A method and system for high-throughput screening of thermodynamically stable molecular crystal forms
By constructing a lattice energy database and a deep graph neural network model, the problem of low efficiency and low accuracy in the screening of thermodynamically stable crystal forms in existing technologies has been solved, achieving efficient and accurate crystal form screening and reducing computational costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN UNIV
- Filing Date
- 2026-03-13
- Publication Date
- 2026-07-03
Smart Images
Figure CN122337360A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of molecular crystal technology, specifically relating to a high-throughput screening method and system for thermodynamically stable molecular crystal forms. Background Technology
[0002] In the pharmaceutical industry, more than half of drugs exist in crystalline form, and over 85% of the production process involves crystallization. Crystallization is the final and crucial step in the preparation of solid active pharmaceutical ingredients (APIs), directly determining the drug's core properties such as crystal form, crystal habit, particle size distribution, and purity. Because drug molecules commonly exhibit polymorphism, different crystal forms differ significantly in solubility, dissolution rate, stability, and hygroscopicity, thus affecting bioavailability and clinical efficacy. Therefore, screening for the thermodynamically most stable crystal form is a key issue in drug development. Currently, traditional methods for crystal form screening primarily rely on experimental trial and error. This method not only consumes significant manpower, resources, and time but also carries the risk of overlooking advantageous crystal forms, making it difficult to meet the demands of efficient research and development. With the improvement of computer performance, CSP methods based on first-principles calculations, such as density functional theory (DFT), have successfully predicted the polymorphism of many molecules and can accurately calculate the energy of molecular crystals. However, due to the large number of crystal space groups (up to 230) and the complexity of molecular conformations, the polymorphic search space of CSP is extremely large. First-principles crystal structure optimization and energy assessment processes are time-consuming, computationally expensive, and severely limit its high-throughput applications. In recent years, machine learning (ML), with its efficient data processing capabilities and low cost, has been introduced into the field of crystal form prediction to improve computational efficiency. However, existing models are mostly designed for atomic crystals and struggle to effectively characterize the complex long-range dispersion interactions and diverse packing patterns in molecular crystals, resulting in limited generalization ability. Overall, existing methods still present a significant trade-off between efficiency and accuracy, making it difficult to achieve efficient and accurate screening of thermodynamically stable molecular crystal forms. Summary of the Invention
[0003] In view of this, the present invention provides a high-throughput screening method and system for thermodynamically stable molecular crystal forms to solve the problems of inefficiency and inaccuracy in the existing technology for screening the most thermodynamically stable crystal forms.
[0004] The technical solution adopted in this invention is as follows: A high-throughput screening method for thermodynamically stable molecular crystal forms includes: Step S1: Calculate the lattice energy of molecular crystals using quantum mechanical methods that consider dispersion interactions, and establish a lattice energy database that includes molecular crystal structures and their corresponding lattice energies. Step S1 specifically includes: Select the crystal structure of the target molecule and obtain the crystal's unit cell and molecular information; Density functional theory was employed, combined with a range-separated self-consistent screening method to handle dispersive interactions, and the lattice energy of the crystal structure was calculated using quantum mechanics. The calculated lattice energy of the crystal structure was achieved through the following equation: .
[0005] in, The total energy of the unit cell. The number of molecules in a unit cell. This represents the energy of a single gas molecule.
[0006] The calculated lattice energy data are preprocessed to establish a database containing different molecular crystal structures and their corresponding lattice energies.
[0007] Step S2: Based on the lattice energy database, a quantitative prediction model between crystal structure and lattice energy is constructed using a deep graph neural network. The quantitative prediction model is used to predict the lattice energy corresponding to an unknown crystal structure. Step S2 specifically includes: Step S21: Extract atomic features, molecular features, and atomic coordinate information from the crystal structure of the lattice energy database; Step S22: After connecting atomic features and molecular features through an embedding layer, they are mapped to form high-dimensional features through a linear layer. The embedding layer is used to associate atomic features and molecular features, and the linear layer is used to map the associated features to a higher-dimensional vector space. Step S23: Calculate the three-dimensional spatial distance between atoms based on atomic coordinate information, and then convert the distance between atoms into edge features using Bessel functions to capture the interaction between atoms; In step S23, the three-dimensional spatial distance between atoms is calculated using the following formula: .
[0008] in, , , For atomic coordinates, This represents the distance between atoms.
