Machine learning driven high-throughput screening method for tumor drug active ingredients

Abstract

The present application relates to the technical field of artificial intelligence and biological medicine, and discloses a machine learning driven high-throughput screening method for active ingredients of tumor drugs, comprising the following steps: S1, constructing a multi-source heterogeneous training dataset: integrating chemical structure information of known active ingredients, molecular characteristic information of tumor cell lines and corresponding biological activity measure information from different data sources, and forming a structured data set for model training after standardization and correlation alignment processing. Compound structure and cell line molecular characteristics are respectively deeply encoded by a graph neural network and a one-dimensional convolution network, and the dynamic interaction and weighted fusion of the two heterogeneous characteristics are realized by using an attention mechanism. This architecture enables the model to learn the complex rules that drug activity depends on specific cell background, thereby showing significantly better prediction accuracy and generalization performance than traditional single-mode or simply spliced models when facing large-scale virtual libraries and diverse tumor cell lines.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and biomedicine, specifically to a machine learning-driven high-throughput screening method for active ingredients in tumor drugs. Background Technology

[0002] In drug development, the key challenge is to quickly and accurately discover lead compounds that are effective against specific tumors from a vast number of compounds. Traditional high-throughput screening is costly and has limited throughput. While computational virtual screening can reduce costs, existing methods have two major drawbacks.

[0003] First, the model's information fusion capability is insufficient, which limits the accuracy of prediction. Most methods are based only on compound structure or simply combined with single target information, failing to deeply integrate the multidimensional molecular features unique to tumor cell lines (such as gene expression and mutation spectrum). This shallow or missing association modeling makes the model unable to accurately capture the specific response relationship between drugs and heterogeneous tumors, resulting in large prediction bias when faced with new structural compounds or different cell lines.

[0004] Secondly, the screening process has a low level of intelligence and lacks continuous optimization capabilities. Existing virtual screening systems mostly adopt a fixed parallel computing mode, which cannot dynamically optimize resource scheduling according to tasks, thus limiting high-throughput processing efficiency. More importantly, the screening model is usually static, and the virtual screening results are disconnected from the subsequent wet experimental verification data. Valuable experimental feedback information cannot be used for model iteration and updates, resulting in the system performance not being able to improve itself with the use of the system, causing a waste of data and knowledge resources. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-driven high-throughput screening method for active ingredients in tumor drugs to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-driven high-throughput screening method for active ingredients in tumor drugs, comprising the following steps:

[0007] S1. Construct a multi-source heterogeneous training dataset: Integrate the chemical structure information of known active ingredients, the molecular feature information of tumor cell lines, and the corresponding bioactivity measurement information from different data sources. After standardization and correlation alignment processing, a structured data set for model training is formed.

[0008] S2. Construct and train a multimodal activity prediction model: Using a structured dataset as training samples, construct a multimodal deep learning model that includes a chemical structure encoding network, a cell line feature encoding network, and a cross-modal fusion network; the cross-modal fusion network is configured to receive and fuse the output features of the first two networks and generate a predicted activity evaluation value; use the structured dataset to train and validate the model;

[0009] S3. Implement high-throughput virtual screening: Pair the molecular feature vector of the target tumor cell line with the chemical structure information of all candidate compounds in the virtual compound library, and input it into the multimodal activity prediction model trained in step S2. Through a batch parallel processing mechanism, obtain a list of predicted activity evaluation values ​​for each candidate compound.

[0010] S4. Determine and verify the high-activity candidate set: Sort all candidate compounds in descending order according to the predicted activity evaluation value list. Select the top K high-value compounds or those above the preset activity threshold according to the value of the predicted activity evaluation value to form a high-activity candidate set. Then, verify the high-activity candidate set by in vitro tumor cell inhibition experiments.

