Drug synergy prediction graph large model for out-of-distribution generalization

Abstract

This application discloses an out-of-distribution generalized drug synergy prediction graph model, belonging to the field of computer technology. In this model, the target adaptive decoupling molecular graph encoding module can obtain richer molecular graph representations and enable the model to explicitly distinguish between the inherent structural properties of drugs and the target-related functional properties; the multi-level contextual cell feature alignment model allows the cellular environment to participate in the message passing process within the molecular graph; the pairwise attention graph neural architecture search module dynamically searches and combines graph neural network structures suitable for the current drug combination to accurately obtain molecular embedding representations; and the DrugSyn-LLM module, which enhances biomedical instruction fine-tuning through retrieval, transforms drug synergy prediction into a structured generation task of a large language model, enabling the model to simultaneously utilize multimodal information such as cell line text descriptions, target representations, individual drug molecular embedding representations, and drug SMILES sequences for prediction.

Description

Technical Field

[0001] This application belongs to the field of computer technology, specifically relating to a large model for predicting the synergistic effects of out-of-distribution generalized drugs. Background Technology

[0002] Current drug synergy prediction techniques typically represent drugs as drug molecular graphs, using graph neural networks to learn the representation of the drug's topological structure, and then combining cell line features or drug-related features to predict the synergy score or category of drug combinations through multilayer perceptrons, attention networks, or other deep learning models.

[0003] However, in the process of implementing this application, the inventors found that the current solutions are mainly aimed at the scenario of predicting drug synergy within the distribution, and usually adopt a fixed graph neural network architecture or a simple multimodal feature fusion method. As a result, the generalization ability is insufficient for the scenario of predicting drug synergy outside the distribution, where there are new drugs, new molecular skeletons or different cell line environments, which leads to low prediction accuracy of drug synergy outside the distribution. Summary of the Invention

[0004] The purpose of this application is to provide a large-scale model for predicting out-of-distribution generalized drug synergy, which can solve the problem that related technologies have insufficient generalization ability for predicting out-of-distribution drug synergy scenarios, resulting in low prediction accuracy.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a large-scale model for predicting the synergistic effects of out-of-distribution generalized drugs, the model comprising: The target-adaptive decoupled molecular graph encoding module is configured to, in response to receiving drug molecular graphs and target protein information for each drug, process the drug molecular graph of each drug into a molecular graph representation through multi-view encoding branches constructed for different chemical structural relationships and semantic functions; decouple the molecular graph representation of each drug into a target-independent molecular representation and a target-related molecular representation, wherein the target-independent molecular representation is used to characterize the intrinsic structural properties of the drug, and the target-related molecular representation is used to characterize the target-related functional properties of the drug; determine the target-adaptive molecular representation of each drug under each target protein associated with each drug based on the target protein information and the target-related molecular representation of each drug; and determine the target-adaptive associated representation of each drug based on the target-adaptive molecular representation of each drug associated with each drug and the target-independent molecular representation of each drug, and input it into the pairwise attention graph neural architecture search module. A multi-level contextualized cell feature alignment module is configured to, in response to receiving cell line gene expression features and drug molecule maps of each drug, add the cell line gene expression features to the drug molecule map of each drug to obtain a feature-enhanced molecular map of each drug; map the cell line gene expression features to a continuous language space representation as a target representation; input the feature-enhanced molecular maps of each drug to a pairwise attention map neural architecture search module, and input the target representation to a DrugSyn-LLM module for retrieving enhanced biomedical instruction fine-tuning; The pairwise attention graph neural architecture search module is configured to select a graph neural network operator based on the pairwise attention results of the adaptive association representation of the target of each drug, and use the selected graph neural network operator to process the featured molecular graph of each drug into a molecular embedding representation and input it into the DrugSyn-LLM module for retrieval enhancement and fine-tuning of biomedical instructions. The DrugSyn-LLM module, which enhances the fine-tuning of biomedical instructions for retrieval, is configured to generate synergistic effect predictions for each drug based on the received cell line text description, the target representation, the molecular embedding representation of each drug, and the drug SMILES sequence.

[0006] Secondly, embodiments of this application provide a method for predicting the synergistic effect of out-of-distribution generalized drugs, applied to a large-scale model for predicting the synergistic effect of out-of-distribution generalized drugs. The method includes: In the large-scale model for predicting out-of-distribution generalized drug synergy, the target-adaptive decoupled molecular graph encoding module, in response to receiving the drug molecular graphs and target protein information of each drug, processes the drug molecular graph of each drug into a molecular graph representation through multi-view encoding branches constructed for different chemical structural relationships and semantic effects; and decouples the molecular graph representation of each drug into a target-independent molecular representation and a target-related molecular representation, wherein the target-independent molecular representation is used to characterize the intrinsic structural properties of the drug, and the target-related molecular representation is used to characterize the target-related functional properties of the drug; and determines the target-adaptive molecular representation of each drug under each target protein associated with each drug based on the target protein information and the target-related molecular representation of each drug; and determines the target-adaptive associated representation of each drug based on the target-adaptive molecular representation of each drug associated with each drug and the target-independent molecular representation of each drug, and inputs it into the pairwise attention graph neural architecture search module; In the large-scale model for predicting drug synergy outside the distribution, a multi-level contextualized cell feature alignment module, in response to receiving cell line gene expression features and drug molecular maps of each drug, adds the cell line gene expression features to the drug molecular map of each drug to obtain a feature-bearing molecular map of each drug; maps the cell line gene expression features to a continuous language space representation as a target representation; and inputs the feature-bearing molecular maps of each drug into a pairwise attention map neural architecture search module, and inputs the target representation into a DrugSyn-LLM module for retrieving enhanced biomedical instruction fine-tuning. The pairwise attention graph neural architecture search module in the large model of out-of-distribution generalized drug synergy prediction graph selects graph neural network operators based on the pairwise attention results of the adaptive association representation of each drug's target. Using the selected graph neural network operators, the featured molecular graph of each drug is processed into a molecular embedding representation and input into the DrugSyn-LLM module for retrieval enhancement and fine-tuning of biomedical instructions. The DrugSyn-LLM module, which enhances biomedical instruction fine-tuning, is retrieved from the large model of out-of-distribution generalized drug synergy prediction graph. Based on the received cell line text description, the target representation, the molecular embedding representation of each drug, and the drug SMILES sequence, the synergy prediction results of each drug are generated.

