A drug efficacy prediction method and system based on an end-to-end deep learning framework

Abstract

This invention belongs to the field of drug efficacy prediction technology and discloses a drug efficacy prediction method and system based on an end-to-end deep learning framework. The method includes: generating a hierarchical phenotypic dataset based on the microscopic changes in the transcriptional profiles of drugs and diseases, and calculating state embedding vectors using the hierarchical phenotypic dataset to construct an end-to-end prediction model for drug efficacy prediction; training the end-to-end prediction model using a binary cross-entropy loss function, evaluating the performance of the trained end-to-end prediction model based on the operating characteristic curve, and outputting a drug clinical efficacy prediction model after the performance results meet the requirements; evaluating the comprehensive recommendation score of drugs for diseases based on the drug clinical efficacy prediction model and drug costs, and outputting the recommended ranking results of drugs corresponding to diseases based on the comprehensive recommendation score. This invention can improve the accuracy of drug efficacy prediction and optimize the drug screening process, providing support for drug discovery for complex diseases.

Description

Technical Field

[0001] This invention relates to the field of drug efficacy prediction technology, and in particular to a drug efficacy prediction method and system based on an end-to-end deep learning framework. Background Technology

[0002] Drug discovery is a high-cost, high-risk process, especially in the study of complex diseases. Traditional methods often have a very high failure rate. Although artificial intelligence (AI) has made some progress in drug discovery, current AI methods mainly focus on a single research and development stage, often neglecting the effective integration of clinical data and prior knowledge, resulting in a significant translational gap in the drug development process.

[0003] Currently, many AI methods rely on preclinical experimental data and cell / animal models for drug prediction. However, these models cannot fully reflect the complex physiological environment of the human body, and the efficacy in clinical trials often cannot be directly matched with laboratory results. Therefore, how to effectively integrate clinical data with existing computing frameworks is the key to solving this translational challenge.

[0004] Therefore, how to provide a drug efficacy prediction method and system based on an end-to-end deep learning framework is an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a drug efficacy prediction method and system based on an end-to-end deep learning framework to solve the problems mentioned above in the prior art.

[0006] According to a first aspect of the present invention, a method for predicting drug efficacy based on an end-to-end deep learning framework is provided.

[0007] In one embodiment, the drug efficacy prediction method based on an end-to-end deep learning framework includes: Based on the microscopic changes in the transcriptional profiles of drugs and diseases, a hierarchical phenotypic dataset is generated, and state embedding vectors are calculated using the hierarchical phenotypic dataset to construct an end-to-end prediction model for drug efficacy prediction. The end-to-end prediction model is trained using the binary cross-entropy loss function, and the performance of the trained end-to-end prediction model is evaluated based on the working characteristic curve. Once the performance results meet the requirements, the drug clinical efficacy prediction model is output. Based on the drug clinical efficacy prediction model and drug cost, the comprehensive recommendation score of drugs for diseases is evaluated, and the recommendation ranking results of drugs corresponding to diseases are output according to the comprehensive recommendation score.

[0008] In one embodiment, a hierarchical phenotypic dataset is generated based on microscopic changes in the transcriptional profiles of drugs and diseases, and state embedding vectors are calculated using the hierarchical phenotypic dataset to construct an end-to-end prediction model for drug efficacy prediction, including: A drug judgment set containing drugs and marker genes is obtained based on a bioinformatics dataset, and the average of the micro-changes in the transcriptional profile of the drugs is taken. The micro-change vector of the drug transcriptional profile is obtained based on the average result. Gene expression was selected from the differential expression feature library, and common genes were extracted based on the intersection of gene expression and marker genes to represent the micro-changes in the transcriptional spectrum of the disease, so as to obtain the vector of micro-changes in the transcriptional spectrum of the disease. The micro-change vectors of drug transcription profiles and disease transcription profiles are integrated, and disease-drug pairs are generated as a hierarchical phenotypic dataset by proportionally sampling from the integration results and the therapeutic target database. Extract the hidden state embedding vectors corresponding to the hierarchical phenotypic dataset, and after projecting the hidden state embedding vectors, output an end-to-end prediction model to describe the probability of predicting the efficacy of drugs for diseases.

[0009] In one embodiment, extracting the latent state embedding vectors corresponding to the hierarchical phenotypic dataset, and after projecting the latent state embedding vectors, outputting an end-to-end prediction model describing the probability of drug efficacy prediction for a disease includes: A drug encoder for drug encoding, a disease encoder for disease encoding, and a decoder for predicting disease-drug relationships are constructed based on a hierarchical phenotypic dataset. By using a drug encoder and a disease encoder, the micro-change vectors of drug transcription profiles and disease transcription profiles are progressively reduced to the target dimension. The dimensionality reduction result is mapped to the shared latent feature space to obtain the drug latent state embedding vector and the disease latent state embedding vector. The feature tensor is then fused with the drug latent state embedding vector and the disease latent state embedding vector through the outer product mapping mechanism to obtain the embedding representation vector. After flattening the embedded representation vector, it is input into the decoder that predicts the relationship between the disease and the drug to construct an end-to-end prediction model. The end-to-end prediction model is then used to output the predicted efficacy probability of the drug for the disease.

[0010] In one embodiment, the end-to-end prediction model is trained using a binary cross-entropy loss function, and the performance of the trained end-to-end prediction model is evaluated based on the operating characteristic curve. Once the performance results meet the requirements, the output drug clinical efficacy prediction model includes: The hierarchical phenotypic dataset is divided into a training set and a test set, and ten-fold cross-validation is performed on the training set to divide the training set into a training subset and a validation subset. Set the total number of training rounds, perform iterative training of the end-to-end prediction model on the training subset, and calculate the loss value of the validation subset using the binary cross-entropy loss function after the training rounds are completed. Select the model parameters corresponding to the minimum loss value as the optimal parameters of the end-to-end prediction model to complete the training optimization. The target drug is predicted using the trained and optimized end-to-end prediction model to determine the predicted efficacy probability of the target drug for the disease, and the area under the working feature curve and the area under the precision and recall curve are plotted based on the predicted efficacy probability. The area under the working feature curve and the area under the precision and recall curve were compared with the connectivity score and the reverse gene expression score, respectively, to verify the performance of the training and optimization of the end-to-end prediction model. Based on the optimization results, the end-to-end prediction model was further optimized until the performance met the requirements, and the drug clinical efficacy prediction model was output.