[0009] In step S23, the inter-atomic distance is converted into edge features using the Bessel function, which is achieved by the following formula: .
[0010] in, The distance between atoms. To cut off the distance, For envelope function, To scale the frequency, The calculated edge features
[0011] Step S24: Aggregate the high-dimensional features obtained in step S22 and the edge features obtained in step S23 to the same dimension through a linear layer, update the node information and capture the relationship between nodes using the TransformerConv layer, enhance the nonlinear expressive power of the model through a multilayer perceptron, and integrate the feature information of all atoms through a global pooling layer to obtain global molecular features. Step S25: Input the global molecular features into the linear layer output lattice energy prediction results, iteratively adjust the model parameters by the difference between the predicted values and the lattice energies in the database, and complete the model training and evaluation by combining the training set, validation set and test set to obtain a quantitative prediction model that can quickly predict lattice energy based on crystal structure.
[0012] In step S25, the performance of the quantitative prediction model is evaluated using the mean absolute percentage error, root mean square error, and coefficient of determination.
[0013] Step S3: Couple the trained quantitative prediction model with the crystal structure prediction process; first, generate candidate crystal structures of the target molecular crystal using the crystal structure prediction method, then use the quantitative prediction model to predict the lattice energy of the candidate crystal structures, and finally screen out thermodynamically stable candidate crystal forms based on the predicted lattice energy.
[0014] Step S3 specifically includes: Step S31: Based on the target molecular structure, set space group symmetry and unit cell size constraints, and use a crystal structure prediction tool to randomly generate potential crystal structures that satisfy the Weikov positions; Step S32: Pre-screen the potential crystal structures to remove unreasonable crystal structures with excessively close intermolecular distances, and obtain candidate crystal structures; Step S33: Using the quantitative prediction model constructed in step S2, predict the corresponding lattice energy based on the characteristic information of each candidate crystal structure; Step S34: Based on the predicted lattice energy and density information of candidate crystal structures, draw a lattice energy-density map, sort the candidate crystal structures based on the lattice energy, and screen for thermodynamically stable molecular crystal forms.
[0015] In step S34, the larger the absolute value of the lattice energy, the higher the thermodynamic stability of the crystal. Thermodynamically stable molecular crystal forms are concentrated in the region with lower lattice energy in the lattice energy-density diagram.
[0016] A high-throughput screening system for thermodynamically stable molecular crystal forms includes: The lattice energy calculation and database establishment module uses a quantum mechanical method that considers dispersion interactions to calculate the lattice energy of molecular crystals and establishes a lattice energy database that includes molecular crystal structures and their corresponding lattice energies. The model building module, based on the lattice energy database, uses a deep graph neural network to build a quantitative prediction model between crystal structure and lattice energy. The quantitative prediction model is used to predict the lattice energy corresponding to an unknown crystal structure. The coupling screening module couples the trained quantitative prediction model with the crystal structure prediction process. First, candidate crystal structures of the target molecular crystal are generated using the crystal structure prediction method. Then, the lattice energy of the candidate crystal structures is predicted using the quantitative prediction model. Finally, thermodynamically stable candidate crystal forms are screened based on the predicted lattice energy.
[0017] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, a lattice energy database is constructed by taking into account the quantum mechanical method of dispersion interaction, and a deep graph neural network (DGNN) is used to capture the complex interactions of molecular crystals. The constructed model can accurately correlate crystal structure with lattice energy, accurately reflect the relative thermal stability between polymorphs, and provide a solid basis for the screening of stable crystal forms.
[0018] 2. In this invention, the DGNN model is used to replace the time-consuming energy calculation process in traditional crystal structure prediction (CSP), which greatly reduces the computational cost. Combined with the structure generation and pre-screening mechanism, a large number of candidate crystal structures can be processed quickly, breaking through the efficiency bottleneck of traditional methods and realizing high-throughput screening of thermodynamically stable crystal forms. Attached Figure Description
[0019] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the process structure of the present invention; Figure 2 This is a schematic diagram comparing the density functional theory simulation value and the experimental value of the lattice energy in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the molecular crystal structure pre-screening process in Embodiment 2 of the present invention; Figure 4 This is a schematic diagram showing the number distribution of molecular crystal structures in the 33 space groups in Example 2 of the present invention; Figure 5 This is a schematic diagram of the lattice energy-density spectra of 1285 candidate crystal structures in Embodiment 2 of the present invention; Figure 6 This is a schematic diagram showing the number distribution of molecular crystal structures in the 27 space groups in Example 2 of the present invention; Figure 7 This is a schematic diagram of the lattice energy-density spectra of 698 candidate crystal structures in Embodiment 2 of the present invention; Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.
