A multi-threaded cooperative method for molecular docking in drug virtual screening

Abstract

The present application relates to the field of information screening, and particularly relates to a multi-thread cooperation method for drug virtual screening molecular docking, the present application obtains target molecular clustering cluster and determines typical molecules by extracting molecular features for clustering analysis, collects the interaction features of typical molecules and computing power nodes in historical docking data, and calculates node demand characteristic values; a centralized demand node set of the molecular clustering cluster is constructed based on the characteristic values, when parallel operation is performed, the clustering cluster called at the same time is analyzed, a computing power cluster hotspot map is constructed based on the centralized demand node set, and the parallel operation path is optimized accordingly. The present application establishes the correlation between the molecular clustering characteristics and the computing resource demand, realizes intelligent task scheduling based on molecular features, effectively solves the multi-task resource cooperation and conflict problems, improves the computing resource utilization rate and the drug virtual screening efficiency, and overcomes the system performance bottleneck in large-scale molecular docking calculation.

Description

Technical Field

[0001] This invention relates to the field of information screening, and more particularly to a multi-threaded collaborative method for molecular docking in virtual drug screening. Background Technology

[0002] Virtual drug screening is a crucial step in the modern drug development process. It efficiently screens large-scale compound databases to rapidly discover lead compounds that may bind to specific target proteins, significantly reducing drug development costs and shortening the development cycle. With the rapid development of combinatorial chemistry and high-throughput synthesis technologies, compound libraries have reached the millions or even hundreds of millions in size, posing a challenge to molecular docking computation in terms of massive data processing and enormous computational resources. Traditional virtual screening methods typically rely on high-performance computing clusters, employing multi-threaded parallel processing to accelerate the docking process. However, in real-world large-scale parallel computing environments, efficiently coordinating and managing tens of thousands of computational tasks, optimizing the allocation of computing resources, and avoiding uneven load distribution and access conflicts among computing nodes have become core technical bottlenecks restricting further improvements in virtual screening efficiency. Therefore, developing a multi-threaded collaborative method capable of intelligently predicting computational load and optimizing task scheduling strategies is of great significance for improving the overall performance of drug virtual screening. Currently, various molecular screening methods have been developed.

[0003] Chinese Patent Publication No. CN114566233B discloses a method, apparatus, electronic device, and storage medium for molecular screening, relating to deep learning, graph theory, and bio-information technologies within the field of artificial intelligence. The method involves: obtaining a first label map of the molecule to be screened and a second label map of a reference molecule; forming a molecular pair between each molecule to be screened and the reference molecule; for each molecular pair, obtaining the mapping between vertices and the conflict information between the mappings in the first and second label maps to generate a mapping map of the molecular pair; sampling the mapping map to obtain the maximum weight fully connected subgraph of the molecular pair; and selecting the molecule with the highest similarity to the reference molecule as the target molecule based on the maximum weight fully connected subgraph of each molecular pair. A framework for molecular screening based on Gaussian Bosonic sampling is proposed, enabling efficient implementation of ligand-based virtual drug screening and improving the efficiency of molecular screening.

[0004] However, the following problems still exist in the existing technology.

[0005] 1. Existing technologies do not adequately consider the resource coordination and conflict avoidance issues among multiple tasks in a large-scale parallel computing environment. Existing solutions mostly focus on improving the algorithm efficiency within a single screening task, but fail to address the resource contention and access conflicts caused by concentrated access requests to specific node resources in the computing cluster when multiple parallel tasks run simultaneously, resulting in task blocking and low utilization of computing resources.

[0006] 2. Existing technologies do not fully consider the correlation between molecular structural similarity and computational resource requirements. They fail to predict the demand for specific computing nodes based on molecular clustering characteristics, making it impossible to achieve intelligent scheduling and resource pre-allocation of computing tasks based on molecular characteristics. This makes it difficult to optimize the overall computational throughput at the system level, resulting in low utilization of computational resources. Summary of the Invention

[0007] To address this, the present invention provides a multi-threaded collaborative method for molecular docking in drug virtual screening, which overcomes the problems of insufficient consideration of resource coordination and conflict avoidance among multiple tasks in the prior art under large-scale parallel computing environment, as well as insufficient consideration of the correlation between molecular structure similarity and computing resource requirements, resulting in task blocking and low utilization of computing resources.