[0011] In one embodiment, based on a drug clinical efficacy prediction model and drug cost, a comprehensive recommendation score for the drug for the disease is evaluated, and a ranking of recommended drugs for the corresponding disease is output based on the comprehensive recommendation score, including: Based on the drug clinical efficacy prediction model, identify drugs whose clinical efficacy probability for the target disease meets the target value, and match the drug procurement cost with illustrations to generate a drug medical dataset. The drug and medical dataset is input into a bidirectional encoder representation model for word vector transformation to obtain a drug text word embedding matrix. The semantic association of the drug text word embedding matrix is ​​then analyzed using a self-attention mechanism. Based on semantic association, determine the semantic feature vector of drug efficacy, evaluate and classify the semantic feature vector of efficacy, and determine the evaluation index vector. Based on the evaluation index vector and drug specification definition, the efficacy comprehensive evaluation feature vector is defined, and the change rate of efficacy cost within a continuous step is analyzed in combination with the cost control effect. The basic comprehensive recommendation score of the drug is calculated based on the change rate. A simulation algorithm is used to sample the basic comprehensive recommendation scores of drugs to analyze the matching degree between users and drugs, calculate the comprehensive recommendation scores of drugs, and sort them in order to obtain the recommendation ranking results of drugs.

[0012] In one embodiment, a comprehensive efficacy evaluation feature vector is defined based on the evaluation index vector and drug specifications. This is combined with cost control effectiveness analysis to determine the rate of change in efficacy costs over consecutive steps. The basic comprehensive recommendation score for the drug is calculated based on this rate of change, including: The dosage form and specification data of the drug are uniquely encoded. Based on the encoding results, the basic feature vector of the drug is determined. The evaluation index vector is then concatenated with the basic feature vector of the drug to obtain the comprehensive efficacy evaluation feature vector. A state space is constructed based on the actual efficacy and cost control effect of historically recommended drugs, a decision state matrix is ​​generated, and the weight update strategy is optimized by combining the efficacy comprehensive evaluation feature vector to obtain a dynamic weight coefficient group of efficacy and cost. Calculate the rate of change of the dynamic weight coefficient group of efficacy cost within a continuous step size, compare the rate of change with a preset convergence threshold, determine the weight convergence when the rate of change is less than the threshold, and generate the weight convergence detection result. The weight coefficients are determined based on the weight convergence test results, and the efficacy prediction score and cost score are weighted and fused using the weight coefficients to output a comprehensive drug recommendation score.

[0013] In one embodiment, a simulation algorithm is used to sample the basic comprehensive recommendation scores of drugs to analyze the matching degree between users and drugs, calculate the comprehensive recommendation scores of drugs, and rank them in order to obtain the drug recommendation ranking results, including: The basic comprehensive recommendation score of drugs and the risk of drug price fluctuations are used as random variables and input into the Monte Carlo simulation algorithm for random sampling to obtain a risk variable sample set. The batch processing engine is then used to perform streaming calculations on the risk variable sample set to dynamically update the drug application profile. The matching degree between users and drugs is analyzed based on the updated drug application profile results, which serves as the real-time matching score. The real-time matching score is then weighted and integrated with the basic comprehensive drug recommendation score to obtain the comprehensive drug recommendation score. The drug recommendation ranking is determined based on the descending order of the comprehensive drug recommendation score.

[0014] According to a second aspect of the present invention, a drug efficacy prediction system based on an end-to-end deep learning framework is provided.

[0015] In one embodiment, the drug efficacy prediction system based on an end-to-end deep learning framework includes: The end-to-end prediction model building module is used to generate a hierarchical phenotypic dataset based on the micro-changes in the transcriptional profile of drugs and diseases, and to use the hierarchical phenotypic dataset to calculate the state embedding vector in order to build an end-to-end prediction model for drug efficacy prediction. The end-to-end prediction model optimization module is used to train the end-to-end prediction model using the binary cross-entropy loss function, evaluate the performance of the trained end-to-end prediction model based on the working characteristic curve, and output the drug clinical efficacy prediction model after the performance results meet the requirements. The drug recommendation generation and ranking module is used to evaluate the comprehensive recommendation score of drugs for diseases based on drug clinical efficacy prediction models and drug costs, and output the recommendation ranking results of drugs corresponding to diseases based on the comprehensive recommendation score.

[0016] According to a third aspect of the present invention, a computer device is provided.

[0017] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0018] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0019] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0020] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention designs an end-to-end deep learning framework that embeds macroscopic clinical trial results, microscopic drug molecular transcriptomic responses, and microscopic disease transcriptomic changes for predicting drug efficacy in the clinical stage. By combining macroscopic clinical phenotypic data, microscopic transcriptomic disease change data, and microscopic drug interference omics change data, it learns the corresponding macroscopic-microscopic data mapping patterns, thereby optimizing the drug screening and efficacy prediction process. This can significantly shorten the drug development cycle and improve the clinical translation rate of drugs, thereby improving the accuracy of drug efficacy prediction and optimizing the drug screening process to support drug discovery for complex diseases.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0023] Figure 1 This is a flowchart illustrating a drug efficacy prediction method based on an end-to-end deep learning framework according to an exemplary embodiment; Figure 2 This is a block diagram illustrating the principle of a drug efficacy prediction system based on an end-to-end deep learning framework, according to an exemplary embodiment. Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment; Figure 4 This is a schematic diagram of the structure of an end-to-end prediction model in a drug efficacy prediction method based on an end-to-end deep learning framework, according to an exemplary embodiment. Figure 5 This is one of the schematic diagrams illustrating the performance results of a drug efficacy prediction method based on an end-to-end deep learning framework under ten-fold cross-validation, according to an exemplary embodiment. Figure 6 This is the second schematic diagram illustrating the performance results of a drug efficacy prediction method based on an end-to-end deep learning framework under ten-fold cross-validation, according to an exemplary embodiment. Figure 7 This is a schematic diagram illustrating the measurement of dorsal skin thickness in normal mice and SSc mice treated with testosterone undecanoate, dinogest, and citalopram in a drug efficacy prediction method based on an end-to-end deep learning framework, according to an exemplary embodiment. Figure 8 This is a schematic diagram of the histopathological staining of the skin thickness on the back of normal mice and SSc mice treated with testosterone undecanoate, dinogest, and citalopram, in a drug efficacy prediction method based on an end-to-end deep learning framework, according to an exemplary embodiment. Detailed Implementation