[0023] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0024] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.
[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.
[0026] Example 1
[0027] like Figure 1As shown, this invention discloses a high-throughput screening method for thermodynamically stable molecular crystal forms. First, based on density functional theory (DFT) + range-separated self-consistent screening many-body dispersion correction (MBD@rsSCS) method, the lattice energy of molecular crystals is predicted considering dispersion interactions, establishing a lattice energy dataset. Second, based on the established dataset, a quantitative prediction model between crystal structure and lattice energy is established using a deep graphical neural network (DGNN) method. Subsequently, the established lattice energy prediction model is coupled with the crystal structure prediction (CSP) method, replacing its computationally expensive energy prediction process. Finally, based on the lattice energy-density map obtained from crystal structure prediction (CSP), high-throughput screening of molecular crystals is achieved. This includes: Step S1: Calculate the lattice energy of molecular crystals using quantum mechanical methods that consider dispersion interactions, and establish a lattice energy database that includes molecular crystal structures and their corresponding lattice energies. Step S1 specifically includes: S11. Select the crystal structure of the target molecule and obtain the crystal cell and molecular information; S12. Density functional theory is used, combined with a range-separated self-consistent screening method to handle dispersive interactions, and the lattice energy of the crystal structure is calculated using quantum mechanics; the calculation of the lattice energy of the crystal structure is achieved by the following formula: (1); in, The total energy of the unit cell. The number of molecules in a unit cell. This represents the energy of a single gas molecule.
[0028] Specifically, all calculations were performed in the CASTEP module of Materials Studio 2019, based on pseudopotential plane waves. Exchange-correlation interactions were performed using the generalized gradient approximation density functional theory (DFT) based on the Perdew-Burke-Ernzerhof (PBE-GGA) method. The generalized gradient approximation DFT is a commonly used method in density functional theory, effectively balancing computational accuracy and cost. Furthermore, effective theoretical calculation methods need to consider various intermolecular interactions, including hydrogen bonds, electrostatic interactions, inductive effects, and van der Waals forces. Dispersion forces, as a type of van der Waals force, exist between all molecules, especially in some macromolecules, where their strength is even comparable to that of hydrogen bonds. Therefore, considering dispersion effects in lattice energy calculations is essential. In the past, researchers have proposed various dispersion-corrected DFT methods, such as Grimme's Dn (n=1~4) correction, Tkatchenko-Scheffler (TS) correction, and many-body dispersion (MBD) correction. The advantage of many-body dispersion correction methods lies in considering many-body dispersion interactions, enabling a more accurate description of the structure and stability of molecular crystals. The many-body dispersion energy is the difference between the zero-point energies of coupled and uncoupled systems. Its calculation equation originates from the central quantum black hole model of isotropic atoms, as shown in the following equation. Wherein, For Hamiltonian eigenvalues, Let be the characteristic frequency of the j-th atom in the system. N This represents the number of atoms in the system.
[0029] .
[0030] The many-body dispersion correction method was later improved into the range-separated self-consistent screening many-body dispersion correction method MBD@rsSCS. This method has been proven to be efficient and accurate in calculating long-range weak interactions in finite bandgap systems, including molecular crystals. This paper uses MBD@rsSCS for dispersion correction. S13. Preprocess the calculated lattice energy data to establish a database containing different molecular crystal structures and their corresponding lattice energies.
[0031] Specifically: All calculations are set to fine precision. Full optimization is performed on the solid-phase unit cell. For gaseous molecules, a single molecule is replicated from a non-relaxed unit cell and placed in a cubic simulation box with a side length of 10 Å; the unit cell is not optimized. During geometry optimization, the convergence thresholds for maximum energy change, maximum force, maximum stress, and maximum displacement are 1 × 10⁻⁶. -5 eV / atom, 0.03eV / Å, 0.05GPaand0.001Å.
