Advanced methods and systems for determining molecular properties through machine learning

Abstract

To determine properties of a molecule in an environment.SOLUTION: Such an embodiment constructs one or more three-dimensional (3D) structure models that indicate positions of atoms of the molecule. For each of the constructed one or more 3D structure models, (i) a surface model is generated that represents the environment, where 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 structure model and the plurality of segments, and (ii) using a machine learning model, charge (e.g., electric charge) and chemical potential of each segment of the plurality of segments are predicted based on the 3D structure model and the generated surface model. An embodiment further predicts, using a supplemental machine learning model, energy corresponding to the 3D structure model based on the 3D structure model and the generated surface model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to advanced methods and systems for determining molecular properties through machine learning. [Background technology]

[0002] Existing approaches for determining quantum chemical properties of molecules in condensed environments are limited to small or medium-sized molecules. Furthermore, existing continuum solvation modeling approaches are computationally demanding and / or less accurate depending on the size and complexity of the molecule. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] [1] A. Klamt, G. Schuurmann (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

[0004] [Non-patent document 2] [2] A. Klamt (2005). “From Quantum Chemistry to Fluid Phase Thermodynamics and Drug Design”. Boston, MA, USA: Elsevier. ISBN 9780444519948

[0005] [Non-patent document 3] [3] P.C. Petris, P. Becherer, J.G.E.M. Fraaije (2021). “Alkane / water partition coefficient calculation based on the modified AM1 method and internal hydrogen bonding sampling using COSMO-RS”. J. Chem. Inf. Model. 61 (7): 3453-3462. doi:10.1021 / acs.jcim.0c01478

[0006] [Non-Patent Document 4] [4] M. Hornig, A. Klamt (2005). “COSMOfrag:a novel tool for high-throughput ADME property prediction and similarity screening based quantum chemistry”. J. Chem. Inf. Model. 45: 1169-1177. doi:10.1021 / ci0501948

[0007] [Non-Patent Document 5] [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

[0008] ​​​[6]Schutt et al. (2018). “SchNet - A deep learning architecture for molecules and materials”. J. Chem. Phys. 148: 241722 (2018); doi:10.1063 / 1.501977 Summary of the Invention

[0009] Therefore, there is a need for improved accuracy and computationally efficient functionality for determining the properties of molecules, e.g., large molecules, in condensed environments. Embodiments provide such functionality.

[0010] One such embodiment provides this functionality by constructing one or more three-dimensional (3D) structural models that indicate the positions of atoms of the molecule. Then, for each of the constructed one or more 3D structural models, (i) generate a surface model that represents the environment, the surface model including a plurality of segments, the generated surface model defining relationships between the indicated positions of the atoms of the 3D structural model and the plurality of segments, and (ii) use a machine learning model to predict the electrical charge (e.g., charge) and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model.

[0011] Additionally, certain embodiments provide functionality for vetting candidate molecules for their properties in an environment. An exemplary embodiment for vetting candidate molecules is directed to a method for receiving one or more user requirements, determining properties of each candidate molecule of a plurality of candidate molecules, and selecting a given molecule from among the plurality of candidate molecules based on the determined properties of the given molecule and the received one or more user requirements.

[0012] Certain embodiments relate to fluid-phase thermodynamics and simulation. For example, some embodiments are generally applicable to, for example, organic and organometallic molecules. The size of molecular systems to which certain embodiments of the described approaches can be applied is not limited to small or medium molecules. Advantageously, certain embodiments can also be used with large molecular systems, such as polymers or biomolecules, among other examples. Furthermore, some embodiments can be used to predict and calculate thermodynamic equilibrium properties of molecular systems in the liquid and vapor / gas phases. Determining such properties is important across many different industries, such as materials science, pharmaceuticals, life sciences, medical care, consumer goods, cosmetics, polymers, and coatings. In particular, some embodiments can be applied to excipient screening, for example, chemical engineering, drug development, formulation design for personal and consumer care products, packaging material design, and plastic recycling.

[0013] Furthermore, certain embodiments can be used to rapidly calculate thermodynamic equilibrium properties of, for example, large organic molecules and ions in the liquid phase with reliable accuracy. Thermodynamic equilibrium properties can include, for example, activity coefficients, vapor pressure, solubility, free energy of solvation, partition coefficients, reactivity, and other relevant properties known in the art. In some embodiments, large molecules can also be modeled and analyzed, which is important for polymer manufacturers for industrial applications such as the development of new compounds and materials, e.g., biologics in the pharmaceutical industry, or biodegradable polymers, among other examples.

[0014] Certain embodiments can use innovative workflows, including machine learning, to construct and calculate surface charge densities and corresponding potentials for organic and organometallic molecules and ions in the dielectric continuum. The method according to one embodiment is computationally efficient and can model and analyze large molecules, including, for example, polymers. Some embodiments can extend the applicability domain of methods that use charge densities to predict thermodynamic properties to, for example, complex polymeric or biochemical systems.

[0015] Furthermore, certain embodiments can rapidly calculate thermodynamic equilibrium properties involving the liquid phase of, for example, conventional and novel large organic molecules and ions with reliable accuracy. Efficient methods according to embodiments reduce computational time by several orders of magnitude, thus enabling high-throughput screening to be performed and the applicable domain of thermodynamic equilibrium prediction to be extended to large molecules, e.g., polymers or biomolecules. Some embodiments provide computational methods for constructing and calculating segmental surface charge densities and potentials of molecules and ions in a dielectric continuum using a workflow involving machine learning with no prior knowledge other than atom types (e.g., represented by element symbols or atomic numbers) and atomic 3D coordinates, while taking into account the segment-specific local 3D chemical environment.

[0016] One such embodiment provides this functionality via a workflow that includes: (i) providing a molecular shape (conformer) or a set of molecular shapes (a set of conformers) as input; (ii) constructing conformer-specific segments on the solvent accessibility surface; (iii) approximating the segment-specific information in an efficient process involving a trained machine learning model and predicting molecular energies; and (iv) exporting the collected information about segments and energies for each conformer under consideration. According to one embodiment, the aforementioned information is described 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 processed by a statistical thermodynamics software package, such as BIOVIA® COSMOtherm®, or other suitable software packages known in the art, to predict the thermodynamic equilibrium properties of molecules in condensed environments.

