A drug response prediction method based on residual graph neural network

Abstract

This invention discloses a drug response prediction method based on residual graph neural networks (ResGNN), comprising: 1) constructing multi-source heterogeneous data and molecular graph features, extracting and preprocessing multi-source heterogeneous raw data of drugs and cell lines from a large-scale database; 2) constructing a dual-path drug feature encoder based on residual connections to extract dynamic drug characterizations that take into account both general chemical prior knowledge and specific tumor microenvironment task specificity; 3) employing a progressive hot-start training strategy to achieve the transition from baseline prior knowledge to specific fine-tuning; 4) constructing a cell line feature encoder to extract deep and robust cell line molecular characterizations from high-dimensional and redundant gene network data; and 5) constructing a registration generative adversarial regularization prediction framework, utilizing the game mechanism of generative adversarial networks to achieve drug efficacy prediction. This invention utilizes ResGNN to extract molecular features and combines a hot-start strategy with a generative adversarial network to predict the half-maximum inhibitory concentration (IC50) of anticancer drugs.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and artificial intelligence, and in particular to a method for predicting drug response based on residual graph neural networks. Background Technology

[0002] With the development of precision medicine, predicting a patient's response to a specific drug based on their genomic characteristics (i.e., drug sensitivity prediction) has become a core aspect of cancer treatment. IC50 is a key quantitative indicator for measuring drug efficacy. However, determining the combination of numerous drugs with cell lines through in vitro cell line experiments is not only costly but also time-consuming. Therefore, computational prediction models based on deep learning (DL) have become a research hotspot.

[0003] Although existing computational methods have improved prediction accuracy to some extent, the following three core technical bottlenecks still exist when dealing with complex drug-cell line interaction tasks:

[0004] 1) The "Static Rigidity" Problem in Drug Feature Representation. Current mainstream methods such as DeepDR and GraphDRP typically employ a "pre-training + frozen" approach to drug molecules. This involves using graph neural networks (GNNs) or graph convolutional networks (GCNs) pre-trained on large-scale chemical databases to extract drug features, and then maintaining these feature parameters unchanged in subsequent IC50 prediction tasks. While this static feature approach preserves general chemical structural information, it ignores task-specific characteristics within the specific cancer context. For example, certain functional groups or stereoconformities of some drugs play a crucial role in lung cancer cells but may be unimportant in breast cancer cells. Static features cannot be fine-tuned through backpropagation, making it difficult for the model to capture these subtle structural features that determine drug efficacy.

[0005] 2) Cold-start instability in end-to-end GNN training. To address the aforementioned feature rigidity problem, some studies have attempted to directly introduce trainable GNN modules into the prediction model. However, due to the complex topological structure of drug molecule maps and the relatively limited amount of labeled data, randomly initialized GNNs often output high-noise features in the early stages of training. This noise disrupts the model's existing feature distribution, making the model highly susceptible to getting trapped in local optima, and even causing gradient vanishing or exploding, resulting in training collapse and severely impacting convergence speed and final accuracy.

[0006] 3) Insufficient distribution fitting ability and poor generalization of regression models. Traditional drug response prediction models mostly treat it as a simple regression problem, using only mean squared error (MSE) as the loss function. MSE tends to predict the mean of the samples to minimize the overall error, which results in a distribution of the predicted results that is often "narrower" than the true distribution, making it difficult to capture extremely sensitive or extremely resistant samples. In addition, this deterministic regression approach ignores the underlying distribution patterns of the data, making the model unable to resist batch effects when faced with independent external datasets, resulting in a significant decrease in generalization ability. Although GANs introduced in existing research can alleviate the distribution problem, they are usually difficult to effectively combine with regression tasks, and the training process is extremely unstable. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a drug response prediction method based on residual graph neural networks. It solves the feature rigidity problem by constructing a dual-path residual drug encoder, uses a hot-start strategy to solve the training instability of graph neural networks (GNNs), and introduces generative adversarial networks (GANs) as regularization terms to solve the problem of poor generalization ability.