[0032] Step S2: Based on the lattice energy database, a quantitative prediction model between crystal structure and lattice energy is constructed using a deep graph neural network. The quantitative prediction model is used to predict the lattice energy corresponding to an unknown crystal structure. Because the interatomic spacing between periodic structures in molecular crystals is typically quite long, traditional graph neural network models cannot accurately represent this information. Therefore, this paper employs a deep graph neural network (DGNN) model for molecular crystal modeling to establish a quantitative prediction model between molecular crystal structure and lattice energy. Specifically, this includes: Step S2 specifically includes: Step S21: Extract atomic features, molecular features, and atomic coordinate information from the crystal structure of the lattice energy database; Step S22: After connecting atomic features and molecular features through an embedding layer, they are mapped to form high-dimensional features through a linear layer. The embedding layer is used to associate atomic features and molecular features, and the linear layer is used to map the associated features to a higher-dimensional vector space. Step S23: Calculate the three-dimensional spatial distance between atoms based on atomic coordinate information, and then convert this distance into edge features using a Bessel function to capture the interactions between atoms. Step S24: Aggregate the high-dimensional features obtained in Step S22 and the edge features obtained in Step S23 to the same dimension through a linear layer. Use a TransformerConv layer to update node information and capture the relationships between nodes. Then, enhance the nonlinear expressive power of the model through a multilayer perceptron. Finally, integrate the feature information of all atoms through a global pooling layer to obtain global molecular features.
[0033] Step S25: Input the global molecular features into the linear layer output lattice energy prediction results, iteratively adjust the model parameters by the difference between the predicted values and the lattice energies in the database, and complete the model training and evaluation by combining the training set, validation set and test set to obtain a quantitative prediction model that can quickly predict lattice energy based on crystal structure.
[0034] Step S3: Couple the trained quantitative prediction model with the crystal structure prediction process. First, generate candidate crystal structures for the target molecular crystal using crystal structure prediction methods. Then, use the quantitative prediction model to predict the lattice energy of the candidate crystal structures. Finally, screen out thermodynamically stable candidate crystal forms based on the predicted lattice energy. Crystal structure prediction (CSP) refers to using theoretical calculations and simulation techniques, combined with the chemical composition, interactions, and physical properties of molecules or atoms, to predict the most stable crystal structure that a substance may form under different external conditions (such as temperature, pressure, etc.), and obtain the most promising candidate structures by calculating their energy, symmetry, lattice constant, and other properties. Specifically: Step S3 specifically includes: Step S31: Based on the target molecular structure, set space group symmetry and unit cell size constraints, and use a crystal structure prediction tool to randomly generate potential crystal structures that satisfy the Weikov positions; Step S3.2: Pre-screen the potential crystal structures to remove unreasonable crystal structures with excessively close intermolecular distances, and obtain candidate crystal structures; Step S33: Using the quantitative prediction model constructed in step S2, predict the corresponding lattice energy based on the characteristic information of each candidate crystal structure; Step S34: Based on the predicted lattice energy and density information of candidate crystal structures, draw a lattice energy-density map, sort the candidate crystal structures based on the lattice energy, and screen for thermodynamically stable molecular crystal forms.
[0035] In step S34, the larger the absolute value of the lattice energy, the higher the thermodynamic stability of the crystal. Thermodynamically stable molecular crystal forms are concentrated in the region with lower lattice energy in the lattice energy-density diagram.
[0036] This paper utilizes the CrySPY tool to randomly generate crystal structures. CrySPY is a powerful crystal structure prediction tool that combines various advanced computational methods, such as random search and evolutionary algorithms, to help researchers quickly and effectively predict the crystal structures of new materials based on theoretical calculations. First, based on a given molecular structure, potential crystal forms with reasonable chemical structures are randomly generated using user-defined symmetry constraints, cell sizes, and other information. However, molecular crystals that satisfy symmetry requirements under a specific space group may have problems with excessively close intermolecular distances, meaning that accurate molecular distances within the cell cannot be obtained when calculating intermolecular bonds. Therefore, to ensure the accuracy of molecular crystal structures, all structures are pre-screened before structure evaluation to remove structures with excessively close intermolecular distances within the cell. Second, for the potential crystal forms obtained through pre-screening, a lattice energy prediction model based on a deep graphical neural network method is used to replace density functional theory calculations to obtain the lattice energies of different crystal structures, thereby reducing computational costs. Finally, density-lattice energy maps are plotted and analyzed to achieve high-throughput screening of polymorphs.
