A multi-objective molecular optimization method and system based on selective sampling learning

Abstract

This invention discloses a multi-objective molecular optimization method based on selective sampling learning, comprising: initializing a molecular generation model; performing molecular sampling based on the initialized molecular generation model, with each sampling sampling a batch of molecules to form a first molecular set; calculating a first index, a second index, and a third index for each molecule in the first molecular set; filtering each molecule in the first molecular set according to the comparison results between preset thresholds for the first, second, and third indices and the first, second, and third indices, retaining molecules that meet all the criteria and discarding molecules that do not meet any of the criteria; repeating the process until the number of molecules obtained after filtering is equal to or higher than the first molecule number threshold, and then stopping the loop to obtain a second molecular set; selecting the second molecular set based on a first evaluation criterion to obtain a third molecular set; and using the third molecular set for training and parameter updating of the molecular generation model. A corresponding system and application are also disclosed.

Description

Technical Field

[0001] This invention belongs to the field of biological gene and artificial intelligence drug research technology, and in particular relates to a multi-objective molecular optimization method and system based on selective sampling learning. Background Technology

[0002] In the field of drug molecule synthesis and optimization, ensuring the safety and efficacy of a drug requires meeting multiple clinical requirements. These indicators include binding activity to specific protein targets, drug toxicity in vivo, and drug metabolism rate. In new drug development, the first step is to identify lead compounds, followed by multi-indicator optimization. This optimization process is often performed by experts based on experience, which is slow and inefficient. Therefore, automating optimization and efficiently improving the various indicators of lead compounds is crucial. With the rise of machine learning, the interdisciplinary field of AI-driven drug discovery has developed significantly, leading to the emergence of numerous methods for molecule optimization. Currently, commonly used methods include reinforcement learning-based and Monte Carlo tree search-based methods. Both focus on key indicators in drug development, using computational methods to transform these indicators into optimization goals and train models to ultimately optimize multiple molecular indicators.

[0003] Reinforcement learning-based methods typically represent molecules as graphs based on molecular diagrams. During training, the model uses molecular subgraphs as "states" in the reinforcement learning process, inducing and completing these subgraphs to ensure the generated molecules meet expectations in key metrics, thereby optimizing the molecule. However, in reality, even minor modifications to a molecule can lead to significant changes in its activity, toxicity, and other properties. Therefore, knowing the properties of the final molecule is insufficient to accurately assess the properties of intermediate molecules, which introduces many limitations to reinforcement learning-based methods.

[0004] The core of Monte Carlo tree search methods is to represent molecules as nodes in a search tree and use Monte Carlo methods to search for molecules that meet the requirements across multiple metrics. However, these methods are computationally intensive and require high computational efficiency. Furthermore, the number of molecules that can be optimized and produced is uncertain, depending on the actual expansion of the search tree; this instability in the output results cannot meet practical application needs.

[0005] Therefore, there is indeed a need to propose better solutions for the aforementioned existing technologies. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, a multi-objective molecular optimization method and system based on selective sampling learning is provided. This method and system do not depend on any specific molecular generation model, have strong transferability, and can be adapted to various molecular generation models, achieving the expected results under various optimization objectives. Furthermore, this strategy can flexibly adjust the optimization objectives, adapting to specific and changing business needs in actual drug development projects and improving multiple indicators of molecules generated by the molecular generation model.

[0007] Through experiments and validation on different molecular generation models, it has been found that for a molecular generation model, the various properties of the generated molecules are similar to those of the training data. Based on this characteristic, this invention proposes a multi-objective molecular optimization strategy based on selective sampling learning. This strategy continuously selects molecules that meet specific optimization objectives through sampling and screening methods, and allows the model to learn, thereby achieving the goal of multi-objective optimization.

[0008] This invention provides a multi-objective molecular optimization method based on selective sampling learning, comprising:

[0009] S1, Initialize the molecular generation model;

[0010] S2, molecular sampling is performed based on the initialized molecular generation model. Each molecular sampling will sample a batch of molecules to form the first molecular set.

[0011] S3, calculate the first index, second index and third index of each molecule in the first molecule set;

[0012] S4, based on the preset first index threshold, second index threshold and third index threshold, the first index, second index and third index are compared with the first index threshold, second index threshold and third index threshold respectively. Based on the comparison results, each molecule in the first molecule set is filtered. Molecules that meet the standards of the first index, second index and third index all meet the standards of the first index threshold, second index threshold and third index threshold are retained. Molecules that do not meet the standards of the first index, second index threshold and third index threshold at any one of the first index, second index and third index threshold are discarded.

