Protein-target affinity prediction method and system based on small sample learning, storage medium and terminal
By using dynamic training and recurrent shallow neural networks, the overfitting problem caused by insufficient samples in protein-target affinity prediction is solved, thereby improving the prediction accuracy and reliability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2024-04-18
- Publication Date
- 2026-05-12
AI Technical Summary
In existing technologies, protein-target affinity prediction is prone to overfitting and reduced generalization ability due to insufficient sample size. Furthermore, increasing the number of iterations in conventional methods exacerbates the overfitting phenomenon.
We employ a dynamic training method and a cyclically running shallow neural network. The training set is supplemented with a validation set. We combine One-Hot encoding and convolutional neural network to extract features, and use the rdkit library to extract drug target features. We use nested loops to train the shallow neural network to reduce the number of iterations and avoid overfitting.
It effectively reduces the risk of overfitting, improves the model's fitting and generalization abilities, and enhances the accuracy and reliability of protein-target affinity prediction.
Smart Images
Figure CN118522343B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of protein prediction technology, and in particular to a protein-target affinity prediction method, system, storage medium, and terminal based on few-sample learning. Background Technology
[0002] In protein-target affinity prediction, insufficient available samples are a common problem. Protein-target affinity prediction requires a large amount of known protein-ligand complex affinity data to train models and learn patterns of protein-ligand interactions. However, experimental determination of protein-ligand complexes is expensive, time-consuming, and challenging, resulting in a relatively small amount of available training datasets.
[0003] Insufficient sample size can lead to the following problems: Due to the small training set, the model may focus too much on noise and special samples in the training set, failing to capture the true features of the data. Common methods like increasing the number of iterations to improve model learning may become inapplicable, as a higher number of training iterations causes the model to memorize more details from the training set, further exacerbating overfitting and reducing the model's generalization ability on new data. Secondly, the limited number of samples available for training means the neural network model cannot be adequately trained, making it prone to overfitting.
[0004] Therefore, in few-sample learning problems such as protein-target affinity prediction, there are two main drawbacks. The first is overfitting due to excessive iterations; the second is overfitting due to insufficient available training data. Summary of the Invention
[0005] The purpose of this invention is to overcome the problems existing in current protein-target affinity prediction and to provide a protein-target affinity prediction method, system, storage medium, and terminal based on few-shot learning. Without significantly increasing the time and space complexity of the original algorithm, it can effectively resist the risk of overfitting and provides a new perspective on resisting overfitting in few-shot learning problems.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] Firstly, a protein-target affinity prediction method based on few-sample learning is provided, comprising the following steps:
[0008] S1. Protein sequence feature extraction;
[0009] S2. Extraction of SMILES code features for drug targets;
[0010] S3. Combine the feature data extracted in steps S1 and S2 to obtain protein-target feature data.
[0011] S4. Construct a protein-target affinity prediction model and train the protein-target affinity prediction model using the protein-target feature data; wherein, the protein-target feature data is divided into a training set, a validation set, and a test set; during the training process, dynamic training is performed by supplementing the training set with the validation set, and when the number of validation sets drops to a certain threshold, the supplementation of the training set is stopped.
[0012] S5. Input the test set data into the trained protein-target affinity prediction model to obtain the protein-target affinity prediction results.
[0013] As a further improvement of the present invention, step S4 further includes:
[0014] A loop is nested outside the protein-target affinity prediction model, and the model is trained using a method of running a shallow neural network in a loop.
[0015] Each iteration records the model's evaluation metric parameters at the last iteration and determines whether the maximum number of iterations has been reached. If it has, the iteration exits and selects the optimal evaluation metric parameters from all iterations as the model's parameters.
[0016] As a further improvement of the present invention, the protein sequence feature extraction includes:
[0017] After quantifying the protein sequence using One-Hot encoding, a convolutional neural network is used to extract its features.
[0018] As a further improvement of the present invention, the extraction of the drug target SMILES code features includes:
[0019] Convert SMILES codes into molecular objects;
[0020] The corresponding functions in the rdkit library are called to extract the molecular descriptor features and molecular geometric features of the drug target.
