A molecular multi-modal data detection method and system based on cross-modal knowledge distillation, a terminal and a storage medium

Abstract

The application relates to the technical field of data processing, and discloses a molecular multi-modal data detection method, system, terminal and storage medium based on cross-modal knowledge distillation, which comprises the following steps: acquiring molecular multi-modal data, constructing a multi-modal teacher model according to text data, sequence data and structure data of the molecular multi-modal data, and performing regression prediction training to obtain a target multi-modal teacher model; constructing a lightweight student model according to wild-type sequences and mutant sequences of the molecular multi-modal data, migrating parameters of the target multi-modal teacher model to the lightweight student model to obtain a target lightweight student model; inputting a test set and an independent test set of to-be-detected multi-modal data into the target lightweight student model, evaluating and testing the test set and the independent test set, and obtaining a molecular interaction prediction result. The application realizes high-precision and robust detection of molecular interaction based on multi-modal teacher and student model distillation and parameter migration according to an alignment strategy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system, terminal, and computer-readable storage medium for detecting molecular multimodal data based on cross-modal knowledge distillation. Background Technology

[0002] Determining intermolecular interactions through experimental methods is typically costly, time-consuming, and low-throughput, and is easily affected by differences in experimental conditions, sample preparation, and detection platforms. Faced with a large number of candidate molecules, protein mutations, or antibody modification spaces, relying solely on experimental verification one by one is insufficient to meet the needs of rapid screening and rational design. Therefore, computational methods can quickly predict and rank intermolecular binding ability, binding free energy changes, and favorable mutations before experiments, thereby narrowing the scope of experimental verification, reducing R&D costs, and improving design efficiency. Computational methods have largely solved the problems of high cost and long time consumption in experimental determination. However, traditional computational methods rely on physical energy functions, empirical features, and single structures for modeling, resulting in low accuracy, generalization, and computational efficiency in predicting intermolecular interactions.

[0003] Traditional molecular interaction prediction methods (such as empirical calculation methods based on physical energy functions, empirical features, and single structures) are limited to a single modality in terms of optimization granularity, cannot fully integrate textual semantics and three-dimensional structural information during the training phase, and introduce high time and economic costs due to over-reliance on experimental determination. As a result, they are difficult to achieve efficient, accurate, and robust molecular detection in scenarios such as large-scale candidate molecule screening, protein mutation analysis, and antibody modification, which has become an urgent problem to be solved.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for molecular multimodal data detection based on cross-modal knowledge distillation, aiming to solve the problem that it is difficult to achieve efficient, accurate, and robust molecular detection in scenarios such as large-scale candidate molecule screening, protein mutation analysis, and antibody modification in the prior art.

[0006] To achieve the above objectives, this invention provides a method for detecting molecular multimodal data based on cross-modal knowledge distillation, the method comprising the following steps: Acquire molecular multimodal data, construct a multimodal teacher model based on the text data, sequence data and structural data of the molecular multimodal data, and train the multimodal teacher model by regression prediction to obtain the target multimodal teacher model; A lightweight student model is constructed based on the wild-type and mutant sequences of the molecular multimodal data. The parameters of the target multimodal teacher model are then transferred to the lightweight student model to obtain the target lightweight student model. Acquire the multimodal data to be tested, input the test set and independent test set of the multimodal data to be tested into the target lightweight student model, the target lightweight student model evaluates and tests the test set and the independent test set to obtain a first prediction result and a second prediction result, and compare the first prediction result and the second prediction result to obtain the molecular interaction detection result.

[0007] Optionally, the molecular multimodal data detection method based on cross-modal knowledge distillation, wherein acquiring molecular multimodal data, constructing a multimodal teacher model based on the text data, sequence data, and structural data of the molecular multimodal data, and performing regression prediction training on the multimodal teacher model to obtain a target multimodal teacher model, specifically includes: Molecular multimodal data is acquired, and the text data, sequence data, and structural data of the molecular multimodal data are standardized to obtain processed multimodal data. A multimodal teacher model is constructed based on the processed multimodal data. The processed multimodal data is divided according to a preset ratio to obtain a training set and a validation set; The multimodal teacher model is trained using regression prediction based on the training set to obtain the trained multimodal teacher model. The trained multimodal teacher model is optimized based on the model parameters of the validation set to obtain the target multimodal teacher model.

[0008] Optionally, the molecular multimodal data detection method based on cross-modal knowledge distillation, wherein acquiring molecular multimodal data, standardizing the text data, sequence data, and structural data of the molecular multimodal data to obtain processed multimodal data, and constructing a multimodal teacher model based on the processed multimodal data, specifically includes: Acquire molecular multimodal data, and convert the text data, sequence data, and structural data into standard format information according to a standard format. The standard format information is removed based on the free energy change labels of the molecular multimodal data to obtain the target dataset; The target dataset is subjected to sample equalization processing according to the undersampling strategy to obtain processed multimodal data; The processed multimodal data is encoded by a multimodal encoder of a large language model to obtain multiple multimodal feature vectors. All the multimodal feature vectors are concatenated to obtain multimodal fusion features. The multimodal fusion features are downsampled in dimension to obtain multiple structural knowledge features. All the structural knowledge features are then used to construct a multimodal teacher model.

[0009] Optionally, the molecular multimodal data detection method based on cross-modal knowledge distillation, wherein the process of encoding the processed multimodal data using a multimodal encoder of a large language model to obtain multiple multimodal feature vectors, and concatenating all the multimodal feature vectors to obtain multimodal fusion features, specifically includes: The processed multimodal data is text-encoded using a text encoder based on a large language model to obtain a text semantic feature vector. The processed multimodal data is sequence encoded using a sequence encoder based on a large language model to obtain a sequence feature vector. The processed multimodal data is structurally encoded using a structural encoder based on a large language model to obtain a three-dimensional structural feature vector. The text semantic feature vector, sequence feature vector, and three-dimensional structural feature vector are concatenated to obtain cross-modal features. The cross-modal features are then linearly mapped to obtain a query matrix, a key matrix, and a value matrix. Multimodal fusion features are obtained by performing residual connections and normalization layer transformations on the query matrix, the key matrix, and the value matrix through a feedforward network.

