Enzyme substrate interaction prediction method, apparatus, device, medium, and product

Abstract

The application discloses an enzyme-substrate interaction prediction method, device, equipment, medium and product, relates to the field of protease catalysis, and the enzyme-substrate interaction prediction method comprises the following steps: extracting enzyme features of a target enzyme and substrate features of a target substrate respectively through a pre-trained encoder; adjusting weight parameters of a preset protein prediction model according to the enzyme features and the substrate features, to obtain a target protein prediction model; inputting the enzyme features and the substrate features into the target protein prediction model, and outputting interaction features of the target enzyme and the target substrate from the target protein prediction model; and inputting the interaction features into a preset decoder, decoding the interaction features from the decoder, and outputting an interaction prediction result of the target enzyme and the target substrate. The application solves the technical problem of poor interaction prediction effect of a traditional enzyme-substrate pair interaction prediction scheme on rare enzyme-substrate pairs.

Description

Technical Field

[0001] This application relates to the field of protein enzyme catalysis technology, and in particular to a method, apparatus, device, computer-readable storage medium, and computer program product for predicting enzyme-substrate interactions. Background Technology

[0002] Predicting enzyme-substrate interactions is crucial for understanding biological processes, accelerating drug development, and driving innovation in biomedical research. However, traditional wet laboratory methods face limitations in terms of time and cost. With the rapid development of deep learning technology, an increasing number of studies are attempting to utilize deep learning computational methods to predict enzyme-substrate interactions.

[0003] However, current deep learning computational methods ignore the severe imbalance at the data level, directly fitting neural networks onto enzyme-substrate interaction data. These methods include training neural network models from scratch and using features extracted from pre-trained models. While these methods show good predictive performance on the test set, they tend to rely on fitting to in-distribution data and have poor predictive ability for out-of-distribution data. This limitation is particularly severe in long-tailed data distributions, as models tend to overfit the head data during training, neglecting the tail data, resulting in poor prediction performance for rare or few enzyme-substrate pairs.

[0004] The information disclosed in this background section is only for understanding the background technology of the present application concept, and therefore may contain information that does not constitute prior art. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, device, computer-readable storage medium, and computer program product for predicting enzyme-substrate interactions, aiming to solve the technical problem that traditional enzyme-substrate interaction prediction schemes have poor prediction performance for rare enzyme-substrate pairs.

[0006] To achieve the above objectives, this application provides a method for predicting enzyme-substrate interactions, the method comprising:

[0007] The enzyme features of the target enzyme and the substrate features of the target substrate are extracted using the pre-trained encoder.

[0008] The weight parameters of the preset protein prediction model are adjusted according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model.

[0009] The enzyme characteristics and the substrate characteristics are input into the target protein prediction model, and the target protein prediction model outputs the interaction characteristics between the target enzyme and the target substrate.

[0010] The interaction features are input into a preset decoder, which decodes the interaction features and outputs the predicted interaction results between the target enzyme and the target substrate.

[0011] In one embodiment, the encoder includes at least a protein language model and a graph neural network;

[0012] The steps of extracting the enzyme features of the target enzyme and the substrate features of the target substrate using the pre-trained encoder include:

[0013] Obtain the enzyme sequence information of the target enzyme and the substrate molecule information of the target substrate;

[0014] The enzyme sequence information is input into a preset protein language model to obtain the enzyme features of the target enzyme, wherein the enzyme features include hierarchical features of the enzyme.

[0015] The substrate molecule information is input into a preset graph neural network to obtain the substrate features of the target substrate, wherein the substrate features include topological structure and chemical properties.

[0016] In one embodiment, after the step of inputting the substrate molecule information into a preset graph neural network to obtain the substrate features of the target substrate, the method further includes:

[0017] The enzyme features and the substrate features are mapped to a unified latent space, wherein the distance between the enzyme features and the substrate features in the latent space is used to characterize the interaction strength between the target enzyme and the target substrate.

[0018] In one embodiment, prior to the step of extracting the enzyme features of the target enzyme and the substrate features of the target substrate using a pre-trained encoder, the method further includes:

[0019] The enzyme sequence information and the substrate molecule information are respectively input into the initial protein language model and the initial graph neural network, and then the initial protein language model and the initial graph neural network respectively output the prediction results corresponding to the enzyme sequence information and the substrate molecule information;

[0020] Based on a preset contrast loss function, the real labels corresponding to the enzyme sequence information and the substrate molecule information, and the prediction results, the initial protein language model and the initial graph neural network are pre-trained in a preset shared embedding space to obtain the pre-trained protein language model and graph neural network.

[0021] In one embodiment, the preset protein prediction model includes at least a weight adjustment layer;

[0022] The step of adjusting the weight parameters of the preset protein prediction model according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model includes:

[0023] The enzyme features and the substrate features are input into a preset weighted supernetwork, and the weighted supernetwork outputs target weight information.

[0024] The weight information of the weight adjustment layer in the preset protein prediction model is adjusted according to the target weight information to obtain the target protein prediction model.

