A method for evaluating exosome-targeted molecules

Abstract

The present application discloses a method for evaluating exosome-targeted molecules, which relates to the fields of data processing and artificial intelligence technology. The method comprises S1. constructing a training set and a validation set based on multi-source biological databases and experimental measurement data; S2. constructing an exosome molecular structure evaluation model; S3. training and saving an adaptive boundary loss and classification accuracy evaluation model based on the validation set to obtain a trained exosome molecular structure evaluation model; S4. obtaining an exosome-targeted molecule evaluation result based on the trained exosome molecular structure evaluation model. The present application further enhances the information interaction between different features through an attention mechanism, effectively improves the classification accuracy of high-efficiency samples by retaining important features related to the target, dynamically adjusts the decision boundary of the classifier, improves the recognition ability of high-efficiency molecules, and avoids the bias of traditional methods towards the majority class.

Description

Technical Field

[0001] The present application relates to the fields of data processing and artificial intelligence technology, and in particular to a method for evaluating exosome-targeted molecules. Background Art

[0002] As an important carrier of intercellular information transmission, exosomes have unique biological functions and broad application prospects, especially in disease diagnosis, drug delivery and cancer treatment.

[0003] The following problems exist in the prior art and still need to be further solved:

[0004] 1. Traditional methods usually use simple splicing or independent branches to process heterogeneous features, which cannot effectively capture the deep topological relationship between structure and function. Simple splicing or independent processing will ignore the interaction between different data modalities, resulting in poor classification results.

[0005] 2. Existing technologies usually use methods such as L1 regularization for feature selection. These methods perform poorly in dealing with target information sparsity and nonlinear relationships. They cannot dynamically adjust feature importance based on the feature differences of different target-molecule combinations, which can easily lead to the suppression of key features or the retention of noise features.

[0006] 3. When dealing with high-dimensional features, existing dimensionality reduction methods easily destroy the discriminative association between features and categories and fail to effectively remove redundant information, making it difficult for the model to extract meaningful features and affecting the classification effect. This is especially challenging when identifying rare categories.

[0007] 4. When faced with the problem of class imbalance, traditional cross entropy loss function and focal loss function often find it difficult to assign sufficient decision boundaries to sparse categories, resulting in poor recognition ability of the model for sparse categories, manifested as low recall rate and poor classification performance.

[0008] The molecular structure and amino acid sequence on the surface of exosomes are closely related to their targeting efficacy. Therefore, accurately evaluating the efficacy of exosome targeting molecules is of great significance for research and application in related fields. Summary of the Invention

[0009] The purpose of this application is to provide an exosome targeting molecule evaluation method, which improves the accuracy and robustness of exosome targeting molecule evaluation by combining the collaborative embedding mechanism of graph convolution and sequence convolution, the target condition gating mechanism, the principle of maximizing mutual information, and the adaptive class boundary loss mechanism.

[0010] To achieve the above objectives, the present application provides the following scheme: a method for evaluating exosome-targeted molecules, the method comprising:

[0011] S1. Construct training and validation sets based on multi-source biological databases and experimental measurement data;

[0012] S2. Construct an exosome molecular structure evaluation model;

[0013] S3. Train and save the adaptive boundary loss and classification accuracy evaluation model based on the validation set to obtain the trained exosome molecular structure evaluation model;

[0014] S4. Obtain exosome targeting molecule evaluation results based on the trained exocrine molecular structure evaluation model;

[0015] Among them, the exosome molecular structure evaluation model constructed in S2 includes:

[0016] S201. Based on the collaborative embedding mechanism obtained through graph convolution and sequence convolution, combined with the attention mechanism fusion feature, the molecular structure data and sequence data are uniformly represented to generate a fusion feature matrix ;

[0017] S202. The fusion feature matrix is ​​sequentially fused through the target condition gating mechanism, the mutual information maximization principle and the dual-channel gated residual network. Processing is performed to generate a fused feature vector;

[0018] S203. Generate an optimized targeting efficacy probability classifier based on the fused feature vector.

[0019] Optionally, the fusion feature matrix is ​​output in S201 , also includes:

[0020] Adjacency matrix based on chemical bonds, degree matrix of adjacency matrix of chemical bonds, and molecular graph structure input matrix Output graph convolution feature matrix ;

[0021] Combining one-dimensional convolution kernel and sequence input matrix Output sequence feature matrix ;

[0022] Based on graph convolution feature matrix and the sequence feature matrix Computational Structure - Sequence Attention Mask Matrix And output the fusion feature matrix through channel splicing operation .

[0023] Optional, graph convolution feature matrix , expressed as:

[0024] ;

[0025] Where, Represents the graph convolution feature matrix; is a set of chemical bond types; is a positive integer; Represents the ELU activation function, which enhances the model's expressiveness through nonlinearity; For the degree matrix of the chemical bond-like adjacency matrix; for The inverse matrix of For the Adjacency matrix of chemical-like bonds; Input matrix for molecular graph structure; is the graph convolution weight matrix.

[0026] Optionally, fused feature matrix , expressed as:

[0027] ;

[0028] Where, represents the fusion feature matrix; is the structure-sequence attention mask matrix; is the global average pooling operation; represents element-wise multiplication; represents the channel concatenation operation, connecting different features along the column dimension; is the fusion coefficient; is the feature alignment matrix.

[0029] Optionally, generating the fused feature vector in S202 further includes:

[0030] S2021. Dynamically modulate the fusion feature matrix using target-conditional gating mechanism The response intensity is used to obtain discriminative features that are strongly correlated with the target;

[0031] S2022. Combine the mutual information maximization principle and the multi-layer perceptron to sequentially aggregate and reorganize the discriminative features to generate a recombined feature vector ;

[0032] S2023. Reconstructing feature vectors via dual-channel gated residual networks Perform dynamic feature fusion to generate a fused feature vector.