[0017] An exemplary embodiment is directed to a computer-implemented method for determining properties of a molecule in an environment, e.g., a condensed-phase environment such as a liquid, solvent, or excipient. The method begins by constructing one or more 3D structural models that indicate the positions of atoms of the molecule. Next, for each of the constructed one or more 3D structural models, the method (i) generates a surface model representing the environment, the surface model including a plurality of segments, the generated surface model defining 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 may be defined by the generated surface model using a relationship between (i) the center positions or coordinates of the constructed segments on the 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 may be taken by the center of a solvent or probe sphere. According to one aspect, the generated surface model stores separate coordinates for each atom position and each segment position. In such implementations, these stored coordinates define the relationship between the positions of atoms and the segments. Furthermore, it should be noted that a surface model including multiple segments may be generated by any suitable cavity construction method known to those skilled in the art, including, for example, the known cavity construction methods described or discussed herein.

[0018] Certain embodiments may predict the charge and chemical potential using a respective machine learning model for each property. For example, in one aspect, the machine learning models include a first machine learning model and a second machine learning model. According to one such aspect, predicting the charge and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model includes using the first machine learning model to predict the charge of each of the plurality of segments based on the 3D structural model and the generated surface model, and using the second machine learning model to predict the chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model.

[0019] According to another embodiment, determining the properties of the molecules in the environment further includes, for each constructed 3D structural model, predicting an energy corresponding to the 3D structural model based on the 3D structural model and the generated surface model using a complementary machine learning model.

[0020] According to another exemplary embodiment, the machine learning model includes a neural network. In one aspect, the neural network includes one or more hidden layers, and the neural network is configured to use an activation function at one or more nodes of the one or more hidden layers. According to one implementation, 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 may employ any suitable activation function known in the art.

[0021] In yet another exemplary embodiment, the method further includes training a machine learning model based on the training dataset. According to one aspect, the machine learning model includes a neural network, and training the machine learning model based on the training dataset includes training the neural network by iteratively updating one or more network weights of the neural network based on the training dataset. In one implementation, iteratively updating the one or more network weights of the neural network based on the training dataset includes employing one or more of an adaptive moment estimation (Adam) solver algorithm and an early stopping algorithm. According to an exemplary embodiment, the training dataset includes data for one or more of an exemplary molecule, an exemplary conformer, an exemplary segment, an exemplary segment charge, an exemplary segment chemical potential, and an exemplary continuum model energy.

[0022] In one embodiment, using 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 includes deriving input feature data based on the 3D structural model. Further, such an embodiment uses the 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 and the derived input feature data. According to one embodiment, the derived input feature data includes an indication of one or more of atom type, interatomic distance, atom-segment distance, bond type, bond angle, torsion angle, formal charge, 3D atom position, and atom-type specific features.

[0023] Exemplary embodiments further include receiving one or more user requirements. Such embodiments then evaluate the candidate molecules in relation to the received user requirements. For each candidate molecule of the plurality of candidate molecules, the exemplary embodiments perform property construction and determination and select 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 include, for example, a molecular input structure or ionic input structure, an input representation, and / or a connection table.

[0024] In another exemplary embodiment, predicting the charge and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model using the 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 exemplary embodiment, the determined overall formal charge is a predicted charge of the plurality of segments.

[0025] In an exemplary embodiment, each 3D structural model of the one or more constructed 3D structural models corresponds to a respective conformer of the molecule.

[0026] Various types of information and / or models may be used to construct one or more 3D structural models. For example, according to yet another exemplary embodiment, one or more 3D structural models showing the positions of atoms of a molecule are constructed based on representations of one or more of atom types, coordinates, and chemical connectivity, among other examples. In another aspect, the one or more 3D structural models may be constructed by employing one or more of rule-based geometric models, force fields, and quantum-chemically derived geometric models, among other non-limiting examples. According to one embodiment, quantum-chemically derived geometric models may include, for example, tight-binding models, semi-empirical models, geometric models derived from density functional theory, or any combination thereof.

[0027] In another implementation, the surface model representing the environment is generated using a cavity construction model. Some embodiments may employ any suitable cavity construction model known in the art, such as the COSMO FINE cavity construction model.

[0028] Another exemplary embodiment is directed to a computer-based system for determining properties of molecules in an environment. The system includes a processor and a memory having computer code instructions stored thereon. In such an embodiment, the processor and memory are configured to use the computer code instructions to cause the system to implement any embodiment or combination of embodiments described herein.

[0029] Yet another exemplary embodiment is directed to a cloud computing implementation for determining properties of molecules in an environment. Such an embodiment is directed to a computer program product executed by a server in communication with one or more clients over a network, the computer program product including 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.

[0030] 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 explanation of the drawings]

[0031] The foregoing will be apparent from the following more particular description of exemplary embodiments, as illustrated in the accompanying drawings, in which like reference characters refer to the same parts throughout the different views. The drawings are not necessarily to scale, emphasis instead being placed upon illustrating embodiments.

[0032] [Figure 1]FIG. 1 is a flowchart of a method for determining properties (eg, continuum solvation model properties) of molecules in an environment, according to one embodiment. [Figure 2] FIG. 2 is a simplified block diagram of a system for determining properties (eg, continuum solvation model properties) of molecules in an environment, according to one embodiment. [Figure 3] FIG. 3 illustrates an exemplary workflow for determining the properties of molecules in an environment, according to one embodiment. [Figure 4] FIG. 4 is a simplified block diagram of a computer system for determining properties of molecules in an environment, according to one embodiment. [Figure 5] FIG. 5 is a simplified block diagram of a computer network environment in which embodiments of the present invention may be implemented. DETAILED DESCRIPTION OF THE INVENTION

[0033] A description of an exemplary embodiment follows.

[0034] Chemical reactions, i.e., chemical interactions / reactions, primarily occur in the liquid or solvent phase. There are known theoretical models that embed molecules in a solvent continuum. For example, existing continuum solvation models approximate the solvent as a dielectric continuum surrounding the solute molecule outside the molecular cavity. The cavity surface is approximated by segments, e.g., hexagons, pentagons, or triangles. Models that approximate the solvent as a dielectric continuum can be called dielectric continuum solvation models (DCSMs), including the widely used polarizable continuum model (PCM). Another example of a conventional DCSM is the conductor-like screening model (COSMO), which derives the continuum polarization charge caused by the solute's polarity from a scaled-conductor approximation [1] (the parenthetical numbers in this document refer to the reference list below in this specification). COSMO is one of the most widely applied computational methods for determining the electrostatic interactions between molecules and the solvent or liquid environment. Output data from COSMO or other known DCSM approaches (generally referred to herein as "COSMO information" or "COSMO-type information") may include, for example, segment-wise surface charge density and chemical potential. In general, each DCSM approximates the solvent effect in a similar way by approximating a dielectric continuum.