[0008] To achieve the above objectives, the present invention provides a multi-threaded collaborative method for molecular docking in drug virtual screening, comprising:

[0009] Molecular features of target molecules are extracted from the molecular library, and the target molecules are clustered based on the molecular features to obtain several target molecule clusters, and several typical molecules within the target molecule clusters are determined.

[0010] When performing docking operations on typical molecules from various molecular clusters within a historical period, the required feature values ​​of the nodes are calculated based on the interaction characteristics of each computing node in the computing power cluster.

[0011] Based on the node demand feature values, a set of centralized demand nodes for molecular clusters is constructed. The set of centralized demand nodes includes several centralized demand nodes and the demand feature values ​​of each centralized demand node.

[0012] In response to the need to perform parallel docking operations on molecules within the target molecule cluster, the parallel operation path is optimized, including:

[0013] Determine the molecules that need to be called under the same call sequence for each parallel computing path, construct a computing power cluster hotspot map based on the set of concentrated demand nodes corresponding to the target molecule cluster to which the molecule belongs, and optimize each of the parallel computing paths based on the computing power cluster hotspot map;

[0014] The interaction features include the interaction frequency of the computing nodes and the load of the computing nodes, and the parallel computing path is the calling order of the target molecular cluster during the docking operation.

[0015] Furthermore, molecular features of target molecules are extracted from the molecular library, and the target molecules are clustered based on these molecular features to obtain several target molecule clusters, including...

[0016] Obtain structural images of each target molecule;

[0017] Calculate the similarity between the structural images of each target molecule;

[0018] Clustering is performed based on the similarity, and target molecules with similarity higher than a preset threshold are grouped into the same cluster.

[0019] Furthermore, the calculation of the similarity between the structural images of each target molecule includes:

[0020] Extract the molecular structure contours from the structural image and construct an image feature vector based on the molecular structure contours;

[0021] Calculate the cosine similarity between the image feature vectors, and determine the cosine similarity as the structural similarity of the corresponding target molecules.

[0022] Furthermore, several typical molecules of the target molecule cluster were identified, including,

[0023] Based on the image feature vector of each target molecule, calculate the average similarity between it and the image feature vectors of all other molecules in the same cluster;

[0024] The calculated average similarity scores are sorted from highest to lowest.

[0025] The target molecules that are selected and ranked at the top according to the preset screening ratio are used as the typical molecules of the corresponding target molecule cluster.

[0026] Furthermore, when performing docking operations on typical molecules from various molecular clusters within historical periods, the interaction characteristics of each computing node in the computing power cluster are considered, and the node requirement characteristic values ​​are calculated, including...

[0027] For each computing node, collect its interaction frequency and load data when processing typical molecules of various molecular clusters within a historical period.

[0028] The ratio of the interaction frequency to the standard value of the interaction frequency is determined as the characteristic value of the interaction frequency;

[0029] The ratio of the load data to the standard value of the load intensity is determined as the characteristic value of the load data;

[0030] Calculate the node demand characteristic value based on the interaction frequency characteristic value and the load data characteristic value.

[0031] Further, calculate the node requirement characteristic values, including,

[0032] The weighted sum of the interaction frequency feature value and the load data feature value is used as the node requirement feature value.

[0033] Furthermore, based on the node demand characteristic values, a set of centralized demand nodes for molecular clusters is constructed, including:

[0034] For each molecular cluster, select computing power nodes from the computing power cluster whose node demand characteristic values ​​exceed a preset demand characteristic value threshold;

[0035] The selected computing power nodes and their corresponding node demand feature values ​​are associated and combined, and then recorded to construct a set of centralized demand nodes for the corresponding molecular clusters.