[0033] Optionally, obtaining discriminative features strongly correlated with the target in S2021 also includes:

[0034] Combined fusion feature matrix and the preset target feature vector Generate dynamic gating vectors ;

[0035] Use dynamic gating vector g to fusion feature matrix Perform weighted masking and combine it with the compensation weight matrix Generate masked feature matrix :

[0036] ;

[0037] Where, is the compensation weight matrix; is a unit vector; is a unit vector Subtract the dynamic gate vector , characterize the generated inversion mask;

[0038] Combined masked feature matrix And the fusion feature matrix Generate discriminative features.

[0039] Optionally, generating an optimized targeting effectiveness probability classifier based on the fused feature vector in S203 further includes:

[0040] S2031. Constructing a targeting efficacy probability classifier based on the fused feature vector;

[0041] S2032. Adopt the adaptive class boundary margin mechanism to dynamically adjust the classification interval of the target effectiveness probability classifier to generate an optimized target effectiveness probability classifier.

[0042] Optionally, the trained exosome molecular structure evaluation model is obtained in S3, which also includes:

[0043] S301. Dynamically iteratively update all trainable parameters of the exosome molecular structure evaluation model based on mutual information loss and adaptive boundary loss;

[0044] S302. Obtain the number of iterations based on the adaptive boundary loss and classification accuracy of the validation set to generate a trained exosome molecular structure evaluation model.

[0045] Optionally, parameter updates are performed on the exosome molecular structure assessment model, including:

[0046] S3011. Dynamically calculate mutual information loss based on mutual information loss and adaptive boundary loss Dynamic weight and adaptive boundary loss Dynamic weight ;

[0047] S3012. Combined mutual information loss Dynamic weight and adaptive boundary loss Dynamic weight Generate fused total gradient ;

[0048] S3013. Adam optimizer with gradient clipping and fused total gradient Iteratively update the parameters of the exosome molecular structure evaluation model.

[0049] Optionally, calculate mutual information loss Dynamic weight and adaptive boundary loss Dynamic weight , also includes:

[0050] ;

[0051] ;

[0052] Where, Mutual information loss Dynamic weight of is the adaptive boundary loss Dynamic weight of is the Frobenius norm; represents the mutual information loss Relative to model parameters gradient; represents the adaptive boundary loss Relative to model parameters gradient; is a natural constant; Control mutual information loss for the decay coefficient Contribution decays in the later stages of training; is the current training cycle; is the total number of training cycles.

[0053] According to the specific embodiments provided in this application, this application has the following technical effects:

[0054] 1. The present invention solves the heterogeneity problem of molecular structure data and amino acid sequence data by combining the collaborative embedding mechanism of graph convolution and sequence convolution. It captures chemical structure features through graph convolution and extracts local patterns of amino acid sequences through sequence convolution. It also further enhances the information interaction between different features through the attention mechanism, thereby improving the classification accuracy of the model for targeting efficacy.

[0055] 2. The present invention adopts a target condition gating mechanism to dynamically adjust the response intensity of different features and automatically adapt to the feature differences of different target-molecule combinations. This avoids the problem that traditional feature selection methods such as L1 regularization cannot dynamically adapt to different target combinations, ensuring that important features related to the target are retained, thereby improving classification accuracy.

[0056] 3. The present invention solves the problem of redundant information in high-dimensional feature space by maximizing the principle of mutual information, enhances the aggregation of similar features and the separability of features of different categories, and effectively improves the classification accuracy of high-performance samples by constructing category prototype centers and optimizing based on intra-class mutual information loss.

[0057] 4. This invention adopts an adaptive class boundary loss mechanism to address the common class imbalance problem in exosome targeted assessment. It dynamically adjusts the decision boundary of the classifier, giving rare classes a larger decision space, thereby improving the ability to identify high-efficiency molecules and avoiding the bias of traditional methods towards the majority class. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0059] Figure 1 A flow chart of an exosome-targeted molecule evaluation method provided in this application;

[0060] Figure 2 A performance comparison chart of different models provided in the embodiments of this application at three energy efficiency levels;

[0061] Figure 3 This is a performance comparison chart under different training samples provided in the embodiment of this application;

[0062] Figure 4 A histogram showing the effectiveness of the heterogeneous feature fusion mechanism provided in the embodiments of this application;

[0063] Figure 5 An analysis diagram showing the impact of adaptive boundary loss on various performance aspects provided in an embodiment of the present application;

[0064] Figure 6 Recombination analysis graph for discriminative features driven by mutual information. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0066] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0067] like Figure 1 As shown, a method for evaluating exosome-targeted molecules comprises:

[0068] S1. Construct training and validation sets based on multi-source biological databases and experimental measurement data;

[0069] S2. Construct an exosome molecular structure evaluation model;

[0070] S3. Train and save the adaptive boundary loss and classification accuracy evaluation model based on the validation set to obtain the trained exosome molecular structure evaluation model;

[0071] S4. Obtain exosome targeting molecule evaluation results based on the trained exocrine molecular structure evaluation model;

[0072] Among them, the exosome molecular structure evaluation model constructed in S2 includes:

[0073] S201. Based on the collaborative embedding mechanism obtained through graph convolution and sequence convolution, combined with the attention mechanism fusion feature, the molecular structure data and sequence data are uniformly represented to generate a fusion feature matrix ;

[0074] S202. The fusion feature matrix is ​​sequentially fused through the target condition gating mechanism, the mutual information maximization principle and the dual-channel gated residual network. Processing is performed to generate a fused feature vector;

[0075] S203. Generate an optimized targeting efficacy probability classifier based on the fused feature vector.

[0076] Furthermore, dataset construction includes: constructing a training set and a validation set by integrating multi-source biological databases and experimental measurement data. In one embodiment, the data is collected from public exosome-target interaction databases (such as ExoCarta and Vesiclepedia) and in vitro binding experiment datasets reported in the literature, covering the molecular structure (SMILES format or three-dimensional coordinates) and amino acid sequence (FASTA format) of exosomes from different sources and their corresponding target biological information (receptor type, tissue distribution).