[0021] As a further improvement to the present invention, the ratio of the training set, validation set and test set is 3:1:1.
[0022] As a further improvement of the present invention, the method of using a validation set to supplement the training set for dynamic training includes:
[0023] In each iteration of the model, s data points are randomly selected from the validation set and added to the training set.
[0024] As a further improvement of the present invention, the step of stopping the replenishment of the training set when the number of validation sets drops to a certain threshold includes:
[0025] When the number of validation samples reaches 5% of the total number of samples, the training set is no longer supplemented.
[0026] Secondly, a protein-target affinity prediction system based on few-sample learning is provided, including:
[0027] The protein feature extraction module is configured to extract protein sequence features.
[0028] The target feature extraction module is configured for drug target SMILES code feature extraction.
[0029] The feature fusion module is configured to fuse protein features and target features to obtain protein-target feature data.
[0030] The prediction model construction and training module is configured to construct a protein-target affinity prediction model and train the protein-target affinity prediction model using the protein-target feature data; wherein, the protein-target feature data is divided into a training set, a validation set, and a test set; during the training process, the training set is dynamically supplemented using the validation set, and the supplementation of the training set is stopped when the number of validation sets drops to a certain threshold;
[0031] The protein-target affinity prediction module is configured to input test set data into a pre-trained protein-target affinity prediction model to obtain protein-target affinity prediction results.
[0032] Thirdly, a computer storage medium is provided, on which computer instructions are stored, wherein the computer instructions, when executed, perform the relevant steps in the protein-target affinity prediction method based on few-sample learning.
[0033] Fourthly, a terminal is provided, including a memory and a processor, wherein the memory stores computer instructions that can be executed on the processor, characterized in that the processor executes relevant steps in the protein-target affinity prediction method based on few-sample learning when executing the computer instructions.
[0034] It should be further noted that the technical features corresponding to the above options can be combined or substituted to form new technical solutions if there is no conflict.
[0035] Compared with the prior art, the beneficial effects of the present invention are:
[0036] (1) This invention proposes a dynamic training method. By using a validation set to supplement the training set during the training process, the data can be better utilized, the model's fitting ability and generalization ability can be improved, and the risk of overfitting due to insufficient training data in small sample learning can be effectively resisted, thereby improving the accuracy and reliability of protein-target affinity prediction.
[0037] (2) This invention designs a method for cyclically running shallow neural networks to replace training methods that use a higher number of iterations. It can achieve better training results with fewer iterations, avoids the overfitting problem caused by too many iterations, and maintains the predictive and generalization abilities of the model, effectively reducing the risk of overfitting and further improving the accuracy and reliability of protein-target affinity prediction. Attached Figure Description
[0038] Figure 1 This is a flowchart illustrating a protein-target affinity prediction method based on few-sample learning, as shown in an embodiment of the present invention.
[0039] Figure 2 This is a schematic diagram illustrating the numerical representation of a protein sequence according to an embodiment of the present invention;
[0040] Figure 3 This is a schematic diagram of One-Hot encoding as shown in an embodiment of the present invention;
[0041] Figure 4 This is a schematic diagram of dynamic training as shown in an embodiment of the present invention;
[0042] Figure 5 This is a schematic diagram of a shallow neural network trained in a loop, as shown in an embodiment of the present invention. Detailed Implementation
[0043] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0044] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0045] Example 1
[0046] In one exemplary embodiment, reference is made to Figure 1 This paper provides a protein-target affinity prediction method based on few-shot learning, which includes the following steps:
[0047] S1. Protein sequence feature extraction;
[0048] S2. Extraction of SMILES code features for drug targets;
[0049] S3. Combine the feature data extracted in steps S1 and S2 to obtain protein-target feature data.
[0050] S4. Construct a protein-target affinity prediction model and train the protein-target affinity prediction model using the protein-target feature data; wherein, the protein-target feature data is divided into a training set, a validation set, and a test set; during the training process, dynamic training is performed by supplementing the training set with the validation set, and when the number of validation sets drops to a certain threshold, the supplementation of the training set is stopped.
[0051] S5. Input the test set data into the trained protein-target affinity prediction model to obtain the protein-target affinity prediction results.