[0013] S5, set a first molecule number threshold, loop S1-S4 until the number of molecules obtained after filtering is equal to or higher than the first molecule number threshold, then stop the loop and obtain a second molecule set;

[0014] S6, Based on the first evaluation criterion, the second molecular set is screened again to obtain the third molecular set;

[0015] S7, the third molecular set is used for training and parameter updating of the molecular generation model.

[0016] Preferably, the first indicator is a physicochemical property, the second indicator is ADMET, and the third indicator is activity.

[0017] Preferably, the first evaluation criterion is: retaining molecules in the second molecular set M1 whose activity is higher than the overall median as the third molecular set.

[0018] Preferably, the initialized molecular generation model is a deep graph generation model (DGMG), a connection tree variational autoencoder (JTNN) for subgraph generation, a GCPN, an MRNN, or a PS-VAE.

[0019] Preferably, before implementing S1, the initialized molecular generation model is pre-trained using the publicly available dataset ZINC.

[0020] Preferably, the method further includes:

[0021] S8, after the training and parameter update of the molecular generation model are completed, multiple molecules are generated using the trained molecular generation model, and the multiple molecules are evaluated by indicators.

[0022] Preferably, S8 includes:

[0023] S81, Determine the optimization objective for this optimization experiment;

[0024] S82, determine the scoring function corresponding to each optimization objective;

[0025] S83, multiple molecules are generated using the trained molecular generation model, and the multiple molecules are evaluated based on the scoring function.

[0026] A second aspect of the present invention provides a multi-objective molecular optimization system based on selective sampling learning, comprising:

[0027] The model initialization module (101) is used to initialize the molecular generation model;

[0028] The first molecular set generation module (102) is used to perform molecular sampling based on the initialized molecular generation model. Each molecular sampling will sample a batch of molecules to form the first molecular set.

[0029] The calculation module (103) is used to calculate the first index, the second index and the third index of each molecule in the first molecule set;

[0030] The comparison filtering module (104) is used to compare the first index, the second index, and the third index with the first index threshold, the second index threshold, and the third index threshold respectively according to the preset first index threshold, the second index threshold, and the third index threshold. Based on the comparison results, each molecule in the first molecule set is filtered. Molecules that meet the standards of the first index, the second index, and the third index are retained, and molecules that do not meet the standards of the first index, the second index threshold, and the third index are discarded.

[0031] The second molecular set generation module (105) is used to set a first molecular number threshold, and stop the loop when the number of molecules obtained after filtering is equal to or higher than the first molecular number threshold, and obtain a second molecular set. In this embodiment, the second molecular set is denoted as M1.

[0032] The third molecular set generation module (106) filters the second molecular set again based on the first evaluation criterion to obtain the third molecular set;

[0033] The model training module (107) is used to use the third molecular set for training and parameter updating of the molecular generation model.

[0034] Preferably, the system further includes:

[0035] The index evaluation module (108) is used to generate multiple molecules using the trained molecular generation model after the training and parameter update of the molecular generation model are completed, and to evaluate the index of the multiple molecules.

[0036] A third aspect of the present invention is to provide an application of a multi-objective molecular optimization method based on selective sampling learning in drug development.

[0037] The method, system, and application provided by this invention have the following beneficial technical effects:

[0038] This invention innovatively proposes a multi-objective molecular optimization strategy based on selective sampling learning, overcoming the shortcomings of traditional molecular optimization methods such as unstable number of generated molecules, poor method transferability, and poor flexibility in adding optimization targets. By adopting a strategy of sampling first, then filtering, and finally learning, the method ensures flexibility and universality. In the verification of actual projects, the bioactivity of the generated molecules has been improved by two to three orders of magnitude. This brings great convenience and efficiency to the molecular optimization and modification process in new drug development, reduces the dependence on knowledge in related chemical fields, and has high practical value. Attached Figure Description

[0039] Figure 1This is a schematic diagram of a multi-objective molecular optimization method based on selective sampling learning according to a preferred embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of the JNK3 activity value enhancement curve according to a preferred embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram illustrating the increase curve of PAPP value with the number of training rounds according to a preferred embodiment of the present invention.