[0077] The data content includes atomic-level molecular graph structure (node ​​features include atom type, hybridization state, and formal charge), sequence residue features (amino acid type, hydrophilicity index, and isoelectric point), and target efficacy level labels (low / medium / high efficiency) measured by surface plasmon resonance (SPR) or flow cytometry. The target efficacy level labels are used as data labels for supervised training of the evaluation model.

[0078] Optionally, the fusion feature matrix is ​​output in S201 , also includes:

[0079] Adjacency matrix based on chemical bonds, degree matrix of adjacency matrix of chemical bonds, and molecular graph structure input matrix Output graph convolution feature matrix ;

[0080] Combining one-dimensional convolution kernel and sequence input matrix Output sequence feature matrix ;

[0081] Based on graph convolution feature matrix and the sequence feature matrix Computational Structure - Sequence Attention Mask Matrix And output the fusion feature matrix through channel splicing operation .

[0082] Furthermore, heterogeneous molecular features are fused and embedded into graph structures:

[0083] In the evaluation of exosome-targeting molecules, the heterogeneity of molecular structure and sequence data leads to inconsistent feature representation. Conventional methods such as simple splicing or independent branch processing struggle to effectively capture the topological relationships between structure and function and tend to overlook interactions at the atomic / residue level, limiting the accuracy of the model's classification of targeting efficacy.

[0084] To address the heterogeneity problem between molecular structure data and sequence data and avoid the limited model accuracy caused by inconsistent feature representation, this paper uses a collaborative embedding mechanism of graph convolution and sequence convolution, combined with an attention mechanism to fuse features, uniformly represent heterogeneous data and enhance information interaction at the atomic or residue level. The specific steps are as follows:

[0085] 1) Multi-relation graph convolutional structure feature extraction:

[0086] Based on the molecular graph structure input matrix, independent adjacency matrices and weight matrices are used for feature transformation for different chemical bond types. Then, by aggregating the neighboring atom information of each atom and applying the ELU activation function, the graph convolution feature matrix is ​​output. The graph convolution feature matrix represents the high-level representation of the molecular structure, thereby capturing the topological associations related to the chemical bond types, avoiding feature fragmentation, and enhancing the model's ability to represent structural features. It is expressed as:

[0087] ;

[0088] Where, Represents the graph convolution feature matrix, the dimension is the number of atoms Multiply by the hidden layer dimension , a high-level representation used to characterize molecular structures, such as the number of atoms in an ethanol molecule , graph convolution aggregates the neighbor information of each atom and outputs the graph convolution feature matrix The dimension is , high-level representations that characterize molecular structures;

[0089] is the atomic number;

[0090] is the hidden layer dimension, such as, ;

[0091] is a set of chemical bond types, defined artificially, such as, For a single key, For double bonds, Corresponding to three keys, for example, For covalent bonds, For hydrogen bonds, For each type of chemical bond, a separate adjacency matrix is ​​used to describe its topological connection relationship in the molecule.

[0092] is a positive integer;

[0093] Represents the ELU activation function, which enhances the model's expressiveness through nonlinearity;

[0094] For the The degree matrix of the chemical bond-like adjacency matrix contains the degree of each atom, that is, the number of neighbor atoms;

[0095] for The inverse matrix of is used to normalize the adjacency weights to stabilize training.

[0096] For the The adjacency matrix of chemical bonds, where the element value is 0 or 1, indicates whether there is such a bond connection between atoms;

[0097] Input matrix for molecular graph structure, dimension is the number of atoms Multiply by the atomic feature dimension , characterizes chemical properties such as atomic type and bonding state. For example, the characteristic vector of a carbon atom may include one-hot encoding of atomic type, electronegativity, valence state, etc.;

[0098] is the atomic feature dimension, such as, ;

[0099] is the graph convolution weight matrix, for the Chemical bonds implement feature transformation and are trainable parameters.

[0100] 2) One-dimensional sequence convolution local feature extraction:

[0101] Based on the sequence input matrix, the convolution kernel is used to slide and scan on the sequence dimension, and the ReLU activation function is applied for nonlinear mapping to output the sequence feature matrix. The sequence feature matrix represents the local semantic information of the sequence, thereby capturing the local pattern of the amino acid sequence, overcoming the problem of simple splicing ignoring the sequence context, and suppressing noise interference. It is expressed as:

[0102] ;

[0103] Where, Represents the sequence feature matrix, the dimension is the sequence length Multiply by the hidden layer dimension , used to represent the local semantic information of the sequence, such as, for a peptide sequence "LYS-ARG-GLU", the sequence length , one-dimensional convolution extracts local patterns, and After the activation function performs nonlinear mapping, the output sequence feature matrix Represent the local semantic information of the sequence;

[0104] is the sequence length;

[0105] for Activation function, which enhances feature sparsity to suppress noise;

[0106] Represents a one-dimensional convolution operation, which extracts features by sliding the convolution kernel on the input sequence, and, The convolution kernel parameters are , input is One-dimensional convolution operation;

[0107] Input matrix for the sequence, dimension is sequence length Multiply by the amino acid feature dimension ;

[0108] is the amino acid feature dimension, including amino acid type and physicochemical properties. For example, amino acid type is a 20-dimensional one-hot encoding, and physicochemical properties include hydrophilicity, molecular weight, etc., a total of 5 dimensions. , the input corresponding to the sequence "ALA" is a row of 25-dimensional vectors;

[0109] Represents the one-dimensional convolution kernel parameters, which are used for feature extraction in the sequence dimension and are trainable parameters.

