Advanced methods and systems for determining properties of molecules by machine learning
By constructing a 3D structural model and surface model of molecules, using machine learning to predict the thermodynamic properties of macromolecules, the problems of low computational efficiency and insufficient accuracy of macromolecules in the existing technology are solved, and efficient and accurate for macromolecules systems are achieved. Properties prediction.
Patent Information
- Application Number
- CN202410607610.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-05-30
- Filing Date
- 2024-05-16
- Publication Date
- 2024-12-03
AI Technical Summary
Existing methods are difficult to accurately calculate the thermodynamic properties of macromolecules in condensing environments, especially in computationally inefficient and insufficient accuracy, limiting their application to the research and development of macromolecules such as polymers and biomolecules.
By constructing a 3D structural model of molecules and generating surface models, using machine learning models to predict the charge and chemical potential of segments, combined with cavity construction models and neural network technology, efficient calculation of the thermodynamic properties of macromolecules can be achieved.
It significantly reduces calculation time, improves accuracy, and can quickly and efficiently calculate the thermodynamic properties of macromolecules in the liquid phase. It is suitable for polymers and biomolecules, expanding the application field of thermodynamic equilibrium prediction.
Smart Images

Figure CN119069023A_ABST
Abstract
Description
Background Art
[0001] Existing methods for determining the quantum chemical properties of molecules in a condensed environment are limited to small or medium sized molecules. Further, depending on the size and complexity of the molecule, existing methods for continuum solvation models are computationally demanding and / or less accurate. Summary of the invention
[0002] Therefore, there is a need for functionality with improved accuracy and computational efficiency to determine properties of molecules, such as macromolecules, in a condensing environment. Embodiments provide such functionality.
[0003] One such embodiment provides this functionality by constructing one or more three-dimensional (3D) structural models indicating the positions of atoms of the molecule. For each of the one or more 3D structural models constructed, such embodiments further: (i) generate a surface model representing the environment, wherein the surface model includes a plurality of segments, and the generated surface model defines a relationship between the indicated positions of atoms of the 3D structural model and the plurality of segments; and (ii) use a machine learning model to predict a charge (e.g., electric charge) and a chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model.
[0004] In addition, some embodiments provide functionality for reviewing candidate molecules for their properties in an environment. Example embodiments for reviewing candidate molecules involve a method that receives one or more user requirements, determines properties of each candidate molecule in a plurality of candidate molecules, and selects a given molecule from the plurality of candidate molecules based on the determined properties of the given molecule and the received one or more user requirements.
[0005] Certain embodiments relate to fluid phase thermodynamics and simulations. For example, some embodiments are generally applicable to, for example, organic and organometallic molecules. The size of the molecular system to which certain embodiments of the described methods can be applied is not limited to small or medium-sized molecules. Advantageously, certain embodiments can also be used for macromolecular systems, such as polymers or biomolecules, and other examples. Further, some embodiments can be used to predict and calculate the thermodynamic equilibrium properties of molecular systems in liquid phase and vapor / gas phase. The determination of such properties is important in many different industries, such as materials science, pharmaceuticals, life sciences, medical treatment, consumer products, cosmetics, polymers and coatings, etc. Specifically, some embodiments can be applied to, for example, chemical engineering, drug development, formulation design of personal and consumer care products, packaging material design and excipient screening for plastic recycling, and other examples.
[0006] Further, certain embodiments can be used to calculate thermodynamic equilibrium properties of, for example, large organic molecules and ions in a liquid phase in a fast manner with reliable accuracy. Thermodynamic equilibrium properties can include, for example, activity coefficients, vapor pressures, solubility, solvation free energy, partition coefficients, electroactivity, and other related properties known in the art. Through some embodiments, macromolecules can also be modeled and analyzed, which is important for industrial applications such as developing new compounds and materials, such as biodegradable polymers for the pharmaceutical industry or polymer manufacturers, and other examples.
[0007] Certain embodiments may use innovative workflows involving machine learning to construct and calculate surface charge densities and corresponding potentials for organic and organometallic molecules and ions in dielectric continua. The method according to an embodiment is computationally efficient and is capable of modeling and analyzing macromolecules including, for example, polymers. Some embodiments may extend the domain of applicability of methods for predicting thermodynamic properties using charge density to, for example, complex polymers or biochemical systems.
[0008] In addition, certain embodiments can calculate thermodynamic equilibrium properties of liquid phases involving conventional and novel large organic molecules and ions in a fast manner with reliable accuracy. The efficient method according to the embodiment reduces the calculation time by several orders of magnitude, thereby enabling high-throughput screening and extending the applicable domain of thermodynamic equilibrium predictions to large molecules, such as polymers or biomolecules. Some embodiments provide a computational method for constructing and calculating segment-by-segment surface charge density and potential of molecules and ions in a dielectric continuum using a workflow involving machine learning, without the need for prior knowledge other than the atom type (e.g., represented by the element symbol or atomic number) and the atomic 3D coordinates, while taking into account the segment-specific local 3D chemical environment.
[0009] One such embodiment provides this functionality through a workflow comprising: (i) providing a molecular geometry (conformer) or a set of molecular geometries (conformer set) as input; (ii) constructing conformer-specific segments on the solvent-accessible surface area; (iii) approximating the segment-specific information, plus predicting molecular energies, in an efficient process involving a trained machine learning model; and (iv) writing out the collected segment information and energies for each conformer considered. According to one embodiment, the aforementioned information is written in a so-called COSMO file or other suitable file format known to those skilled in the art. Finally, the COSMO (or other format) file information can be used by statistical thermodynamics software packages, such as or other suitable software packages known in the art to predict the thermodynamic equilibrium properties of molecules in a condensing environment.
[0010] Example embodiments relate to a computer-implemented method for determining properties of a molecule in an environment, such as a condensed phase environment such as a liquid, solvent, or excipient. The method begins by constructing one or more 3D structural models indicating the positions of atoms of the molecule. Next, for each of the one or more constructed 3D structural models, the method: (i) generates a surface model representing the environment, wherein the surface model includes a plurality of segments, and the generated surface model defines a relationship between the indicated positions of the atoms of the 3D structural model and the plurality of segments; and (ii) uses a machine learning model to predict the charge and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model. In one embodiment, the relationship between the indicated positions of the atoms of the 3D structural model and the plurality of segments can be defined by the generated surface model using (i) the center position or coordinates of the constructed segments on a solvent accessible surface and (ii) the indicated positions of the atoms of the 3D structural model. According to one embodiment, the solvent accessible surface is defined as the boundary of all positions in space, which can be acquired by the center of the solvent or the probe ball. According to one aspect, the generated surface model stores individual coordinates for each atomic position and each segment position. In such embodiments, these stored coordinates define the relationship between the positions of the atoms and the segments. Further, it should be noted that the surface model comprising multiple segments can be generated by any suitable cavity construction method known to those skilled in the art, including known cavity construction methods such as described or discussed herein.
[0011] Certain embodiments may predict charge and chemical potential using a corresponding machine learning model for each property. For example, in one aspect, the machine learning model includes a first machine learning model and a second machine learning model. According to one such aspect, predicting the charge and the chemical potential of each of the plurality of segments based on the 3D structure model and the generated surface model includes: predicting the charge of each of the plurality of segments based on the 3D structure model and the generated surface model using the first machine learning model; and predicting the chemical potential of each of the plurality of segments based on the 3D structure model and the generated surface model using the second machine learning model.