[0052] Specifically, before model training, feature extraction of the protein sequence is necessary. Convolutional Neural Networks (CNNs) are used for feature extraction. Through convolutional and pooling layers, CNNs can capture local features within the protein sequence. To extract features using a CNN, the protein sequence must first be converted into numerical data; here, One-Hot encoding is used for numerical conversion. The encoded sequence is then input into the CNN. By stacking multiple convolutional and fully connected layers, the CNN can extract higher-level features and perform classification or regression tasks. Convolutional layers use a set of learnable filters to convolve the input sequence, generating a series of feature maps. Pooling layers are used to reduce the dimensionality of the feature maps and extract higher-level features; multi-channel convolution can capture features of different dimensions within the protein sequence.
[0053] Before model training, feature extraction of the SMILES code for the drug target is necessary. For this, we can use the third-party Python library rdkit to process the SMILES code and extract the molecular descriptor and geometric features of the drug target. These features are very useful for describing the physicochemical properties and structural characteristics of molecules, including common features such as LogP value, polar surface area (TPSA), exact molecular weight, number of rotatable bonds, number of hydrogen bond acceptors, number of aromatic rings, and ring number. By extracting these features, we can obtain important information about the drug target molecule, providing a foundation for subsequent model training and prediction.
[0054] Furthermore, it is necessary to effectively fuse protein sequence features with drug target features to construct a complete protein-target feature representation, which can then be used to predict the interaction or affinity between the two. The specific fusion process is as follows:
[0055] The preprocessed protein and target feature data (numerical protein feature analysis and chemical structuring of drug target features) are directly concatenated to form a new joint feature vector. For example, in a specific study, if the protein feature vector is n-dimensional and the target feature vector is m-dimensional, then the fused protein-target feature vector will be an (n+m)-dimensional feature vector.
[0056] Furthermore, in few-shot learning, the limited amount of data available for training makes overfitting highly likely. Typically, a validation set is used to check for overfitting during model training; however, separating the validation set from the total available data further reduces the amount of data actually used for training. Since the validation set only needs to reflect the trend of the loss function, it doesn't need to precisely characterize the function. To address this, we employ a dynamic training method that supplements the training set with the validation set during training. As training progresses, the size of the training set dynamically increases, while the size of the validation set dynamically decreases. A dynamic stopping condition is set in advance, for example, when the size of the validation set reaches 5% of the total samples (the percentage can be adjusted appropriately according to specific circumstances), the dynamic process stops, meaning the size of both the validation and training sets no longer changes. By using the validation set to supplement the training set during training, we can better utilize the data, improve the model's fitting and generalization abilities, effectively reduce the risk of overfitting, and improve the accuracy and reliability of protein-target affinity prediction.
[0057] Example 2
[0058] Based on a further improvement of Embodiment 1, step S4 further includes:
[0059] A loop is nested outside the protein-target affinity prediction model, and the model is trained using a method of running a shallow neural network in a loop.
[0060] Each iteration records the model's evaluation metric parameters at the last iteration and determines whether the maximum number of iterations has been reached. If it has, the iteration exits and selects the optimal evaluation metric parameters from all iterations as the model's parameters.
[0061] Specifically, in few-shot learning, the limited training data often prevents the model from being fully trained. In such cases, increasing the number of iterations is often used to improve the training effect. However, this approach introduces potential problems that require careful handling. For example, increasing the number of iterations may lead to overfitting the training data, resulting in decreased performance on test data. Therefore, this invention proposes a novel method for training neural network models: using iterative training of shallow neural networks instead of higher iteration counts to mitigate the problems caused by data limitations. Specifically, after each regular training iteration (loop), the evaluation metric parameters are calculated, and it is determined whether the maximum number of iterations (rounds) has been reached. If the maximum number is reached, training stops. Finally, after training stops, the optimal set of evaluation metric parameter values is taken as the training result.
[0062] This embodiment achieves good training results with fewer iterations by using a shallow neural network that runs in a loop. This avoids overfitting caused by excessive iterations, while maintaining the model's predictive and generalization abilities. It effectively reduces the risk of overfitting and further improves the accuracy and reliability of protein-target affinity prediction.