[0110] 3) Attention-guided cross-modal feature fusion:

[0111] The attention mask matrix is ​​calculated based on the graph convolution feature matrix and the sequence feature matrix, and the mask dimension is adjusted through global average pooling. Then, the graph convolution features are projected into the sequence feature space using element-by-element multiplication and weighted concatenation operations combined with the feature alignment matrix, and the fused feature matrix is ​​output. This ensures that the interactions at the atomic or residue level are preserved, solves the problem of conventional methods ignoring cross-modal associations, and realizes the effective fusion of heterogeneous features, which can be expressed as:

[0112] ;

[0113] Where, represents the fusion feature matrix;

[0114] is the structure-sequence attention mask matrix, with dimension of number of atoms Multiply by the sequence length , the calculation method is expressed as ;

[0115] is the global average pooling operation, Used to transform the structure-sequence attention mask matrix The dimension is converted to the fusion feature matrix The same dimension, thus achieving element-by-element multiplication;

[0116] represents element-wise multiplication;

[0117] represents the channel concatenation operation, connecting different features along the column dimension;

[0118] is the fusion coefficient, , balancing the contributions of sequence and structural features, e.g. ;

[0119] is the feature alignment matrix, used to convolve the graph feature matrix Projected into the sequence feature space to achieve dimensionality matching, it is a trainable parameter.

[0120] Optionally, generating the fused feature vector in S202 further includes:

[0121] S2021. Dynamically modulate the fusion feature matrix using target-conditional gating mechanism The response intensity is used to obtain discriminative features that are strongly correlated with the target;

[0122] S2022. Combine the mutual information maximization principle and the multi-layer perceptron to sequentially aggregate and reorganize the discriminative features to generate a recombined feature vector ;

[0123] S2023. Reconstructing feature vectors via dual-channel gated residual networks Perform dynamic feature fusion to generate a fused feature vector.

[0124] Furthermore, target conditional gating feature dynamic masking includes:

[0125] Target biological information is often sparse and has complex nonlinear correlations with molecular features. Conventional feature selection methods, such as L1 regularization, cannot dynamically adapt to the differences between target-molecule combinations, resulting in the suppression of key correlation features or the retention of noise features, reducing the accuracy of target efficacy classification.

[0126] To address the sparsity and nonlinear correlation issues of target biological information and avoid suppressing key features or retaining noise, the present invention adopts a target-conditional gating mechanism to dynamically modulate the response intensity of fusion features, achieve feature adaptive masking, and highlight discriminative features that are strongly correlated with the target. The specific steps are as follows:

[0127] 1) Target feature-dependent gating vector generation:

[0128] After concatenating the fusion feature matrix with the target feature vector, a linear transformation is performed using learnable parameters, and a Sigmoid activation function is applied to generate a dynamic gating vector. This quantifies the importance of the corresponding feature channel to the current target, solving the problem of being unable to dynamically adapt to different target combinations and ensuring that the feature importance is dynamically adjusted according to the target biological information, which can be expressed as:

[0129] ;

[0130] Where, Represents a dynamic gating vector, where each element has a value range of [0, 1] and quantifies the importance of the corresponding feature channel;

[0131] is the matrix flattening operation, The fused feature matrix Stretched to a one-dimensional vector;

[0132] express Activation function;

[0133] is the gate weight matrix, which is used for linear transformation and splicing features and is a trainable parameter;

[0134] The target feature vector contains biological information such as receptor type and affinity baseline, which are obtained through experimental measurement or biological database. The receptor protein expression level is determined by ELISA, and the affinity baseline is obtained by SPR technology. They are encoded into a feature vector, for example, by one-hot encoding of the receptor type and quantization value.

[0135] is the gate bias vector, which is a trainable parameter.

[0136] In one embodiment, the target feature vector is constructed by one-hot encoding the receptor type and quantizing the value. , the specific method is as follows:

[0137] For receptor type one-hot encoding, the purpose is to convert discrete receptor categories into numerical form to facilitate model processing. For example, discrete receptor categories are different protein receptors. Assuming there are three receptor types, including EGFR, CD44 and HER2, the receptor type set is {EGFR, CD44, HER2}. A binary vector of length 3 is generated for each receptor type. If the current target belongs to the first Type receptor, then the vector Position , the rest are , a is a positive integer, then the encoding of each type of receptor is:

[0138] The encoding of EGFR is [1,0,0], the encoding of CD44 is [0,1,0], and the encoding of HER2 is [0,0,1];

[0139] For scalar values, the purpose is to convert continuous biological features into numerical values. For example, continuous biological features are "receptor protein expression", "affinity baseline" and "tissue specificity score", etc., and directly use experimental measurement values. For example, assuming that the biological feature "receptor protein expression" is measured by "ELISA measured concentration value", the specific scalar value is 5.2ng / mL, and the multi-dimensional features are normalized to avoid dimensional differences. Assume that the normalized value is 0.7, then, the one-hot encoding vector and the scalar value are spliced ​​along the one-dimensional direction to form a complete target feature vector. , then for the EGFR receptor type, its target feature vector is .

[0140] 2) Soft masking compensation feature selection:

[0141] The dynamic gating vector is used to perform weighted masking on the fusion feature matrix. At the same time, the low response features are processed by the compensation weight matrix to generate a masked feature matrix, thereby avoiding the complete filtering of key signals, overcoming the information loss defect of conventional binary masking, and compensating for the information of weak response features, which is expressed as:

[0142] ;

[0143] Where, represents the masked feature matrix;

[0144] The compensation weight matrix is ​​used to retain part of the information flow of the feature and is a trainable parameter;

[0145] is a unit vector;

[0146] is a unit vector Subtract the dynamic gate vector , characterizes the generated inverted mask, identifies the low-importance feature channel, and each element is minus The corresponding elements of , by generating a non-binary soft mask, allow the low-importance feature channel to partially pass through, avoiding complete information filtering.

[0147] Mutual information driven discriminative feature recombination includes:

[0148] There is a large amount of redundant information in high-dimensional fusion features. Traditional dimensionality reduction methods will destroy the discriminative association between features and categories, easily leading to the loss of key classification information and reducing the accuracy of targeted efficacy classification.