[0012] According to another embodiment, for each constructed 3D structural model, determining the property of the molecule in the environment further includes using a supplementary machine learning model to predict an energy corresponding to the 3D structural model based on the 3D structural model and the generated surface model.
[0013] According to another example embodiment, the machine learning model comprises a neural network. In one aspect, the neural network comprises one or more hidden layers, and the neural network is configured to employ an activation function at one or more nodes of the one or more hidden layers. According to one embodiment, the activation function is one of a rectified linear unit (ReLU) activation function and a softmax function. However, embodiments are not limited to the specific activation functions listed above, and instead any suitable activation function known in the art may be employed.
[0014] In yet another example embodiment, the method further comprises training the machine learning model based on a training data set. According to one aspect, the machine learning model comprises a neural network, and training the machine learning model based on the training data set comprises training the neural network by iteratively updating one or more network weights of the neural network based on the training data set. In one embodiment, iteratively updating the one or more network weights of the neural network based on the training data set comprises employing one or more of an adaptive moment estimation (Adam) solver algorithm and an early stopping algorithm. According to an example embodiment, the training data set comprises data of one or more of the following: example molecules, example conformers, example segments, example segment charges, example segment chemical potentials, and example continuum model energies.
[0015] In one embodiment, using the machine learning model to predict the charge and the chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model includes deriving input feature data based on the 3D structural model. Further, such embodiments use the machine learning model to predict the charge and the chemical potential of each of the plurality of segments based on the 3D structural model, the generated surface model, and the derived input feature data. According to one such embodiment, the derived input feature data includes an indication of one or more of: atom type, atom-atom distance, atom-segment distance, bond type, bond angle, torsion angle, formal charge, 3D atomic position, and atom type specific features.
[0016] Example embodiments further include receiving one or more user requirements. Such embodiments then evaluate candidate molecules with respect to the received user requirements. For each candidate molecule in a plurality of candidate molecules, example embodiments perform said constructing and said determining said properties, and select said given molecule from said plurality of candidate molecules based on the determined properties of the given molecule and the received one or more user requirements. According to one aspect, said one or more user requirements may include, for example, a molecular or ionic input structure, an input symbol, and / or a connectivity table.
[0017] In another example embodiment, predicting the charge and the chemical potential of each of the plurality of segments based on the 3D structure model and the generated surface model using the machine learning model comprises: correcting one or more residual charges of the plurality of segments; and determining an overall formal charge of the plurality of segments based on the corrected one or more residual charges of the plurality of segments. According to one such example embodiment, the determined overall formal charge is a predicted charge of the plurality of segments.
[0018] In an exemplary embodiment, each of the constructed one or more 3D structural models corresponds to a corresponding conformer of the molecule.
[0019] Various types of information and / or models can be used to construct the one or more 3D structural models. For example, according to yet another exemplary embodiment, one or more 3D structural models indicating the positions of the atoms of the molecule are constructed based on indications of one or more of atom types, coordinates, and chemical connectivity, as well as other examples. On the other hand, the one or more 3D structural models can be constructed by adopting one or more of the following: rule-based geometric models, force fields, and quantum chemically derived geometric models, as well as other non-limiting examples. According to an embodiment, the quantum chemically derived geometric model can include, for example, a tightly bound model, a semi-empirical model, a density functional theory derived geometric model, or any combination thereof.
[0020] In another embodiment, a cavity building model is used to generate a surface model representing the environment. Some embodiments may use any suitable cavity building model known in the art, such as a COSMO fine cavity building model.
[0021] Another example embodiment relates to a computer-based system for determining properties of molecules in an environment. The system comprises: a processor; and a memory having computer code instructions stored thereon. In such embodiments, the processor and the memory having the computer code instructions are configured to cause the system to implement any embodiment or combination of embodiments described herein.
[0022] Yet another example embodiment relates to a cloud computing implementation for determining properties of molecules in an environment. Such embodiments relate to a computer program product executed by a server communicating with one or more clients over a network, wherein the computer program product includes instructions that, when executed by one or more processors, cause the one or more processors to implement any embodiment or combination of embodiments described herein.
[0023] It should be noted that the method, system and computer program product embodiments may be configured to implement any embodiment or combination of embodiments described herein. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] The foregoing will be apparent from the following more particular description of example embodiments, as illustrated in the accompanying drawings, in which like reference numerals refer to the same parts throughout the different views. The drawings are not necessarily drawn to scale, emphasis instead being placed upon illustrating the embodiments.
[0025] Figure 1 is a flow chart of a method for determining properties of a molecule in an environment (eg, continuum solvation model properties), according to an embodiment.
[0026] Figure 2 is a simplified block diagram of a system for determining properties of a molecule in an environment (eg, continuum solvation model properties), according to an embodiment.
[0027] Figure 3 An exemplary workflow for determining properties of a molecule in an environment according to one embodiment is presented.
[0028] Figure 4 is a simplified block diagram of a computer system for determining properties of molecules in an environment according to one embodiment.
[0029] Figure 5 is a simplified block diagram of a computer network environment in which embodiments of the present invention may be implemented. DETAILED DESCRIPTION
[0030] The following is a description of example embodiments.
[0031] Chemistry, i.e. chemical interactions / reactions, occurs primarily in a liquid or solvent phase. Theoretical models for embedding molecules in a solvent continuum are known. For example, existing continuum solvation models approximate the solvent as a dielectric continuum of solute molecules surrounding the outside of a molecular cavity. The cavity surface is approximated by segments, such as hexagons, pentagons, or triangles. Models that approximate the solvent as a dielectric continuum can be referred to as dielectric continuum solvation models (DCSMs), and include widely used polarizable continuum models (PCMs). Another example conventional DCSM is the conductor-like screening model (COSMO), which derives the polarization charge of the continuum caused by the polarity of the solute from the scaled conductor approximation [1] (the numbers in brackets in this document refer to the reference list listed below). COSMO is one of the most applied computational methods for determining the electrostatic interactions of molecules with solvents or liquid environments. Output data from COSMO or other known DCSM methods (generally referred to herein as "COSMO information" or "COSMO type information") can include, for example, segment-by-segment surface charge density and chemical potential. In general, DCSMs each approximate the solvent effects in a similar manner by approximating the dielectric continuum.
[0032] The COSMO information can then be used to calculate the chemical potential of the molecule in a solvent or solvent mixture. This can be performed, for example, using the existing COSMO-RS (where “RS” stands for “real solvent”) method, which takes as input previously calculated and stored COSMO information. The COSMO-RS method involves modeling a set of physicochemical interaction terms between a molecule and its liquid environment in terms of the pairwise charge density of specific segments of the molecular cavity [2].
[0033] The chemical potential data generated by, for example, COSMO-RS can then form the basis for calculating general thermodynamic equilibrium properties. Such thermodynamic properties can include, for example, activity coefficients, solubility, partition coefficients, vapor pressures, and solvation free energies. The COSMO-RS method was developed to provide a general prediction method that does not require system-specific adjustments. The method is widely used in academia as well as in the chemical and pharmaceutical industries.