[0011] As a preferred embodiment of the present invention, step S1 specifically includes:

[0012] S11. Obtain the standardized molecular formula of the active ingredient and its corresponding standardized activity value from the compound activity database;

[0013] S12. Obtain gene expression profile data of multiple cell lines from the tumor cell line characteristic database;

[0014] S13. Based on the known correlation mapping relationship between components, cell lines and activity, match and integrate the data of S11 and S12 to form an initial dataset consisting of multiple triplet records.

[0015] S14. Perform molecular fingerprinting or graph structure initialization on the chemical structure information in the initial dataset, reduce the dimensionality of the cell line molecular feature information, and generate a numerical feature matrix that the model can directly read.

[0016] As a preferred technical solution of the present invention, in step S2:

[0017] The chemical structure encoding network is a graph neural network, whose input is a molecular diagram represented by atoms and chemical bonds;

[0018] The cell line feature encoding network is a one-dimensional convolutional neural network, whose input is a one-dimensional vector of gene expression profiles;

[0019] The cross-modal fusion network includes attention interaction units to establish the association between chemical features and cell line features, and outputs predicted activity evaluation values ​​through fully connected layers.

[0020] As a preferred embodiment of the present invention, the batch parallel processing mechanism in step S3 is specifically as follows:

[0021] The virtual compound library is divided into multiple compound subsets of equal size;

[0022] Multimodal activity prediction models are deployed synchronously on multiple parallel processing nodes;

[0023] The master node distributes the compound subset index, and each processing node independently undertakes the task of screening a compound subset. After combining the features of all compounds in the compound subset with the features of the target cell line, the complete forward processing flow of the model is completed.

[0024] The processing results from all processing nodes are aggregated and rearranged according to the initial index to form the final list of predicted activity evaluation values.

[0025] As a preferred embodiment of the present invention, in step S3, before executing the batch parallel processing mechanism, the system performs a resource pre-scheduling step:

[0026] Based on the overall size of the virtual compound library, the number of target cell lines, and the real-time load and memory status of each processing node, the optimal size of each compound subset and the task sequence allocated to each processing node are dynamically determined to optimize the overall screening throughput.

[0027] As a preferred technical solution of the present invention, in step S4, the top K compounds with predicted activity evaluation values ​​are selected, or all compounds with evaluation values ​​higher than the activity threshold determined dynamically based on the overall distribution are selected.

[0028] As a preferred embodiment of the present invention, the method further includes the following step after step S4:

[0029] S5. Model Iteration and Optimization: The new chemical structure, cell line and activity data pairs obtained in this experiment are added as new training samples to the structured data set in step S1, and the parameters of the multimodal activity prediction model in step S2 are updated and retrained based on these samples.

[0030] As a preferred embodiment of the present invention, the method further includes the following steps before step S1:

[0031] S0. Screening task initialization: Based on the target tumor type or specific drug resistance phenotype, retrieve and select one or more target tumor cell lines and their corresponding molecular feature vectors from the pre-constructed cell line feature library.

[0032] As a preferred embodiment of the present invention, the standardization and correlation alignment processing in step S1 follows the following rules:

[0033] For chemical structure information, it is uniformly converted into a standardized SMILES string and the chiral information is standardized.

[0034] For bioactivity measurement information, it is uniformly converted into pIC50 values ​​at the same concentration unit;

[0035] For cell line molecular feature information from different platforms, quantile normalization was used to correct the distribution and then perform correlation alignment.

[0036] As a preferred embodiment of the present invention, in step S4, while experimentally verifying the highly active candidate set, the following steps are also performed:

[0037] Compounds from the high-activity candidate set were compared with known antitumor drugs with well-defined mechanisms of action based on chemical structure similarity.