[0008] The objective of this invention is achieved as follows: a drug response prediction method based on residual graph neural networks, comprising the following steps:

[0009] 1) Construct multi-source heterogeneous data and molecular graph features, extract and preprocess multi-source heterogeneous raw data of drugs and cell lines from large-scale databases;

[0010] 2) Construct a dual-path drug feature encoder based on residual connections to extract dynamic drug characterizations that take into account both general chemical prior knowledge and specific tumor microenvironment task specificity;

[0011] 3) A progressive hot-start training strategy is adopted to achieve the transition from baseline prior knowledge to specific fine-tuning;

[0012] 4) Construct a cell line feature encoder to extract deep and robust cell line molecular characterizations from high-dimensional and redundant gene network data.

[0013] 5) Construct a registration-based generative adversarial regularization prediction framework and utilize the game mechanism of generative adversarial networks to achieve drug efficacy prediction.

[0014] Furthermore, step 1) specifically includes: constructing a gene co-expression network reflecting dynamic regulatory relationships by calculating the correlation between genes, and converting a one-dimensional drug sequence into a two-dimensional topological graph structure containing rich physicochemical properties; obtaining a large-scale cancer drug sensitivity database, including a cell line gene expression profile data matrix. And the set of simplified molecular linear input specifications (SMILES) for drugs;

[0015] 1-1) Network Construction: Calculate the expression of any two genes in the gene expression data. and The Pearson correlation coefficient (PCC) between the two pairs is used to construct the co-expression adjacency matrix. :

[0016]

[0017] 1-2) Drug Molecular Map Construction: Using cheminformatics tools to convert drug samples into molecular maps. ;in For a set of atomic nodes, For the set of chemical bond edges; construct the adjacency matrix of drug molecules. and atomic feature matrix .

[0018] Furthermore, step 2) specifically includes: by designing a dual-path parallel graph neural network architecture, combining the frozen main path for extracting general features with the dynamic branch path for specific fine-tuning, and performing adaptive residual fusion;

[0019] 2-1) Static main path: Extract general features using the pre-trained model MolGNet; parameters of this branch Freeze during training;

[0020]

[0021] in, This represents the general feature matrix of drug molecules extracted from static branches. This represents the input drug molecule diagram. This represents the frozen network weight parameters in the pre-trained MolGNet model;

[0022] 2-2) Dynamic Residual Branch: Construct an L-layer Graph Convolutional Network (GCN). For the l-th layer, its feature propagation formula is defined as:

[0023]

[0024] in, To add a self-loop adjacency matrix, for The degree matrix, For the first The trainable weight matrix of the layer, The function is a linear rectified function ReLU; after propagation through L layers, the residual feature vector is obtained through global max pooling. :

[0025]

[0026] 2-3) Adaptive residual fusion: This involves weighted fusion of static and dynamic features to obtain the final drug characterization. :

[0027]

[0028] in, This indicates vector concatenation or element-wise addition. This is the hot start adjustment coefficient.

[0029] Furthermore, step 3) specifically includes: introducing a time-dependent adjustment function based on training epochs to achieve the transition from baseline prior knowledge to specific fine-tuning; designing the time-dependent adjustment function. To control the intervention intensity of residual branches, Defined as a piecewise function with respect to training epochs t:

[0030]

[0031] in, Here, k represents the preset freeze period, and k is the growth slope. The center point of the sigmoid function is used; this strategy ensures that only the center point is used in the early stages of training. Benchmark training is performed, followed by a smooth introduction of residual feature vectors. Make minor adjustments.

[0032] Furthermore, step 4) specifically includes: injecting noise into the original gene expression vector to damage it, and using an autoencoder mechanism to minimize the reconstruction error, so that the model learns low-dimensional manifold features with strong generalization ability; and using a denoising autoencoder (DAE) to extract low-dimensional manifold features of gene expression.