[0024] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0025] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Figure 1 An embodiment of a drug efficacy prediction method based on an end-to-end deep learning framework of the present invention is shown.

[0028] In this optional embodiment, the drug efficacy prediction method based on an end-to-end deep learning framework includes: Step S101: Based on the microscopic changes in the transcriptional profile of drugs and diseases, a hierarchical phenotypic dataset is generated, and the state embedding vector is calculated using the hierarchical phenotypic dataset to construct an end-to-end prediction model for drug efficacy prediction. Step S102: The end-to-end prediction model is trained using the binary cross-entropy loss function, and the performance of the trained end-to-end prediction model is evaluated based on the working characteristic curve. After the performance results meet the requirements, the drug clinical efficacy prediction model is output. Step S103: Based on the drug clinical efficacy prediction model and drug cost, evaluate the comprehensive recommendation score of the drug for the disease, and output the recommendation ranking result of the corresponding drug for the disease according to the comprehensive recommendation score.

[0029] In this optional embodiment, a hierarchical phenotypic dataset is generated based on the micro-changes in the transcriptional profiles of drugs and diseases. The state embedding vectors are then calculated using the hierarchical phenotypic dataset to construct an end-to-end prediction model for drug efficacy prediction. This includes: obtaining a drug judgment set containing drugs and marker genes based on a bioinformatics dataset; averaging the micro-changes in the transcriptional profiles of drugs to obtain a drug transcriptional profile micro-change vector; selecting gene expressions from a differential expression feature library and extracting common genes as micro-changes in the transcriptional profiles of diseases based on the intersection of gene expressions and marker genes to obtain a disease transcriptional profile micro-change vector; integrating the drug transcriptional profile micro-change vector and the disease transcriptional profile micro-change vector, and generating disease-drug pairs as a hierarchical phenotypic dataset based on the integration result and proportional random sampling from a therapeutic target database; extracting the latent state embedding vectors corresponding to the hierarchical phenotypic dataset; and after projecting the latent state embedding vectors, outputting an end-to-end prediction model describing the probability of drug efficacy prediction for diseases.

[0030] In this optional embodiment, extracting the latent state embedding vectors corresponding to the hierarchical phenotypic dataset and, after projecting the latent state embedding vectors, outputting an end-to-end prediction model describing the predicted probability of drug efficacy for a disease includes: constructing a drug encoder for drug encoding, a disease encoder for disease encoding, and a decoder for predicting the drug-disease relationship based on the hierarchical phenotypic dataset; using the drug encoder and disease encoder, progressively reducing the dimensionality of the drug transcription spectrum micro-change vector and the disease transcription spectrum micro-change vector to the target dimension; mapping the dimensionality reduction result to a shared latent feature space to obtain the drug latent state embedding vector and the disease latent state embedding vector, and performing feature tensor fusion on the drug latent state embedding vector and the disease latent state embedding vector through an outer product mapping mechanism to obtain the embedding representation vector; flattening the embedding representation vector and inputting it into the decoder for predicting the drug-disease relationship to construct an end-to-end prediction model, and using the end-to-end prediction model to output the predicted efficacy probability of drug efficacy for a disease.

[0031] In this optional embodiment, the end-to-end prediction model is trained using a binary cross-entropy loss function, and the performance of the trained end-to-end prediction model is evaluated based on the operating characteristic curve. The output of the drug clinical efficacy prediction model after the performance results meet the requirements includes: dividing the hierarchical phenotypic dataset into training and testing sets, and performing 10-fold cross-validation on the training set to divide the training set into training and validation subsets; setting the total number of training rounds, iteratively training the end-to-end prediction model on the training subset, and calculating the loss value of the validation subset using the binary cross-entropy loss function after the training rounds are completed, selecting the minimum loss. The corresponding model parameters are used as the optimal parameters of the end-to-end prediction model to complete the training and optimization. The trained and optimized end-to-end prediction model is used to predict the target drug and determine the predicted efficacy probability of the target drug for the disease. The area under the working feature curve and the area under the precision and recall curve are plotted based on the predicted efficacy probability. The area under the working feature curve and the area under the precision and recall curve are compared with the connectivity score and the reverse gene expression score, respectively, to verify the performance of the trained and optimized end-to-end prediction model. The end-to-end prediction model is then optimized a second time based on the optimization results until the performance meets the requirements, and the drug clinical efficacy prediction model is output.

[0032] In this optional embodiment, based on the drug clinical efficacy prediction model and drug cost, the comprehensive recommendation score of the drug for the disease is evaluated, and the recommended ranking result of the drug corresponding to the disease is output according to the comprehensive recommendation score. This includes: determining the drugs whose clinical efficacy probability for the target disease meets the target value according to the drug clinical efficacy prediction model; screening all drugs whose efficacy probability is greater than or equal to the target value by preset a clinical efficacy probability target value (usually set to 0.7 to 0.8); excluding drugs whose efficacy does not meet the target value to narrow the scope of subsequent calculations; and matching the latest purchase unit price, unit dose cost, annual purchase volume and other cost data of the screened drugs from the hospital drug procurement system and the National Drug Centralized Procurement Platform database. At the same time, matching the standardized illustrative text of the drug's generic name, trade name, dosage form, specification, indications, contraindications, adverse reactions, usage and dosage from the database of the National Medical Products Administration.