[0038] The output is a validation report containing the following information: predicted and measured activity values ​​of each compound, its similarity score with known drugs, and a proposed potential mechanism of action based on similarity analysis.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. The multimodal deep learning model constructed in this invention deeply encodes the compound structure and cell line molecular features through graph neural networks and one-dimensional convolutional networks, respectively, and uses an attention mechanism to achieve dynamic interaction and weighted fusion of the two heterogeneous features. This architecture enables the model to learn the complex rules that drug activity depends on specific cellular backgrounds. Thus, when faced with large-scale virtual libraries and diverse tumor cell lines, it exhibits significantly better prediction accuracy and generalization performance than traditional single-modal or simple splicing models, directly improving the reliability and hit rate of virtual screening.

[0041] 2. The dynamic resource pre-scheduling mechanism proposed in this invention can intelligently decide on the optimal task partitioning and allocation strategy based on the size of the virtual compound library, the number of targets, and the real-time status of computing nodes. This technology ensures the efficient and balanced utilization of computing resources, maximizes the throughput of ultra-large-scale compound library screening tasks, and makes the processing time predictable, providing key technical support for stable and efficient industrial-grade high-throughput virtual screening.

[0042] 3. This invention automatically feeds back novel, high-quality compound-cell line-activity data generated from experimental verification to the training set and performs targeted iterative updates to the prediction model, establishing a complete closed loop of computational screening → experimental verification → model enhancement. This mechanism enables the system's predictive ability to continuously evolve with data accumulation, overcoming the shortcomings of traditional static model performance solidification and significantly improving the long-term application value and knowledge accumulation efficiency of the method.

[0043] 4. Based on the output of a high-activity candidate set, this invention further integrates chemical structure similarity comparison analysis with known drugs and automatically generates a verification report containing inferences about potential mechanisms of action. This function links computational prediction with expert knowledge in the field of medicinal chemistry, providing researchers with a multi-dimensional decision-making basis that goes beyond simple activity ranking and is more biologically and pharmaceutically insightful, thus accelerating the transformation process from active compounds to candidate drugs with clear mechanisms of action. Attached Figure Description

[0044] Figure 1 This is an overall flowchart of the machine learning-driven high-throughput screening method for active ingredients in tumor drugs according to the present invention. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1: Screening of active ingredients in a gastric cancer cell line with a high incidence rate in an eastern region

[0047] This embodiment aims to fully demonstrate the application process of the method of the present invention, targeting the human gastric cancer AGS cell line to screen for novel potential active ingredients.

[0048] S0, Initialization of the filtering task:

[0049] The user explicitly selects gastric cancer as the target and specifies cell lines that are sensitive to the epidermal growth factor receptor signaling pathway. The system searches the pre-built cell line feature library, matches AGS cell lines derived from gastric cancer, and extracts their corresponding molecular feature vectors, including gene expression profiles and key mutation information.

[0050] S1. Construct a multi-source heterogeneous training dataset:

[0051] S11: Download all compound data labeled with clear antitumor activity from the public compound activity database ChEMBL, obtain their normalized SMILES strings and corresponding normalized activity values, such as half-maximal inhibitory concentration (IC50). For example, approximately 150,000 compound-activity records were obtained.

[0052] S12: Download gene expression profile data covering nearly a thousand tumor cell lines from the Cancer Drug Sensitivity Genomics Database.

[0053] S13: Based on the cell line names and test compounds shared in GDSC and ChEMBL, establish a component-cell line-activity association mapping relationship. Through cross-referencing of the database, integrate about 80,000 valid triplet records to form the initial dataset. Each record is, for example: (SMILES of compound A, expression profile vector of AGS cell line, IC50=1.2μM).

[0054] S14: Process the initial dataset:

[0055] Chemical structure information: Using the RDKit cheminformatics toolkit, all SMILES strings were converted into Morgan molecular fingerprints as an initialization representation for the graph neural network.

[0056] Cell line molecular characteristic information: Dimensional reduction of gene expression profile data. First, about 5,000 core genes that are stably expressed in more than 80% of cell lines in the GDSC data set are selected. Then, principal component analysis is used to reduce the 5,000-dimensional expression vector of each cell line to a 100-dimensional principal component feature vector.

