Protein point mutation effect zero sample prediction system based on self-distillation contrast learning

Abstract

The invention discloses a protein point mutation effect zero sample prediction system based on self-distillation contrast learning. The system comprises a data import module, a deep neural network model loading module and a mutation effect prediction module. A deep neural network model loaded by the deep neural network model loading module is composed of a feature processor and a feature fusion device, wherein the feature processor comprises a WDB module and an improved Transform module, and the deep neural network model loading module is used for generating deep features highly sensitive to mutation; and a mutation effect prediction module calls the deep neural network model to calculate a log-likelihood difference value of wild-type and mutant amino acids on a mutation site so as to realize efficient zero sample prediction of any unseen protein. According to the method, an innovative deep learning architecture is combined with zero sample prediction, fine adjustment is not needed, and accurate and efficient technical support with more generalization ability is provided for protein engineering and disease mechanism research.

Description

Technical Field

[0001] This invention relates to the technical field of protein point mutation effect prediction, and in particular to a zero-sample prediction system for protein point mutation effects based on self-distillation contrastive learning. Background Technology

[0002] Proteins are the primary carriers of life activities, and their amino acid sequences determine their three-dimensional structure and biological functions. A single point mutation in the sequence, i.e., the substitution of an amino acid, can severely affect a protein's stability, activity, or interactions with other molecules, thereby triggering various major diseases such as genetic disorders and cancer, or producing desired beneficial effects in protein engineering. Therefore, accurately predicting the harmful or functional effects of point mutations is a key issue in modern biomedical research and protein engineering, possessing significant scientific and applied value.

[0003] To address this issue, various prediction methods have been developed. Early methods primarily relied on prior biological knowledge, such as using conservation scores like BLOSUM or PAM matrices, or calculating changes in the physicochemical properties of proteins before and after mutations to assess their impact. Subsequently, traditional machine learning methods, such as Support Vector Machines (SVM) and Random Forests, were used to combine multiple hand-designed features for comprehensive prediction. In recent years, with the development of deep learning, end-to-end models, represented by Convolutional Neural Networks (CNN) or Transformers, have been introduced into this field. These models can automatically learn features and make predictions directly from one-dimensional sequences or even three-dimensional structures of proteins.

[0004] However, despite the progress made in existing technologies, their inherent limitations are becoming increasingly apparent when addressing the need for accurate and efficient prediction of mutation effects. Firstly, at the feature extraction level, existing deep learning models lack sufficient feature extraction capabilities and their sensitivity to key mutation information needs improvement. For example, the fixed convolutional kernels of standard CNNs struggle to effectively capture subtle local changes in sequences caused by single amino acid substitutions; while the standard Transformer model excels at capturing global long-range dependencies, its relatively fixed nonlinear mapping in the feedforward network may not optimally balance local details and global context. This results in existing methods underutilizing low-level semantic information and being insufficiently sensitive to subtle perturbations caused by mutation sites, thus limiting the performance ceiling of prediction models.

[0005] Secondly, at the model training level, existing methods employ relatively simplistic learning strategies, generally lacking explicit alignment of wild-type (WT) and mutant (MUT) features and effective model regularization. When processing wild-type and mutant sequences, they often treat them as two independent inputs, lacking a mechanism to explicitly compare and align their feature representations at multiple levels. This makes it difficult for the model to focus on the key semantic differences between the two, resulting in insufficient feature alignment. Simultaneously, deep models are prone to overfitting, especially when labeled data on protein mutation effects is relatively scarce. The model's generalization ability is severely affected, making it difficult to achieve robust training results in small-sample environments.

[0006] More fundamentally, most current advanced prediction models heavily rely on supervised learning paradigms and lack zero-shot prediction capabilities, resulting in high application costs and limited generalization. These models must be trained on mutation datasets containing a large number of labels, which not only makes building high-quality labeled datasets a time-consuming and labor-intensive task, but more importantly, the model's predictive ability is constrained by its training data. When encountering protein families or novel mutations not seen in the training set, their performance often drops significantly. Therefore, existing technologies generally lack zero-shot prediction capabilities—that is, the ability to directly predict the effects of any mutation on any protein without relying on any mutation effect labels for fine-tuning. This greatly limits the universality and application efficiency of these methods in real-world scientific research and industrial scenarios.

[0007] In view of this, developing a protein point mutation effect prediction system that can deeply mine sequence features, effectively align semantics before and after mutation, and ultimately achieve high-precision zero-sample prediction is a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a zero-sample prediction system for protein point mutation effects based on self-distillation contrastive learning. This system can more sensitively capture small sequence changes caused by point mutations, explicitly align wild-type and mutant features, enhance the generalization ability of the model, and achieve accurate and efficient prediction of mutation effects for any unseen mutations.

[0009] To achieve the above objectives, the technical solution provided by this invention is: a zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning, comprising:

[0010] The data import module is used to acquire and load wild-type amino acid sequences and specified mutation information, and to preprocess them to obtain mutant amino acid sequences.

[0011] A deep neural network model loading module is used to load a pre-trained deep neural network model. This model includes a feature processor and a feature fusion unit. The feature processor includes a Wavelet-DenseBlock module (WDB module) and an improved Transformer module. The WDB module replaces traditional convolution operations with learnable wavelet transforms, introduces first-order Gaussian derivative wavelets to enhance the model's sensitivity to local changes, and normalizes and compensates for features using learnable scale and translation parameters. This allows the model to capture edge information in the sequence at multiple scales, thereby improving the feature's expressive power. The improved Transformer module replaces the original feedforward neural network layer with a Kolmogorov-Arnold Network module. The Kolmogorov-Arnold Network module, by introducing learnable wavelet transforms, changes the nonlinear feature mapping from a fixed mapping method to a data-driven, weighted, and flexible method. This enables more efficient collaborative capture of local details and long-range dependencies in the sequence, improving the quality and information density of the final generated features.

[0012] The mutation effect prediction module is used to call the pre-trained deep neural network model loaded by the deep neural network model loading module, take the wild-type amino acid sequence and the mutant amino acid sequence as input to the deep neural network model, detect the log-likelihood change caused by a single amino acid substitution based on the deep neural network model, calculate and output a score quantifying the mutation effect based on the log-likelihood change, and finally obtain the mutation effect prediction result.

[0013] Furthermore, the data import module includes a data loading module and a data preprocessing module, wherein:

[0014] The data loading module reads wild-type amino acid sequence data and mutation information data without real label data from the public protein databases UniProt and PDB. The mutation information data includes the protein ID, the amino acid before the mutation, the amino acid sequence position where the mutation occurred (hereinafter referred to as the mutation site), and the amino acid after the mutation.

