Active cliff data set construction method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of data processing, in particular to an active cliff data set construction method and device, electronic equipment and a storage medium, and aims to improve the accuracy and consistency of active cliff recognition. The method comprises the following steps: for each molecule which is in a pre-constructed molecular data set, can be combined with the same biological target spot and is subjected to biological activity measurement under the same experimental condition, obtaining the similarity between molecular fingerprints of every two molecules in each molecule, and analyzing the similarity of every two molecules on substructures to obtain the bioactivity of the molecules. Selecting a matching molecule pair meeting a preset similarity evaluation condition; for each biological target spot, selecting a similarity as a molecular activity cliff threshold value corresponding to the biological target spot according to distribution conditions of the corresponding matched molecule pairs under different similarities, and selecting a target molecule pair from the molecules according to the molecular activity cliff threshold value; and constructing an activity cliff data set based on the biological activity parameter difference corresponding to each selected target molecule pair.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, electronic device and storage medium for constructing an active cliff dataset. Background Technology

[0002] Activity cliffs (ACs) refer to the phenomenon where a small change in the structure of a group of structurally similar compounds leads to a significant difference in their biological activities, which has important implications for drug discovery and optimization. For example, adding a hydroxyl group (-OH) to a benzene ring to form phenol allows it to form hydrogen bonds with water, greatly increasing its solubility by several orders of magnitude. However, the uncertainties in the definition and assessment of activity cliffs pose significant challenges to computational methods for predicting them.

[0003] Specifically, the definition and assessment of the activity cliff need to consider the similarity of multiple molecules. In related technologies, a commonly used metric between compounds is the Tanimoto similarity coefficient, calculated using structural descriptors (especially 2D molecular fingerprints).

[0004] However, this method relies on a subjective threshold setting, i.e., manually selecting a threshold based on the Tanimoto coefficient to classify compounds based on similarity. However, using a manually set Tanimoto threshold to classify the similarity between molecules under different molecular characterization methods may lead to discrepancies in the activity cliff assessment results, making the analysis of the activity cliff with large molecular datasets more challenging and prone to more problems.

[0005] In summary, the subjectivity and uncertainty of similarity threshold setting affect the accuracy and consistency of active cliff identification. Summary of the Invention

[0006] This application provides a method, apparatus, electronic device, and storage medium for constructing an active cliff dataset to improve the accuracy and consistency of active cliff identification.

[0007] This application provides a method for constructing an active cliff dataset, including:

[0008] Obtain a pre-constructed molecular dataset; the molecular dataset contains multiple molecules, at least one biological target that each molecule can bind to, and a bioactivity parameter for each molecule, the bioactivity parameter being measured under at least one experimental condition;

[0009] For each molecule in the molecular dataset that can bind to the same biological target and whose bioactivity can be measured under the same experimental conditions, the similarity between the molecular fingerprints of each pair of molecules is obtained; and, by analyzing the similarity of each pair of molecules in their substructures, matching molecule pairs that meet the preset similarity evaluation conditions are selected.

[0010] For each biological target, based on the distribution of each matching molecule pair under different similarities, a similarity is selected as the molecular activity cliff threshold corresponding to the biological target, and target molecule pairs are selected from each molecule based on the molecular activity cliff threshold.

[0011] Based on the differences in the corresponding bioactivity parameters of each selected target molecule pair, we analyzed whether there was an activity cliff relationship between the target molecule pairs, and constructed an activity cliff dataset based on the analysis results.

[0012] This application provides an apparatus for constructing an active cliff dataset, comprising:

[0013] The first acquisition unit is used to acquire a pre-constructed molecular dataset; the molecular dataset contains multiple molecules, at least one biological target that each molecule can bind to, and a bioactivity parameter for each molecule, wherein the bioactivity parameter is measured under at least one experimental condition;

[0014] The second acquisition unit is used to acquire the similarity between the molecular fingerprints of each pair of molecules in the molecular dataset that can bind to the same biological target and whose bioactivity can be measured under the same experimental conditions; and to select matching molecule pairs that meet the preset similarity evaluation conditions by analyzing the similarity of each pair of molecules in their substructures.

[0015] The threshold determination unit is used to select a similarity as the molecular activity cliff threshold corresponding to the biological target for each biological target based on the distribution of each matching molecule pair under multiple similarities, and select the target molecule pair from each molecule based on the molecular activity cliff threshold.

[0016] The construction unit is used to analyze whether there is an activity cliff relationship between the selected target molecule pairs based on the differences in the corresponding bioactivity parameters, and to construct an activity cliff dataset based on the analysis results.

[0017] Optionally, the threshold determination unit is specifically used for:

[0018] Count the number of matching pairs at each similarity level;

[0019] The number of matched pairs at each similarity level is accumulated in descending order of similarity.

[0020] The similarity determined when the proportion of accumulated matched molecule pairs reaches the target proportion is used as the molecular activity cliff threshold corresponding to the biological target; the proportion of accumulated matched molecule pairs represents the ratio of the number of accumulated matched molecule pairs to the total number of matched molecule pairs under the biological target.

[0021] Optionally, the threshold determination unit is specifically used for:

[0022] For each pair of molecules, if the similarity between the two molecules is greater than the molecular activity cliff threshold, then the two molecules are considered as a target molecule pair.

[0023] Optionally, the building unit is specifically used for:

[0024] For each target molecule pair, if the difference in bioactivity parameters between the molecules is greater than a preset difference threshold, then it is determined that there is an activity cliff relationship between the molecules of the target molecule pair, and the target molecule pair is taken as a positive sample.

[0025] If the difference in bioactivity parameters between molecules is not greater than a preset difference threshold, then it is determined that there is no activity cliff relationship between the molecules of the target molecule pair, and the target molecule pair is taken as a negative sample.

[0026] Based on the positive and negative samples, an active cliff dataset is constructed for training an active cliff prediction model. The active cliff prediction model is used to predict whether an active cliff relationship exists between input molecular pairs.

[0027] Optionally, the device further includes:

[0028] The post-processing unit is configured to perform at least one of the following post-processing steps on each biological target:

[0029] If the number of target molecule pairs selected under the biological target is less than the first preset number, then each target molecule pair under the biological target is deleted;

[0030] If the number of positive samples corresponding to the biological target is less than the second preset number, then each target molecule pair under the biological target is deleted;

[0031] The second preset quantity is less than the first preset quantity.

[0032] Optionally, the first acquisition unit is specifically used for:

[0033] Obtain known molecular databases; the molecular databases are open-source databases used for molecular biological activity research;

[0034] Based on the first and second screening criteria, molecules and related data that meet the criteria are selected from the molecular database and the molecular dataset is formed.

[0035] The first screening condition is used to limit the data quality and experimental type of the required molecules; the second screening condition is used to limit the activity index and purity of the required molecules; the relevant data includes the biological targets that each molecule can bind to, and the biological activity parameters of each molecule; the activity index represents the data type of the biological activity parameters.

[0036] Optionally, before assembling the molecular dataset, the first acquisition unit is further configured to:

[0037] For a single molecule targeting a biological site:

[0038] If the molecule has multiple bioactivity parameters under the same activity index, and the variance of the multiple bioactivity parameters is greater than a preset difference threshold, then the multiple bioactivity parameters are deleted.

[0039] If the molecule has multiple bioactivity parameters under the same activity index, and the variance of the multiple bioactivity parameters is not greater than the preset difference threshold, then the mean of the multiple bioactivity parameters is taken as the bioactivity parameter of the molecule under the activity index.

[0040] Optionally, the second acquisition unit is specifically used for:

[0041] Each molecule is cut by performing acyclic single bond cleavage, dividing each molecule into two parts: an invariant core substructure and a region molecular structure. The invariant core substructure is the common part between molecules, and the region molecular structure is the unique part of each molecule.

[0042] Each pair of molecules with the same invariant core substructure is grouped into a candidate molecule pair.

[0043] By using preset similarity assessment conditions, each candidate molecule pair is screened to obtain matching molecule pairs; the similarity assessment conditions are used to evaluate the atomic differences and differences in bioactivity parameters between the molecular structures in the candidate molecule pairs.

[0044] An electronic device provided in this application includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of any of the above-described methods for constructing an active cliff dataset.

[0045] This application provides a computer-readable storage medium including a computer program. When the computer program is run on an electronic device, the computer program is used to cause the electronic device to perform the steps of any of the above-described methods for constructing an active cliff dataset.

[0046] This application provides a computer program product, which includes a computer program stored in a computer-readable storage medium. When a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, causing the electronic device to perform the steps of any of the above-described methods for constructing an active cliff dataset.

[0047] The beneficial effects of this application are as follows:

[0048] This application provides a method, apparatus, electronic device, and storage medium for constructing an activity cliff dataset, and establishes a set of molecular activity cliff screening and evaluation criteria. Specifically, the molecular dataset pre-constructed in this application includes the biological targets that each molecule can bind to and the bioactivity parameters of each molecule. Based on this, molecules that can bind to the same biological target and whose bioactivity is measured under the same experimental conditions are compared and their similarity is calculated and molecular pairs are divided. Specifically, molecular fingerprints are used to obtain the similarity between every two molecules, and by analyzing the similarity of every two molecules in their substructures, matching molecular pairs that meet the preset similarity evaluation criteria are selected. In this step, this application comprehensively considers the fingerprint similarity criteria between molecules and the substructure-based similarity evaluation criteria. The fingerprint similarity criteria between molecules simplifies complex molecular structures into binary or numerical vectors, enabling efficient calculation of molecular similarity, while the substructure-based similarity evaluation criteria focus on molecular functional regions, accurately identifying and matching key pharmacophores or functional units in chemical structures, which can improve the accuracy of active compound discovery.

[0049] Based on the above, for each biological target, a molecular activity cliff threshold is set by matching the distribution of molecular pairs under different similarities. Different biological targets correspond to different thresholds, which effectively avoids the subjective method of setting a uniform threshold in related technologies. It also overcomes the limitation of focusing only on changes at a single site in the substructure similarity evaluation standard of related technologies by combining the fingerprint similarity criteria between molecules.

[0050] Finally, for each molecular pair under each biological target, target molecular pairs are selected based on the molecular activity cliff threshold. On this basis, an activity cliff dataset is constructed, which effectively improves the accuracy and consistency of activity cliff identification.

[0051] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0052] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0053] Figure 1 This is a schematic diagram illustrating an application scenario of a method for constructing an active cliff dataset according to an embodiment of this application.

[0054] Figure 2 A flowchart illustrating the implementation of a method for constructing an active cliff dataset, as provided in this application embodiment;

[0055] Figure 3A A schematic diagram illustrating the similarity between molecular fingerprints provided in an embodiment of this application;

[0056] Figure 3B A schematic diagram illustrating the similarity of a stent provided in an embodiment of this application;

[0057] Figure 3C A schematic diagram illustrating the similarity of SMILES strings provided in an embodiment of this application;

[0058] Figure 4A A schematic diagram of a topological cliff provided for an embodiment of this application;

[0059] Figure 4B A schematic diagram of a chiral cliff provided for an embodiment of this application;

[0060] Figure 4C A schematic diagram of a substituent cliff provided in an embodiment of this application;

[0061] Figure 4D A schematic diagram of a support cliff provided in an embodiment of this application;

[0062] Figure 4E A schematic diagram of a support / topological cliff provided in an embodiment of this application;

[0063] Figure 5A A schematic diagram of a matching molecule pair cliff provided for an embodiment of this application;

[0064] Figure 5B A schematic diagram of another matching molecule pair cliff provided for an embodiment of this application;

[0065] Figure 6A A schematic diagram of a matching molecule pair provided in an embodiment of this application. Figure 1 ;

[0066] Figure 6B A schematic diagram of a matching molecule pair provided in an embodiment of this application. Figure 2 ;

[0067] Figure 6C Schematic diagram three of a matching molecule pair provided for an embodiment of this application;

[0068] Figure 6D Schematic diagram four of a matching molecule pair provided for an embodiment of this application;

[0069] Figure 6E Schematic diagram five of a matching molecule pair provided for an embodiment of this application;

[0070] Figure 6F A schematic diagram of a matching molecule pair provided for an embodiment of this application is shown in Figure 6.

[0071] Figure 7 A schematic diagram illustrating the number of matched molecule pairs provided in an embodiment of this application;

[0072] Figure 8A A histogram for determining the molecular activity cliff threshold is provided in an embodiment of this application;

[0073] Figure 8B Another histogram for determining the molecular activity cliff threshold provided in this application embodiment;

[0074] Figure 8C Another histogram for determining the molecular activity cliff threshold provided in this application embodiment;

[0075] Figure 9 Histograms of different methods for calculating intermolecular similarity at different molecular activity cliff thresholds are provided for embodiments of this application.

[0076] Figure 10 A histogram showing the distribution of similarity at different biological targets, provided for embodiments of this application;

[0077] Figure 11 A scatter plot of target molecule pairs at different target ratios is provided for embodiments of this application;

[0078] Figure 12 A schematic diagram illustrating a method for constructing an active cliff dataset provided in an embodiment of this application;

[0079] Figure 13 This is a schematic diagram of the structure of an active cliff prediction model provided in an embodiment of this application;

[0080] Figure 14 Histograms of the performance of two datasets, ACDataBase and ACNet, in a prediction liveness cliff task are provided in an embodiment of this application.