[0035] The COSMO information can then be used to calculate the chemical potential of the molecule in a solvent or solvent mixture. For example, this may be done via the existing COSMO-RS ("RS" stands for "real solvent") approach, which takes as input previously calculated and stored COSMO information. The COSMO-RS approach involves modeling a set of physicochemical interaction terms between a molecule and its liquid environment as a function of the pairwise charge density of a specific segment of the molecular cavity [2].

[0036] For example, chemical potential data generated by 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 pressure, and free energy of solvation. The COSMO-RS method was developed to provide a general predictive method without the need for system-specific tuning. The method is widely applied by academia and the chemical and pharmaceutical industries.

[0037] Currently used approaches to generate COSMO information rely either on quantum chemical calculations [1,2] partially complemented by calibrated machine learning [3], or on fragmentation approaches [4].

[0038] Quantum chemical calculations, for example, have the drawback of being computationally very demanding, especially for larger molecular systems. Typically, quantum chemical calculations are performed on high-performance computing (HPC) clusters for sets of related conformers. Therefore, routine application of quantum chemical methods is limited to setups by trained users with access to HPC clusters. These computationally intensive calculations limit quantum chemical methods to small molecules, with molecular weights significantly below 1 kilodalton (kDa). In contrast, many industrially relevant chemicals and biochemicals, such as polymers, surfactants, proteins, or biologicals, are in the 10-100 kDa range or even larger. Modern drug molecules also reach this molecular weight limit very quickly.

[0039] Although some existing quantum chemical methods, such as semi-empirical methods, have reduced computational demands, employing such methods leads to poor polarity distributions, which impair subsequent calculations of thermodynamic properties.

[0040] Fragmentation approaches do not adequately reflect the 3D conformational space and its effect on charge distribution. For example, in certain 3D configurations, the effect of conformational space on charge distribution is influenced by the formation of intramolecular hydrogen bonds.

[0041] Some embodiments described herein provide the advantage of reducing the computational effort required to calculate surface charge densities from days or weeks to seconds or even less, with accuracy comparable to that of quantum chemical calculations, for example. In certain embodiments, quantum chemical calculations are replaced by an efficient workflow including a set of machine learning procedures. Additionally, numerical artifacts resulting from the derivation of quantum chemical equations can be avoided or smoothed by the machine learning procedures implemented by some embodiments. For this reason, certain embodiments provide thermodynamic property calculations with increased predictive accuracy compared to existing approaches. The fast machine learning procedures of some embodiments take a molecular shape or a set of molecular shapes as input. Certain embodiments efficiently predict the charges, e.g., screening charges, and chemical potentials, of molecular surface segments by reflecting the local chemical environment of the segments in a particular molecule or conformer. In one embodiment, the segments are constructed in 3D space by employing any suitable cavity construction model known in the art, such as the COSMO FINE cavity construction model described in [5] (incorporated herein by reference in its entirety), or any other existing cavity construction method, for example, as outlined by [5] (incorporated herein by reference in its entirety).

[0042] Exemplary Method Embodiments FIG. 1 illustrates an example of such an exemplary method embodiment 100. Method 100 is a computer-implemented method for determining properties of a molecule in an environment. Method 100 begins in step 101 by constructing, e.g., in computer memory, one or more 3D structural models showing the positions of atoms of the molecule. In one embodiment, constructing the one or more 3D structural models in step 101 may be performed using any suitable technique known to those of skill in the art. For example, embodiments may construct the 3D structural models using known software platforms, such as BIOVIA® platforms such as Pipeline Pilot®, COSMOquick®, Materials Studio®, Discovery Studio®, or COSMOconf®, or other platforms such as TURBOMOLE. According to one embodiment, each of the one or more constructed 3D structural models corresponds to a respective conformer of the molecule. Various types of information and / or models may be used in constructing the one or more 3D structural models in step 101. For example, in one embodiment of method 100, one or more 3D structural models are constructed (101) based on representations of one or more of atom types, coordinates, and chemical connectivity. Chemical connectivity refers to the way atoms are spatially bonded to one another. Furthermore, in some embodiments, the models may be constructed based on measurements / observations of real-world molecules, such that the resulting models reflect the measured / observed properties of the real world. According to another exemplary embodiment of method 100, constructing (101) the one or more 3D structural models includes employing one or more of a rule-based geometric model, a force field, and a quantum-chemically derived geometric model, among other non-limiting examples. In one embodiment, the quantum-chemically derived geometric model may include, for example, a tight-binding model, a semi-empirical model, a geometric model derived from density functional theory, or any combination thereof.

[0043] To continue, method 100 determines properties of the molecule in the environment in step 102. The properties of the molecule are determined in step 102 by, for each of the constructed one or more 3D structural models (from step 101), (i) generating, for example, in computer memory, a surface model representing the environment, the surface model including a plurality of segments, the generated surface model defining relationships between indicated positions of atoms of the 3D structural model and the plurality of segments, and (ii) using a machine learning model to predict the electrical charge (e.g., 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, generating the surface model in step 102 can be performed using any suitable technique known to those of skill in the art.

[0044] As noted, method 100 is computer-implemented, such that the functionality and effective operations, e.g., constructing (101) and determining (102), may be implemented automatically by one or more digital processors. Furthermore, method 100 may be implemented using any computer device or combination of computing devices known in the art. Among other examples, method 100 may be implemented using computer system 440, described herein below in connection with FIG. 4, and computer network environment 550, described below in connection with FIG. 5.

[0045] At step 102, method 100 may predict the charge using a first machine learning model and predict the chemical potential using a second machine learning model. In one embodiment of method 100, at step 102, the first machine learning model predicts the charge of each of the plurality of segments based on the 3D structural model and the generated surface model. In such an embodiment, the 3D structural model (and / or its features) and the generated surface model (and / or its features) are provided as input to the first machine learning model, and the first machine learning model is configured to output the charge of each segment in response to the input. Similarly, in one embodiment, at step 102, the second machine learning model predicts the chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model. In such an embodiment, the 3D structural model (and / or its features) and the generated surface model (and / or its features) are input to the second machine learning model, and the second machine learning model is configured to output the chemical potential of each segment in response to the input.

[0046] In embodiments of method 100, determining a property of the molecule in the environment in step 102 further includes predicting an energy corresponding to the 3D structural model based on the 3D structural model and the generated surface model using a supplemental machine learning model. Such embodiments input the 3D structural model (and / or features thereof) and the generated surface model (and / or features thereof) into the supplemental machine learning model, and the supplemental machine learning model is configured to output an energy in response to the input.