[0015] The data preprocessing module receives wild-type amino acid sequence data and specified mutation information data from the data loading module. Based on each mutation information, it performs an amino acid substitution operation at the mutation site on the corresponding wild-type amino acid sequence, replacing the unmutated amino acid with the mutated amino acid, thereby generating the corresponding mutant amino acid sequence.

[0016] Furthermore, the deep neural network model is generated through a specialized training system, which includes:

[0017] The training data import module is used to acquire and load wild-type amino acid sequences containing real label data, specified mutation information, and corresponding real label data. Based on each mutation information, an amino acid substitution operation is performed on the corresponding wild-type amino acid sequence to generate the corresponding mutant amino acid sequence.

[0018] The feature extraction module is used to integrate the wild-type amino acid sequence containing real label data and the mutant amino acid sequence. Using a fixed pre-trained protein language model (PLM) as the base model, the wild-type amino acid sequence and the corresponding mutant amino acid sequence are converted into feature vector form to obtain wild-type shallow feature vector and mutant shallow feature vector. The feature processor of the deep neural network model is used to process the wild-type shallow feature vector and mutant shallow feature vector to obtain wild-type deep feature vector and mutant deep feature vector.

[0019] The feature fusion module uses the feature fusioner of the deep neural network model to fuse the wild-type deep feature vector and the mutant deep feature vector output by the feature extraction module into an enhanced fused feature vector.

[0020] The self-distillation contrastive learning module is used to perform self-distillation learning on the wild-type shallow feature vector and mutant shallow feature vector, as well as the wild-type deep feature vector and mutant deep feature vector output by the feature extraction module, and the fused feature vector output by the feature fusion module to obtain self-distillation loss; and to perform contrastive learning on the wild-type shallow feature vector and mutant shallow feature vector, as well as the wild-type deep feature vector and mutant deep feature vector output by the feature extraction module to obtain shallow contrastive loss and deep contrastive loss.

[0021] The supervised loss calculation module includes a Kolmogorov-Arnold Network (KAN) classifier, which is connected after the feature fusion module. It is used to predict the mutation effect of protein point mutations. The supervised loss is calculated using the cross-entropy loss function through real label data and prediction results.

[0022] The model optimization module is used to perform end-to-end training on the deep neural network model using the self-distillation loss, shallow contrast loss, and deep contrast loss obtained by the self-distillation contrast learning module, as well as the supervised loss, until the model converges, and finally generates a pre-trained deep neural network model, that is, the deep neural network model with optimal performance.

[0023] Furthermore, the feature extraction module includes a shallow feature extraction module and a deep feature extraction module, wherein:

[0024] The shallow feature extraction module uses a fixed pre-trained protein language model as the base model to extract wild-type amino acid sequences. and the corresponding mutant amino acid sequence The formula for converting it into feature vector form is as follows:

[0025] ;

[0026] In the formula, Represents the amino acid sequence. ,in Represents the wild type. Represents a mutant type; This refers to the pre-trained protein language model; This represents the output feature vector, with dimension 1. ,in It refers to the batch size. It is the sequence length. It is the embedding dimension; the shallow feature extraction module ultimately outputs a wild-type shallow feature vector. and mutant shallow feature vectors ;

[0027] The deep feature extraction module utilizes the feature processor of the deep neural network model, and processes the wild-type shallow feature vector and the mutant shallow feature vector through the WDB module and the improved Transformer module to obtain the wild-type deep feature vector and the mutant deep feature vector. The processing steps include:

[0028] 1.1) The wild-type shallow feature vector output by the shallow feature extraction module and mutant shallow feature vectors The features are respectively used as initial inputs to the WDB module, and the processed features are output as feature vectors. and ;

[0029] 1.2) The feature vector and The inputs are then fed into the improved Transformer module for sequence modeling to obtain wild-type deep feature vectors. and mutant deep feature vectors ;

[0030] The WDB module consists of n Dense Blocks and one Transition Layer. The input of the m-th (m≤n)th Dense Block is obtained by concatenating the initial features and the outputs of the first m-1 layers in the channel dimension, performing wavelet transform, batch normalization, activation function, and random deactivation to finally obtain the output of the m-th layer. The initial features and the outputs of the first n Dense Blocks are concatenated in the channel dimension and used as the input of the Transition Layer. The Transition Layer consists of convolution operation, batch normalization, nonlinear activation function, and random deactivation.

[0031] The improved Transformer module includes a multi-layer multi-head self-attention mechanism and a Kolmogorov-Arnold Network module that introduces learnable wavelet transform. Each layer uses the standard multi-head self-attention mechanism to model the long-range dependencies of the input features. The Kolmogorov-Arnold Network module, which introduces learnable wavelet transform, performs nonlinear feature mapping. The final output is passed to the next layer or used for downstream tasks.

[0032] Furthermore, the feature fusion module utilizes the feature fusioner of the deep neural network model and employs a dynamic gating fusion mechanism to combine the wild-type deep feature vector output by the deep feature extraction module. and mutant deep feature vectors The fusion is performed to obtain the enhanced fused feature vector. It includes the following steps:

[0033] 2.1) Feature Projection: Projecting the wild-type deep feature vector and mutant deep feature vectors Each wild-type deep feature vector is obtained by mapping it to a common feature space of the same dimension using an independent learnable projection matrix. and mutant deep feature vectors ;

[0034] 2.2) Dynamic gating weight calculation: The projected wild-type depth feature vector is... and mutant deep feature vectors The features are concatenated along the feature dimension, and then a gating vector with importance weights is calculated using a gating network and a softmax function. and The process is described by the following formula:

[0035] ;

[0036] In the formula, Indicates feature concatenation operation; Here is the weight matrix of the gated network; A gated vector representing the importance weights corresponding to different feature vectors, including the projected wild-type deep feature vector. and the projected mutant depth feature vector Importance weights of the gating vector and ;

[0037] 2.3) Weighted Fusion: The projected feature vectors are weighted and summed using the calculated importance weights to obtain preliminary fused features. :

[0038] ;

[0039] In the formula, Represents the projected wild-type depth feature vector The importance weighting gate vector, Represents the projected mutant depth feature vector The gating vector for importance weights;

[0040] 2.4) Nonlinear enhancement and residual connection: Initially fusing features Enhanced features are obtained by nonlinear enhancement using a gated linear unit (GLU). Then enhance the features By summing all the projected feature vectors in step 2.1) and performing residual concatenation, the enhanced fused feature vector is finally obtained. :

[0041] ;

[0042] ;

[0043] In the formula, The weight matrix for the feature enhancement network; Represents the activation function of the gated linear unit; Presentation layer normalization operation.