[0007] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the out-of-distribution generalization drug synergy prediction method as described in the second aspect.

[0008] Fourthly, embodiments of this application provide a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the out-of-distribution generalized drug synergy prediction method as described in the second aspect.

[0009] Fifthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the out-of-distribution generalized drug synergy prediction method as described in the second aspect.

[0010] In this embodiment, the target-adaptive decoupled molecular graph encoding module utilizes multi-perspective encoding branches constructed for different chemical structural relationships and semantic effects to obtain richer molecular graph representations. These representations are further divided into target-independent and target-related molecular representations, enabling the model to explicitly distinguish between the inherent structural properties of drugs and target-related functional properties, thereby enhancing the model's adaptability in different contexts. The multi-level contextual cell feature alignment model simultaneously injects gene expression features into the molecular graph structure encoding process and the semantic input space of the large language model, allowing the cellular environment to participate in the message passing process within the molecular graph. The pairwise attention graph neural architecture search module dynamically searches for and combines graph neural network structures suitable for the current drug combination to accurately obtain molecular embedding representations. The DrugSyn-LLM module, which enhances biomedical instruction fine-tuning through retrieval, transforms drug synergy prediction into a structured generation task for the large language model, enabling the model to simultaneously utilize multimodal information such as cell line text descriptions, target representations, individual drug molecular embedding representations, and drug SMILES sequences for prediction. This improves the model's generalization ability in out-of-distribution drug synergy prediction scenarios, thereby enhancing prediction accuracy in such scenarios. Attached Figure Description

[0011] Figure 1 A schematic diagram of a large-scale model for predicting the synergistic effect of out-of-distribution generalized drugs, provided in an embodiment of this application; Figure 2 A flowchart illustrating an out-of-distribution generalization drug synergy prediction method provided in this application embodiment; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0014] The following detailed description, in conjunction with the accompanying drawings, of a large-scale model for predicting the synergistic effects of out-of-distribution generalized drugs provided in this application, through specific embodiments and application scenarios, will be provided in detail.

[0015] See Figure 1 The image shown is a schematic diagram of a large-scale model for predicting the synergistic effect of out-of-distribution generalized drugs provided in an embodiment of this application. The model includes: The target-adaptive decoupled molecular graph encoding module is configured to, in response to receiving drug molecular graphs and target protein information for each drug, process the drug molecular graph of each drug into a molecular graph representation through multi-view encoding branches constructed for different chemical structural relationships and semantic functions; decouple the molecular graph representation of each drug into a target-independent molecular representation and a target-related molecular representation, wherein the target-independent molecular representation is used to characterize the intrinsic structural properties of the drug, and the target-related molecular representation is used to characterize the target-related functional properties of the drug; determine the target-adaptive molecular representation of each drug under each target protein associated with each drug based on the target protein information and the target-related molecular representation of each drug; and determine the target-adaptive associated representation of each drug based on the target-adaptive molecular representation of each drug associated with each drug and the target-independent molecular representation of each drug, and input it into the pairwise attention graph neural architecture search module. A multi-level contextualized cell feature alignment module is configured to, in response to receiving cell line gene expression features and drug molecule maps of each drug, add the cell line gene expression features to the drug molecule map of each drug to obtain a feature-enhanced molecular map of each drug; map the cell line gene expression features to a continuous language space representation as a target representation; input the feature-enhanced molecular maps of each drug to a pairwise attention map neural architecture search module, and input the target representation to a DrugSyn-LLM module for retrieving enhanced biomedical instruction fine-tuning; The pairwise attention graph neural architecture search module is configured to select a graph neural network operator based on the pairwise attention results of the adaptive association representation of the target of each drug, and use the selected graph neural network operator to process the featured molecular graph of each drug into a molecular embedding representation and input it into the DrugSyn-LLM module for retrieval enhancement and fine-tuning of biomedical instructions. The DrugSyn-LLM module, which enhances the fine-tuning of biomedical instructions for retrieval, is configured to generate synergistic effect predictions for each drug based on the received cell line text description, the target representation, the molecular embedding representation of each drug, and the drug SMILES sequence.

[0016] In this embodiment of the application, considering that traditional methods often lack an effective distinction between molecular structural information that is "related to the cell target" and "unrelated to the cell target", it is difficult to obtain a target-adaptive drug representation for a specific cellular environment. Therefore, a target-adaptive decoupled molecular graph encoding module is constructed to extract the inherent structural information of the drug and the functional information related to the target protein from the drug molecular graph.

[0017] Specifically, for any drug d, its drug molecular diagram can be represented as follows: , and This represents the set of atomic nodes and the set of chemical bonds of drug d.

[0018] Unlike traditional methods that statically encode molecular graphs, the target-adaptive decoupled molecular graph encoding module constructs multi-perspective encoding branches oriented towards different chemical structural relationships and semantic interactions, thereby obtaining a richer molecular graph representation for each drug. For example, multiple heterogeneous graph neural networks can be used to represent each atom node in the drug molecule graph. Encode and use a feature fusion function and graph-level pooling functions Molecular diagram representation obtained The process can be represented as follows:

[0019] Building upon this, unlike traditional methods that directly encode drug molecules as a single vector representation, the target-adaptive decoupled molecular graph encoding module further divides the drug molecule representation into target-independent molecular representation and target-related molecular representation. This enables the model to explicitly distinguish between the inherent structural properties of the drug and the target-related functional properties, thereby enhancing the model's adaptability in different contexts.