[0034] Current methods for generating COSMO information rely on quantum chemical calculations [1,2], partially supplemented by machine learning post-correction [3] or fragmented methods [4].
[0035] Quantum chemical calculations have the disadvantage that they may require very high computational requirements, particularly for larger molecular systems. Typically, quantum chemical calculations are performed on high performance computing (HPC) clusters for sets of related conformers. Therefore, the conventional application of quantum chemical methods is limited to the settings of trained users with access to HPC clusters. This type of computationally intensive calculation limits quantum chemical methods to small molecules, with molecular weights significantly lower than 1 kilodalton (kDa). In contrast, the scope of chemical and biochemical products related to many industries such as polymers, surfactants, proteins or biopharmaceuticals is 10 to 100kDa or even larger. Modern drug molecules also quickly reach this molecular weight limit.
[0036] Although some existing quantum chemical methods, such as semiempirical methods, have reduced computational requirements, employing such methods results in insufficient polarity distribution, which in turn impairs subsequent calculations of thermodynamic properties.
[0037] The fragmentation method cannot properly reflect the 3D conformational space and its influence on the charge distribution. For example, in certain 3D configurations, the influence of the conformational space on the charge distribution is affected by the formation of intramolecular hydrogen bonds.
[0038] As an example, some embodiments described herein provide the advantage of reducing the computational workload of calculating surface charge density from days or even weeks to seconds or even less time, with an accuracy comparable to that of quantum chemical calculations. In certain embodiments, quantum chemical calculations are replaced by an efficient workflow involving a set of machine learning programs. In addition, digital artifacts generated by the derivation of quantum chemical equations can be avoided or eliminated by machine learning programs implemented by some embodiments. Therefore, compared with existing methods, certain embodiments provide thermodynamic property calculations with increased prediction accuracy. The fast machine learning programs of some embodiments take molecular geometry or a set of molecular geometries as input. Certain embodiments efficiently predict the charge of a molecular surface segment by reflecting the local chemical environment of a segment in a specific molecule or conformer, such as screening charge and chemical potential. In one embodiment, a segment is constructed in 3D space by adopting any suitable cavity construction model known in the art, such as the COSMO fine cavity construction model described in [5] (which is incorporated herein by reference in its entirety), or any other existing cavity construction method reviewed by [5] (which is incorporated herein by reference in its entirety).
[0039] Example Method Embodiments
[0040] Figure 1One such example method embodiment 100 is shown. Method 100 is a computer-implemented method for determining properties of a molecule in an environment. Method 100 begins at step 101 by constructing one or more 3D structural models indicating the positions of atoms of the molecule, for example in a computer memory. In one embodiment, constructing the one or more 3D structural models at step 101 can be performed using any suitable technique known to those skilled in the art. For example, embodiments may use known software platforms, such as Platforms, such as Pipeline Materials Discovery or Or other platforms such as TURBOMOLE to build 3D structural models. According to one embodiment, each of the one or more 3D structural models constructed corresponds to the corresponding conformer of the molecule. At step 101, various types of information and / or models can be used to build the one or more 3D structural models. For example, in an embodiment of method 100, the one or more 3D structural models are constructed (101) based on the indication of one or more of the atomic type, coordinates and chemical connectivity. Chemical connectivity refers to the way atoms are connected to each other in space. Further, in some embodiments, the model can be constructed based on the measurement results / observations of real-world molecules, and the resulting model reflects the properties of real-world measurements / observations. According to another example embodiment of method 100, constructing the one or more 3D structural models (101) includes adopting one or more of the following: rule-based geometric models, force fields and quantum chemically derived geometric models, and other non-limiting examples. In one embodiment, the quantum chemically derived geometric model can include, for example, a tightly bound model, a semi-empirical model, a density functional theory derived geometric model or any combination thereof.
[0041] To continue, method 100 determines properties of the molecule in the environment at step 102. For each of the constructed one or more 3D structural models (from step 101), the properties of the molecule are determined at step 102 by: (i) generating a surface model representing the environment, for example in a computer memory, wherein the surface model includes a plurality of segments, and the generated surface model defines a relationship between the indicated positions of atoms of the 3D structural model and the plurality of segments; and (ii) using a machine learning model to predict a charge (e.g., electric charge) and a chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model. In an embodiment, generating the surface model at step 102 can be performed using any suitable technique known to those skilled in the art.
[0042] As described above, method 100 is computer-implemented, and thus functions and effective operations, such as constructing (101) and determining (102), are automatically performed by one or more digital processors. In addition, method 100 can be implemented using any computer device or combination of computing devices known in the art. In other examples, method 100 can be implemented using the following description of Figure 4 The computer system 440 described below with respect to Figure 5 The described computer network environment 550 is implemented.
[0043] At step 102, method 100 may use a first machine learning model to predict charge, and use a second machine learning model to predict chemical potential. An embodiment of method 100 uses a first machine learning model at step 102 to predict the charge of each of the multiple segments based on the 3D structure model and the generated surface model. In such embodiments, the 3D structure model (and / or its characteristics) and the generated surface model (and / or its characteristics) are provided as inputs to the first machine learning model, and the first machine learning is configured to output the charge of each segment in response to the input. Similarly, at step 102, an embodiment uses a second machine learning model to predict the chemical potential of each of the multiple segments based on the 3D structure model and the generated surface model. Such embodiments input the 3D structure model (and / or its characteristics) and the generated surface model (and / or its characteristics) to the second machine learning model, and the second machine learning is configured to output the chemical potential of each segment in response to the input.
[0044] In an embodiment of method 100, determining the properties of the molecule in the environment at step 102 further comprises predicting an energy corresponding to the 3D structure model based on the 3D structure model and the generated surface model using a supplementary machine learning model. Such embodiments input the 3D structure model (and / or its characteristics) and the generated surface model (and / or its characteristics) to the supplementary machine learning model, and the supplementary machine learning model is configured to output an energy corresponding to the 3D structure model in response to the input.
[0045] In an embodiment of the method 100, using a machine learning model to predict a charge (e.g., electric charge) and a chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model (102) includes deriving input feature data based on the 3D structural model. Such embodiments use a machine learning model to predict a charge and a chemical potential of each of the plurality of segments based on the 3D structural model, the generated surface model, and the derived input feature data. According to an embodiment, the derived input feature data includes an indication of one or more of: atom type, atom-atom distance, atom-segment distance, bond type, bond angle, torsion angle, formal charge, 3D atomic position, and atom type specific features.
[0046] In an embodiment of the method 100, predicting a charge (e.g., electric charge) and a chemical potential of each of the plurality of segments based on the 3D structure model and the generated surface model (102) using a machine learning model includes: correcting one or more residual charges of the plurality of segments; and determining an overall formal charge of the plurality of segments based on the corrected one or more residual charges of the plurality of segments. According to one such example embodiment, the determined overall formal charge is a predicted charge of the plurality of segments.