[0057] Bioactivity measurement information: Convert all IC50 values ​​to pIC50 values ​​to meet the needs of model regression prediction.

[0058] Finally, a structured dataset containing 80,000 records is generated, each consisting of a 2048-dimensional fingerprint vector, a 100-dimensional cell line feature vector, and a pIC50 value. This dataset is then randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0059] S2. Construct and train a multimodal activity prediction model:

[0060] The model architecture is constructed as follows:

[0061] Chemical structure encoding network: A 4-layer graph convolutional network is used. The input is the atomic features of the molecule and the adjacency matrix. After passing through the GCN layer and skip connections, a 256-dimensional chemical feature vector is finally obtained through global average pooling.

[0062] Cell line feature encoding network: A three-layer one-dimensional convolutional neural network is used, with kernel sizes of 3, 5, and 3, and channel numbers of 128, 256, and 256 respectively. The input is a 100-dimensional cell line PCA feature vector, and the output is a 256-dimensional cell line feature vector.

[0063] Cross-modal fusion network: First, the two 256-dimensional vectors mentioned above are input into a multi-head attention interaction unit. This unit maps the chemical feature vector to a query matrix and the cell line feature vector to a key and value matrix through linear transformation. The weighted fused 256-dimensional feature vector is calculated by scaling dot product attention formula. This vector then passes through two fully connected layers and finally outputs the predicted pIC50 value.

[0064] Model training: The PyTorch framework was used to train the model on a server equipped with an NVIDIA A100 GPU, using the Adam optimizer and mean squared error loss function. Early stopping was performed on the validation set to obtain a well-trained multimodal activity prediction model.

[0065] S3. Implement high-throughput virtual screening:

[0066] Target cell line: AGS cells.

[0067] Virtual compound library: A subset from the ZINC15 database, containing approximately 1.2 million classes of pharmaceutically active small molecule compounds, provided in SMILES format.

[0068] Screening process:

[0069] 1. Resource pre-scheduling: The system detects that there are 8 available GPU nodes in the current computing cluster, each with 32GB of memory. Based on the size of the compound library and the single objective, the system dynamically decides to divide the compound library into 240 sub-batches, each containing 5000 compounds, and generates a task allocation sequence according to the load of each node.

[0070] 2. Batch Parallel Processing: Trained model instances are loaded onto 8 GPU nodes. Compound subset indices are distributed via the master node, with each node receiving 30 sub-batch tasks. Each node converts the SMILES of 5000 compounds in each batch into Morgan fingerprints, concatenates them with the 100-dimensional feature vector of AGS cells, and inputs this data into the model for forward propagation inference. Each node independently completes the processing of all its sub-batch tasks.

[0071] 3. Results Summary: The master node aggregates the results of all child nodes and rearranges them according to the initial index to form a list containing 1.2 million predicted pIC50 values.

[0072] S4. Identify and validate the high-activity candidate set:

[0073] Candidate set determination: The list of 1.2 million predicted pIC50 values ​​was sorted in descending order. In this embodiment, the top 50 compounds were selected as the high-activity candidate set.

[0074] Experimental validation: A third-party biotechnology service company was commissioned to conduct in vitro MTT assay on the above 50 candidate compounds. 5-fluorouracil and doxorubicin, first-line clinical chemotherapy drugs for gastric cancer, were used as positive controls, and DMSO was used as a negative control. The test concentration was 10 μM. Cell viability was detected after treating AGS cells for 72 hours.

[0075] Extended Analysis: Simultaneously, the Tanimoto coefficient was used to calculate the chemical structural similarity of each candidate compound with known anti-gastric cancer drugs and EGFR inhibitors. The system automatically generated a validation report, which included: compound number, predicted pIC50, measured cell viability, similarity score with each reference drug, and preliminary speculation on potential mechanisms of action based on high similarity.