[0020] As one possible implementation, the target-adaptive decoupled molecular graph encoding module can be based on a decoupled representation learning mechanism to divide the drug molecule representation into target-independent molecular representation and target-related molecular representation. In this case, the target-independent molecular representation and the target-related molecular representation can be determined by the following formula: in, and This represents the target-independent molecules and the target-related molecules of drug d. The molecular diagram representing drug d. and This represents the mapping matrix of target-independent features and target-related features learned in advance during the training phase.

[0021] Then, the target adaptive decoupling molecular graph encoding module determines the target adaptive molecular representation of each drug under each target protein based on the target protein information and target-related molecular representation of each drug, using methods such as cross-attention mechanism interaction.

[0022] Finally, the target-adaptive decoupling molecular graph encoding module determines the target-adaptive association representation for each drug based on the target-adaptive molecular representations associated with each drug and the target-independent molecular representations for each drug, and inputs this representation into the pairwise attention graph neural architecture search module. For example, the target-adaptive decoupling molecular graph encoding module can concatenate the target-independent molecular representations of each drug with the target-adaptive molecular representations associated with each drug to form the final drug representation for each drug (i.e., the target-adaptive association representation). The target-adaptive association representation can then be represented as follows:

[0023] in, This represents the adaptive association between the target and drug d. This represents the target-independent molecule representation of drug d. Let K represent the target-adaptive molecular representation of the k-th target protein associated with drug d, where K represents the number of target proteins of drug d.

[0024] Therefore, the target-adaptive decoupled molecular graph encoding module can provide a more target-adaptive and functionally separable molecular representation for subsequent drug pair modeling and out-of-distribution prediction.

[0025] Furthermore, considering that fixed graph neural network architectures are difficult to adapt to the structural differences of different drugs, and that simple feature splicing or shallow fusion methods cannot fully leverage the complementary effects between molecular graph structural information, biomedical semantic information, and cell line context information, a pairwise attention graph neural architecture search module was constructed to dynamically search for and combine graph neural network structures suitable for the current drug combination.

[0026] Specifically, the pairwise attention graph neural architecture search module can select graph neural network operators based on the pairwise molecular attention mechanism and the pairwise attention results of the adaptive association representation of each drug's target. For example, it can determine the weight of each graph neural network operator for each drug, and then select and combine different graph neural network operators according to the weights to process the featured molecular graph of the corresponding drug into a molecular embedding representation and input it into the DrugSyn-LLM module for retrieval enhancement and fine-tuning of biomedical instructions.

[0027] Furthermore, considering that drug synergy has a significant cell line dependence, the same drug combination may exhibit different synergistic or antagonistic effects in different cell lines. Therefore, a multi-level contextualized cell feature alignment module was also constructed to simultaneously inject cell line information (i.e. cell line gene expression features) into the molecular graph structure encoding process and the semantic input space of the large language model.

[0028] As one possible implementation, in structural-level alignment, for cell line c, let its gene expression characteristics be: The multi-level contextualized cell feature alignment module first determines the corresponding cell context representation. For example, this can be achieved through context projection functions. Will Convert to The process can be represented as follows:

[0029] Then, for any atomic node in the drug molecule diagram of drug d. Let its initial atomic characteristics be The multi-level contextualized cell feature alignment module represents the cell context. With atomic characteristics By splicing the images together, we can obtain a molecular map with distinctive features. This process can be represented as follows:

[0030] in, The features of atomic node v in the molecular diagram representing drug d. This represents the set of atomic nodes in the drug molecule diagram of drug d.

[0031] Therefore, by directly injecting the cell line context into the atomic node representation, the cellular environment can participate in the message passing process within the molecular graph.

[0032] As one possible implementation, in semantic alignment, a multi-level contextualized cell feature alignment module maps cell line gene expression features to a large language model input space. For example, it maps cell line gene expression features to a continuous language space representation as the target representation through a pre-trained projector, so that the DrugSyn-LLM module, which enhances biomedical instruction fine-tuning, can process it.

[0033] Optionally, the multi-level contextualized cell feature alignment module can further map the cell line text description to the large language model input space. For example, it can convert the cell line description text into discrete tokens and input them into the DrugSyn-LLM module for retrieval enhancement and biomedical instruction fine-tuning, replacing the direct input of the cell line text description. This eliminates the need for the DrugSyn-LLM module to process the cell line description text.

[0034] Furthermore, this application also constructs a DrugSyn-LLM module for enhanced biomedical instruction fine-tuning to further integrate molecular structure information, biomedical semantic knowledge, and task reasoning capabilities.

[0035] As one possible implementation, the DrugSyn-LLM module for retrieving enhanced biomedical instruction fine-tuning is also configured to perform the following steps: The molecular embedding representation of each drug is mapped to the language model input space by a drug projector to obtain the drug structure representation of each drug. The cell line text descriptions and drug SMILES sequences for each drug were converted into text lexical units. By inputting the text terms describing the cell line, the drug structure representation of each drug, and the text terms of the drug SMILES sequence into the biomedical big language model, the synergistic effect prediction results of each drug are obtained.

[0036] In practical implementation, the DrugSyn-LLM module, which enhances biomedical instruction fine-tuning, utilizes a drug projector to map the molecular embedding representation obtained by the pairwise attention map neural architecture search module to the language model input space, thus obtaining a drug structure representation. It can also use the large model's segmenter to convert drug SMILES sequences into text tokens (e.g., CC(=O)NC1=CC=C(O)C=C1, acetaminophen). Furthermore, it can obtain cell line-related text descriptions from external biomedical knowledge and process them into text tokens describing cell lines using the large model's segmenter. Subsequently, the DrugSyn-LLM module, which enhances biomedical instruction fine-tuning, inputs the aforementioned information into the large biomedical language model (i.e., the large model after retrieval-enhanced biomedical instruction fine-tuning) to generate drug synergy prediction results.