[0010] Optionally, the molecular multimodal data detection method based on cross-modal knowledge distillation, wherein the step of constructing a lightweight student model based on the wild-type and mutant sequences of the molecular multimodal data, and transferring the parameters of the target multimodal teacher model to the lightweight student model to obtain the target lightweight student model, specifically includes: Wild-type and mutant sequences of molecular multimodal data are obtained, and the wild-type and mutant sequences are encoded by a sequence encoder to obtain sequence encoding output features. A lightweight student model is constructed based on the sequence encoding output features. Based on representation distillation, attention distillation, and structural knowledge distillation, the intermediate representations, attention maps, and structural space relations of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

[0011] Optionally, the molecular multimodal data detection method based on cross-modal knowledge distillation, wherein acquiring the wild-type and mutant sequences of the molecular multimodal data, encoding the wild-type and mutant sequences using a sequence encoder to obtain sequence encoding output features, and constructing a lightweight student model based on the sequence encoding output features, specifically includes: Wild-type and mutant sequences of molecular multimodal data are obtained, and the wild-type and mutant sequences are encoded by a sequence encoder to obtain sequence-encoded output features; Obtain the position code, perform residual aggregation on the sequence code output feature and the position code to obtain the normalized feature; Obtain the feature mapping matrix, and map the normalized features according to the feature mapping matrix to obtain the lightweight student model.

[0012] Optionally, the molecular multimodal data detection method based on cross-modal knowledge distillation, wherein the transfer of intermediate representations, attention maps, and structural spatial relationships of the target multimodal teacher model to the lightweight student model based on representation distillation, attention distillation, and structural knowledge distillation to obtain the target lightweight student model specifically includes: The multimodal fusion features are aligned using the alignment projection layer of the target multimodal teacher model to obtain representation distillation, attention distillation, and structural knowledge distillation. Obtain the projection matrix, and project the representation distillation, the attention distillation, and the structural knowledge distillation according to the projection matrix to obtain the target representation distillation, the target attention distillation, and the target structural knowledge distillation. Based on the target representation distillation, target attention distillation, and target structural knowledge distillation, the intermediate representations, attention maps, and structural space relationships of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a molecular multimodal data detection system based on cross-modal knowledge distillation, wherein the molecular multimodal data detection system based on cross-modal knowledge distillation: The model parameter training module is used to acquire molecular multimodal data, construct a multimodal teacher model based on the text data, sequence data and structural data of the molecular multimodal data, and perform regression prediction training on the multimodal teacher model to obtain the target multimodal teacher model; The model data transfer module is used to construct a lightweight student model based on the wild-type and mutant sequences of the molecular multimodal data, and to transfer the parameters of the target multimodal teacher model to the lightweight student model to obtain the target lightweight student model. The molecular data detection module is used to acquire the multimodal data to be tested, input the test set and independent test set of the multimodal data to be tested into the target lightweight student model, the target lightweight student model evaluates and tests the test set and the independent test set to obtain a first prediction result and a second prediction result, and compares the first prediction result and the second prediction result to obtain the molecular interaction detection result.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a molecular multimodal data detection program based on cross-modal knowledge distillation, and when the molecular multimodal data detection program based on cross-modal knowledge distillation is executed by a processor, it implements the steps of the molecular multimodal data detection method based on cross-modal knowledge distillation as described above.

[0015] This invention acquires molecular multimodal data, constructs a multimodal teacher model based on the text, sequence, and structural data of the molecular multimodal data, and trains the multimodal teacher model through regression prediction to obtain a target multimodal teacher model. A lightweight student model is constructed based on the wild-type and mutant sequences of the molecular multimodal data, and the parameters of the target multimodal teacher model are transferred to the lightweight student model to obtain a target lightweight student model. Test multimodal data is acquired, and the test set and independent test sets of the test multimodal data are input into the target lightweight student model. The target lightweight student model evaluates and tests the test set and the independent test sets to obtain a first prediction result and a second prediction result. The first prediction result and the second prediction result are compared to obtain the molecular interaction prediction result. This invention is based on multimodal teacher-student model distillation and transfers parameters according to a cross-modal knowledge alignment strategy to achieve high-precision and robust detection of molecular interactions. Attached Figure Description

[0016] Figure 1 This is a flowchart of a preferred embodiment of the molecular multimodal data detection method based on cross-modal knowledge distillation of the present invention; Figure 2 This is a schematic diagram illustrating molecular interaction prediction, representing a preferred embodiment of the molecular multimodal data detection method based on cross-modal knowledge distillation according to the present invention. Figure 3 This is a preferred embodiment of the molecular multimodal data detection method based on cross-modal knowledge distillation of the present invention, showing the sampling process before and after sampling. A schematic diagram of the G distribution; Figure 4 This is a schematic diagram of the molecular interaction prediction architecture based on cross-modal knowledge distillation, a preferred embodiment of the molecular multimodal data detection method based on cross-modal knowledge distillation of the present invention. Figure 5 This is a structural diagram of a preferred embodiment of the molecular multimodal data detection system based on cross-modal knowledge distillation of the present invention; Figure 6 This is a structural diagram of a preferred embodiment of the terminal of the device of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0018] Traditional molecular interaction prediction methods (such as empirical calculation methods based on physical energy functions, empirical features, and single structures) are limited to a single modality in terms of optimization granularity, cannot fully integrate textual semantics and three-dimensional structural information during the training phase, and introduce high time and economic costs due to over-reliance on experimental determination. As a result, they are difficult to achieve efficient, accurate, and robust molecular detection in scenarios such as large-scale candidate molecule screening, protein mutation analysis, and antibody modification. Therefore, there is a need for a molecular multimodal data detection method based on cross-modal knowledge distillation. This method is based on multimodal teacher and student model distillation and parameter transfer according to a cross-modal knowledge alignment strategy to achieve high-precision and robust detection of molecular interactions.