[0025] In one embodiment, the preset protein prediction model further includes a projection layer;

[0026] Before the step of adjusting the weight parameters of the preset protein prediction model according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model, the method further includes:

[0027] Acquire enzyme feature training data and substrate feature training data;

[0028] The enzyme feature training data and the substrate feature training data are respectively input into the initial weighted supernetwork and the protein pre-training model. The projection layer in the protein pre-training model and the initial weighted supernetwork are trained to obtain the weighted supernetwork and the preset protein prediction model.

[0029] The projection layer is used to process the enzyme features and substrate features received by the preset protein prediction model.

[0030] Furthermore, to achieve the above objectives, this application also provides an enzyme-substrate interaction prediction device, the enzyme-substrate interaction prediction device comprising:

[0031] The feature extraction module is used to extract the enzyme features of the target enzyme and the substrate features of the target substrate through the pre-trained encoder.

[0032] The weight adjustment module is used to adjust the weight parameters of the preset protein prediction model according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model.

[0033] The feature prediction module is used to input the enzyme features and the substrate features into the target protein prediction model, and the target protein prediction model outputs the interaction features between the target enzyme and the target substrate.

[0034] The feature decoding module is used to input the interaction features into a preset decoder, and the decoder decodes the interaction features to output the interaction prediction results between the target enzyme and the target substrate.

[0035] In addition, to achieve the above objectives, this application also provides an enzyme substrate interaction prediction device, which includes at least: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the enzyme substrate interaction prediction method as described above.

[0036] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the enzyme substrate interaction prediction method as described above.

[0037] In addition, to achieve the above objectives, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the enzyme substrate interaction prediction method as described above.

[0038] This application provides a method for predicting enzyme-substrate interactions. First, the enzyme features of the target enzyme and the substrate features of the target substrate are extracted by a pre-trained encoder. Then, the weight parameters of a preset protein prediction model are adjusted according to the enzyme features and the substrate features to obtain a target protein prediction model. This target protein prediction model can be adapted to the interaction prediction of the target enzyme and the target substrate in a targeted and rapid manner. The enzyme features and substrate features are input into the target protein prediction model, and the target protein prediction model outputs the interaction features of the target enzyme and the target substrate. Finally, the interaction features are input into a preset decoder, and the decoder decodes the interaction features to output the interaction prediction result of the target enzyme and the target substrate. In the technical solution of this application, the target protein prediction model adaptively adjusts the weights based on the characteristics of the enzyme and substrate, enabling the model to quickly learn the characteristics of the target enzyme and substrate. Compared with traditional enzyme-substrate pair interaction prediction schemes, it maintains high prediction accuracy for out-of-distribution data, enabling the model to achieve better out-of-distribution generalization ability on sparse data with imbalanced data volume. It overcomes the shortcomings of existing schemes in long-tailed distributed data and sparse and few-sample scenarios, and improves the generalization ability of the protein prediction model and the prediction effect of rare enzyme-substrate pair interactions. Attached Figure Description

[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram illustrating the principle of traditional enzyme substrate pair interaction prediction schemes for predicting out-of-distribution data examples.

[0042] Figure 2 This is a flowchart illustrating an embodiment of the enzyme-substrate interaction prediction method in this application.

[0043] Figure 3 This is a schematic diagram illustrating the principle of the enzyme-substrate interaction prediction method in this application for predicting out-of-distribution data examples;

[0044] Figure 4 This is a schematic diagram of the data flow of a feasible enzyme-substrate interaction prediction scheme in the embodiments of this application;

[0045] Figure 5 This is a schematic diagram of the enzyme-substrate interaction prediction device in the embodiments of this application;

[0046] Figure 6 This is a schematic diagram of the hardware operating environment involved in the enzyme-substrate interaction prediction method in the embodiments of this application.

[0047] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0048] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0049] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0050] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0051] Traditional methods for predicting enzyme-substrate interactions have many limitations. Wet experiments are not only time-consuming and costly, but also limited by experimental conditions and techniques, making it difficult to establish complete enzyme-substrate mappings. Although high-throughput technologies have accelerated this process, the diversity of enzymes and substrates and the complexity of their interactions still pose challenges to data acquisition. Existing enzyme substrate data suffers from severe long-tail distribution problems, and data imbalance limits researchers' comprehensive understanding of enzyme-substrate interactions, posing a significant challenge to the development of deep learning computational models. Although deep learning technology has been applied to some extent in this field, its performance and generalization ability are significantly insufficient when dealing with imbalanced data distributions and few-shot or zero-shot learning scenarios. Figure 1 As shown, when dealing with out-of-distribution data (latent data), the representational distance between the substrate and the enzyme is relatively large, and the model cannot accurately identify these rare enzyme-substrate pairs, resulting in poor prediction performance. This is because traditional deep learning models often rely on fitting within-distribution data and have poor predictive ability for out-of-distribution data. Especially in long-tailed data distributions, the model is prone to overfitting the head data and ignoring the tail data, leading to poor prediction performance for rare enzyme-substrate pairs.