[0081] Figure 15 Histogram of the performance of two datasets, ACDataBase and ACNet, in the prediction liveness cliff task under a scenario two provided in this application embodiment;

[0082] Figure 16 This is a schematic diagram of the composition structure of another active cliff dataset construction device in an embodiment of this application;

[0083] Figure 17 This is a schematic diagram of the hardware structure of an electronic device using an embodiment of this application. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0085] The following describes some of the concepts involved in the embodiments of this application.

[0086] Biological targets are molecules or structures within an organism that possess specific functions, such as proteins (e.g., enzymes, receptors, ion channels), nucleic acids (deoxyribonucleic acid (DNA), ribonucleic acid (RNA)), or other molecular complexes. In drug development, biological targets are the objects to which drugs act; drugs bind to or modulate their activity to achieve the goal of treating diseases.

[0087] Molecular fingerprints are binary vectors where each site represents the presence of a specific structural feature in a molecule, such as a particular atom type, bond type, or substructure. Values ​​typically range from 0 to 1. In this embodiment, by representing molecules as molecular fingerprints, the similarity between molecular pairs can be determined.

[0088] Molecular activity cliff threshold: This is a standard used to determine whether two molecules are sufficiently similar. It is generally based on the similarity between the molecular fingerprints of the two molecules. If the similarity is greater than the molecular activity cliff threshold, the two molecules are considered sufficiently similar; otherwise, they are considered not similar enough. In this application, a combination of similarity criteria based on molecular fingerprints and similarity criteria based on substructure similarity is used to determine different molecular activity cliff thresholds for different biological targets.

[0089] Non-cyclic single bond cleavage: In cheminformatics, non-cyclic single bond cleavage is a method for analyzing molecular structure, primarily used to identify and separate core substructures and region structures within molecules to study their roles in chemical reactions, drug design, or molecular recognition. This application utilizes this process to construct matched molecular pairs (MMPs) and conduct activity cliff analysis to understand how structural changes affect the biological activity of molecules.

[0090] The invariant core is the part shared by the molecules after a non-cyclic single bond is cleaved. This core remains unchanged after cleavage and does not contain any of the cleaved bonds. The invariant core is a structurally stable and functionally conserved region within the molecule, significantly influencing its function and properties, and playing a fundamental role in its chemical properties and biological activity.

[0091] Differentiating fragments (GFs) are the unique portions of a molecule pair formed after cleavage at non-cyclic single bonds; they are molecular fragments not contained within the invariant core substructure. Even small structural changes in these GFs can lead to significant differences in molecular properties (such as solubility, metabolic stability, and biological activity). In activity cliff analysis, identifying GFs is crucial for understanding why structurally similar molecules exhibit significantly different biological activities. By comparing the invariant core substructure and GFs, we can gain a deeper understanding of the impact of structural changes on molecular properties, which is significant for drug design, chemical synthesis pathway planning, and understanding chemical reaction mechanisms. For example, in drug discovery, identifying the invariant core and GFs in the activity cliff phenomenon can help design new molecules with optimized biological activity.

[0092] Bioactivity parameters are used in drug design, pharmacology, and biochemistry to describe the strength of the interaction between molecules and biological targets. These include, but are not limited to, half-maximal inhibitory concentration (IC50), dissociation constant for inhibition (Ki), dissociation constant (Kd), half-maximal effective concentration (EC50), effective dose (ED50), median lethal dose (LD50), etc. In this application's embodiments, Ki, Kd, ​​and IC50 values ​​are generally used as examples; other bioactivity parameters are similar and will not be elaborated further.

[0093] IC50 is an important parameter obtained in bioactivity assays, representing the concentration at which an inhibitor achieves 50% of its maximum inhibitory effect. A lower IC50 value indicates a stronger inhibitory effect on the target organism. IC50 is commonly used to determine the activity of enzyme inhibitors, receptor antagonists, or cell proliferation inhibitors.

[0094] Ki: This describes the affinity between the inhibitor and the enzyme, and is a parameter that more directly reflects the strength of the binding between the inhibitor and the enzyme. The lower the Ki value, the stronger the binding between the inhibitor and the enzyme, and the stronger the inhibitory effect. Ki can be obtained through biochemical experiments on different types of inhibition (such as competitive, non-competitive, or uncompetitive inhibition).

[0095] Kd is an important parameter describing the binding strength between ligands (such as drug molecules) and receptors, antibodies and antigens, enzymes and substrates, and other biomolecules. The Kd value is defined as the ratio of the product of the free ligand concentration and the ligand-receptor complex concentration to the total ligand-receptor complex concentration at equilibrium. A lower Kd value indicates that a higher binding saturation can be achieved at a lower ligand concentration; that is, the ligand and receptor bind more tightly and the interaction is stronger. Conversely, a higher Kd value indicates weaker ligand-receptor binding and easier dissociation.

[0096] Bioassay methods are experimental techniques used to assess the strength of a compound's effects on biological systems. These methods can be used to measure a variety of biological effects of a compound, including but not limited to its ability to inhibit, activate, modulate, or interfere with specific biological targets (such as enzymes, receptors, ion channels, etc.). Experimental measurements, such as IC50, Ki, and Kd, are key values ​​obtained in these bioassays, and are the bioassay parameters discussed in this paper. They quantify the inhibitory effect of a compound on a biological target. IC50, Ki, and Kd are all obtained through bioassays, but they represent different measurements of biological effects.

[0097] The design concept of the embodiments in this application is briefly introduced below:

[0098] The compound optimization phase in early drug discovery is a crucial part of the drug development process. The main goal of this phase is to transform initially screened biologically active but potentially defective compounds (i.e., lead compounds) into drug candidates with better pharmacological properties, safety, and drugability. The structure-activity relationship (SAR) study within the compound optimization phase involves systematically altering the chemical structure of the original lead compounds through chemical modification and observing how these changes affect the compound's biological activity. This aims to reveal the relationship between structure and activity and identify key structural features contributing to the activity.

[0099] The discontinuity of SAR and the activity cliff can often reveal key factors determining SAR, especially during the compound optimization stage. Therefore, in the compound optimization stage of early drug discovery, accurate prediction of the activity cliff becomes particularly important to capture subtle chemical changes that have a significant impact on specific biological activities.

[0100] However, the uncertainties in the definition and evaluation of the activity cliff pose a significant challenge to computational methods in predicting the activity cliff.

[0101] The definition and evaluation of activity cliffs require consideration of differences in similarity and potency among multiple molecules. Unlike predicting individual molecules, the challenge in predicting compound activity cliffs lies in considering compound pairs rather than individual molecules. Therefore, determining the similarity among multiple molecules is crucial. In related techniques, the similarity among multiple molecules is defined using Tanimoto coefficient similarity values ​​calculated from structural descriptors (especially 2D molecular fingerprints). However, this method requires manually setting thresholds for the Tanimoto coefficient similarity values, and different thresholds can significantly impact the evaluation of activity cliffs. These uncertainties in evaluation mean that computational analysis of activity cliffs in large compound datasets is challenging and potentially problematic.

[0102] In view of this, embodiments of this application provide a method, apparatus, electronic device, and storage medium for constructing an activity cliff dataset, and establish a set of molecular activity cliff screening and evaluation criteria. Specifically, the molecular dataset pre-constructed in this application includes the biological targets that each molecule can bind to and the bioactivity parameters of each molecule; based on this, molecules that have undergone bioactivity measurement under the same experimental conditions for the same biological target are evaluated for similarity between molecules and molecular pairs are divided. Specifically, molecular fingerprints are used to obtain the similarity between every two molecules, and matching molecular pairs are constructed based on the substructure similarity of each molecule. In this step, this application comprehensively considers the fingerprint similarity criteria between molecules and the substructure-based similarity evaluation criteria; the fingerprint similarity criteria between molecules simplifies complex molecular structures into binary or numerical vectors, enabling efficient calculation of molecular similarity, while the substructure-based similarity evaluation criteria focus on molecular functional regions, accurately identifying and matching key pharmacophores or functional units in chemical structures, which can improve the accuracy of active compound discovery.

[0103] Based on the above, for each biological target, a molecular activity cliff threshold is set by matching the distribution of molecular pairs under different similarities. Different biological targets correspond to different thresholds, which effectively avoids the subjective method of setting a uniform threshold in related technologies. It also overcomes the limitation of focusing only on changes at a single site in the substructure similarity evaluation standard of related technologies by combining the fingerprint similarity criteria between molecules.

[0104] Finally, for each molecular pair under each biological target, target molecular pairs are selected based on the molecular activity cliff threshold. On this basis, an activity cliff dataset is constructed, which effectively improves the accuracy and consistency of activity cliff identification.

[0105] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0106] like Figure 1 The diagram shown is an application scenario illustration of an embodiment of this application. The application scenario illustration includes a terminal device 110 and a server 120.

[0107] In this embodiment, the terminal device 110 includes, but is not limited to, mobile phones, tablets, laptops, desktop computers, e-book readers, smart voice interaction devices, smart home appliances, and in-vehicle terminals. The terminal device may have a client installed related to building the active cliff dataset. This client can be software (e.g., a browser, data processing software), a webpage, or a mini-program. The server 120 is the backend server corresponding to the software, webpage, or mini-program, or a server specifically used for building the active cliff dataset; this application does not impose specific limitations. The server 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0108] It should be noted that the method for constructing the active cliff dataset in each embodiment of this application can be executed by an electronic device, which can be a terminal device 110 or a server 120. That is, the method can be executed by the terminal device 110 or the server 120 alone, or by the terminal device 110 and the server 120 together.

[0109] For example, when executed jointly by terminal device 110 and server 120, the object selects a molecular database on the data processing software installed on terminal device 110 that can be used to construct the activity cliff dataset, and enters the screening criteria for the molecular database before sending it to server 120. Server 120 filters the molecular database based on the above screening criteria to obtain the molecular dataset used for constructing the activity cliff dataset. This molecular dataset contains multiple molecules, the biological targets that each molecule can bind to, and the bioactivity parameters of each molecule. Next, server 120 obtains each pair of molecules that have undergone bioactivity measurement under the same experimental conditions for the same biological target. The server 120 assesses the similarity between molecular fingerprints of molecules and selects matching molecule pairs that meet preset similarity evaluation criteria by analyzing the similarity of substructures between each pair of molecules. Then, for each biological target, the server 120 selects a similarity threshold as the molecular activity cliff threshold based on the distribution of each matching molecule pair under different similarity levels, and selects target molecule pairs from each molecule based on the molecular activity cliff threshold. Finally, the server 120 analyzes whether there is an activity cliff relationship between the selected target molecule pairs based on the differences in bioactivity parameters, and constructs an activity cliff dataset based on the analysis results. This activity cliff dataset can be used for subsequent tasks such as model training, drug discovery, and compound optimization. The server 120 can send it to the terminal device 110 for viewing.

[0110] In one alternative implementation, the terminal device 110 and the server 120 can communicate via a communication network.

[0111] In one alternative implementation, the communication network is a wired network or a wireless network.

[0112] It should be noted that, Figure 1 The examples shown are merely illustrative; in reality, the number of terminal devices and servers is unlimited and is not specifically limited in the embodiments of this application.

[0113] In this embodiment of the application, when there are multiple servers, the multiple servers can form a blockchain, and the server is a node on the blockchain; as disclosed in the embodiment of the application, the molecular data involved can be stored on the blockchain, such as molecular datasets, similarity between molecular fingerprints, molecular activity cliff thresholds, activity cliff datasets, etc.

[0114] Furthermore, the embodiments of this application can be applied to various scenarios, including not only pharmaceutical scenarios, but also scenarios such as cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0115] Specifically, the embodiments of this application are mainly applied to pharmaceutical scenarios such as drug discovery, drug optimization, and drug synthesis. The application scenarios of the embodiments of this application are briefly described below:

[0116] (1) Compound screening and optimization.

[0117] In the early stages of drug discovery, it is necessary to screen candidate compounds with potential efficacy from a large compound library. Activity cliff prediction can help identify compounds that are structurally very similar but have significantly different activities. By analyzing activity cliff datasets, it is possible to identify tiny structures that cause significant changes in molecular activity, thus allowing for greater focus on these structures during compound screening. This helps avoid overlooking key active compounds and improves screening efficiency.

[0118] (2) Optimization of lead compounds.

[0119] Lead compounds are compounds that have been preliminarily discovered to possess biological activity. Further optimization of their structure can improve efficacy and reduce side effects. Activity cliff prediction can provide guidance during the optimization process, avoiding unnecessary or ineffective modifications to compounds, thus saving time and resources. Analyzing activity cliff datasets can help understand the relationship between molecular activity and molecular structure, avoid activity traps (compounds that may cause a sudden loss of activity during further optimization), explore the diversity of molecular structures, and more.

[0120] (3) Toxicity prediction and avoidance.

[0121] In drug discovery, identifying and avoiding potentially toxic compounds is crucial. Molecular activity cliff datasets can help identify compounds that exhibit significant variations between activity and toxicity, thereby enabling the early avoidance of potentially toxic drug candidates and improving their safety.

