Molecular activity prediction method, device, equipment, storage medium and program product

Abstract

The invention discloses a molecular activity prediction method, device and equipment, a storage medium and a program product, and relates to the technical field of drug research and development. In the application, the feature representation of the target protein corresponding to the candidate molecule is additionally considered, the feature representation of the candidate molecule is optimized from the macroscopic protein global structure and the microscopic binding pocket local structure of the target protein, and then the molecular activity of the candidate molecule is predicted according to the perceptual interaction feature of the candidate molecule obtained after optimization. Therefore, by introducing the feature representation of the target protein corresponding to the candidate molecule, not only can the activity mutation phenomenon caused by the micro change of the structure be more accurately explained and predicted, but also the adaptability to different target systems is realized, and the cross-target and multi-task unified modeling is realized; and the prediction efficiency and decision reliability in the actual drug design process are improved.

Description

Technical Field

[0001] This application relates to the technical field of drug development, and in particular to a molecular activity prediction method, a molecular activity prediction method apparatus, a molecular activity prediction method device, a storage medium, and a computer program product. Background Technology

[0002] In drug development, molecular activity prediction is a crucial step in lead compound screening and structure optimization. Molecular activity prediction refers to using computational models to predict the interaction activity between small molecules (compounds) and specific biological targets (such as proteins and enzymes). Among these, the activity cliff (AC) problem is particularly complex. The activity cliff refers to the phenomenon where structurally very similar molecules exhibit significantly different biological activities. Due to the existence of the activity cliff phenomenon, many traditional modeling methods based on molecular similarity or molecular maps have low accuracy and limited generalization ability in practical applications, and may even mislead molecular design decisions.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide a molecular activity prediction method, a molecular activity prediction method apparatus, a molecular activity prediction method device, a storage medium, and a computer program product, aiming to solve the technical problems of inaccurate and inefficient molecular activity prediction caused by the activity cliff.

[0005] To achieve the above objectives, this application proposes a molecular activity prediction method, which includes: Structural feature encoding is performed on candidate molecules and their corresponding target proteins to obtain feature representations of the candidate molecules and the target proteins, respectively. Based on the feature representation of the target protein, the feature representation of the candidate molecule is optimized from the macroscopic global structure of the target protein and the microscopic local structure of the binding pocket, so as to obtain the perceptual interaction features of the candidate molecule. Based on the aforementioned perceptual interaction features, molecular activity prediction is performed on the candidate molecules.

[0006] In one embodiment, the feature representations of the candidate molecule and the target protein are respectively the initial feature distribution of the candidate molecule and the global structural features of the target protein; the step of optimizing the feature representation of the candidate molecule based on the feature representation of the target protein, from the macroscopic global protein structure and the microscopic binding pocket local structure of the target protein, includes: In the macroscopic target semantic guidance module, overall structural guidance information is extracted from the global structural features of the target protein, and the initial feature distribution of the candidate molecules is adjusted based on the overall structural guidance information so that the adjusted candidate molecules are aligned with the overall semantics of the target protein. In the micro-pocket semantic guidance module, local structural guidance information of the binding pocket region of the target protein is extracted from the global structural features of the target protein. The local structural guidance information is then fused with the feature distribution of the candidate molecules adjusted in the macro-target semantic guidance module, so that the candidate molecules adjusted in the micro-pocket semantic guidance module are aligned with the local semantics of the binding pocket region of the target protein.

[0007] In one embodiment, the step of extracting global structural guidance information from the global structural features of the target protein and adjusting the initial feature distribution of the candidate molecule based on the global structural guidance information includes: The global structural features of the target protein are averaged and pooled to extract the global structural features of the target protein and compress them into a semantic vector of uniform dimension. The semantic vector is processed by a weighted generation network to output normalized parameters; The normalization parameters are input as modulation conditions into the normalization layer to perform layer normalization processing on the initial feature distribution of the candidate molecules.