[0037] Example 2
[0038] This embodiment is based on the experiment in embodiment 1, as detailed below: 1. Lattice energy prediction based on density functional theory (DFT): This paper uses density functional theory (DFT) combined with range-separated self-consistent screening many-body dispersion correction (MBD@rsSCS) to calculate 230 crystal structures of 40 common molecules, including compounds with various functional groups such as acids, alcohols, amides, amino acids, and acid anhydrides. Among these, there are 10 acids, 4 alcohols, 3 amides, 2 amino acids, 2 acid anhydrides, and the remainder are aromatic and heterocyclic compounds. To further explore the polymorphism of drug molecules, the dataset also includes aspirin, metronidazole, and acetaminophen. Therefore, this dataset can reflect the structural characteristics of most organic molecular crystals to a certain extent.
[0039] Lattice energy is a crucial indicator for high-throughput screening of crystal structures and requires accurate prediction. To improve the accuracy of the proposed lattice energy prediction method, the predicted lattice energy value obtained based on density functional theory (DFT) is compared with... ) and experimental values in existing literature ( The accuracy of the predicted lattice energy was compared using mean absolute percentage error (MAPE) and coefficient of determination (R²). The regression results are as follows: Figure 2 As shown, the coefficient of determination (R²) = 0.9572 and the mean absolute percentage error (MAPE) = 5.51%, indicating that the proposed density functional theory (DFT) + range separation self-consistent screening many-body dispersion correction (MBD@rsSCS) method can estimate lattice energy with high accuracy.
[0040] 2. Lattice energy prediction model based on deep graphical neural network (DGNN): Based on the established lattice energy database, a quantitative prediction model between crystal structure and lattice energy was built using a deep graphical neural network (DGNN) for molecular crystals. The dataset was divided into training, validation, and test sets in a ratio of 7:1:2. The validation set was used to select hyperparameters for the model, and the test set was used to evaluate model performance. Optimal values for key parameters were determined through multiple simulations.
[0041] 3. Crystal structure prediction based on deep graphical neural networks (DGNN): This paper couples a lattice energy prediction model based on the deep graphical neural network (DGNN) method with crystal structure prediction (CSP) to perform high-throughput crystal structure screening. Using glycine and aspirin, two typical polymorphic substances, as examples, this paper conducts a full-space crystal structure search.
[0042] Case 1: Glycine Glycine (C2H5NO2) is the simplest amino acid, containing both acidic and basic functional groups in its molecule. It is a nonpolar neutral amino acid, widely distributed in organisms, and is an important component of proteins and an indispensable intermediate in many biosynthetic processes. Glycine has extensive applications in the food, pharmaceutical, pesticide, and chemical industries. For example, it can be used as a food additive to improve food flavor; in the pharmaceutical field, glycine is used to synthesize antibiotics, antitumor drugs, and inosine; in agriculture, it is an important raw material for the synthesis of the herbicide glyphosate. Furthermore, glycine is an important precursor in the human body for the synthesis of creatine, glutathione, porphyrin, and purine nucleotides, and acts as an inhibitory neurotransmitter in the central nervous system, helping to improve sleep and relieve anxiety. Due to its good biocompatibility and wide range of functions, glycine has broad application prospects in the biomedical and health fields.
[0043] First, a random algorithm was used to perform a full space group search on glycine, generating 6400 possible glycine crystal structures. Then, the generated structures were pre-screened by considering factors such as order, lattice position, number of atoms, number of atoms (z') in the asymmetric unit cell, and space group. The process is as follows: Figure 3 As shown; ultimately, 1285 candidate crystal structures were obtained, covering 33 space groups. The distribution of the number of each space group is as follows. Figure 4 As shown, by Figure 4 It can be seen that the top 10 space groups in terms of the number of crystals are P21(4), P-1(2), Pc(7), P212121(19), P31(144), P32(145), Pna21(33), P21 / c(14), Cc(9), and P-1(1), with the numbers in parentheses indicating the order of crystals. Figure 4 The space group numbering in the model is used. Finally, the deep graphical neural network (DGNN) model established in Section 3.2 is used to predict the lattice energy of 1285 candidate crystal structures, resulting in lattice energy-density maps, as shown below. Figure 5 As shown.