[0044] Furthermore, the self-distillation contrastive learning module includes a self-distillation module and a multi-level contrastive learning module, wherein:

[0045] The self-distillation module is configured to: employ a knowledge distillation framework to process the enhanced fused feature vector output by the feature fusion module. As a source of knowledge for teachers, this translates into shallow feature vectors for students. , and deep feature vectors , The process of transmitting information includes the following steps:

[0046] 3.1) Soft Label Calculation: Project, temperature scaling, and Softmax normalization are applied to the features of the teacher model and the features of each student model to generate soft labels in the form of probability distributions, including teacher soft labels. and student soft tags ;

[0047] Teacher soft labels The calculation is as follows:

[0048] ;

[0049] For the shallow feature vector of the student , and deep feature vectors , Each feature in, using This indicates the corresponding student soft tag. The calculation is as follows:

[0050] ;

[0051] In the formula, and These are the projection functions used by the teacher model and the student model, respectively. Temperature hyperparameters for knowledge distillation are used to adjust the smoothness of soft tag distribution;

[0052] 3.2) Calculation of distillation loss: The consistency of knowledge transfer is measured by calculating the mean squared error loss between each student's soft tag and the teacher's soft tag using the mean squared error loss function. The final self-distillation loss is obtained by summing the mean squared error losses between each student's soft tag and the teacher's soft tag. :

[0053] ;

[0054] In the formula, This represents the mean squared error loss function;

[0055] The multi-level contrastive learning module is configured to enhance the semantic alignment and consistency between wild-type features and corresponding mutant features by performing contrastive learning on shallow feature vectors and deep feature vectors respectively, including the following steps:

[0056] 4.1) Feature projection: Set up an independent shallow projection head and depth projection head This maps feature vectors from different levels to their respective common contrast spaces;

[0057] ;

[0058] ;

[0059] In the formula, These are the wild-type and mutant characteristics projected onto the shallow public space, respectively. These are the wild-type and mutant characteristics projected onto the deep common space, respectively;

[0060] 4.2) L2 normalization: Perform L2 norm normalization on all projected feature vectors so that the cosine similarity can be efficiently calculated through the vector inner product later.

[0061] ;

[0062] ;

[0063] In the formula, These are the shallow and deep feature vectors of the normalized wild-type and mutant amino acid sequences, respectively.

[0064] 4.3) Contrast Loss Calculation: The symmetric InfoNCE loss function is used to calculate the shallow contrast loss separately. Contrast loss with depth The InfoNCE loss function aims to maximize the similarity of corresponding wild-type-mutant pairs (positive sample pairs) within a batch, while minimizing the similarity with all other non-corresponding samples (negative sample pairs).

[0065] Shallow contrast loss The calculation formula is as follows:

[0066] ;

[0067] Depth Contrast Loss The calculation formula is as follows:

[0068] ;

[0069] In the formula, Batch size; and These represent the first in the batch. Normalized shallow feature vectors of wild-type and mutant types for each sample and These represent the first in the batch. Normalized shallow feature vectors of wild-type and mutant types for each sample and These represent the first in the batch. Normalized wild-type deep feature vectors and mutant deep feature vectors for each sample and These represent the first in the batch. Normalized wild-type deep feature vectors and mutant deep feature vectors for each sample; This represents the dot product of vectors, i.e., cosine similarity. This is a temperature hyperparameter used to adjust the sharpness of the similarity distribution.

[0070] Furthermore, the supervised loss calculation module includes the following steps:

[0071] 5.1) The enhanced fused feature vector Perform global average pooling to obtain the input features of the classifier. ;

[0072] 5.2) The input is fed into the KAN network classifier, which is composed of multiple stacked KAN layers. Each layer uses a learnable B-spline basis function as the activation function, replacing the fixed activation function in the traditional multilayer perceptron (MLP).

[0073] 5.3) The KAN network classifier outputs the classification probability and calculates the supervised loss by comparing it with the real labels. .

[0074] Furthermore, the model optimization module includes the following steps:

[0075] 6.1) By jointly optimizing a total loss function To update the parameters of the deep neural network model, the total loss function Loss due to supervision Shallow contrast loss Depth contrast loss and self-distillation loss The weighted sum is obtained by:

[0076] ;

[0077] In the formula, The weighting hyperparameters for each loss;

[0078] 6.2) Use backpropagation to train the model and obtain the best-performing deep neural network model.

[0079] Furthermore, the mutation effect prediction module is structurally configured to include: a log-likelihood change calculation module, a mutation effect score calculation module, and a result output module, wherein:

[0080] The log-likelihood change calculation module includes the following steps:

[0081] 7.1) Call the pre-trained deep neural network model to process the wild-type amino acid sequence. and the mutant amino acid sequence described above A single forward propagation is performed; during this propagation, the feature extraction backbone inside the model works in concert with the deep self-attention network to automatically integrate sequence information with implicit structural and functional context by modeling the global dependencies between all amino acid residues in the sequence, thereby generating a deep contextual feature representation containing rich semantics.

[0082] 7.2) Based on this deep contextual feature representation, and utilizing the powerful decoding capabilities formed by the deep neural network model after joint optimization through self-distillation contrastive learning and classification tasks, at each specified mutation site... The above steps involve obtaining a complete conditional probability distribution of amino acids, from which wild-type amino acids are extracted. and mutant amino acids log-likelihood change value , :

[0083] ;

[0084] ;

[0085] In the formula, This represents the conditional probability defined by the pre-trained deep neural network model. Represents the location amino acid variables, This is the representation of the deep context features;

[0086] The mutation effect score calculation module includes the following steps:

[0087] 8.1) Receive the log-likelihood change values ​​for all mutation sites from the log-likelihood change calculation module. and ;

[0088] 8.2) Sum the log-likelihood changes at all mutation sites to obtain the final mutation effect score. And output it as the prediction result:

[0089] ;

[0090] In the formula, It is the set of all mutation sites;

[0091] The result output module includes the following steps:

[0092] 9.1) Receive the final score from the mutation effect score calculation module. ;

[0093] 9.2) This As a quantitative prediction of the mutation effect, and further based on preset thresholds or ranking results, mutations predicted to be beneficial or harmful are screened out, and finally structured data containing mutation information and prediction scores are provided to downstream applications.