[0052] This application provides a method for predicting enzyme-substrate interactions, referring to... Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the enzyme-substrate interaction prediction method of this application. The enzyme-substrate interaction prediction method includes:

[0053] Step S10: Extract the enzyme features of the target enzyme and the substrate features of the target substrate using the pre-trained encoder.

[0054] In this context, the target enzyme and target substrate refer to the enzyme and substrate for which interaction prediction is required. Specifically, when extracting the corresponding enzyme and substrate features, the enzyme sequence of the target enzyme and the molecular map of the substrate can be used as the data basis. The pre-trained encoder refers to a pre-trained encoder capable of accurately extracting the corresponding features from the enzyme sequence of the target enzyme and the molecular map of the substrate. This can be a deep learning model or a neural network. The goal of pre-training is to ensure that the extracted enzyme features of the target enzyme and the substrate features of the target substrate accurately reflect the interaction strength between the target enzyme and the target substrate. For example, the higher the interaction strength, the closer the enzyme features and substrate features are, thus enabling the quantification of the interaction strength of the enzyme-substrate pair through a simple distance metric. Compared to the traditional manual feature extraction and direct splicing fusion strategy, this significantly enhances interpretability and final prediction accuracy, and lays the foundation for few-sample learning in data-scarce scenarios.

[0055] Step S20: Adjust the weight parameters of the preset protein prediction model according to the enzyme characteristics and substrate characteristics to obtain the target protein prediction model.

[0056] The key aspect of this application's embodiments lies in the adaptive adjustment of the weight parameters of the preset protein prediction model based on the extracted enzyme and substrate features. It is understood that due to the long-tailed nature of the enzyme-substrate interaction dataset, its data distribution exhibits significant class imbalance. Therefore, to prevent the preset protein prediction model from becoming inaccurate in its predictions of the target enzyme and substrate due to model overfitting when predicting the input enzyme and substrate features, this application's embodiments introduce a dynamic weight generation mechanism.

[0057] The preset protein prediction model is a pre-trained model used to predict enzyme-substrate interactions. It has some predictive ability, but may not be highly effective for every enzyme-substrate pair. Based on the input enzyme and substrate features, the preset protein prediction model dynamically and adaptively adjusts its parameters to ultimately obtain a target protein prediction model with stronger predictive ability for the current input enzyme and substrate features. It should be noted that each time new enzyme and substrate features are input, the weight parameters are adaptively generated and updated based on the preset protein prediction model, thus obtaining a target protein prediction model specific to that enzyme-substrate pair, in order to obtain more accurate enzyme-substrate interaction prediction results.

[0058] Step S30: Input the enzyme features and substrate features into the target protein prediction model, and the target protein prediction model outputs the interaction features between the target enzyme and the target substrate;

[0059] After adjusting the weights of the preset protein prediction model and obtaining the target protein prediction model, the target protein prediction model can perform more accurate and in-depth processing and analysis of enzyme characteristics and substrate characteristics, thereby obtaining interaction characteristics that reflect the interaction between the target enzyme and the target substrate. These interaction characteristics can directly reflect whether the target enzyme and the target substrate can interact and react. If so, the interaction characteristics can also reflect the reaction rate between the two.

[0060] Step S40: Input the interaction features into a preset decoder, and the decoder decodes the interaction features to output the predicted interaction results between the target enzyme and the target substrate.

[0061] The preset decoder is a pre-trained model or network that can decode the interaction features output by the target protein prediction model. The decoder can decode and output the interaction prediction results of the target enzyme and the target substrate more intuitively. The interaction prediction results can include whether the interaction can occur or not. If the interaction prediction results include whether the interaction can occur, it can also include the interaction reaction rate between the two.

[0062] Interaction prediction results can be applied to downstream tasks (such as enzyme function annotation and substrate-specific prediction), providing an interpretable feature space for downstream tasks.

[0063] This application provides an enzyme-substrate interaction prediction method. First, the enzyme features of the target enzyme and the substrate features of the target substrate are extracted by a pre-trained encoder. Then, the weight parameters of a preset protein prediction model are adjusted according to the enzyme features and the substrate features to obtain a target protein prediction model. This target protein prediction model can be adapted to the interaction prediction of the target enzyme and the target substrate in a targeted and rapid manner. The enzyme features and substrate features are input into the target protein prediction model, and the target protein prediction model outputs the interaction features of the target enzyme and the target substrate. Finally, the interaction features are input into a preset decoder, and the decoder decodes the interaction features to output the interaction prediction result of the target enzyme and the target substrate. In the technical solution of this application embodiment, the target protein prediction model performs adaptive weight adjustment based on the characteristics of the enzyme and substrate, enabling the model to quickly learn the characteristics of the target enzyme and substrate. Compared with traditional enzyme-substrate pair interaction prediction schemes, it maintains high prediction accuracy for out-of-distribution data, enabling the model to achieve better out-of-distribution generalization ability on sparse data with imbalanced data volume. It overcomes the shortcomings of existing schemes in long-tailed distributed data and sparse and few sample scenarios, and improves the generalization ability of the protein prediction model and the interaction prediction effect for rare enzyme-substrate pairs.