[0122] (4) SAR research.

[0123] Analyzing activity cliff datasets can enhance research on molecular structure-activity relationships, leading to a better understanding of the connection between compound structure and biological activity. By identifying how small structural changes can lead to significant changes in activity, compounds can be designed and optimized more effectively, and new chemical skeletons or pharmacophores can be discovered, facilitating the development of novel drugs.

[0124] (5) Target selection and verification.

[0125] In the early stages of drug discovery, selecting appropriate biological targets is crucial. Molecular activity cliff datasets can help identify biological targets that may be significantly associated with changes in compound activity, thus facilitating target selection and validation processes and improving the relevance and effectiveness of target selection.

[0126] (6) Accelerate the identification of candidate drugs.

[0127] Analyzing activity cliff datasets can predict which compounds are likely to become activity cliffs, enabling faster identification of drug candidates. This process can accelerate the overall drug discovery process, allowing promising candidates to enter preclinical research and development stages more quickly.

[0128] (7) Assisted virtual screening.

[0129] In virtual screening, computational models screen potential drugs by predicting the bioactivity of compounds. By incorporating an activity cliff dataset, the accuracy and reliability of virtual screening can be significantly improved, resulting in the selection of higher-quality candidate compounds.

[0130] (8) Construct and optimize the quantitative structure-activity relationship (QSAR) model.

[0131] Analyzing the activity cliff dataset can provide further insight into the functional groups or chemical bonds that cause significant changes in activity, which is crucial for building more accurate QSAR models and improving their predictive power.

[0132] (9) Drug repositioning.

[0133] Analyzing activity cliff datasets can identify compounds with similar structures but different activities, thereby uncovering potential new indications for existing drugs. For example, a drug may be ineffective against one disease, but if an activity cliff exists, its structural analogues may be effective against another disease.

[0134] It should be noted that the application scenarios listed above are just simple examples. Other application scenarios are also applicable to the embodiments of this application, and will not be described in detail here.

[0135] The following describes the method for constructing an active cliff dataset provided by an exemplary embodiment of this application, in conjunction with the application scenarios described above and with reference to the accompanying drawings. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of this application, and the implementation of this application is not limited in any way in this respect.

[0136] It should be noted that the method for constructing the active cliff dataset provided in this application embodiment can be based on the Python language and deployed on a server equipped with a Linux or Windows operating system and CPU computing resources; it can also be based on the C language and deployed on a server equipped with a macOS operating system, etc. This article does not make any specific limitations on this.

[0137] See Figure 2 The diagram shown is an implementation flowchart of a method for constructing an active cliff dataset according to an embodiment of this application. Taking the server as the execution entity as an example, the specific implementation flow of this method is as follows: S21 to S24:

[0138] S21: Obtain a pre-constructed molecular dataset; the molecular dataset contains multiple molecules, at least one biological target that each molecule can bind to, and bioactivity parameters of each molecule, which are measured under at least one experimental condition.

[0139] Each experimental condition corresponds to at least one activity index, and each activity index includes at least one bioactivity parameter. If a molecule is subjected to multiple bioactivity tests under the same experimental conditions, then each activity index corresponding to that experimental condition can correspond to multiple bioactivity parameters.

[0140] Among these, experimental conditions, also known as control conditions, are required for bioactivity testing. This includes setting appropriate temperatures, pH values, and solvent systems, and ensuring the stability of the compound within the test concentration range in order to accurately assess its biological effects.

[0141] Each biological target can bind to multiple molecules, which may have similar or different structures, but can all specifically interact with the target, affecting its biological function and thus achieving therapeutic effects or producing side effects. These molecules include drugs, metabolites, natural molecules, and signaling molecules.

[0142] Accordingly, each molecule can bind to at least one biological target. For example, molecule a can bind to both biological target 1 and biological target 2, and so on. Taking protein targets as an example, the imatinib molecule can bind to and inhibit two different protein targets: the BCR-ABL fusion protein and the c-KIT receptor tyrosine kinase. These two proteins play key roles in the treatment of different diseases. By inhibiting the activity of these proteins, imatinib can effectively control the progression of these diseases.

[0143] In the embodiments of this application, bioactivity parameters such as IC50, Ki, Kd, ​​EC50, ED50, and LD50 are key indicators for evaluating molecules as potential drug candidates. They are directly related to whether the molecule can effectively and stably bind to the target biological site, thereby exerting the expected biological effect. In other words, bioactivity parameters directly quantify the strength and stability of the interaction between the molecule and the target, reflecting the affinity of the molecule for the biological target. The activity index in this application represents the data type of bioactivity parameters, which can be Ki type, Kd type, IC50 type, etc., and can therefore also be called binding affinity index. For a molecule pair, the specific numerical difference of the bioactivity parameter under the same activity index can reflect the difference in efficacy of the molecule pair.

[0144] Common activity indicators include, but are not limited to, Ki type, Kd type, IC50 type, LD50 type, Hill Slope type, etc.

[0145] It should be noted that the molecular dataset in this application embodiment is used to construct an activity cliff dataset. An open-source database used for molecular biological activity research can be directly used as the molecular dataset, or a molecular dataset that better meets actual business needs can be obtained by filtering existing open-source databases based on certain conditions. Taking the filtered molecular dataset as an example, the construction method of the activity cliff dataset provided in this application embodiment can be divided into the following four steps: (1) data preprocessing; (2) constructing matching molecular pairs based on molecular fingerprints and substructure similarity; (3) calculating the molecular activity cliff threshold for each biological target; and (4) data postprocessing.

[0146] Among them, the molecular dataset in S21 is obtained based on the data preprocessing step (1), S22 is the step of constructing matching molecular pairs based on molecular fingerprint and substructure similarity (2), S23 is the step of calculating the molecular activity cliff threshold of each biological target (3), and the activity cliff dataset in S24 is obtained based on the data postprocessing step (4).

[0147] The data preprocessing steps will be described in detail below:

[0148] In the data preprocessing step, an open-source database for molecular biological activity research is first selected, which contains a large amount of molecular data. Examples of open-source databases include BindingDB, PubChem, DrugBank, and CHEMBL. After downloading molecular data from the open-source database, it is initially screened using pre-set criteria to obtain a more reliable molecular dataset. The following description uses CHEMBL as an example to illustrate the construction method of the activity cliff dataset provided in this application.

[0149] The CHEMBL (Chemical Blueprint Library) database contains a large amount of data on small molecule compounds and their biological activities, and is widely used in drug discovery and chemical biology research. This application specifically uses version 33 of this database (ChEMBL version 33, ChEMBL v33).

[0150] Specifically, ChEMBL is a widely used open-source database that collects and organizes medicinal chemistry and bioactivity data for drug discovery and chemical biology research. ChEMBL v33 data entries typically include, but are not limited to, the following types of information: (1) Compound data: the chemical structure of the compound (represented in Simplified Molecular Input Line Entry System (SMILES) or Structure Data File (SDF) format), molecular weight, molecular formula, hydrophobicity parameter (logP), number of hydrogen bond donors and acceptors, and other physicochemical properties. In this article, "molecule" refers to the molecules that make up the compound, and each molecule can be understood as a compound. (2) Bioactivity data: data on the interaction between the compound and specific biological targets (such as proteins, enzymes, receptors, etc.), including but not limited to binding affinity (such as Ki, Kd, ​​IC50, etc.), activity concentration (such as EC50, IC50), mode of action (such as agonist, antagonist, inhibitor) and experimental conditions. (3) Target information: Provide detailed target descriptions, including target identifier (ID), species, sequence, gene name, mechanism of action, target classification and functional annotation.

[0151] In this application embodiment, when obtaining molecular datasets through data preprocessing, an optional implementation method is as follows:

[0152] Obtain known molecular databases; molecular databases are open-source databases used for molecular biological activity research; based on the first and second screening criteria, select molecules and related data that meet the criteria from the molecular databases and form a molecular dataset.

[0153] The first screening criterion is used to limit the data quality and experimental type of the required molecules; the second screening criterion is used to limit the activity index and purity of the required molecules; the relevant data includes the biological targets that each molecule can bind to, and the biological activity parameters of each molecule; the activity index represents the data type of the biological activity parameters.

[0154] Specifically, the first screening criterion can specify data quality, i.e., data reliability, to ensure that the selected molecular data is highly reliable. Furthermore, the first screening criterion can also specify the experimental type, including the type of experiment the molecule belongs to, such as dose-response experiments, concentration-effect experiments, time-effect experiments, etc., and the type of experiment, such as conjugate experiments, control experiments, crossover experiments, etc., to select experimental data that meets actual business needs. Additionally, the first screening criterion can also specify the concentration unit of the molecule to select experimental data that meets affinity and potency requirements, and so on.

[0155] The second screening criteria can specify activity indicators to select experimental data with activity indicators that meet actual business needs. In addition, the second screening criteria can also specify purity to ensure that the molecule does not contain certain substances, such as salts, polymers, heavy metals, etc. The second screening criteria also limit the degree of measurement variability of the data. The degree of measurement variability represents the difference between multiple measurements of the same molecule under the same experimental conditions. It can reflect the coefficient of variation or data fluctuation of the molecule during the experiment. For example, it can be limited by variance, standard deviation, etc.

[0156] The relevant data includes the biological targets that each molecule can bind to, as well as the biological activity parameters of each molecule. For example, for a molecule that binds to multiple targets, there will be a corresponding record for each biological target that it can bind to. Each record will contain detailed information about the interaction between the molecule and the specific biological target, such as binding affinity (e.g., Ki, Kd, ​​IC50, etc.), mechanism of action, experimental conditions, biological activity parameters, etc.

[0157] In this application embodiment, when limiting the degree of measurement variability of data through a second screening condition, an optional implementation method is as follows:

[0158] For a molecule targeting a biological target: if the molecule has multiple biological activity parameters under the same activity index, and the variance of the multiple biological activity parameters is greater than a preset difference threshold, then the multiple biological activity parameters are deleted; if the molecule has multiple biological activity parameters under the same activity index, and the variance of these multiple biological activity parameters is not greater than the preset difference threshold, then the mean of these multiple biological activity parameters is taken as the biological activity parameter of the molecule under this activity index.

[0159] Specifically, the preset difference threshold is an empirical value, such as 8, 9, or 10. The activity index can be of type IC50, Ki, or Kd, etc. The activity index corresponds to the experimental conditions, and each experimental condition includes at least one activity index. A bioactivity parameter refers to the experimental result obtained by measuring a specific activity index.

[0160] When limiting the degree of measurement variability, the affinity values ​​of activity indicators can be aggregated based on information such as experimental conditions (assay), biological target (target), and molecule (molecule). If an activity indicator of a molecule under a biological target has multiple biological activity parameters, then for that molecule, the average value and variance of the multiple biological activity parameters measured under multiple batches of experiments are calculated. Then, by analyzing the variance, it is determined whether the degree of measurement variability is within the range of difference from the average value indicated by the preset difference threshold.

[0161] If the variance is less than or equal to a preset difference threshold (let's say 10), the efficacy difference is considered to be within one order of magnitude, and these measurements are considered reliable. Therefore, the average of these multiple bioactivity parameters is used as the bioactivity parameter of the molecule under that activity index to construct the molecular dataset and the subsequent activity cliff dataset. If the variance is greater than the preset difference threshold, it indicates that the efficacy difference of these multiple bioactivity parameters is large, meaning that these measurements are unreliable. In this case, these multiple bioactivity parameters can be directly deleted.

[0162] For example, if the activity index is of the IC50 type and the preset difference threshold is 10, if the mean of multiple bioactivity parameters of molecule B under biological target A in the IC50 type is calculated to be 1 nanomolar (nM) and the variance is 11.3, then multiple bioactivity parameters are deleted; if the variance is 9.5, then the mean of 1 nM is taken as the bioactivity parameter of the molecule in the IC50 index.

[0163] Furthermore, if the variance exceeds a preset difference threshold, instead of directly deleting all bioactivity parameters under that activity index, only deleting some with significant differences can ensure that the variance of the remaining bioactivity parameters is less than the preset difference threshold. Therefore, in this application embodiment, when limiting the measurement variability of data through the second screening condition, another optional implementation method is:

[0164] If a molecule has multiple bioactivity parameters under the same activity index, and the variance of the multiple bioactivity parameters is greater than a preset difference threshold, then the bioactivity parameters whose measured variability exceeds the difference from the mean indicated by the preset difference threshold are deleted, and the mean of the remaining bioactivity parameters is taken as the bioactivity parameter of the molecule under this activity index.

[0165] Specifically, if the variance of multiple bioactivity parameters is greater than a preset difference threshold, then the bioactivity parameters whose measurement variability exceeds the preset difference threshold are deleted, so that the variance of the remaining multiple bioactivity parameters is less than the preset difference threshold.

[0166] For example, if the activity index is of the IC50 type and the preset difference threshold is 10, if the mean of multiple bioactivity parameters of molecule B under biological target A in the IC50 type is calculated to be 1 nanomolar (nM) and the variance is 11.3, then the bioactivity parameters with the larger degree of measurement variability are deleted until the variance of the remaining bioactivity parameters is less than or equal to the preset difference threshold. Then the mean of the remaining bioactivity parameters is taken as the bioactivity parameter of the molecule under this activity index.