[0042] Figure 4 This is a schematic diagram illustrating the increase curve of the KS value with the number of training rounds according to a preferred embodiment of the present invention;

[0043] Figure 5 This is a diagram illustrating the architecture of a multi-objective molecular optimization system based on selective sampling learning according to a preferred embodiment of the present invention. Detailed Implementation

[0044] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0045] The terminology used in the embodiments is as follows:

[0046] SMILES: Simplified Molecular Linear Input Specification

[0047] ADMET: Indicators related to drug absorption, distribution, metabolism, efflux, and toxicity.

[0048] DGMG: A depth map generation model

[0049] Ubuntu: A Linux operating system

[0050] Example 1

[0051] like Figure 1 As shown, this embodiment provides a multi-objective molecule optimization method based on selective sampling learning, including:

[0052] S1, Initialize the molecular generation model;

[0053] S2, molecular sampling is performed based on the initialized molecular generation model. Each molecular sampling will sample a batch of molecules to form the first molecular set.

[0054] S3, calculate the first index, second index and third index of each molecule in the first molecule set;

[0055] In a preferred embodiment, the first indicator is a physicochemical property, the second indicator is ADMET, and the third indicator is activity;

[0056] S4, based on the preset first index threshold, second index threshold and third index threshold, the first index, second index and third index are compared with the first index threshold, second index threshold and third index threshold respectively. Based on the comparison results, each molecule in the first molecule set is filtered. Molecules that meet the standards of the first index, second index and third index all meet the standards of the first index threshold, second index threshold and third index threshold are retained. Molecules that do not meet the standards of the first index, second index threshold and third index threshold at any one of the first index, second index and third index threshold are discarded.

[0057] S5, set a first molecule number threshold, loop S1-S4 until the number of molecules obtained after filtering is equal to or higher than the first molecule number threshold, then stop the loop and obtain a second molecule set. In this embodiment, the second molecule set is denoted as M1.

[0058] S6, Based on the first evaluation criterion, the second molecular set is screened again to obtain the third molecular set;

[0059] In this embodiment, the first evaluation criterion is: retain the molecules in the second molecular set M1 whose activity is higher than the overall median, and denot them as the third molecular set M2;

[0060] S7, the third molecular set is used for training and parameter updating of the molecular generation model.

[0061] As a preferred embodiment, the initialized molecular generation model is DGMG (Deep Generative Models of Graphs). DGMG is a general framework for learning graph distributions by progressively adding atoms and bonds.

[0062] Of course, those skilled in the art will readily recognize that JTNN (Junction Tree Variational Autoencoder for Molecular Graph Generation) can also be used as the initial molecular generation model. JTNN can progressively expand the molecule while maintaining the chemical valence at each step, thus enabling its use for molecular generation and optimization.

[0063] Of course, those skilled in the art may also use GCPN, MRNN or PS-VAE as the three small molecule generation models as the initial molecular generation model.

[0064] GCPN is essentially a reinforcement learning framework where each action comprises four values: the two nodes to be connected, the type of the edge, and whether to terminate. First, GCN extracts features from the graph, including the already generated subgraphs and the atoms and subgraphs to be selected. Then, an action is derived through four steps: selecting the first node, the second node, the form of the edge connection, and the graph's features to determine whether to terminate.

[0065] MRNN (molecular) is a small extension of GRNN (graph). GRNN has two RNNs: Node-RNN (horizontal) and Edge-RNN (vertical). That is, at each step, a node is added to the current subgraph (one step horizontally), and the edges between the newly added node and existing nodes are output (i-1 steps vertically), until there are no more edges to add in the (n+1)th step, at which point the process terminates. Compared to GRNN, MRNN only classifies the output nodes, and edges also have four choices: {0, 1, 2, 3}. Furthermore, MRNN uses RL to optimize property.

[0066] In PS-VAE, PS stands for principle subgraph, meaning the subgraph that appears frequently in the dataset. These subgraphs are extracted, and during the generation process, PSs are generated directly and edges are added, which is much more efficient than adding nodes one by one. The main framework is VAE. First, a GNN is used to obtain the graph's features h_G as input to the encoder. A reparameterization trick is used to obtain the latent variable z, which is then used as input to a GRU (similar to RNN). Each step outputs a PS, and after outputting the end, all PSs are obtained. For any two points on any two PSs, an MLP is used to determine the attributes of these two points and whether an edge needs to be added. The final loss function consists of three parts: two are built into the VAE, and the third is the loss introduced by generating specific attributes.