[0008] In one embodiment, the step of fusing the local structural guidance information and the feature distribution of candidate molecules adjusted in the macroscopic target semantic guidance module includes: The local structure guidance information and the feature distribution of candidate molecules adjusted in the macro-target semantic guidance module are input into the cross-attention mechanism for fusion. The feature distribution of candidate molecules adjusted in the macro-target semantic guidance module is used as the query vector, and the combined pocket features corresponding to the local structure guidance information are used as the key value vector.

[0009] In one embodiment, the molecular activity prediction method further includes: A deep fusion structure is constructed by alternately stacking the macroscopic target semantic guidance module and the microscopic pocket semantic guidance module; wherein, the input of the first layer is the initial feature distribution of the candidate molecule, and the input of any layer after the first layer is the latest feature distribution of the candidate molecule output by the previous layer; The target molecular features of the candidate molecule are output through the deep fusion structure.

[0010] In one embodiment, the molecular activity prediction method further includes: Based on the optimized perceptual interaction characteristics of candidate molecules, predict the activity changes of the candidate molecules at different target proteins; and / or, Based on the optimized perceptual interaction features of two structurally similar candidate molecules, it is determined whether there is an activity cliff in the two structurally similar candidate molecules.

[0011] Furthermore, to achieve the above objectives, this application also proposes a molecular activity prediction device, which includes: The encoding module is used to encode the structural features of the candidate molecules and the target proteins corresponding to the candidate molecules, so as to obtain the feature representations of the candidate molecules and the target proteins respectively. An optimization module is used to optimize the feature representation of the candidate molecule based on the feature representation of the target protein, from the macroscopic global structure of the target protein and the microscopic local structure of the binding pocket, to obtain the perceptual interaction features of the candidate molecule. The prediction module is used to predict the molecular activity of the candidate molecules based on the perceived interaction features.

[0012] In addition, to achieve the above objectives, this application also proposes a molecular activity prediction device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the molecular activity prediction method as described above.

[0013] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the molecular activity prediction method described above.

[0014] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the molecular activity prediction method described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: The activity cliff refers to the phenomenon in drug design where, despite only minor differences in molecular structure, the corresponding biological activities can fluctuate dramatically. This phenomenon makes it difficult for traditional modeling methods to accurately assess changes in the activity of candidate compounds, severely limiting the precision and efficiency of drug screening and optimization. Specifically, the nonlinear changes in biological activity corresponding to the activity cliff usually stem from alterations in key modes of action triggered by minor structural differences during binding with target proteins. Traditional activity cliff prediction methods mostly focus on modeling only the structural information of the small molecule itself, neglecting its dynamic binding relationship with receptor proteins. However, changes in the activity of drug molecules are essentially determined by the complex spatial coupling and chemical interactions between them and specific target proteins. Therefore, ignoring target protein characteristics prevents models from perceiving the key mechanisms inducing activity mutations. Especially when facing multiple different targets, traditional models often lack generalization ability and struggle to achieve unified modeling.

[0016] In this application, the feature representation of the target protein corresponding to the candidate molecule is additionally considered. The feature representation of the candidate molecule is optimized from both the macroscopic global structure of the target protein and the microscopic local structure of the binding pocket. Then, based on the sensory interaction characteristics of the optimized candidate molecule, molecular activity is predicted. Thus, by introducing the feature representation of the target protein corresponding to the candidate molecule, not only can the phenomenon of activity mutation caused by minor structural changes be explained and predicted more accurately, but it also has the ability to adapt to different target systems, achieving unified modeling across targets and multiple tasks, and improving the prediction efficiency and decision reliability in the actual drug design process. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the molecular activity prediction method of this application. Figure 2 A schematic diagram of the macroscopic target semantic guidance module provided in the first embodiment of the molecular activity prediction method of this application; Figure 3 A schematic diagram of the micro-pocket semantic guidance module provided in the first embodiment of the molecular activity prediction method of this application; Figure 4 This is a schematic diagram of the architecture provided in the first embodiment of the molecular activity prediction method of this application; Figure 5 This is a flowchart illustrating the second embodiment of the molecular activity prediction method of this application. Figure 6 This is a flowchart illustrating the third embodiment of the molecular activity prediction method of this application. Figure 7 This is a schematic diagram of the module structure of the molecular activity prediction device according to an embodiment of this application; Figure 8 This is a schematic diagram of the hardware operating environment involved in the molecular activity prediction method in the embodiments of this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] In traditional activity cliff (AC) prediction methods, which aim to identify and predict significantly different biological activities between structurally similar small molecules, the determination of the activity cliff is typically achieved by modeling the structural information of the small molecule itself. These methods largely rely on the decomposition and encoding of molecular substructures to capture intermolecular similarities. In other words, traditional methods aim to predict the activity cliff by modeling the structural features of small molecules, relying heavily on internal structural information (such as core structures and substituents). However, these methods do not deeply consider the dynamics and complexity of protein-drug interactions. Most current activity cliff prediction models focus only on the structural information of the small molecule itself, neglecting its binding relationship and mechanism of action with receptor proteins. This is particularly problematic when dealing with multiple targets or different types of protein receptors, where the models are prone to failure. Since molecular activity is essentially a function of "molecule-target" interactions, modeling only the molecular end while ignoring the receptor end makes it difficult for the model to capture the changes in biological response resulting from decisive structural differences.