[0033] All the above data are linked and integrated according to the drug's unique identifier to generate a structured drug and medical dataset. All textual data in the drug and medical dataset (including indications, contraindications, adverse reactions, usage and dosage in drug instructions) are input into a pre-trained bidirectional encoder representation model (BERT). This model adopts a multi-layer bidirectional Transformer architecture, which can simultaneously capture the contextual semantic information of each word in the text. The input text is preprocessed by word segmentation, adding special tags, truncation and padding, etc., to convert the text into an input format acceptable to the model. Through the forward propagation process of the BERT model, a 768-dimensional or 024-dimensional word vector corresponding to each word is output. All word vectors are arranged in order to form a drug text word embedding matrix. At the same time, the self-attention mechanism built into the BERT model is used to calculate the attention weight between each word in the word embedding matrix and all other words. The calculated attention weight can quantify the semantic association strength between different text fragments. For example, it can capture the causal semantic association between the contraindication of the drug for patients with renal insufficiency and the fact that the drug is mainly excreted through the kidneys, thereby mining the deep semantic relationships hidden in the text.

[0034] The attention weight matrix and word embedding matrix output by the self-attention mechanism are weighted and summed to obtain the global semantic feature vector for each drug. Then, an average pooling layer is used to reduce the dimensionality of the global semantic feature vector, generating a uniformly dimensional efficacy semantic feature vector. This vector condenses all semantic information related to efficacy and safety in the drug text. Simultaneously, the efficacy semantic feature vector is input into a pre-trained fully connected classification network. This network uses historical clinical evaluation data as its training set and is trained using the cross-entropy loss function. Its output layer has multiple neurons, each corresponding to a preset evaluation index, including clinical effectiveness rate, etc. The classification network outputs probability scores for each evaluation indicator, including the incidence of adverse reactions, the incidence of serious adverse reactions, the risk of drug interactions, patient compliance, and ease of administration. These scores are then arranged sequentially to form an evaluation indicator vector. Based on this vector and drug specifications, a comprehensive efficacy evaluation feature vector is defined. The rate of change in efficacy cost within a continuous step is analyzed in conjunction with cost control effectiveness, and a basic comprehensive recommendation score for the drug is calculated based on this rate of change. A simulation algorithm is used to sample the basic comprehensive recommendation scores to analyze the matching degree between users and drugs, calculate the comprehensive recommendation score, and then rank the results to obtain the drug recommendation ranking.

[0035] In this optional embodiment, a comprehensive efficacy evaluation feature vector is defined based on the evaluation index vector and drug specifications. Combined with cost control effect analysis, the rate of change in efficacy cost within a continuous step is analyzed. The calculation of the drug's basic comprehensive recommendation score based on the rate of change includes: extracting dosage form and specification data from the drug medical dataset; converting discrete dosage form and specification data using one-hot encoding; normalizing continuous specification data and mapping it to the [0,1] interval; and concatenating all encoded and normalized results to form a basic drug feature vector. This vector represents the drug's physical form and dosage characteristics. Simultaneously, the generated evaluation index vector and the basic drug feature vector are concatenated dimensionally to obtain a comprehensive efficacy evaluation feature vector with dimensions equal to the number of evaluation indicators plus the number of basic drug features. This vector simultaneously contains multiple aspects of the drug's efficacy, safety, applicability, and physical form. Furthermore, by integrating multi-source heterogeneous feature data, a unified comprehensive evaluation feature representation is formed.

[0036] Historical drug recommendation data from the past 3-5 years is extracted from the hospital information system, including actual clinical efficacy data and actual cost control effect data for each drug. A state space is constructed using this historical data, where each state consists of two dimensions: actual efficacy level and actual cost control effect level. Each state corresponds to a decision value, representing the optimal weight allocation between efficacy and cost in that state. All states and their corresponding decision values ​​constitute a decision state matrix. The comprehensive efficacy evaluation feature vector is then input into a reinforcement learning model. This model employs a Deep Q-Network (DQN) architecture, using the decision state matrix as the environment, the comprehensive efficacy evaluation feature vector as the state input, the weight coefficient adjustment amount as the action output, and the overall satisfaction of historical recommendation results as the reward function. By continuously interacting with the environment, the model optimizes the weight update strategy, gradually adjusting the weight coefficients of efficacy and cost, generating a dynamic weight coefficient set for efficacy and cost. This allows the model to automatically learn the optimal weight allocation for efficacy and cost based on historical data.

[0037] The rate of change of the dynamic weight coefficient group of efficacy and cost is calculated within N consecutive iteration steps (usually set to 10-20 steps). The rate of change is calculated using Euclidean distance and compared with a preset convergence threshold (usually set to 1e-5). If the rate of change is less than the convergence threshold, the weight coefficients are determined to have converged, a weight convergence detection result is generated, and the weight update process stops. If the rate of change is greater than or equal to the convergence threshold, the next round of iteration update continues until the weights converge. After the weights converge, the final efficacy weight coefficient α and cost weight coefficient β are extracted, where α+β=1. The efficacy prediction score and cost score of the drug are calculated respectively. The efficacy prediction score is obtained by normalizing the efficacy probability value output by the drug clinical efficacy prediction model, and the cost score is obtained by reciprocally normalizing the drug procurement cost (i.e., the lower the cost, the higher the cost score). Finally, the basic comprehensive recommendation score of each drug is calculated according to the formula: Basic Comprehensive Recommendation Score of Drug = α × Efficacy Prediction Score + β × Cost Score.