[0067] As a preferred implementation, before implementing S1, the initialized molecular generation model is pre-trained using the publicly available dataset ZINC.

[0068] In a preferred embodiment, the method further includes:

[0069] S8, after the training and parameter update of the molecular generation model are completed, multiple molecules are generated using the trained molecular generation model, and the multiple molecules are evaluated by indicators.

[0070] In a preferred embodiment, S8 includes:

[0071] S81, Determine the optimization objective for this optimization experiment;

[0072] In this embodiment, the specific optimization objectives are as follows:

[0073] (1) Molecular activity of JNK3, a target associated with Alzheimer's disease;

[0074] (2) Apparent penetration rate PAPP in the ADMET index;

[0075] (3) Kinetic solubility KS in the ADMET index.

[0076] S82, determine the scoring function corresponding to each optimization objective;

[0077] S83, multiple molecules are generated using the trained molecular generation model, and the multiple molecules are evaluated based on the scoring function.

[0078] Based on the multi-objective molecule optimization strategy based on selective sampling learning proposed in this invention, we designed a specific example for verification, as follows:

[0079] I. Experimental Preparation

[0080] (I) First, determine the optimization objective of this optimization experiment and prepare the corresponding scoring function. The specific optimization objective is as follows:

[0081] 1. Molecular activity of JNK3, a target associated with Alzheimer's disease;

[0082] 2. Apparent penetration rate (PAPP) in the ADMET metric;

[0083] 3. Kinetic solubility KS in the ADMET index.

[0084] (ii) Next, the molecular generation model to be optimized is determined to be DGMG, and the model is pre-trained in advance using the public dataset ZINC.

[0085] II. Optimization Experiment

[0086] This optimization experiment was conducted on a server running Ubuntu 20.04, using 20 CPU cores for model training, with a total training time of 7 hours. The trend graphs of the changes in the target JNK3 molecule activity, PAPP value, and KS value for each training epoch are shown below. Figure 2 , Figure 3 and Figure 4 As shown, from Figure 2-4 As can be seen, all indicators have been significantly improved.

[0087] III. Result Evaluation

[0088] After the model training was completed, 20,000 molecules were generated using the trained DGMG generative model, and various indicators were evaluated. The evaluation results are shown in Table 1 below. The activity of the JNK3 target was increased by nearly 4 orders of magnitude, and PAPP and HERG were increased by 52% and 66% respectively, showing significant optimization effects.

[0089] Table 1 Comparison of Multi-Objective Indicators

[0090]

[0091] Example 2

[0092] like Figure 5 As shown, this embodiment provides a multi-objective molecular optimization system based on selective sampling learning, including:

[0093] The model initialization module (101) is used to initialize the molecular generation model;

[0094] The first molecular set generation module (102) is used to perform molecular sampling based on the initialized molecular generation model. Each molecular sampling will sample a batch of molecules to form the first molecular set.

[0095] The calculation module (103) is used to calculate the first index, the second index and the third index of each molecule in the first molecule set;

[0096] In a preferred embodiment, the first indicator is a physicochemical property, the second indicator is ADMET, and the third indicator is activity;

[0097] The comparison filtering module (104) is used to compare the first index, the second index, and the third index with the first index threshold, the second index threshold, and the third index threshold respectively according to the preset first index threshold, the second index threshold, and the third index threshold. Based on the comparison results, each molecule in the first molecule set is filtered. Molecules that meet the standards of the first index, the second index, and the third index are retained, and molecules that do not meet the standards of the first index, the second index threshold, and the third index are discarded.

[0098] The second molecular set generation module (105) is used to set a first molecular number threshold, and stop the loop when the number of molecules obtained after filtering is equal to or higher than the first molecular number threshold, and obtain a second molecular set. In this embodiment, the second molecular set is denoted as M1.

[0099] The third molecular set generation module (106) filters the second molecular set again based on the first evaluation criterion to obtain the third molecular set;

[0100] In this embodiment, the first evaluation criterion is: retain the molecules in the second molecular set M1 whose activity is higher than the overall median, and denot them as the third molecular set M2;

[0101] The model training module (107) is used to use the third molecular set for training and parameter updating of the molecular generation model.