[0024] To address the aforementioned issues, this application proposes an activity cliff prediction architecture that incorporates a multi-granularity target protein semantic perception mechanism. This architecture integrates receptor protein semantic information and possesses multi-granularity feature perception capabilities. Specifically, it introduces a receptor information fusion mechanism at both the macroscopic (global protein structure) and microscopic (local binding pocket) levels, dynamically regulating molecular characterization through a "conditional guidance" approach to effectively capture key changes related to target binding patterns. Specifically, this architecture introduces a "Multi-Grained Target Perception (MTP) module," which guides the dynamic optimization modeling of molecular features by fusing macroscopic-level global semantic information (Macro-level Target Semantic, MTS) with microscopic-level binding pocket semantic information (Micro-level Pocket Semantic, MPS). In this architecture, the molecule and its corresponding receptor protein are encoded as feature representations, and the MTP module guides the model from the global protein structure to local active sites through a hierarchical attention mechanism, progressively enhancing the model's ability to perceive key structural differences.

[0025] This application is not only applicable to modeling the activity cliff of a single receptor protein target, but also enables unified modeling for multiple receptor protein targets, thus possessing good versatility and scalability. It not only improves the model's ability to identify the AC phenomenon, but also enhances its interpretability and broadens its applicability, providing a unified and efficient technical approach for key tasks such as drug screening and structure optimization. Experimental results show that this method significantly outperforms existing mainstream graph neural network methods on several typical datasets, exhibiting lower root mean square error and stronger predictive interpretability, effectively improving the model's ability to identify functional groups, chemical bonds, and active sites.

[0026] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or molecular activity prediction device capable of performing the above functions. The following description uses a molecular activity prediction device as an example to illustrate this embodiment and the subsequent embodiments.

[0027] Based on this, embodiments of this application provide a method for predicting molecular activity, referring to... Figure 1 , Figure 1 This is a schematic flowchart of the first embodiment of the molecular activity prediction method of this application.

[0028] In this embodiment, the molecular activity prediction method includes steps S10 to S30: Step S10: Structural feature encoding is performed on the candidate molecules and their corresponding target proteins to obtain the feature representations of the candidate molecules and target proteins, respectively. Candidate molecules, i.e., small molecule compounds, and their corresponding target proteins are respectively encoded with structural features. In one embodiment, referencing Figure 4 For the small molecule portion, it is first transformed into a graph structure representation, where nodes represent atoms and edges represent chemical bonds, containing topological connectivity information and chemical properties. Then, a graph neural network (such as GIN, Mole-BERT, etc.) is used to encode this graph, generating a molecular embedding representation with semantic features, representing the structural information of the small molecule portion, thus obtaining the feature representation of the candidate molecule. For the target protein portion, the input is its amino acid sequence, which is encoded using a pre-trained protein language model (such as ESM2) to obtain an embedding vector expressing the global semantic information of the protein, thus obtaining the feature representation of the target protein. Therefore, the initial feature representations of small molecules and receptor proteins are constructed, laying the foundation for subsequent cross-modal interaction modeling.