[0047] In one embodiment of method 100, predicting an electrical charge (e.g., charge) and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model (102) using a machine learning model includes deriving input feature data based on the 3D structural model. In one such embodiment, predicting an electrical charge and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model and the derived input feature data using a machine learning model. According to one such embodiment, the derived input feature data includes an indication of one or more of atom type, interatomic distance, atom-segment distance, bond type, bond angle, torsion angle, formal charge, 3D atom position, and features specific to the atom type.

[0048] In one embodiment of method 100, predicting the charge (e.g., electric charge) and chemical potential of each of the plurality of segments based on the 3D structural model and the generated surface model (102) using the 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 exemplary embodiment, the determined overall formal charge is the predicted charge of the plurality of segments.

[0049] In one embodiment of method 100, generating 102 a surface model representing the environment includes employing a cavity construction model. An embodiment may employ any suitable cavity construction model known in the art, such as the COSMO FINE cavity construction model described in [5], or any other existing cavity construction method outlined by [5]. In one embodiment of method 100, the machine learning model used in step 102 includes a neural network. Furthermore, 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 and may instead employ any suitable activation function known in the art. While neural networks are discussed herein, those skilled in the art will recognize that embodiments of method 100 are not limited to such technology. Rather, embodiments of method 100 may utilize any suitable known machine learning or statistical learning method to determine the charge and chemical potential in step 102.

[0050] According to one embodiment, the method 100 further includes training a machine learning model based on the training dataset. In one embodiment, the machine learning model includes a neural network, and training the machine learning model based on the training dataset includes training the neural network by iteratively updating one or more network weights of the neural network based on the training dataset. Furthermore, in yet another embodiment, iteratively updating one or more network weights of the neural network based on the training dataset includes employing one or more of an adaptive moment estimation (Adam) solver algorithm and an early stopping algorithm. It should be noted that embodiments are not limited to the specific algorithms listed above; instead, any suitable algorithm known in the art may be employed. According to an exemplary embodiment, the training dataset includes data for one or more of exemplary molecules, exemplary conformers, exemplary segments, exemplary segment charges, exemplary segment chemical potentials, and exemplary continuum model energies. In yet another embodiment, the training dataset may include a large collection of COSMO files covering organic chemical space, for example, across solvents, industrial chemicals, pharmaceuticals, ion protonation states, ions, and ionic liquids (cations and anions, including multivalent ions).

[0051] According to one embodiment, method 100 further includes receiving one or more user requirements. In one such exemplary embodiment, for each candidate molecule of the plurality of candidate molecules, method 100 then performs property building (101) and determination (102) to select 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 include, for example, a molecular input structure or ion input structure, an input representation, and / or a connection table.

[0052] Additionally, in another embodiment of method 100, properties of candidate molecules are determined from real-world measurements of the candidate molecules. These measured properties are then used in step 101 to construct computer-based models of the candidate molecules with said properties. These models (which reflect the real-world measured / observed properties of the candidate molecules) are then used in step 102 to determine the charge and chemical potential of each candidate molecule. Given candidate molecules that meet desired criteria can then be selected for use in real-world applications such as developing research plans, optimizing formulations to improve product properties in desired ways, designing sustainable polymers for recycling and other processes involving polymers, identifying appropriate pharmaceutical excipients in the context of drug development, focusing a set of experiments, and reducing material waste by improving experimental efficiency, to name a few.

[0053] Particular embodiments of method 100 may use 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., electrical charge), chemical potential, energy, and any other data / values ​​described herein. Some embodiments may also generate and store data in the cloud, such as, for example, the 3DEXPERIENCE® platform. However, embodiments are not limited to a particular file format / structure or cloud platform; instead, any suitable file format / structure or cloud platform known in the art may be used.

[0054] Exemplary System Embodiments FIG. 2 is a simplified block diagram of a system 220 for determining properties of molecules in an environment, according to one embodiment.

[0055] As shown in FIG. 2, in one embodiment, 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.

[0056] In an embodiment, data source 221 may include molecular data, i.e., data about the molecule whose properties are being determined, as well as environmental data, i.e., data about the environment of the molecule. According to one implementation, data source 221 may be provided by a user of system 220. In an exemplary embodiment, data source 221 may be used, for example, as input to molecular model generator 222 and surface model generator 223.

[0057] According to an embodiment of system 220, using data about molecules provided by data source 221, molecular model generator 222 may construct one or more 3D structural models showing the positions of atoms of the molecules.

[0058] Similarly, in one aspect, using data about the environment provided by data source 221, surface model generator 223 may generate a surface model representing the environment for each 3D structural model of interest constructed by molecular model generator 222. In such an embodiment, each surface model includes a plurality of segments, and each generated surface model defines relationships between indicated positions of atoms of the 3D structural model and the plurality of segments.

[0059] According to one implementation, based on each 3D structural model of the one or more 3D structural models constructed by the model generator 222 and each surface model generated by the surface model generator 223, the machine learning model 224 may 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, where the first machine learning model predicts the charge and the second machine learning model predicts the chemical potential. Furthermore, according to one aspect, the machine learning model 224 may include a supplemental machine learning model that predicts the energy corresponding to the 3D structural model based on the 3D structural model and the generated surface model. In one embodiment, the machine learning model 224 may include a neural network.

[0060] In an exemplary embodiment, properties, such as charge, chemical potential, and / or energy, predicted by machine learning model 224 may be output to storage device 225. Furthermore, according to one embodiment, once recorded in storage device 225, such output data may be used for further processing, such as calculation of thermodynamic properties, among other examples.

[0061] It should be noted that the system 220 may implement any embodiment described herein, such as the method 100 described herein above with respect to FIG. 1, to determine the properties of molecules in an environment.

[0062] Example Workflow 3 illustrates an exemplary workflow 330 for determining the properties of molecules in an environment, according to one embodiment. Note that workflow 330 may implement any of the embodiments described herein, such as method 100 and system 220 described herein above with respect to FIGS. 1 and 2, respectively, to determine the properties of molecules in an environment.

[0063] In one embodiment, an input structure is obtained at step 331 of workflow 330, e.g., in response to a user action. According to an exemplary embodiment, the input structure may be the structure of a molecule in its environment whose properties are to be determined. Figure 3 shows an exemplary two-dimensional (2D) molecular structure 340 that may be obtained at step 331 of workflow 330.

[0064] Next, according to one embodiment, 3D structure generation of the molecule can occur in step 332. FIG. 3 shows the constructed 3D structure 341a of the 2D molecular structure 340. In one implementation, optional conformer generation of the molecule can occur in step 333. For example, FIG. 3 shows additional conformers 341b-c. According to one embodiment, the 3D structure generation in step 332, and the optional conformer structure generation in step 333, can include, for example, constructing, in computer memory, one or more 3D structural models (341a-c) that indicate the positions of atoms of the molecule. 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 exemplary embodiment, each of the constructed one or more 3D structural models (341a-c) corresponds to a respective conformer (340) of the molecule.