[0094] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0095] 1. The feature extraction module employed in this invention optimizes the network structure, achieving more sensitive capture of mutation information and higher-density feature representation, thus solving the problem of existing models being insensitive to minor sequence perturbations. Existing technologies, such as standard convolutional neural networks (CNNs) or Transformer models, due to their fixed convolutional kernels or feedforward network structures, struggle to effectively capture minute changes in local sequence "edges" or "gradients" caused by single amino acid substitutions. This invention overcomes this deficiency through the following two improvements: First, the Wavelet-DenseBlock module (WDB module) in the feature processor of its deep neural network model replaces the fixed-kernel convolution operation with a learnable wavelet transform, enabling the model to accurately capture edge and gradient changes caused by local mutations in the amino acid sequence at multiple scales, thereby exhibiting higher sensitivity to minor sequence perturbations. Secondly, the improved Transformer module in the feature processor of its deep neural network model innovatively replaces the standard feedforward neural network layer module with a Kolmogorov-Arnold Network module that introduces learnable wavelet transform. At the same time, by introducing learnable wavelet transform, the Kolmogorov-Arnold Network module transforms the nonlinear feature mapping from a fixed mapping method to a data-driven, weighted and flexible method, thereby enabling more efficient collaborative capture of local details and long-range dependencies in the sequence, greatly improving the quality and information density of the final generated features.

[0096] 2. This invention enhances model training through the following two strategies: First, by performing comparative learning on independently extracted wild-type shallow feature vectors and mutant shallow feature vectors, as well as wild-type deep feature vectors and mutant deep feature vectors, the model learns and aligns the semantic information of the wild type with its corresponding mutant, thereby significantly improving the model's ability to recognize subtle differences caused by single point mutations. Second, through self-distillation learning, the more information-rich fused feature vectors generated through a dynamic gating mechanism are used as teachers to transfer knowledge to the shallow and deep feature vectors, which act as students. This process not only achieves efficient knowledge transfer across levels and modalities, but more importantly, it plays a powerful regularization role, effectively preventing model overfitting and enhancing the model's generalization ability and robustness when facing noise or unseen sequences.

[0097] 3. The fundamental advantage of this invention lies in its model optimization, which does not rely on a single loss function. Instead, it weights and sums the supervised loss, shallow contrast loss, deep contrast loss, and self-distillation loss to form a total loss function for joint optimization. In the prediction phase, the system requires no fine-tuning with labeled data regarding mutation effects. It directly infers the effect of the mutation simply by calculating the difference in log-likelihood between wild-type and mutant amino acids at the mutated amino acid sequence position. This essentially measures whether the mutation conforms to the natural laws learned by the model. Compared to traditional models that rely on extensive labeled data for fine-tuning, this first-principles-based reasoning capability allows it to directly generalize to any unseen protein and its mutations, providing unparalleled advantages in handling data-sparse protein engineering tasks.

[0098] In summary, this invention, through a series of tightly coupled technical solutions including the aforementioned feature extractor, self-distillation contrastive learning, and zero-sample mutation effect prediction, ultimately achieves deep learning of the intrinsic patterns of protein sequences. This system no longer relies on supervised learning of specific mutation effect tags, but instead predicts by evaluating the impact of mutations on the natural conditional probability of the sequence. This enables more accurate and efficient zero-sample mutation effect prediction, providing more fundamental and generalizable technical support for fields such as protein engineering, disease mechanism research, and drug design. Attached Figure Description

[0099] Figure 1 This is a diagram showing the overall functional modules of the system of this invention.

[0100] Figure 2 This is a schematic diagram showing the relationship between the various modules of the training system used in the present invention for generating deep neural network models.

[0101] Figure 3This is a schematic diagram of the feature extraction module of the training system used to generate deep neural network models in the system of the present invention.

[0102] Figure 4 This is a schematic diagram of the WDB module and the improved Transformer module of the feature processor of the deep neural network model in the system of the present invention; in the figure, Positional Encoding represents the positional encoding layer in the traditional Transformer model, Multi-Head Attention represents the multi-head self-attention mechanism in the traditional Transformer model, LayerNorm represents the layer normalization operation in the traditional Transformer model, and Add&Norm represents the residual connection and normalization operation in the traditional Transformer model.

[0103] Figure 5 This is a schematic diagram of the feature fusion unit of the deep neural network model in the system of this invention.

[0104] Figure 6 This is a schematic diagram illustrating the internal working mechanism of the self-distillation contrastive learning module in the training system for generating deep neural network models in this embodiment of the invention, demonstrating the process of self-distillation and multi-level contrastive learning. In the diagram, "Student" represents the feature vectors participating in the self-distillation learning process as wild-type shallow feature vectors, mutant shallow feature vectors, wild-type deep feature vectors, and mutant deep feature vectors; "Teacher" represents the feature vectors participating in the self-distillation learning process as enhanced fused feature vectors.

[0105] Figure 7 This is a schematic diagram of the joint loss function used by the model optimization module of the training system for generating deep neural network models in an embodiment of the present invention.

[0106] Figure 8 This is a schematic diagram of the workflow of the mutation effect prediction module in the prediction stage in an embodiment of the present invention. Detailed Implementation

[0107] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0108] This embodiment discloses a zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning. It is a software system preferably developed using Python and deep learning frameworks such as PyTorch, and can be deployed and run on servers running mainstream operating systems such as Linux or Windows. Figure 1 As shown, it includes the following functional modules:

[0109] The data import module is used to acquire and load wild-type amino acid sequences and specified mutation information, and to preprocess them to obtain mutant amino acid sequences.

[0110] A deep neural network model loading module is used to load a pre-trained deep neural network model. This model includes a feature processor and a feature fusion unit. The feature processor includes a Wavelet-DenseBlock module (WDB module) and an improved Transformer module. The WDB module replaces traditional convolution operations with learnable wavelet transforms, introduces first-order Gaussian derivative wavelets to enhance the model's sensitivity to local changes, and normalizes and compensates for features using learnable scale and translation parameters. This allows the model to capture edge information in the sequence at multiple scales, thereby improving the feature's expressive power. The improved Transformer module replaces the original feedforward neural network layer with a Kolmogorov-Arnold Network module. The Kolmogorov-Arnold Network module, by introducing learnable wavelet transforms, changes the nonlinear feature mapping from a fixed mapping method to a data-driven, weighted, and flexible method. This enables more efficient collaborative capture of local details and long-range dependencies in the sequence, improving the quality and information density of the final generated features.

[0111] The mutation effect prediction module is used to call the pre-trained deep neural network model loaded by the deep neural network model loading module. The wild-type amino acid sequence and the mutant amino acid sequence are used as inputs to the deep neural network model. Based on this deep neural network model, the log-likelihood change caused by a single amino acid substitution is detected, and a score quantifying the mutation effect is calculated and output based on this log-likelihood change, ultimately obtaining the mutation effect prediction result. Specifically, the data import module includes a data loading module and a data preprocessing module, wherein:

[0112] The data loading module reads wild-type amino acid sequence data and mutation information data without real label data from the public protein databases UniProt and PDB. The mutation information data includes the protein ID, the amino acid before the mutation, the amino acid sequence position where the mutation occurred (hereinafter referred to as the mutation site), and the amino acid after the mutation.