[0038] In this optional embodiment, a simulation algorithm is used to sample the basic comprehensive recommendation score of drugs to analyze the matching degree between users and drugs, calculate the comprehensive recommendation score of drugs, and sort them in order to obtain the recommendation ranking results of drugs. The basic comprehensive recommendation score of drugs and drug price fluctuation risk are used as two random variables. The drug price fluctuation risk is characterized by analyzing the historical drug price data of the past 1-2 years, calculating the price volatility, and constructing a log-normal distribution probability model. These two random variables are input into a Monte Carlo simulation algorithm, and the number of simulations is set to 100 to 1000. Each simulation randomly selects a sample value from the probability distribution of the two random variables to obtain a set of risk variable samples. All samples are combined to form a risk variable sample set, and a batch processing engine (such as Apache Flink) is used to perform parallel streaming computation on the risk variable sample set. The batch processing engine can process batch historical data and real-time incremental data at the same time. It will perform comprehensive calculations on each sample and statistically analyze the distribution, expected value, variance, and 95% confidence interval of the basic comprehensive recommendation score of drugs under different price fluctuation conditions.

[0039] Simultaneously, the drug application profile is dynamically updated based on these statistical indicators. This profile includes information such as the drug's average comprehensive score, score fluctuation range, price sensitivity, applicable population characteristics, and clinical application risk level. Furthermore, by introducing uncertainty analysis, the impact of drug price fluctuations on recommendation results is quantified, and the comprehensive application characteristics of the drug are dynamically updated. Additionally, the system obtains the current user's personal characteristic data from the electronic medical record system, including age, gender, weight, liver and kidney function indicators, comorbidities, concomitant medications, and drug allergy history. This user personal characteristic data is then matched and calculated with the applicable population, contraindications, and drug interaction risks information in the dynamically updated drug application profile. The cosine similarity algorithm is used to calculate the similarity between the user's feature vector and the feature vector of the target population for the drug, serving as a real-time matching score between the user and the drug. A higher matching score indicates that the drug is more suitable for the user. A base score weight γ and a matching score weight δ are set, where γ + δ = 1. These weights can be adjusted according to clinical needs, typically γ = 0.6-0.7 and δ = 0.3-0.4. The final comprehensive recommendation score for each drug is calculated using the formula: Drug Comprehensive Recommendation Score = γ × Basic Comprehensive Recommendation Score + δ × Real-time Matching Score. Finally, all drugs are sorted in descending order of their comprehensive recommendation scores, generating and outputting the drug recommendation ranking results.

[0040] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows: Step 1: Data Collection and Preprocessing; To support DEEP in predicting drug clinical efficacy, a hierarchical phenotypic dataset was constructed, which includes drug- and disease-induced transcriptional micro-changes (CTP) as well as macro-level clinical trial results.

[0041] Drug CTPs: These are derived from the LINCS L1000 bioinformatics dataset, containing 978 marker genes and over 20,000 drugs. Drugs with ≥5 replicates are retained. The average of all CTPs for each drug is used to obtain the drug CTP vector. .

[0042] Disease CTP: The differential expression feature library CREEDS (based on gene expression data of disease clinical samples and normal tissues in GEO) was extracted through crowdsourcing. The intersection of CREEDS genes and L1000 marker genes was taken to obtain 970 common genes as the gene set of disease CTP, resulting in the disease CTP vector. .

[0043] Clinical Results: Results are expressed as disease-drug pairs, derived from the TTD (Therapeutic Target Database). All approved disease-drug pairs or those in Phase I / II / III clinical trials were extracted. After integrating the aforementioned CTP information, samples lacking CTPs were excluded, resulting in 9033 positive samples (labeled 1). For each disease, a proportional random sampling was performed from the remaining drugs, generating 9033 negative samples (labeled 0). This resulted in 18066 disease-drug pairs, involving 1155 drugs and 106 diseases, with a positive-to-negative sample ratio of 1:1 for each disease.

[0044] Step 2: Model Construction and Feature Representation; like Figure 4 As shown, the DEEP model consists of a disease CTP encoder. Drug CTP encoder and disease-drug relationship decoder All three components are multilayer perceptrons (MLPs).

[0045] End-to-end prediction of drug efficacy (DEEP). This framework integrates drug-induced transcriptomic changes, disease-related transcriptomic changes, and clinical trial results to train a multilayer feedforward neural network (MLP) to learn the intrinsic relationships between drug efficacy.

[0046] The end-to-end prediction model comprises: an AI drug encoder for drug encoding, an AI disease encoder for disease encoding, and an AI decoder for predicting disease-drug relationships. Transcriptome profile changes (CTPs) caused by disease and drug perturbations are fed into a multilayer perceptron (MLP) based deep learning model, which is trained using a global clinical trial results dataset to predict the therapeutic effects of drugs on the corresponding diseases.

[0047] In this embodiment, both the disease feature encoder and the drug feature encoder are constructed using three-layer linear transformation units. For the disease CTP feature encoding branch, a hierarchical feature dimensionality reduction mechanism is used to gradually map the 970-dimensional high-dimensional gene transcription original features to 128-dimensional and 48-dimensional features, and finally to compactly encode them into an 8-dimensional low-dimensional representation. Correspondingly, the drug CTP feature encoding branch adopts a symmetric dimensionality compression strategy to perform isomorphic feature extraction on the 978-dimensional input vector, and the hidden layer dimension setting is consistent with the former.