[0033] 4-1) Noise injection: injecting noise into the original gene expression vector Add Gaussian noise Damaged input ;

[0034] 4-2) Encoding and Decoding:

[0035]

[0036]

[0037] Where z represents the feature vector of the cell line extracted by the encoder. This represents the reconstructed gene expression vector output by the decoder. Represents a non-linear activation function. and Let these represent the trainable weight matrices of the encoder network and the decoder network, respectively. and These represent the bias vectors of the encoder network and the decoder network, respectively;

[0038] 4-3) Pre-training loss: Optimize the DAE parameters by minimizing the reconstruction error:

[0039]

[0040] Finally, the encoder output z is taken as the cell line feature. .

[0041] Furthermore, step 5) specifically includes: constructing a minimax game framework between the generator and the discriminator, using the distribution consistency constraint generated by the adversarial mechanism as a strong regularization term, and combining it with traditional regression error for joint optimization to achieve drug efficacy prediction; using a generative adversarial network (GAN) for regression prediction, the model includes a generator G, a discriminator D, and an objective function;

[0042] Generator G: Input is drug features cell line characteristics The concatenation of these values ​​outputs the predicted drug sensitivity values. :

[0043]

[0044] Discriminator D: Input is the actual drug sensitivity value y or the predicted drug sensitivity value Output the probability that the sample belongs to the true distribution. ;

[0045] Objective function: The model's total loss function From regression loss and combat losses Composition and overall optimization objective:

[0046] Regression loss: Mean squared error is used to ensure the accuracy of the predicted values;

[0047]

[0048] Adversarial Loss: Employing either Least Squares Generative Adversarial Network (LSGAN) loss or standard Generative Adversarial Network (GAN) loss, the distribution of predicted values ​​is optimized. Distribution that approximates the true value ;

[0049]

[0050] Overall optimization objective: The generator G and the discriminator D engage in a minimax game.

[0051]

[0052] Here, γ is the balance coefficient, used to adjust the strength of the resistance to regularization.

[0053] Compared with existing technologies, the beneficial effects of this invention are as follows: 1) This invention introduces a registration-based generative adversarial RegGAN as an inference engine, using a discriminator to adversarially constrain the drug efficacy distribution output by the generator. This distribution consistency constraint acts as a powerful regularization term, forcing the model to learn the intrinsic manifold distribution of real biological data rather than rote memorization of a small number of samples. Experiments show that even in extremely sparse scenarios using only 10% of the training data, this model can still maintain a high prediction PCC, effectively solving the generalization problem under small sample conditions. It significantly improves the robustness and prediction accuracy of the model in sparse data scenarios, solving the problem that existing deep learning models heavily rely on massive paired data and are prone to overfitting when high-quality clinical samples are scarce.

[0054] 2) To address the challenge of training new molecular entities or rare subtype targets lacking prior interaction data, this invention designs a dual-path ResGNN encoder with a "hot start" strategy. By providing a general chemical knowledge base through a pre-trained static branch that relies on frozen parameters in the early stages of training, and by smoothly and gradually introducing task-specific dynamic branches using dynamic coefficients, the model achieves stable initialization and rapid convergence even without sufficient historical data. This successfully realizes an effective "cold start" under zero-sample or few-sample conditions, effectively overcoming the technical bottleneck of "cold start" for new drugs and targets.

[0055] 3) The ResGNN dynamic branching of this invention can be fine-tuned on specific breast cancer task data to capture conformational changes or specific target affinity characteristics that may occur after drug molecules enter a specific tumor microenvironment. This task-driven dynamic characterization method significantly improves the biological relevance of features, thereby enhancing the accuracy of the final drug efficacy prediction. It overcomes the limitations of traditional methods that treat drugs as static, invariant graph structures, and achieves dynamic and precise characterization of drug molecules in a specific tumor microenvironment.

[0056] 4) This invention abandons the traditional approach of using only single gene expression levels, innovatively constructing a gene co-expression network to capture the dynamic regulatory relationships between genes, and utilizing DAE for deep feature extraction. By deeply fusing this patient omics representation containing rich topological information with drug graph representation in a multimodal manner, the model can more comprehensively understand the complex interaction mechanisms between "drug-gene-disease". This enables the model to not only perform well on internal datasets, but also demonstrate stronger batch effect resistance and generalization performance in cross-independent dataset validation. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0058] Figure 1 This is a diagram of the overall architecture of the ResGNN model in the example.