[0029] It should be noted that the semantics mentioned in this application refer to the deep-level information contained in a molecule and related to its chemical structure and function, focusing on characterizing the biological significance or functional attributes of a molecule through structural or molecular features.

[0030] Step S20: Based on the feature representation of the target protein, the feature representation of the candidate molecule is optimized from the macroscopic global structure of the target protein and the microscopic local structure of the binding pocket, so as to obtain the perceptual interaction features of the candidate molecule. The binding pocket region of a receptor protein is a key functional structure. The structural features of the binding pocket region include its three-dimensional structure, the internal chemical environment of the pocket, and key residues within the pocket.

[0031] In the existing molecular representation learning framework, target proteins are introduced as conditional information to perceive the regulatory patterns of protein interactions on molecules at different scales, from global structures (such as amino acid sequences) to local key regions (such as binding pockets). Molecular representations are dynamically optimized at the semantic level to enhance their ability to perceive structural differences.

[0032] Step S30: Based on the perceptual interaction features, predict the molecular activity of candidate molecules.

[0033] In this embodiment, by fully utilizing the information of target proteins, a multi-level semantic association between molecules and their binding proteins is systematically modeled. Compared with traditional methods that only consider molecular features and focus on homogeneous and less important parts (motif patterns), by incorporating receptor protein information, the interaction between molecules and receptors is better captured, focusing on substitution differences in small molecules. This effectively reveals the potential mechanism of the formation of the activity cliff, effectively improves the identification ability and prediction accuracy of activity mutation-sensitive regions, and has strong generalization and adaptability to multi-target scenarios.

[0034] In one feasible implementation, the feature representations of the candidate molecule and the target protein are respectively the initial feature distribution of the candidate molecule and the global structural features of the target protein; step S20 includes: In the macroscopic target semantic guidance module, the overall structural guidance information is extracted from the global structural features of the target protein. Based on the overall structural guidance information, the initial feature distribution of candidate molecules is adjusted so that the adjusted candidate molecules are aligned with the overall semantics of the target protein. In the micro pocket semantic guidance module, local structural guidance information of the binding pocket region of the target protein is extracted from the global structural features of the target protein. The local structural guidance information is then fused with the feature distribution of the candidate molecules adjusted in the macro target semantic guidance module, so that the candidate molecules adjusted in the micro pocket semantic guidance module are aligned with the local semantics of the binding pocket region of the target protein.

[0035] The macroscopic target semantic guidance module (MTS) is used to guide the global semantic modeling process of small molecules, enabling its representation to adapt to the global structural semantics of different target proteins.

[0036] The Micropocket Semantic Guidance Module (MPS) is dedicated to extracting local information about binding pockets from protein structures and using it as a semantic guidance factor to regulate the expression of key structures in small molecules.

[0037] In another feasible implementation, step S20 includes: Average pooling is performed on the global structural features of the target protein to extract the global structural features of the target protein and compress them into a semantic vector of uniform dimension. The semantic vector is processed by a weighted generation network to output normalized parameters; The normalization parameter is input as a modulation condition into the normalization layer to perform layer normalization on the initial feature distribution of candidate molecules.

[0038] Reference Figure 2First, the semantic representation (Ftarget) of the receptor protein is averaged and pooled. A feature abstractor extracts its global structural features and compresses them into a semantic vector (Fcompress) of uniform dimension. Next, a weight regressor processes this vector, outputting a set of dynamically normalized parameters (scaling and shift parameters γ and β, which are trainable parameters in the normalization process). These parameters serve as overall structural guidance information and are input as modulation conditions into subsequent normalization layers (LayerNorm). These parameters are used to perform layer normalization on the intermediate representation (Fmol) of the molecule, including the initial and latest feature distributions of candidate molecules, thereby achieving semantic regulation of "molecules adjusted according to proteins." The normalized molecular representation then enters the All-Atom Attention module to further extract its global structural features within the protein context. This module achieves global regulation and feature integration of small molecules by proteins, ensuring that the adjusted candidate molecules maintain semantic alignment with the target protein and enabling subsequent predictions to more accurately reflect differences in molecular behavior within the protein context.