[0063] Example 3
[0064] Based on Example 2, this invention uses a test set to evaluate the trained model. Features from the test set are input into the model to obtain the predicted protein-target affinity. Then, the model's predictions are compared with the true labels in the test set, and accuracy and other evaluation metrics, such as precision, recall, and F1 score, are calculated. These can be calculated using a confusion matrix or other relevant metrics. Finally, the model's performance is evaluated by analyzing the prediction accuracy and other evaluation metrics. By comprehensively evaluating the model, calculating the prediction accuracy, and analyzing the results, the performance of the protein-target affinity prediction model on the test set can be fully assessed and explained, providing guidance and reference for subsequent model improvement and application.
[0065] For example, high precision and recall indicate a good ability to predict protein-target affinity. Meanwhile, the F1 score, which considers both precision and recall, is the harmonic mean of the two. A higher F1 score indicates a better balance between precision and recall. In protein-target affinity prediction, a high F1 score means the model can accurately predict positive samples and capture true positive samples well.
[0066] The above steps allow us to evaluate the prediction accuracy of the protein-target affinity prediction model on the test set. This helps us understand the model's generalization ability and performance, and allows for model improvement and optimization.
[0067] Example 4
[0068] Based on Example 2, specific implementation examples for each step are given. These include the following:
[0069] 1. Protein sequence feature extraction
[0070] First, the protein sequence is numerically converted using One-Hot encoding. Protein sequences are combinations of English letters, which is non-numerical data. To extract its features using a convolutional neural network, it must first be numerically converted. For example... Figure 2 As shown. One-hot encoding, also known as one-bit valid encoding, uses an N-bit state register to encode N states. Each state has its own independent register bit, and at any given time, only one bit is valid. Since each protein sequence consists of 20 or fewer amino acids, N = 20 in the numeration process using One-Hot encoding. Specifically, One-Hot encoding represents each amino acid as a sparse vector (by a 20-bit state register), where only one element is 1 and the rest are 0. The encoding process is as follows. Figure 3 As shown.
[0071] After encoding, the numerical data of the protein sequence is fed into a convolutional neural network (CNN) for feature extraction. The protein sequence is input into the CNN model, and the convolutional neural network model, including convolutional layers, pooling layers, and fully connected layers, is designed and constructed to extract features from the protein sequence.
[0072] 2. Drug target SMILES code feature extraction
[0073] Feature extraction of drug target SMILES codes is based on the rdkit library, a widely used Python third-party library in cheminformatics and computational chemistry. When using rdkit for extraction, the SMILES codes must first be converted into molecular objects:
[0074] molecule=Chem.MolFromSmiles(smiles)
[0075] Next, the corresponding functions in the library can be called to extract the molecular descriptor features and molecular geometric features of the drug target. By extracting these features of the drug target molecule, we can obtain important information about the physicochemical properties and structural characteristics of the drug molecule, providing a foundation for subsequent model training and prediction.
[0076] 3. Training is performed using the dynamic neural network training method and the iterative shallow network training method.
[0077] (1) Dynamic training method for neural networks
[0078] During the training of an artificial neural network, the dynamic process of the validation set and training set is as follows: Figure 4 As shown.
[0079] In machine learning, when available data is limited, a 3:1:1 ratio is typically used to divide the training, validation, and test sets. Therefore, we input 300 protein-target sample data points and 100 protein-target validation data points into the network model. The following are some important parameter settings:
[0080] • P = 25 (Exit dynamic training when the number of data points in the validation set reaches 5% of the total data volume)
[0081] • s = 1 (One data point is drawn from the validation set and added to the training set each time)
[0082] • epochs = 100 (model iterations 100 times)
[0083] Compared to conventional neural network model training methods, the dynamic training method proposed in this invention randomly selects *s* data points from the validation set and adds them to the training set in each model iteration. After the model completes N full forward and backward propagations, the transformation between the training and validation sets is as follows:
[0084] Number of new training sets = Number of old training sets + N*s
[0085] Number of new validation sets = Number of old validation sets - N*s
[0086] By adjusting parameter P to set a dynamic stopping condition, once the pre-set threshold for the number of validation sets is reached, if the epoch progress has reached 80%, the remaining 20% of iterations can be performed according to the normal training mode. Theoretically, this dynamic training method can effectively reduce the risk of overfitting from small-sample learning and improve the accuracy and reliability of protein-target affinity prediction.
