Immune protein ligand receptor function prediction method, device and storage medium

Abstract

The present invention provides a method, device and storage medium for predicting the function of immune protein ligand receptor, which relates to the field of artificial intelligence, including: building an artificial neural network that supports multiple inputs, the artificial neural network is used to receive ligand and receptor vectors, and output the prediction score of a specific function, the artificial intelligence network at least includes a feature extractor, a feature fusion device and a classifier, wherein the feature extractor is used to extract feature vectors in amino acid sequence vectors, the feature fusion device is used to fuse the feature vectors of protein ligands and receptors, and the classifier is used to classify the fused feature vectors to output the prediction results; the artificial neural network is trained using transfer learning strategies and meta-learning strategies; and the receptor ligand data to be predicted is predicted based on the trained artificial neural network. The method provided by the present invention can predict the functions of different ligands and receptor proteins, and generalize the prediction of the interaction functional properties of unseen ligand and receptor combinations.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method, device and storage medium for predicting the function of an immune protein ligand receptor. Background Art

[0002] The application of deep learning in predicting the functions of various receptors (immunoreceptors) of immune cells has great application potential. As a powerful artificial intelligence method, deep learning can process complex bioinformatics data, extract features, and make accurate predictions. The interaction between ligands and immune receptors is the key to determining immune responses. Deep learning can assist in analyzing immune receptor-ligand interactions, thereby systematically analyzing immune functions mediated by immune receptors. The advantage of deep learning models is that they can find the rules of ligand-receptor interactions by learning the interaction data of known ligands and receptors. This is of great significance for understanding the mechanism of immune response mediated by immune ligands, developing immunotherapy, and drug optimization. With the improvement of computing performance and the accumulation of biological data, the application of deep learning in this field will continue to expand and deepen. However, the currently known technical means can only predict the interaction between various receptors of immune cells (such as TCR and BCR, referred to as immune receptors) and specific ligands, while the prediction performance of unknown ligands and unknown immune receptor combinations is very poor. Existing deep learning methods for predicting specific immune receptor functions usually adopt end-to-end training directly on the target task data. However, due to the limitations of data acquisition, such training strategies are prone to overfitting problems. Therefore, how to achieve direct prediction of broad-spectrum neutralizing antibodies has become a technical problem that needs to be solved urgently. Summary of the invention

[0003] The present invention aims to solve at least one of the technical problems existing in the prior art or related technology, and discloses a method, device and storage medium for predicting the function of immune protein ligand receptor, which can simultaneously predict the functions of different ligands and different immune receptor combinations, and can generalize the prediction of the interaction between unknown ligands and unknown immune receptor pairing combinations and the ligand function.

[0004] The first aspect of the present invention discloses a method for predicting the function of an immune protein ligand receptor, comprising: building an artificial neural network that supports multiple inputs, the artificial neural network is used to receive ligand and receptor vectors, and output a prediction score for a specific function, the artificial intelligence network at least includes a feature extractor, a feature fusion device, and a classifier, wherein the feature extractor is used to extract feature vectors from an amino acid sequence vector, the feature fusion device is used to fuse the feature vectors of a protein ligand and a receptor, and the classifier is used to classify the fused feature vectors to output a prediction result; the artificial neural network is trained using a transfer learning strategy and a meta-learning strategy; and the receptor ligand data to be predicted is predicted based on the trained artificial neural network.

[0005] In this technical solution, the feature fusion of the artificial neural network uses global maximum pooling to remove sequence space features, and stacks the extracted feature vectors of ligands and receptors to fuse the two. This method can also be replaced by an attention mechanism to align feature dimensions, or simply fuse the features of the two by adding them. The classifier of the artificial neural network uses a fully connected network.

[0006] According to the immune protein ligand receptor function prediction method disclosed in the present invention, preferably, it also includes: encoding each amino acid in the labeled ligand and receptor amino acid sequences to obtain an amino acid vector sequence; inserting a zero-filled vector to make the amino acid vector sequence reach a fixed length; sliding the non-zero sequence part on the vector sequence to reduce the neural network's dependence on the absolute position characteristics of the sequence dimension.