[0044] Depend on Figure 5 It can be seen that the crystal structures of the 33 space groups are relatively dispersed, with densities ranging from 1.2 to 2.2 g / cm³ and lattice energies ranging from -175 to 165 kJ / mol. Structures with lower lattice energies (i.e., the most thermodynamically stable) are concentrated in the P212121 space group, which accounts for 61.54% of the 1% of structures with the lowest lattice energies. Of the 5% of structures with lower lattice energies, 67.69% belong to the same crystal system (orthorhombic system) as the P212121 space group. Further verification is needed. Figure 5 The reliability of the results was determined by comparing the candidate crystal structures obtained from Crystal Structure Prediction (CSP) with the four experimental crystal forms of glycine contained in the Cambridge Structure Database (CSD). -、 -、 -and -) Compare them. By iterating... Figure 5 The candidate crystal structures were analyzed, and four candidate crystal forms were ultimately determined to be highly consistent with the experimental crystal forms. The results are shown in Table 1. Here, a, b, and c are cell constants, corresponding to the side lengths of the cell in the three spatial dimensions; α, β, and γ are lattice angles, where α is the angle between the b-axis and c-axis, β is the angle between the a-axis and c-axis, and γ is the angle between the a-axis and b-axis. Table 3 shows that the simulated parameters of the four glycine cell structures are in good agreement with the experimental parameters.
[0045] Table 1 Comparison of simulated and experimental parameters of glycine unit cell: .
[0046] In summary, the glycine lattice energy-density map obtained based on the proposed deep graph neural network (DGNN) and crystal structure prediction (CSP) coupling method can provide a theoretical basis and strong support for screening more stable glycine crystal structures. Further experimental exploration can be conducted in the future to obtain more stable crystal structures.
[0047] Case 2: Aspirin Aspirin (C9H8O4) is a classic nonsteroidal anti-inflammatory drug (NSAID) with analgesic, antipyretic, anti-inflammatory, and antiplatelet aggregation effects. It is widely used to treat headaches, toothaches, arthritis, and fever caused by colds. It also plays a crucial role in cardiovascular disease prevention; low doses can effectively reduce the risk of thrombosis and prevent myocardial infarction and stroke. However, long-term use of aspirin may cause gastrointestinal discomfort and even ulcer bleeding, therefore it should be taken rationally under the guidance of a doctor. As one of the most widely used drugs in the world, aspirin not only plays a significant role in the medical field but also drives the development of medicinal chemistry.
[0048] First, a randomized algorithm was used to perform a full space group search on aspirin, generating 13,000 possible aspirin crystal structures; then... Figure 3 The screening process shown yielded 698 aspirin crystal structures, encompassing 27 space groups, such as... Figure 6 As shown. By Figure 6 It can be seen that the top 10 space groups in terms of the number of crystals are P21 / c (14), Cc (9), Pna21 (33), P212121 (19), Pca21 (9), P43 (78), Pnn2 (34), P41 (76), P21212 (18), and P42 (77), with the numbers in parentheses indicating the number of crystals. Figure 6 The space group numbering in the model is then determined. Finally, the established deep graphical neural network (DGNN) model is used to predict the lattice energy of 698 candidate crystal structures, resulting in lattice energy-density maps, such as... Figure 7 As shown.
[0049] Depend on Figure 7 It can be seen that the crystal structures of the 27 space groups are relatively dispersed, with densities ranging from 1.0 to 1.7 g / cm³ and lattice energies ranging from -150 to -145 kJ / mol. Structures with lower lattice energies (i.e., more thermodynamically stable) are concentrated in... of
[0050] P21(4) Space group. Observe all groups that have this characteristic ( The energy ranking of candidate crystal structures revealed that over 60% of these structures ranked in the top 100 and possessed relatively low lattice energies, suggesting they could be used as one of the indicators for qualitatively assessing the thermodynamic stability of crystal structures. Further verification is needed. Figure 7 The reliability of the results was assessed by comparing the candidate crystal structures obtained from Crystal Structure Prediction (CSP) with the two experimental crystal forms of aspirin contained in the Cambridge Structure Database (CSD). -and -) Compare them. By iterating... Figure 7 The candidate crystal structures were analyzed, and two candidate crystal forms were ultimately determined to be highly consistent with the experimental crystal forms. The results are shown in Table 4. The meanings of a, b, c, α, β, and γ are the same as those mentioned above. As can be seen from Table 2, the simulated parameters of the two aspirin unit cells are in good agreement with the experimental parameters, and the space groups are completely consistent.
[0051] Table 2 Comparison of simulated and experimental parameters of aspirin unit cell: .
[0052] In summary, the aspirin lattice energy-density spectrum obtained in this paper can provide a theoretical basis and strong support for screening more stable aspirin crystal structures. Further experimental exploration can be conducted in the future to obtain more stable crystal structures.