[0087] (2) Cyclic shallow network method
[0088] This invention proposes a novel method for training artificial neural network models, namely, using cyclic training of shallow neural networks to replace the use of a high number of iterations, thereby alleviating the problems caused by data limitations.
[0089] The specific process is as follows: Figure 5 As shown:
[0090] In conventional neural network models, when the number of epochs is 100, to obtain better parameters, the number of iterations is usually increased, for example, by directly changing the number of epochs to 1000. In this invention, a primary neural network model is encapsulated, and then a loop is nested inside it, with the number of loops set to 10. In this way, we obtain a pseudo-iteration count:
[0091] 10 * 100 = 1000 times
[0092] In each iteration, record the model parameters and evaluation metric parameters up to the last iteration. Upon exiting the loop, compare the evaluation metric parameters 10 times to obtain the optimal model parameters. Here, we define a method for comparing the evaluation metric parameters:
[0093] Min f=a1×ASE+(-a2)×R2_score+a3×bias+...+a n ×other criteria
[0094]
[0095] a i >0, i=1, 2, 3,..., n
[0096] In the above formula, a1~a n The parameters represent the importance of each evaluation metric: MSE (mean squared error), R²_score (coefficient of determination), bias (bias), and other criteria (other criteria). The parameters before R²_score are negative because a higher R²_score (closer to 1) indicates better model performance. However, this function is designed to find the minimum value, i.e., to standardize the objective function, so the parameter 'a' of R²_score is artificially set. i It became -a i In practical applications, the values can be set as needed. Before using the above formula, all evaluation indicators need to be normalized to avoid the influence of small-value indicators being eliminated by evaluation indicators with excessively large values.
[0097] By using this cyclical shallow network method, overfitting caused by excessive iterations can be avoided. Theoretically, this can effectively reduce the risk of overfitting in small-sample learning and improve the accuracy and reliability of protein-target affinity prediction.
[0098] Example 5
[0099] Based on the same inventive concept as in Example 1, a protein-target affinity prediction system based on few-sample learning is provided, comprising:
[0100] The protein feature extraction module is configured to extract protein sequence features.
[0101] The target feature extraction module is configured for drug target SMILES code feature extraction.
[0102] The feature fusion module is configured to fuse protein features and target features to obtain protein-target feature data.
[0103] The prediction model construction and training module is configured to construct a protein-target affinity prediction model and train the protein-target affinity prediction model using the protein-target feature data; wherein, the protein-target feature data is divided into a training set, a validation set, and a test set; during the training process, the training set is dynamically supplemented using the validation set, and the supplementation of the training set is stopped when the number of validation sets drops to a certain threshold;
[0104] The protein-target affinity prediction module is configured to input test set data into a pre-trained protein-target affinity prediction model to obtain protein-target affinity prediction results.
[0105] Example 6
[0106] Based on the same inventive concept as in Embodiment 1, the present invention provides a computer storage medium storing computer instructions, which, when executed, perform relevant steps in the protein-target affinity prediction method based on few-sample learning.
[0107] Based on this understanding, the technical solution of this embodiment, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0108] Example 7
[0109] Based on the same inventive concept as in Embodiment 1, the present invention provides a terminal, including a memory and a processor. The memory stores computer instructions that can be executed on the processor. When the processor executes the computer instructions, it performs the relevant steps in the protein-target affinity prediction method based on few-sample learning.
[0110] The processor may be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement the present invention.
[0111] The embodiments of the subject matter and functional operation described in this specification can be implemented in: tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing device or for controlling the operation of a data processing device. Alternatively or additionally, the program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information and transmit it to a suitable receiving device for execution by the data processing device.
[0112] The processing and logic flow described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform corresponding functions by operating on input data and generating output. The processing and logic flow can also be executed by dedicated logic circuitry—such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits), and the device can also be implemented as dedicated logic circuitry.