[0149] To solve the problem of redundant information in high-dimensional fusion features and avoid the destruction of discriminative associations by traditional dimensionality reduction methods, this paper uses the principle of maximizing mutual information to force similar features to cluster and reorganize feature channels by constructing category prototype centers and calculating losses. The specific steps are as follows:

[0150] 1) Dynamic category prototype center construction:

[0151] Based on the label distribution of the current batch of samples, the feature mean of each category is dynamically calculated as the prototype center vector. The masked feature vectors of samples belonging to the same category are averaged and the prototype center vector is output. This provides a category reference benchmark, avoids the distribution deviation caused by a fixed center, and adapts to dynamic changes in data. It is expressed as:

[0152] ;

[0153] Where, represents the prototype center vector of the jth category, and different categories correspond to different levels of targeting effectiveness;

[0154] is a positive integer;

[0155] is a small constant to prevent the denominator from being 0, such as, ;

[0156] is the set of sample indexes belonging to the jth category in the batch;

[0157] Representing a collection The cardinality is the number of samples of the j-th category in the batch;

[0158] is the masked feature vector of the i-th sample;

[0159] i is a positive integer.

[0160] 2) Calculation of intra-class aggregate mutual information loss:

[0161] The cosine similarity function is used to measure the correlation strength between the sample masked feature vector and the true category prototype center vector. The negative logarithmic loss is used to maximize the mutual information between the feature and the category. The mutual information loss is calculated based on batch samples to strengthen the aggregation of similar features and force the model to learn category-related patterns. It is expressed as:

[0162] ;

[0163] Where, Represents mutual information loss, which is used for back-propagation optimization of feature distribution;

[0164] is the batch size, that is, the number of single training samples;

[0165] is the true category label of the i-th sample;

[0166] is the prototype center vector corresponding to the true category label of the i-th sample;

[0167] For the The prototype center vector of each category;

[0168] is the cosine similarity function, which measures the directional consistency of two vectors;

[0169] represents the natural exponential operation;

[0170] Indicates logarithmic operation, the default base is 10;

[0171] It represents the sum of three categories, which refer to the target performance level, such as, Indicates low efficiency, Indicates medium performance, Indicates high efficiency.

[0172] 3) Feature dimensionality reduction and reorganization guided by class center:

[0173] The masked feature vector of the sample is concatenated with the center vector of its true category prototype, mapped into low-dimensional discriminative features through a multi-layer perceptron, and the reconstructed feature vector is output to eliminate redundant information and retain category-related patterns, which can be expressed as:

[0174] ;

[0175] Where, represents the reorganized feature vector;

[0176] Represents a multilayer perceptron, specifically a two-layer network structure, including linear transformation and ReLU activation function. The weights and biases of the multilayer perceptron are trainable parameters.

[0177] is the prototype center vector corresponding to the true label;

[0178] Indicates that the feature matrix will be masked The prototype center vector corresponding to the true label Concatenate along the column dimension.

[0179] The dual-channel gated residual network construction includes:

[0180] Conventional fully connected networks have difficulty in effectively modeling the complex nonlinear relationships between high-dimensional features, which can easily lead to gradient vanishing or insufficient feature representation capabilities, affecting the accuracy of targeted efficacy classification.

[0181] To solve the problem that conventional fully connected networks have difficulty modeling complex nonlinear relationships and avoid gradient vanishing or insufficient representation capabilities, this paper adopts a dual-channel residual structure and dynamically fuses features through a gating mechanism to enhance information flow. The specific steps are as follows:

[0182] 1) Nonlinear high-order feature transformation of the main channel:

[0183] Perform a linear transformation on the input feature vector and apply the ELU activation function to extract high-order nonlinear feature representations and output the main channel feature vector, thereby enhancing the model's expressiveness, overcoming the representation bottleneck of the fully connected network, and extracting richer feature representations, which can be expressed as:

[0184] ;

[0185] Where, Indicates the main channel Output features of the layer;

[0186] Main channel The weight matrix of the layer is a trainable parameter;

[0187] Main channel The output features of the layer, that is, the The input features of the layer, when When , the output feature of the first layer of the main channel ;

[0188] Main channel The bias vector of the layer is a trainable parameter.

[0189] 2) Input feature adaptive gate generation:

[0190] Based on the current layer input feature vector, a linear transformation is performed through learnable parameters, and a Sigmoid activation function is applied to generate a gating vector. This dynamically allocates the information ratio between the main channel and the skip path, adaptively adjusts the feature update strength, and avoids gradient disappearance. It is expressed as:

[0191] ;

[0192] Where, Indicates the gate control module The gating vector of the layer;

[0193] For the gate control module The weight matrix of the layer is a trainable parameter;

[0194] For the gate control module The gated bias vector of the layer is a trainable parameter;

[0195] 3) Gated weighted residual feature fusion:

[0196] The gating vector and the complementary gating vector are used to perform element-by-element weighted operations. The main channel feature vector and the skip connection feature vector are weightedly fused through the gating vector, and the fused feature vector is output. This preserves the original feature information, ensures stable training of the deep network, and alleviates the gradient problem by dynamically balancing the information of the main channel and the skip path. It is expressed as:

[0197] ;

[0198] Where, Indicates the The fused feature vector output by the layer;

[0199] For jump connection The weight matrix of the layer is a trainable parameter, and it is necessary to make Dimensions and characteristics The dimensions are forced to be consistent;

[0200] Represents the generation of a complementary gating vector, each element of which is minus The corresponding elements of , It provides complementary gating, dynamically balances the information of the main channel and the skipping path, and avoids gradient vanishing.

[0201] Furthermore, generating an optimized targeting effectiveness probability classifier based on the fused feature vector in S203 also includes:

[0202] S2031. Constructing a target efficacy probability classifier:

[0203] The high-dimensional fusion features extracted by the dual-channel residual network need to be mapped into target performance class probabilities. Traditional single-layer fully connected classifiers struggle to adequately model the complex decision boundaries between features and classes, and fixed weight assignments cannot adapt to dynamic changes in feature distribution, potentially reducing classification robustness.