[0053] in conclusion: First, this paper uses density functional theory (DFT) with range separation self-consistent screening and many-body dispersion correction (MBD@rsSCS) to predict the lattice energy of crystal structures, considering dispersion forces. A database of lattice energies containing 40 molecules and 230 crystal structures is established. The prediction results of DFT with range separation self-consistent screening and many-body dispersion correction (MBD@rsSCS) are compared with experimental values, with a coefficient of determination (R²) of 0.9572 and a mean absolute percentage error (MAPE) of 5.51%. The accuracy of the relative thermal stability ranking of the lattice energies predicted by this method is verified using imidazole and benzamide as examples. Second, based on the established database, a quantitative prediction model between crystal structure and lattice energy is established using a deep graph neural network (DGNN) considering molecular information. The training set has a coefficient of determination (R²) of 0.9905 and a mean absolute percentage error (MAPE) of 2.31%, while the test set has a coefficient of determination (R²) of 0.9460 and a mean absolute percentage error (MAPE) of 2.31%. The quantitative prediction model (APE) achieved good prediction results on both the validation set (R² = 0.9798, MAPE = 2.99%) and the verification set (APE = 3.69%). Subsequently, the proposed quantitative prediction model was coupled with the Crystal Structure Prediction (CSP) method. Under given molecular structure and a series of constraints (such as cell size, space group symmetry, etc.), the CSP method first obtained a chemically reasonable crystal structure. Then, by considering information such as orderliness, lattice position, and number of atoms, the generated crystal structure was pre-screened to obtain candidate crystal structures. Next, a quantitative prediction method based on deep graphical neural network (DGNN) was used to predict the lattice energy of each candidate structure, resulting in a lattice energy-density map, thus achieving high-throughput screening of crystal structures. Finally, the proposed method was verified using glycine and aspirin as case studies. The candidate crystal structures generated by the CSP method included the experimentally obtained crystal structures, and these experimental crystal structures were distributed at the bottom of the lattice energy-density map. The method proposed in this paper provides a theoretical basis and strong support for future exploration and acquisition of more thermodynamically stable crystal structures.
[0054] Example 3
[0055] This embodiment proposes a high-throughput screening system for thermodynamically stable molecular crystal forms, including: The lattice energy calculation and database establishment module uses a quantum mechanical method that considers dispersion interactions to calculate the lattice energy of molecular crystals and establishes a lattice energy database that includes molecular crystal structures and their corresponding lattice energies. The model building module, based on the lattice energy database, uses a deep graph neural network to build a quantitative prediction model between crystal structure and lattice energy. The quantitative prediction model is used to predict the lattice energy corresponding to an unknown crystal structure. The coupling screening module couples the trained quantitative prediction model with the crystal structure prediction process. First, candidate crystal structures of the target molecular crystal are generated using the crystal structure prediction method. Then, the lattice energy of the candidate crystal structures is predicted using the quantitative prediction model. Finally, thermodynamically stable candidate crystal forms are screened based on the predicted lattice energy.
[0056] The circuits, electronic components, and modules involved are all existing technologies, which can be fully implemented by those skilled in the art, and need not be elaborated upon. The scope of protection of this invention does not involve any improvement to the software and methods.
[0057] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0058] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for high-throughput screening of thermodynamically stable molecular crystal forms, characterized in that, include: Step S1: Calculate the lattice energy of molecular crystals using quantum mechanical methods that consider dispersion interactions, and establish a lattice energy database that includes molecular crystal structures and their corresponding lattice energies. Step S2: Based on the lattice energy database, a quantitative prediction model between crystal structure and lattice energy is constructed using a deep graph neural network. The quantitative prediction model is used to predict the lattice energy corresponding to an unknown crystal structure. Step S3: Couple the trained quantitative prediction model with the crystal structure prediction process; first, generate candidate crystal structures of the target molecular crystal using the crystal structure prediction method, then use the quantitative prediction model to predict the lattice energy of the candidate crystal structures, and finally screen out thermodynamically stable candidate crystal forms based on the predicted lattice energy.
2. The high-throughput screening method for thermodynamically stable molecular crystal forms according to claim 1, characterized in that, Step S1 specifically includes: Select the crystal structure of the target molecule and obtain the crystal's unit cell and molecular information; Density functional theory was used in conjunction with a range-separated self-consistent screening method to handle dispersive interactions, and the lattice energy of the crystal structure was calculated using quantum mechanics. The calculated lattice energy data are preprocessed to establish a database containing different molecular crystal structures and their corresponding lattice energies.