[0036] Further, determining the numerator to be called for each parallel computing path under the same call sequence includes:

[0037] Analyze the task scheduling sequence of each parallel computing path to identify the target molecular clusters that need to be invoked within the same time period;

[0038] The task scheduling sequence is sorted by the sequence number of the molecule to be called.

[0039] Furthermore, the construction of a computing power cluster hotspot graph based on the set of concentrated demand nodes corresponding to the target molecule cluster to which the molecule belongs includes:

[0040] Determine the number of each node in each centralized demand node set and the corresponding node demand characteristic value;

[0041] Calculate the hotspot value for a single node based on the number of nodes and the node demand characteristics.

[0042] Determine the hotspot values ​​of each node and generate a hotspot map of the computing power cluster.

[0043] Furthermore, the parallel computing paths are optimized based on the hotspot graph of the computing cluster, including:

[0044] Change the task scheduling sequence of each parallel computing path to generate several computing cluster hotspot maps;

[0045] Analyze the hotspot graphs of each computing cluster, calculate the average hotspot value, and record the task scheduling sequence of several parallel computing paths corresponding to the minimum average hotspot value.

[0046] The docking operation is performed based on the recorded task scheduling sequence of each operation path.

[0047] Compared with existing technologies, this invention extracts molecular features for cluster analysis to obtain target molecular clusters and identify typical molecules. It collects the interaction features between typical molecules and computing nodes from historical docking data and calculates node demand feature values. Based on these feature values, it constructs a set of nodes with concentrated demand for molecular clusters. During parallel computation, it analyzes clusters called within the same time period and constructs a computing cluster hotspot graph based on the set of nodes with concentrated demand, thereby optimizing parallel computation paths. This invention establishes the correlation between molecular clustering characteristics and computing resource requirements, realizing intelligent task scheduling based on molecular features. It effectively solves the problems of multi-task resource coordination and conflict, improves computing resource utilization and drug virtual screening efficiency, and overcomes the system performance bottleneck in large-scale molecular docking computation.

[0048] In particular, this invention considers the construction and characterization mechanism of target molecule clusters. In practice, traditional virtual screening methods often directly perform parallel computations on massive numbers of molecules, neglecting the inherent correlation between molecular structural similarity and computational resource requirements, leading to unreasonable allocation of computational resources and low system efficiency. This invention extracts molecular features from target molecules in a molecular library, clusters each target molecule based on these features, obtains several target molecule clusters, and identifies several typical molecules within each cluster. This unique mechanism first acquires structural images of each target molecule, extracts features representing the molecular structural contours and the spatial distribution of functional groups, constructs image feature vectors, and calculates cosine similarity. Target molecules with similarity higher than a preset threshold are grouped into the same cluster. The clustering method based on structural similarity ensures that molecules within each cluster have highly consistent structural features and computational requirements, laying the foundation for establishing a requirement mapping relationship between molecular clusters and computing nodes, improving the accuracy of computational resource allocation and overall system efficiency.

[0049] In particular, this invention considers a typical molecule screening mechanism based on molecular clusters. In practice, traditional molecule screening methods often neglect the representative features of the internal structure of molecular clusters, leading to insufficient accuracy in subsequent computational demand prediction. This invention extracts the structural image features of each target molecule, generates image feature vectors, and calculates cosine similarity. Based on the average similarity ranking, it selects the typical molecules that best represent the cluster structure. The typical molecule screening mechanism ensures that the selected typical molecules can accurately reflect the structural features of the entire molecular cluster, providing a reliable data foundation for subsequent computational demand prediction, improving the accuracy of node demand feature value calculation, and optimizing the overall resource allocation efficiency.

[0050] In particular, this invention considers a node requirement feature value calculation mechanism based on multi-dimensional interaction characteristics. In practice, the load assessment of computing nodes is often simplistic and cannot fully reflect the true performance requirements of nodes when processing specific types of molecules. This invention comprehensively collects the interaction frequency and load data of computing nodes, calculates the interaction frequency feature value and load data feature value respectively, and obtains the node requirement feature value by weighted summation. This unique mechanism realizes a multi-dimensional quantitative assessment of the processing capacity of computing nodes, establishes the correlation between molecular structure characteristics and computing resource requirements, and provides a basis for constructing an accurate set of nodes with concentrated requirements.