[0167] In the above implementation, by setting a preset difference threshold for the variance of multiple bioactivity parameters, data with large variations are removed. This removes abnormal data caused by unstable experimental conditions and inaccurate measurement methods in the molecular dataset, improving the reliability and consistency of the data in the molecular dataset. This reduces the risk of drawing erroneous conclusions, enhances the accuracy and reproducibility of the research, and ensures that subsequent analyses are based on stable and reliable data.

[0168] The first and second screening conditions are further described below through specific examples.

[0169] For example, taking the open-source database CHEMBL version 33 as an example, assuming 20,334,684 molecules were downloaded from CHEMBL version 33, the first screening criteria include, but are not limited to:

[0170] The data is presented in the following format: Relationship = D, Confidence score = 9, concentration unit = nanomolar, and Assay type = B.

[0171] Wherein, Relationship = D indicates the labeling of the relationship between activity data and biological targets. D stands for dose-response, indicating that the measured data should come from a dose-response experiment, which typically provides a series of molecular concentrations and observes their effects on the biological target; Confidence_score = 9 indicates a data confidence score of 9, which is highly reliable data. Confidence_score is used to indicate the confidence level of the association between the target and molecular activity data, with a score range from 0 to 9, where 9 represents the highest confidence level, i.e., the highest reliability and accuracy of the data; Standard_units = nM indicates that the test concentration unit is nM, which is the standard unit for activity data, representing nanomoles (10⁻⁶) per liter of solution. -9The number of molecules is (in moles); Assay_type = B indicates that the experiment type is a binding experiment (B), that is, the experimental data comes from a binding experiment, which is usually used to measure the binding affinity between a compound and a biomolecule (such as a receptor, enzyme, etc.).

[0172] After screening using the first screening criterion, 2,285,354 molecules remained out of 20,334,684, covering 6,674 biological targets (Uniprot ID) and 197,659 experimental conditions (Bio-Assay, or Assay for short). Statistical analysis was performed on the molecules corresponding to different biological targets (e.g., protein targets) and Assays, following statistical methods in relevant techniques, and the above data was analyzed according to biological targets:

[0173] (1) If the number of molecules of a biological target exceeds 20,000, it is divided into large subsets;

[0174] (2) If the number of molecules for a biological target is between 1,000 and 20,000, it is classified as a medium subset.

[0175] (3) If the number of molecules of a biological target is between 100 and 1000, it is divided into small subsets;

[0176] (4) If the number of molecules of a certain biological target is less than 100, it is divided into a small subset.

[0177] The statistical results are shown in Table 1 below:

[0178] Table 1

[0179] Molecular number threshold Biological Target (Uniprot ID) Experimental conditions (Assay) >100000 1 0 [10000，100000] 13 5 [1000，10000] 423 13 [100，1000] 1573 1528 <100 4664 196113

[0180] Table 1 shows the number of Unirot IDs and Bio-Assays covered under different molecular number thresholds after filtering by the first screening condition.

[0181] The second screening criteria include, but are not limited to: activity indicators of Ki type, Kd type, IC50 type (Standard_type = ['IC50', 'Ki', 'Kd']), desalting, selecting the IC50 value (Relation = '="), and deleting molecules with large measurement variance.

[0182] Here, Standard_type = ['IC50', 'Ki', 'Kd'] indicates that only data with activity indicators of IC50, Ki, and Kd types are selected; Delete salts indicates that some unnecessary ions are deleted to remove interfering values; Relation = '=' indicates that IC50 is selected as a certain value, which can be greater than or less than a certain value. Delete compounds with large measurement variance indicates that compounds with variances greater than or equal to a preset difference threshold are deleted to ensure that the efficacy differences of the remaining bioactivity parameters are within one order of magnitude. For specific implementation methods, please refer to the above embodiments, and repeated details will not be elaborated further.

[0183] After screening using the second screening criterion, 1,350,920 molecules remained out of 2,285,354, covering 6,061 biological targets and 140,068 experimental conditions. The molecules corresponding to different biological targets and experimental conditions were statistically analyzed, and the results are shown in Table 2 below.

[0184] Table 2

[0185] Molecular number threshold Biological Target (Uniprot ID) Experimental conditions (Assay) [10000，100000] 5 0 [1000，10000] 307 4 [100，1000] 1196 767 <100 4553 139297

[0186] Table 2 shows the number of Uniprot IDs and Bio-Assays covered under different molecular number thresholds after filtering by the second screening condition.

[0187] In the above implementation scheme, molecular data in the molecular database is preprocessed using the first and second screening conditions to remove abnormal data and obtain a molecular dataset that meets the requirements, thereby improving the reliability and consistency of the data in the molecular dataset.

[0188] After performing the data preprocessing steps as described above, this application embodiment implements the steps of constructing matching molecule pairs based on molecular fingerprints and substructure similarity in the following manner:

[0189] S22: For each molecule in the molecular dataset that can bind to the same biological target and whose bioactivity can be measured under the same experimental conditions, obtain the similarity between the molecular fingerprints of every two molecules in each molecule; and, by analyzing the similarity of the substructures of every two molecules, select matching molecule pairs that meet the preset similarity evaluation conditions.

[0190] Specifically, molecules that can bind to the same biological target and whose bioactivity can be measured under the same experimental conditions can be grouped into the same molecular data subset. Then, for each molecular data subset, the similarity between the molecular fingerprints of every two molecules is obtained through the following method, and the similarity of the substructures of every two molecules is analyzed to obtain matching molecule pairs within each molecular data subset. The process of "analyzing the similarity of the substructures of every two molecules to obtain matching molecule pairs for each molecular data subset" can be implemented based on substructure-based similarity criteria.

[0191] Molecular fingerprinting is a method that transforms molecular structures into fixed-length binary vectors, where each site represents the presence or absence of a specific structural feature in the molecule. By calculating the similarity between fingerprints (such as the Tanimoto coefficient), the degree of structural similarity between two molecules can be quantified. For the similarity criteria in activity cliff studies, this application proposes a novel set of molecular activity cliff screening and evaluation criteria, which comprehensively considers the similarity between molecular fingerprints and substructure-based similarity evaluation standards.

[0192] First, we will introduce how to obtain the similarity between molecular fingerprints of every two molecules in the same molecular data subset:

[0193] Before calculating the similarity of molecular fingerprints, the number of molecules corresponding to each assertion for each biological target was first counted, and asserts containing only one molecule were removed. Then, based on the molecules corresponding to the same assertion under the same biological target—that is, based on molecules within the same molecular subset—pairwise matching molecular pairs were constructed. Finally, the Tanimoto coefficient between molecular pairs was calculated using the molecular fingerprints.

[0194] The process of evaluating molecular pair similarity by calculating the Tanimoto coefficient of molecular fingerprints is as follows:

[0195] (1) Obtaining Molecular Fingerprint Representation: First, a molecular fingerprint is generated for each molecule using a molecular fingerprinting algorithm (such as Morgan Fingerprints, Topological Fingerprints, Path-based Fingerprints, etc.). A molecular fingerprint is a form of encoding molecular structural information into a fixed-length binary vector or numerical vector, where each bit or element represents the presence or absence of a specific substructure or chemical feature in the molecule.

[0196] (2) Calculate the Tanimoto coefficient between molecular fingerprints: Next, for each pair of molecules, calculate the Tanimoto coefficient between their molecular fingerprints. The Tanimoto coefficient measures the similarity between two molecular fingerprints by comparing their intersection (i.e., bits present in both fingerprints) and union (i.e., bits present in either fingerprint).

[0197] (3) Capturing global differences: By calculating the Tanimoto coefficient, the global structural differences between molecular pairs can be quantified, that is, the entire set of substructures contained in the molecule is considered. The value of the Tanimoto coefficient ranges from 0 to 1. The closer the value is to 1, the more similar the molecular pairs are, and the closer the value is to 0, the less similar the molecular pairs are.

[0198] In summary, by calculating the Tanimoto coefficient of molecular fingerprints, we can effectively capture the shared substructures between molecular pairs and assess their global structural similarity.

[0199] like Figure 3A As shown, it is a schematic diagram of the similarity between molecular fingerprints provided in an embodiment of this application. Figure 3A It contains two molecules, with the position 301 of the left molecule being different from that of the right molecule at position 302. The Ki of the left molecule is 101 nM, while the Ki of the right molecule is 3213 nM. This indicates that although the two molecules have similar structures, their potency differs by more than 10 times.

[0200] In addition to using molecular fingerprints to calculate the Tanimoto coefficient between molecular pairs, scaffold similarity, similarity of SMILES strings, and so on can also be used.

[0201] When using Scaffold Similarity, it is determined by calculating the molecular fingerprint on the scaffold and the corresponding Tanimoto coefficient. Scaffold Similarity allows for the identification of compound pairs with minor differences in their molecular cores, or based on differences in scaffold decoration.

[0202] like Figure 3B As shown, it is a schematic diagram of a stent similarity provided in an embodiment of this application. Figure 3B It contains two molecules. The molecule on the left has a different Ki at position 303 compared to the molecule on the right at position 304. The Ki of the molecule on the left is 3 Nm, while the Ki of the molecule on the right is 95 nM. This indicates that although the two molecules have similar structures, their potency differs by more than 10 times.

[0203] When using Similarity of SMILES Strings, the similarity of SMILES strings is captured by calculating the Levenshtein distance. This metric can detect character insertions, deletions, and shifts.

[0204] like Figure 3C As shown, it is a schematic diagram of the similarity of SMILES strings provided in an embodiment of this application. Figure 3C It contains two molecules. The molecule on the left has a different Ki at position 305 compared to the molecule on the right at position 306. The Ki of the molecule on the left is 3 nM, while the Ki of the molecule on the right is 126 nM. This indicates that although the two molecules have similar structures, their potency differs by more than 10 times.

[0205] In addition, Tanimoto coefficients can be calculated by computing the molecular access system keys (MACCS keys), daylight fingerprints, and extended-connectivity fingerprints (ECEP) for each molecule, thereby assessing the similarity of molecule pairs.

[0206] Specifically, MACCS Keys consist of 166 bits, each corresponding to a specific chemical structural feature. If a molecule contains that feature, the corresponding bit is set to 1; otherwise, it is set to 0. In this way, complex molecular structures can be transformed into a 166-bit binary vector, allowing the determination of the Tanimoto coefficients and molecular pair similarity between two molecules based on their vectors.

[0207] Daylight Fingerprints is a dynamically generated chemical fingerprinting method based on the atomic connectivity of molecules, including structural information such as rings, branches, and functional groups. It uses a path-based approach, considering the paths between all atomic pairs within the molecule. Each path is encoded into a specific bit; if a corresponding path is found in the molecule, the corresponding bit is activated. Thus, based on the chemical fingerprints of two molecules, the Tanimoto coefficients and the similarity between the molecular pairs can be determined.

[0208] The core idea of ​​ECEP is to generate a fixed-length bit vector by calculating the neighborhood information of each atom in a molecule. This bit vector reflects the structural features of the molecule. The generation of ECEP typically involves the following steps: (1) For each atom in the molecule, calculate the connectivity information of atoms within a certain radius (called the "radius" or "order") around it. (2) Convert this connectivity information into a unique hash value. (3) Map these hash values ​​to a fixed-length bit vector, typically 1024 or 2048 bits long. (4) Repeat the above process until the information of all atoms is encoded into the bit vector. Based on the bit vectors of the two molecules, the Tanimoto coefficients and the similarity between the molecules are determined.

[0209] In the above embodiments, the present application examples determine the similarity between each pair of molecules through molecular fingerprints, so as to facilitate subsequent screening of molecular pairs based on the similarity of molecular pairs through molecular activity cliffs.

[0210] Furthermore, the similarity evaluation criteria based on substructure are introduced:

[0211] Substructure-based similarity criteria can serve as an alternative to the Tanimoto coefficient for calculating molecular fingerprints. This approach often uses the logarithm of Ki (pKi) to compare molecular activity; therefore, when two molecules have a pKi difference of 1, their Ki differs by a factor of 1. In related technologies, this indicates an activity cliff relationship between the two molecules, which can be denoted as an activity cliff molecular pair. However, this method typically produces a binary result: two drug compounds either share the same predefined substructure and are classified as similar, or they do not share it and are classified as dissimilar, resulting in a limited number of activity cliff molecular pairs screened. To overcome this problem, this application comprehensively considers both molecular fingerprint similarity and substructure-based similarity assessment criteria.

[0212] Substructure-based similarity can be evaluated in different ways. The following section introduces active cliff classification methods based on different substructures:

[0213] (1) Topology Cliffs: These refer to the phenomenon in chemical space where two or more molecules, despite having very similar topological structures (i.e., the way atoms are connected and arranged), exhibit significantly different biological activities. These differences may stem from minute variations in the position, orientation, or spatial arrangement of certain key functional groups within the molecule, which can have a significant impact on the interaction between the molecule and biological targets.