[0064] Furthermore, in one feasible embodiment, the encoder includes at least a protein language model and a graph neural network;

[0065] The steps of extracting the enzyme features of the target enzyme and the substrate features of the target substrate using the pre-trained encoder include:

[0066] Step S11: Obtain the enzyme sequence information of the target enzyme and the substrate molecule information of the target substrate;

[0067] Step S12: Input the enzyme sequence information into a preset protein language model to obtain the enzyme features of the target enzyme, wherein the enzyme features include the hierarchical features of the enzyme.

[0068] Step S13: Input the substrate molecule information into a preset graph neural network to obtain the substrate features of the target substrate, wherein the substrate features include topological structure and chemical properties.

[0069] This application discloses a method for extracting hierarchical features from enzyme sequence information using a protein language model to obtain enzyme features; and for extracting topological structure and chemical properties from substrate molecule information using a graph neural network to obtain substrate features.

[0070] The pre-set protein language model and the pre-set graph neural network are both pre-trained. The purpose of pre-training is to ensure that enzyme-substrate pairs participating in the same biochemical reaction obtain consistent representations in the latent space, while unrelated combinations produce significant feature differences. Enzyme proteins and substrate molecules are considered as two complementary modalities of the biochemical reaction, and cross-modal semantic associations are constructed through contrastive learning.

[0071] Furthermore, after the step of inputting substrate molecule information into a preset graph neural network to obtain substrate features of the target substrate, the method further includes:

[0072] Step S13: Map enzyme features and substrate features to a unified latent space, wherein the distance between enzyme features and substrate features in the latent space is used to characterize the interaction strength between the target enzyme and the target substrate.

[0073] After extracting the enzyme features of the target enzyme and the substrate features of the target substrate, these features are mapped to a unified latent space. The purpose of this step is to reflect the specific rules governing the interactions between different enzymes and substrates through the distances between enzyme and substrate features in the latent space. That is, enzymes with similar catalytic functions will be characterized closer to their substrate molecules in the latent space, while enzyme-substrate pairs with different functional classes will maintain a significant distance. This geometric property of spatial distance provides an interpretable feature space for downstream tasks such as enzyme function annotation and substrate-specific prediction.

[0074] It is understandable that when applying the enzyme-substrate interaction prediction method of the embodiments of this application to out-of-distribution data (latent data), such as Figure 3 As shown, the representational distance between the substrate and the enzyme is closer. Before inputting enzyme features and substrate features into the model, the model is weighted using a supermodel that combines enzyme features and substrate features. This allows the model to adapt to the currently input enzyme features and substrate features, resulting in more accurate predictions.

[0075] Furthermore, in one feasible embodiment, prior to the step of extracting the enzyme features of the target enzyme and the substrate features of the target substrate respectively using a pre-trained encoder, the method further includes:

[0076] Step A10: Input the enzyme sequence information and substrate molecule information into the initial protein language model and the initial graph neural network, respectively, and then output the prediction results corresponding to the enzyme sequence information and substrate molecule information from the initial protein language model and the initial graph neural network, respectively.

[0077] Step A20: Based on the preset contrast loss function, the real labels corresponding to the enzyme sequence information and the substrate molecule information, and the prediction results, the initial protein language model and the initial graph neural network are pre-trained in the preset shared embedding space to obtain the pre-trained protein language model and graph neural network.

[0078] This application provides a method for pre-training a protein speech model and a graph neural network. In the pre-training stage, this application constructs a comparative learning framework of enzyme protein sequence-substrate molecular graph, such as... Figure 4 As shown, GNN (Graph Neural Network) and ESM-2 (Evolution Scale Modeling v2, a protein sequence pre-training model) process the substrate map and enzyme sequence (e.g., GASP) respectively, extracting the corresponding substrate features and enzyme features. These features are then processed through an adapter layer and combined before being input into the weight adjustment layer of the target protein prediction model. During training, for paired enzyme substrates, the distance between them is reduced; for unpaired enzyme substrates, the distance between them is increased.

[0079] Furthermore, during training, a contrastive loss function is used to measure the encoder's loss to guide the pre-training process. Specifically, this symmetric loss design forces bidirectional modal alignment, ensuring that enzyme proteins can predict their corresponding substrates, and substrates can trace back to their source enzyme proteins, thereby enhancing the interoperability of the representations. This stage mainly achieves the alignment of enzyme and molecule modalities through contrastive learning, and employs a dual-encoder architecture to extract features from enzyme sequence and substrate molecule graphs respectively. In a shared embedding space, enzyme-substrate pairs are trained to bring matching enzyme-substrate pairs closer together, while unmatched pairs repel each other. It should be noted that the initial protein language model and the initial graph neural network are the encoders to be trained. Enzyme sequence information and substrate molecule information correspond to real labels used for the real enzyme features and real substrate features corresponding to enzyme sequence information and substrate molecule information, respectively. The prediction results reflect the predicted enzyme features and predicted substrate features. The contrastive loss function is a loss function that optimizes the feature space by comparing the similarity of sample pairs. Its core objective is to make similar samples closer in the embedding space and dissimilar samples farther apart.