[0037] Understandably, unlike traditional methods that directly output prediction results through classifiers or regressors, this application transforms the prediction of drug synergistic effects into a structured generation task of a large language model, thereby enabling the model to make predictions using multimodal information such as molecular structure information, cellular context information, and biomedical semantic knowledge.

[0038] As can be seen from the above technical solutions, the target-adaptive decoupled molecular graph encoding module utilizes multi-perspective encoding branches constructed for different chemical structural relationships and semantic effects to obtain richer molecular graph representations. It further divides these representations into target-independent and target-related molecular representations, enabling the model to explicitly distinguish between the inherent structural attributes of drugs and target-related functional attributes, thereby enhancing the model's adaptability in different contexts. The multi-level contextual cell feature alignment model simultaneously injects gene expression features into the molecular graph structure encoding process and the semantic input space of the large language model, allowing the cellular environment to participate in the message passing process within the molecular graph. The pairwise attention graph neural architecture search module dynamically searches for and combines graph neural network structures suitable for the current drug combination to accurately obtain molecular embedding representations. The DrugSyn-LLM module, which enhances biomedical instruction fine-tuning through retrieval, transforms drug synergy prediction into a structured generation task for the large language model, enabling the model to simultaneously utilize multimodal information such as cell line text descriptions, target representations, individual drug molecular embedding representations, and drug SMILES sequences for prediction. This improves the model's generalization ability in out-of-distribution drug synergy prediction scenarios, thereby enhancing prediction accuracy in such scenarios.

[0039] In some embodiments, the target-adaptive decoupling molecular graph encoding module is further configured to perform the following steps: Using a pre-trained protein encoder, the target protein information of each drug is processed into target protein representations under each target protein associated with each drug. By interacting with the target protein representations associated with each drug and the target-related molecular representations of each drug through a cross-attention mechanism, adaptive molecular representations of the target proteins associated with each drug are obtained.

[0040] In practical implementation, for the K related target proteins corresponding to drug d, the target adaptive decoupling molecular graph encoding module can use a pre-trained protein encoder to obtain the target protein representation under the K related target proteins. Furthermore, it interacts with the target-related molecular representation of drug d through a cross-attention mechanism to obtain a target-adaptive molecular representation, which can be represented as follows:

[0041] in, This represents the target-adaptive molecular representation of drug d at the k-th target protein. This indicates cross-attention mechanism interaction. This represents the target-related molecules of drug d.

[0042] Understandably, this approach, based on a pre-trained protein encoder, can fully extract features from long protein sequences, thus avoiding the use of lengthy protein sequences.

[0043] In some embodiments, the loss term associated with the target adaptive decoupling molecular graph encoding module in the loss function used during the model training phase. It is expressed as follows:

[0044] in, express and The corresponding correlation matrix, and This indicates the different target conditions k of the two drugs. The decoupling representation composed of the K target site adaptive molecular representations obtained below, ||·|| F This represents the operation for calculating the F-norm.

[0045] In this embodiment of the application, in order to further reduce the information redundancy between different target adaptive representations, the application adds decorrelation constraints between molecular representations under different target conditions based on the correlation matrix. Thus, the above-mentioned loss term is designed for the target adaptive decoupled molecular graph encoding module to enable molecular representations under different target conditions to learn complementary and mutually distinguishable target-related information, and avoid multiple target representations from collapsing into a similar space.

[0046] As one possible implementation, for any two different target conditions k, The K target adaptive molecular representations obtained below can be used to calculate their correlation matrix using the following formula. This describes the statistical correlation or information overlap between different decoupling representations:

[0047] in, and This indicates the different target conditions k of the two drugs. The decoupling representation is composed of the K target site adaptive molecular representations obtained below. This represents the degrees of freedom of the target condition k.

[0048] In some embodiments, the pairwise attention graph neural architecture search module is further configured to perform the following steps: The multi-head attention mechanism is used to process the adaptive association representation of each drug's target, and the context interaction vector of each drug is used as the pairwise attention result. For each candidate operator associated with a pre-constructed multi-layer graph neural network, the weight of the candidate operator for each drug is determined based on the pre-learned latent vector corresponding to the candidate operator and the context interaction vector of each drug. Each layer of the multi-layer graph neural network is configured with a set of candidate message passing operators, and different candidate operators in the set of candidate message passing operators correspond to different graph information propagation methods. Based on each candidate operator associated with the multilayer graph neural network and the weight of each candidate operator for each drug, the molecular representation of the featured molecular graph of each drug is iteratively updated, and the molecular representation obtained by the final update of the featured molecular graph of each drug is used as the molecular embedding representation of each drug.

[0049] In practical implementation, for a drug combination (d1, d2), a multi-head attention mechanism is used to achieve bidirectional contextual interaction between the two drugs. This process can be represented as the superposition of the original drug vector and the attention mechanism result vector, i.e.:

[0050] in, Represents the context interaction vector of drug d1; This represents the adaptive correlation representation of drug targets; FFN represents a feedforward neural network. The multi-head attention mechanism is represented by Q, K, and V, which represent the query vector, key vector, and value vector, respectively, and are defined as two entities in the drug pair interaction. That is, in this embodiment of the application, the bidirectional context interaction is achieved by using two drug representations as query vectors and another drug as key and value vectors to perform attention calculation.

[0051] Furthermore, this application constructs a latent operator space parameterization mechanism, namely, configuring a set of candidate message passing operators in the l-th layer graph neural network. Each candidate operator corresponds to a different graph information propagation method, and in order to make the architecture search process differentiable, a learnable latent vector is assigned to each candidate operator, thereby mapping it to a continuous latent operator space.