[0047] In an embodiment of method 100, generating a surface model representing the environment (102) includes employing a cavity building model. Embodiments may employ any suitable cavity building model known in the art, such as the COSMO fine cavity building model described in [5], or any other existing cavity building method reviewed, for example, by [5]. In an embodiment of method 100, the machine learning model used at step 102 includes a neural network. Further, in yet another embodiment, the neural network includes one or more hidden layers, and the neural network is configured to employ an activation function at one or more nodes of the one or more hidden layers. According to one embodiment, the activation function is one of a rectified linear unit (ReLU) activation function and a softmax function. However, embodiments are not limited to the specific activation functions listed above; instead, any suitable activation function known in the art may be employed. Although neural networks are discussed herein, those skilled in the art will recognize that embodiments of method 100 are not limited to such techniques. Instead, embodiments of method 100 may utilize any suitable known machine learning or statistical learning method to determine charge and chemical potential at step 102.
[0048] According to an embodiment, method 100 further includes training a machine learning model based on a training data set. In an embodiment, the machine learning model includes a neural network, and training the machine learning model based on the training data set includes training the neural network by iteratively updating one or more network weights of the neural network based on the training data set. Further, in yet another embodiment, iteratively updating the one or more network weights of the neural network based on the training data set includes employing one or more of an adaptive moment estimation (Adam) solver algorithm and an early stopping algorithm. It should be noted that the embodiments are not limited to the specific algorithms listed above; instead, any suitable algorithm known in the art may be employed. According to an example embodiment, the training data set includes data for one or more of the following: example molecules, example conformers, example segments, example segment charges, example segment chemical potentials, and example continuum model energies. In yet another embodiment, the training data set may include a large collection of COSMO files covering organic chemical space spanning, for example, solvents, industrial chemicals, pharmaceuticals, ion protonation states, ions, and ionic liquids (cations and anions, also including multiply charged ions).
[0049] According to an embodiment, method 100 further comprises receiving one or more user requirements. In one such example embodiment, for each candidate molecule in the plurality of candidate molecules, method 100 then performs construction (101) and determination of properties (102), and selection of the given molecule from the plurality of candidate molecules based on the determined properties of the given molecule and the received one or more user requirements. According to one aspect, the one or more user requirements may comprise, for example, a molecular or ionic input structure, an input symbol, and / or a connectivity table.
[0050] In addition, in another embodiment of method 100, the properties of the candidate molecules are determined based on the real-world measurements of the candidate molecules. These measured properties are then used at step 101 to build a computer-based model of the candidate molecules having the characteristics. In turn, these models (which reflect the real-world measured / observed properties of the candidate molecules) are used to determine the charge and chemical potential of each candidate molecule at step 102. A given candidate molecule that meets the desired criteria can then be selected for use in real-world applications, such as developing a research plan, optimizing formulations to improve product properties in a desired manner, designing sustainable polymers for recycling and other processes involving polymers, identifying suitable pharmaceutical excipients in the context of drug development, focusing a group of experiments, and reducing material waste by improving experimental efficiency, as well as other examples.
[0051] Certain embodiments of method 100 may use the file formats / structures of various known software tools such as COSMO to store various properties determined / predicted by method 100, such as charge (e.g., electric charge), chemical potential, energy, and any other data / values described herein. Some embodiments may also be stored in the cloud, such as However, the embodiments are not limited to a specific file format / structure or cloud platform; instead, any suitable file format / structure or cloud platform known in the art may be used.
[0052] Example System Embodiments
[0053] Figure 2 is a simplified block diagram of a system 220 for determining properties of molecules in an environment, according to one embodiment.
[0054] like Figure 2 As shown, on the one hand, the system 220 includes one or more data sources 221, a molecular model generator 222, a surface model generator 223, one or more machine learning models 224 and an output storage device 225.
[0055] In an embodiment, data sources 221 may include molecular data, i.e., data about a molecule whose properties are to be determined, and environmental data, i.e., data about the environment of the molecule. According to an embodiment, data sources 221 may be provided by a user of system 220. In an example embodiment, data sources 221 may be used as input to, for example, molecular model generator 222 and surface model generator 223.
[0056] According to an embodiment of the system 220 , using data about the molecule provided by the data source 221 , the molecular model generator 222 may construct one or more 3D structural models indicating the positions of atoms of the molecule.
[0057] Likewise, in one aspect, using data about the environment provided by data source 221, surface model generator 223 can generate a surface model representing the environment for each 3D structural model of interest constructed by molecular model generator 222. In such embodiments, each surface model includes a plurality of segments, and each generated surface model defines a relationship between the indicated positions of atoms of the 3D structural model and the plurality of segments.
[0058] According to one embodiment, based on each 3D structure model of the one or more 3D structure models constructed by the model generator 222 and the corresponding surface model generated by the surface model generator 223, the machine learning model 224 can predict, for example, the charge (e.g., electric charge) and chemical potential of each of the multiple segments of the surface model. In one embodiment, the machine learning model 224 may include a first machine learning model and a second machine learning model, wherein the first machine learning model predicts the charge and the second machine learning model predicts the chemical potential. Further, according to one aspect, the machine learning model 224 may include a supplementary machine learning model, which predicts the energy corresponding to the 3D structure model based on the 3D structure model and the generated surface model. In one embodiment, the machine learning model 224 may include a neural network.
[0059] In an example embodiment, properties predicted by the machine learning model 224, such as charge, chemical potential, and / or energy, may be output to the storage device 225. Further, according to an embodiment, once recorded in the storage device 225, such output data may then be used for additional processing, such as calculating thermodynamic properties, among other examples.
[0060] It should be noted that the system 220 may implement any of the embodiments described herein, for example, Figure 1 A method 100 is described for determining properties of molecules in an environment.
[0061] Example Workflow
[0062] Figure 3 An exemplary workflow 330 for determining the properties of a molecule in an environment according to an embodiment is shown. The workflow 330 may implement any of the embodiments described herein, for example, Figure 1 and 2 The method 100 and system 220 are described to determine properties of molecules in an environment.
[0063] In one embodiment, at step 331 of the workflow 330, an input structure is obtained, for example, in response to a user action. According to an example embodiment, the input structure may be a structure of a molecule, the property of which in an environment is to be determined. Figure 3 An example two-dimensional (2D) molecular structure 340 that may be obtained at step 331 of workflow 330 is shown.
[0064] Next, according to an embodiment, at step 332, 3D structure generation of the molecule may occur. Figure 3The constructed 3D structure 341a of the 2D molecular structure 340 is shown. In one embodiment, optional conformational isomer generation of the molecule may also be performed at step 333. For example, Figure 3 Additional conformers 341b-c are depicted. According to an embodiment, the 3D structure generation of step 332 and the optional conformer generation of step 333 can include, for example, constructing one or more 3D structural models (341a-c) indicating the positions of atoms of the molecule in a computer memory. In one aspect, constructing the one or more 3D structural models (341a-c) can be performed using any suitable technique known to those skilled in the art. According to an example embodiment, each of the one or more 3D structural models (341a-c) constructed corresponds to a corresponding conformer of the molecule (340).