[0007] According to the immune protein ligand receptor function prediction method disclosed in the present invention, preferably, the step of using a transfer learning strategy to train an artificial neural network specifically includes: using a transfer learning strategy to separately train a feature extractor, and performing unsupervised optimization on protein sequence data through an encoder-decoder structure, so that the feature extractor can act as an encoder to learn how to convert from an input sequence vector to a latent space feature vector.

[0008] In this technical solution, the feature extractor is pre-trained and transferred based on the unsupervised learning method of mask reconstruction based on the universal protein sequence. There are many alternatives: the pre-training data can be replaced by more targeted amino acid sequences of multiple immune proteins, and the training strategy can also be replaced by other unsupervised or supervised methods such as clustering.

[0009] According to the immune protein ligand receptor function prediction method disclosed in the present invention, preferably, the training data is derived from generalized protein data so as to increase the depth of understanding of the feature extractor for the protein sequence.

[0010] According to the immune protein ligand receptor function prediction method disclosed in the present invention, preferably, the step of using a meta-learning strategy to train an artificial neural network specifically includes: using a meta-learning strategy based on task sampling to train the artificial neural network to balance the network's understanding depth between different ligands or different receptors, and to improve the generalization of unseen receptors or unseen ligands; using a meta-learning strategy based on hyperparameters to fine-tune the artificial neural network, targeting the interaction function of the paired receptor to be predicted and the ligand, to improve the overall generalization ability of the artificial neural network and avoid overfitting.

[0011] In this technical solution, the meta-learning method based on task sampling adopts the reptile algorithm, which can also be replaced by the MAML algorithm or other methods based on sampling tasks to improve the generalization of the model for new tasks (unseen ligands or receptors).

[0012] According to the immune protein ligand receptor function prediction method disclosed in the present invention, preferably, the meta-learning strategy based on task sampling and the meta-learning strategy based on hyper-parameters are combined and used in the same training cycle, thereby improving the generalization of the model from both the hyper-parameter and data sampling perspectives.

[0013] According to the immune protein ligand receptor function prediction method disclosed in the present invention, preferably, the L1 regularization term loss function weight of the model weight and the dropout probability of the fully connected layer are selected as the hyperparameters adjusted in the meta-learning strategy based on the hyperparameter, and the hyperparameters are adjusted by the difference in accuracy or F1 index between the meta-training set and the meta-validation set, thereby reducing the overfitting problem in network fine-tuning.

[0014] In this technical solution, the selected hyperparameters can be replaced by other hyperparameters such as L2 regularization term weight, data augmentation parameters, network short-circuit parameters, etc., and the indicators for measuring the prediction performance of the meta-training set and the meta-validation set can also be replaced by other evaluation indicators.

[0015] According to the immune protein ligand receptor function prediction method disclosed in the present invention, preferably, the feature extractor is implemented based on the ProteinBERT model.

[0016] The second aspect of the present invention discloses an immune protein ligand receptor function prediction device, comprising: a memory for storing program instructions; a processor for calling the program instructions stored in the memory to implement an immune protein ligand receptor function prediction method as described in any of the above technical solutions.

[0017] The third aspect of the present invention discloses a computer-readable storage medium storing a program code, wherein the program code is used to implement the method for predicting the function of an immune protein ligand receptor as described in any of the above technical solutions.

[0018] The beneficial effects of the present invention include at least: most of the existing technologies can only input ligand or receptor sequences separately to predict the specific functions of specific ligands or receptors, but cannot provide functional predictions of different ligand receptor combinations at the same time. The present invention designs a multi-input network to input the amino acid sequences of ligands and receptors, thereby predicting the functions of different ligand and receptor proteins, and can generalize and predict the interaction functional properties of unseen ligand and receptor combinations. The prior art usually only inputs the amino acid sequence of the interaction region, but ignores the effect of changes in other regions on the properties of the protein, thereby limiting the prediction accuracy. In the present invention, the amino acid sequence of the entire variable region is directly input to ensure the integrity of the input information. The existing deep learning methods for predicting the specific functions of ligand receptors usually directly adopt end-to-end training directly on the target data, but due to the limitations of data acquisition, this training strategy is prone to overfitting. The present invention adopts transfer learning to pre-train the feature extractor module based on an unsupervised training method to deepen the model's understanding of proteins. In addition, the present invention uses a newly designed meta-learning strategy to train the model, which improves the generalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic flow chart of a method for predicting immune protein ligand receptor function according to an embodiment of the present invention is shown.