[0048] To effectively improve model convergence efficiency, enhance training robustness, and suppress gradient vanishing, each linear mapping unit is followed by a cascaded Sigmoid nonlinear activation module and a batch normalization layer. This achieves feature distribution standardization and enhanced nonlinear expressive power. Through the aforementioned dual-branch isomorphic coding structure, two types of heterogeneous input data are uniformly mapped to a shared latent feature space, outputting a compact embedding vector with a dimension of 8 (k=8). The decoding module employs a two-layer multilayer perceptron structure, achieving feature tensor fusion through an outer product mapping mechanism. It projects two sets of 8-dimensional latent features into a 64-dimensional joint feature representation, performs stepwise reconstruction mapping on the 64-dimensional fused deep features, and outputs them to the prediction node after a 12-dimensional latent layer nonlinear transformation (with simultaneous Sigmoid activation and batch normalization processing). At the end, a Sigmoid constraint unit is introduced to numerically normalize the prediction results, ensuring that the output drug clinical efficacy prediction values ​​converge stably within the [0,1] interval, thus improving the interpretability and practicality of the prediction results.

[0049] Disease CTP vector and drug CTP vector Input the corresponding encoders and compute the hidden state embeddings. , In this embodiment, the encoder has 3 layers, the hidden layer dimension k=8, batch normalization is used to accelerate convergence, and the activation function is Sigmoid, where: , .

[0050] Will and Feature tensor fusion is achieved through an outer product mapping mechanism, projecting two sets of 8-dimensional latent features into a 64-dimensional joint feature representation, resulting in... Flattened input to decoder (2 layers) Output the predicted efficacy probability of the drug for the disease. The details are as follows: ; in, That is, the embedded representation vector.

[0051] Step 3: Model training; The drug efficacy prediction task is defined as a supervised binary classification problem, using clinical trial outcomes as the true labels, and employing a binary cross-entropy loss function L, where... Let p be the true label and p be the predicted probability. The Adam optimizer is used with an initial learning rate of 0.001, as detailed below: ; To comprehensively evaluate the model's generalization ability, the following data partitioning strategy was adopted: Random partitioning: The 18066 disease-drug pairs were randomly partitioned into training and testing sets at a ratio of 9:1, and 10-fold cross-validation was performed. Figure 5 and Figure 6 As shown.

[0052] The training and test sets were divided in a 9:1 ratio, and 1 / 10 of the training set was randomly selected as the validation set. The model was trained for 100 epochs, and the model parameters with the minimum loss on the validation set were saved for evaluation on the test set.

[0053] Step 4: Calculate the evaluation indicators; The following metrics were used to evaluate model performance: AUROC represents the area under the receiver operating characteristic curve, reflecting the model's overall ability to distinguish between positive and negative samples; AUPR represents the area under the precision-recall curve, which focuses more on the precision and recall of positive samples and provides a more robust evaluation.

[0054] True positive rate at low false positive rate (TPR): TPR is calculated at thresholds such as FPR=0.01 or 0.05 to simulate the scenario in actual drug screening where only top-ranked candidates are considered.

[0055] Step 5: Knowledge-based drug screening is used for experimental evaluation; For the disease to be evaluated, a model obtained from 10-fold cross-validation was used to predict the efficacy of all drugs, excluding those known to be effective in the dataset. Drugs were then sorted in descending order of their average predicted probability, and the top 5% were selected as candidates.

[0056] Step 6: Model performance verification; In randomized 10-fold cross-validation, the area under the ROC curve (AUPR) and the area under the PR curve (AUROC) of the DEEP model reached 0.76 and 0.77, respectively, significantly outperforming the connectivity score. The Connectivity Score and the Reverse Gene Expression Score (RGES) (both close to 0.5, close to the random level) are used to measure the similarity between drug treatment and disease expression profile (negative scores indicate the opposite), while the Reverse Gene Expression Score (RGES) is specifically used to quantify the ability of a drug to reverse the disease expression profile back to normal (the lower the negative score, the stronger the effect). In the disease-blinded experiment, 10 diseases were randomly selected as the test set, with an average AUPR of 0.816, an AUROC of 0.798, and an average recall rate of more than 20 times higher than the baseline when FPR=1%. The average hit rate of the first 20 diseases reached 72%, as shown in Table 1. The drug-blinded and double-blind experiments also showed that the model still had significantly better predictive performance than the baseline when no drug or disease-drug combination was seen. Therefore, this embodiment can capture the potential association between drugs and diseases from CTP data, accurately predict drug efficacy, and can be extended to a variety of diseases and drugs, showing application potential.

[0057] Table 1. Results of 10-fold cross-validation for the three models on the training set.

[0058] Systemic sclerosis (SSc), commonly known as scleroderma, is a chronic autoimmune disease of unknown etiology. Its main pathological features include immune system dysfunction, microvascular damage, and progressive fibrosis of the skin and internal organs. The disease has a chronic, progressive course and is clinically classified into localized cutaneous SSc (lcSSc) and diffuse cutaneous SSc (dcSSc) based on the extent of skin involvement. It can further affect multiple internal organs, such as the lungs, heart, kidneys, and digestive tract. The etiology of SSc involves the interaction of multiple factors, including genetic susceptibility, environmental factors (such as viral infections and chemical exposure), and epigenetic modifications, which are not yet fully understood. Based on current technology, there are no drugs or therapies worldwide that can cure SSc. Existing treatment strategies mainly target symptom relief, suppression of the immune inflammatory response, and slowing the progression of fibrosis, but their efficacy is limited and exhibits significant individual differences. Therefore, the development of novel anti-fibrotic, immunomodulatory, or vascular protective agents is of significant clinical importance and urgently needed for the treatment of this disease. To address the aforementioned needs, this embodiment employs the DEEP model for drug prediction and screening, selecting testosterone undecanoate, dinogest, and citalopram as candidate drugs for the treatment of scleroderma.