[0076] Example 2: Screening and Model Iteration of Drug-Resistant Nasopharyngeal Carcinoma Cell Lines with High Incidence in a Certain Region of South China

[0077] This embodiment demonstrates the application of the present invention in addressing drug resistance issues and in continuous model learning.

[0078] S0: The user's goal is to screen for active ingredients in cisplatin-resistant nasopharyngeal carcinoma cells. The system selects the cisplatin-resistant nasopharyngeal carcinoma cell line 5-8F / DDP and its parental sensitive cell line 5-8F as a pair of control targets.

[0079] S1 / S2: The same dataset construction and initial model training methods as in Example 1 are used.

[0080] S3: Using the trained model, perform virtual screening in parallel on the 5-8F and 5-8F / DDP cell lines respectively.

[0081] S4: For each cell line, the top 30 compounds with predicted pIC50 were selected individually. Further, compounds with significantly higher predicted activity on the drug-resistant strain 5-8F / DDP than on the sensitive strain 5-8F were screened, resulting in 15 potential drug resistance reversal candidate compounds. MTT assays were performed to verify their inhibitory effects on the two cell lines.

[0082] S5. Model Iteration and Optimization: The new activity data of the 15 compounds obtained in this validation were used as 30 new triplet records to supplement the original structured dataset. The original model was fine-tuned using the expanded dataset to obtain the optimized model V2.0, which was used for screening of subsequent related tasks.

[0083] Comparison with Example 1: Traditional single-modality screening method

[0084] To highlight the advantages of the multimodal fusion of the present invention, this comparative example is provided, in which the method predicts activity based solely on the chemical structure of the compound.

[0085] 1. Data and Model: Using the same training set as in Example 1, but only the Morgan fingerprint of the compound is used as the input feature, and cell line features are ignored. A deep neural network with a parameter scale similar to that of the chemical structure encoding network of this invention is constructed, and pIC50 is directly predicted from the fingerprint. The network is trained in the same way.

[0086] 2. Screening and validation: The AGS cell line was virtually screened using this single-modality model, and the top 50 compounds predicted were selected.

[0087] 3. Experiment: The 50 compounds were subjected to the same MTT assay as in Example 1 for verification.

[0088] Compare with Example 2: Parallel filtering without resource pre-scheduling

[0089] This comparative example is set up to verify the technical effect of the resource pre-scheduling step of the present invention.

[0090] 1. Method: In step S3 of Example 1, the resource pre-scheduling step is cancelled, and a fixed-size sub-batch is used and randomly allocated to 8 GPU nodes.

[0091] 2. Observation indicators: Record the total time taken from the start of the task to the completion of the summary of all results, and compare it with the time taken in Example 1.

[0092] Example 3: Validation of the impact of standardization rules on model performance

[0093] This embodiment aims to verify the effectiveness of specific standardized rules.

[0094] 1. Set up three different sets of data preprocessing rules:

[0095] Group A (Rules of this Invention): Chemical structure: Standardized SMILES and chiral information; Activity: Standardized pIC50; Cell line characteristics: Quantile normalization.

[0096] Group B (partially normalized): Chemical structure: only SMILES are normalized, chirality is ignored; Activity: uniformly pIC50; Cell line characteristics: Z-score normalized.

[0097] Group C (Simplified Treatment): Chemical Structure: Original SMILES retained; Activity: Original IC50 (nM) values ​​used; Cell Line Characteristics: Minimum-maximum normalization performed only.

[0098] 2. Model Training and Evaluation: Using three sets of structured datasets generated by different preprocessing rules, three multimodal activity prediction models with identical structures (architecture as in Example 1) were trained respectively, and the model performance was evaluated on the same independent test set.