[0019] The preferred embodiment of the present invention describes a molecular multimodal data detection method based on cross-modal knowledge distillation, such as... Figure 1 and Figure 2 As shown, the molecular multimodal data detection method based on cross-modal knowledge distillation includes the following steps: Step S10: Obtain molecular multimodal data, construct a multimodal teacher model based on the text data, sequence data and structural data of the molecular multimodal data, and perform regression prediction training on the multimodal teacher model to obtain the target multimodal teacher model.

[0020] Step S10 includes: Step S11: Obtain molecular multimodal data, standardize the text data, sequence data and structural data of the molecular multimodal data to obtain processed multimodal data, and construct a multimodal teacher model based on the processed multimodal data; Step S12: Divide the processed multimodal data according to a preset ratio to obtain a training set and a validation set; Step S13: Perform regression prediction training on the multimodal teacher model based on the training set to obtain the trained multimodal teacher model; Step S14: Optimize the trained multimodal teacher model based on the model parameters of the validation set to obtain the target multimodal teacher model.

[0021] Specifically, molecular multimodal data is acquired, including text data (antigen-antibody complex ID: for example, the PDB ID corresponding to a certain antigen-antibody complex is 1MHP; antigen chain and antibody chain identifiers: for example, the antigen chain identifier is A, and the antibody chain identifier is H; mutation information: for example, the sixth amino acid of the A chain of the antigen-antibody complex is mutated from G to N, which can be represented as A:G6N; protein function information: for example, the complex 1MHP acts as a cell surface receptor, performing physiological functions related to neurite growth, neuronal adhesion, and axonogenesis on the surface of neurons), sequence data (wild-type antigen or antibody sequence: for example, the wild-type antigen sequence is PDVDLGDISGINAS, and the antibody sequence is MVPRRPASLEVTVACIWLLT; context window of the mutation site; used to extract local sequence features near the mutation, represented as...) ,in The mutation site The window radius is usually taken as... The structural data (corresponding 3D complex structure; the 3D structure of the complex is obtained from the PDB and preprocessed to remove water molecules, ions, or unnecessary ligands) are standardized to obtain processed multimodal data. A multimodal teacher model is constructed based on the processed multimodal data. The processed multimodal data is divided according to a preset ratio to obtain a training set and a validation set (the total number of training epochs for the teacher model is defined as Epoch=30; before the start of each training epoch, the training set is...). All samples are randomly shuffled to avoid the sample order affecting model training and to improve the stability and generalization ability of the model training. That is, all samples after sampling are randomly shuffled and divided into training sets in an 8:1:1 ratio. Validation set and test set Each contains the number of samples. The multimodal teacher model is trained using regression prediction based on the training set to obtain a trained multimodal teacher model. The trained multimodal teacher model is then optimized based on the model parameters from the validation set to obtain a target multimodal teacher model. After each epoch of training, the validation set is... Input the teacher model to obtain the validation set predictions and the Mean Squared Error (MSE). During training, select the model parameters with the lowest MSE on the validation set as the optimal teacher model parameters. If the MSE on the validation set no longer decreases after 5 consecutive epochs, an early stopping strategy is used to stop training in order to prevent overfitting.

[0022] In this embodiment, the regression prediction training process is as follows: the batch training size BS = 64 is defined for each Epoch, and the total number of iterations for each Epoch is: In each iteration, from the training set Read a batch of samples ,in Tags representing changes in binding free energy before and after mutation The teacher model predicts for each sample The value is represented as Define the loss function for the teacher model; the main training objective is to make the predicted values ​​close to the true labels, therefore, the supervised regression loss for each batch is defined as the mean squared error loss (MSE): The teacher model parameters are updated using the AdamW optimizer, with an initial learning rate of [missing information]. The weight decay coefficient is set to After 30 training epochs, the learning rate gradually decreases to a minimum through linear decay: To prevent abnormally large gradients, a gradient clipping threshold is set. This is to improve the stability of model training.

[0023] like Figure 2 As shown, firstly, a multimodal dataset is constructed, containing text, sequence, and structural information, and the data is preprocessed. Then, a multimodal teacher model is defined, using different encoders to encode the multimodal data, and multimodal feature fusion is achieved through a cross-modal attention mechanism. Next, a loss function for the teacher model is constructed, and an early stopping strategy is used to train the teacher model, resulting in a trained multimodal teacher model. Subsequently, a lightweight student model is constructed, which uses only sequence representations and fuses wild-type and mutant sequence features. Based on this, cross-modal knowledge distillation is performed, constructing a cross-modal knowledge distillation loss function to transfer the knowledge from the trained multimodal teacher model to the lightweight student model, completing the training of the student model. Finally, prediction performance is evaluated, constructing metrics such as MSE, PCC, speed, stability, and robustness to comprehensively evaluate the trained student model, obtaining the final molecular interaction prediction results.

[0024] Step S11 includes: Step S111: Obtain molecular multimodal data, and convert the text data, sequence data and structural data according to the standard format to obtain standard format information; Step S112: Based on the free energy change labels of the molecular multimodal data, the standard format information is removed to obtain the target dataset; Step S113: Perform sample equalization processing on the target dataset according to the undersampling strategy to obtain processed multimodal data; Step S114: Encode the processed multimodal data using a multimodal encoder of a large language model to obtain multiple multimodal feature vectors, and concatenate all the multimodal feature vectors to obtain multimodal fusion features; Step S115: Perform dimensional downsampling on the multimodal fusion features to obtain multiple structural knowledge features, and construct all the structural knowledge features to obtain a multimodal teacher model.