[0059] To verify the efficacy of the above-mentioned candidate drugs, this embodiment constructed a bleomycin-induced SSc mouse model and evaluated its in vivo efficacy. Six- to eight-week-old female C57BL / 6 mice were used. A 1cm × 1cm area was exposed on the back, and injection points were established at the four corners of this area. Using rotating needles, bleomycin saline solution (1 mg / mL, 50 μL / time) was injected intradermally at each point once daily for 6 weeks to induce local scleroderma-like changes. Mice with successful model establishment were randomly divided into 5 groups: a normal group, a model control group, and 3 treatment groups, with 5 mice in each group. The treatment groups were administered testosterone undecanoate (15 mg / kg) or dinogest (1 mg / kg) via gavage (ig), or citalopram (10 mg / kg) via intraperitoneal injection (ip), once daily for 21 days. The model control group received an equal volume of solvent. After the treatment, skin tissue from the injection area was collected for hematoxylin-eosin (HE) staining and Masson staining. The results showed that... Figure 7 and Figure 8 As shown, compared with the model control group, skin thickness was improved. At the same time, the content of hydroxyproline (or collagen deposition) in the skin tissue of each drug administration group was significantly reduced, and collagen fiber deposition was reduced. Among them, the citalopram group showed a statistically significant difference in reducing mouse skin thickness.

[0060] The above results demonstrate that testosterone undecanoate, dinogest, and citalopram have a clear therapeutic effect on bleomycin-induced scleroderma in mice, and also verify the accuracy of the DEEP model in drug prediction.

[0061] Figure 2 An embodiment of a drug efficacy prediction system based on an end-to-end deep learning framework according to the present invention is shown.

[0062] In this optional embodiment, the drug efficacy prediction system based on an end-to-end deep learning framework includes: The end-to-end prediction model building module 201 is used to generate a hierarchical phenotypic dataset based on the micro-changes in the transcriptional profile of drugs and diseases, and to calculate the state embedding vector using the hierarchical phenotypic dataset in order to build an end-to-end prediction model for drug efficacy prediction. The end-to-end prediction model optimization module 202 is used to train the end-to-end prediction model using the binary cross-entropy loss function, evaluate the performance of the trained end-to-end prediction model based on the working characteristic curve, and output the drug clinical efficacy prediction model after the performance results meet the requirements. The drug recommendation generation and ranking module 203 is used to evaluate the comprehensive recommendation score of drugs for diseases based on the drug clinical efficacy prediction model and drug cost, and output the recommendation ranking result of the drugs corresponding to the diseases according to the comprehensive recommendation score.

[0063] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0064] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0065] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0066] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0067] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0068] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A drug efficacy prediction method based on an end-to-end deep learning framework, characterized in that, The method includes: Based on the microscopic changes in the transcriptional profiles of drugs and diseases, a hierarchical phenotypic dataset is generated, and state embedding vectors are calculated using the hierarchical phenotypic dataset to construct an end-to-end prediction model for drug efficacy prediction. The end-to-end prediction model is trained using the binary cross-entropy loss function, and the performance of the trained end-to-end prediction model is evaluated based on the working characteristic curve. Once the performance results meet the requirements, the drug clinical efficacy prediction model is output. Based on the drug clinical efficacy prediction model and drug cost, the comprehensive recommendation score of drugs for diseases is evaluated, and the recommendation ranking results of drugs corresponding to diseases are output according to the comprehensive recommendation score.

2. The drug efficacy prediction method based on an end-to-end deep learning framework according to claim 1, characterized in that, The process of generating a hierarchical phenotypic dataset based on microscopic changes in the transcriptional profiles of drugs and diseases, and using this dataset to calculate state embedding vectors to construct an end-to-end prediction model for drug efficacy prediction includes: A drug judgment set containing drugs and marker genes is obtained based on a bioinformatics dataset, and the average of the micro-changes in the transcriptional profile of the drugs is taken. The micro-change vector of the drug transcriptional profile is obtained based on the average result. Gene expression was selected from the differential expression feature library, and common genes were extracted based on the intersection of gene expression and marker genes to represent the micro-changes in the transcriptional spectrum of the disease, so as to obtain the vector of micro-changes in the transcriptional spectrum of the disease. The micro-change vectors of drug transcription profiles and disease transcription profiles are integrated, and disease-drug pairs are generated as a hierarchical phenotypic dataset by proportionally sampling from the integration results and the therapeutic target database. Extract the hidden state embedding vectors corresponding to the hierarchical phenotypic dataset, and after projecting the hidden state embedding vectors, output an end-to-end prediction model to describe the probability of predicting the efficacy of drugs for diseases.

3. The drug efficacy prediction method based on an end-to-end deep learning framework according to claim 2, characterized in that, The step of extracting the hidden state embedding vectors corresponding to the hierarchical phenotypic dataset, and then projecting the hidden state embedding vectors to output an end-to-end prediction model that describes the probability of predicting the efficacy of a drug for a disease, includes: A drug encoder for drug encoding, a disease encoder for disease encoding, and a decoder for predicting disease-drug relationships are constructed based on a hierarchical phenotypic dataset. By using a drug encoder and a disease encoder, the micro-change vectors of drug transcription profiles and disease transcription profiles are progressively reduced to the target dimension. The dimensionality reduction result is mapped to the shared latent feature space to obtain the drug latent state embedding vector and the disease latent state embedding vector. The feature tensor is then fused with the drug latent state embedding vector and the disease latent state embedding vector through the outer product mapping mechanism to obtain the embedding representation vector. After flattening the embedded representation vector, it is input into the decoder that predicts the relationship between the disease and the drug to construct an end-to-end prediction model. The end-to-end prediction model is then used to output the predicted efficacy probability of the drug for the disease.

4. The drug efficacy prediction method based on an end-to-end deep learning framework according to claim 3, characterized in that, Both the drug encoder and the disease encoder are constructed using a three-layer linear transformation technique, and the decoder is constructed using a two-layer perceptron. The drug encoder employs a hierarchical feature dimensionality reduction mechanism to progressively map the micro-change vectors of the drug transcription spectrum to the target dimension; the disease encoder employs a symmetrical dimensionality compression mechanism to progressively map the micro-change vectors of the disease transcription spectrum to the target dimension.