[0020] Figure 2 A schematic diagram of the network structure of a method for predicting immune protein ligand receptor function according to an embodiment of the present invention is shown.

[0021] Figure 3 A schematic diagram of the model training process of a method for predicting immune protein ligand receptor function according to an embodiment of the present invention is shown.

[0022] Figure 4 A schematic diagram of a process for training an artificial neural network based on a meta-learning strategy of task sampling according to an embodiment of the present invention is shown.

[0023] Figure 5 A schematic diagram of a process for fine-tuning an artificial neural network based on a meta-learning strategy based on hyperparameters according to an embodiment of the present invention is shown.

[0024] Figure 6 A schematic block diagram of an apparatus for predicting immune protein ligand receptor function according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0025] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0027] like Figure 1 and Figure 2 As shown, according to one embodiment of the present invention, the immune protein ligand receptor function prediction method disclosed in the present invention includes: inputting annotated multi-receptor (or ligand) specific ligand data; processing the original sequence data; building a trainable neural network model; training the neural network model based on transfer learning and meta-learning strategies; inputting multi-receptor (or ligand) specific ligand data to be predicted; using the neural network model to predict; and outputting the prediction results. Specifically:

[0028] Step 1: Processing of raw sequence data, including standardization of ligand and receptor amino acid sequences, vector encoding, and data enhancement. The purpose of the first two is to encode the amino acid sequence into a data form that can be processed by the neural network, and the purpose of the latter is to improve the stability of deep learning to data changes. Among them, sequence data standardization first intercepts the variable region amino acid sequence of the ligand receptor, and screens out amino acid sequences with abnormal sequencing results through indicators such as length and sequencing markers. Sequence vector encoding encodes the amino acid sequence into a set of vectors by encoding each amino acid in the amino acid sequence as a vector, and then inserts a zero-filled vector to make the vector sequence reach a fixed length. Data enhancement reduces the dependence of the neural network on the absolute position characteristics of the sequence dimension by sliding the non-zero sequence part on the vector sequence.

[0029] Step 2: Build a multi-input artificial neural network, which can input a set of ligand and receptor vectors and output a prediction score for its specific function. The network consists of three parts, a feature extractor (module 1), a feature fusion device (module 2) and a classifier (module 3). Among them, the feature extractor (module 1) is used to extract the features in the amino acid sequence vector described in step 1, which is called a feature vector. In the present invention, this module 1 is divided into two or three branches, which are used to extract the feature vectors corresponding to the protein ligand and the receptor. The feature vectors of the ligand and the receptor can be aligned by pooling and deformation, and then the corresponding feature vectors are fused by splicing or addition to obtain a fused feature vector classifier (module 3) based on the fused feature vector for outputting the prediction result.

[0030] Step 3: Use transfer learning and meta-learning strategies to train the artificial neural network, the purpose of which is to optimize the parameters of the artificial neural network built in step 2 through the feature vector data corresponding to the labeled immune protein receptor ligand processed in step 1.

[0031] Step 4: Use the trained artificial neural network model to predict the receptor ligand data to be predicted.

[0032] According to the above embodiments, the present invention designs a multi-input network for inputting the amino acid sequences of ligands and receptors, so as to make predictions for the interaction functions between paired ligands and receptor proteins. The present invention adopts transfer learning, and pre-trains the feature extractor module based on an unsupervised training method, so as to deepen the model's understanding of proteins. The present invention adopts a meta-learning strategy training model based on task sampling, so as to balance the prediction performance of different receptor-specific functions and improve the generalization of new receptors. The present invention adopts a hyperparameter meta-learning strategy training model, and adjusts the hyperparameters through the prediction performance of the training set and the validation set, so as to improve the generalization of the model for new ligand receptor combinations. Among them, the immune protein ligand receptor includes B cells (antibody light, heavy chains and antigen sequences) and T cells (TCR and MHC and antigen fragments), and the functions described in the present invention include binding function and neutralization function.