[0065] Continuing with reference to FIG. 3 , according to one embodiment, workflow 330 determines properties of molecules in an environment. In one aspect, the properties of the molecules are determined by, for each of one or more constructed 3D structural models 341 a-c (from steps 332 and, optionally, 333), (i) generating, for example, in computer memory, a surface model (342 a-c) representing the environment, where surface models 342 a-c include a plurality of segments (which may be, for example, hexagonal tiles such as tile 345 shown in FIG. 3 ), and generated surface models 342 a-c define relationships between indicated positions of atoms of 3D structural models 342 a-c and the plurality of segments; and (ii) predicting, for example, in step 335, using a machine learning model, the charge and chemical potential of each of the plurality of segments based on 3D structural models 342 a-c and generated surface models 342 a-c. According to an exemplary embodiment, generating the surface model may include performing cavity construction and tiling into segments in step 334. It should be noted that generating the surface model, including, for example, cavity construction and tiling into segments in step 334, may be performed using any suitable technique known to those skilled in the art. According to another exemplary embodiment, when visually presented, as shown, for example, by region 343 in Figure 3, individual segments may be shaded / colored to distinguish segments that map to or correspond to specific atoms of the molecule.

[0066] According to one embodiment, step 335 of workflow 330 may include using one or more machine learning models to predict various charge and chemical potentials for each segment and energy of each conformer. For example, an implementation may use a first machine learning model to predict charge and a second machine learning model to predict chemical potential. According to one aspect, a supplemental machine learning model may be used to predict energy. In one embodiment, when visually presented, individual segments may be shaded / colored to distinguish between segments with negative, positive, or neutral predicted charges, as shown, for example, by region 344 in FIG. 3 .

[0067] Continuing with reference to Figure 3, in an exemplary embodiment, a processing option may involve geometric optimization in step 336 of workflow 330 by evolving the 3D structural model and its optimization towards 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 exemplary embodiment, after properties, such as charge, chemical potential, and / or energy, are predicted by one or more machine learning models, the property values ​​may be written or saved to one or more output files 346 in step 337 of workflow 330. Furthermore, according to one embodiment, after being saved in step 337, such output data may be used for further processing in step 338, such as calculating thermodynamic properties and / or generating thermodynamic predictions 347, among other examples.

[0069] Molecular or macromolecular geometry Certain embodiments may, for example, in step 101 of method 100, construct one or more 3D structural models for a particular molecule of interest.

[0070] Some embodiments may take as input a molecular geometry (a single conformer) or a set of molecular geometries (a set of conformers). Conformers may include, for example, tautomers and / or protonation states. Molecular geometries may be defined by individual atomic coordinates and atom types.

[0071] Additionally, in certain embodiments, any one or more of the following operations may be performed to construct a model showing the positions of atoms: (i) a conformer search to generate a respective set of 3D structural models of conformers; (ii) a tautomer search to generate, for each tautomer, a respective set of 3D structural models of a single conformer or set of conformers; (iii) a search for relevant protonation states to generate, for each protonation state, a respective set of 3D structural models of a single conformer or set of conformers; (iv) generating different substitution patterns; and (v) generating different mutant or variant structures.

[0072] According to one embodiment, constructing the 3D structural model includes receiving one or more molecular geometries from any source known in the art, such as industry-standard rapid 3D structure generators, X-ray structures, force fields, tight-binding approximation methods, or semi-empirical or density functional theory-based calculations.

[0073] For example, in one embodiment, BIOVIA® Pipeline Pilot® or BIOVIA® COSMOquick® software may be utilized to generate molecular geometries and sets of conformers from chemical connectivity information in molecular files stored, for example, as SDF (Structure Data File), MOL (Molfile), PDB (Protein Data Bank), or related molecular file formats, or as molecular input line representations, SMILES (Simple Molecular Input Line Entry System), among other examples. BIOVIA® Materials Studio® or BIOVIA® Discovery Studio® molecular builders, which utilize force fields to generate reliable 3D geometries for molecules and macromolecular systems, such as CHARMM (Chemistry at Harvard Macromolecular Mechanics) or COMPASS (Condensed-phase Optimized Molecular Potentials for Atomistic Simulation Studies) force fields, may also be utilized by some embodiments to build 3D structural models. Additionally, BIOVIA® COSMOconf® and TURBOMOLE support complementary methods for generating 3D structures, e.g., RDKit-related force fields, MOPAC (Molecular Orbital Package) semi-empirical methods, or xTB semi-empirical extended tight-binding approximations, which may be employed by certain embodiments. In some embodiments, density functional theory-based calculations may be performed by quantum chemistry packages, e.g., TURBOMOLE, to generate 3D structures. It should be noted that embodiments are not limited to the particular software platforms or file formats described herein; instead, any suitable software platform or file format known in the art may be used.

[0074] Molecular 3D geometries can be generated, for example, by the ETKDG (Experimental-Torsion basic Knowledge Distance Geometry) method of the RDKit cheminformatics library. Other known sources and techniques for molecular shapes may be used in addition to those discussed herein.

[0075] It should also be noted that in some embodiments, atomic coordinates and atom types are used to determine molecular properties, e.g., properties of continuum solvation models. Unlike existing approaches, these embodiments do not explicitly rely on, e.g., bond information, connectivity information, functional groups, fingerprints, segment-specific information, molecular surface-specific information, energies, or potentials.

[0076] Construction of solvent accessible face segments Certain embodiments may generate a surface model representing an environment, such as a continuum solvent environment, for example, at step 102 of method 100. The generated surface model may include multiple segments. Furthermore, the generated surface model may define a relationship between the positions of indicated 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) within a tiled cavity, where the tiled cavity represents a 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 implementation, the multiple segments of a given surface model may represent a solvent accessible surface (SAS) specific to the conformer.

[0077] In principle, the electrostatic principles underlying DCSMs are accurate. In reality, the electron densities of solute and solvent molecules overlap, so there is no well-defined surface separating the solute from the solvent environment. Therefore, cavity definition and molecule-specific cavity construction are crucial steps.

[0078] In some embodiments, the solvent accessible surface segments may be constructed in 3D space. Such embodiments may employ any suitable cavity construction method known in the art, such as the existing cavity construction method outlined by [5].