[0052] Subsequently, based on the adaptive routing mechanism, the weights of the candidate operators for each drug are determined according to the pre-learned latent vectors corresponding to the candidate operators and the context interaction vectors of each drug, so as to dynamically select the appropriate graph neural network operator for the current drug.

[0053] Optionally, the weight of each candidate operator for each drug is determined by the following formula:

[0054] in, Let represent the weight of the i-th candidate operator associated with the l-th layer graph neural network for drug d, and m represent the number of candidate operators in the candidate message passing operator set configured in the l-th layer graph neural network. This represents the context interaction vector of drug d. , Let represent the pre-learned latent vectors corresponding to the i-th and j-th candidate operators associated with the l-th layer graph neural network.

[0055] Finally, based on each candidate operator associated with the multilayer graph neural network and the weight of each candidate operator for each drug, the molecular representation of each drug is updated iteratively using the featured molecular graph, thus accurately obtaining the molecular embedding representation of each drug.

[0056] Optionally, the molecular representation update process performed by the l-th layer graph neural network is as follows:

[0057] in, This represents the molecular representation of drug d output by the l-th layer graph neural network. This represents the molecular representation of drug d as input to the l-th layer graph neural network. The molecular representation of the initial drug d. The features of atomic node v in the molecular diagram representing drug d. This represents the i-th candidate operator associated with the l-th layer of the graph neural network. For the weight of drug d, m represents the number of candidate operators in the candidate message passing operator set configured in the l-th layer graph neural network.

[0058] Optionally, in the loss function used during the model training phase, the loss term associated with the pairwise attention map neural architecture search module may be included. It is expressed as follows:

[0059] in, , Let represent the learnable latent vectors corresponding to the i-th and j-th candidate operators in the candidate message passing operator set configured in the l-th layer graph neural network, and m represent the number of candidate operators in the candidate message passing operator set configured in the l-th layer graph neural network.

[0060] In this embodiment of the application, in order to ensure that different candidate operators have functional distinction in the latent space, the application further designs the above-mentioned operator separation loss for the pairwise attention graph neural architecture search module to suppress the excessively high similarity between the latent representations of different candidate operators, avoid spatial degradation, and thus overcome the defect that the molecular representation learned by the traditional graph neural network model is prone to failure when the test drug is significantly different from the training drug in topology, resulting in a decline in prediction performance.

[0061] In some embodiments, the loss function used during the model training phase includes the loss term associated with the DrugSyn-LLM module for retrieving enhanced biomedical instructions. It is expressed as follows:

[0062] in, This represents the loss term associated with the fine-tuning phase of retrieval-enhanced biomedical instructions, used to guide the large language model in learning external biomedical knowledge. This represents the loss term associated with the training phase of the drug synergy prediction task, used to guide the large language model to complete the prediction of drug synergy in a structured generation task manner. This indicates a stage indicator variable, and The value is 1 during the refinement phase of the enhanced biomedical instruction and 0 during the training phase of the drug co-prediction task.

[0063] In this embodiment, a two-stage training strategy is adopted for the DrugSyn-LLM module, which enhances the fine-tuning of biomedical instructions for retrieval.

[0064] The first stage is the retrieval enhancement and fine-tuning stage for biomedical instructions, which is used to inject external pharmacological, biological, and target mechanism knowledge into the large language model, giving it a stronger understanding of the biomedical domain. In this stage, for each training sample, this application retrieves domain knowledge related to the drug, target protein, cell line, or disease background from external biomedical databases and organizes it into a structured biomedical description as input prompts.

[0065] The second stage is the drug synergy prediction task training stage, which guides the large language model to complete the drug synergy prediction in a structured generation task manner. In this stage, the large language model receives task prompts, drug SMILES tokens, cell line tokens, continuous representations of cell lines, and continuous structural representations of the two drugs, and generates a structured prediction result R={R_label, R_score}, where R_label and R_score represent the category and score of drug combination synergy.

[0066] Optionally, the loss term associated with the retrieval enhancement biomedical instruction fine-tuning phase and the drug co-prediction task training phase. and It is expressed as follows:

[0067]

[0068] Where R represents the target generated response. , This indicates input prompts for the refinement phase of enhanced biomedical instructions and the training phase of the drug co-prediction task. Text terms representing cell line descriptions Text terms representing the drug SMILES sequence of drugs d1 and d2. Indicates the target representation, This represents the drug structure representation of d1 and d2.

[0069] As one possible implementation, during the model training phase, the various modules in the model can be jointly optimized. For example, the overall training can be performed with the final output of the large language model as the target, while simultaneously optimizing all parameters. In this case, the overall loss function used during the model training phase can be expressed as follows:

[0070] in, , Indicates the weight.

[0071] Therefore, the large-scale model for predicting out-of-distribution generalized drug synergies combines molecular graph representation learning of cellular context with a large language model through the above four jointly optimized modules, thereby ensuring that the model can accurately predict out-of-distribution generalized drug synergies.

[0072] Based on the above embodiments, this application mainly addresses the technical problem of predicting out-of-distribution generalized drug synergy, namely, how to accurately predict the synergistic effect of drug combinations (such as two drugs) in specific cell lines or disease cell environments under distribution offset scenarios such as new drugs, new molecular skeletons, or different molecular sizes. It can overcome the shortcomings of related technologies, such as drug synergy prediction models often relying heavily on in-distribution assumptions (i.e., relying on the assumption that the drug structure distributions in the training set and the test set are similar), difficulty in generalizing to new compounds that continuously appear in the real drug discovery process (i.e., there are out-of-distribution offset problems caused by new molecular skeletons, new compound sizes, and unknown drug structures), insufficient structural semantic fusion, insufficient cell context modeling, and weak model adaptability.