[0059] Figure 2 The diagram shows the core innovative mechanism and verification results of this invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] A drug response prediction method based on residual graph neural networks includes the following steps:

[0062] 1) Construct multi-source heterogeneous data and molecular graph features, extract and preprocess multi-source heterogeneous raw data of drugs and cell lines from large-scale databases;

[0063] Gene co-expression networks reflecting dynamic regulatory relationships were constructed by calculating inter-gene correlations, and one-dimensional drug sequences were transformed into two-dimensional topological graph structures containing rich physicochemical properties; a large-scale cancer drug sensitivity database was obtained, including cell line gene expression profile data matrices. And the set of molecular formulas of the Simplified Molecular Input Line Entry System (SMILES);

[0064] 1-1) Network Construction: Calculate the expression of any two genes in the gene expression data. and The Pearson Correlation Coefficient (PCC) between them is used to construct the co-expression adjacency matrix. :

[0065]

[0066] 1-2) Drug Molecular Map Construction: Using cheminformatics tools to convert drug samples into molecular maps. ;in For a set of atomic nodes, For the set of chemical bond edges; construct the adjacency matrix of drug molecules. and atomic feature matrix .

[0067] 2) Construct a dual-path (ResGNN) drug feature encoder based on residual connections to extract dynamic drug characterizations that take into account both general chemical prior knowledge and specific tumor microenvironment task specificity;

[0068] To extract dynamic drug characterizations that balance general chemical prior knowledge with the specificity of the tumor microenvironment, an innovative dual-parallel network architecture is designed. This architecture combines a frozen main path for extracting general features with a dynamic branch for specific fine-tuning, and performs adaptive residual fusion. To address the issue of the inability to fine-tune pre-trained features, a dual-parallel graph neural network architecture is designed.

[0069] 2-1) Static main path: Extract general features using the pre-trained model MolGNet; parameters of this branch Freeze during training;

[0070]

[0071] in, This represents the general feature matrix of drug molecules extracted from static branches. This represents the input drug molecule diagram. This represents the frozen network weight parameters in the pre-trained MolGNet model;

[0072] 2-2) Dynamic Residual Branch: Construct an L-layer Graph Convolutional Network (GCN). For the l-th layer, its feature propagation formula is defined as:

[0073]

[0074] in, To add a self-loop adjacency matrix, for The degree matrix, For the first The trainable weight matrix of the layer, The function is a Rectified Linear Unit (ReLU); after propagation through L layers, the residual feature vector is obtained through global max pooling. :

[0075]

[0076] 2-3) Adaptive residual fusion: This involves weighted fusion of static and dynamic features to obtain the final drug characterization. :

[0077]

[0078] in, This indicates vector concatenation or element-wise addition. This is the hot start adjustment coefficient.

[0079] 3) A progressive hot-start training strategy is adopted to achieve the transition from baseline prior knowledge to specific fine-tuning;

[0080] By introducing a time-dependent adjustment function based on training epochs, the model's learning ability of the dynamic residual branch is smoothly and progressively released, achieving a safe transition from baseline prior knowledge to specific fine-tuning. To address the gradient noise problem caused by the cold start of GNNs, a time-dependent adjustment function is designed. To control the intervention intensity of residual branches, Defined as a piecewise function with respect to training epochs t:

[0081]

[0082] in, Here, k represents the preset freeze period, and k is the growth slope. The center point of the sigmoid function is used; this strategy ensures that only the center point is used in the early stages of training. Benchmark training is performed, followed by a smooth introduction of residual feature vectors. Make minor adjustments.

[0083] 4) Construct a cell line feature encoder to extract deep and robust cell line molecular characterizations from high-dimensional and redundant gene network data.

[0084] Noise is injected into the original gene expression vector to disrupt it, and the reconstruction error is minimized by an autoencoder mechanism, enabling the model to learn low-dimensional manifold features with strong generalization ability; low-dimensional manifold features of gene expression are extracted using a denoising autoencoder (DAE).