[0079] According to one embodiment, for example, in step 102 of method 100, the surface model may be generated using the COSMO FINE cavity construction model [5]. The marching tetrahedron algorithm used by the FINE model provides a technique for triangulation to arrive at surface segments. The FINE model further utilizes an iso-density cavity construction algorithm based on atom-type-specific COSMO radii, which results in smooth, fully paved cavities in molecular shapes. This is particularly important for the COSMO-RS model, which uses the screening charge density on the surface as the primary descriptor for defining intermolecular interactions. Details of the procedure are provided in [5].

[0080] Machine learning models for predicting charge and chemical potential For each surface segment of the generated surface model, certain embodiments calculate the charge (e.g., electric charge) and chemical potential as typical inputs for the COSMO-RS thermodynamic model. In existing methods, the charge distribution and chemical potential are determined via quantum chemistry calculations. In contrast, some embodiments use one or more machine learning models to predict the charge and chemical potential of the segment. Compared to existing approaches, certain embodiments therefore provide a significant speed increase, by several orders of magnitude. Some embodiments reduce the computation time per central processing unit (CPU) core to a few seconds, even at the most accurate level of consideration. Similar to quantum chemistry, the machine learning model-based methods of certain embodiments reflect the specific atomistic environment of the segment in a specific spatial molecular arrangement of neighboring atoms.

[0081] Some embodiments may apply one or more machine learning models to predict target properties, i.e., segmental charge and segmental chemical potential. In certain embodiments, quantum chemical information as input for training the model may be taken or constructed from molecular geometries optimized at the density functional level, e.g., from the B88-VWN-P86 functional and the def-TZVP (valence triple zeta polarization with diffuse functions) basis set, by subsequent single-point calculations using, for example, TURBOMOLE software and the def2-TZVPD (valence triple zeta polarization with diffuse functions) basis set, where the activated scaled-conductor approximation has an infinite permittivity ε=∞. The segmental charge may include, for example, a screening charge or charge density. According to some embodiments, the two quantum chemical levels may be abbreviated as BP-TZVP and BP-TZVPD-FINE in the name of the BIOVIA® COSMOtherm® software. Chemical potential in this context refers to the segmental chemical potential in response to polarity changes.

[0082] One or more machine learning models of certain embodiments may be trained against a large collection of COSMO files that serve as a training dataset. In one aspect, the training dataset may include one or more of the following data: 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 test sets. The COSMO file collection used to train the machine learning models of certain embodiments covers organic chemical space, spanning, for example, solvents, industrial chemicals, pharmaceuticals, ionic protonation states, ions, and ionic liquids (cations and anions, including multivalent ions). According to some embodiments, the training dataset may include a set of conformers containing representatives of both intramolecular hydrogen bonds and open (intermolecular) hydrogen bond candidates. In one aspect, the COSMO file collection used for the training dataset may include approximately 16,000 compounds, represented by approximately 65,000 conformers per quantum chemical level. According to one implementation, each conformer contains approximately 10 conformers, with specific values ​​for segment area, charge, and potential, depending on the size and quantum chemical level of the molecule represented by the COSMO file. 3 In some embodiments, the average number of conformers in the entire dataset can be approximately 10. 8 Only a percentage of segments are used in the training process. According to one such embodiment, on average, only 3% of segments related to carbon and hydrogen atoms may be used for training and validation (thus leaving 97% for the test set), and only 25% of segments may be related to nitrogen, oxygen, and fluorine atoms (thus leaving 75% for the test set). In one aspect, no threshold is applied to other atom types. According to one embodiment, during training and validation, a percentage of 0.9 of all used segments may be used for training, while a percentage of 0.1 may be reserved for validation.

[0083] In one embodiment, the segments in the COSMO file already reflect the out-of-range charge correction, so no explicit out-of-range charge correction is necessary. According to one implementation, after all segments are processed, a residual charge correction may be performed to arrive at an overall formal charge. Certain embodiments may also apply corrections to the predicted charges using various techniques. These techniques may include, but are not limited to, (i) charge alignment based on formal charge and overall predicted conformer, and (ii) out-of-range charge correction, if necessary.

[0084] For example, in step 102 of method 100, the one or more machine learning models applied by certain embodiments may include, but are not limited to, an artificial neural network. In one embodiment, the artificial neural network architecture may include one or more hidden layers, which may begin with an input layer, a dense hidden layer consisting of, for example, 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. Furthermore, neural networks according to embodiments may include additional layers, such as convolutional layers. According to one aspect, the neural network may employ a rectified linear unit (ReLU) activation function, an adaptive moment estimation (Adam) solver algorithm, a 10 -8, tolerances, and early stopping algorithms may be used. Furthermore, it should be noted that 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 implementation, the hyperparameters of the artificial neural network architecture may be subjected to further optimization. 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 and scaling to unit variance of the training set. While neural networks are discussed herein, those skilled in the art will recognize that embodiments are not limited to such techniques; rather, any suitable known machine learning or statistical learning method may be used.

[0085] In some embodiments, one or more machine learning models may use the same types of information for input features as those used by traditional quantum chemical approaches. These features may include, for example, atom types, interatomic distances, atomic segment distances, bond types, bond angles, torsion angles, formal charges, 3D atom positions, and various atom-type-specific features. According to one embodiment, the 3D atom positions may be rotationally invariant representations of the 3D positioning of atoms relative to one another. Furthermore, implementations may apply arbitrary cutoff radii. Certain embodiments offer the advantage of not having to rely on features such as molecular topology or connectivity, molecular fragments, chemical functional groups, and fingerprints based on classical cheminformatics, among others. By avoiding such dependencies, some embodiments may determine very general chemical descriptions of, for example, possible charge distributions. Certain embodiments are not required to consider these features, but may optionally use them. In one aspect, there is no direct cutoff radius, but up to 24 nearest atoms may be considered the chemical environment of a particular segment. According to one embodiment, atom types are not directly coded, but rather indirectly coded by their quantum chemical characteristics, such as electron affinity, ionization potential, allowed orbital configurations, atom-type-specific radii, etc. This technique further increases the applicability domain of one or more machine learning models according to some embodiments. However, in certain embodiments, atom types may be directly coded. All atom types up to Radon (excluding atomic number 86, i.e., g-orbitals), as well as ionic structures, are supported by the construction of one or more machine learning models according to embodiments.

[0086] Machine learning models for energy forecasting Certain embodiments may further use a supplemental machine learning model (i.e., a machine learning model in addition to the one or more models used in step 102 to predict charges and chemical potentials) to predict energies corresponding to the 3D structural model based on the 3D structural model and the generated surface model. As described herein, in embodiments, the 3D structural model may correspond to a conformer of the molecule.