[0214] like Figure 4AAs shown, it is a schematic diagram of a topological cliff provided in an embodiment of this application. Figure 4A The sample contains two molecules with identical scaffolds and substituents. However, the 401 position of the left molecule differs from the 402 position of the right molecule, indicating a different topological connection. Under this topological difference, the pKi of the left molecule is 6.27, while that of the right molecule is 8.44.

[0215] (2) Chirality Cliffs: These refer to a pair of molecules with identical chemical compositions, but due to their different spatial configurations (the molecules have two non-hyper-overlapping mirror isomers, like the left and right hands, which cannot overlap each other in three-dimensional space by rotation), their binding ability or mode of action to targets in organisms may be very different, resulting in huge differences in biological activity.

[0216] like Figure 4B As shown, it is a schematic diagram of a chiral cliff provided in an embodiment of this application. Figure 4B The sample contains two molecules with identical scaffolds and substituents, but the structure at position 403 of the left molecule differs from that at position 404 of the right molecule, meaning they have different configurations at their stereocenters. In this case, the pKi of the left molecule is 5.60, and the pKi of the right molecule is 7.89.

[0217] (3) R-group Cliffs: Substituents are functional groups attached to the main chain or ring; they can be simple atoms or complex functional groups. R-group cliffs refer to the phenomenon that in structurally similar molecules, changing just one or more R-groups can lead to significant differences in potency.

[0218] like Figure 4C As shown, it is a schematic diagram of a substituent cliff provided in an embodiment of this application. Figure 4C The sample contains two molecules with the same scaffold, but the substituent at position 405 of the left molecule is different from the substituent at position 406 of the right molecule. In this case, the pKi of the left molecule is 5.89 and the pKi of the right molecule is 8.62.

[0219] (4) Scaffold Cliffs: The phenomenon that, while keeping R-groups similar, molecules with similar core structures or scaffolds can have significantly different efficiencies by changing only a single atom or functional group on the core structure or scaffold.

[0220] like Figure 4D As shown, it is a schematic diagram of a support cliff provided in an embodiment of this application. Figure 4DThe sample contains two molecules with the same substituents, but the molecular structure at position 407 of the left molecule is different from that at position 408 of the right molecule. In this case, the pKi of the left molecule is 6.89 and the pKi of the right molecule is 9.22.

[0221] like Figure 4E As shown, it is a schematic diagram of a support / topological cliff provided in an embodiment of this application. Figure 4D The sample contains two molecules with the same substituents, but the substituents at position 409 on the left molecule are different from those at position 410 on the right molecule. In this case, the pKi of the left molecule is 6.99 and the pKi of the right molecule is 9.05.

[0222] (5) Matched Molecular Pair Cliffs (MMP-Cliffs): MMPs refer to a pair of molecules that differ only in a single site of chemical modification. By limiting the size of the chemical modification, it can be ensured that the compounds that form MMPs are mostly structurally similar. Applying this similarity criterion, MMP-based activity cliffs are generated, which are called "MMP-Cliffs".

[0223] like Figure 5A As shown, this is a schematic diagram of a matching molecule pair cliff provided in an embodiment of this application. Figure 5A It contains two molecules. After these two molecules are randomly broken, the lower molecule has an additional substructure at position 501 compared to the upper molecule. In this case, the pKi of the upper molecule is 5.60 and the pKi of the lower molecule is 9.00.

[0224] like Figure 5B As shown, this is a schematic diagram of another matching molecule pair cliff provided in an embodiment of this application. Figure 5B It contains two molecules. After the nicotinamide is cleaved under the RECAP rule, the substructure at position 502 of the upper molecule is different from the substructure at position 503 of the lower molecule. In this case, the pKi of the upper molecule is 5.60 and the pKi of the lower molecule is 9.00.

[0225] The steps for constructing matching molecule pairs in the embodiments of this application are described below, taking MMPs as an example.

[0226] When constructing matching molecule pairs based on the substructure similarity of each molecule in this application embodiment, one optional implementation method is as follows:

[0227] Each molecule was cleaved using acyclic single bonds, resulting in two parts: an invariant core substructure and a region molecular structure. The invariant core substructure is the common part between molecules, while the region molecular structure is unique to each molecule. Every two molecules with the same invariant core substructure were grouped into a candidate molecule pair. The candidate molecule pairs were then screened using preset similarity assessment conditions to obtain matching molecule pairs. The similarity assessment conditions were used to evaluate the atomic differences and differences in bioactivity parameters between the region molecular structures of the candidate molecule pairs.

[0228] Among these, the constraints can be restrictions on the complexity or efficacy differences of candidate molecule pairs. For example, the complexity can be restricted by the number of differences in the number of non-hydrogen atoms in the molecular structure of the restricted region, and the efficacy difference can be restricted by the order of magnitude of the difference in activity between molecules, and so on.

[0229] Specifically, in the substructure similarity calculation, the number of molecules corresponding to each Assay in each biological target was first counted, and Assays with only one molecule were removed; then, for each molecular data subset, MMPs were constructed according to the fragment indexing algorithm.

[0230] The specific process of constructing MMPs is as follows:

[0231] First, perform all possible acyclic single-bond cleavage operations on each molecule corresponding to the same Assay for the same biological target. Second, index the cleaved parts and label them as keys and values, where the key represents the common part, i.e., the invariant core substructure, and the value represents the unique part, i.e., the molecular structure. Finally, identify molecular pairs with the same bonds through key-value pairs.

[0232] Furthermore, in identifying molecular pairs with the same bonds, the embodiments of this application set the following restrictions on MMPs: (1) the difference in the number of non-hydrogen atoms between the exchanged molecular structures (i.e., values) shall not exceed 8; (2) the number of non-hydrogen atoms in the exchanged molecular structures shall not exceed 13; and (3) the difference in activity between molecules shall be at least one order of magnitude.

[0233] After further filtering by the above constraints, the remaining molecular pairs are the matching molecular pairs. That is, among molecular pairs with the same bonds, the molecular pairs that satisfy the above constraints are the matching molecular pairs.

[0234] like Figure 6A The diagram shown illustrates a matching molecule pair provided in an embodiment of this application. Figure 1 . Figure 6AIt contains a pair of matching molecules, in which the a1, b1, a2, c1 positions of the upper molecule are different from the a3, b2, a4, b3, a5, c2 positions of the lower molecule.

[0235] like Figure 6B The diagram shown illustrates a matching molecule pair provided in an embodiment of this application. Figure 2 . Figure 6A It contains a pair of matching molecules, in which the d1, e1, d2, c1 positions of the upper molecule are different from the d3, e2, d4, e3, d5 positions of the lower molecule.

[0236] like Figure 6C As shown, this is a schematic diagram of a matching molecule pair provided in an embodiment of this application. Figure 6A It contains a pair of matching molecules, in which the h1, f1, f2, g1 positions of the upper molecule are different from the f3, g2, f4, g3, f5, h2 positions of the lower molecule.

[0237] like Figure 6D As shown, this is a schematic diagram of a matching molecule pair provided in an embodiment of this application. Figure 6A It contains a pair of matching molecules, in which the k1, i1, i2, j1 positions of the upper molecule are different from the j4, j2, i5, j3, i6, k2 positions of the lower molecule.

[0238] like Figure 6E As shown, this is a schematic diagram of a matching molecule pair provided in an embodiment of this application. Figure 6A It contains a pair of matching molecules, in which the n1, l1, m1, l2, n2 positions of the upper molecule are different from the l3, m2, l4, m3, l5, n3 positions of the lower molecule.

[0239] like Figure 6F As shown, this is a schematic diagram of a matching molecule pair provided in an embodiment of this application. Figure 6A It contains a pair of matching molecules, in which the o1, p1, o2, q1 positions of the upper molecule are different from the p2, o3, o4, q2 positions of the lower molecule.

[0240] After obtaining the similarity between each pair of molecules determined by molecular fingerprints using the above methods, and constructing matching molecule pairs based on substructure similarity, considering that there may be molecule pairs with poor quality or that do not meet the requirements, this application embodiment filters these matching molecule pairs in the following way:

[0241] (1) Eliminate matched molecular pairs with inconsistent experimental conditions (such as Ki, Kd, ​​IC50) to ensure that the comparison of matched molecular pairs is carried out under the same experimental conditions. This is because different experimental conditions (such as Ki and IC50) cannot be directly compared numerically. For example, for a matched molecular pair containing molecules A and B, if the experimental condition for molecule A is Ki, while the experimental condition for molecule B is IC50, then the matched molecular pair should be eliminated.

[0242] (2) Eliminate matching molecular pairs where the first molecular entity (molecular1) and the second molecular entity (molecular2) are the same. If molecular1 and molecular2 are the same, it means that the first molecule and the second molecule are completely identical, that is, they are the same molecule.

[0243] (3) Eliminate any molecule pair in which either the first bioactivity parameter (standard value1) or the second bioactivity parameter (standard value2) is missing (NAN). Standard value1 is the bioactivity parameter of molecular 1, and standard value2 corresponds to the bioactivity parameter of molecular 2. If any molecule has an invalid bioactivity parameter, then that molecule cannot be used to analyze molecular activity.

[0244] (4) Remove duplicate matching pairs.

[0245] Table 3 below serves as an example. Table 3 shows the number of matching molecule pairs constructed based on molecular fingerprint similarity and matching molecule pairs constructed based on substructure similarity before and after the above filtering process:

[0246] Table 3

[0247] Construction method Number of matched molecule pairs (before filtering) Number of matched molecule pairs (after filtering) Based on molecular fingerprint 39,902,900 33,397,003 Based on substructure 18,911,399 6,058,106

[0248] like Figure 7 As shown, it is a schematic diagram of the number of matching molecule pairs provided in an embodiment of this application. Figure 7 This is a Venn diagram showing the number of matched molecules after filtering in Table 3 above. It presents common and unique molecular pairs in the two datasets, calculated based on molecular fingerprint similarity and substructure similarity, among the molecular pairs calculated using two similarity criteria. Figure 7 It can be clearly seen that the molecular pairs obtained based on molecular fingerprints, namely Figure 7 The middle 701 (i.e., the large circle) covers all matching molecule pairs obtained based on substructure construction, i.e. Figure 7The middle 702 (i.e., the small circle). Therefore, the number of molecular pairs constructed based on molecular fingerprints far exceeds the number of matching molecular pairs constructed based on substructures. This indicates that the molecular fingerprint similarity criterion has a wider coverage and can identify more matching molecular pairs. However, the MMPs method is more stringent, identifying a relatively smaller number of matching molecular pairs, but these matching molecular pairs have more similar chemical structures.

[0249] therefore, Figure 7 The effectiveness of the novel molecular activity cliff screening and evaluation criteria proposed in the embodiments of this application has been verified. That is, by combining two similarity evaluation methods, activity cliffs with chemical similarity and biological activity differences can be accurately identified in a wider range of molecular pairs.

[0250] In the above embodiments, the present application's embodiments determine the similarity between every two molecules based on molecular fingerprints and construct matching molecule pairs based on substructure similarity, which comprehensively considers the advantages of the two methods and makes up for the advantages of the two approaches. This is beneficial for subsequently determining the similarity threshold of molecular fingerprints based on substructure similarity.

[0251] After determining the similarity between each pair of molecules based on molecular fingerprints using the above method, and constructing matching molecule pairs for each molecular data subset based on substructure similarity, the embodiments of this application can calculate the molecular activity cliff threshold for each biological target in the following manner:

[0252] S23: For each biological target, based on the distribution of each matching molecule pair under different similarities, select a similarity as the molecular activity cliff threshold corresponding to the biological target, and select the target molecule pair from each molecule based on the molecular activity cliff threshold.

[0253] Among them, the activity cliff refers to the phenomenon that molecular pairs with highly similar structures have significantly different biological activities.

[0254] Specifically, to combine the advantages of the two methods for constructing matching molecule pairs, this application embodiment determines the similarity threshold between the molecular fingerprints of every two molecules using MMPs. For each biological target, the Tanimoto coefficient calculated based on the molecular fingerprint is calculated, which covers all or part of the MMPs corresponding to the biological target. This Tanimoto coefficient is used as the molecular activity cliff threshold. Thus, molecule pairs whose similarity meets the molecular activity cliff threshold can be screened from every two molecules whose similarity is determined based on the molecular fingerprint.

[0255] Therefore, in the embodiments of this application, when determining the molecular activity cliff, an optional implementation method is as follows:

[0256] The number of matching molecule pairs at each similarity level is counted; the number of matching molecule pairs at each similarity level is accumulated sequentially in descending order of similarity; the similarity level determined when the proportion of the accumulated number of matching molecule pairs reaches the target ratio is used as the molecular activity cliff threshold corresponding to the biological target.

[0257] The percentage represents the proportion of the total number of matched molecule pairs at the biological target. The target percentage is an empirical value, but the target can also be set according to individual needs, such as 80%, 90%, 85%, 95%, 100%, etc.

[0258] Specifically, for each biological target, all its matching molecule pairs are statistically analyzed according to their similarity to obtain a histogram with similarity on the horizontal axis and the number of matching molecule pairs on the vertical axis. On this histogram, the number of matching molecule pairs is accumulated from right to left until the proportion of the accumulated number of matching molecule pairs in the total number of matching molecule pairs under the biological target reaches the target proportion. At this point, the horizontal axis value that reaches the target proportion is used as the molecular activity cliff threshold.