[0204] The present invention converts the high-dimensional fusion features output by the dual-channel residual network into target efficacy category probabilities, overcoming the problem that traditional single-layer classifiers are difficult to model complex decision boundaries. The specific steps are as follows:

[0205] 1) Feature space to category score mapping:

[0206] The fusion feature vector finally output by the dual-channel residual network is multiplied by the classifier weight matrix to generate the original score vector of each category. Then, the high-dimensional feature space is mapped to the category score space. Each score corresponds to the possibility of a specific target effectiveness level, establishing a linear association basis between features and classification decisions, avoiding the lack of adaptability of fixed weight allocation to dynamic feature distribution, which is expressed as:

[0207] ;

[0208] Where, Indicates the The category score vector of samples, dimension is ;

[0209] is the total number of categories, i.e. the number of target efficacy categories, e.g. , corresponding to low, medium and high efficiency;

[0210] is the classifier weight matrix, whose The row vector represents the classifier weight vector of the jth category , is a trainable parameter;

[0211] is the fused feature vector output by the dual-channel residual network for the i-th sample.

[0212] 2) Probability normalized decision output:

[0213] Apply the Softmax function to the raw score vector to transform it and calculate the probability value of the sample belonging to each target efficacy category. This operation uses exponential operation and normalization to ensure that the sum of the probabilities of each category is 1 and conforms to the probability distribution characteristics. Effect: The raw score obtained by linear transformation is converted into an interpretable probability distribution, forming the basis for the final classification decision, improving the robustness and interpretability of the classification results:

[0214] ;

[0215] Where, Indicates the The samples belong to The probability of the category is calculated in the same way as the Softmax function;

[0216] For the The sample in The score on the category is calculated as ;

[0217] For the The sample in The score on the category is calculated as ;

[0218] is the classifier weight vector of the jth category, which is a trainable parameter;

[0219] The classifier weight vector for the kth category is a trainable parameter;

[0220] for The transpose of .

[0221] S2032, Adaptive Class Boundary Loss Optimization:

[0222] There is a class imbalance problem in the exosome targeting efficacy classification task. High-efficiency samples are scarce, and the standard cross-entropy loss function has difficulty distinguishing the boundaries of similar categories, which easily leads to the classification decision boundary being biased towards the majority class, thereby ignoring the key features of the rare class.

[0223] To address the class imbalance problem and prevent the classification decision boundary from being biased towards the majority class and ignoring the rare class, this paper adopts a learnable adaptive class boundary margin mechanism. By dynamically adjusting the classification interval, it forces the model to generate a larger decision boundary for the rare class, thereby improving the recognition ability of high-performance samples. The specific steps are as follows:

[0224] 1) Sample distribution-aware margin calculation:

[0225] According to the distribution of sample numbers in each category in the training set, based on the logarithmic proportional relationship of the number of category samples, the adjustment strength is controlled by the margin scaling factor, and the category-related boundary margin value is dynamically calculated. A larger decision boundary is automatically assigned to categories with sparse samples, forcing the model to pay more attention to the key features of high-performance samples and alleviate the classification bias caused by category imbalance. It is expressed as:

[0226] ;

[0227] Where, represents the boundary margin of the jth category;

[0228] is the number of samples of the jth category in the training set, reflecting the sparsity of category distribution;

[0229] is the total number of samples in the training set;

[0230] is the margin scaling factor, which controls the interval adjustment strength, e.g. .

[0231] 2) Margin-constrained cross entropy loss construction:

[0232] The adaptive margin value is embedded in the standard cross entropy loss function. The class differentiation constraint is imposed when calculating the classification loss. The indicator function is used to control the margin to act only on non-real classes. Class-specific interval adjustment is used in the score calculation to enhance the model's ability to identify rare classes and improve the classification accuracy of high-performance samples while maintaining the clear distinction of the decision boundaries of each class. It is expressed as:

[0233] ;

[0234] Where, represents the adaptive boundary loss;

[0235] is the classifier weight vector corresponding to the true category of the i-th sample, which is a trainable parameter;

[0236] for The transpose of

[0237] is the classifier weight vector of the jth category, which is a trainable parameter;

[0238] for The transpose of

[0239] is the fusion feature vector output by the dual-channel residual network of the i-th sample;

[0240] is the total number of layers of the dual-channel residual network;

[0241] For the The margin corresponding to the true category of the sample;

[0242] represents the boundary margin of the jth category;

[0243] is an indicator function, which has a value of 1 when the condition is met and 0 otherwise, i.e. Indicates when Not equal to The time value is , otherwise ;

[0244] is the true category label of the i-th sample.

[0245] Optionally, the trained exosome molecular structure evaluation model is obtained in S3, which also includes:

[0246] S301. Dynamically iteratively update all trainable parameters of the exosome molecular structure evaluation model based on mutual information loss and adaptive boundary loss;

[0247] S302. Obtain the number of iterations based on the adaptive boundary loss and classification accuracy of the validation set to generate a trained exosome molecular structure evaluation model.

[0248] Specifically, the performance of the validation set is continuously monitored during model training to determine when to stop iteration. Specifically, the adaptive boundary loss and classification accuracy of the validation set are calculated after each training cycle. When the validation loss does not decrease and the accuracy fluctuation range is less than 0.5% within 15 consecutive cycles, the model is determined to have reached a convergence state. If the number of training cycles reaches the preset upper limit and still has not converged, such as 500 cycles, the training is forcibly terminated to prevent overfitting.

[0249] Model saving follows the principle of optimal performance: only when the classification accuracy of the validation set exceeds the historical highest record, the current model parameters and its complete architecture are serialized into binary files for storage, and the training environment configuration is synchronously recorded when saving.

[0250] Optionally, parameter updates are performed on the exosome molecular structure assessment model, including:

[0251] S3011. Dynamically calculate mutual information loss based on mutual information loss and adaptive boundary loss Dynamic weight and adaptive boundary loss Dynamic weight ;

[0252] S3012. Combined mutual information loss Dynamic weight and adaptive boundary loss Dynamic weight Generate fused total gradient ;

[0253] S3013. Adam optimizer with gradient clipping and fused total gradient Iteratively update the parameters of the exosome molecular structure evaluation model.

