system of any one of claims 1 to 22, wherein the first neural network comprises a 1x10 6 parameters.

24. The computer system of any one of claims 1 to 23, wherein the corresponding initial embedding comprises a data structure comprising 100 or more values.

25. The computer system of any one of claims 1 to 24, wherein the plurality of initial embeddings comprises a first plurality of values, and wherein applying the attention mechanism comprises: (i) inputting the first plurality of values ​​into an attention neural network to obtain a first plurality of weights, wherein each weight in the first plurality of weights corresponds to a corresponding value in the first plurality of values, and (ii) weighting each corresponding value in the first plurality of values ​​by a corresponding weight in the plurality of weights to obtain the attention embedding.

26. The computer system of claim 25, wherein the first plurality of weights sums to one and each weight in the first plurality of weights is a scalar value between zero and one.

27. A computer system according to any one of claims 1 to 26, wherein the pooling function folds the attention embedding into the pooled embedding by applying a statistical function to combine each part of the attention embedding representing a different posture in the multiple postures to form the pooled embedding.

28. A computer system according to claim 27, wherein the attention embedding includes corresponding multiple values ​​of corresponding multiple elements of each corresponding posture in the multiple postures, and the statistical function is a maximum value function that takes the maximum value across the corresponding elements of each corresponding posture represented in the attention embedding to form the pooled embedding.

29. A computer system according to claim 27, wherein the attention embedding includes corresponding multiple values ​​of corresponding multiple elements of each corresponding posture in the multiple postures, and the statistical function is an average function that averages the corresponding elements of each corresponding posture represented in the attention embedding to form the pooled embedding.

30. The computer system of claim 1, wherein the first model is a regression task and the first interaction score quantifies the interaction between the test compound and the target polymer.

31. The computer system of claim 1, wherein the first model is a classification task and the first interaction score classifies the interaction between the test compound and the target polymer.

32. The computer system of any one of claims 1 to 31, wherein each respective set of the plurality of sets of atomic coordinates comprises atomic coordinates of at least 15 atoms, at least 20 atoms, at least 25 atoms, or at least 30 atoms.

33. The computer system of any one of claims 1 to 32, wherein the first neural network performs at least 100,000 computations to compute each of the plurality of initial embeddings.

34. The computer system of any one of claims 1 to 32, wherein the first neural network performs at least 1x10 6 The method further comprises performing a calculation to calculate each initial embedding of the plurality of initial embeddings.

35. The computer system of any one of claims 1 to 34, wherein the first model performs at least 10,000 calculations to calculate the first interaction score.

36. The computer system of any one of claims 1 to 34, wherein the first model performs at least 100,000 calculations to calculate the first interaction score.

37. The computer system of any one of claims 1 to 34, wherein the first model performs at least 1x10 6 A second calculation is performed to calculate the first interaction score.

38. The computer system of any one of claims 1 to 37, wherein the first model comprises more than 400 parameters, and wherein the first model performs more than 1000 calculations to calculate the first interaction score.

39. The computer system of any one of claims 1 to 37, wherein the first model comprises more than 400 parameters, and wherein the first model performs more than 10,000 calculations to calculate the first interaction score.

40. A method for characterizing an interaction between a test compound and a target polymer, comprising: (A) obtaining a plurality of sets of atomic coordinates, wherein each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates comprises the test compound bound to the target polymer in a corresponding pose in a plurality of poses, wherein each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates comprises atomic coordinates of at least 5 atoms; (B) for each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates, inputting the corresponding set of atomic coordinates or an encoding of the corresponding set of atomic coordinates into a first neural network to obtain a corresponding initial embedding as an output of the first neural network to obtain a plurality of initial embeddings, wherein each initial embedding in the plurality of initial embeddings corresponds to a set of atomic coordinates in the plurality of sets of atomic coordinates, and wherein the first neural network includes more than 400 parameters and performs at least 10,000 calculations to calculate each initial embedding in the plurality of initial embeddings; (C) applying an attention mechanism to the multiple initial embeddings in a cascaded manner to obtain an attention embedding; (D) applying a pooling function to the attention embedding to derive a pooled embedding; and (E) inputting the pooled embedding into a first model to obtain a first interaction score for the interaction between the test compound and the target polymer.

41. A non-transitory computer readable storage medium, wherein the non-transitory computer readable storage medium stores instructions that, when executed by a computer system, cause the computer system to perform a method of characterizing an interaction between a test compound and a target polymer, the method comprising: (A) obtaining a plurality of sets of atomic coordinates, wherein each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates comprises the test compound bound to the target polymer in a corresponding pose in a plurality of poses, wherein each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates comprises atomic coordinates of at least 5 atoms; (B) for each corresponding set of atomic coordinates in the plurality of sets of atomic coordinates, inputting the corresponding set of atomic coordinates or an encoding of the corresponding set of atomic coordinates into a first neural network to obtain a corresponding initial embedding as an output of the first neural network to obtain a plurality of initial embeddings, wherein each initial embedding in the plurality of initial embeddings corresponds to a set of atomic coordinates in the plurality of sets of atomic coordinates, and wherein the first neural network includes more than 400 parameters and performs at least 10,000 calculations to calculate each initial embedding in the plurality of initial embeddings; (C) applying an attention mechanism to the multiple initial embeddings in a cascaded manner to obtain an attention embedding; (D) applying a pooling function to the attention embedding to derive a pooled embedding; and (E) inputting the pooled embedding into a first model to obtain a first interaction score for the interaction between the test compound and the target polymer.

Images