3. The high-throughput screening method for thermodynamically stable molecular crystal forms according to claim 2, characterized in that, The lattice energy of the crystal structure is calculated using the following formula: ； in, The total energy of the unit cell. The number of molecules in a unit cell. This represents the energy of a single gas molecule.
4. The high-throughput screening method for thermodynamically stable molecular crystal forms according to claim 1, characterized in that, Step S2 specifically includes: Step S21: Extract atomic features, molecular features, and atomic coordinate information from the crystal structure of the lattice energy database; Step S22: After connecting atomic features and molecular features through an embedding layer, they are mapped to form high-dimensional features through a linear layer. The embedding layer is used to associate atomic features and molecular features, and the linear layer is used to map the associated features to a higher-dimensional vector space. Step S23: Calculate the three-dimensional spatial distance between atoms based on atomic coordinate information, and then convert the distance between atoms into edge features using Bessel functions to capture the interaction between atoms; Step S24: Aggregate the high-dimensional features obtained in step S22 and the edge features obtained in step S23 to the same dimension through a linear layer, update the node information and capture the relationship between nodes using the TransformerConv layer, enhance the nonlinear expressive power of the model through a multilayer perceptron, and integrate the feature information of all atoms through a global pooling layer to obtain global molecular features. Step S25: Input the global molecular features into the linear layer output lattice energy prediction results, iteratively adjust the model parameters by the difference between the predicted values and the lattice energies in the database, and complete the model training and evaluation by combining the training set, validation set and test set to obtain a quantitative prediction model that can quickly predict lattice energy based on crystal structure.
5. The high-throughput screening method for thermodynamically stable molecular crystal forms according to claim 4, characterized in that, In step S23, the three-dimensional spatial distance between atoms is calculated using the following formula: ； in, , , For atomic coordinates, This represents the distance between atoms.
6. The high-throughput screening method for thermodynamically stable molecular crystal forms according to claim 4, characterized in that, In step S23, the inter-atomic distance is converted into edge features using the Bessel function, which is achieved by the following formula: ； in, The distance between atoms. To cut off the distance, For envelope function, To scale the frequency, The edge features are calculated.
7. The high-throughput screening method for thermodynamically stable molecular crystal forms according to claim 4, characterized in that, In step S25, the performance of the quantitative prediction model is evaluated using the mean absolute percentage error, root mean square error, and coefficient of determination.
8. The high-throughput screening method for thermodynamically stable molecular crystal forms according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Based on the target molecular structure, set space group symmetry and unit cell size constraints, and use a crystal structure prediction tool to randomly generate potential crystal structures that satisfy the Weikov positions; Step S32: Pre-screen the potential crystal structures to remove unreasonable crystal structures with excessively close intermolecular distances, and obtain candidate crystal structures; Step S33: Using the quantitative prediction model constructed in step S2, predict the corresponding lattice energy based on the characteristic information of each candidate crystal structure; Step S34: Based on the predicted lattice energy and density information of candidate crystal structures, draw a lattice energy-density map, sort the candidate crystal structures based on the lattice energy, and screen for thermodynamically stable molecular crystal forms.
9. The high-throughput screening method for thermodynamically stable molecular crystal forms according to claim 8, characterized in that, In step S34, the larger the absolute value of the lattice energy, the higher the thermodynamic stability of the crystal. Thermodynamically stable molecular crystal forms are concentrated in the region with lower lattice energy in the lattice energy-density diagram.
10. A high-throughput screening system for thermodynamically stable molecular crystal forms, characterized in that, include: The lattice energy calculation and database establishment module uses a quantum mechanical method that considers dispersion interactions to calculate the lattice energy of molecular crystals and establishes a lattice energy database that includes molecular crystal structures and their corresponding lattice energies. The model building module, based on the lattice energy database, uses a deep graph neural network to build a quantitative prediction model between crystal structure and lattice energy. The quantitative prediction model is used to predict the lattice energy corresponding to an unknown crystal structure. The coupling screening module couples the trained quantitative prediction model with the crystal structure prediction process. First, candidate crystal structures of the target molecular crystal are generated using the crystal structure prediction method. Then, the lattice energy of the candidate crystal structures is predicted using the quantitative prediction model. Finally, thermodynamically stable candidate crystal forms are screened based on the predicted lattice energy.