[0113] The data preprocessing module receives wild-type amino acid sequence data and specified mutation information data from the data loading module. Based on each mutation information, it performs an amino acid substitution operation at the corresponding mutation site in the wild-type amino acid sequence, replacing the unmutated amino acid with the mutated one, thereby generating the corresponding mutant amino acid sequence. Specifically, the deep neural network model is generated through a specialized training system, such as... Figure 2 As shown, the training system includes:

[0114] The training data import module is used to acquire and load wild-type amino acid sequences containing real label data, specified mutation information, and corresponding real label data. Based on each mutation information, an amino acid substitution operation is performed on the corresponding wild-type amino acid sequence to generate the corresponding mutant amino acid sequence.

[0115] The feature extraction module is used to integrate the wild-type amino acid sequence containing real label data and the mutant amino acid sequence. Using a fixed pre-trained protein language model (PLM) as the base model, the wild-type amino acid sequence and the corresponding mutant amino acid sequence are converted into feature vector form to obtain wild-type shallow feature vector and mutant shallow feature vector. The feature processor of the deep neural network model is used to process the wild-type shallow feature vector and mutant shallow feature vector to obtain wild-type deep feature vector and mutant deep feature vector.

[0116] The feature fusion module uses the feature fusioner of the deep neural network model to fuse the wild-type deep feature vector and the mutant deep feature vector output by the feature extraction module into an enhanced fused feature vector.

[0117] The self-distillation contrastive learning module is used to perform self-distillation learning on the wild-type shallow feature vector and mutant shallow feature vector, as well as the wild-type deep feature vector and mutant deep feature vector output by the feature extraction module, and the fused feature vector output by the feature fusion module to obtain self-distillation loss; and to perform contrastive learning on the wild-type shallow feature vector and mutant shallow feature vector, as well as the wild-type deep feature vector and mutant deep feature vector output by the feature extraction module to obtain shallow contrastive loss and deep contrastive loss.

[0118] The supervised loss calculation module includes a Kolmogorov-Arnold Network (KAN) classifier, which is connected after the feature fusion module. It is used to predict the mutation effect of protein point mutations. The supervised loss is calculated using the cross-entropy loss function through real label data and prediction results.

[0119] The model optimization module is used to perform end-to-end training on the deep neural network model using the self-distillation loss, shallow contrastive loss, and deep contrastive loss obtained from the self-distillation contrastive learning module, as well as the supervised loss, until the model converges, ultimately generating a pre-trained deep neural network model, i.e., the optimal-performing deep neural network model. Specifically, as shown... Figure 3 As shown, the feature extraction module includes a shallow feature extraction module and a deep feature extraction module, wherein:

[0120] The shallow feature extraction module uses a fixed pre-trained protein language model as the base model to extract wild-type amino acid sequences. and the corresponding mutant amino acid sequence The formula for converting it into feature vector form is as follows:

[0121] ;

[0122] In the formula, Represents the amino acid sequence. ,in Represents the wild type. Represents a mutant type; This refers to the pre-trained protein language model; This represents the output feature vector, with dimension 1. ,in It refers to the batch size. It is the sequence length. It is the embedding dimension; the shallow feature extraction module ultimately outputs a wild-type shallow feature vector. and mutant shallow feature vectors ;

[0123] The deep feature extraction module utilizes the feature processor of the deep neural network model, and processes the wild-type shallow feature vector and the mutant shallow feature vector through the WDB module and the improved Transformer module to obtain the wild-type deep feature vector and the mutant deep feature vector. The processing steps include:

[0124] 1.1) The wild-type shallow feature vector output by the shallow feature extraction module and mutant shallow feature vectors The features are respectively used as initial inputs to the WDB module, and the processed features are output as feature vectors. and ;

[0125] 1.2) The feature vector and The inputs are then fed into the improved Transformer module for sequence modeling to obtain wild-type deep feature vectors. and mutant deep feature vectors ;

[0126] The WDB module consists of n Dense Blocks and one Transition Layer. The input of the m-th (m≤n) Dense Block is the initial feature and the output of the first m-1 layers concatenated along the channel dimension. After wavelet transform, batch normalization, activation function, and random deactivation, the output of the m-th layer is finally obtained. The initial feature and the output of the first n Dense Blocks concatenated along the channel dimension are used as the input of the Transition Layer. The Transition Layer consists of convolution operation, batch normalization, nonlinear activation function, and random deactivation.

[0127] The improved Transformer module includes a multi-layer multi-head self-attention mechanism and a Kolmogorov-Arnold Network module incorporating learnable wavelet transform. Each layer uses a standard multi-head self-attention mechanism to model long-range dependencies of the input features, and the Kolmogorov-Arnold Network module with learnable wavelet transform performs nonlinear feature mapping. The final output is passed to the next layer or used for downstream tasks. Specifically, as... Figure 5 As shown, the feature fusion module utilizes the feature fusioner of the deep neural network model and employs a dynamic gating fusion mechanism to combine the wild-type deep feature vector output by the deep feature extraction module. and mutant deep feature vectors The fusion is performed to obtain the enhanced fused feature vector. It includes the following steps:

[0128] 2.1) Feature Projection: Projecting the wild-type deep feature vector and mutant deep feature vectors Each wild-type deep feature vector is obtained by mapping it to a common feature space of the same dimension using an independent learnable projection matrix. and mutant deep feature vectors ;

[0129] 2.2) Dynamic gating weight calculation: The projected wild-type depth feature vector is... and mutant deep feature vectors The features are concatenated along the feature dimension, and then a gating vector with importance weights is calculated using a gating network and a softmax function. and The process is described by the following formula:

[0130] ;

[0131] In the formula, Indicates feature concatenation operation; Here is the weight matrix of the gated network; A gated vector representing the importance weights corresponding to different feature vectors, including the projected wild-type deep feature vector. and the projected mutant depth feature vector Importance weights of the gating vector and ;

[0132] 2.3) Weighted Fusion: The projected feature vectors are weighted and summed using the calculated importance weights to obtain preliminary fused features. :

[0133] ;

[0134] In the formula, Represents the projected wild-type depth feature vector The importance weighting gate vector, Represents the projected mutant depth feature vector The gating vector for importance weights;

[0135] 2.4) Nonlinear enhancement and residual connection: Initially fusing features Enhanced features are obtained by nonlinear enhancement using a gated linear unit (GLU). Then enhance the features By summing all the projected feature vectors in step 2.1) and performing residual concatenation, the enhanced fused feature vector is finally obtained. :

[0136] ;

[0137] ;

[0138] In the formula, The weight matrix for the feature enhancement network; Represents the activation function of the gated linear unit; Presentation layer normalization operation.