[0085] 4-1) Noise injection: injecting noise into the original gene expression vector Add Gaussian noise Damaged input ;

[0086] 4-2) Encoding and Decoding:

[0087]

[0088]

[0089] Where z represents the feature vector of the cell line extracted by the encoder. This represents the reconstructed gene expression vector output by the decoder. Represents a non-linear activation function. and Let these represent the trainable weight matrices of the encoder network and the decoder network, respectively. and These represent the bias vectors of the encoder network and the decoder network, respectively;

[0090] 4-3) Pre-training loss: Optimize the DAE parameters by minimizing the reconstruction error:

[0091]

[0092] Finally, the encoder output z is taken as the cell line feature. .

[0093] 5) Construct a registration generative adversarial (RegGAN) regularized prediction framework and use the game mechanism of generative adversarial networks to achieve drug efficacy prediction; solve the problems of model overfitting and insufficient distribution fitting ability under sparse sample conditions.

[0094] By constructing a minimax game framework between the generator and the discriminator, the distribution consistency constraint generated by the adversarial mechanism is used as a strong regularization term and combined with the traditional regression error for joint optimization to achieve drug efficacy prediction; regression prediction is performed using a generative adversarial network (GAN), and the model includes a generator G, a discriminator D, and an objective function.

[0095] Generator G: Input is drug features cell line characteristics The concatenation of these values ​​outputs the predicted drug sensitivity values. :

[0096]

[0097] Discriminator D: Input is the actual drug sensitivity value y or the predicted drug sensitivity value Output the probability that the sample belongs to the true distribution. ;

[0098] Objective function: The model's total loss function From regression loss and combat losses Composition and overall optimization objective:

[0099] Regression loss: Mean squared error is used to ensure the accuracy of the predicted values;

[0100]

[0101] Adversarial Loss: Employing either Least Squares Generative Adversarial Network (LSGAN) or standard Generative Adversarial Network (GAN) loss, the distribution of predicted values ​​is optimized. Distribution that approximates the true value ;

[0102]

[0103] Overall optimization objective: The generator G and the discriminator D engage in a minimax game.

[0104]

[0105] Here, γ is the balance coefficient, used to adjust the strength of the resistance to regularization.

[0106] Example 1:

[0107] This will be illustrated using the construction and application of the RegGAN-ResGNN model as an example. Figure 1 As shown, the model adopts an end-to-end deep learning architecture, and its processing flow mainly includes three core layers: the first layer is multi-source data perception and preprocessing, the second layer is ResGNN-based dual dynamic feature encoding, and the third layer is a RegGAN generative pharmacodynamics inference engine. The specific implementation methods of each part are described in detail below:

[0108] Step 1: Multi-source data perception and graph structure construction; This step aims to transform heterogeneous biochemical data into graph-structured data that the model can process, such as... Figure 1 The first layer shown;

[0109] 1.1) Construction of drug molecule maps:

[0110] Obtain the SMILES sequence of the target drug from drug databases such as PubChem or DrugBank; use a cheminformatics toolkit to resolve the SMILES sequence into a two-dimensional molecular map. Among them, the node set V d Characterizing drug atoms includes physicochemical features such as atom type, hybrid orbital type, implicit valence, and formal charge; edge set E d Characterize chemical bonds, including single bonds, double bonds, triple bonds, and aromatic bonds.

[0111] 1.2) Construction of patient gene co-expression network:

[0112] This study obtains RNA-seq gene expression profiles from breast cancer patients' omics data. To capture dynamic regulatory relationships between genes rather than relying solely on single expression levels, this embodiment constructs a gene co-expression network. Specifically, the top-N highly variable core genes are screened, and the expression levels of any two genes are calculated. and PCC in the population sample. A significance threshold was set: PCC > 0.7 and p-value < 0.05. When the correlation coefficient was higher than the threshold, a connection edge was established between the two gene nodes. This ultimately forms a patient-specific gene co-expression network. The node feature is the gene expression value, and the edge feature is the co-expression intensity.