[0025] Specifically, molecular multimodal data is acquired, and the text data, sequence data, and structural data are format-converted according to a standard format to obtain standard format information (converting mutation descriptions from different sources into a standard format of chain ID: wild-type amino acid, site number, and mutant amino acid, e.g., H:Y33A). The standard format information is then processed by removing data based on the free energy change labels of the molecular multimodal data to obtain the target dataset. (Subsequently, the consistency between the mutation site and the wild-type sequence and the residue numbering of the three-dimensional structure is checked; if the mutation site cannot be matched in the sequence and structure, the sample is deleted or the numbering is corrected.) Labels were standardized in units, outliers were removed, and finally duplicate samples were removed. For duplicate records of the same complex, the same mutation site, and the same mutation type, if... If the labels are the same, keep one sample; if the label differences exceed a preset threshold, such as... If the sample is abnormal, it is marked as an abnormal sample and removed, resulting in a cleaned antigen-antibody mutation sample. Dataset, total number of samples in the dataset The target dataset is subjected to sample balancing processing according to an undersampling strategy to obtain processed multimodal data (the obtained dataset undergoes data sampling and sample balancing processing to reduce data distribution bias and improve the model's performance on different mutation types, different complexes, and different...). (Generalization ability of intervals) The processed multimodal data is encoded by a multimodal encoder of a large language model to obtain multiple multimodal feature vectors. All the multimodal feature vectors are concatenated to obtain multimodal fusion features. The multimodal fusion features are downsampled in dimension to obtain multiple structural knowledge features. All the structural knowledge features are constructed to obtain a multimodal teacher model.

[0026] In this embodiment, as Figure 3 As shown, the data sample distribution before and after sampling is based on... The numerical values ​​divide the samples into enhanced binding samples, weakened binding samples, and samples with minimal impact. <0 indicates that the mutation may enhance binding. A value >0 indicates that the mutation may weaken the binding. ≈0 indicates a small impact from mutations; subsequently, stratified sampling is performed on samples from different categories and from different complex sources to avoid excessive bias towards high-frequency complexes, high-frequency mutation sites, or a particular class during model training. Distribution range; for mutation types with a small sample size or For each interval, oversampling, data augmentation, or weighted sampling strategies are employed; for categories with an excessive number of samples, undersampling strategies are used to maintain a relatively balanced distribution of training data.

[0027] Step S114 includes: Step S1141: The processed multimodal data is text encoded using a text encoder of a large language model to obtain a text semantic feature vector; Step S1142: Sequence encoding is performed on the processed multimodal data using a sequence encoder of a large language model to obtain a sequence feature vector; Step S1143: Perform structural encoding on the processed multimodal data using the structural encoder of the large language model to obtain a three-dimensional structural feature vector; Step S1144: Concatenate the text semantic feature vector, sequence feature vector and the three-dimensional structural feature vector to obtain cross-modal features, and perform linear mapping on the cross-modal features to obtain query matrix, key matrix and value matrix; Step S1145: Perform residual connection and normalization layer transformation on the query matrix, the key matrix and the value matrix through a feedforward network to obtain multimodal fusion features.

[0028] Specifically, the processed multimodal data is text-encoded using a text encoder of a large language model to obtain text semantic feature vectors (samples of each processed multimodal data). ,in This represents the combined text description of all textual content, including complex number, functional information, and mutation information. These represent the antigen-antibody sequences before and after the mutation. These represent the three-dimensional structures of the antigen and antibody before and after the mutation, respectively. Encoding is performed using a life science text encoder (such as BioGPT) or a large language model to obtain the text semantic feature vector, i.e. The processed multimodal data is sequence encoded using a sequence encoder in a large language model to obtain sequence feature vectors (using protein sequence characterization models such as ESM2 to encode the antigen-antibody sequences before and after mutation, i.e.) The processed multimodal data is then structure-encoded using a structural encoder of a large language model to obtain a three-dimensional structural feature vector (using graph neural networks such as GAT or GNN for structural encoding, i.e., ...). The three-dimensional structural feature vector is obtained by concatenating the text semantic feature vector, the sequence feature vector, and the three-dimensional structural feature vector to obtain cross-modal features. The cross-modal features are linearly mapped to obtain a query matrix, a key matrix, and a value matrix. The query matrix, the key matrix, and the value matrix are then subjected to residual connections and normalization layer transformations through a feedforward network to obtain multimodal fusion features.

[0029] In this embodiment, the multimodal features (the text semantic feature vector, the sequence feature vector, and the three-dimensional structural feature vector) are first... The multimodal feature calculation formula is as follows: ; in, , representing the total number of tokens represented by each sample in the teacher model. Construct a cross-modal attention mask matrix for the feature vector dimension of the teacher model. This mask matrix is ​​used to control the attention connection relationship between different tokens, thereby achieving gated fusion of different modal features.

[0030] Cross-modal features Perform a linear mapping to obtain the query matrix. Key matrix Sum matrix : ; in, The learnable parameter matrix is ​​then used to compute a cross-modal attention map using Mask Attention. : ; To ensure feature stability, a feedforward network (FFN), residual connections, and normalization layer transformation are introduced to ultimately obtain the multimodal fusion features of Stage 1. The formula for calculating multimodal fusion features is as follows: ; Furthermore, in order to realize a multimodal teacher model The prediction is that token-dimensional downsampling is achieved by gradually expanding the receptive field. Taking Stage 1 to Stage 2 as an example, the token-dimensional downsampling will be implemented. The features of two adjacent tokens are concatenated along their feature dimensions to obtain the result. The structural knowledge features of Stage 2 are obtained through nonlinear transformation. The formula for calculating structural knowledge features is as follows: ; in, Representing nonlinear transformations, we can select functions such as convolution functions, self-attention functions, feedforward neural networks with activation functions, and so on, to obtain features respectively. Finally passed Mapping matrix and pooling operations, etc., to obtain Predicted value The formula for calculating the predicted value is as follows: ; Step S20: Construct a lightweight student model based on the wild-type and mutant sequences of the molecular multimodal data, and transfer the parameters of the target multimodal teacher model to the lightweight student model to obtain the target lightweight student model.