[0051] In particular, this invention considers a dynamic optimization mechanism for parallel computing paths based on heatmaps. In practice, when performing docking operations on molecules, multiple computing nodes are used. Within a computing cluster, different nodes may be responsible for different functions. Due to differences in molecule structure, the computational requirements involved in docking operations vary, potentially leading to differences in computational power requirements for different nodes. Consequently, under multiple parallel computing paths, docking operations on multiple molecules may be performed simultaneously, potentially resulting in competition for the same computing node resources, leading to severe access conflicts and system congestion. This invention analyzes the task scheduling sequences of each parallel computing path to identify possible target molecule cluster combinations that may be called within the same time period. Based on a set of nodes with concentrated demand, a heatmap reflecting the expected load distribution is constructed, and the scheduling sequence of computing tasks is dynamically adjusted. This unique mechanism achieves off-peak scheduling of computing tasks for different target molecule clusters, effectively avoiding concurrent access conflicts and improving computing resource utilization and system throughput. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the steps of a multi-threaded collaborative method for molecular docking in virtual drug screening, as described in an embodiment of the invention.

[0053] Figure 2 A logic block diagram for determining the screening ratio in an embodiment of the invention;

[0054] Figure 3 The following is a logical block diagram for filtering computing power nodes whose node demand feature values ​​exceed a preset demand feature value threshold, as an embodiment of the invention.

[0055] Figure 4 This is a logic block diagram for optimizing parallel computing paths according to an embodiment of the invention. Detailed Implementation

[0056] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0057] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0058] Please see Figure 1 As shown, Figure 1 This is a schematic diagram illustrating the steps of a multi-threaded collaborative method for molecular docking in drug virtual screening, according to an embodiment of the invention. The multi-threaded collaborative method for molecular docking in drug virtual screening of the present invention includes:

[0059] Step S1: Extract molecular features of target molecules from the molecular library, cluster each target molecule based on the molecular features, obtain several target molecule clusters, and determine several typical molecules within the target molecule clusters.

[0060] Step S2: Collect typical molecules from various molecular clusters in the historical period and perform docking operations based on the interaction characteristics of each computing node in the computing power cluster, and calculate the node requirement feature value.

[0061] Step S3: Construct a set of centralized demand nodes for molecular clusters based on the node demand feature values. The set of centralized demand nodes includes several centralized demand nodes and the demand feature values ​​of each centralized demand node.

[0062] Step S4, in response to the need to call the molecules of the target molecule cluster for parallel docking operations, optimize the parallel operation path, including,

[0063] Determine the molecules that need to be called under the same call sequence for each parallel computing path, construct a computing power cluster hotspot map based on the set of concentrated demand nodes corresponding to the target molecule cluster to which the molecule belongs, and optimize each of the parallel computing paths based on the computing power cluster hotspot map;

[0064] The interaction features include the interaction frequency of the computing nodes and the load of the computing nodes, and the parallel computing path is the calling order of the target molecular cluster during the docking operation.

[0065] Specifically, molecular docking is a core technology in computational chemistry and drug design. It uses computer simulations to predict the optimal binding mode and binding strength between small molecules (ligands) and biomolecules (receptors, usually proteins). There are no limitations on the computational tools used for molecular docking; those skilled in the art can choose existing molecular docking tools to achieve the corresponding functions. In practice, distributed servers or computing clusters containing multiple computing nodes can be used to run the tools or software required for molecular docking, which will not be elaborated further here.

[0066] Specifically, molecular features of target molecules are extracted from the molecular library, and the target molecules are clustered based on these molecular features to obtain several target molecule clusters, including...

[0067] Obtain structural images of each target molecule;

[0068] Calculate the similarity between the structural images of each target molecule;

[0069] Clustering is performed based on the similarity, and target molecules with similarity higher than a preset threshold are grouped into the same cluster.