[0065] Continue About Figure 3 According to one embodiment, workflow 330 determines the properties of a molecule in an environment. For each of the one or more constructed 3D structure models 341a-c (from step 332, and optionally step 333), the properties of the molecule are determined by: (i) generating, for example, in a computer memory, a surface model (342a-c) representing the environment, wherein the surface model 342a-c comprises a plurality of segments (e.g., tiles that may be hexagonal, such as Figure 3 (ii) predicting the charge and chemical potential of each of the plurality of segments based on the 3D structure models 342a-c and the generated surface models 342a-c, for example at step 335. According to an example embodiment, generating the surface model may include performing cavity construction and tiling into segments at step 334. It should be noted that generating the surface model including, for example, cavity construction and tiling into segments at step 334 may be performed using any suitable technique known to those skilled in the art. According to another example embodiment, such as, for example Figure 3 As shown in region 343 , when visually presented, individual segments may be shaded / colored to distinguish segments that map to or correspond to specific atoms of a molecule.
[0066] According to one embodiment, step 335 of workflow 330 can include using one or more machine learning models to variously predict the charge and chemical potential of each segment and the energy of each conformer. For example, in one embodiment, a first machine learning model can be used to predict charge, and a second machine learning model can be used to predict chemical potential. According to one aspect, a supplemental machine learning model can be used to predict energy. In one embodiment, such as, for example, Figure 3As shown in region 344 of FIG. 1 , when visually presented, individual segments may be shaded / colored to distinguish segments having a negative, positive, or neutral predicted charge.
[0067] To continue Figure 3 In an example embodiment, as a processing option, a geometry optimization may be performed at step 336 of workflow 330 by changing the 3D structure model and its optimization toward an energy minimum, followed by repeating steps 333, 334, and 335. According to one such embodiment, this processing option may be iterated one or more times.
[0068] In yet another example embodiment, after the one or more machine learning models predict a property, such as charge, chemical potential, and / or energy, the value of the property may be written or saved to one or more output files 346 at step 337 of the workflow 330. Further, according to an embodiment, after being saved at step 337, such output data may then be used for additional processing at step 338, such as calculating thermodynamic properties and / or generating thermodynamic predictions 347, among other examples.
[0069] Molecular or macromolecular geometry
[0070] Certain embodiments may construct one or more 3D structural models associated with a particular molecule of interest, for example at step 101 of method 100 .
[0071] Some embodiments may take as input a molecular geometry (single conformer) or a set of molecular geometries (set of conformers). Conformers may include, for example, tautomerism and / or protonation states. Molecular geometry may be defined by individual atomic coordinates and atom types.
[0072] In addition, certain embodiments may perform any one or more of the following operations to construct a model indicating the positions of atoms: (i) conformational search to generate a set of corresponding 3D structural models of conformers; (ii) tautomerism search to generate a set of corresponding 3D structural models or a set of conformers for a single conformer of each tautomer; (iii) search for relevant protonation states to generate a set of corresponding 3D structural models or a set of conformers for a single conformer of each protonation state; (iv) create different substitution patterns; and (v) create different mutant or variant structures.
[0073] According to one embodiment, constructing a 3D structure model comprises receiving one or more molecular geometries from any source known in the art, such as an industry standard fast 3D structure generator, X-ray structure, force field, tight binding method, or semi-empirical or density functional calculations.
[0074] For example, in one embodiment, Pipeline or The software can be used to generate molecular geometries and conformer sets from chemical connectivity information stored in molecular files such as SDF (Structure Data File), MOL (Molfile), PDB (Protein Data Bank) or related molecular file formats or molecular input line notation, SMILES (Simplified Molecular Input Line Entry System), and other examples. In some embodiments, the use of force fields can also be used Materials or Discovery Generate reliable 3D geometries for molecules and macromolecular systems using molecular builders such as CHARMM (Chemistry at Harvard University Macromolecular Mechanics) or COMPASS (Condensed Phase Optimized Molecular Potentials for Atomistic Simulation Studies) to build 3D structural models. TURBOMOLE supports complementary methods for generating 3D structures, such as RDKit related force fields, MOPAC (molecular orbital package) semi-empirical methods, or xTB semi-empirical extensions, which can be used by certain embodiments. In some embodiments, density functional calculations can be performed by quantum chemistry packages such as TURBOMOLE to generate 3D structures. It should be noted that the embodiments are not limited to the specific software platforms or file formats described herein; instead, any suitable software platform or file format known in the art can be used.
[0075] Molecular 3D geometry can be generated by, for example, the ETKDG (Experimental Torsion Elementary Knowledge Distance Geometry) method of the RDKit chemical information library. Other known sources and techniques of molecular geometry can be used in addition to those discussed herein.
[0076] It should also be noted that in some embodiments, atomic coordinates and atom types are used to determine molecular properties, such as continuum solvation model properties. Unlike existing methods, such embodiments do not explicitly rely on, for example, bond information, connectivity information, functional groups, fingerprints, segment-specific information, molecular surface-specific information, energy, or potential.
[0077] Construction of solvent accessible surface segments
[0078] Certain embodiments may, for example, generate a surface model representing an environment such as a continuum solvent environment at step 102 of method 100. The generated surface model may include multiple segments. Further, the generated surface model may define the relationship between the indicated positions of atoms of the 3D structural model and the multiple segments. According to one embodiment, the surface model may encapsulate a conformer (represented by the 3D structural model) in a tiled cavity, wherein the tiled cavity represents the continuum solvent environment. In some embodiments, the tiled cavity representing the continuum solvent environment may be calculated by any suitable technique known in the art, such as the COSMO solvation model. According to one embodiment, multiple segments of a given surface model may represent a conformer-specific solvent accessible surface (SAS).
[0079] In principle, the underlying electrostatic principles of DCSM are accurate. In practice, there is no well-defined surface separating the solute from the solvent environment because the electron density of the solute and solvent molecules overlap. Therefore, cavity definition and molecule-specific cavity construction are key steps.
[0080] In some embodiments, the solvent accessible surface segments can be constructed in 3D space. Such embodiments can employ any suitable cavity construction method known in the art, for example, existing cavity construction methods reviewed by [5].
[0081] According to one aspect, for example, at step 102 of method 100, a surface model can be generated using a COSMO refined cavity construction model [5]. The marching tetrahedron algorithm employed by the refined model provides a triangulation technique to reach surface segments. The refined model further utilizes an isodense cavity construction algorithm based on atom-type specific COSMO radii, which produces a smooth, fully tiled cavity of the molecular shape. This is particularly important for the COSMO-RS model, which uses the screening charge density on the surface as the primary descriptor for the definition of intermolecular interactions. Details of the procedure are given in [5].
[0082] Machine learning models for predicting charge and chemical potential
[0083] For each surface segment of the generated surface model, certain embodiments calculate charge (e.g., electric charge) and chemical potential as typical inputs to the COSMO-RS thermodynamic model. In existing methods, charge distribution and chemical potential are determined by quantum chemical calculations. In contrast, some embodiments use one or more machine learning models to predict the charge and chemical potential of a segment. Compared to existing methods, certain embodiments therefore provide significant acceleration of several orders of magnitude. Some embodiments reduce the computation time from hours per central processing unit (CPU) core to a few seconds, even at the most precise level of consideration. Similar to quantum chemistry, the machine learning model-based methods of certain embodiments reflect the special and atomic environment of a segment in a specific spatial molecular arrangement of adjacent atoms.