[0031] Step S20 includes: Step S21: Obtain wild-type and mutant sequences of molecular multimodal data, encode the wild-type and mutant sequences using a sequence encoder to obtain sequence encoding output features, and construct a lightweight student model based on the sequence encoding output features; Step S22: Based on representation distillation, attention distillation, and structural knowledge distillation, the intermediate representations, attention maps, and structural space relations of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

[0032] Specifically, wild-type and mutant sequences of molecular multimodal data are acquired, and the wild-type and mutant sequences are encoded using a sequence encoder to obtain sequence encoding output features. A lightweight student model is then constructed based on these sequence encoding output features. The student model uses fewer inputs than the teacher model, employing only wild-type and mutant sequences. It does not rely on complete 3D structural or functional description information and utilizes a lightweight sequence encoder. For example, parameter-trimmed ESM, lightweight Transformer, CNN-Transformer, or shallow protein language model), based on representation distillation, attention distillation, and structural knowledge distillation, the intermediate representations, attention maps, and structural space relations of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

[0033] Step S21 includes: Step S211: Obtain the wild-type and mutant sequences of molecular multimodal data, and encode the wild-type and mutant sequences using a sequence encoder to obtain sequence encoding output features; Step S212: Obtain the position code, perform residual aggregation on the sequence code output feature and the position code to obtain the normalized feature; Step S213: Obtain the feature mapping matrix, and map the normalized features according to the feature mapping matrix to obtain the lightweight student model.

[0034] Specifically, wild-type and mutant sequences of molecular multimodal data are acquired, and the wild-type and mutant sequences are encoded using a sequence encoder to obtain sequence-encoded output features (first adding position encoding). ： The positional encoding is obtained, and residual aggregation is performed on the sequence encoding output features and the positional encoding to obtain normalized features (residual aggregation yields layer normalized features). ： ), obtain the feature mapping matrix, and map the normalized features according to the feature mapping matrix to obtain the lightweight student model (through the feature mapping matrix). Mapping the student model's representation to the same token dimension as the teacher model's representation, i.e., the student model representation. ).

[0035] In this embodiment, the input of the student model Sequence encoding output features : ; ; ; in, point Characterization of wild-type and mutant sequences, The dimension of the feature representation vector for the student model is typically much smaller than the dimension of the representation vector for the teacher model described in the steps above. ,Right now To achieve lightweight student models, superscript Represents the relevant features or variables of the student model.

[0036] Furthermore, a prediction head for the student model is constructed, and through token aggregation and feature mapping, the features of each stage of the student model are finally obtained. and prediction results .

[0037] Step S22 includes: Step S221: Align the multimodal fusion features through the alignment projection layer of the target multimodal teacher model to obtain representation distillation, attention distillation and structural knowledge distillation; Step S222: Obtain the projection matrix, and project the representation distillation, the attention distillation, and the structural knowledge distillation according to the projection matrix to obtain the target representation distillation, the target attention distillation, and the target structural knowledge distillation; Step S223: Based on the target representation distillation, the target attention distillation, and the target structural knowledge distillation, the intermediate representations, attention maps, and structural space relationships of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

[0038] Specifically, the multimodal fusion features are aligned using the alignment projection layer of the target multimodal teacher model to obtain representation distillation, attention distillation, and structural knowledge distillation (firstly, a dimension alignment projection layer needs to be constructed for each stage to achieve feature alignment, and then...). There are stages, and the projection matrix is ​​defined as follows: The parameters of the projection matrix are trainable. After projection, the teacher model features are consistent with the student model in both the token and feature dimensions; that is, the aligned representation of the student model is... The process involves obtaining a projection matrix, projecting the representation distillation, attention distillation, and structural knowledge distillation onto the projection matrix to obtain target representation distillation, target attention distillation, and target structural knowledge distillation. Based on the target representation distillation, target attention distillation, and target structural knowledge distillation, the intermediate representations, attention maps, and structural space relations of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model. (The intermediate representations of the target multimodal teacher model are transferred according to the target representation distillation to obtain target intermediate representations; the attention maps of the target multimodal teacher model are transferred according to the target attention distillation to obtain target attention maps; the structural space relations of the target multimodal teacher model are transferred according to the structural space relations to obtain target structural space relations; and the lightweight student model is updated according to the target intermediate representations, target attention maps, and target structural space relations to obtain the target lightweight student model.)

[0039] In this embodiment, a cross-modal knowledge distillation loss function is defined, which includes characterization distillation (Stage 1). Attention distillation (Stage 2) Structural knowledge distillation (Stage 3, Stage 4) Output distillation The various distillation loss functions are expressed as follows: ; ;

[0040] .

[0041] Furthermore, the student model is jointly optimized, and the teacher model parameters are obtained from the model. And the parameters remain fixed, the student model training objective To jointly optimize all cross-modal knowledge distillation functions and align them with the teacher model, the teacher model loss function... Replaced with student model training objective This allows for the training of student models and the realization of cross-modal knowledge distillation. ; in, These are the weighting coefficients for each distillation loss function, all set to 1 here.

[0042] Step S30: Obtain the multimodal data to be tested, input the test set and independent test set of the multimodal data to be tested into the target lightweight student model, the target lightweight student model evaluates and tests the test set and the independent test set to obtain a first prediction result and a second prediction result, compare the first prediction result and the second prediction result to obtain the molecular interaction detection result.

[0043] Specifically, the multimodal data to be tested is acquired, and the test set and independent test set of the multimodal data to be tested are input into the target lightweight student model. The target lightweight student model evaluates and tests the test set and the independent test set to obtain a first prediction result and a second prediction result. The first prediction result and the second prediction result are compared to obtain the molecular interaction detection result.

[0044] In this embodiment, the multimodal data to be tested is randomly divided into test sets. And the independent test set, the evaluation metrics in the evaluation test are as follows, mean squared error: ,in Pearson correlation coefficient: (This represents the total number of samples in the test set.) ,in, These are the mean of the predicted result and the mean of the true value, respectively. The closer the value is to 1, the more consistent the model's predicted trend is with the actual results. Speed ​​evaluation metric: Define model throughput. The unit is The test environment is a single card. Stability: Five different random seeds are set, and the student model is used to conduct multiple rounds of testing on the test set. The standard deviation (STD) of the test results between different rounds is compared to verify the stability of the model's prediction results under different random seeds. Robustness: Robustness is verified by comparing whether the prediction performance of the student model is consistent on randomly divided test sets and independent test sets.