[0099] Experimental Results and Analysis

[0100] Table 1: Comparison of experimental validation results for different screening methods (for AGS cell lines)

[0101]

[0102] Data Analysis:

[0103] Of the 50 candidate compounds screened by the method of this invention, 18 were confirmed by in vitro experiments to inhibit the growth of AGS cells by more than 50% at a concentration of 10 μM, with an experimental hit rate of 36.0%. This hit rate is significantly higher than the 14.0% of the traditional single-modality method and far exceeds the theoretical expectation based on random screening. This strongly demonstrates that the present invention, by integrating chemical structure and cell line molecular characteristics through a multimodal deep learning model, can more accurately predict the activity of compounds on specific tumor cell lines, thereby greatly improving the accuracy and efficiency of high-throughput virtual screening. The effective results of the positive control drug also ensure the credibility of the experimental system itself.

[0104] Table 2: Impact of resource pre-scheduling steps on screening efficiency (for a 1.2 million compound library)

[0105]

[0106] Data Analysis:

[0107] After adopting the resource pre-scheduling strategy of this invention, the total time to screen a compound library of 1.2 million compounds is only 95 minutes. Compared with the fixed allocation strategy without pre-scheduling, the efficiency is improved by about 25.2%. At the same time, the average peak utilization of GPU reaches 92%, indicating that computing resources are more fully utilized. This verifies the technical feature of dynamically determining the optimal scale and task sequence, which can effectively optimize the overall processing throughput and is an important guarantee for achieving stable, efficient and high-throughput screening.

[0108] Table 3: Impact of different data standardization rules on model prediction performance (independent test set)

[0109]

[0110] Data Analysis:

[0111] The model trained using the specific standardized rules defined in this invention exhibits the best predictive performance on the independent test set. It has the highest Pearson correlation coefficient between the predicted and actual measured values ​​and the lowest root mean square error. Compared with rule groups B and C, its performance is significantly improved. This confirms that the specific standardized processing of chemical structure chiral information, bioactivity dimensions, and cross-platform cell line data is not a conventional choice in the field, but a key technical means that can substantially improve the learning effect and final prediction accuracy of multimodal models. These rules provide a high-quality data foundation for the model's high-precision prediction.

[0112] Table 4: Screening and model iteration effects for drug-resistant cell lines (5-8F vs 5-8F / DDP)

[0113]

[0114] Data Analysis:

[0115] The initial model successfully screened 15 candidates from a database of 1.2 million samples, showing significantly better predictive activity against drug-resistant strains than against sensitive strains. Among them, 6 were verified to have clear inhibitory activity against the drug-resistant strain 5-8F / DDP, demonstrating the model's ability to discover potential components that reverse drug resistance. More importantly, after adding this new data to the training set for iterative model optimization, the resulting V2.0 model improved its predictive relevance from 0.79 to 0.85 on a test set containing other drug resistance-related data. This vividly demonstrates that the screening-verification-learning closed loop constituted by step S5 of this invention enables the system to possess intelligent characteristics of continuous evolution and increasing refinement with use, a significant advantage that traditional static screening methods do not have.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A machine learning-driven high-throughput screening method for active ingredients in tumor drugs, characterized in that, Includes the following steps: S1. Construct a multi-source heterogeneous training dataset: Integrate the chemical structure information of known active ingredients, the molecular feature information of tumor cell lines, and the corresponding bioactivity measurement information from different data sources. After standardization and correlation alignment processing, a structured data set for model training is formed. S2. Construct and train a multimodal activity prediction model: Using a structured dataset as training samples, construct a multimodal deep learning model that includes a chemical structure encoding network, a cell line feature encoding network, and a cross-modal fusion network; the cross-modal fusion network is configured to receive and fuse the output features of the first two networks and generate a predicted activity evaluation value; use the structured dataset to train and validate the model; S3. Implement high-throughput virtual screening: Pair the molecular feature vector of the target tumor cell line with the chemical structure information of all candidate compounds in the virtual compound library, and input it into the multimodal activity prediction model trained in step S2. Through a batch parallel processing mechanism, obtain a list of predicted activity evaluation values ​​for each candidate compound. S4. Determine and verify the high-activity candidate set: Sort all candidate compounds in descending order according to the predicted activity evaluation value list. Select the top K high-value compounds or those above the preset activity threshold according to the value of the predicted activity evaluation value to form a high-activity candidate set. Then, verify the high-activity candidate set by in vitro tumor cell inhibition experiments.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S11. Obtain the standardized molecular formula of the active ingredient and its corresponding standardized activity value from the compound activity database; S12. Obtain gene expression profile data of multiple cell lines from the tumor cell line characteristic database; S13. Based on the known correlation mapping relationship between components, cell lines and activity, match and integrate the data of S11 and S12 to form an initial dataset consisting of multiple triplet records. S14. Perform molecular fingerprinting or graph structure initialization on the chemical structure information in the initial dataset, reduce the dimensionality of the cell line molecular feature information, and generate a numerical feature matrix that the model can directly read.