[0080] In this embodiment, only the enzyme sequence and substrate molecular map are used as input, avoiding complex feature engineering and thus improving the model's versatility and scalability. A contrastive learning strategy is employed during the pre-training stage to make the enzyme more closely resemble its catalyzed substrate in the latent feature space, thereby emphasizing the specificity of enzyme catalysis. This contrastive learning method not only enhances the model's understanding of enzyme-substrate interactions but also lays a solid foundation for the subsequent model weight adjustment stage.

[0081] In one feasible embodiment, the preset protein prediction model includes at least a weight adjustment layer;

[0082] The step of adjusting the weight parameters of the preset protein prediction model based on enzyme characteristics and substrate characteristics to obtain the target protein prediction model includes:

[0083] Step S21: Input the enzyme features and substrate features into a preset weighted supernetwork, and the weighted supernetwork outputs the target weight information;

[0084] Step S22: Adjust the weight information of the weight adjustment layer in the preset protein prediction model according to the target weight information to obtain the target protein prediction model.

[0085] This application's embodiments introduce a dynamic weight generation mechanism, primarily based on a hypernetwork parameter adaptive framework. This framework can overcome the limitations of traditional static fully connected layers, achieving instance-level adaptive modeling (e.g., through a dual-network collaborative architecture) Figure 4 (As shown). Specifically, the main network of the protein prediction model can be defined as the main structure that performs forward inference. The weight parameters of its final weight adjustment layer are no longer fixed, but are dynamically generated by the supernetwork based on the enzyme-substrate features of the current input. The weight adjustment layer can be an MLP (Multilayer Perceptron).

[0086] In this embodiment, by employing the enzyme and substrate features output by the aforementioned encoder, the connections represented by enzymes and substrates are used as meta-information to generate corresponding weights for each data instance, thereby adaptively updating the initial protein prediction model. This dynamic weight generation mechanism essentially constructs a smooth mapping from the input space to the parameter space. This allows the model to automatically adjust the decision boundary based on input similarity rather than simply performing extrapolation predictions when faced with out-of-distribution samples. Considering that connections contain coarse enzyme-substrate interaction patterns, this embodiment can rapidly adapt the catalytic interaction of each enzyme-substrate pair at the instance level. Furthermore, from the perspective of weight generation, out-of-distribution data that is out of distribution in terms of data may not be entirely considered out of distribution from the perspective of optimal weight optimization, which helps improve the model's predictive performance.

[0087] It should be noted that the embodiments of this application use a two-stage alignment method, using enzyme substrate features as input to the supernetwork to enable the model to quickly adjust weights. Contrastive learning allows the enzyme substrate features to be extracted more effectively by the supernetwork.

[0088] like Figure 4 As shown, enzyme features and substrate features are input into a weight-adjusted supernetwork (i.e., a weighted supernetwork). The supernetwork outputs a weight stream (i.e., target weight information) to the weight adjustment layer of the initial protein prediction model for weight adjustment, resulting in the target protein prediction model. After obtaining the target protein prediction model, interaction features are predicted using the input enzyme and substrate features, and then output to the decoder to obtain the interaction prediction results, which are then applied to downstream tasks.

[0089] This application introduces a supernetwork structure to generate model weights adapted to the current task (i.e., the input enzyme features and substrate weights). The supernetwork utilizes the enzyme and substrate features output from the pre-trained enzyme and substrate encoders in the first stage to efficiently learn the most generalizable weights, rather than simply fitting poorly conditional distributions in the training data. In this way, the target protein prediction model can maintain high prediction performance even when facing out-of-distribution data. Furthermore, the introduction of the supernetwork not only improves the model's generalization ability but also significantly reduces the number of training iterations, enabling the model to converge with fewer iterations. Based on the above design, the technical solution of this application can achieve better out-of-distribution generalization on imbalanced data, overcoming the shortcomings of existing methods in long-tailed distribution data and few-sample scenarios, providing a more powerful tool for the biomedical field, and promoting innovative development in enzyme engineering, drug discovery, and other related fields.

[0090] In one feasible embodiment, the preset protein prediction model further includes a projection layer;

[0091] Before the step of adjusting the weight parameters of the preset protein prediction model based on enzyme characteristics and substrate characteristics to obtain the target protein prediction model, the method further includes:

[0092] Step B10: Obtain enzyme feature training data and substrate feature training data;

[0093] Step B20: Input the enzyme feature training data and substrate feature training data into the initial weighted supernetwork and the protein pre-training model, respectively. Train the projection layer and the initial weighted supernetwork in the protein pre-training model to obtain the weighted supernetwork and the preset protein prediction model.