[0045] For example, such as Figure 4 As shown, firstly, a Teacher (multimodal teacher model) is constructed: the textual description, amino acid sequence, and three-dimensional structure of the molecule are independently encoded using a text encoder (BioGPT), a sequence encoder (ESM-2), and a structural encoder (GAT / GNN). The features of each modality are then input into a cross-modal fusion layer, sequentially processed through feature concatenation, Cross-Attention, and gating fusion to obtain a Stage 1 fusion representation with dimensions L×d. Next, through a token aggregation method that progressively expands the receptive field, multi-scale feature downsampling is performed sequentially through Stage 2 (dimensions L / 2×2d), Stage 3 (dimensions L / 4×4d), and Stage 4 (dimensions L / 8×8d), finally inputting the final representation. Regression head, output prediction Simultaneously, intermediate representations, attention maps, and structural spatial relationships are extracted from each stage of the teacher model as knowledge to be transferred. Then, a lightweight student model is constructed: wild-type and mutant sequences are encoded using only the ESM student sequence encoder. After processing by a feature aggregation layer (including residual aggregation, position encoding, and layer normalization), the sequences undergo multi-stage feature downsampling through Stage 1 (L×ds), Stage 2 (L / 2×2ds), Stage 3 (L / 4×4ds), and Stage 4 (L / 8×8ds). Finally, the results are processed... Prediction head output prediction Finally, cross-modal knowledge distillation is performed: in Stage 1, teacher and student feature alignment is achieved through representation distillation (feature alignment); in Stage 2, attention distribution alignment is achieved through attention distillation (attention alignment); in Stages 3 and 4, structural spatial relationship alignment is achieved through structural knowledge distillation (structural relationship alignment); and at the output end, prediction consistency constraints are achieved through output distillation (prediction consistency). This effectively transfers the multimodal knowledge of the teacher model to the target lightweight student model that only depends on sequence input.

[0046] Furthermore, the performance on a randomly partitioned test set was evaluated. Five random seeds were selected. Using the student model on the test set Five rounds of testing were conducted to obtain the test results. The total test duration was recorded, and the test results of each group were compared with the true values. The performance of the target lightweight student model on the test set is shown in Table 1. The results show that the student model has high prediction accuracy (high PCC MEAN, low MSE MEAN) and high stability (small PCC STD, small MSE STD).

[0047] Specifically, the performance of the target lightweight student model on the test set is shown in Table 1: Table 1: Performance of the Target Lightweight Student Model on the Test Set

[0048] As an example, multiple rounds of testing were conducted on the independent test set and evaluation metrics were calculated. The results are shown in Table 2. Comparing the results in Table 1 and Table 2, it can be seen that the student model still maintains high accuracy on the independent test set and is basically consistent with the results of the randomly divided test set. Therefore, it is verified that the method described in this invention has strong robustness. The student model obtained by cross-modal knowledge distillation can achieve high-precision, high-stability, and robust high-speed inference in the antigen-antibody interaction prediction task.

[0049] Specifically, the performance of the target lightweight student model on the independent test set is shown in Table 2: Table 2: Performance of the Target Lightweight Student Model on the Independent Test Set

[0050] Furthermore, such as Figure 5 As shown, based on the above-mentioned molecular multimodal data detection method based on cross-modal knowledge distillation, the present invention also provides a molecular multimodal data detection system based on cross-modal knowledge distillation, wherein the molecular multimodal data detection system based on cross-modal knowledge distillation includes: The model parameter training module 51 is used to acquire molecular multimodal data, construct a multimodal teacher model based on the text data, sequence data and structural data of the molecular multimodal data, and perform regression prediction training on the multimodal teacher model to obtain the target multimodal teacher model. The model data transfer module 52 is used to construct a lightweight student model based on the wild-type and mutant sequences of the molecular multimodal data, and to transfer the parameters of the target multimodal teacher model to the lightweight student model to obtain the target lightweight student model. The molecular data detection module 53 is used to acquire the multimodal data to be tested, input the test set and independent test set of the multimodal data to be tested into the target lightweight student model, the target lightweight student model evaluates and tests the test set and the independent test set to obtain a first prediction result and a second prediction result, and compares the first prediction result and the second prediction result to obtain the molecular interaction detection result.

[0051] Furthermore, such as Figure 6 As shown, based on the above-mentioned molecular multimodal data detection method and system based on cross-modal knowledge distillation, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 6 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0052] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a molecular multimodal data detection program 40 based on cross-modal knowledge distillation, which can be executed by the processor 10 to implement the molecular multimodal data detection method based on cross-modal knowledge distillation in this application.

[0053] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the molecular multimodal data detection method based on cross-modal knowledge distillation.

[0054] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminals communicate with each other via a system bus.

[0055] In one embodiment, when the processor 10 executes the molecular multimodal data detection program 40 based on cross-modal knowledge distillation in the memory 20, the following steps are performed: Acquire molecular multimodal data, construct a multimodal teacher model based on the text data, sequence data and structural data of the molecular multimodal data, and train the multimodal teacher model by regression prediction to obtain the target multimodal teacher model; A lightweight student model is constructed based on the wild-type and mutant sequences of the molecular multimodal data. The parameters of the target multimodal teacher model are then transferred to the lightweight student model to obtain the target lightweight student model. Acquire the multimodal data to be tested, input the test set and independent test set of the multimodal data to be tested into the target lightweight student model, the target lightweight student model evaluates and tests the test set and the independent test set to obtain a first prediction result and a second prediction result, compare the first prediction result and the second prediction result to obtain the molecular interaction detection result; The process of acquiring molecular multimodal data, constructing a multimodal teacher model based on the text data, sequence data, and structural data of the molecular multimodal data, and training the multimodal teacher model through regression prediction to obtain a target multimodal teacher model specifically includes: Molecular multimodal data is acquired, and the text data, sequence data, and structural data of the molecular multimodal data are standardized to obtain processed multimodal data. A multimodal teacher model is constructed based on the processed multimodal data. The processed multimodal data is divided according to a preset ratio to obtain a training set and a validation set; The multimodal teacher model is trained using regression prediction based on the training set to obtain the trained multimodal teacher model. The trained multimodal teacher model is optimized based on the model parameters of the validation set to obtain the target multimodal teacher model.