[0070] Specifically, the structural images of each target molecule are two-dimensional molecular structure diagrams. The structural images fully present the atomic composition, chemical bond connections, and spatial arrangement of functional groups of the molecule. Standardized drawing specifications ensure that all molecular images have a uniform size ratio and resolution, accurately reflecting the molecular topological structure features, and providing a visual foundation for subsequent image feature extraction and similarity calculation.

[0071] Specifically, calculating the similarity between the structural images of each target molecule includes:

[0072] Extract the molecular structure contours from the structural image and construct an image feature vector based on the molecular structure contours;

[0073] Calculate the cosine similarity between the image feature vectors, and determine the cosine similarity as the structural similarity of the corresponding target molecules.

[0074] Specifically, there are no restrictions on the method of collecting molecular structure contours. It can be understood that the structure is presented as a two-dimensional molecular structure diagram, and the contour can be extracted through image segmentation algorithm after standardization.

[0075] When calculating contour similarity, the image contours are vectorized and then the cosine similarity is calculated. There are no restrictions on the vectorization method. Calculating similarity for image contours is an existing technology, which will not be elaborated here.

[0076] This invention considers the construction and characterization mechanism of target molecule clusters. In practice, traditional virtual screening methods often directly perform parallel computation on massive numbers of molecules, neglecting the inherent correlation between molecular structural similarity and computational resource requirements, leading to unreasonable allocation of computational resources and low system efficiency. This invention extracts molecular features from target molecules in a molecular library, clusters each target molecule based on these features, obtains several target molecule clusters, and identifies several typical molecules within each cluster. This unique mechanism first acquires structural images of each target molecule, extracts features representing the molecular structural contour and spatial distribution of functional groups, constructs image feature vectors, and calculates cosine similarity. Target molecules with similarity higher than a preset threshold are grouped into the same cluster. The clustering method based on structural similarity ensures that molecules within each cluster have highly consistent structural features and computational requirements, laying the foundation for establishing a requirement mapping relationship between molecular clusters and computing nodes, improving the accuracy of computational resource allocation and overall system efficiency.

[0077] Specifically, several typical molecules of the target molecular cluster are identified, including,

[0078] Based on the image feature vector of each target molecule, calculate the average similarity between it and the image feature vectors of all other molecules in the same cluster;

[0079] The calculated average similarity scores are sorted from highest to lowest.

[0080] The target molecules that are selected and ranked at the top according to the preset screening ratio are used as the typical molecules of the corresponding target molecule cluster.

[0081] Please refer to Figure 2 As shown, Figure 2 This is a logic block diagram illustrating the determination of the screening ratio in an embodiment of the invention. The preset screening ratio is determined based on the number of target molecules within the target molecule cluster. If the number of target molecules within the target molecule cluster is greater than a preset threshold, a smaller first screening ratio is used; if the number of target molecules is less than or equal to the preset threshold, a larger second screening ratio is used. The recommended values ​​in this embodiment are: the preset threshold range is [500, 1000]; the first screening ratio ranges from [1.5%, 2.5%]; and the second screening ratio ranges from [2.5%, 5%]. For large-scale clusters, a screening ratio of 1.5%-2.5% ensures representativeness of typical molecules while effectively controlling the computational scale; for small-scale clusters, a higher ratio of 2.5%-5% ensures a sufficient number of typical molecules, avoiding insufficient characterization due to insufficient samples.

[0082] This invention considers a typical molecule screening mechanism based on molecular clusters. In practice, traditional molecule screening methods often neglect the representative features of the internal structure of molecular clusters, leading to insufficient accuracy in subsequent computational demand prediction. This invention extracts the structural image features of each target molecule, generates image feature vectors, and calculates cosine similarity. Based on the average similarity ranking, it selects the typical molecules that best represent the cluster structure. The typical molecule screening mechanism ensures that the selected typical molecules accurately reflect the structural features of the entire molecular cluster, providing a reliable data foundation for subsequent computational demand prediction, improving the accuracy of node demand feature value calculation, and optimizing the overall resource allocation efficiency.

[0083] Specifically, when performing docking operations on typical molecules from various molecular clusters within a historical period, the interaction characteristics of each computing node in the computing power cluster are considered, and the node requirement feature values ​​are calculated, including...