[0113] Suitable processors for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, the central processing unit receives instructions and data from read-only memory and / or random access memory. The basic components of a computer include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as disks, magneto-optical disks, or optical disks, or the computer will be operatively coupled to such mass storage devices to receive data from or transfer data to them, or both. However, a computer is not required to have such devices. Furthermore, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.
[0114] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.
[0115] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.
[0116] The above detailed embodiments are a description of the present invention. It should not be considered that the specific embodiments of the present invention are limited to these descriptions. For those skilled in the art, several simple deductions and substitutions can be made without departing from the concept of the present invention, and all of these should be considered to fall within the protection scope of the present invention.
Claims
1. A protein-target affinity prediction method based on few-sample learning, characterized in that, Includes the following steps: S1. Protein sequence feature extraction; S2. Extraction of SMILES code features for drug targets; S3. Combine the feature data extracted in steps S1 and S2 to obtain protein-target feature data. S4. Construct a protein-target affinity prediction model and train the model using the protein-target feature data; wherein, the protein-target feature data is divided into a training set, a validation set, and a test set; during training, the training set is dynamically supplemented using the validation set, and supplementation is stopped when the number of validation sets drops to a certain threshold; step S4 further includes: A loop is nested outside the protein-target affinity prediction model, and the model is trained using a method of running a shallow neural network in a loop. Each iteration records the model's evaluation metric parameters at the last iteration and determines whether the maximum number of iterations has been reached. If it has, the iteration exits and selects the optimal evaluation metric parameters from all iterations as the model's parameters. S5. Input the test set data into the trained protein-target affinity prediction model to obtain the protein-target affinity prediction results.
2. The protein-target affinity prediction method based on few-sample learning according to claim 1, characterized in that, The protein sequence feature extraction includes: After quantifying the protein sequence using One-Hot encoding, a convolutional neural network is used to extract its features.
3. The protein-target affinity prediction method based on few-sample learning according to claim 1, characterized in that, The extraction of SMILES code features for drug targets includes: Convert SMILES codes into molecular objects; The corresponding functions in the rdkit library are called to extract the molecular descriptor features and molecular geometric features of the drug target.
4. The protein-target affinity prediction method based on few-sample learning according to claim 1, characterized in that, The ratio of training set, validation set, and test set is 3:1:
1.
5. The protein-target affinity prediction method based on few-sample learning according to claim 1, characterized in that, The method of using a validation set to supplement the training set for dynamic training includes: In each iteration of the model, s data points are randomly selected from the validation set and added to the training set.
6. The protein-target affinity prediction method based on few-sample learning according to claim 5, characterized in that, The step of stopping the replenishment of the training set when the number of validation sets drops to a certain threshold includes: When the number of validation samples reaches 5% of the total number of samples, the training set is no longer supplemented.
7. A protein-target affinity prediction system based on few-shot learning, characterized in that, include: The protein feature extraction module is configured to extract protein sequence features. The target feature extraction module is configured for drug target SMILES code feature extraction. The feature fusion module is configured to fuse protein features and target features to obtain protein-target feature data. The prediction model construction and training module is configured to construct a protein-target affinity prediction model and train it using the protein-target feature data. The protein-target feature data is divided into a training set, a validation set, and a test set. During training, the training set is dynamically supplemented using the validation set; when the number of validation sets drops to a certain threshold, the supplementation stops. The prediction model construction and training module is also used to nest a loop around the protein-target affinity prediction model and train it using a method of looping through a shallow neural network. Each iteration records the model's evaluation metric parameters at the last iteration and determines whether the maximum number of iterations has been reached. If it has, the iteration exits and selects the optimal evaluation metric parameters from all iterations as the model's parameters. The protein-target affinity prediction module is configured to input test set data into a pre-trained protein-target affinity prediction model to obtain protein-target affinity prediction results.
8. A computer storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed, they perform the relevant steps in the protein-target affinity prediction method based on few-sample learning as described in any one of claims 1-6.
9. A terminal, comprising a memory and a processor, wherein the memory stores computer instructions that run on the processor, characterized in that, When the processor executes computer instructions, it performs the relevant steps in the protein-target affinity prediction method based on few-sample learning as described in any one of claims 1-6.