[0084] Some embodiments may apply one or more machine learning models to predict target properties: segment charge and segment chemical potential. For certain embodiments, quantum chemical information used as input to the training model may be obtained or constructed from, for example, molecular geometries optimized at the density functional level, such as through, for example, the B88-VWN-P86 functional set of the TURBOMOLE software and the def-TZVP (valence triple-ζ polarization) basis set and subsequent single-point calculations in the def2-TZVPD (valence triple-ζ polarization with diffusion function) basis set, where the activated scaled conductors approximate to infinite dielectric constant ε=∞. The segment charge may include, for example, a screening charge or a charge density. According to some embodiments, in In the naming of the software, the two quantum chemical levels can be abbreviated as BP-TZVP and BP-TZVPD-FINE. In this case, the chemical potential is the chemical potential of the segment that responds to the polarization change.
[0085] The one or more machine learning models of certain embodiments may be trained on a large number of COSMO file collections serving as training data sets. On the one hand, the training data set may include data for one or more of the following: example molecules, example conformers, example segments, example segment charges, example segment chemical potentials, and example continuum model energies. In one embodiment, the one or more machine learning models may be further validated using known methods and / or using a test set. The COSMO file collection used to train the machine learning models of certain embodiments may cover organic chemical spaces including, for example, solvents, industrial chemicals, pharmaceuticals, ion protonation states, ions, and ionic liquids (cations and anions, also including multiply charged ions). According to some embodiments, the training data set may include a set of conformers representing both intramolecular hydrogen bonds and open (intermolecular) hydrogen bond candidates. On the one hand, the COSMO file collection used for the training data set may include approximately 16,000 compounds, each quantum chemical level represented by approximately 65,000 conformers. According to one embodiment, depending on the size of the molecule and the quantum chemical level represented by the COSMO file, each conformer may be represented by approximately 10 on average. 3 In some embodiments, only the available approximately 10 segments of all conformers across the entire dataset are used during training. 8 According to one such embodiment, on average, only 3% of the segments associated with carbon and hydrogen atoms can be used for training and validation (thus, 97% are retained for the test set), and only 25% of the segments can be associated with nitrogen, oxygen, and fluorine atoms (thus, 75% are retained for the test set). In one aspect, no threshold is applied to other atom types. According to one embodiment, during training and validation, a fraction of 0.9 of all used segments can be used for training, while a fraction of 0.1 can be reserved for validation.
[0086] In one embodiment, since the segments in the COSMO file already reflect the external charge correction, no explicit external charge correction is required. According to one embodiment, after all segments are processed, a correction of residual charge can be performed to obtain the overall formal charge. Certain embodiments can also apply correction to the predicted charge using various techniques. Such techniques can include, but are not limited to: (i) alignment of formal charge and overall predicted conformer-based charge; and (ii) correction of external charge as needed.
[0087] The one or more machine learning models applied by certain embodiments, for example at step 102 of method 100, may include, but are not limited to, artificial neural networks. In one embodiment, the artificial neural network architecture may include an input layer, one or more hidden layers that may start with, for example, a dense hidden layer consisting of 256 nodes, and an output layer. Embodiments may employ any neural network architecture known in the art. For example, embodiments may utilize various types of architectures, including, but not limited to, artificial neural networks and deep neural networks. In addition, a neural network according to an embodiment may include additional layers such as convolutional layers. According to one aspect, a neural network may use a rectified linear unit (ReLU) activation function, an adaptive moment estimation (Adam) solver algorithm, 10 -8 Tolerances and early stopping algorithms. Further, it should be noted that the embodiments are not limited to the specific number of layers, number of nodes, activation functions, algorithms, or tolerances described herein; instead any suitable number of layers, number of nodes, activation functions, algorithms, or tolerances known in the art may be employed. In one embodiment, the hyperparameters of the artificial neural network architecture may be further optimized. For example, network parameters may be selected to achieve a desired balance between network size and model performance. In some embodiments, input features may be scaled by removing the mean of the training set and scaling to unit variance. Although neural networks are discussed herein, those skilled in the art will recognize that the embodiments are not limited to such techniques; instead any suitable known machine learning or statistical learning methods may be used.
[0088] In some embodiments, the one or more machine learning models may use the same information types of input features as those used by conventional quantum chemical methods. Such features may include, for example, atom types, atom-atom distances, atom-segment distances, bond types, bond angles, torsion angles, formal charges, 3D atomic positions, and various atom type specific features. According to one embodiment, the 3D atomic position may be a rotationally invariant representation of the 3D positioning of atoms relative to each other. Further, in one embodiment, an optional cutoff radius may be applied. Certain embodiments provide, in particular, the benefits of not having to rely on features such as molecular topology or connectivity, molecular fragments, chemical functional groups, and classical chemical information fingerprints. By avoiding such dependencies, some embodiments may determine, for example, very general chemical descriptions of possible charge distributions. Although certain embodiments do not need to consider these features, they may be used optionally. On the one hand, there is no direct cutoff radius, but up to 24 closest atoms may be considered as the chemical environment of a particular segment. According to one embodiment, the atom type is not directly encoded, but indirectly encoded by its quantum chemical features, such as electron affinity, ionization potential, allowed orbital configuration, atom type ratio radius, etc. According to some embodiments, this technique even further increases the applicable domain of the one or more machine learning models. However, in some embodiments, the atom type can be encoded directly. All atom types up to Radon (atomic number 86, i.e., not containing g orbitals) and ionic structures are supported by the construction of the one or more machine learning models according to the embodiments.
[0089] Machine learning models for energy prediction
[0090] Certain embodiments may further use a supplemental machine learning model (i.e., a machine learning model other than the one or more models used to predict charge and chemical potential at step 102) to predict energies corresponding to the 3D structural model based on the 3D structural model and the generated surface model. As described herein, in one embodiment, the 3D structural model may correspond to a conformer of a molecule.
[0091] In some embodiments, the predicted energy can be one or more of the following: (i) total molecular or ion energy; (ii) dielectric energy, such as dielectric energy in a conductor-like dielectric continuum; and (iii) gas phase energy. According to one aspect, for each target property and calculation level (e.g., BP-TZVP and BP-TZVPD-FINE), the same set of COSMO files (discussed above under the heading "Machine Learning Model for Predicting Charge and Chemical Potential") can be used to train the supplementary machine learning model. Similarly, in one embodiment, a score of 0.9 for each target property can be used for training, and a score of 0.1 can be reserved for validation. Some embodiments can employ the same artificial neural network architecture and apply hyperparameters as described above for the supplementary machine learning model. However, in one aspect, the neural network used to predict energy (e.g., total molecular energy, ion energy, dielectric energy, or gas phase energy) can include a first hidden layer consisting of 512 nodes instead of 256 nodes for total energy and gas phase energy. Certain embodiments can use similar input features as the supplementary machine learning model, such as atom type, distance, angle, torsion, and formal charge, etc. to predict energy. According to one embodiment, the atom type is not encoded directly, but indirectly encoded by its quantum chemical characteristics, such as electron affinity, ionization potential, allowed orbital configurations, and atom type specific radius, etc. Some embodiments can predict the energy (e.g., conductor or gas phase state) of atom types such as H (hydrogen), Li (lithium), Be (beryllium), B (boron), C (carbon), N (nitrogen), O (oxygen), F (fluorine), Na (sodium), Mg (magnesium), Si (silicon), P (phosphorus), S (sulfur), K (potassium), Ca (calcium), Cl (chlorine), Se (selenium), Br (bromine), and I (iodine), as well as other examples.