[0084] For each computing node, collect its interaction frequency and load data when processing typical molecules of various molecular clusters within a historical period.

[0085] The ratio of the interaction frequency to the standard value of the interaction frequency is determined as the characteristic value of the interaction frequency;

[0086] The ratio of the load data to the standard value of the load intensity is determined as the characteristic value of the load data;

[0087] Calculate the node demand characteristic value based on the interaction frequency characteristic value and the load data characteristic value.

[0088] Specifically, the interaction frequency refers to the ratio of the number of requests made by a computing node to exchange data with an external database to the processing time when handling docking tasks for a certain type of molecular cluster.

[0089] Specifically, the interaction frequency standard value is the arithmetic mean of the interaction frequencies of all computing nodes processing typical molecules of various molecular clusters within a historical period. This standard value is calculated by statistically analyzing all historical interaction frequency data and reflects the overall interaction level of the computing cluster in handling molecular docking tasks.

[0090] Specifically, load data refers to the GPU computing power utilization rate when computing nodes perform docking operations.

[0091] Specifically, the standard value of load intensity is the arithmetic mean of the proportion of computing power load occupied by each computing node within a historical period.

[0092] This invention considers a node requirement feature value calculation mechanism based on multi-dimensional interaction characteristics. In practice, the load assessment of computing nodes is often simplistic and cannot fully reflect the true performance requirements of nodes when processing specific types of molecules. This invention comprehensively collects the interaction frequency and load data of computing nodes, calculates the interaction frequency feature value and load data feature value respectively, and obtains the node requirement feature value by weighted summation. This unique mechanism realizes a multi-dimensional quantitative assessment of the processing capacity of computing nodes, establishes the correlation between molecular structure characteristics and computing resource requirements, and provides a basis for constructing a precise set of nodes with concentrated requirements.

[0093] Specifically, the computational node requires characteristic values, including:

[0094] The weighted sum of the interaction frequency feature value and the load data feature value is used as the node requirement feature value.

[0095] Specifically, the weighted summation weight of the interaction frequency feature value is 0.35, and the weighted summation weight of the load data feature value is 0.65. The load data feature value directly reflects the real-time processing capability and resource utilization of the computing nodes, and has a higher indicative value for predicting node performance; while the interaction frequency feature value reflects the communication activity of the nodes, its volatility is relatively large, and its data representation is slightly worse. By using a weight ratio of 0.35:0.65, both the node communication needs and the dominant role of the load data are considered, so that the final calculated node demand feature value can more accurately reflect the actual processing capability of the nodes.

[0096] Specifically, a set of centralized demand nodes for molecular clusters is constructed based on the node demand characteristic values, including:

[0097] For each molecular cluster, select computing power nodes from the computing power cluster whose node demand characteristic values ​​exceed a preset demand characteristic value threshold;

[0098] The selected computing power nodes and their corresponding node demand feature values ​​are associated and combined, and then recorded to construct a set of centralized demand nodes for the corresponding molecular clusters.

[0099] Specifically, under normal circumstances, the node demand characteristic value is around 1. To characterize the situation where the node demand characteristic value is too high, the preset demand characteristic value threshold is selected within the range [1.25, 1.5]. In practice, it is preferred to be 1.4.

[0100] Please refer to Figure 3 As shown, Figure 3 The present invention provides a logical block diagram for filtering computing power nodes whose node demand characteristic values ​​exceed a preset demand characteristic value threshold.

[0101] If the node demand characteristic value of a computing power node is greater than the characteristic value threshold, then the computing power node will be selected as a computing power node in the centralized demand node set.

[0102] Specifically, determining the numerators to be called for each parallel computing path under the same call sequence includes:

[0103] Analyze the task scheduling sequence of each parallel computing path and identify the molecule that needs to be called under the same call sequence;

[0104] The task scheduling sequence is sorted by the sequence number of the molecule to be called.

[0105] It is understandable that the computation path consists of multiple parallel molecules arranged in sequence, so as to enable the computing power cluster to perform docking operations on multiple molecules simultaneously.