[0056] The acquisition of molecular multimodal data, including standardizing the text data, sequence data, and structural data of the molecular multimodal data to obtain processed multimodal data, and constructing a multimodal teacher model based on the processed multimodal data, specifically includes: Acquire molecular multimodal data, and convert the text data, sequence data, and structural data into standard format information according to a standard format. The standard format information is removed based on the free energy change labels of the molecular multimodal data to obtain the target dataset; The target dataset is subjected to sample equalization processing according to the undersampling strategy to obtain processed multimodal data; The processed multimodal data is encoded by a multimodal encoder of a large language model to obtain multiple multimodal feature vectors. All the multimodal feature vectors are concatenated to obtain multimodal fusion features. The multimodal fusion features are downsampled in dimension to obtain multiple structural knowledge features. All the structural knowledge features are then used to construct a multimodal teacher model.

[0057] Specifically, the process of encoding the processed multimodal data using a multimodal encoder based on a large language model to obtain multiple multimodal feature vectors, and concatenating all the multimodal feature vectors to obtain multimodal fusion features, includes: The processed multimodal data is text-encoded using a text encoder based on a large language model to obtain a text semantic feature vector. The processed multimodal data is sequence encoded using a sequence encoder based on a large language model to obtain a sequence feature vector. The processed multimodal data is structurally encoded using a structural encoder based on a large language model to obtain a three-dimensional structural feature vector. The text semantic feature vector, sequence feature vector, and three-dimensional structural feature vector are concatenated to obtain cross-modal features. The cross-modal features are then linearly mapped to obtain a query matrix, a key matrix, and a value matrix. Multimodal fusion features are obtained by performing residual connections and normalization layer transformations on the query matrix, the key matrix, and the value matrix through a feedforward network.

[0058] Specifically, the step of constructing a lightweight student model based on the wild-type and mutant sequences of the molecular multimodal data, and transferring the parameters of the target multimodal teacher model to the lightweight student model to obtain the target lightweight student model, includes: Wild-type and mutant sequences of molecular multimodal data are obtained, and the wild-type and mutant sequences are encoded by a sequence encoder to obtain sequence encoding output features. A lightweight student model is constructed based on the sequence encoding output features. Based on representation distillation, attention distillation, and structural knowledge distillation, the intermediate representations, attention maps, and structural space relations of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

[0059] The process of acquiring wild-type and mutant sequences from molecular multimodal data, encoding these sequences using a sequence encoder to obtain sequence-encoded output features, and constructing a lightweight student model based on these features, specifically includes: Wild-type and mutant sequences of molecular multimodal data are obtained, and the wild-type and mutant sequences are encoded by a sequence encoder to obtain sequence-encoded output features; Obtain the position code, perform residual aggregation on the sequence code output feature and the position code to obtain the normalized feature; Obtain the feature mapping matrix, and map the normalized features according to the feature mapping matrix to obtain the lightweight student model.

[0060] Specifically, the method of transferring the intermediate representations, attention maps, and structural space relations of the target multimodal teacher model to the lightweight student model based on representation distillation, attention distillation, and structural knowledge distillation to obtain the target lightweight student model includes: The multimodal fusion features are aligned using the alignment projection layer of the target multimodal teacher model to obtain representation distillation, attention distillation, and structural knowledge distillation. Obtain the projection matrix, and project the representation distillation, the attention distillation, and the structural knowledge distillation according to the projection matrix to obtain the target representation distillation, the target attention distillation, and the target structural knowledge distillation. Based on the target representation distillation, target attention distillation, and target structural knowledge distillation, the intermediate representations, attention maps, and structural space relationships of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

[0061] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a molecular multimodal data detection program based on cross-modal knowledge distillation, and the molecular multimodal data detection program based on cross-modal knowledge distillation, when executed by a processor, implements the steps of the molecular multimodal data detection method based on cross-modal knowledge distillation as described above.

[0062] In summary, this invention provides a method, system, terminal, and storage medium for molecular multimodal data detection based on cross-modal knowledge distillation. The method includes: acquiring molecular multimodal data; constructing a multimodal teacher model based on textual, sequence, and structural data of the molecular multimodal data; training the multimodal teacher model through regression prediction to obtain a target multimodal teacher model; constructing a lightweight student model based on wild-type and mutant sequences of the molecular multimodal data; transferring the parameters of the target multimodal teacher model to the lightweight student model to obtain a target lightweight student model; acquiring multimodal data to be tested; inputting the test set and independent test sets of the multimodal data to be tested into the target lightweight student model; evaluating the target lightweight student model against the test set and the independent test sets to obtain a first prediction result and a second prediction result; comparing the first prediction result and the second prediction result to obtain a molecular interaction prediction result. This invention, based on multimodal teacher-student model distillation and transferring parameters according to a cross-modal knowledge alignment strategy, achieves high-precision and robust detection of molecular interactions.

[0063] 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 terminal system 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 terminal system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal system that includes that element.

[0064] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0065] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for detecting molecular multimodal data based on cross-modal knowledge distillation, characterized in that, The molecular multimodal data detection method based on cross-modal knowledge distillation includes: Acquire molecular multimodal data, construct a multimodal teacher model based on the text data, sequence data and structural data of the molecular multimodal data, and train the multimodal teacher model by regression prediction to obtain the target multimodal teacher model; A lightweight student model is constructed based on the wild-type and mutant sequences of the molecular multimodal data. The parameters of the target multimodal teacher model are then transferred to the lightweight student model to obtain the target lightweight student model. Acquire the multimodal data to be tested, input the test set and independent test set of the multimodal data to be tested into the target lightweight student model, the target lightweight student model evaluates and tests the test set and the independent test set to obtain a first prediction result and a second prediction result, compare the first prediction result and the second prediction result to obtain the molecular interaction detection result.