[0039] In another feasible implementation, step S20 includes: The local structure guidance information and the feature distribution of candidate molecules adjusted in the macro-target semantic guidance module are input into the cross-attention mechanism for fusion. The feature distribution of candidate molecules adjusted in the macro-target semantic guidance module is used as the query vector, and the combined pocket features corresponding to the local structure guidance information are used as the key value vector.

[0040] Reference Figure 3First, sub-features of the active pocket region of the target protein are extracted from the global structural features (Ftarget) of the target protein. These sub-features serve as local structural guidance information for the binding pocket region of the target protein. Protein 3D structure analysis tools (such as CavityPlus or pocket prior extractor) can be used to automatically locate binding sites in the protein that may form strong interactions with ligands; these binding sites are the active pocket regions. Then, this local feature (Fpocket, i.e., the sub-feature of the active pocket region) and the structural features of small molecules (Fmol, where Fmol is the Fmol after layer normalization, All-Atom Attention, and layer normalization, i.e., the feature distribution of candidate molecules adjusted in the macroscopic target semantic guidance module) are input together into a cross-attention mechanism for fusion. The molecular feature (Fmol) serves as the query vector, and the pocket feature (Fpocket) serves as the key-value vector; a significant correlation between the two is established through cross-attention operations. The final output is an enhanced molecular feature that incorporates pocket regulatory information (the result after pocket-Attention), used to model the impact of minute structural changes on receptor binding ability. This module enhances the model's sensitivity to local mechanisms of action, especially demonstrating higher recognition accuracy in tasks related to active mutations.

[0041] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 5 The molecular activity prediction method further includes steps T10~T20: Step T10: The macroscopic target semantic guidance module and the microscopic pocket semantic guidance module are stacked alternately to construct a deep fusion structure; wherein, the input of the first layer is the initial feature distribution of the candidate molecules, and the input of any layer after the first layer is the latest feature distribution of the candidate molecules output by the previous layer; Step T20: Through deep fusion of structures, the target molecular features of the candidate molecules are output.

[0042] To achieve multi-round enhancement and optimization of molecular representation in semantic space, this embodiment constructs a deeply fused structure by alternately stacking macroscopic target semantic guidance (MTS) and microscopic pocket semantic guidance (MPS) modules. At each layer, the MTS module is first applied for global semantic modulation, aligning the overall molecular structural features with the global background semantics of the protein. Then, the MPS module is executed to locally align and fuse pocket information with the molecule, emphasizing atomic-level features related to the action site. This alternating structure can be stacked in multiple layers, progressively optimizing the molecular representation capability through layer-by-layer semantic enhancement and residual connections. The final output target molecule features are deeply fused with the semantic guidance of the receptor protein, possessing the ability to accurately capture activity mutation patterns.

[0043] It should be noted that there is no fixed limit to the number of stacking layers; the specific number of stacking layers depends on the complexity of the task and the requirements of the model design.

[0044] Based on the first embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to that in the first embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 6 The molecular activity prediction method further includes steps P10-P20: Step P10: Based on the optimized sensory interaction characteristics of candidate molecules, predict the activity changes of candidate molecules at different target proteins; and / or, Step P20: Based on the optimized perceptual interaction features of the two structurally similar candidate molecules, determine whether there is an activity cliff between the two structurally similar candidate molecules.

[0045] Reference Figure 4 The feedforward layer is typically used for final decision-making and output. That is, the optimized molecular representation, i.e., the optimized perceptual interaction features of candidate molecules, can be input into the output header for prediction and execution of specific tasks. In this embodiment, two mainstream task types are supported: activity cliff regression and activity cliff classification. In the regression task, the activity change value (e.g., ΔpKi) of a molecule under different receptor protein conformations is predicted; in the classification task, it determines whether two structurally similar molecules have a significant activity difference, i.e., whether they constitute an activity cliff (AC / Non-AC). It employs a standard fully connected layer and loss function design, resulting in stable output, fast training convergence, and both practicality and engineering deployability.