[0094] The projection layer is used to process the enzyme features and substrate features received by the preset protein prediction model.

[0095] It is understandable that both the initial protein prediction model and the preset protein prediction model include a projection layer. In deep learning models, the projection layer is an important structural design, its core function being to perform feature mapping and dimensionality adjustment on the input data through linear or nonlinear transformations. In this embodiment, the projection layer is used to receive enzyme features and substrate features output by the encoder, and to perform feature mapping and dimensionality reduction on the enzyme features and substrate features, enabling the protein prediction model to perform further analysis and processing. The enzyme feature training data and substrate feature training data are pre-prepared data with real labels used to train the weighted supernetwork and the pre-trained protein model.

[0096] Considering the constraints of computational resources, this embodiment freezes the parameters of the pre-trained protein model during the training phase, training only the projection layers to reduce computation and improve training efficiency. Given the inherent complexity of hypernetwork training, this embodiment employs a learning rate restart strategy during training. Specifically, a learning rate decay strategy is used to reset the learning rate to its initial value after a certain training step. This training strategy effectively mitigates the potential for local optima during training and improves the model's convergence speed, enabling rapid convergence. For example, mean squared error (MSE) loss can be used to optimize model parameters during iterative training.

[0097] The purpose of the above training steps is to train a supernetwork that can output accurate target weight information based on the input enzyme and substrate features, and a preset protein prediction model that can output accurate interaction features based on the input enzyme and substrate features (this model still needs further weight adjustment).

[0098] Exemplarily, the enzyme-substrate interaction prediction method of this application can be implemented in two stages. The first stage is pre-training using a contrastive learning strategy. By using the enzyme substrate sequence as input, the enzyme is brought closer to its catalyzed substrate in the latent feature space. This results in an enzyme substrate encoding model with stronger recognition capabilities, enabling it to better capture the complex relationship between the enzyme and its substrate. The second stage is training the corresponding hypernetwork, utilizing its weight generation ability and generalization ability to perform targeted weight adjustments on the model. This allows the model to generate targeted weights for different data, improving the generalization ability of the target protein prediction model and making it more suitable for prediction tasks involving rare or few enzyme-substrate pairs.

[0099] This application utilizes the emerging technology of hypernetworks to address the challenge of predicting enzyme-substrate interactions with limited sample data. Hypernetworks, as meta-network models capable of generating network parameters, possess excellent generalization capabilities, enabling them to learn from multiple tasks and quickly adapt to new ones. This characteristic makes them highly promising in limited-sample learning scenarios, effectively improving the learning efficiency and prediction accuracy of models with few samples. The enzyme-substrate interaction prediction method provided in this application not only enhances the accuracy and efficiency of molecular interaction prediction, providing a more powerful tool for the biomedical field and helping researchers make faster decisions in new drug development, but also achieves effective prediction of enzyme-substrate interactions even with scarce data, thereby promoting the full utilization of biological data.

[0100] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the enzyme substrate interaction prediction method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0101] This application also provides an enzyme substrate interaction prediction device, such as... Figure 5 As shown, the enzyme-substrate interaction prediction device includes:

[0102] The feature extraction module 10 is used to extract the enzyme features of the target enzyme and the substrate features of the target substrate through the pre-trained encoder.

[0103] The weight adjustment module 20 is used to adjust the weight parameters of the preset protein prediction model according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model.

[0104] Feature prediction module 30 is used to input the enzyme features and the substrate features into the target protein prediction model, and the target protein prediction model outputs the interaction features between the target enzyme and the target substrate;

[0105] The feature decoding module 40 is used to input the interaction features into a preset decoder, and the decoder decodes the interaction features to output the interaction prediction results between the target enzyme and the target substrate.

[0106] In one embodiment, the encoder includes at least a protein language model and a graph neural network; the feature extraction module 10 is further configured to:

[0107] Obtain the enzyme sequence information of the target enzyme and the substrate molecule information of the target substrate;

[0108] The enzyme sequence information is input into a preset protein language model to obtain the enzyme features of the target enzyme, wherein the enzyme features include hierarchical features of the enzyme.

[0109] The substrate molecule information is input into a preset graph neural network to obtain the substrate features of the target substrate, wherein the substrate features include topological structure and chemical properties.

[0110] In one embodiment, the feature extraction module 10 is further configured to:

[0111] The enzyme features and the substrate features are mapped to a unified latent space, wherein the distance between the enzyme features and the substrate features in the latent space is used to characterize the interaction strength between the target enzyme and the target substrate.

[0112] In one embodiment, the enzyme substrate interaction prediction device further includes a pre-training module, which is used for:

[0113] The enzyme sequence information and the substrate molecule information are respectively input into the initial protein language model and the initial graph neural network, and then the initial protein language model and the initial graph neural network respectively output the prediction results corresponding to the enzyme sequence information and the substrate molecule information;

[0114] Based on a preset contrast loss function, the real labels corresponding to the enzyme sequence information and the substrate molecule information, and the prediction results, the initial protein language model and the initial graph neural network are pre-trained in a preset shared embedding space to obtain the pre-trained protein language model and graph neural network.