[0033] like Figure 3 , Figure 4 and Figure 5 As shown, according to the above embodiment, preferably, the training process of the artificial neural network specifically includes: inputting processed sequence data; training the feature extractor by transfer learning (module 1); training the neural network by a meta-learning strategy based on task sampling; fine-tuning the neural network by a meta-learning strategy based on hyperparameters; and outputting the trained model. Specifically:

[0034] Step 3.1: Use transfer learning strategy to train the feature extractor (module 1) separately, and optimize the encoder-decoder structure unsupervised on the protein sequence data, so that the feature extractor (module 1) can act as an encoder to learn how to convert from the input sequence vector to the latent space feature vector. The training data of the protein sequence used can come from more general protein data, thereby increasing the depth of understanding of the feature extractor for the protein sequence.

[0035] Step 3.2: Train the artificial neural network based on the meta-learning strategy of task sampling, the purpose of which is to balance the network's understanding depth between different ligands or different receptors, and improve the generalization of different unseen receptors or unseen ligands. This method can be implemented using algorithms such as MAML or Reptile. The principle is as follows Figure 4As shown: Step 3.2.1, divide multiple sub-datasets according to different receptor (or ligand) combinations; Step 3.2.2, copy the training model (meta-model) to obtain a set of temporary model parameters; Step 3.2.3, randomly select a sub-dataset and sample from it; Step 3.2.4, optimize the temporary model parameters based on the sampled data and labels; Step 3.2.5, optimize the meta-model parameters based on the temporary model parameters.

[0036] Step 3.3: Fine-tune the artificial neural network based on the meta-learning strategy of hyperparameters, the purpose of which is to improve the overall generalization ability of the artificial neural network and avoid overfitting for the interaction function of the paired receptor and ligand to be predicted. The principle is as follows Figure 5 As shown: Step 3.3.1, divide the meta-training set and meta-validation set as needed; Step 3.3.2, initialize the hyperparameters and set the hyperparameter adjustment rate; Step 3.3.3, sample the meta-training data to optimize the network model parameters; Step 3.3.4, optimize the hyperparameters according to the indicators of the meta-training set and the meta-validation data set.

[0037] According to the above embodiment, preferably, the feature extractor module 1 of the artificial neural network adopts the ProteinBERT model, which may also be replaced by other network architectures or related models.

[0038] According to the above embodiment, preferably, in step 3.1, the feature extractor (module 1) is pre-trained and transferred by unsupervised learning based on the general protein sequence through mask reconstruction. There are multiple alternatives: the pre-training data can be replaced by more targeted multiple immune protein amino acid sequences, and the training strategy can also be replaced by other unsupervised or supervised methods such as clustering.

[0039] According to the above embodiment, preferably, the feature fusion module 2 of the artificial neural network uses global maximum pooling to remove sequence space features, and stacks the extracted feature vectors of ligands and receptors to fuse the two. This method can also be replaced by an attention mechanism to align feature dimensions, or simply fuse the features of the two by addition or the like.

[0040] According to the above embodiment, preferably, the classifier module 3 of the artificial neural network adopts a fully connected network, and the network structure can also be replaced by other artificial neural network structures.

[0041] According to the above embodiment, preferably, the meta-learning method based on task sampling described in step 3.2 adopts the reptile algorithm, and can also be replaced by other methods such as MAML that are based on sampling tasks to improve the generalization of the model for new tasks (unseen ligands or receptors).

[0042] According to the above embodiment, preferably, the hyperparameter meta-learning method described in step 3.3 selects the L1 regularization term loss function weight of the model weight and the dropout probability of the fully connected layer as the hyperparameters adjusted in the meta-learning strategy, and adjusts these hyperparameters by the difference in accuracy or F1 index between the meta-training set and the meta-validation set, thereby reducing the overfitting problem in network fine-tuning. The selected hyperparameters can be replaced by other hyperparameters such as L2 regularization term weights, data enhancement parameters, network short-circuit parameters, and the indicators for measuring the prediction performance of the meta-training set and the meta-validation set can also be replaced by other evaluation indicators.