[0139] Specifically, such as Figure 6 As shown, the self-distillation contrastive learning module includes a self-distillation module and a multi-level contrastive learning module, wherein:

[0140] The self-distillation module is configured to: employ a knowledge distillation framework to process the enhanced fused feature vector output by the feature fusion module. As a source of knowledge for teachers, this translates into shallow feature vectors for students. , and deep feature vectors , The process of transmitting information includes the following steps:

[0141] 3.1) Soft Label Calculation: Project, temperature scaling, and Softmax normalization are applied to the features of the teacher model and the features of each student model to generate soft labels in the form of probability distributions, including teacher soft labels. and student soft tags ;

[0142] Teacher soft labels The calculation is as follows:

[0143] ;

[0144] For the shallow feature vector of the student , and deep feature vectors , Each feature in, using This indicates the corresponding student soft tag. The calculation is as follows:

[0145] ;

[0146] In the formula, and These are the projection functions used by the teacher model and the student model, respectively. Temperature hyperparameters for knowledge distillation are used to adjust the smoothness of soft tag distribution;

[0147] 3.2) Calculation of distillation loss: The consistency of knowledge transfer is measured by calculating the mean squared error loss between each student's soft tag and the teacher's soft tag using the mean squared error loss function. The final self-distillation loss is obtained by summing the mean squared error losses between each student's soft tag and the teacher's soft tag. :

[0148] ;

[0149] In the formula, This represents the mean squared error loss function;

[0150] The multi-level contrastive learning module is configured to enhance the semantic alignment and consistency between wild-type features and corresponding mutant features by performing contrastive learning on shallow feature vectors and deep feature vectors respectively, including the following steps:

[0151] 4.1) Feature projection: Set up an independent shallow projection head and depth projection head This maps feature vectors from different levels to their respective common contrast spaces;

[0152] ;

[0153] ;

[0154] In the formula, These are the wild-type and mutant characteristics projected onto the shallow public space, respectively. These are the wild-type and mutant characteristics projected onto the deep common space, respectively;

[0155] 4.2) L2 normalization: Perform L2 norm normalization on all projected feature vectors so that the cosine similarity can be efficiently calculated through the vector inner product later.

[0156] ;

[0157] ;

[0158] In the formula, These are the shallow and deep feature vectors of the normalized wild-type and mutant amino acid sequences, respectively.

[0159] 4.3) Contrast Loss Calculation: The symmetric InfoNCE loss function is used to calculate the shallow contrast loss separately. Contrast loss with depth The InfoNCE loss function aims to maximize the similarity of corresponding wild-type-mutant pairs (positive sample pairs) within a batch, while minimizing the similarity with all other non-corresponding samples (negative sample pairs).

[0160] Shallow contrast loss The calculation formula is as follows:

[0161] ;

[0162] Depth Contrast Loss The calculation formula is as follows:

[0163] ;

[0164] In the formula, Batch size; and These represent the first in the batch. Normalized shallow feature vectors of wild-type and mutant types for each sample and These represent the first in the batch. Normalized shallow feature vectors of wild-type and mutant types for each sample and These represent the first in the batch. Normalized wild-type deep feature vectors and mutant deep feature vectors for each sample and These represent the first in the batch. Normalized wild-type deep feature vectors and mutant deep feature vectors for each sample; This represents the dot product of vectors, i.e., cosine similarity. This is a temperature hyperparameter used to adjust the sharpness of the similarity distribution.

[0165] Specifically, the supervised loss calculation module includes the following steps:

[0166] 5.1) The enhanced fused feature vector Perform global average pooling to obtain the input features of the classifier. ;

[0167] 5.2) The input is fed into the KAN network classifier, which is composed of multiple stacked KAN layers. Each layer uses a learnable B-spline basis function as the activation function, replacing the fixed activation function in the traditional multilayer perceptron (MLP).

[0168] 5.3) The KAN network classifier outputs the classification probability and calculates the supervised loss by comparing it with the real labels. .

[0169] Specifically, the model optimization module includes the following steps:

[0170] 6.1) By jointly optimizing a total loss function To update the parameters of the deep neural network model, such as Figure 7 As shown, the total loss function Loss due to supervision Shallow contrast loss Depth contrast loss and self-distillation loss The weighted sum is obtained by:

[0171]

[0172] In the formula, The weighting hyperparameters for each loss;

[0173] 6.2) Use backpropagation to train the model and obtain the best-performing deep neural network model.

[0174] Specifically, such as Figure 8 As shown, the mutation effect prediction module is structurally configured to include: a log-likelihood change calculation module, a mutation effect score calculation module, and a result output module, wherein:

[0175] The log-likelihood change calculation module includes the following steps:

[0176] 7.1) Call the pre-trained deep neural network model to process the wild-type amino acid sequence. and the mutant amino acid sequence described above A single forward propagation is performed; during this propagation, the feature extraction backbone inside the model works in concert with the deep self-attention network to automatically integrate sequence information with implicit structural and functional context by modeling the global dependencies between all amino acid residues in the sequence, thereby generating a deep contextual feature representation containing rich semantics.

[0177] 7.2) Based on this deep contextual feature representation, and utilizing the powerful decoding capabilities formed by the deep neural network model after joint optimization through self-distillation contrastive learning and classification tasks, at each specified mutation site... The above steps involve obtaining a complete conditional probability distribution of amino acids, from which wild-type amino acids are extracted. and mutant amino acids log-likelihood change value , :

[0178] ;

[0179] ;

[0180] In the formula, This represents the conditional probability defined by the pre-trained deep neural network model. Represents the location amino acid variables, This is the representation of the deep context features;

[0181] The mutation effect score calculation module includes the following steps:

[0182] 8.1) Receive the log-likelihood change values ​​for all mutation sites from the log-likelihood change calculation module. and ;

[0183] 8.2) Sum the log-likelihood changes at all mutation sites to obtain the final mutation effect score. And output it as the prediction result:

[0184] ;

[0185] In the formula, It is the set of all mutation sites;

[0186] The result output module includes the following steps:

[0187] 9.1) Receive the final score from the mutation effect score calculation module. ;

[0188] 9.2) This As a quantitative prediction of the mutation effect, and further based on preset thresholds or ranking results, mutations predicted to be beneficial or harmful are screened out, and finally structured data containing mutation information and prediction scores are provided to downstream applications.