[0115] In one embodiment, the preset protein prediction model includes at least a weight adjustment layer; the weight adjustment module 20 is further configured to:

[0116] The enzyme features and the substrate features are input into a preset weighted supernetwork, and the weighted supernetwork outputs target weight information.

[0117] The weight information of the weight adjustment layer in the preset protein prediction model is adjusted according to the target weight information to obtain the target protein prediction model.

[0118] In one embodiment, the preset protein prediction model further includes a projection layer, and the enzyme-substrate interaction prediction device further includes a model training module, which is further used for:

[0119] Acquire enzyme feature training data and substrate feature training data;

[0120] The enzyme feature training data and the substrate feature training data are respectively input into the initial weighted supernetwork and the protein pre-training model. The projection layer in the protein pre-training model and the initial weighted supernetwork are trained to obtain the weighted supernetwork and the preset protein prediction model.

[0121] The projection layer is used to process the enzyme features and substrate features received by the preset protein prediction model.

[0122] The enzyme substrate interaction prediction device provided in this application, employing the enzyme substrate interaction prediction method described in the above embodiments, can solve the technical problem of poor prediction performance of traditional enzyme substrate pair interaction prediction schemes for rare enzyme substrate pairs. Compared with the prior art, the beneficial effects of the enzyme substrate interaction prediction device provided in this application are the same as those of the enzyme substrate interaction prediction method provided in the above embodiments, and other technical features in this enzyme substrate interaction prediction device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0123] This application also provides an enzyme substrate interaction prediction device, which includes at least: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the enzyme substrate interaction prediction method in the above embodiments.

[0124] The following is for reference. Figure 6 It shows a schematic diagram of the structure of a control unit suitable for implementing the embodiments of this application. Figure 6 The control unit shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this application.

[0125] like Figure 6As shown, the control unit may include a processing device 101 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 102 or a program loaded from storage device 103 into random access memory (RAM) 104. RAM 104 also stores various programs and data required for the operation of the control unit. The processing device 101, ROM 102, and RAM 104 are interconnected via bus 105. Input / output (I / O) interface 106 is also connected to the bus. Typically, the following systems can be connected to I / O interface 106: input devices 107 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 108 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 103 including, for example, magnetic tapes, hard disks, etc.; and communication devices 109. Communication device 109 allows the control unit to communicate wirelessly or wiredly with other devices to exchange data. Although the diagram shows control units with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented alternatively.

[0126] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 103, or installed from ROM 102. When the computer program is executed by processing device 101, it performs the functions defined in the methods of the embodiments of this application.

[0127] The enzyme substrate interaction prediction device provided in this application, employing the enzyme substrate interaction prediction method described in the above embodiments, can solve the technical problem of poor prediction performance for rare enzyme substrate pairs using traditional enzyme substrate pair interaction prediction schemes. Compared with the prior art, the beneficial effects of the enzyme substrate interaction prediction device provided in this application are the same as those of the enzyme substrate interaction prediction method provided in the above embodiments, and other technical features in this enzyme substrate interaction prediction device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0128] It should be understood that various parts of the embodiments of this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0129] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the above claims.

[0130] This application also provides a computer-readable storage medium storing a computer program that can run on a processor. The computer program is used to execute the enzyme substrate interaction prediction method in the above embodiments.

[0131] The computer-readable storage medium provided in this application embodiment may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0132] The aforementioned computer-readable storage medium may be included in the enzyme substrate interaction prediction device; or it may exist independently and not assembled into the enzyme substrate interaction prediction device.

[0133] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the enzyme-substrate interaction prediction device, the enzyme-substrate interaction prediction device performs the following actions: extracting enzyme features of the target enzyme and substrate features of the target substrate using a pre-trained encoder; adjusting the weight parameters of a preset protein prediction model based on the enzyme features and substrate features to obtain a target protein prediction model; inputting the enzyme features and substrate features into the target protein prediction model, and having the target protein prediction model output the interaction features between the target enzyme and the target substrate; inputting the interaction features into a preset decoder, and having the decoder decode the interaction features to output the prediction result of the interaction between the target enzyme and the target substrate.

[0134] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0135] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0136] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0137] The computer-readable storage medium provided in this application embodiment stores computer-readable program instructions for executing the above-described enzyme substrate interaction prediction method, which can solve the technical problem that traditional enzyme substrate interaction prediction schemes have poor prediction effects on rare enzyme substrate pairs. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application embodiment are the same as the beneficial effects of the enzyme substrate interaction prediction method provided in the above embodiments, and will not be repeated here.

[0138] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the enzyme substrate interaction prediction method as described above.