[0092] Although neural networks are discussed herein, those skilled in the art will recognize that embodiments are not limited to such techniques; rather embodiments may use any suitable known machine learning or statistical learning methods to predict energy. Disclosed machine learning models include deep learning frameworks for molecules and materials, such as those described by [6].
[0093] Certain embodiments provide single point results for predefined molecular geometries. Because some embodiments predict energies, such as total molecular energy or ion energy, such embodiments can also be extended to run geometry optimization in a condensed environment by changing the 3D structural model and optimizing it towards an energy minimum.
[0094] Writing output files
[0095] Certain embodiments may collect the results, including, for example, predicted segment charges, segment-wise chemical potentials, and / or energies as described above, and write the results to an output file, such as a dedicated conformer-specific COSMO file that stores all COSMO type information. For example, the predicted charges and chemical potentials determined at step 102 of method 100 may be written to an output file. For a set of conformers, some embodiments may generate multiple COSMO files by running the segment build (described above under the heading "Building Solvent Accessible Surface Segments") and all predictions (described above under the headings "Machine Learning Model for Predicting Charge and Chemical Potential" and "Machine Learning Model for Predicting Energies") for each specific conformer separately. Certain embodiments may also be stored in a cloud, such as The output file, such as COSMO file, is generated and stored in the platform. The embodiments are not limited to a specific output file format or cloud platform; instead, any suitable output file format or cloud platform known in the art may be used.
[0096] Use of output files in additional calculations
[0097] To demonstrate the application of the novel techniques described herein, output files according to some embodiments, such as COSMO files, can be used as The software can calculate the fluid thermodynamic properties by COSMO-RS method. Some embodiments can also be used, for example, by Platform (e.g. Virtual ) or any suitable platform known to those skilled in the art to perform the calculations.
[0098] Embodiment can successfully predict thermodynamic equilibrium properties, such as solubility, partition coefficient and / or liquid density. Further, the result shows that such properties determined using the embodiment are closely matched with known experimentally determined values. Based on the conformational isomer set (e.g., COSMO file set) generated by the novel machine learning technology according to the embodiment, the prediction of various chemical substances is very consistent with the prediction generated by conventional quantum chemical calculation (e.g., density functional theory). In addition, the result of the machine learning method according to the embodiment is superior to the result generated by the existing method.
[0099] In addition to thermodynamic property predictions, other examples of additional computations based on the output files according to embodiments include, but are not limited to: machine learning models that use the generated prediction data as input features (e.g., charge density distribution, energy, and σ-moment, etc.), software applications related to quantum chemistry and / or materials science (e.g., TURBOMOLE and Materials ) and prediction of biological and / or biochemical properties (e.g. Discovery ).
[0100] advantage
[0101] Embodiments determine the properties of molecules in an environment and offer numerous advantages.
[0102] For example, the fast machine learning procedures of the embodiments reduce the time required to construct and calculate COSMO information for a complete set of conformers from many days or even weeks to seconds or even less.
[0103] As another example advantage, embodiments do not require an HPC cluster. Embodiments facilitate automation of high throughput screening predictions and democratization of solutions for thermodynamic equilibrium calculations in liquid phases. Cloud-based platforms such as the platform benefit from the results provided by the embodiments that are essentially immediately available. In this way, the embodiments reduce costs and carbon footprint; in addition, the embodiments increase the platform's expectations for customers.
[0104] Further, as yet another example advantage, embodiments can extend the applicability of COSMO and COSMO-RS to macromolecular systems such as biomolecules and polymers, as well as other examples. Embodiments can also be used to efficiently guide experiments. In addition, embodiments provide solubility predictions for polymers, which can be used to design sustainable polymers, and thereby embodiments can improve real-world processes utilizing such polymers, such as recycling processes. Since excipients are typically polymer-based, embodiments can additionally enhance virtual screening of drugs in drug development. In such settings, efficient methods for screening properties, such as solubility, are needed to focus experiments and reduce material waste due to inefficient experimental testing. Embodiments provide the desired methods.
[0105] Advantageously, embodiments also improve the accuracy of thermodynamic property calculations by avoiding numerical artifacts produced by the derivation of quantum chemical equations through efficient machine learning procedures.
[0106] Computer Support
[0107] Embodiments can be implemented in existing software and computer-aided design and computer-aided engineering platforms. For example, embodiments can use 3DS features and functions of the software.
[0108] Figure 4It is a simplified block diagram of a computer-based system 440, which can be used to calculate the properties of molecules in an environment according to various embodiments of the present invention described herein. System 440 includes a bus 443. Bus 443 serves as an interconnect between the various components of system 440. Connected to bus 443 is an input / output device interface 446, which is used to connect various input and output devices such as keyboard, mouse, touch screen, display, speaker, etc. to system 440. CPU 442 is connected to bus 443 and provides the execution of computer instructions. Memory 445 provides volatile storage of data for executing computer instructions. Storage device 444 provides non-volatile storage for software instructions such as operating system (not shown). System 440 also includes a network interface 441 for connecting to any multiple networks known in the art, including wide area network (WAN) and local area network (LAN).
[0109] It should be understood that the example embodiments described herein can be implemented in many different ways. In some cases, the various methods and machines described herein can each be a physical, virtual, or hybrid general-purpose computer such as computer system 440 or as described below with respect to Figure 4 The computer system 440 may be implemented in a computer network environment such as the computer environment 440 described herein. For example, by loading software instructions implementing the method 100 into the memory 445 or the non-volatile storage device 444 for execution by the CPU 442, the computer system 440 may be converted into a machine that performs the methods described herein. It should be further understood by those of ordinary skill in the art that the system 440 and its various components may be configured to perform any embodiment or combination of embodiments described herein. Further, the system 440 may implement the various embodiments described herein using any combination of hardware, software, and firmware modules operatively coupled internally or externally to the system 440.
[0110] Figure 5 A computer network environment 550 is shown in which embodiments of the present invention may be implemented. In the computer network environment 550, a server 551 is linked to clients 553a-n via a communication network 552. The environment 550 may be used alone or in combination with the server 551 to enable the clients 553a-n to perform any of the embodiments described herein. As non-limiting examples, the computer network environment 550 provides a cloud computing embodiment, a software as a service (SaaS) embodiment, and the like.
[0111] Embodiments or aspects thereof may be implemented in the form of hardware, firmware, or software. If implemented in software, the software may be stored on any non-transitory computer-readable medium that is configured to enable a processor to load the software or a subset of its instructions. The processor then executes the instructions and is configured to operate or cause the device to operate in the manner described herein.
[0112] Further, firmware, software, routines or instructions may be described herein as performing certain actions and / or functions of a data processor. However, it should be understood that such descriptions included herein are only for convenience and such actions are actually caused by a computing device, processor, controller or other device that executes the firmware, software, routines, instructions.