[0254] Furthermore, S3, evaluation model training and saving also includes:

[0255] S301, Multi-objective dynamic gradient collaborative training:

[0256] Traditional backpropagation uses a single loss gradient to update parameters, while the present invention requires simultaneous optimization of mutual information loss and adaptive margin loss. Directly summing these losses can lead to training oscillations due to the dimensional differences between the two types of losses and the conflicting optimization objectives. Furthermore, fixed weights cannot adapt to changes in parameter sensitivity across different training stages, reducing model convergence efficiency and classification accuracy.

[0257] In order to coordinate the joint optimization of mutual information loss and boundary loss and solve the dimensional conflict and training oscillation problems of multi-objective loss, this paper designs a dynamic gradient gating collaborative mechanism to achieve gradient fusion and parameter update, which can be expressed as:

[0258] 1) Loss sensitivity normalization weighting:

[0259] Based on the gradient amplitude of the two types of losses in the current batch, the normalized weight coefficient is dynamically calculated. Combined with the training cycle attenuation factor, the contribution strength of each loss in different training stages is adaptively adjusted, thereby balancing the dimensional differences of multi-objective optimization, adapting to the parameter sensitivity changes during training, and preventing a single loss from dominating the update direction. It is expressed as:

[0260] ;

[0261] ;

[0262] Where, Mutual information loss Dynamic weight of

[0263] is the adaptive boundary loss Dynamic weight of

[0264] is the Frobenius norm;

[0265] represents the mutual information loss Relative to model parameters gradient;

[0266] represents the adaptive boundary loss Relative to model parameters gradient;

[0267] is a natural constant;

[0268] Control mutual information loss for the decay coefficient The contribution decays in the later stages of training, e.g. ;

[0269] is the current training cycle, representing the current number of iterations;

[0270] is the total number of training cycles, representing the total number of iterations.

[0271] 2) Target-gated gradient fusion:

[0272] Dynamic weights are used to perform weighted fusion of the gradients of the two types of losses to generate a unified total gradient signal. Through the target condition gating mechanism, parameters that are strongly related to the target are updated first, thereby coordinating the joint optimization of feature reconstruction and classification boundaries, ensuring the consistency of gradient update direction, and improving training stability. It is expressed as:

[0273] ;

[0274] Where, is the total fusion gradient.

[0275] 3) Stability constraint parameter update:

[0276] The Adam optimizer with gradient clipping is used to perform parameter updates to limit the impact of abnormal gradients. At the same time, spectral norm constraints are imposed on graph convolution weights to control the complexity of feature extraction, thereby preventing gradient explosion or overfitting during training and ensuring the stability of model convergence and the robustness of topological feature learning. It can be expressed as:

[0277] ;

[0278] Where, is the set of all trainable parameters;

[0279] For the The model parameters of the iteration represent the updated model parameters;

[0280] For the The model parameters of the iteration represent the model parameters before the update;

[0281] is the learning rate, such as, ;

[0282] Represents the Adam optimization algorithm;

[0283] Represents the gradient clipping operation;

[0284] is the gradient clipping threshold.

[0285] Furthermore, the evaluation results of exosome-targeted molecules obtained in S4 also include:

[0286] When using the trained model to evaluate the targeting efficacy of new molecules, the exosome molecules to be tested are first preprocessed to analyze their molecular structure into a graph input matrix, the sequence data is encoded into a sequence input matrix, and the target biological features are converted into target feature vectors.

[0287] Then, end-to-end reasoning is performed, and through data processing of S201-S205, the three-category targeting effectiveness probability distribution calculated by the classifier is obtained. The evaluation results are output in the form of probability vectors, and the category corresponding to the maximum probability is taken as the targeting effectiveness level.

[0288] Generate exosome targeting molecule assessment results based on targeting efficacy rating.

[0289] Specifically, when using the trained model to evaluate the targeting efficacy of new molecules, the exosome molecules to be tested are first preprocessed, their molecular structures are parsed into a graph input matrix, the sequence data is encoded into a sequence input matrix, and the target biological features are converted into target feature vectors;

[0290] Then, end-to-end reasoning is performed, and through data processing of S201-S205, the three-category targeting effectiveness probability distribution calculated by the classifier is obtained. The evaluation results are output in the form of probability vectors, and the category corresponding to the maximum probability is taken as the targeting effectiveness level.

[0291] Example 2:

[0292] like Figure 2 As shown in the figure, in order to evaluate the recognition ability of various models for low-, medium-, and high-efficiency exosome molecules, the classification performance (F1 score) of the five models at three targeting efficacy levels is intuitively presented through grouped bar charts. The experimental results show that conventional technologies (independent branch processing, L1 regularization, traditional full connection, and standard cross entropy) perform significantly worse in the high-efficiency category than in the medium- and low-efficiency categories, reflecting the insufficient recognition of rare high-efficiency samples by conventional methods. In contrast, the technology of the present invention maintains excellent performance at all three efficacy levels, especially leading significantly in the high-efficiency category, indicating that the heterogeneous feature fusion mechanism and adaptive class boundary optimization of the technology of the present invention effectively solve the class imbalance problem in the evaluation of exosome-targeted molecules and enhance the recognition ability of high-efficiency molecules.

[0293] like Figure 3As shown in the figure, in order to analyze the impact of training data scale on the generalization ability of the model, a line graph with error bars is used to show the verification accuracy change trend of the two technologies under different training sample sizes. Experiments show that conventional technologies perform poorly when the sample size is small. Although there is an improvement with the increase of samples, the growth rate is slow and the fluctuation is large, reflecting that its data utilization efficiency is limited. The technology of the present invention shows strong performance in small sample scenarios, and shows a more stable and rapid improvement with the increase of samples, and finally achieves a higher accuracy rate, which verifies the effectiveness of the collaborative embedding mechanism of graph convolution and sequence convolution of the technology of the present invention, and can extract richer feature representations from limited data. The error bar shows that the fluctuation range of the technology of the present invention is smaller, which proves that the training stability is better, and shows the balancing effect of the dynamic gradient gating collaborative mechanism on multi-target losses.