[0139] The computer program product provided in this application can solve the technical problem that traditional enzyme substrate interaction prediction schemes have poor prediction performance for rare enzyme substrate pairs. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the enzyme substrate interaction prediction method provided in the above embodiments, and will not be repeated here.

[0140] The above are merely exemplary solutions of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent processing scope of this application.

Claims

1. A method for predicting enzyme-substrate interactions, characterized in that, The enzyme-substrate interaction prediction method includes: The enzyme features of the target enzyme and the substrate features of the target substrate are extracted using the pre-trained encoder. The weight parameters of the preset protein prediction model are adjusted according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model. The enzyme characteristics and the substrate characteristics are input into the target protein prediction model, and the target protein prediction model outputs the interaction characteristics between the target enzyme and the target substrate. The interaction features are input into a preset decoder, which decodes the interaction features and outputs the predicted interaction results between the target enzyme and the target substrate. The preset protein prediction model includes at least a weight adjustment layer; The step of adjusting the weight parameters of the preset protein prediction model according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model includes: The enzyme features and the substrate features are input into a preset weighted supernetwork, and the weighted supernetwork outputs target weight information. The weight information of the weight adjustment layer in the preset protein prediction model is adjusted according to the target weight information to obtain the target protein prediction model.

2. The enzyme-substrate interaction prediction method as described in claim 1, characterized in that, The encoder includes at least a protein language model and a graph neural network; The steps of extracting the enzyme features of the target enzyme and the substrate features of the target substrate using the pre-trained encoder include: Obtain the enzyme sequence information of the target enzyme and the substrate molecule information of the target substrate; The enzyme sequence information is input into a preset protein language model to obtain the enzyme features of the target enzyme, wherein the enzyme features include hierarchical features of the enzyme. The substrate molecule information is input into a preset graph neural network to obtain the substrate features of the target substrate, wherein the substrate features include topological structure and chemical properties.

3. The enzyme-substrate interaction prediction method as described in claim 2, characterized in that, After the step of inputting the substrate molecule information into a preset graph neural network to obtain the substrate features of the target substrate, the method further includes: The enzyme features and the substrate features are mapped to a unified latent space, wherein the distance between the enzyme features and the substrate features in the latent space is used to characterize the interaction strength between the target enzyme and the target substrate.

4. The enzyme-substrate interaction prediction method as described in claim 2, characterized in that, Prior to the step of extracting the enzyme features of the target enzyme and the substrate features of the target substrate using a pre-trained encoder, the method further includes: The enzyme sequence information and the substrate molecule information are respectively input into the initial protein language model and the initial graph neural network, and then the initial protein language model and the initial graph neural network respectively output the prediction results corresponding to the enzyme sequence information and the substrate molecule information; Based on a preset contrast loss function, the real labels corresponding to the enzyme sequence information and the substrate molecule information, and the prediction results, the initial protein language model and the initial graph neural network are pre-trained in a preset shared embedding space to obtain the pre-trained protein language model and graph neural network.

5. The enzyme-substrate interaction prediction method as described in claim 1, characterized in that, The preset protein prediction model also includes a projection layer; Before the step of adjusting the weight parameters of the preset protein prediction model according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model, the method further includes: Acquire enzyme feature training data and substrate feature training data; The enzyme feature training data and the substrate feature training data are respectively input into the initial weighted supernetwork and the protein pre-training model. The projection layer in the protein pre-training model and the initial weighted supernetwork are trained to obtain the weighted supernetwork and the preset protein prediction model. The projection layer is used to process the enzyme features and substrate features received by the preset protein prediction model.

6. An enzyme-substrate interaction prediction device, characterized in that, The enzyme-substrate interaction prediction device includes: The feature extraction module is used to extract the enzyme features of the target enzyme and the substrate features of the target substrate through the pre-trained encoder. The weight adjustment module is used to adjust the weight parameters of the preset protein prediction model according to the enzyme characteristics and the substrate characteristics to obtain the target protein prediction model. The feature prediction module is used to input the enzyme features and the substrate features into the target protein prediction model, and the target protein prediction model outputs the interaction features between the target enzyme and the target substrate. The feature decoding module is used to input the interaction features into a preset decoder, and the decoder decodes the interaction features to output the interaction prediction results between the target enzyme and the target substrate; The preset protein prediction model includes at least a weight adjustment layer. The weight adjustment module is further configured to: input the enzyme features and the substrate features into a preset weight supernetwork, and output target weight information from the weight supernetwork; adjust the weight information of the weight adjustment layer in the preset protein prediction model according to the target weight information to obtain the target protein prediction model.

7. An enzyme-substrate interaction prediction device, characterized in that, The enzyme substrate interaction prediction device includes at least: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the enzyme substrate interaction prediction method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for implementing an enzyme-substrate interaction prediction method, the program for implementing the enzyme-substrate interaction prediction method being executed by a processor to implement the steps of the enzyme-substrate interaction prediction method as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the enzyme substrate interaction prediction method as described in any one of claims 1 to 5.

Images