[0259] like Figure 8A As shown, it is a histogram for determining the molecular activity cliff threshold provided in an embodiment of this application. Figure 8A The middle section shows the histogram for the biological target with ID Q00796. The horizontal axis of this histogram represents similarity, and the vertical axis represents the number of matching molecule pairs. If the target ratio is set to 90%, then... Figure 8A The number of matched molecule pairs to the right of the dashed line accounts for 90% of the total number of matched molecule pairs. The x-coordinate of the intersection of the dashed line and the horizontal axis is 0.56, which means the molecular activity cliff threshold is 0.56.

[0260] like Figure 8B As shown, it is another histogram for determining the molecular activity cliff threshold provided in the embodiments of this application. Figure 8B The middle section shows a histogram for the biological target with ID P21266. The horizontal axis of this histogram represents similarity, and the vertical axis represents the number of matching molecule pairs. If the target ratio is also set to 90%, then... Figure 8B The number of matched molecule pairs to the right of the dashed line accounts for 90% of the total number of matched molecule pairs. The x-coordinate of the intersection of the dashed line and the horizontal axis is 0.58, which means the molecular activity cliff threshold is 0.58.

[0261] Clearly, different biological targets have different molecular activity cliff thresholds when the same target ratio is set.

[0262] Furthermore, in this embodiment, the number of matching molecule pairs at each similarity level can be accumulated sequentially in ascending order of similarity. The similarity level determined when the proportion of accumulated matching molecule pairs reaches the anti-target ratio is used as the molecular activity cliff threshold corresponding to the biological target. At this time, the sum of the anti-target ratio and the target ratio mentioned above is one.

[0263] For example, the anti-target ratio can be set to 10%, 15%, 20%, etc.

[0264] like Figure 8C As shown, it is another histogram for determining the molecular activity cliff threshold provided in the embodiments of this application. Figure 8B The middle section shows a histogram for the biological target with ID P19217. The horizontal axis of this histogram represents similarity, and the vertical axis represents the number of matching molecule pairs. If the target ratio is also set to 10%, then... Figure 8B The number of matched molecule pairs to the left of the dashed line accounts for 10% of the total number of matched molecule pairs. The x-coordinate of the intersection of the dashed line and the horizontal axis is 0.53, which means the molecular activity cliff threshold is 0.53.

[0265] In the above embodiments, the molecular activity cliff threshold is set by matching the distribution of molecular pairs under different similarities, and different biological targets correspond to different molecular activity cliff thresholds. This effectively avoids the subjective method of setting a uniform threshold in related technologies. Since it combines the fingerprint similarity criteria between molecules, it also overcomes the limitation of related technologies that only focus on changes at a single site in the substructure similarity evaluation criteria.

[0266] After determining the molecular activity cliff threshold using the above method, an optional implementation method for determining the target molecule pair in this application embodiment is as follows:

[0267] For any two molecules in each molecule, if the similarity between the two molecules is greater than the molecular activity cliff threshold, then the two molecules are considered as a target molecule pair.

[0268] Specifically, for each molecule under the same biological indicator at each biological target, based on the molecular activity cliff threshold corresponding to that biological target, molecules with a similarity greater than the molecular activity cliff threshold are identified as target molecule pairs.

[0269] Continue Figure 8A The assumption, Figure 8A If the molecular activity cliff threshold is 0.56, then molecules with a similarity greater than 0.56 between any two molecules at biological target Q00796 are considered as target molecule pairs.

[0270] Figure 8B The assumption, Figure 8BIf the molecular activity cliff threshold is 0.58, then molecules with a similarity greater than 0.58 between any two molecules at biological target P21266 are considered as target molecule pairs.

[0271] Figure 8C The assumption, Figure 8C If the molecular activity cliff threshold is 0.53, then molecules with a similarity greater than 0.53 between any two molecules at biological target P19217 are considered as target molecule pairs.

[0272] The following examples illustrate the results of different methods for calculating intermolecular similarity under different molecular activity cliff thresholds for target molecule pairs.

[0273] like Figure 9 As shown, it is a histogram of different molecular activity cliff thresholds for a different method of calculating intermolecular similarity provided in the embodiments of this application. Figure 9 The histogram in the image shows the number of target molecule pairs on the vertical axis and the molecular activity cliff threshold on the horizontal axis. It includes three similarity calculation methods: the Tanimoto coefficient between molecule pairs calculated based on molecular fingerprints, scaffold similarity, and the similarity of the SMILES string. Figure 9 It is evident that when the molecular activity cliff threshold is different, the number of target molecule pairs varies depending on the method used to calculate similarity.

[0274] Considering that the same molecular activity cliff threshold is used for different biological targets in related technologies, while different molecular activity cliff thresholds may exist for different biological targets if the same target ratio is used in the embodiments of this application, the embodiments of this application can more accurately identify the activity cliff for different biological targets.

[0275] like Figure 10 As shown, it is a histogram of similarity distribution under different biological targets provided in an embodiment of this application. Figure 10 The distribution of Tanimoto coefficients for eight biological targets is presented, with each target representing 80% of the total target population. Figure 10 It can be seen that the distribution of Tanimoto coefficients is not the same for different biological targets, which verifies the differences in the molecular activity cliff threshold of different biological targets.

[0276] like Figure 11 As shown, it is a scatter plot of target molecule pairs under different target ratios provided in an embodiment of this application. Figure 11 For a given biological target, the number of target molecule pairs obtained under different target ratios, i.e., under different molecular activity cliff thresholds, is presented.

[0277] In the above embodiments, this application's examples determine different molecular activity cliff thresholds for different biological targets by integrating molecular fingerprint-based similarity criteria and substructure similarity-based similarity criteria. This method for generating molecular activity cliff thresholds not only avoids the subjective threshold setting and chemical interpretation difficulties inherent in molecular fingerprint-based similarity criteria but also overcomes the limitations of substructure similarity-based evaluation standards that only focus on changes at a single site. By integrating these two similarity criteria, activity cliffs can be identified more accurately, enabling effective computational analysis in large compound datasets, compensating for deficiencies in related technologies, and avoiding misjudgments caused by imperfections in evaluation methods.

[0278] After calculating the molecular activity cliff threshold for each biological target using the above method, the activity cliff dataset and data post-processing can be constructed as follows:

[0279] S24: Based on the differences in the corresponding bioactivity parameters of each selected target molecule pair, analyze whether there is an activity cliff relationship between the target molecule pairs, and construct an activity cliff dataset based on the analysis results.

[0280] In this embodiment of the application, in order to improve the accuracy and generalization ability of the live cliff prediction, a live cliff prediction model is constructed. Therefore, in order to train the model, it is also necessary to construct a live cliff dataset (ACDataBase) containing positive and negative samples.

[0281] One possible implementation for constructing the live cliff dataset is as follows:

[0282] For each target molecule pair, if the difference in bioactivity parameters between the molecules is greater than a preset difference threshold, then the molecule pair is determined to have an activity cliff relationship and is designated as a positive sample; if the difference in bioactivity parameters between the molecules is not greater than the preset difference threshold, then the molecule pair is determined to have no activity cliff relationship and is designated as a negative sample; based on the positive and negative samples, an activity cliff dataset is constructed to train the activity cliff prediction model.

[0283] The activity cliff prediction model is used to predict whether an activity cliff relationship exists between input molecular pairs. The preset difference threshold can be the power difference factor between the two molecules in the molecular pair, which is generally 10 times (it can also be set to 100 times or other factors, but this paper does not make a specific limit on this).

[0284] Specifically, for molecular pairs with a power difference greater than a preset difference threshold, there is an activity cliff relationship between the two molecules in the molecular pair, which is regarded as a positive sample, namely a cliff pair (Cliffs-Pairs). For molecular pairs with a power difference less than or equal to the preset difference threshold, there is no activity cliff relationship between the two molecules in the molecular pair, which is regarded as a negative sample, namely a non-cliff pair (Non-AC Pairs).

[0285] For example, if the preset difference threshold is a power difference of 10 times, then for two target molecule pairs, in target molecule pair 1, one molecule has a Ki of 3 nM and the other molecule has a Ki of 95 nM, then the power difference between the two molecules in the target molecule pair is greater than 10 times, and target molecule pair 1 is a positive sample. In target molecule pair 2, one molecule has a Ki of 3 nM and the other molecule has a Ki of 23 nM, then the power difference between the two molecules in target molecule pair 2 is less than 10 times, and target molecule pair 2 is a negative sample.

[0286] After constructing the active cliff dataset, in order to make the active cliff dataset better suited for training the active cliff prediction model, this application embodiment further performs post-processing on the active cliff dataset in the following manner.

[0287] For each biological target, perform at least one of the following post-processing steps:

[0288] If the number of target molecule pairs selected under a biological target is less than the first preset number, then each target molecule pair under the biological target is deleted; if the number of positive samples corresponding to a biological target is less than the second preset number, then each target molecule pair under the biological target is deleted.

[0289] The second preset quantity is less than the first preset quantity.

[0290] Specifically, considering the problem that if the number of target molecule pairs for each biological target is too small, the training effect will be poor when using the target molecule pairs for the activity cliff prediction model. Therefore, this application embodiment sets a first preset number, which is an empirical value, generally 100, 120, 140, etc., and needs to comprehensively consider the training effect and the distribution of the number of target molecules for each biological target. In addition, this application embodiment also limits the number of positive samples for each target by a second preset number to avoid all samples being negative samples when training the model. This second preset number can be 10, 15, 20, etc., and needs to comprehensively consider the training effect and the distribution of the number of positive samples for each biological target.

[0291] For example, if the first preset quantity is 100 and the second preset quantity is 10, then for biological target 1 and biological target 2, since the number of target molecule pairs under biological target 1 is 58, which is less than the first preset quantity, the 58 target molecule pairs under biological target 1 are deleted. The number of target molecule pairs under biological target 2 is 120, which is greater than the first preset quantity, but the number of positive samples under biological target 2 is 3, which is less than the second preset quantity, therefore the 120 target molecule pairs under biological target 2 are deleted.

[0292] In the above implementation, different molecular activity cliff thresholds are generated for different biological targets. Positive and negative samples are extracted from matching molecular pairs obtained based on the similarity between molecular fingerprints and substructure similarity. Finally, an activity cliff dataset is constructed so that an activity cliff prediction model can be trained based on the activity cliff dataset to predict the activity cliff.

[0293] In summary, the following is a combination of... Figure 12 A brief description of the construction process of the active cliff dataset in the embodiments of this application:

[0294] like Figure 12 As shown, this is a schematic diagram of a method for constructing an active cliff dataset provided in an embodiment of this application. For example, the open-source database in 12 is CHEMBL version 33. The first screening criteria are Relationship = D, Confidence_score = 9, Standard_units = nM, and Assay_type = B. For details, please refer to the above embodiment, which will not be repeated here. After screening using the first screening criteria, 2,285,354 molecules remain out of 20,334,684.

[0295] The second screening criteria are Standard_type = ['IC50', 'Ki', 'Kd'], Delete salts, Relation = '=', and Delete compounds with large measurement variance. For details, please refer to the above examples, which will not be repeated here. After screening using the second screening criteria, 1,350,920 molecules remain out of 2,285,354 molecules.

[0296] For multiple biological targets, such as Figure 12In the examples Q00796, P21266, and P19217, similarity calculations and molecular pair identification are performed, respectively. Similarity can be calculated using molecular fingerprint similarity, scaffold similarity, and SMILES string similarity. For molecular pair identification, the number of heavy atoms must be less than 100, and the substituent size is limited to less than 10 heavy atoms and less than half the size of the molecular core. Furthermore, based on 80% MMP coverage, a molecular activity cliff threshold is calculated for each biological target. Target molecular pairs are obtained based on the activity cliff threshold, and then, according to a preset difference threshold of 10, the target molecular pairs are divided into cliff molecular pairs and non-cliff molecular pairs. See the above embodiments for details; repeated points will not be elaborated upon.

[0297] The above process, by combining two similarity criteria—molecular fingerprint similarity and substructure similarity—can more accurately identify activity cliffs and perform effective computational analysis on large compound datasets.

[0298] Table 4 below shows the number of UniprotIDs, experimental conditions, and target molecule pairs in the ACDataBase obtained based on different MMPs coverage thresholds, i.e., target proportions (e.g., 90%, 80%, 70%).

[0299] Table 4

[0300]

[0301] Table 4 shows the activity cliff datasets composed of target molecule pairs that meet different target ratios. It can be clearly seen that the higher the target ratio is set, the larger the number of target molecule pairs in the activity cliff dataset, and the more biological targets and experimental conditions are corresponding to it.

[0302] Furthermore, to test ACDataBase's performance in predicting molecular activity cliffs, it is necessary to select a dataset of similar type from existing publicly available activity cliff datasets. Currently, publicly available and large-scale activity cliff datasets include MoleculeACE (Molecule Enhanced Activity Cliff) and the neural network-based activity cliff dataset (ACNet). However, MoleculeACE does not use molecular pairs as input to the model; instead, it distributes highly similar molecules with significant activity differences between the training and test sets during the construction of the training and test sets. Therefore, ACNet is selected for comparison with ACDataBase constructed in this application. Referring to the statistical methods used for ACNet in related technologies, ACDataBase is statistically analyzed according to target points:

[0303] (1) If the number of target molecule pairs in a dataset of a biological target exceeds 20,000, it is divided into large subsets.