[0189] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning, characterized in that, include: The data import module is used to acquire and load wild-type amino acid sequences and specified mutation information, and to preprocess them to obtain mutant amino acid sequences. A deep neural network model loading module is used to load a pre-trained deep neural network model. This model includes a feature processor and a feature fusion unit. The feature processor includes a Wavelet-DenseBlock module (WDB module) and an improved Transformer module. The WDB module replaces traditional convolution operations with learnable wavelet transforms, introduces first-order Gaussian derivative wavelets to enhance the model's sensitivity to local changes, and normalizes and compensates for features using learnable scale and translation parameters. This allows the model to capture edge information in the sequence at multiple scales, thereby improving the feature's expressive power. The improved Transformer module replaces the original feedforward neural network layer with a Kolmogorov-Arnold Network module. The Kolmogorov-Arnold Network module, by introducing learnable wavelet transforms, changes the nonlinear feature mapping from a fixed mapping method to a data-driven, weighted, and flexible method. This enables more efficient collaborative capture of local details and long-range dependencies in the sequence, improving the quality and information density of the final generated features. The mutation effect prediction module is used to call the pre-trained deep neural network model loaded by the deep neural network model loading module, take the wild-type amino acid sequence and the mutant amino acid sequence as input to the deep neural network model, detect the log-likelihood change caused by a single amino acid substitution based on the deep neural network model, calculate and output a score quantifying the mutation effect based on the log-likelihood change, and finally obtain the mutation effect prediction result.

2. The zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning according to claim 1, characterized in that, The data import module includes a data loading module and a data preprocessing module, wherein: The data loading module reads wild-type amino acid sequence data and mutation information data without real label data from the public protein databases UniProt and PDB. The mutation information data includes the protein ID, the amino acid before the mutation, the amino acid sequence position where the mutation occurred (hereinafter referred to as the mutation site), and the amino acid after the mutation. The data preprocessing module receives wild-type amino acid sequence data and specified mutation information data from the data loading module. Based on each mutation information, it performs an amino acid substitution operation at the mutation site on the corresponding wild-type amino acid sequence, replacing the unmutated amino acid with the mutated amino acid, thereby generating the corresponding mutant amino acid sequence.

3. The zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning according to claim 2, characterized in that, The deep neural network model is generated through a specialized training system, which includes: The training data import module is used to acquire and load wild-type amino acid sequences containing real label data, specified mutation information, and corresponding real label data. Based on each mutation information, an amino acid substitution operation is performed on the corresponding wild-type amino acid sequence to generate the corresponding mutant amino acid sequence. The feature extraction module is used to integrate the wild-type amino acid sequence containing real label data and the mutant amino acid sequence. Using a fixed pre-trained protein language model (PLM) as the base model, the wild-type amino acid sequence and the corresponding mutant amino acid sequence are converted into feature vector form to obtain wild-type shallow feature vector and mutant shallow feature vector. The feature processor of the deep neural network model is used to process the wild-type shallow feature vector and mutant shallow feature vector to obtain wild-type deep feature vector and mutant deep feature vector. The feature fusion module uses the feature fusioner of the deep neural network model to fuse the wild-type deep feature vector and the mutant deep feature vector output by the feature extraction module into an enhanced fused feature vector. The self-distillation contrastive learning module is used to perform self-distillation learning on the wild-type shallow feature vector and mutant shallow feature vector, as well as the wild-type deep feature vector and mutant deep feature vector output by the feature extraction module, and the fused feature vector output by the feature fusion module to obtain self-distillation loss; and to perform contrastive learning on the wild-type shallow feature vector and mutant shallow feature vector, as well as the wild-type deep feature vector and mutant deep feature vector output by the feature extraction module to obtain shallow contrastive loss and deep contrastive loss. The supervised loss calculation module includes a Kolmogorov-Arnold Network (KAN) classifier, which is connected after the feature fusion module. It is used to predict the mutation effect of protein point mutations. The supervised loss is calculated using the cross-entropy loss function through real label data and prediction results. The model optimization module is used to perform end-to-end training on the deep neural network model using the self-distillation loss, shallow contrast loss, and deep contrast loss obtained by the self-distillation contrast learning module, as well as the supervised loss, until the model converges, and finally generates a pre-trained deep neural network model, that is, the deep neural network model with optimal performance.

4. The zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning according to claim 3, characterized in that, The feature extraction module includes a shallow feature extraction module and a deep feature extraction module, wherein: The shallow feature extraction module uses a fixed pre-trained protein language model as the base model to extract wild-type amino acid sequences. and the corresponding mutant amino acid sequence The formula for converting it into feature vector form is as follows: ； In the formula, Represents the amino acid sequence. ,in Represents the wild type. Represents a mutant type; This refers to the pre-trained protein language model; This represents the output feature vector, with dimension 1. ,in It refers to the batch size. It is the sequence length. It is the embedding dimension; the shallow feature extraction module ultimately outputs a wild-type shallow feature vector. and mutant shallow feature vectors ; The deep feature extraction module utilizes the feature processor of the deep neural network model, and processes the wild-type shallow feature vector and the mutant shallow feature vector through the WDB module and the improved Transformer module to obtain the wild-type deep feature vector and the mutant deep feature vector. The processing steps include: 1.1) The wild-type shallow feature vector output by the shallow feature extraction module and mutant shallow feature vectors The features are respectively used as initial inputs to the WDB module, and the processed features are output as feature vectors. and ; 1.2) The feature vector and The inputs are then fed into the improved Transformer module for sequence modeling to obtain wild-type deep feature vectors. and mutant deep feature vectors ; The WDB module consists of n Dense Blocks and one Transition Layer. The input of the m-th (m≤n)th Dense Block is the initial features and the outputs of the first m-1 layers concatenated along the channel dimension, then subjected to wavelet transform, batch normalization, activation function, and random deactivation to finally obtain the output of the m-th layer. The initial features and the outputs of the first n Dense Blocks concatenated along the channel dimension serve as the input of the Transition Layer, which consists of convolution operations, batch normalization, nonlinear activation function, and random deactivation. The improved Transformer module includes a multi-layer multi-head self-attention mechanism and a Kolmogorov-Arnold Network module that introduces learnable wavelet transform. Each layer uses the standard multi-head self-attention mechanism to model the long-range dependency of the input features, and performs nonlinear feature mapping through the Kolmogorov-Arnold Network module that introduces learnable wavelet transform. The final output is passed to the next layer or used for downstream tasks.