[0087] In some embodiments, the predicted energy may be one or more of: (i) total molecular or ionic energy; (ii) dielectric energy, e.g., dielectric energy within a conductor-like dielectric continuum; and (iii) gas-phase energy. According to one aspect, the same COSMO file collection (discussed herein under the heading "Machine Learning Models for Predicting Charge and Chemical Potential") may be used to train a supplemental machine learning model for each target property and calculation level (e.g., BP-TZVP and BP-TZVPD-FINE). Similarly, in one implementation, 0.9 percent of each target property may be used for training, and 0.1 percent may be reserved for validation. Some embodiments may employ the same artificial neural network architecture and applied hyperparameters as described above for the supplemental machine learning model. However, in one aspect, the neural network used to predict energy (e.g., total molecular energy, ionic energy, dielectric energy, or gas-phase energy) may include a first hidden layer consisting of 512 nodes instead of 256 for the total energy and gas-phase energy. Certain embodiments may use similar input features, such as atom type, distance, angle, torsion, and formal charge, with complementary machine learning models to predict energies. According to one implementation, atom types are not directly coded, but rather indirectly coded by their quantum chemical characteristics, such as electron affinity, ionization potential, allowed orbital configurations, and atom-type-specific radii. Some embodiments may predict energies (e.g., conductor or gas-phase states) 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), among other examples.

[0088] While neural networks are discussed herein, those skilled in the art will recognize that embodiments are not limited to such techniques; rather, any suitable known machine learning or statistical learning method may be used by embodiments to predict energies. Published machine learning models include, for example, deep learning architectures for molecules and materials described by [6].

[0089] Certain embodiments provide single-point results for predefined molecular geometries. In some embodiments, to predict energies, e.g., total molecular energies or ionic energies, such embodiments can also be extended to perform geometry optimization in a condensed environment by varying the 3D structural model and its optimization toward an energy minimum.

[0090] Writing the Output File Certain embodiments may collect results, including predicted segment charges, segment-wise chemical potentials, and / or energies, as described herein above, and write the results to an output file, e.g., a dedicated conformer-specific COSMO file that stores information for all COSMO types. For example, the predicted charges and chemical potentials determined in 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 separately running segment construction (described herein above under the heading "Construction of Solvent Accessible Face Segments") and all predictions (described herein above under the heading "Machine Learning Models for Predicting Charges and Chemical Potentials for Predicting Energies") for each specific conformer. Certain embodiments may also generate and store output files, e.g., COSMO files, in the cloud, e.g., the 3DEXPERIENCE® platform. Embodiments are not limited to a particular output file format or cloud platform; instead, any suitable output file format or cloud platform known in the art may be used.

[0091] Applying the output file in further calculations To demonstrate application of the novel techniques described herein, output files from some embodiments, e.g., COSMO files, may be used as input for BIOVIA® COSMOtherm® software. The software may calculate fluid thermodynamic properties via the COSMO-RS method. Certain embodiments may also perform calculations via applications provided by, for example, the 3DEXPERIENCE® platform (e.g., BIOVIA® Virtual Bench®), or any suitable platform known to those skilled in the art.

[0092] Embodiments can successfully predict thermodynamic equilibrium properties, such as solubility, partition coefficient, and / or liquid density. Furthermore, results show that such properties determined using embodiments closely match known, experimentally determined values. Predictions for multiple chemical substances based on a set of conformers (e.g., a set of COSMO files) generated by novel machine learning techniques according to embodiments match well with predictions generated by traditional quantum chemistry calculations (e.g., density functional theory). Furthermore, results from machine learning methods according to embodiments exceed those generated by existing approaches.

[0093] In addition to thermodynamic property predictions, other examples of further calculations 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 profiles, energies, σ-moments, etc.), software applications involving quantum chemistry and / or materials science (e.g., TURBOMOLE and BIOVIA® Materials Studio®), and prediction of biological and / or biochemical properties (e.g., BIOVIA® Discovery Studio®).

[0094] advantage Embodiments determine the properties of molecules in an environment and offer numerous advantages.

[0095] For example, the rapid machine learning procedures of embodiments reduce the time required to build and calculate COSMO information for a complete set of conformers from days or weeks to seconds or even less.

[0096] As another exemplary advantage, embodiments do not require an HPC cluster. Embodiments facilitate the automation of high-throughput screening predictions and the democratization of solutions for thermodynamic equilibrium calculations in the liquid phase. Cloud-based platforms, such as Dassault Systemes' 3DEXPERIENCE® platform, benefit from results provided by embodiments that are essentially instantly available. In this way, embodiments reduce both costs and carbon emissions; further, embodiments increase the platform's desirability to customers.

[0097] Furthermore, as yet another exemplary advantage, embodiments can extend the applicability of COSMO and COSMO-RS to large molecular systems, such as biomolecules and polymers, among other examples. Embodiments can also be used to efficiently guide experiments. Furthermore, embodiments provide polymer solubility predictions that can be leveraged to design sustainable polymers, thereby improving real-world processes that utilize such polymers, such as recycling processes. Because excipients are typically polymer-based, embodiments can additionally enhance pharmaceutical virtual screening 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 necessary methods.

[0098] Advantageously, embodiments also improve the accuracy of thermodynamic property calculations by avoiding numerical artifacts arising from the derivation of quantum chemical equations through efficient machine learning procedures.

[0099] Computer Support Embodiments may be implemented in existing software and computer-aided design and computer-aided engineering platforms. For example, embodiments may be implemented using the features and functionality of 3DS BIOVIA® software.

[0100] FIG. 4 is a simplified block diagram of a computer-based system 440 that can be used to determine the properties of molecules in an environment, according to any of the various embodiments of the invention described herein. System 440 includes a bus 443. Bus 443 serves as an interconnection between the various components of system 440. Connected to bus 443 is an input / output device interface 446 for connecting various input and output devices, such as a keyboard, mouse, touchscreen, display, and speakers, to system 440. CPU 442 is connected to bus 443 and provides for the execution of computer instructions. Memory 445 provides volatile storage for data used to execute the computer instructions. Storage 444 provides non-volatile storage for software instructions of an operating system (not shown). System 440 also includes a network interface 441 for connecting to any of a variety of networks known in the art, including wide area networks (WANs) and local area networks (LANs).

[0101] It should be appreciated that the exemplary embodiments described herein may be implemented in many different ways. In some instances, the various methods and systems described herein each utilize computer system 440, or the computer system described herein below in connection with FIG. systemThe system 440 may be implemented by a physical, virtual, or hybrid general-purpose computer, such as a computer network environment, such as 440. The computer system 440 may be converted into a machine that performs the methods described herein, for example, by loading software instructions implementing the method 100 into either the memory 445 or the non-volatile storage 444 for execution by the CPU 442. Those skilled in the art should further appreciate that the system 440 and its various components may be configured to perform any embodiment or combination of embodiments described herein. Furthermore, the system 440 may implement the various embodiments described herein utilizing any combination of hardware, software, and firmware modules operably coupled internally or externally to the system 440.