[0304] (2) If the number of target molecule pairs in a dataset of a certain biological target is between 1,000 and 20,000, it is divided into medium subsets;

[0305] (3) If the number of target molecule pairs in a dataset of a biological target is between 100 and 1000, it is divided into small subsets.

[0306] (4) If the number of target molecule pairs in a dataset of a biological target is less than 100, it is divided into a small subset.

[0307] In this way, the dataset can be organized into different subsets according to its size, thereby better evaluating and comparing the model's performance on tasks of different sizes. Table 5 below shows the detailed information of ACDataBase obtained at different MMPs coverage thresholds (90%, 80%, 70%).

[0308] Table 5

[0309]

[0310] In the above embodiments, this application provides a novel set of similarity assessment criteria. By combining substructure-based similarity assessment standards and molecular fingerprint-based similarity criteria, it avoids subjective threshold settings and single threshold settings for different targets, enabling more accurate identification of activity cliffs. Furthermore, artificial intelligence algorithms for predicting activity cliffs can be developed based on data, capturing minute chemical changes that have a significant impact on specific biological activities and improving the predictive ability of activity cliffs. This will ultimately promote the effective optimization from active molecules to lead compounds, as well as the identification of activity cliffs during lead compound optimization. This is of great significance for advancing the potential of AI in the drug optimization stage.

[0311] The beneficial effects of the method for constructing the active cliff dataset provided in this application will be demonstrated below through specific embodiments.

[0312] 1. Dataset partitioning strategy.

[0313] To comprehensively test the performance of ACDataBase in predicting molecular activity cliffs, this application's embodiments designed two data partitioning scenarios.

[0314] Scenario 1: Drug Out Of Distribution (DrugOOD) extracts common and unique biological targets from ACDataBase and ACNet to ensure that molecules in the training and testing sets act on the same biological targets, in order to test the model's generalization ability under known biological targets.

[0315] Scenario 2: Target Out Of Distribution (TargetOOD) – This scenario extracts unique biological targets from ACDatabase and ACNet and selects the top ten biological targets by sample size (i.e., the number of target molecule pairs). This ensures that molecules in the training and testing sets act on different biological targets, testing the model's predictive ability under unknown biological targets. Both scenarios are designed to comprehensively evaluate the model's performance under different dataset construction schemes, ensuring the model has broader application potential.

[0316] Scenario 1 (DrugOOD): Common biological targets are extracted from the ACDataBase and ACNet datasets. 20% of molecules from each common biological target are selected and placed into the test dataset, while the remaining molecules are placed into the training dataset. Additionally, molecules corresponding to the unique biological targets in each dataset are also placed into the training dataset. To avoid data leakage, if a molecule in the test set appears in the training set, it is removed from the test set. In this scenario, molecules in both the training and test sets act on the same biological targets, but molecules in the test set do not appear in the training set. This allows testing whether the activity cliff prediction model can expand its molecular activity prediction capability for known biological targets by using a larger amount of biological target and molecule data in the activity cliff dataset constructed in this application embodiment. The information for the training and test sets under the data partitioning strategy in Scenario 1 is shown in Table 6 below.

[0317] Table 6

[0318]

[0319] Scenario 2 (TargetOOD): Unique biological targets (UniprotIDs) are extracted from both the ACDataBase and ACNet datasets. The top ten biological targets by sample size are selected, and their corresponding molecules are extracted to generate the test dataset. The remaining data is used to generate the training dataset. To avoid data leakage, if a molecule in the test set appears in the training set, it is removed from the test set. In this scenario, the molecules in the training and test sets act on different targets, allowing the model to test whether it can learn more prior knowledge from the activity cliff dataset constructed in this embodiment, thus expanding its predictive ability for unknown targets. Under the data partitioning strategy in Scenario 2, the information for the training and test sets is shown in Table 7 below.

[0320] Table 7

[0321]

[0322] 2. Active cliff prediction model

[0323] To improve the accuracy and generalization ability of activity cliff prediction, this application utilizes a deep learning model as an encoder to extract molecular representations. Unlike traditional feature engineering preprocessing methods, it directly encodes the complete molecule, avoiding information loss caused by splitting substructures. Specifically, as... Figure 13 The diagram shown is a structural schematic of an activity cliff prediction model provided in an embodiment of this application. This embodiment uses a deep learning molecular representation encoder to extract molecular representations of the two compounds in a molecular pair; these representations are concatenated and used as input to a multilayer perceptron prediction head; the multilayer perceptron prediction head predicts the activity cliff relationship between the two compounds based on the concatenated representation.

[0324] 3. Prediction Results

[0325] The model was trained on training sets obtained under different thresholds for both Scenario 1 and Scenario 2, and then tested on the test set. The prediction results on the test set are shown in Tables 8 and 9 below. Evaluation metrics included accuracy, area under the curve of receiver operating characteristic (AUC_ROC), precision, recall, and F1 score.

[0326] Table 8

[0327]

[0328] Table 9

[0329]

[0330] Table 8 shows the performance of the ACDataBase and ACNet datasets in predicting the liveness cliff under different target ratios in Scenario 1 (DrugOOD). The liveness cliff prediction model trained with ACDataBase significantly outperforms ACNet on the test set across all target ratios, especially in Accuracy and AUC_ROC, demonstrating stronger predictive power and reliability. The liveness cliff prediction model trained with ACDataBase exhibits better generalization ability, as evidenced by its stable performance on the test set across different thresholds, with high Accuracy and AUC_ROC. In contrast, the liveness cliff prediction model trained with ACNet has poor generalization ability, resulting in low prediction accuracy across various evaluation metrics on the test set, particularly Recall and F1Score. This is especially important for evaluating class imbalance. As the target ratio decreases, the metrics of ACDataBase (especially Accuracy and F1Score) decline slightly, but the overall change is not significant, indicating relatively stable performance. ACNet's Recall and F1Score values ​​are low, indicating poor overall performance. This suggests that the live cliff prediction model trained with ACDataBase has stronger stability and predictive ability across different target proportions. Similarly, as... Figure 14 As shown, it is a histogram of the performance of two datasets, ACDataBase and ACNet, in the prediction of the liveness cliff task in a scenario provided by an embodiment of this application. Figure 14 This is the histogram statistical form shown in Table 8.

[0331] Table 9 shows the performance of the liveness cliff prediction models trained on the ACDataBase and ACNet datasets in the liveness cliff prediction task under different target ratios in Scenario 2 (TargetOOD). ACDataBase significantly outperforms ACNet on all evaluation metrics, demonstrating stronger generalization ability and stability. ACDataBase's performance is relatively stable across all target ratios, but its AUC_ROC and precision are low. Although ACNet has higher accuracy, its AUC_ROC and recall are very low, especially the recall, which is almost zero, indicating that the liveness cliff prediction model can hardly identify the correct samples, and its generalization ability is even worse. Similarly, as... Figure 15As shown, it is a histogram of the performance of two datasets, ACDataBase and ACNet, in the prediction of the liveness cliff task under a scenario two provided in an embodiment of this application. Figure 15 This is the histogram statistical form shown in Table 8.

[0332] In both scenarios, the liveness cliff prediction model trained with ACDataBase performed exceptionally well, particularly in terms of accuracy and AUC_ROC, demonstrating stronger predictive and generalization capabilities. Its performance remained stable across different target ratios, especially in cases of class imbalance. ACNet underperformed ACDataBase in both scenarios, particularly in recall and F1Score, indicating insufficient generalization ability. While its accuracy was higher at certain thresholds, its AUC_ROC and recall metrics showed poor ability to identify correct samples.

[0333] This application provides a method, apparatus, electronic device, and storage medium for constructing an activity cliff dataset. This application also establishes a set of criteria for screening and evaluating molecular activity cliffs. Specifically, the pre-constructed molecular dataset includes the biological targets that each molecule can bind to and the bioactivity parameters of each molecule. Based on this, molecules that can bind to the same biological target and whose bioactivity can be measured under the same experimental conditions are grouped into the same molecular data subset, with each subset calculating the similarity between molecules and dividing them into molecular pairs. Specifically, molecular fingerprinting is used to obtain the similarity between every two molecules in each molecular data subset. By analyzing the similarity of every two molecules in their substructures, matching molecule pairs that meet the preset similarity evaluation conditions are selected. In this step, this application comprehensively considers the fingerprint similarity criteria between molecules and the substructure-based similarity evaluation criteria. The fingerprint similarity criteria between molecules simplifies complex molecular structures into binary or numerical vectors, enabling efficient calculation of molecular similarity. The substructure-based similarity evaluation criteria, on the other hand, focuses on molecular functional regions, which can accurately identify and match key pharmacophores or functional units in chemical structures, thereby improving the accuracy of active compound discovery.

[0334] Based on the above, for each biological target, a molecular activity cliff threshold is set by matching the distribution of molecular pairs under different similarities. This achieves different thresholds for different biological targets, effectively avoiding the subjective method of setting a uniform threshold in related technologies. Furthermore, by combining the fingerprint similarity criteria between molecules, it overcomes the limitation of related technologies that only focus on changes at a single site in substructure similarity evaluation standards.

[0335] Finally, for each molecular pair under each biological target, target molecular pairs are selected based on the molecular activity cliff threshold. On this basis, an activity cliff dataset is constructed, which effectively improves the accuracy and consistency of activity cliff identification.

[0336] In summary, this application sets a molecular activity cliff threshold by matching the distribution of molecular pairs under different similarities, avoiding the subjective and artificial threshold setting methods in related technologies, and also overcoming the limitation of related technologies that only focus on changes at a single site in substructure similarity evaluation standards.

[0337] Based on the same inventive concept, embodiments of this application also provide an apparatus for constructing an active cliff dataset. For example... Figure 16 As shown, this is a schematic diagram of the structure of the active cliff dataset construction device 1600, which may include:

[0338] The first acquisition unit 1601 is used to acquire a pre-constructed molecular dataset; the molecular dataset contains multiple molecules, at least one biological target that each molecule can bind to, and bioactivity parameters of each molecule, the bioactivity parameters being measured under at least one experimental condition;

[0339] The second acquisition unit 1602 is used to acquire the similarity between molecular fingerprints of each pair of molecules in a molecular dataset that can bind to the same biological target and whose bioactivity is measured under the same experimental conditions; and to select matching molecule pairs that meet preset similarity evaluation criteria by analyzing the similarity of each pair of molecules in their substructures.

[0340] The threshold determination unit 1603 is used to select a similarity as the molecular activity cliff threshold corresponding to the biological target for each biological target based on the distribution of each matching molecule pair under different similarities, and select the target molecule pair from each molecule based on the molecular activity cliff threshold; the activity cliff refers to the phenomenon that molecule pairs with highly similar structures have significant differences in biological activity.

[0341] Unit 1604 is used to analyze whether there is an activity cliff relationship between the selected target molecule pairs based on the differences in the corresponding bioactivity parameters, and to construct an activity cliff dataset based on the analysis results.

[0342] Optionally, the threshold determination unit 1603 is specifically used for:

[0343] Count the number of matching pairs at each similarity level;

[0344] The number of matched pairs at each similarity level is accumulated in descending order of similarity.

[0345] The similarity determined when the proportion of accumulated matched molecule pairs reaches the target proportion is used as the molecular activity cliff threshold corresponding to the biological target; the proportion of accumulated matched molecule pairs represents the ratio of the number of accumulated matched molecule pairs to the total number of matched molecule pairs under the biological target.

[0346] Optionally, the threshold determination unit 1603 is specifically used for:

[0347] For any two molecules in each molecule, if the similarity between the two molecules is greater than the molecular activity cliff threshold, then the two molecules are considered as a target molecule pair.

[0348] Optionally, building block 1604 is specifically used for:

[0349] For each target molecule pair, if the difference in bioactivity parameters between the molecules is greater than a preset difference threshold, then it is determined that there is an activity cliff relationship between the molecules of the target molecule pair, and the target molecule pair is regarded as a positive sample.

[0350] If the difference in bioactivity parameters between molecules is not greater than a preset difference threshold, then it is determined that there is no activity cliff relationship between the molecules of the target molecule pair, and the target molecule pair is used as a negative sample.

[0351] Based on positive and negative samples, an active cliff dataset is constructed to train the active cliff prediction model. The active cliff prediction model is used to predict whether an active cliff relationship exists between input molecular pairs.

[0352] Optionally, the device also includes:

[0353] Post-processing unit 1605 is used to perform at least one of the following post-processing steps on each biological target:

[0354] If the number of target molecule pairs selected under the biological target is less than the first preset number, then each target molecule pair under the biological target will be deleted.

[0355] If the number of positive samples corresponding to a biological target is less than the second preset number, then delete each target molecule pair under the biological target.

[0356] The second preset quantity is less than the first preset quantity.