[0106] The same call sequence refers to sequences with the same number of bits, and for multiple operation paths, molecular docking operations are performed sequentially.

[0107] Specifically, the construction of a computing power cluster hotspot graph based on the set of concentrated demand nodes corresponding to the target molecule cluster to which the molecule belongs includes:

[0108] Determine the number of each node in each centralized demand node set and the corresponding node demand characteristic value;

[0109] Calculate the hotspot value for a single node based on the number of nodes and the node demand characteristics.

[0110] Determine the hotspot values ​​of each node and generate a hotspot map of the computing power cluster.

[0111] Specifically, the hotspot value is the product of the number of nodes and the node demand characteristic value, reflecting the tendency of the node to be demanded when performing docking operations with multiple molecules. The larger the hotspot value, the greater the tendency of the node to be demanded.

[0112] Specifically, the corresponding serial numbers of each node can be arranged in the graph, and hotspot values ​​can be marked on the corresponding serial numbers of the nodes to generate a hotspot graph of the computing power cluster, which reflects the heat value of each node.

[0113] Specifically, please refer to Figure 4 As shown, Figure 4 This is a logic block diagram illustrating the optimization of parallel computing paths according to an embodiment of the invention. The optimization of each parallel computing path is based on a computing cluster hotspot map, including:

[0114] Change the task scheduling sequence of each parallel computing path to generate several computing cluster hotspot maps within several time periods;

[0115] Analyze the hotspot graphs of each computing cluster, calculate the average hotspot value, and record the task scheduling sequence of several parallel computing paths corresponding to the minimum average hotspot value.

[0116] The docking operation is performed based on the recorded task scheduling sequence of each operation path.

[0117] It is understandable that by changing the arrangement of the task scheduling sequence, the concentrated computing power demand on certain nodes can be avoided, and tasks that might have called the same node at the same time can be executed at different times, ultimately generating an optimized parallel computing path that can avoid concurrent access conflicts.

[0118] This invention considers a dynamic optimization mechanism for parallel computing paths based on heatmaps. In practice, when performing docking operations on molecules, multiple computing nodes are used. Within a computing cluster, different nodes may be responsible for different functions. Due to differences in molecule structure, the computational requirements for docking operations vary, potentially leading to differences in computational power requirements for different nodes. Consequently, under multiple parallel computing paths, docking operations on multiple molecules may be performed simultaneously, potentially resulting in competition for the same computing node resources, leading to severe access conflicts and system congestion. This invention analyzes the task scheduling sequences of each parallel computing path to identify possible target molecule cluster combinations that may be called within the same time period. Based on a set of nodes with concentrated demand, a heatmap reflecting the expected load distribution is constructed, and the scheduling sequence of computing tasks is dynamically adjusted. This unique mechanism achieves off-peak scheduling of computing tasks for different target molecule clusters, effectively avoiding concurrent access conflicts and improving computing resource utilization and system throughput.

[0119] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A multi-threaded cooperative method for drug virtual screening molecular docking, characterized in that, The method comprises the following steps: extracting molecular features of target molecules in a molecular library, clustering each of the target molecules based on the molecular features, obtaining a plurality of target molecular clusters, and determining a plurality of typical molecules in each of the target molecular clusters; collecting interaction features of each computing node in the computing cluster when the typical molecules in each type of molecular cluster in the historical period are subjected to docking operation, and calculating node demand characteristic values; constructing a centralized demand node set for each molecular cluster based on the node demand characteristic values, wherein the centralized demand node set comprises a plurality of centralized demand nodes and demand characteristic values of each of the centralized demand nodes; in response to the need to call target molecules in the molecular cluster for parallel docking operation, optimizing the parallel operation path, comprising: determining the required molecules corresponding to each parallel operation path under the same calling sequence, constructing a computing cluster hotspot map based on the centralized demand node set corresponding to the target molecular cluster to which the molecules belong, and optimizing each of the parallel operation paths based on the computing cluster hotspot map; wherein the interaction features include interaction frequency and load of the computing node, and the parallel operation path is the calling sequence of the target molecular cluster in the docking operation process.