3. The method according to claim 1, characterized in that, In step S2: The chemical structure encoding network is a graph neural network, whose input is a molecular diagram represented by atoms and chemical bonds; The cell line feature encoding network is a one-dimensional convolutional neural network, whose input is a one-dimensional vector of gene expression profiles; The cross-modal fusion network includes attention interaction units to establish the association between chemical features and cell line features, and outputs predicted activity evaluation values ​​through fully connected layers.

4. The method according to claim 1, characterized in that, The batch parallel processing mechanism in step S3 is specifically as follows: The virtual compound library is divided into multiple compound subsets of equal size; Multimodal activity prediction models are deployed synchronously on multiple parallel processing nodes; The master node distributes the compound subset index, and each processing node independently undertakes the task of screening a compound subset. After combining the features of all compounds in the compound subset with the features of the target cell line, the complete forward processing flow of the model is completed. The processing results from all processing nodes are aggregated and rearranged according to the initial index to form the final list of predicted activity evaluation values.

5. The method according to claim 1, characterized in that, In step S3, before executing the batch parallel processing mechanism, the system performs a resource pre-scheduling step: Based on the overall size of the virtual compound library, the number of target cell lines, and the real-time load and memory status of each processing node, the optimal size of each compound subset and the task sequence allocated to each processing node are dynamically determined to optimize the overall screening throughput.

6. The method according to claim 1, characterized in that, In step S4, the top K compounds with predicted activity evaluation values ​​are selected, or all compounds with evaluation values ​​higher than the activity threshold dynamically determined based on the overall distribution are selected.

7. The method according to claim 1, characterized in that, The process following step S4 also includes: S5. Model Iteration and Optimization: The new chemical structure, cell line and activity data pairs obtained in this experiment are added as new training samples to the structured data set in step S1, and the parameters of the multimodal activity prediction model in step S2 are updated and retrained based on these samples.

8. The method according to claim 1, characterized in that, The procedure preceding step S1 also includes: S0. Screening task initialization: Based on the target tumor type or specific drug resistance phenotype, retrieve and select one or more target tumor cell lines and their corresponding molecular feature vectors from the pre-constructed cell line feature library.

9. The method according to claim 1, characterized in that, The standardization and correlation alignment processing in step S1 follows these rules: For chemical structure information, it is uniformly converted into a standardized SMILES string and the chiral information is standardized. For bioactivity measurement information, it is uniformly converted into pIC50 values ​​at the same concentration unit; For cell line molecular feature information from different platforms, quantile normalization was used to correct the distribution and then perform correlation alignment.

10. The method according to claim 1, characterized in that, In step S4, while experimentally verifying the high-activity candidate set, the following steps are also performed: Compounds from the high-activity candidate set were compared with known antitumor drugs with well-defined mechanisms of action based on chemical structure similarity. The output is a validation report containing the following information: predicted and measured activity values ​​of each compound, its similarity score with known drugs, and a proposed potential mechanism of action based on similarity analysis.

Images