[0357] Optionally, the first acquisition unit 1601 is specifically used for:

[0358] Obtain known molecular datasets; molecular datasets are open-source datasets used for molecular biological activity research;

[0359] Based on the first and second screening criteria, molecules and related data that meet the criteria are selected from the molecular dataset and then combined to form a molecular dataset.

[0360] The first screening criterion is used to limit the data quality and experimental type of the required molecules; the second screening criterion is used to limit the activity index and purity of the required molecules; the relevant data includes the biological targets that each molecule can bind to, and the biological activity parameters of each molecule; the activity index represents the data type of the biological activity parameters.

[0361] Optionally, before assembling the molecular dataset, the first acquisition unit 1601 is also used for:

[0362] For a single molecule targeting a biological site:

[0363] If a molecule has multiple bioactivity parameters under the same activity index, and the variance of the multiple bioactivity parameters is greater than a preset difference threshold, then the multiple bioactivity parameters are deleted.

[0364] If a molecule has multiple bioactivity parameters under the same activity index, and the variance of the multiple bioactivity parameters is not greater than a preset difference threshold, then the mean of the multiple bioactivity parameters is taken as the bioactivity parameter of the molecule under the activity index.

[0365] Optionally, the second acquisition unit 1602 is specifically used for:

[0366] Each molecule was cut by performing acyclic single bond cleavage, dividing each molecule into two parts: an invariant core substructure and a region molecular structure. The invariant core substructure is the common part between molecules, while the region molecular structure is the unique part of each molecule.

[0367] Each pair of molecules with the same invariant core substructure is grouped into a candidate molecule pair.

[0368] By using preset constraints, each candidate molecule pair is screened to obtain matching molecule pairs; the constraints are used to limit the differences in atomic structure and bioactivity parameters between the molecular structures.

[0369] In summary, this application establishes a set of molecular activity cliff screening and evaluation criteria. Specifically, the pre-constructed molecular dataset includes the biological targets that each molecule can bind to and the bioactivity parameters of each molecule. Based on this, molecules that have undergone bioactivity measurement under the same experimental conditions for the same biological target are evaluated on a per-target basis to calculate the similarity between molecules and to classify them into molecular pairs. Specifically, molecular fingerprints are used to obtain the similarity between every two molecules in each molecule, and matching molecular pairs are constructed based on the substructure similarity of each molecule. In this step, this application comprehensively considers both the fingerprint similarity criterion between molecules and the substructure-based similarity evaluation standard. The fingerprint similarity criterion between molecules simplifies complex molecular structures into binary or numerical vectors, enabling efficient calculation of molecular similarity, while the substructure-based similarity evaluation standard focuses on molecular functional regions, accurately identifying and matching key pharmacophores or functional units in chemical structures, thereby improving the accuracy of active compound discovery.

[0370] Based on the above, for each biological target, a molecular activity cliff threshold is set by matching the distribution of molecular pairs under different similarities. Different biological targets correspond to different thresholds, which effectively avoids the subjective method of setting a uniform threshold in related technologies. It also overcomes the limitation of focusing only on changes at a single site in the substructure similarity evaluation standard of related technologies by combining the fingerprint similarity criteria between molecules.

[0371] Finally, for each molecular pair under each biological target, target molecular pairs are selected based on the molecular activity cliff threshold. On this basis, an activity cliff dataset is constructed, which effectively improves the accuracy and consistency of activity cliff identification.

[0372] For ease of description, the above sections are divided into modules (or units) according to their functions and described separately. Of course, in implementing this application, the functions of each module (or unit) can be implemented in one or more software or hardware components.

[0373] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0374] Having introduced the method and apparatus for constructing an active cliff dataset according to an exemplary embodiment of this application, we will now introduce an electronic device according to another exemplary embodiment of this application.

[0375] Those skilled in the art will understand that various aspects of this application can be implemented as a system, method, or program product. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "system."

[0376] Based on the same inventive concept as the above-described method embodiments, this application also provides an electronic device. In one embodiment, the electronic device may be a server, such as... Figure 1 The server 120 is shown. In this embodiment, the structure of the electronic device can be as follows: Figure 17 As shown, it includes a memory 1701, a communication module 1703, and one or more processors 1702.

[0377] The memory 1701 is used to store computer programs executed by the processor 1702. The memory 1701 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and programs required to run instant messaging functions, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.

[0378] Memory 1701 may be volatile memory, such as random-access memory (RAM); memory 1701 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 1701 may be any other medium capable of carrying or storing a desired computer program having the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 1701 may be a combination of the above-described memories.

[0379] Processor 1702 may include one or more central processing units (CPUs) or digital processing units, etc. Processor 1702 is used to implement the above-described method for constructing the active cliff dataset when it calls the computer program stored in memory 1701.

[0380] The communication module 1703 is used to communicate with terminal devices and other servers.

[0381] This application embodiment does not limit the specific connection medium between the memory 1701, communication module 1703, and processor 1702. This application embodiment... Figure 17 The memory 1701 and the processor 1702 are connected via a bus 1704, and the bus 1704 is in Figure 17 The diagram uses thick lines to describe the connections between other components; these are for illustrative purposes only and should not be considered limiting. The 1704 bus can be divided into address bus, data bus, control bus, etc. For ease of description, Figure 17 It is described using only a thick line, but does not indicate that there is only one bus or one type of bus.

[0382] The memory 1701 stores a computer storage medium containing computer-executable instructions for implementing the method for constructing the active cliff dataset according to embodiments of this application. The processor 1702 executes the aforementioned method for constructing the active cliff dataset, such as... Figure 2 As shown.

[0383] In some possible implementations, various aspects of the method for constructing the active cliff dataset provided in this application can also be implemented in the form of a program product, which includes a computer program. When the program product is run on an electronic device, the computer program is used to cause the electronic device to perform the steps in the method for constructing the active cliff dataset according to various exemplary embodiments of this application described above. For example, the electronic device can perform actions such as... Figure 2 The steps are shown in the figure.

[0384] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0385] The program product of the embodiments of this application may employ a portable compact disc read-only memory (CD-ROM) and include a computer program, and may run on an electronic device. However, the program product of this application is not limited thereto. In this document, the readable storage medium may be any tangible medium that contains or stores a program that may be used by or in conjunction with a command execution system, apparatus, or device.

[0386] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a readable computer program. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with a command execution system, apparatus, or device.

[0387] Computer programs contained on readable media may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0388] Computer programs for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The computer program can execute entirely on the user's electronic device, partially on the user's electronic device, as a standalone software package, partially on the user's electronic device and partially on a remote electronic device, or entirely on a remote electronic device or server. In cases involving remote electronic devices, the remote electronic device can be connected to the user's electronic device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external electronic device (e.g., via the Internet using an Internet service provider).

[0389] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0390] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0391] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing a computer-usable computer program.

[0392] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0393] These computer program commands may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the commands stored in the computer-readable storage medium produce an article of manufacture including command means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0394] These computer program commands can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing the commands executed on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0395] Although preferred embodiments of this application 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 the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0396] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for constructing an active cliff dataset, characterized in that, The method includes: Obtain a pre-constructed molecular dataset; the molecular dataset contains multiple molecules, at least one biological target that each molecule can bind to, and a bioactivity parameter for each molecule, the bioactivity parameter being measured under at least one experimental condition; For each molecule in the molecular dataset that can bind to the same biological target and whose bioactivity can be measured under the same experimental conditions, the similarity between the molecular fingerprints of each pair of molecules is obtained; and, by analyzing the similarity of each pair of molecules in their substructures, matching molecule pairs that meet the preset similarity evaluation criteria are selected. For each biological target, based on the distribution of each matching molecule pair under multiple similarities, a similarity is selected as the molecular activity cliff threshold corresponding to the biological target, and target molecule pairs are selected from each molecule based on the molecular activity cliff threshold. Based on the differences in the corresponding bioactivity parameters of each selected target molecule pair, we analyzed whether there was an activity cliff relationship between the target molecule pairs, and constructed an activity cliff dataset based on the analysis results.

2. The method as described in claim 1, characterized in that, The step of selecting a similarity level as the molecular activity cliff threshold corresponding to the biological target based on the distribution of each matching molecule pair under multiple similarity levels includes: Count the number of matching pairs at each similarity level; The number of matched pairs at each similarity level is accumulated in descending order of similarity. The similarity determined when the proportion of accumulated matched molecule pairs reaches the target proportion is used as the molecular activity cliff threshold corresponding to the biological target; the proportion of accumulated matched molecule pairs represents the ratio of the number of accumulated matched molecule pairs to the total number of matched molecule pairs under the biological target.

3. The method as described in claim 1, characterized in that, The step of selecting target molecule pairs from the molecules based on the molecular activity cliff threshold includes: For each pair of molecules, if the similarity between the two molecules is greater than the molecular activity cliff threshold, then the two molecules are considered as a target molecule pair.

4. The method as described in claim 1, characterized in that, The method involves analyzing the differences in bioactivity parameters of each selected target molecule pair to determine whether an activity cliff relationship exists, and constructing an activity cliff dataset based on the analysis results, including: For each target molecule pair, if the difference in bioactivity parameters between the molecules is greater than a preset difference threshold, then it is determined that there is an activity cliff relationship between the molecules of the target molecule pair, and the target molecule pair is taken as a positive sample. If the difference in bioactivity parameters between molecules is not greater than a preset difference threshold, then it is determined that there is no activity cliff relationship between the molecules of the target molecule pair, and the target molecule pair is taken as a negative sample. Based on the positive and negative samples, an active cliff dataset is constructed for training an active cliff prediction model. The active cliff prediction model is used to predict whether an active cliff relationship exists between input molecular pairs.

5. The method as described in claim 4, characterized in that, The method further includes: For each biological target, at least one of the following post-processing steps is performed: If the number of target molecule pairs selected under the biological target is less than the first preset number, then each target molecule pair under the biological target is deleted; If the number of positive samples corresponding to the biological target is less than the second preset number, then each target molecule pair under the biological target is deleted; The second preset quantity is less than the first preset quantity.

6. The method according to any one of claims 1 to 5, characterized in that, The acquisition of the pre-constructed molecular dataset includes: Obtain known molecular databases; the molecular databases are open-source databases used for molecular biological activity research; Based on the first and second screening criteria, molecules and related data that meet the criteria are selected from the molecular database and the molecular dataset is formed. The first screening condition is used to limit the data quality and experimental type of the required molecules; the second screening condition is used to limit the activity index and purity of the required molecules; the relevant data includes the biological targets that each molecule can bind to, and the biological activity parameters of each molecule; the activity index represents the data type of the biological activity parameters.

7. The method as described in claim 6, characterized in that, Before assembling the molecular dataset, the method further includes: For a single molecule targeting a biological site: If the molecule has multiple bioactivity parameters under the same activity index, and the variance of the multiple bioactivity parameters is greater than a preset difference threshold, then the multiple bioactivity parameters are deleted. If the molecule has multiple bioactivity parameters under the same activity index, and the variance of the multiple bioactivity parameters is not greater than the preset difference threshold, then the mean of the multiple bioactivity parameters is taken as the bioactivity parameter of the molecule under the activity index.

8. The method according to any one of claims 1 to 5, characterized in that, The process of analyzing the similarity of substructures between any two molecules and selecting matching molecule pairs that meet preset similarity evaluation criteria includes: Each molecule is cut by performing acyclic single bond cleavage, dividing each molecule into two parts: an invariant core substructure and a region molecular structure. The invariant core substructure is the common part between molecules, and the region molecular structure is the unique part of each molecule. Each pair of molecules with the same invariant core substructure is grouped into a candidate molecule pair. By using preset similarity assessment conditions, each candidate molecule pair is screened to obtain matching molecule pairs; the similarity assessment conditions are used to evaluate the atomic differences and differences in bioactivity parameters between the molecular structures in the candidate molecule pairs.

9. An apparatus for constructing an active cliff dataset, characterized in that, The device includes: The first acquisition unit is used to acquire a pre-constructed molecular dataset; the molecular dataset contains multiple molecules, at least one biological target that each molecule can bind to, and a bioactivity parameter for each molecule, wherein the bioactivity parameter is measured under at least one experimental condition; The second acquisition unit is used to acquire the similarity between the molecular fingerprints of each pair of molecules in the molecular dataset that can bind to the same biological target and whose bioactivity can be measured under the same experimental conditions; and to select matching molecule pairs that meet the preset similarity evaluation conditions by analyzing the similarity of each pair of molecules in their substructures. The threshold determination unit is used to select a similarity as the molecular activity cliff threshold corresponding to the biological target for each biological target based on the distribution of each matching molecule pair under multiple similarities, and select the target molecule pair from each molecule based on the molecular activity cliff threshold. The construction unit is used to analyze whether there is an activity cliff relationship between the selected target molecule pairs based on the differences in the corresponding bioactivity parameters, and to construct an activity cliff dataset based on the analysis results.

10. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any one of the methods described in claims 1 to 8.

11. A computer-readable storage medium, characterized in that, It includes a computer program that, when run on an electronic device, causes the electronic device to perform the steps of any of the methods described in claims 1 to 8.

12. A computer program product, characterized in that, The method includes a computer program stored in a computer-readable storage medium; when a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, causing the electronic device to perform the steps of any one of claims 1 to 8.

Images