2. The multi-threaded cooperative approach for molecular docking in virtual screening of drugs according to claim 1, wherein, extracting molecular features of target molecules in a molecular library, clustering each of the target molecules based on the molecular features, obtaining a plurality of target molecular clusters, comprising obtaining structural images of each target molecule; calculating the similarity between the structural images of each target molecule; clustering based on the similarity, and merging target molecules with a similarity higher than a preset threshold into the same cluster.

3. The multi-threaded cooperative approach for molecular docking of virtual screening of drugs according to claim 2, wherein, The calculation of the similarity between the structural images of each target molecule comprises: extracting the molecular structure contour in the structural image, and constructing an image feature vector based on the molecular structure contour; calculating the cosine similarity between the image feature vectors, and determining the cosine similarity as the structural similarity of the corresponding target molecules.

4. The multi-threaded cooperative approach for molecular docking in virtual screening of drugs according to claim 1, wherein, determining a plurality of typical molecules in each target molecular cluster, comprising: calculating the average similarity of each target molecule with all other molecular image feature vectors in the cluster based on the image feature vector of each target molecule; sorting the calculated average similarity from high to low; selecting the target molecules in the front as the typical molecules of the corresponding target molecular cluster according to a preset screening ratio.

5. The multi-threaded cooperative approach for molecular docking in virtual screening of drugs according to claim 1, wherein, collecting interaction features of each computing node in the computing cluster when the typical molecules in each type of molecular cluster in the historical period are subjected to docking operation, and calculating node demand characteristic values, comprising: for each computing node, collecting interaction frequency and load data when processing the typical molecules of each type of molecular cluster in the historical period; determining the ratio of the interaction frequency to the interaction frequency standard value as the interaction frequency characteristic value; determining the ratio of the load data to the load intensity standard value as the load data characteristic value; calculating the node demand characteristic value based on the interaction frequency characteristic value and the load data characteristic value.

6. The multi-threaded cooperative approach for molecular docking of virtual screening of drugs according to claim 5, wherein, calculating the node demand characteristic value, comprising the weighted sum of the interaction frequency characteristic value and the load data characteristic value is taken as the node demand characteristic value.

7. The multi-threaded cooperative approach for molecular docking in virtual screening of drugs according to claim 1, wherein, constructing a centralized demand node set for each molecular cluster based on the node demand characteristic values, comprising For each molecular cluster, filter out computing nodes whose node demand characteristic values exceed a preset demand characteristic value threshold from the computing node cluster; Combine the filtered computing nodes and their corresponding node demand characteristic values, record and build a centralized demand node set for the corresponding molecular cluster.

8. The multi-threaded cooperative approach for molecular docking in virtual screening of drugs according to claim 1, wherein, The determination of the molecules required by each parallel operation path under the same calling sequence includes: Analyzing the task scheduling sequence of each parallel operation path, and identifying the molecules required under the same calling sequence; Wherein, the task scheduling sequence is the sequence number sorting corresponding to the required molecules.

9. The multi-threaded cooperative approach for molecular docking in virtual screening of drugs according to claim 1, wherein, The construction of the computing node cluster hotspot map based on the centralized demand node set corresponding to the target molecular cluster to which the molecules belong includes: Determine the number of each node in each centralized demand node set and the corresponding node demand characteristic value; Calculate the hotspot value for a single node based on the number of nodes and the node demand characteristic value; Determine the hotspot value of each node and generate a computing node cluster hotspot map.

10. The multi-threaded collaborative method for molecular docking of virtual screening of drugs according to claim 1, wherein, Optimizing each parallel operation path based on the computing node cluster hotspot map includes, Change the task scheduling sequence of each parallel operation path to generate several computing node cluster hotspot maps; Analyze each computing node cluster hotspot map, calculate the hotspot mean, and record the task scheduling sequence of several parallel operation paths corresponding to the minimum hotspot mean; According to the recorded task scheduling sequence of each operation path, perform the docking operation.

Images