[0113] It should be understood that flow charts, block diagrams, and network diagrams may include more or fewer elements, may be arranged in different ways, or may be presented in different ways. However, it should be further understood that certain embodiments may indicate block diagrams and network diagrams and the number of block diagrams and network diagrams that demonstrate the execution of an embodiment implemented in a particular manner.
[0114] Therefore, additional embodiments may also be implemented using a variety of computer architectures, physical, virtual, cloud computers, and / or some combination thereof, and therefore the data processors described herein are intended for illustration purposes only and not as limitations of the embodiments.
[0115] The teachings of all patents, published applications, and references cited herein are incorporated by reference in their entirety.
[0116] While example embodiments have been particularly shown and described, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the embodiments as encompassed by the following claims.
[0117] References
[0118] [1] A. Klamt, G. Schüürmann (1993). “COSMO: a new approach to dielectric screening in solvents with explicit expressions for the screening energy and its gradient”. J. Chem. Soc. Perkin Trans. 2 (5): 799-805. doi: 10.1039 / P29930000799
[0119] [2] A. Klamt (2005). “From Quantum Chemistry to Fluid Phase Thermodynamics and Drug Design”. Boston, MA, USA: Elsevier. ISBN 9780444519948
[0120] [3] PC Petris, P. Becherer, J. GEM Fraaije (2021). “Alkane / water partition coefficient calculation based on the modified AM1 method and internal hydrogen bonding sampling using COSMO-RS”. Journal of Chemical Information and Modeling (J. Chem. Inf. Model.), 61(7): 3453-3462. doi:10.1021 / acs.jcim.0c01478
[0121] [4] M. Hornig, A. Klamt (2005). “COSMOfrag: a novel tool for high-throughput ADME property prediction and similarity screening based on quantum chemistry”. Journal of Chemical Information and Modeling, 45: 1169-1177. doi: 10.1021 / ci0501948
[0122] [5] A. Klamt, M. Diedenhofen (2018). “A refined cavity construction algorithm for the conductor-like screening model”. J. Comput. Chem. 39: 1648-1655. doi: 10.1002 / jcc.25342
[0123] [6] Schütt et al. (2018). “SchNet-A deeplearning architecture for molecules and materials”. J. Chem. Phys. 148: 241722 (2018); doi: 10.1063 / 1.5019779
Claims
1. A computer-implemented method for determining a property of a molecule in an environment, the method comprising: constructing one or more three-dimensional (3D) structural models indicating positions of atoms of the molecule; as well as For each of the one or more constructed 3D structural models, the property of the molecule in the environment is determined by: generating a surface model representing the environment, wherein the surface model comprises a plurality of segments, and the generated surface model defines a relationship between the indicated positions of the atoms of the 3D structure model and the plurality of segments; as well as A machine learning model is used to predict the charge and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model.
2. The method of claim 1 , wherein the machine learning model comprises a first machine learning model and a second machine learning model, and wherein predicting the charge and the chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model comprises: predicting, using the first machine learning model, a charge of each of the plurality of segments based on the 3D structural model and the generated surface model; as well as The chemical potential of each of the plurality of segments is predicted using the second machine learning model based on the 3D structural model and the generated surface model.
3. The method according to claim 1, wherein for each of the one or more constructed 3D structural models, determining the property of the molecule in the environment further comprises: An energy corresponding to the 3D structural model is predicted based on the 3D structural model and the generated surface model using a supplemental machine learning model. 4 . The method according to claim 1 , wherein each of the one or more constructed 3D structural models corresponds to a corresponding conformer of the molecule.
5. The method of claim 1, wherein the machine learning model comprises a neural network.
6. The method of claim 5, wherein the neural network comprises one or more hidden layers, and the neural network is configured to employ an activation function at one or more nodes of the one or more hidden layers.
7. The method of claim 6, wherein the activation function is one of a rectified linear activation function and a softmax function.
8. The method of claim 1, further comprising training the machine learning model based on a training data set.
9. The method of claim 8, wherein the machine learning model comprises a neural network, and wherein training the machine learning model based on the training data set comprises: The neural network is trained by iteratively updating one or more network weights of the neural network based on the training data set.
10. The method of claim 9, wherein iteratively updating the one or more network weights of the neural network based on the training data set comprises employing one or more of an adaptive moment estimation solver algorithm and an early stopping algorithm.
11. The method of claim 8, wherein the training data set comprises data for one or more of: example molecules, example conformers, example segments, example segment charges, example segment chemical potentials, and example continuum model energies.
12. The method of claim 1 , wherein using the machine learning model to predict the charge and the chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model comprises: deriving input feature data based on the 3D structural model; as well as The charge and the chemical potential of each of the plurality of segments are predicted using the machine learning model based on the 3D structural model, the generated surface model, and the derived input feature data.
13. A method according to claim 12, wherein the derived input feature data includes indications of one or more of the following: atom type, atom-atom distance, atom-segment distance, bond type, bond angle, torsion angle, formal charge, 3D atomic position and atom type specific features.
14. The method of claim 1, further comprising: receiving one or more user requests; for each candidate molecule in a plurality of candidate molecules, performing said constructing and said determining said property; as well as The given molecule is selected from the plurality of candidate molecules based on the determined properties of the given molecule and the received one or more user requirements.
15. The method of claim 1 , wherein using the machine learning model to predict the charge and the chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model comprises: correcting one or more residual charges of the plurality of segments; as well as An overall formal charge for the plurality of segments is determined based on the corrected one or more residual charges for the plurality of segments, wherein the determined overall formal charge is a predicted charge for the plurality of segments.
16. The method of claim 1, wherein constructing the one or more 3D structural models indicating the positions of the atoms of the molecule is based on an indication of one or more of: atom types, coordinates, and chemical connectivity.
17. The method of claim 1, wherein constructing the one or more 3D structural models indicating the positions of the atoms of the molecule comprises employing one or more of: a rule-based geometric model, a force field, and a quantum chemically derived geometric model.
18. The method of claim 1, wherein generating the surface model representing the environment comprises constructing the model using a cavity.
19. A computer-based system for determining properties of molecules in an environment, the system comprising: processor; as well as a memory having computer code instructions stored thereon, the processor and the memory having the computer code instructions being configured to cause the system to: constructing one or more three-dimensional (3D) structural models indicating positions of atoms of the molecule; and For each of the one or more constructed 3D structural models, the property of the molecule in the environment is determined by: generating a surface model representing the environment, wherein the surface model comprises a plurality of segments, and the generated surface model defines a relationship between the indicated positions of the atoms of the 3D structure model and the plurality of segments; as well as A machine learning model is used to predict the charge and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model.
20. A non-transitory computer program product for determining a property of a molecule in an environment, the computer program product being executed by a server in communication with one or more clients over a network and comprising: A computer readable medium comprising program instructions which, when executed by one or more processors, cause the one or more processors to: constructing one or more three-dimensional (3D) structural models indicating positions of atoms of the molecule; and For each of the one or more constructed 3D structural models, the property of the molecule in the environment is determined by: generating a surface model representing the environment, wherein the surface model comprises a plurality of segments, and the generated surface model defines a relationship between the indicated positions of the atoms of the 3D structure model and the plurality of segments; as well as A machine learning model is used to predict the charge and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model.