[0073] Specifically, this application provides a large-scale model for predicting out-of-distribution generalized drug synergy, which is intended for products or projects such as computational drug discovery, anticancer combination drug screening, personalized precision treatment, and drug combination optimization. It can be applied to drug synergy prediction platforms, AI-assisted drug combination screening systems, combination drug recommendation systems based on cell line context, and intelligent drug development systems that combine molecular graph structure and biomedical semantic information.

[0074] This application improves the prediction accuracy and generalization ability of the model in out-of-distribution drug combination scenarios by jointly modeling the topological structure of drug molecules, cell line context features, and biomedical semantic information through a graph-large model framework. Compared with related technologies, this application has at least the following advantages: 1. Enhance predictive capabilities in new drug scenarios: improve the product's generalization ability for new molecular skeletons, new drug combinations, and out-of-distribution samples.

[0075] 2. Improve the accuracy of drug synergistic prediction: Combine molecular graph structure, target protein, cell line context and biomedical semantic knowledge to enhance the model's expressive ability.

[0076] 3. Enhance product competitive barriers: Compared with fixed graph neural networks (GNN) or simple feature splicing schemes, this application has the advantages of target adaptive modeling, dynamic architecture search and graph large model fusion.

[0077] This application provides a method for predicting the synergistic effects of out-of-distribution generalized drugs, applied to a large-scale model for predicting the synergistic effects of out-of-distribution generalized drugs, such as... Figure 2 As shown, the method includes: Step S101: In the large-scale model of out-of-distribution generalized drug synergy prediction graph, the target adaptive decoupling molecular graph encoding module, in response to receiving the drug molecular graph and target protein information of each drug, processes the drug molecular graph of each drug into a molecular graph representation through multi-view encoding branches constructed for different chemical structure relationships and semantic effects; and decouples the molecular graph representation of each drug into a target-independent molecular representation and a target-related molecular representation, wherein the target-independent molecular representation is used to characterize the intrinsic structural properties of the drug, and the target-related molecular representation is used to characterize the target-related functional properties of the drug; and determines the target adaptive molecular representation of each drug under each target protein associated with each drug based on the target protein information and the target-related molecular representation of each drug; and determines the target adaptive association representation of each drug based on the target adaptive molecular representation of each drug associated with each drug and the target-independent molecular representation of each drug, and inputs it into the pairwise attention graph neural architecture search module; Step S102: In the large-scale model of out-of-distribution generalized drug synergy prediction graph, the multi-level contextualized cell feature alignment module, in response to receiving cell line gene expression features and drug molecule graphs of each drug, adds the cell line gene expression features to the drug molecule graph of each drug to obtain a feature-bearing molecular graph of each drug; maps the cell line gene expression features to a continuous language space representation as the target representation; and inputs the feature-bearing molecular graphs of each drug to the pairwise attention graph neural architecture search module, and inputs the target representation to the DrugSyn-LLM module for retrieval enhancement of biomedical instruction fine-tuning; Step S103: The pairwise attention graph neural architecture search module in the large model of out-of-distribution generalized drug synergy prediction graph selects graph neural network operators based on the pairwise attention results of the adaptive association representation of each drug's target, and uses the selected graph neural network operators to process the featured molecular graph of each drug into a molecular embedding representation and input it into the DrugSyn-LLM module for retrieval enhancement and fine-tuning of biomedical instructions. Step S104: Retrieve the DrugSyn-LLM module for enhanced biomedical instruction fine-tuning from the large model of the out-of-distribution generalized drug synergy prediction graph. Based on the received cell line text description, the target representation, and the molecular embedding representation and drug SMILES sequence of each drug, generate the synergy prediction results of each drug.

[0078] Optionally, determining the target-adaptive molecular representation for each drug's associated target protein based on the target protein information and target-related molecular representations includes: Using a pre-trained protein encoder, the target protein information of each drug is processed into target protein representations under each target protein associated with each drug. By interacting with the target protein representations associated with each drug and the target-related molecular representations of each drug through a cross-attention mechanism, adaptive molecular representations of the target proteins associated with each drug are obtained.

[0079] Optionally, in the loss function used during the model training phase, the loss term associated with the target adaptive decoupling molecular graph encoding module may be included. It is expressed as follows:

[0080] in, express and The corresponding correlation matrix, and This indicates the different target conditions k of the two drugs. The decoupling representation composed of the K target site adaptive molecular representations obtained below, ||·|| F This represents the operation for calculating the F-norm.

[0081] Optionally, the molecular embedding representation is determined through the following steps: The multi-head attention mechanism is used to process the adaptive association representation of each drug's target, and the context interaction vector of each drug is used as the pairwise attention result. For each candidate operator associated with a pre-constructed multi-layer graph neural network, the weight of the candidate operator for each drug is determined based on the pre-learned latent vector corresponding to the candidate operator and the context interaction vector of each drug. Each layer of the multi-layer graph neural network is configured with a set of candidate message passing operators, and different candidate operators in the set of candidate message passing operators correspond to different graph information propagation methods. Based on each candidate operator associated with the multilayer graph neural network and the weight of each candidate operator for each drug, the molecular representation of the featured molecular graph of each drug is iteratively updated, and the molecular representation obtained by the final update of the featured molecular graph of each drug is used as the molecular embedding representation of each drug.

[0082] Optionally, the molecular representation update process performed by the l-th layer graph neural network is as follows:

[0083]

[0084] in, This represents the molecular representation of drug d output by the l-th layer graph neural network. This represents the molecular representation of drug d as input to the l-th layer graph neural network. The molecular representation of the initial drug d. The features of atomic node v in the molecular diagram representing drug d. This represents the i-th candidate operator associated with the l-th layer of the graph neural network. For the weights of drug d, m represents the number of candidate operators in the set of candidate message passing operators configured in the l-th layer graph neural network. This represents the context interaction vector of drug d. , Let represent the pre-learned latent vectors corresponding to the i-th and j-th candidate operators associated with the l-th layer graph neural network.