[0046] In this application, reference is made to Figure 4The Multi-Granularity Target Perception Module (MTP Module) consists of two key components: First, the Macro-level Target Semantic Module (MTS): This module extracts high-level structural guidance from the global semantic information of the protein. By dynamically generating conditional normalization parameters, it adjusts the feature distribution of small molecules to align them with the overall semantics of the target protein, thereby achieving global-level adaptation and perception. Second, the Micro-level Pocket Semantic Module (MPS): This module focuses on the binding pocket region of the receptor protein. It uses a cross-attention mechanism to model local interaction patterns, coupling fine-grained features between small molecules and receptor pockets to effectively capture subtle structural differences that lead to activity mutations.

[0047] In its implementation, MTPNet first encodes molecular and protein sequences separately. Proteins are processed using pre-trained models (such as ESM2) to obtain their semantic features, while molecules are processed using graph neural networks (such as Mole-BERT) to obtain their structural features. Subsequently, the MTP module constructs a molecular representation that integrates global and local protein information through alternating MTS and MPS operations. This representation is continuously iterated and optimized in a multi-layered stacked structure, ultimately outputting a prediction result of perceptual interaction features.

[0048] This three-stage pathway of "receptor sensing - molecular regulation - predictive output" achieves a fundamental leap in activity cliff modeling from a single molecular dimension to "molecule-receptor joint representation." Even when faced with target proteins of different types and binding mechanisms, it can automatically adapt and output a unified structure, demonstrating excellent task adaptability and cross-dataset transferability, truly realizing the generalization and intelligence of activity cliff prediction.

[0049] In this application, to improve the model's versatility and scalability, the MTP module can be designed as an independently pluggable semantic enhancement component, seamlessly integrated into various existing graph neural network architectures such as GCN, GAT, GIN, MolCLR, and Mole-BERT. After integration, the main structure of the original network remains unchanged; the MTP module is simply used as an intermediate processing layer for molecular representation optimization, forming a combined structure of "basic graph network + multi-granularity receptor guidance," significantly improving model performance without requiring complex structural reconstruction. To enhance its practicality and reliability in research and drug design processes, an attention-score-based interpretation module, such as an interpretation module (interpretability and visualization module), can be designed to track the molecular sites and receptor regions of interest during prediction. This module can annotate the atoms or functional groups of interest on the molecular graph using color heatmaps, highlighting their key roles in activity changes. It can also display which amino acid residues in the protein's pocket regions significantly affect the prediction. This interpretability makes the model predictions not only accurate, but also provides structural evidence, making it easier for researchers to understand the prediction logic and further guide drug structure optimization.

[0050] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the molecular activity prediction method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0051] This application also provides a molecular activity prediction device, please refer to... Figure 7 The molecular activity prediction device includes: Encoding module 10 is used to encode the structural features of candidate molecules and the target proteins corresponding to the candidate molecules, so as to obtain the feature representations of the candidate molecules and the target proteins respectively. The optimization module 20 is used to optimize the feature representation of the candidate molecule based on the feature representation of the target protein, from the macroscopic global structure of the target protein and the microscopic local structure of the binding pocket, to obtain the perceptual interaction features of the candidate molecule. The prediction module 30 is used to predict the molecular activity of the candidate molecules based on the perceived interaction features.

[0052] In one embodiment, the feature representations of the candidate molecule and the target protein are the initial feature distribution of the candidate molecule and the global structural features of the target protein, respectively; the optimization module 20 is further configured to: In the macroscopic target semantic guidance module, the overall structural guidance information is extracted from the global structural features of the target protein. Based on the overall structural guidance information, the initial feature distribution of candidate molecules is adjusted so that the adjusted candidate molecules are aligned with the overall semantics of the target protein. In the micro pocket semantic guidance module, local structural guidance information of the binding pocket region of the target protein is extracted from the global structural features of the target protein. The local structural guidance information is then fused with the feature distribution of the candidate molecules adjusted in the macro target semantic guidance module, so that the candidate molecules adjusted in the micro pocket semantic guidance module are aligned with the local semantics of the binding pocket region of the target protein.