[0113] Step Two: Dual Dynamic Feature Encoding Based on ResGNN; This step is the core step to overcome the limitations of "cold start" and "static representation," aiming to extract dynamic features that integrate general knowledge and task-specific knowledge, such as... Figure 1 The second layer is shown.

[0114] 2.1) Construction of the ResGNN drug encoder:

[0115] This embodiment designs a dual-path residual graph neural network structure that includes "static branches" and "dynamic branches".

[0116] Static branch: This branch uses a graph neural network model, MolGNet, pre-trained on a large-scale general compound database as its base. During the training process for the current breast cancer drug efficacy prediction task, all model parameters are frozen, and no gradient updates occur. Its function is to extract stable and universal underlying chemical structural features H of drug molecules. static .

[0117] Dynamic branch: Construct a graph neural network layer with a similar or lighter structure to the static branch. During training, the parameters of this branch are fine-tuned on a specific breast cancer dataset. Its role is to capture the specific activity characteristics of drug molecules in a specific tumor microenvironment. dynamic .

[0118] Hot-start residual fusion: Introducing a learnable or pre-defined strategy-controlled hot-start controller coefficient λ. Final drug characterization H drug The result is obtained by fusing the residuals using the following formula:

[0119]

[0120] The warm-start strategy is implemented as follows: In the initial training phase, the value of λ=0.01 is set close to 0. At this time, the model mainly relies on the general prior knowledge of the static branch for inference, thereby avoiding gradient oscillations in the cold-start phase. As the number of training iterations increases, the value of λ is gradually increased according to a preset non-linear growth curve, gradually releasing the learning ability of the dynamic branch, allowing the model to smoothly transition to learning specific task features.

[0121] 2.2) Construction of the patient feature encoder:

[0122] A DAE-concatenated Multilayer Perceptron (MLP) structure was adopted. The gene co-expression network data (which can be flattened into an adjacency matrix and a feature matrix) constructed in step 1.2) was input into the DAE for dimensionality reduction and noise denoising. Subsequently, high-dimensional latent space features were extracted using the MLP to obtain the patient's deep molecular characterization Z. patient .

[0123] 2.3) Multimodal feature fusion:

[0124] The drug characterization obtained above H drug Patient Characteristic Z patient The fusion process can be performed by concatenating feature vectors or by introducing an attention mechanism to automatically learn the interaction weights between drug substructures and specific gene modules, ultimately generating a joint feature vector. This serves as the input for the next layer of the inference engine.

[0125] Step 3: Construction of the RegGAN generative pharmacodynamics inference engine ( Figure 1 Third layer)

[0126] This step utilizes the game-theoretic mechanism of generative adversarial networks to address the overfitting problem under sparse data and achieve robust drug efficacy prediction.

[0127] 3.1) Generator G Construction: The generator is a deep feedforward neural network. It receives the joint feature vector output from step 2.3). Through several fully connected layers and the nonlinear activation function ReLU, the final output is a continuous value, namely the predicted drug sensitivity value. .

[0128] 3.2) Discriminator D Construction: The discriminator is also a neural network classifier. Its input consists of two parts: one is the actual drug sensitivity value y from clinical trials, and the other is the predicted drug sensitivity value output by the generator. The discriminator's task is to distinguish, as accurately as possible, whether the input value comes from the real data distribution or is generated by the generator.

[0129] 3.3) Joint Loss Function and Adversarial Training: The model is trained using an adversarial training strategy that alternates between optimizing the generator and discriminator. The total loss function consists of two parts:

[0130] Regression loss: Calculate the generator's predicted value The mean squared error between the predicted value and the true value y is used to constrain the numerical accuracy of the predicted result: .

[0131] Adversarial loss: A loss form based on the original GAN ​​or Wasserstein Generative Adversarial Network (WGAN). The generator aims to minimize the probability that its generated fake data will be recognized by the discriminator, while the discriminator aims to maximize its recognition ability. As a strong regularization term, the adversarial loss constrains the IC50 value predicted by the generator to approximate the distribution of real biological data on the overall distribution manifold, thus preventing the model from rote memorization on small sample data.