2. The molecular multimodal data detection method based on cross-modal knowledge distillation according to claim 1, characterized in that, The process of acquiring molecular multimodal data, constructing a multimodal teacher model based on the text data, sequence data, and structural data of the molecular multimodal data, and training the multimodal teacher model through regression prediction to obtain the target multimodal teacher model specifically includes: Molecular multimodal data is acquired, and the text data, sequence data, and structural data of the molecular multimodal data are standardized to obtain processed multimodal data. A multimodal teacher model is constructed based on the processed multimodal data. The processed multimodal data is divided according to a preset ratio to obtain a training set and a validation set; The multimodal teacher model is trained using regression prediction based on the training set to obtain the trained multimodal teacher model. The trained multimodal teacher model is optimized based on the model parameters of the validation set to obtain the target multimodal teacher model.

3. The molecular multimodal data detection method based on cross-modal knowledge distillation according to claim 1, characterized in that, The process of acquiring molecular multimodal data involves standardizing the text data, sequence data, and structural data of the molecular multimodal data to obtain processed multimodal data. A multimodal teacher model is then constructed based on the processed multimodal data, specifically including: Acquire molecular multimodal data, and convert the text data, sequence data, and structural data into standard format information according to a standard format. The standard format information is removed based on the free energy change labels of the molecular multimodal data to obtain the target dataset; The target dataset is subjected to sample equalization processing according to the undersampling strategy to obtain processed multimodal data; The processed multimodal data is encoded by a multimodal encoder of a large language model to obtain multiple multimodal feature vectors. All the multimodal feature vectors are concatenated to obtain multimodal fusion features. The multimodal fusion features are downsampled in dimension to obtain multiple structural knowledge features. All the structural knowledge features are then used to construct a multimodal teacher model.

4. The molecular multimodal data detection method based on cross-modal knowledge distillation according to claim 3, characterized in that, The process involves encoding the processed multimodal data using a multimodal encoder based on a large language model to obtain multiple multimodal feature vectors. These multimodal feature vectors are then concatenated to obtain a multimodal fusion feature, specifically including: The processed multimodal data is text-encoded using a text encoder based on a large language model to obtain a text semantic feature vector. The processed multimodal data is sequence encoded using a sequence encoder based on a large language model to obtain a sequence feature vector. The processed multimodal data is structurally encoded using a structural encoder based on a large language model to obtain a three-dimensional structural feature vector. The text semantic feature vector, sequence feature vector, and three-dimensional structural feature vector are concatenated to obtain cross-modal features. The cross-modal features are then linearly mapped to obtain a query matrix, a key matrix, and a value matrix. Multimodal fusion features are obtained by performing residual connections and normalization layer transformations on the query matrix, the key matrix, and the value matrix through a feedforward network.

5. The molecular multimodal data detection method based on cross-modal knowledge distillation according to claim 1, characterized in that, The step of constructing a lightweight student model based on the wild-type and mutant sequences of the molecular multimodal data, and transferring the parameters of the target multimodal teacher model to the lightweight student model to obtain the target lightweight student model, specifically includes: Wild-type and mutant sequences of molecular multimodal data are obtained, and the wild-type and mutant sequences are encoded by a sequence encoder to obtain sequence encoding output features. A lightweight student model is constructed based on the sequence encoding output features. Based on representation distillation, attention distillation, and structural knowledge distillation, the intermediate representations, attention maps, and structural space relations of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

6. The molecular multimodal data detection method based on cross-modal knowledge distillation according to claim 5, characterized in that, The process of acquiring wild-type and mutant sequences from molecular multimodal data involves encoding these sequences using a sequence encoder to obtain sequence-encoded output features. A lightweight student model is then constructed based on these sequence-encoded output features, specifically including: Wild-type and mutant sequences of molecular multimodal data are obtained, and the wild-type and mutant sequences are encoded by a sequence encoder to obtain sequence-encoded output features; Obtain the position code, perform residual aggregation on the sequence code output feature and the position code to obtain the normalized feature; Obtain the feature mapping matrix, and map the normalized features according to the feature mapping matrix to obtain the lightweight student model.

7. The molecular multimodal data detection method based on cross-modal knowledge distillation according to claim 6, characterized in that, The method based on representation distillation, attention distillation, and structural knowledge distillation transfers the intermediate representations, attention maps, and structural space relations of the target multimodal teacher model to the lightweight student model, resulting in the target lightweight student model. Specifically, this includes: The multimodal fusion features are aligned using the alignment projection layer of the target multimodal teacher model to obtain representation distillation, attention distillation, and structural knowledge distillation. Obtain the projection matrix, and project the representation distillation, the attention distillation, and the structural knowledge distillation according to the projection matrix to obtain the target representation distillation, the target attention distillation, and the target structural knowledge distillation. Based on the target representation distillation, target attention distillation, and target structural knowledge distillation, the intermediate representations, attention maps, and structural space relationships of the target multimodal teacher model are transferred to the lightweight student model to obtain the target lightweight student model.

8. A molecular multimodal data detection system based on cross-modal knowledge distillation, characterized in that, The molecular multimodal data detection system based on cross-modal knowledge distillation includes: The model parameter training module is used to acquire molecular multimodal data, construct a multimodal teacher model based on the text data, sequence data and structural data of the molecular multimodal data, and perform regression prediction training on the multimodal teacher model to obtain the target multimodal teacher model; The model data transfer module is used to construct a lightweight student model based on the wild-type and mutant sequences of the molecular multimodal data, and to transfer the parameters of the target multimodal teacher model to the lightweight student model to obtain the target lightweight student model. The molecular data detection module is used to acquire the multimodal data to be tested, input the test set and independent test set of the multimodal data to be tested into the target lightweight student model, the target lightweight student model evaluates and tests the test set and the independent test set to obtain a first prediction result and a second prediction result, and compares the first prediction result and the second prediction result to obtain the molecular interaction detection result.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a molecular multimodal data detection program based on cross-modal knowledge distillation stored in the memory and executable on the processor. When the molecular multimodal data detection program based on cross-modal knowledge distillation is executed by the processor, it implements the steps of the molecular multimodal data detection method based on cross-modal knowledge distillation as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a molecular multimodal data detection program based on cross-modal knowledge distillation, which, when executed by a processor, implements the steps of the molecular multimodal data detection method based on cross-modal knowledge distillation as described in any one of claims 1-7.

Images