[0085] Optionally, in the loss function used during the model training phase, the loss term associated with the pairwise attention map neural architecture search module may be included. It is expressed as follows:

[0086] in, , Let represent the learnable latent vectors corresponding to the i-th and j-th candidate operators in the candidate message passing operator set configured in the l-th layer graph neural network, and m represent the number of candidate operators in the candidate message passing operator set configured in the l-th layer graph neural network.

[0087] Optionally, the characteristic molecular map of each drug is determined by the following formula:

[0088] in, The features of atomic node v in the molecular diagram representing drug d. Atom nodes in the drug molecule diagram representing drug d Features This represents the cell context corresponding to cell line c. This represents the set of atomic nodes in the drug molecule diagram of drug d.

[0089] Optionally, the synergistic effect prediction results are generated through the following steps: The molecular embedding representation of each drug is mapped to the language model input space by a drug projector to obtain the drug structure representation of each drug. The cell line text descriptions and drug SMILES sequences for each drug were converted into text lexical units. By inputting the text terms describing the cell line, the drug structure representation of each drug, and the text terms of the drug SMILES sequence into the biomedical big language model, the synergistic effect prediction results of each drug are obtained.

[0090] Optionally, in the loss function used during the model training phase, the loss term associated with the DrugSyn-LLM module for retrieving enhanced biomedical instructions may be included. It is expressed as follows:

[0091] in, This represents the loss term associated with the fine-tuning phase of retrieval-enhanced biomedical instructions, used to guide the large language model in learning external biomedical knowledge. This represents the loss term associated with the training phase of the drug synergy prediction task, used to guide the large language model to complete the prediction of drug synergy in a structured generation task manner. This indicates a stage indicator variable, and The value is 1 during the refinement phase of the enhanced biomedical instruction and 0 during the training phase of the drug co-prediction task.

[0092] Optionally, the loss term associated with the retrieval enhancement biomedical instruction fine-tuning phase and the drug co-prediction task training phase. and It is expressed as follows:

[0093]

[0094] Where R represents the target generated response. , This indicates input prompts for the refinement phase of enhanced biomedical instructions and the training phase of the drug co-prediction task. Text terms representing cell line descriptions Text terms representing the drug SMILES sequence of drugs d1 and d2. Indicates the target representation, This represents the drug structure representation of d1 and d2.

[0095] As can be seen from the above technical solutions, the target-adaptive decoupled molecular graph encoding module utilizes multi-perspective encoding branches constructed for different chemical structural relationships and semantic effects to obtain richer molecular graph representations. It further divides these representations into target-independent and target-related molecular representations, enabling the model to explicitly distinguish between the inherent structural attributes of drugs and target-related functional attributes, thereby enhancing the model's adaptability in different contexts. The multi-level contextual cell feature alignment model simultaneously injects gene expression features into the molecular graph structure encoding process and the semantic input space of the large language model, allowing the cellular environment to participate in the message passing process within the molecular graph. The pairwise attention graph neural architecture search module dynamically searches for and combines graph neural network structures suitable for the current drug combination to accurately obtain molecular embedding representations. The DrugSyn-LLM module, which enhances biomedical instruction fine-tuning through retrieval, transforms drug synergy prediction into a structured generation task for the large language model, enabling the model to simultaneously utilize multimodal information such as cell line text descriptions, target representations, individual drug molecular embedding representations, and drug SMILES sequences for prediction. This improves the model's generalization ability in out-of-distribution drug synergy prediction scenarios, thereby enhancing prediction accuracy in such scenarios.

[0096] Optionally, this application embodiment also provides an electronic device, including a processor 110, a memory 109, and a program or instructions stored in the memory 109 and executable on the processor 110. When the program or instructions are executed by the processor 110, they implement the various processes of the above-described off-distribution generalization drug synergy prediction method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0097] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0098] Figure 3 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of this application.

[0099] The electronic device 100 includes, but is not limited to, components such as: radio frequency unit 101, network module 102, audio output unit 103, input unit 104, sensor 105, display unit 106, user input unit 107, interface unit 108, memory 109, and processor 110.

[0100] Those skilled in the art will understand that the electronic device 100 may also include a power supply (such as a battery) for supplying power to various components. The power supply may be logically connected to the processor 110 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 3 The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.

[0101] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described method for predicting the synergistic effect of out-of-distribution generalized drugs and achieve the same technical effect. To avoid repetition, these will not be described again here.

[0102] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0103] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described embodiments of the method for predicting the synergistic effect of out-of-distribution generalized drugs, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0104] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0105] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0106] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0107] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A large-scale model for predicting the synergistic effects of out-of-distribution generalized drugs, characterized in that, The model includes: The target-adaptive decoupled molecular graph encoding module is configured to, in response to receiving drug molecular graphs and target protein information for each drug, process the drug molecular graph of each drug into a molecular graph representation through multi-view encoding branches constructed for different chemical structural relationships and semantic functions; decouple the molecular graph representation of each drug into a target-independent molecular representation and a target-related molecular representation, wherein the target-independent molecular representation is used to characterize the intrinsic structural properties of the drug, and the target-related molecular representation is used to characterize the target-related functional properties of the drug; determine the target-adaptive molecular representation of each drug under each target protein associated with each drug based on the target protein information and the target-related molecular representation of each drug; and determine the target-adaptive associated representation of each drug based on the target-adaptive molecular representation of each drug associated with each drug and the target-independent molecular representation of each drug, and input it into the pairwise attention graph neural architecture search module. A multi-level contextualized cell feature alignment module is configured to, in response to receiving cell line gene expression features and drug molecule maps of each drug, add the cell line gene expression features to the drug molecule map of each drug to obtain a feature-enhanced molecular map of each drug; map the cell line gene expression features to a continuous language space representation as a target representation; input the feature-enhanced molecular maps of each drug to a pairwise attention map neural architecture search module, and input the target representation to a DrugSyn-LLM module for retrieving enhanced biomedical instruction fine-tuning; The pairwise attention graph neural architecture search module is configured to select a graph neural network operator based on the pairwise attention results of the adaptive association representation of the target of each drug, and use the selected graph neural network operator to process the featured molecular graph of each drug into a molecular embedding representation and input it into the DrugSyn-LLM module for retrieval enhancement and fine-tuning of biomedical instructions. The DrugSyn-LLM module, which enhances the fine-tuning of biomedical instructions for retrieval, is configured to generate synergistic effect predictions for each drug based on the received cell line text description, the target representation, the molecular embedding representation of each drug, and the drug SMILES sequence.