[0053] In one embodiment, the optimization module 20 is further configured to: Average pooling is performed on the global structural features of the target protein to extract the global structural features of the target protein and compress them into a semantic vector of uniform dimension. The semantic vector is processed by a weighted generation network to output normalized parameters; The normalization parameter is input as a modulation condition into the normalization layer to perform layer normalization on the initial feature distribution of candidate molecules.

[0054] In one embodiment, the optimization module 20 is further configured to: The local structure guidance information and the feature distribution of candidate molecules adjusted in the macro-target semantic guidance module are input into the cross-attention mechanism for fusion. The feature distribution of candidate molecules adjusted in the macro-target semantic guidance module is used as the query vector, and the combined pocket features corresponding to the local structure guidance information are used as the key value vector.

[0055] In one embodiment, the molecular activity prediction device further includes a stacking module for: A deep fusion structure is constructed by alternately stacking the macro-target semantic guidance module and the micro-pocket semantic guidance module; wherein, the input of the first layer is the initial feature distribution of the candidate molecules, and the input of any layer after the first layer is the latest feature distribution of the candidate molecules output by the previous layer; By deeply fusing the structure, the target molecular features of the candidate molecules are output.

[0056] In one embodiment, the molecular activity prediction device further includes an application module for: Based on the optimized perceptual interaction characteristics of candidate molecules, predict the activity changes of candidate molecules at different target proteins; and / or, Based on the optimized perceptual interaction features of two structurally similar candidate molecules, it is determined whether there is an activity cliff in the two structurally similar candidate molecules.

[0057] The molecular activity prediction device provided in this application, employing the molecular activity prediction method described in the above embodiments, can solve the technical problems of inaccurate and inefficient molecular activity prediction caused by the activity cliff. Compared with the prior art, the beneficial effects of the molecular activity prediction device provided in this application are the same as those of the molecular activity prediction method provided in the above embodiments, and other technical features in the molecular activity prediction device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0058] This application provides a molecular activity prediction device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the molecular activity prediction method in Embodiment 1 above.

[0059] The following is for reference. Figure 8 The diagram illustrates a structural schematic suitable for implementing the molecular activity prediction device in the embodiments of this application. The molecular activity prediction device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 8 The molecular activity prediction device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application.

[0060] like Figure 8As shown, the molecular activity prediction device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the molecular activity prediction device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the molecular activity prediction device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows molecular activity prediction devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0061] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0062] The molecular activity prediction device provided in this application, employing the molecular activity prediction method described in the above embodiments, can solve the technical problems of inaccurate and inefficient molecular activity prediction caused by the activity cliff. Compared with the prior art, the beneficial effects of the molecular activity prediction device provided in this application are the same as those of the molecular activity prediction method provided in the above embodiments, and other technical features of this molecular activity prediction device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0063] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0065] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to perform the molecular activity prediction method in the above embodiments.

[0066] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer 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. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0067] The aforementioned computer-readable storage medium may be included in the molecular activity prediction device; or it may exist independently and not assembled into the molecular activity prediction device.

[0068] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the molecular activity prediction device, the molecular activity prediction device performs the following: structural feature encoding on candidate molecules and their corresponding target proteins to obtain feature representations of candidate molecules and target proteins, respectively; based on the feature representations of target proteins, it optimizes the feature representations of candidate molecules from the macroscopic global structure of the target protein and the microscopic local structure of the binding pocket to obtain the perceptual interaction features of candidate molecules; and based on the perceptual interaction features, it performs molecular activity prediction on candidate molecules.

[0069] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0070] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0071] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0072] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described molecular activity prediction method. This solves the technical problem of inaccurate and inefficient molecular activity prediction caused by the activity cliff. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the molecular activity prediction method provided in the above embodiments, and will not be repeated here.

[0073] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the molecular activity prediction method described above.

[0074] The computer program product provided in this application can solve the technical problem of inaccurate and inefficient molecular activity prediction caused by the activity cliff. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the molecular activity prediction method provided in the above embodiments, and will not be repeated here.