[0102] In a preferred embodiment, the system further includes:

[0103] The index evaluation module (108) is used to generate multiple molecules using the trained molecular generation model after the training and parameter update of the molecular generation model are completed, and to evaluate the index of the multiple molecules.

[0104] Example 3

[0105] This paper presents the application of a multi-objective molecular optimization method based on selective sampling learning in drug development.

[0106] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A multi-objective molecular optimization method based on selective sampling learning, characterized in that, include: S1, Initialize the molecular generation model; S2, molecular sampling is performed based on the initialized molecular generation model. Each molecular sampling will sample a batch of molecules to form the first molecular set. S3, calculate the first index, second index and third index of each molecule in the first molecule set; S4, based on the preset first index threshold, second index threshold and third index threshold, the first index, second index and third index are compared with the first index threshold, second index threshold and third index threshold respectively. Based on the comparison results, each molecule in the first molecule set is filtered. Molecules that meet the standards of the first index, second index and third index all meet the standards of the first index threshold, second index threshold and third index threshold are retained. Molecules that do not meet the standards of the first index, second index threshold and third index threshold at any one of the first index, second index and third index threshold are discarded. S5, set a first molecule number threshold, loop S1-S4 until the number of molecules obtained after filtering is equal to or higher than the first molecule number threshold, then stop the loop and obtain a second molecule set; S6, based on the first evaluation criterion, the second molecular set is screened again to obtain the third molecular set; the first evaluation criterion is: retain molecules in the second molecular set M1 whose activity is higher than the overall median as the third molecular set; S7, the third molecular set is used for training and parameter updating of the molecular generation model.

2. The multi-objective molecular optimization method based on selective sampling learning according to claim 1, characterized in that, The first indicator is a physicochemical property, the second indicator is ADMET, and the third indicator is activity.

3. The multi-objective molecular optimization method based on selective sampling learning according to claim 2, characterized in that, The initialized molecular generation model is a deep graph generation model (DGMG), a subgraph generation connection tree variational autoencoder (JTNN), GCPN, MRNN, or PS-VAE.

4. The multi-objective molecular optimization method based on selective sampling learning according to claim 1, characterized in that, Before implementing S1, the initial molecular generation model is pre-trained using the publicly available dataset ZINC.

5. The multi-objective molecular optimization method based on selective sampling learning according to claim 1, characterized in that, The method further includes: S8, after the training and parameter update of the molecular generation model are completed, multiple molecules are generated using the trained molecular generation model, and the multiple molecules are evaluated by indicators.

6. The multi-objective molecular optimization method based on selective sampling learning according to claim 5, characterized in that, S8 includes: S81, Determine the optimization objective for this optimization experiment; S82, determine the scoring function corresponding to each optimization objective; S83, multiple molecules are generated using the trained molecular generation model, and the multiple molecules are evaluated based on the scoring function.

7. A multi-objective molecular optimization system based on selective sampling learning, used to implement the method according to any one of claims 1-6, characterized in that, include: The model initialization module (101) is used to initialize the molecular generation model; The first molecular set generation module (102) is used to perform molecular sampling based on the initialized molecular generation model. Each molecular sampling will sample a batch of molecules to form the first molecular set. The calculation module (103) is used to calculate the first index, the second index and the third index of each molecule in the first molecule set; The comparison filtering module (104) is used to compare the first index, the second index, and the third index with the first index threshold, the second index threshold, and the third index threshold respectively according to the preset first index threshold, the second index threshold, and the third index threshold. Based on the comparison results, each molecule in the first molecule set is filtered. Molecules that meet the standards of the first index, the second index, and the third index are retained, and molecules that do not meet the standards of the first index, the second index, and the third index are discarded. The second molecular set generation module (105) is used to set a first molecular number threshold, and stop the loop when the number of molecules obtained after filtering is equal to or higher than the first molecular number threshold, thereby obtaining a second molecular set. The third molecular set generation module (106) filters the second molecular set again based on the first evaluation criterion to obtain the third molecular set; The model training module (107) is used to use the third molecular set for training and parameter updating of the molecular generation model.

8. The multi-objective molecular optimization system based on selective sampling learning according to claim 7 is characterized in that, The system also includes: The index evaluation module (108) is used to generate multiple molecules using the trained molecular generation model after the training and parameter update of the molecular generation model are completed, and to evaluate the index of the multiple molecules.

9. The application of the selective sampling learning multi-objective molecular optimization method according to any one of claims 1-6 in drug development.

Images