[0132] Overall optimization objective: The total loss of the generator is , where γ is the balance coefficient. Through joint optimization, the model significantly improves its generalization ability and robustness while maintaining numerical accuracy.

[0133] Using the RegGAN-ResGNN model trained above for precision diagnosis and treatment decision support in breast cancer:

[0134] 1. Collect tumor tissue from patients with breast cancer to be predicted for gene sequencing, obtain gene expression profile data, and construct the co-expression network of the patient according to step 1.2).

[0135] 2. Determine the list of candidate drugs to be evaluated, obtain their SMILES, and construct a molecular map as per step 1.1).

[0136] 3. Input patient data and data for a specific drug into the trained model. The model undergoes feature extraction and fusion via a dual encoder, and finally, the generator outputs the predicted IC50 value for that drug for that patient.

[0137] 4. Repeat step 3) for all drugs in the candidate list, sort the drugs from low to high according to the predicted IC50 value, generate a recommended drug list, and assist clinicians in developing individualized treatment plans.

[0138] like Figure 2 As shown, this invention comprehensively verifies the technical advantages of the proposed method (RegGAN-ResGNN model) through experiments in three core dimensions:

[0139] (1) Regarding robustness to sparse data, as shown in the line graph, even under the extremely sparse condition of using only 10% of the training data, this method can still maintain a prediction Pearson correlation coefficient (PCC) higher than 0.88, while the traditional benchmark model shows severe performance degradation. This proves that the introduction of the RegGAN adversarial regularization mechanism can effectively prevent overfitting to small samples.

[0140] (2) Regarding the verification of the hot-start residual mechanism, as shown in the network comparison in the figure, the randomly initialized cold-start model is prone to slow convergence and instability during training; while this method gradually releases the learning ability of the dynamic residual branch by introducing the hot-start adjustment coefficient λ, thus achieving rapid and stable convergence of the model and successfully overcoming the "cold-start" bottleneck of drug molecules on new tasks.

[0141] (3) In terms of cross-database generalization ability, as shown in the bar chart in the figure, when conducting zero-sample cross-dataset tests (from the GDSC dataset to the independent CCLE dataset), the prediction performance of this method is significantly better than that of traditional comparative models such as CNN, GNN and MLP, which fully verifies the strong anti-batch effect ability and excellent cross-database generalization performance brought about by multimodal deep fusion of drug graph representation and cell line gene co-expression network.

[0142] The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A drug response prediction method based on residual graph neural network, characterized in that, Includes the following steps: 1) Construct multi-source heterogeneous data and molecular graph features, extract and preprocess multi-source heterogeneous raw data of drugs and cell lines from large-scale databases; 2) Construct a dual-path drug feature encoder based on residual connections to extract dynamic drug characterizations that take into account both general chemical prior knowledge and specific tumor microenvironment task specificity; 3) A progressive hot-start training strategy is adopted to achieve the transition from baseline prior knowledge to specific fine-tuning; 4) Construct a cell line feature encoder to extract deep and robust cell line molecular characterizations from high-dimensional and redundant gene network data. 5) Construct a registration-based generative adversarial regularization prediction framework and utilize the game mechanism of generative adversarial networks to achieve drug efficacy prediction.

2. The drug response prediction method based on residual graph neural network according to claim 1, characterized in that, Step 1) specifically includes: constructing a gene co-expression network reflecting dynamic regulatory relationships by calculating the correlation between genes, and converting a one-dimensional drug sequence into a two-dimensional topological graph structure containing rich physicochemical properties; obtaining a large-scale cancer drug sensitivity database, including a cell line gene expression profile data matrix. And the set of simplified molecular linear input specifications (SMILES) for drugs; 1-1) Network Construction: Calculate the expression of any two genes in the gene expression data. and The Pearson correlation coefficient (PCC) between the two pairs is used to construct the co-expression adjacency matrix. : ； 1-2) Drug Molecular Map Construction: Using cheminformatics tools to convert drug samples into molecular maps. ;in For a set of atomic nodes, For the set of chemical bond edges; construct the adjacency matrix of drug molecules. and atomic feature matrix .