[0043] According to the above embodiment, preferably, the task sampling-based meta-learning method of step 3.2 and the hyperparameter meta-learning method of step 3.3 can be combined and used in the same training cycle, thereby improving the generalization of the model from both the hyperparameter and data sampling perspectives.

[0044] like Figure 6 As shown, according to another embodiment of the present invention, an immune protein ligand receptor function prediction device 600 is also disclosed, including: a memory 601 for storing program instructions; a processor 602, for calling the program instructions stored in the memory to implement the immune protein ligand receptor function prediction method as described in the above embodiment.

[0045] According to another embodiment of the present invention, a computer-readable storage medium is disclosed. The computer-readable storage medium stores program codes, and the program codes are used to implement the immune protein ligand receptor function prediction method as described in the above embodiment.

[0046] All or part of the steps in the various methods of the above embodiments can be completed by controlling the relevant hardware through a program, and the program can be stored in a readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other readable medium that can be used to carry or store data.

[0047] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting immune protein ligand receptor function, characterized in that: include: Building an artificial neural network supporting multiple inputs, the artificial neural network is used to receive ligand and receptor vectors, and output prediction scores of specific functions, the artificial neural network at least includes a feature extractor, a feature fusion device and a classifier, wherein the feature extractor is used to extract feature vectors from amino acid sequence vectors, the feature fusion device is used to fuse feature vectors of protein ligands and receptors, and the classifier is used to classify the fused feature vectors to output prediction results; Using transfer learning strategy and meta-learning strategy to train the artificial neural network; Predicting the receptor ligand data to be predicted based on the trained artificial neural network; The steps of training the artificial neural network using a transfer learning strategy specifically include: A transfer learning strategy is used to train the feature extractor separately, and an encoder-decoder structure is used to perform unsupervised optimization on protein sequence data, so that the feature extractor can act as an encoder to learn how to convert from an input sequence vector to a latent space feature vector; The steps of using a meta-learning strategy to train the artificial neural network specifically include: The artificial neural network is trained using a meta-learning strategy based on task sampling to balance the network's understanding depth between different ligands or different receptors and to improve generalization to unseen receptors or unseen ligands; A meta-learning strategy based on hyperparameters is used to fine-tune the artificial neural network, targeting the interaction function of paired receptors and ligands to be predicted, improving the overall generalization ability of the artificial neural network and avoiding overfitting.

2. The method for predicting immune protein ligand receptor function according to claim 1, characterized in that: Also includes: Encode each amino acid of the annotated ligand and receptor amino acid sequences to obtain an amino acid vector sequence; Inserting a zero-filled vector to make the amino acid vector sequence reach a fixed length; The non-zero sequence part is slid on the vector sequence to reduce the neural network's dependence on the absolute position features of the sequence dimension.

3. The method for predicting immune protein ligand receptor function according to claim 1, characterized in that: The training data comes from generalized protein data to increase the depth of the feature extractor's understanding of protein sequences.

4. The method for predicting immune protein ligand receptor function according to claim 1, characterized in that: The meta-learning strategy based on task sampling and the meta-learning strategy based on hyper-parameters are combined and used in the same training cycle, thereby improving the generalization of the model from both the hyper-parameter and data sampling perspectives.

5. The method for predicting immune protein ligand receptor function according to claim 1, characterized in that: The hyper-parameter-based meta-learning strategy selects the L1 regularization term loss function weight of the model weight and the dropout probability of the fully connected layer as the hyper-parameters adjusted in the meta-learning strategy, and adjusts the hyper-parameters by the difference in accuracy or F1 index between the meta-training set and the meta-validation set, thereby reducing the overfitting problem in network fine-tuning.

6. The method for predicting immune protein ligand receptor function according to any one of claims 1 to 3, characterized in that: The feature extractor is implemented based on the ProteinBERT model.

7. A device for predicting immune protein ligand receptor function, characterized in that: include: A memory for storing program instructions; A processor, configured to call the program instructions stored in the memory to implement the immune protein ligand receptor function prediction method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program code, and the program code is used to implement the immune protein ligand receptor function prediction method according to any one of claims 1 to 6.

Images