[0075] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for predicting molecular activity, characterized in that, The molecular activity prediction method includes: Structural feature encoding is performed on candidate molecules and their corresponding target proteins to obtain feature representations of the candidate molecules and the target proteins, respectively. Based on the feature representation of the target protein, the feature representation of the candidate molecule is optimized from the macroscopic global structure of the target protein and the microscopic local structure of the binding pocket, so as to obtain the perceptual interaction features of the candidate molecule. Based on the aforementioned perceptual interaction features, molecular activity prediction is performed on the candidate molecules.

2. The molecular activity prediction method as described in claim 1, characterized in that, The feature representations of the candidate molecule and the target protein are respectively the initial feature distribution of the candidate molecule and the global structural features of the target protein; the step of optimizing the feature representation of the candidate molecule based on the feature representation of the target protein, from the macroscopic global protein structure and the microscopic binding pocket local structure of the target protein, includes: In the macroscopic target semantic guidance module, overall structural guidance information is extracted from the global structural features of the target protein, and the initial feature distribution of the candidate molecules is adjusted based on the overall structural guidance information so that the adjusted candidate molecules are aligned with the overall semantics of the target protein. In the micro-pocket semantic guidance module, local structural guidance information of the binding pocket region of the target protein is extracted from the global structural features of the target protein. The local structural guidance information is then fused with the feature distribution of the candidate molecules adjusted in the macro-target semantic guidance module, so that the candidate molecules adjusted in the micro-pocket semantic guidance module are aligned with the local semantics of the binding pocket region of the target protein.

3. The molecular activity prediction method as described in claim 2, characterized in that, The step of extracting global structural guidance information from the global structural features of the target protein and adjusting the initial feature distribution of the candidate molecules based on the global structural guidance information includes: The global structural features of the target protein are averaged and pooled to extract the global structural features of the target protein and compress them into a semantic vector of uniform dimension. The semantic vector is processed by a weighted generation network to output normalized parameters; The normalization parameters are input as modulation conditions into the normalization layer to perform layer normalization processing on the initial feature distribution of the candidate molecules.

4. The molecular activity prediction method as described in claim 2, characterized in that, The step of fusing the local structural guidance information and the feature distribution of candidate molecules adjusted in the macroscopic target semantic guidance module includes: The local structure guidance information and the feature distribution of candidate molecules adjusted in the macro-target semantic guidance module are input into the cross-attention mechanism for fusion. The feature distribution of candidate molecules adjusted in the macro-target semantic guidance module is used as the query vector, and the combined pocket features corresponding to the local structure guidance information are used as the key value vector.

5. The molecular activity prediction method as described in claim 2, characterized in that, The molecular activity prediction method further includes: A deep fusion structure is constructed by alternately stacking the macroscopic target semantic guidance module and the microscopic pocket semantic guidance module; wherein, the input of the first layer is the initial feature distribution of the candidate molecule, and the input of any layer after the first layer is the latest feature distribution of the candidate molecule output by the previous layer; The target molecular features of the candidate molecule are output through the deep fusion structure.

6. The molecular activity prediction method as described in claim 1, characterized in that, The molecular activity prediction method further includes: Based on the optimized perceptual interaction characteristics of candidate molecules, predict the activity changes of the candidate molecules at different target proteins; and / or, Based on the optimized perceptual interaction features of two structurally similar candidate molecules, it is determined whether there is an activity cliff in the two structurally similar candidate molecules.

7. A molecular activity prediction device, characterized in that, The molecular activity prediction device includes: The encoding module is used to encode the structural features of the candidate molecules and the target proteins corresponding to the candidate molecules, so as to obtain the feature representations of the candidate molecules and the target proteins respectively. An optimization module is used to optimize the feature representation of the candidate molecule based on the feature representation of the target protein, from the macroscopic global structure of the target protein and the microscopic local structure of the binding pocket, to obtain the perceptual interaction features of the candidate molecule. The prediction module is used to predict the molecular activity of the candidate molecules based on the perceived interaction features.

8. A molecular activity prediction device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the molecular activity prediction method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the molecular activity prediction method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the molecular activity prediction method as described in any one of claims 1 to 6.

Images