3. The drug response prediction method based on residual graph neural network according to claim 1, characterized in that, Step 2) specifically includes: designing a dual-path parallel graph neural network architecture, combining the frozen main path for extracting general features with the dynamic branch path for specific fine-tuning, and performing adaptive residual fusion. 2-1) Static main path: Extract general features using the pre-trained model MolGNet; parameters of this branch. Freeze during training; ； in, This represents the general feature matrix of drug molecules extracted from static branches. This represents the input drug molecule diagram. This represents the frozen network weight parameters in the pre-trained MolGNet model; 2-2) Dynamic Residual Branch: Construct an L-layer Graph Convolutional Network (GCN). For the l-th layer, its feature propagation formula is defined as: ； in, To add a self-loop adjacency matrix, for The degree matrix, For the first The trainable weight matrix of the layer, The function is a linear rectified function ReLU; after propagation through L layers, the residual feature vector is obtained through global max pooling. : ； 2-3) Adaptive residual fusion: This involves weighted fusion of static and dynamic features to obtain the final drug characterization. : ； in, This indicates vector concatenation or element-wise addition. This is the hot start adjustment coefficient.

4. The drug response prediction method based on residual graph neural network according to claim 1, characterized in that, Step 3) specifically includes: introducing a time-dependent adjustment function based on training epochs to achieve the transition from baseline prior knowledge to specific fine-tuning; designing the time-dependent adjustment function. To control the intervention intensity of residual branches, Defined as a piecewise function with respect to training epochs t: ； in, Here, k represents the preset freeze period, and k is the growth slope. The center point of the sigmoid function is used; this strategy ensures that only the center point is used in the early stages of training. Benchmark training is performed, followed by a smooth introduction of residual feature vectors. Make minor adjustments.

5. The drug response prediction method based on residual graph neural network according to claim 1, characterized in that, Step 4) specifically includes: injecting noise into the original gene expression vector to damage it, and using an autoencoder mechanism to minimize the reconstruction error, so that the model learns low-dimensional manifold features with strong generalization ability; and using a denoising autoencoder (DAE) to extract low-dimensional manifold features of gene expression. 4-1) Noise injection: injecting noise into the original gene expression vector Add Gaussian noise Damaged input ; 4-2) Encoding and Decoding: ； ； Where z represents the feature vector of the cell line extracted by the encoder. This represents the reconstructed gene expression vector output by the decoder. Represents a non-linear activation function. and Let these represent the trainable weight matrices of the encoder network and the decoder network, respectively. and These represent the bias vectors of the encoder network and the decoder network, respectively; 4-3) Pre-training loss: Optimize the DAE parameters by minimizing the reconstruction error: ； Finally, the encoder output z is taken as the cell line feature. .

6. The drug response prediction method based on residual graph neural network according to claim 1, characterized in that, Step 5) specifically includes: constructing a minimax game framework between the generator and the discriminator, using the distribution consistency constraint generated by the adversarial mechanism as a strong regularization term, and combining it with traditional regression error for joint optimization to achieve drug efficacy prediction; using a generative adversarial network (GAN) for regression prediction, the model includes a generator G, a discriminator D, and an objective function. Generator G: Input is drug features cell line characteristics The concatenation of these values ​​outputs the predicted drug sensitivity values. : ； Discriminator D: Input is the actual drug sensitivity value y or the predicted drug sensitivity value Output the probability that the sample belongs to the true distribution. ; Objective function: The model's total loss function From regression loss and combat losses Composition and overall optimization objective: Regression loss: Mean squared error is used to ensure the accuracy of the predicted values; ； Adversarial Loss: Employing either Least Squares Generative Adversarial Network (LSGAN) loss or standard Generative Adversarial Network (GAN) loss, the distribution of predicted values ​​is optimized. Distribution that approximates the true value ; ； Overall optimization objective: The generator G and the discriminator D engage in a minimax game. ； Here, γ is the balance coefficient, used to adjust the strength of the resistance to regularization.

Images