[0102] 5 illustrates a computer network environment 550 in which embodiments of the present invention may be implemented. In computer network environment 550, server 551 is linked to clients 553a-n via communication network 552. Environment 550 may be used to enable clients 553a-n, alone or in combination with server 551, to perform any of the embodiments described herein. As non-limiting examples, computer network environment 550 may provide cloud computing embodiments, software as a service (SaaS) embodiments, etc.

[0103] The 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 configured to enable a processor to load the software, or a subset of its instructions. The processor is then configured to execute the instructions to operate a device or cause it to operate in a method described herein.

[0104] Furthermore, firmware, software, routines, or instructions may be described herein as performing certain operations and / or functions of a data processor, although it will be understood that such descriptions contained herein are merely for convenience and that such operations actually result from a computing device, processor, controller, or other device executing the firmware, software, routines, instructions, etc.

[0105] It will be understood that the flow diagrams, block diagrams, and network diagrams may include more or fewer elements, may be arranged differently, or may be represented differently, but it will also be understood that a particular implementation may implement the block diagrams and network diagrams, and the number of block diagrams and network diagrams illustrating the implementation of an embodiment, in a particular way.

[0106] Accordingly, further embodiments may also be implemented in various computer architectures, physical computers, virtual computers, cloud computers, and / or some combination thereof, and therefore the data processors described herein are intended to be illustrative only and not limiting of the embodiments.

[0107] The teachings of all patents, published applications, and references cited herein are incorporated by reference in their entirety.

[0108] While exemplary embodiments have been particularly shown and described, those skilled in the art will understand that various changes in form and details can be made therein without departing from the scope of the embodiments encompassed by the appended claims.

Claims

1. 1. A computer-implemented method for determining a property of a molecule in an environment, said method comprising: constructing one or more three-dimensional (3D) structural models showing the positions of the atoms of the molecule; determining, for each 3D structural model of the constructed one or more 3D structural models, the properties of the molecule in the environment; generating a surface model representing the environment, the surface model including a plurality of segments, the generated surface model defining relationships between the indicated positions of the atoms of the 3D structural model and the plurality of segments; and using 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 and the generated surface model, wherein the 3D structural model and the generated surface model are provided as inputs to the machine learning model, and the machine learning model is configured to output a charge and a chemical potential of each of the plurality of segments in response to the inputs.

2. the machine learning model includes a first machine learning model and a second machine learning model, and predicting the charge and the chemical potential of each segment of the plurality of segments based on the 3D structural model and the generated surface model is predicting a charge of each segment of the plurality of segments based on the 3D structural model and the generated surface model using the first machine learning model; and using the second machine learning model to predict the chemical potential of each segment of the plurality of segments based on the 3D structural model and the generated surface model.

3. determining, for each 3D structural model of the constructed one or more 3D structural models, the properties of the molecule in the environment; 10. The method of claim 1, further comprising: using a complementary machine learning model to predict an energy corresponding to the 3D structural model based on the 3D structural model and the generated surface model.

4. The method of claim 1 , wherein each 3D structural model of the constructed one or more 3D structural models corresponds to a respective conformer of the molecule.

5. The method of claim 1 , wherein the machine learning model comprises a neural network.

6. 6. The method of claim 5, wherein the neural network includes one or more hidden layers, and the neural network is configured to employ activation functions 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 regularized linear activation function and a softmax function.

8. The method of claim 1 , further comprising training the machine learning model based on a training dataset.

9. the machine learning model includes a neural network, and training the machine learning model based on the training dataset 9. The method of claim 8, comprising training the neural network by iteratively updating one or more network weights of the neural network based on the training data set.

10. 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. 9. The method of claim 8, wherein the training data set includes 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. using the machine learning model to predict the charge and the chemical potential of each segment of the plurality of segments based on the 3D structural model and the generated surface model; deriving input feature data based on the 3D structural model; and using the machine learning model to predict the charge and the chemical potential of each segment of the plurality of segments based on the 3D structural model and the generated surface model and the derived input feature data.

13. 13. The method of claim 12, wherein the derived input feature data comprises representations of one or more of atom types, interatomic distances, atomic segment distances, bond types, bond angles, torsion angles, formal charges, 3D atom positions, and atom type specific features.

14. receiving one or more user requirements; performing said constructing and said determining said properties for each candidate molecule of a plurality of candidate molecules; 10. The method of claim 1, further comprising: selecting the given molecule from among the plurality of candidate molecules based on the determined characteristics of the given molecule and the received one or more user requirements.

15. using the machine learning model to predict the charge and the chemical potential of each segment of the plurality of segments based on the 3D structural model and the generated surface model; correcting residual charge on one or more of the plurality of segments; 2. The method of claim 1, comprising: determining an overall formal charge of the plurality of segments based on the corrected one or more residual charges of the plurality of segments, wherein the determined overall formal charge is the predicted charge of the plurality of segments.

16. 10. 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 representations of one or more of atom types, coordinates, and chemical connectivity.

17. 2. 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 said generating said surface model representing said environment comprises employing a cavity construction model.

19. 1. A computer-based system for determining properties of molecules in an environment, said system comprising: a processor; a memory having computer code instructions stored thereon, said processor and said memory using said computer code instructions to cause said system to: constructing one or more three-dimensional (3D) structural models showing the positions of the atoms of the molecule; determining, for each 3D structural model of the constructed one or more 3D structural models, the properties of the molecule in the environment; generating a surface model representing the environment, the surface model including a plurality of segments, the generated surface model defining relationships between the indicated positions of the atoms of the 3D structural model and the plurality of segments; and a memory configured to cause the system to use 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, wherein the 3D structural model and the generated surface model are provided as inputs to the machine learning model, and the machine learning model is configured to output the charge and chemical potential of each of the plurality of segments in response to the inputs.

20. A computer-readable medium containing 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 showing the positions of the atoms of the molecule; determining, for each 3D structural model of the constructed one or more 3D structural models, a property of the molecule in an environment; generating a surface model representing the environment, the surface model including a plurality of segments, the generated surface model defining relationships between the indicated positions of the atoms of the 3D structural model and the plurality of segments; A computer-readable medium for causing a computer to: 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 and the generated surface model, wherein the 3D structural model and the generated surface model are provided as inputs to the machine learning model, and the machine learning model is configured to output a charge and a chemical potential of each of the plurality of segments in response to the inputs.

Images