5. The drug efficacy prediction method based on an end-to-end deep learning framework according to claim 1, characterized in that, The process of training the end-to-end prediction model using the binary cross-entropy loss function and evaluating the performance of the trained end-to-end prediction model based on the operating characteristic curve, followed by outputting the drug clinical efficacy prediction model after the performance results meet the requirements, includes: The hierarchical phenotypic dataset is divided into a training set and a test set, and ten-fold cross-validation is performed on the training set to divide the training set into a training subset and a validation subset. Set the total number of training rounds, perform iterative training of the end-to-end prediction model on the training subset, and calculate the loss value of the validation subset using the binary cross-entropy loss function after the training rounds are completed. Select the model parameters corresponding to the minimum loss value as the optimal parameters of the end-to-end prediction model to complete the training optimization. The target drug is predicted using the trained and optimized end-to-end prediction model to determine the predicted efficacy probability of the target drug for the disease, and the area under the working feature curve and the area under the precision and recall curve are plotted based on the predicted efficacy probability. The area under the working feature curve and the area under the precision and recall curve were compared with the connectivity score and the reverse gene expression score, respectively, to verify the performance of the training and optimization of the end-to-end prediction model. Based on the optimization results, the end-to-end prediction model was further optimized until the performance met the requirements, and the drug clinical efficacy prediction model was output.

6. The drug efficacy prediction method based on an end-to-end deep learning framework according to claim 5, characterized in that, The expression for the binary cross-entropy loss function is: ； ； In the formula, L Indicates the loss value. y Indicates the true label, p This indicates the probability of a drug's predicted therapeutic effect on a disease. A decoder representing the predictive drug-disease relationship. H This represents the embedded representation vector.

7. The drug efficacy prediction method based on an end-to-end deep learning framework according to claim 1, characterized in that, The process of evaluating a drug's comprehensive recommendation score for a disease based on a drug clinical efficacy prediction model and drug cost, and outputting a ranking of recommended drugs for the corresponding disease based on the comprehensive recommendation score, includes: Based on the drug clinical efficacy prediction model, identify drugs whose clinical efficacy probability for the target disease meets the target value, and match the drug procurement cost with illustrations to generate a drug medical dataset. The drug and medical dataset is input into a bidirectional encoder representation model for word vector transformation to obtain a drug text word embedding matrix. The semantic association of the drug text word embedding matrix is ​​then analyzed using a self-attention mechanism. Based on semantic association, determine the semantic feature vector of drug efficacy, evaluate and classify the semantic feature vector of efficacy, and determine the evaluation index vector. Based on the evaluation index vector and drug specification definition, the efficacy comprehensive evaluation feature vector is defined, and the change rate of efficacy cost within a continuous step is analyzed in combination with the cost control effect. The basic comprehensive recommendation score of the drug is calculated based on the change rate. A simulation algorithm is used to sample the basic comprehensive recommendation scores of drugs to analyze the matching degree between users and drugs, calculate the comprehensive recommendation scores of drugs, and sort them in order to obtain the recommendation ranking results of drugs.

8. The drug efficacy prediction method based on an end-to-end deep learning framework according to claim 7, characterized in that, The efficacy comprehensive evaluation feature vector is defined based on the evaluation index vector and drug specifications, and combined with the cost control effect analysis to analyze the rate of change of efficacy cost within a continuous step. The basic comprehensive recommendation score of the drug is calculated based on the rate of change, including: The dosage form and specification data of the drug are uniquely encoded. Based on the encoding results, the basic feature vector of the drug is determined. The evaluation index vector is then concatenated with the basic feature vector of the drug to obtain the comprehensive efficacy evaluation feature vector. A state space is constructed based on the actual efficacy and cost control effect of historically recommended drugs, a decision state matrix is ​​generated, and the weight update strategy is optimized by combining the efficacy comprehensive evaluation feature vector to obtain a dynamic weight coefficient group of efficacy and cost. Calculate the rate of change of the dynamic weight coefficient group of efficacy cost within a continuous step size, compare the rate of change with a preset convergence threshold, determine the weight convergence when the rate of change is less than the threshold, and generate the weight convergence detection result. The weight coefficients are determined based on the weight convergence test results, and the efficacy prediction score and cost score are weighted and fused using the weight coefficients to output a comprehensive drug recommendation score.

9. The drug efficacy prediction method based on an end-to-end deep learning framework according to claim 8, characterized in that, The process involves sampling the basic comprehensive recommendation scores of drugs using a simulation algorithm to analyze the matching degree between users and drugs, calculating the comprehensive recommendation scores, and arranging them in order to obtain the drug recommendation ranking results, including: The basic comprehensive recommendation score of drugs and the risk of drug price fluctuations are used as random variables and input into the Monte Carlo simulation algorithm for random sampling to obtain a risk variable sample set. The batch processing engine is then used to perform streaming calculations on the risk variable sample set to dynamically update the drug application profile. The matching degree between users and drugs is analyzed based on the updated drug application profile results, which serves as the real-time matching score. The real-time matching score is then weighted and integrated with the basic comprehensive drug recommendation score to obtain the comprehensive drug recommendation score. The drug recommendation ranking is determined based on the descending order of the comprehensive drug recommendation score.

10. A drug efficacy prediction system based on an end-to-end deep learning framework, characterized in that, include: The end-to-end prediction model building module is used to generate a hierarchical phenotypic dataset based on the micro-changes in the transcriptional profile of drugs and diseases, and to use the hierarchical phenotypic dataset to calculate the state embedding vector in order to build an end-to-end prediction model for drug efficacy prediction. The end-to-end prediction model optimization module is used to train the end-to-end prediction model using the binary cross-entropy loss function, evaluate the performance of the trained end-to-end prediction model based on the working characteristic curve, and output the drug clinical efficacy prediction model after the performance results meet the requirements. The drug recommendation generation and ranking module is used to evaluate the comprehensive recommendation score of drugs for diseases based on drug clinical efficacy prediction models and drug costs, and output the recommendation ranking results of drugs corresponding to diseases based on the comprehensive recommendation score.

Images