5. The zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning according to claim 4, characterized in that, The feature fusion module utilizes the feature fusion unit of the deep neural network model and employs a dynamic gating fusion mechanism to combine the wild-type deep feature vectors output by the deep feature extraction module. and mutant deep feature vectors The fusion is performed to obtain the enhanced fused feature vector. It includes the following steps: 2.1) Feature Projection: Projecting the wild-type deep feature vector and mutant deep feature vectors Each wild-type deep feature vector is obtained by mapping it to a common feature space of the same dimension using an independent learnable projection matrix. and mutant deep feature vectors ; 2.2) Dynamic gating weight calculation: The projected wild-type depth feature vector is... and mutant deep feature vectors The features are concatenated along the feature dimension, and then a gating vector with importance weights is calculated using a gating network and a softmax function. and The process is described by the following formula: ； In the formula, Indicates feature concatenation operation; Here is the weight matrix of the gated network; A gated vector representing the importance weights corresponding to different feature vectors, including the projected wild-type deep feature vector. and the projected mutant depth feature vector Importance weights of the gating vector and ; 2.3) Weighted Fusion: The projected feature vectors are weighted and summed using the calculated importance weights to obtain preliminary fused features. : ； In the formula, Represents the projected wild-type depth feature vector The importance weighting gate vector, Represents the projected mutant depth feature vector The gating vector for importance weights; 2.4) Nonlinear enhancement and residual connection: Initially fusing features Enhanced features are obtained by nonlinear enhancement using a gated linear unit (GLU). Then enhance the features By summing all the projected feature vectors in step 2.1) and performing residual concatenation, the enhanced fused feature vector is finally obtained. : ； ； In the formula, The weight matrix for the feature enhancement network; Represents the activation function of the gated linear unit; Presentation layer normalization operation.

6. The zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning according to claim 5, characterized in that, The self-distillation contrastive learning module includes a self-distillation module and a multi-level contrastive learning module, wherein: The self-distillation module is configured to: employ a knowledge distillation framework to process the enhanced fused feature vector output by the feature fusion module. As a source of knowledge for teachers, this translates into shallow feature vectors for students. , and deep feature vectors , The process of transmitting information includes the following steps: 3.1) Soft Label Calculation: Project, temperature scaling, and Softmax normalization are applied to the features of the teacher model and the features of each student model to generate soft labels in the form of probability distributions, including teacher soft labels. and student soft tags ; Teacher soft labels The calculation is as follows: ； For the shallow feature vector of the student , and deep feature vectors , Each feature in, using This indicates the corresponding student soft tag. The calculation is as follows: ； In the formula, and These are the projection functions used by the teacher model and the student model, respectively. Temperature hyperparameters for knowledge distillation are used to adjust the smoothness of soft tag distribution; 3.2) Calculation of distillation loss: The consistency of knowledge transfer is measured by calculating the mean squared error loss between each student's soft tag and the teacher's soft tag using the mean squared error loss function. The final self-distillation loss is obtained by summing the mean squared error losses between each student's soft tag and the teacher's soft tag. : ； In the formula, This represents the mean squared error loss function; The multi-level contrastive learning module is configured to enhance the semantic alignment and consistency between wild-type features and corresponding mutant features by performing contrastive learning on shallow feature vectors and deep feature vectors respectively, including the following steps: 4.1) Feature projection: Set up an independent shallow projection head and depth projection head This maps feature vectors from different levels to their respective common contrast spaces; ； ； In the formula, These are the wild-type and mutant characteristics projected onto the shallow public space, respectively. These are the wild-type and mutant characteristics projected onto the deep common space, respectively; 4.2) L2 normalization: Perform L2 norm normalization on all projected feature vectors so that the cosine similarity can be efficiently calculated through the vector inner product later. ； ； In the formula, These are the shallow and deep feature vectors of the normalized wild-type and mutant amino acid sequences, respectively. 4.3) Contrast Loss Calculation: The symmetric InfoNCE loss function is used to calculate the shallow contrast loss separately. Contrast loss with depth The InfoNCE loss function aims to maximize the similarity of corresponding wild-type-mutant pairs (positive sample pairs) within a batch, while minimizing the similarity with all other non-corresponding samples (negative sample pairs). Shallow contrast loss The calculation formula is as follows: ； Depth Contrast Loss The calculation formula is as follows: ； In the formula, Batch size; and These represent the first in the batch. Normalized shallow feature vectors of wild-type and mutant types for each sample and These represent the first in the batch. Normalized shallow feature vectors of wild-type and mutant types for each sample and These represent the first in the batch. Normalized wild-type deep feature vectors and mutant deep feature vectors for each sample and These represent the first in the batch. Normalized wild-type deep feature vectors and mutant deep feature vectors for each sample; This represents the dot product of vectors, i.e., cosine similarity. This is a temperature hyperparameter used to adjust the sharpness of the similarity distribution.

7. The zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning according to claim 6, characterized in that, The supervision loss calculation module includes the following steps: 5.1) The enhanced fused feature vector Perform global average pooling to obtain the input features of the classifier. ; 5.2) The input is fed into the KAN network classifier, which is composed of multiple stacked KAN layers. Each layer uses a learnable B-spline basis function as the activation function, replacing the fixed activation function in the traditional multilayer perceptron (MLP). 5.3) The KAN network classifier outputs the classification probability and calculates the supervised loss by comparing it with the real labels. .

8. The zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning according to claim 7, characterized in that, The model optimization module includes the following steps: 6.1) By jointly optimizing a total loss function To update the parameters of the deep neural network model, the total loss function Loss due to supervision Shallow contrast loss Depth contrast loss and self-distillation loss The weighted sum is obtained by: ； In the formula, The weighting hyperparameters for each loss; 6.2) Use backpropagation to train the model and obtain the best-performing deep neural network model.

9. The zero-shot prediction system for protein point mutation effects based on self-distillation contrastive learning according to claim 8, characterized in that, The mutation effect prediction module is structurally configured to include: a log-likelihood change calculation module, a mutation effect score calculation module, and a result output module, wherein: The log-likelihood change calculation module includes the following steps: 7.1) Call the pre-trained deep neural network model to process the wild-type amino acid sequence. and the mutant amino acid sequence described above A single forward propagation is performed; during this propagation, the feature extraction backbone inside the model works in concert with the deep self-attention network to automatically integrate sequence information with implicit structural and functional context by modeling the global dependencies between all amino acid residues in the sequence, thereby generating a deep contextual feature representation containing rich semantics. 7.2) Based on this deep contextual feature representation, and utilizing the powerful decoding capabilities formed by the deep neural network model after joint optimization through self-distillation contrastive learning and classification tasks, at each specified mutation site... The above steps involve obtaining a complete conditional probability distribution of amino acids, from which wild-type amino acids are extracted. and mutant amino acids log-likelihood change value , : ； ； In the formula, This represents the conditional probability defined by the pre-trained deep neural network model. Represents the location amino acid variables, This is the representation of the deep context features; The mutation effect score calculation module includes the following steps: 8.1) Receive the log-likelihood change values ​​for all mutation sites from the log-likelihood change calculation module. and ; 8.2) Sum the log-likelihood changes at all mutation sites to obtain the final mutation effect score. And output it as the prediction result: ； In the formula, It is the set of all mutation sites; The result output module includes the following steps: 9.1) Receive the final score from the mutation effect score calculation module. ; 9.2) This As a quantitative prediction of the mutation effect, and further based on preset thresholds or ranking results, mutations predicted to be beneficial or harmful are screened out, and finally structured data containing mutation information and prediction scores are provided to downstream applications.

Images