2. The model according to claim 1, characterized in that, The target-adaptive decoupling molecular graph encoding module is also configured to perform the following steps: Using a pre-trained protein encoder, the target protein information of each drug is processed into target protein representations under each target protein associated with each drug. By interacting with the target protein representations associated with each drug and the target-related molecular representations of each drug through a cross-attention mechanism, adaptive molecular representations of the target proteins associated with each drug are obtained.

3. The model according to claim 1, characterized in that, In the loss function used during model training, the loss term associated with the target adaptive decoupling molecular graph encoding module... It is expressed as follows: in, express and The corresponding correlation matrix, and This indicates the different target conditions k of the two drugs. The decoupling representation composed of the K target site adaptive molecular representations obtained below, ||·|| F This represents the operation for calculating the F-norm.

4. The model according to claim 1, characterized in that, The pairwise attention graph neural architecture search module is also configured to perform the following steps: The multi-head attention mechanism is used to process the adaptive association representation of each drug's target, and the context interaction vector of each drug is used as the pairwise attention result. For each candidate operator associated with a pre-constructed multi-layer graph neural network, the weight of the candidate operator for each drug is determined based on the pre-learned latent vector corresponding to the candidate operator and the context interaction vector of each drug. Each layer of the multi-layer graph neural network is configured with a set of candidate message passing operators, and different candidate operators in the set of candidate message passing operators correspond to different graph information propagation methods. Based on each candidate operator associated with the multilayer graph neural network and the weight of each candidate operator for each drug, the molecular representation of the featured molecular graph of each drug is iteratively updated, and the molecular representation obtained by the final update of the featured molecular graph of each drug is used as the molecular embedding representation of each drug.

5. The model according to claim 4, characterized in that, The molecular representation update process performed by the l-th layer graph neural network is as follows: in, This represents the molecular representation of drug d output by the l-th layer graph neural network. This represents the molecular representation of drug d as input to the l-th layer graph neural network. The molecular representation of the initial drug d. The features of atomic node v in the molecular diagram representing drug d. This represents the i-th candidate operator associated with the l-th layer of the graph neural network. For the weights of drug d, m represents the number of candidate operators in the set of candidate message passing operators configured in the l-th layer graph neural network. This represents the context interaction vector of drug d. , Let represent the pre-learned latent vectors corresponding to the i-th and j-th candidate operators associated with the l-th layer graph neural network.

6. The model according to claim 4, characterized in that, In the loss function used during model training, the loss term associated with the pairwise attention map neural architecture search module... It is expressed as follows: in, , Let represent the learnable latent vectors corresponding to the i-th and j-th candidate operators in the candidate message passing operator set configured in the l-th layer graph neural network, and m represent the number of candidate operators in the candidate message passing operator set configured in the l-th layer graph neural network.

7. The model according to claim 1, characterized in that, The characteristic molecular map of each drug is determined by the following formula: in, The features of atomic node v in the molecular diagram representing drug d. Atom nodes in the drug molecule diagram representing drug d Features This represents the cell context corresponding to cell line c. This represents the set of atomic nodes in the drug molecule diagram of drug d.

8. The model according to any one of claims 1-7, characterized in that, The DrugSyn-LLM module, which enhances the retrieval of biomedical command fine-tuning, is also configured to perform the following steps: The molecular embedding representation of each drug is mapped to the language model input space by a drug projector to obtain the drug structure representation of each drug. The cell line text descriptions and drug SMILES sequences for each drug were converted into text lexical units. By inputting the text terms describing the cell line, the drug structure representation of each drug, and the text terms of the drug SMILES sequence into the biomedical big language model, the synergistic effect prediction results of each drug are obtained.

9. The model according to claim 8, characterized in that, In the loss function used during model training, the loss term associated with the DrugSyn-LLM module for retrieving enhanced biomedical instructions is... It is expressed as follows: in, This represents the loss term associated with the fine-tuning phase of retrieval-enhanced biomedical instructions, used to guide the large language model in learning external biomedical knowledge. This represents the loss term associated with the training phase of the drug synergy prediction task, used to guide the large language model to complete the prediction of drug synergy in a structured generation task manner. This indicates a stage indicator variable, and The value is 1 during the refinement phase of the enhanced biomedical instruction and 0 during the training phase of the drug co-prediction task.

10. The model according to claim 8, characterized in that, The loss term associated with the retrieval enhancement biomedical instruction fine-tuning phase and the drug co-prediction task training phase. and It is expressed as follows: Where R represents the target generated response. , This indicates input prompts for the refinement phase of enhanced biomedical instructions and the training phase of the drug co-prediction task. Text terms representing cell line descriptions Text terms representing the drug SMILES sequence of drugs d1 and d2. Indicates the target representation, This represents the drug structure representation of d1 and d2.

Images