[0294] like Figure 4 As shown in the figure, in order to verify the superiority of the cross-modal attention fusion mechanism in integrating molecular structure graph data and amino acid sequence data, by comparing the performance of five methods including the use of structural features alone, the use of sequence features alone, simple feature splicing, traditional attention fusion and the cross-modal attention fusion proposed by the technology of the present invention in three key indicators: accuracy, high-efficiency recall rate and feature mutual information. The experimental results show that the cross-modal attention fusion method of the technology of the present invention is significantly ahead in all three indicators, especially in the high-efficiency recall rate. The different colored bars in the figure represent different methods, which highly intuitively reflect the performance differences. The technology of the present invention effectively solves the problem of inconsistent representation of heterogeneous features through attention interaction at the atom / residue level, enhances the ability to recognize high-efficiency targeted molecules, and thus comprehensively surpasses traditional methods in key indicators.

[0295] like Figure 5 To evaluate the effectiveness of the adaptive margin loss mechanism in addressing class imbalance, the classification performance and overall accuracy of four methods—traditional cross-entropy loss, focal loss, a non-adaptive version of the proposed technique, and the full version—are compared across low-efficiency, medium-efficiency, and high-efficiency categories. Four subfigures display the F1 score and overall accuracy for each category, with different colored bars used to distinguish between methods. Experimental results demonstrate that the full version of the proposed technique achieves significant improvements in the high-efficiency category, attributed to the dynamic margin mechanism's automatic allocation of a larger decision space for rare classes. As clearly visible in the figure, the full version's bar for the high-efficiency category is significantly taller than that of the other methods, while the overall accuracy bar also maintains a leading position. This demonstrates that the proposed mechanism improves the ability to identify rare classes without sacrificing overall performance, effectively addressing the critical issue of the scarcity of high-efficiency samples in targeted exosome assessment.

[0296] like Figure 6As shown in the figure, the optimization effect of the mutual information recombination mechanism is intuitively displayed through feature space visualization. The original feature space distribution, the feature space distribution after reorganization and the change of decision boundary are compared. The original feature space distribution diagram shows that the three types of samples in the original feature space are seriously mixed. The feature space diagram after mutual information reorganization shows that the same type of samples are obviously clustered and different types of samples are effectively separated after reorganization. The experimental results show that the mutual information recombination mechanism effectively eliminates redundant information through feature dimensionality reduction guided by class centers, so that the three types of low, medium and high efficiency samples form obvious clusters in the feature space, especially the aggregation degree of high efficiency samples is significantly improved.

[0297] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0298] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for evaluating exosome targeting molecules, characterized in that: The method comprises: S1. Construct training and validation sets based on multi-source biological databases and experimental measurement data; S2. Build an exosome molecular structure evaluation model based on the training set; S3. Train and save the adaptive boundary loss and classification accuracy evaluation model based on the validation set to obtain the trained exosome molecular structure evaluation model; S4. Obtain exosome targeting molecule evaluation results based on the trained exocrine molecular structure evaluation model; Among them, the exosome molecular structure evaluation model constructed in S2 includes: S201. Based on the collaborative embedding mechanism obtained through graph convolution and sequence convolution, combined with the attention mechanism fusion feature, the molecular structure data and sequence data are uniformly represented to generate a fusion feature matrix ; S202. The fusion feature matrix is ​​sequentially fused through the target condition gating mechanism, the mutual information maximization principle and the dual-channel gated residual network. Processing is performed to generate a fused feature vector; S203. Generate an optimized targeting efficacy probability classifier based on the fused feature vector.

2. The method for evaluating exosome targeting molecules according to claim 1, wherein: Output fusion feature matrix in S201 , also includes: Adjacency matrix based on chemical bonds, degree matrix of adjacency matrix of chemical bonds, and molecular graph structure input matrix Output graph convolution feature matrix ; Combining one-dimensional convolution kernel and sequence input matrix Output sequence feature matrix ; Based on graph convolution feature matrix and the sequence feature matrix Computational Structure - Sequence Attention Mask Matrix And output the fusion feature matrix through channel splicing operation .

3. The method for evaluating exosome targeting molecules according to claim 2, wherein: Graph convolution feature matrix , expressed as: ； Where, Represents the graph convolution feature matrix; is a set of chemical bond types; is a positive integer; Represents the ELU activation function, which enhances the model's expressiveness through nonlinearity; For the degree matrix of the chemical bond-like adjacency matrix; for The inverse matrix of For the Adjacency matrix of chemical-like bonds; Input matrix for molecular graph structure; is the graph convolution weight matrix.

4. The method for evaluating exosome targeting molecules according to claim 2, wherein: Fusion feature matrix , expressed as: ； Where, represents the fusion feature matrix; is the structure-sequence attention mask matrix; is the global average pooling operation; represents element-wise multiplication; represents the channel concatenation operation, connecting different features along the column dimension; is the fusion coefficient; is the feature alignment matrix.

5. The method for evaluating exosome targeting molecules according to claim 4, wherein: Generating the fused feature vector in S202 also includes: S2021. Dynamically modulate the fusion feature matrix using target-conditional gating mechanism The response intensity is used to obtain discriminative features that are strongly correlated with the target; S2022. Combine the mutual information maximization principle and the multi-layer perceptron to sequentially aggregate and reorganize the discriminative features to generate a recombined feature vector ; S2023. Reconstructing feature vectors via dual-channel gated residual networks Perform dynamic feature fusion to generate a fused feature vector.

6. The method for evaluating exosome targeting molecules according to claim 5, wherein: S2021 obtains discriminative features that are strongly correlated with the target, including: Combined fusion feature matrix and the preset target feature vector Generate dynamic gating vectors ; Use dynamic gating vector g to fusion feature matrix Perform weighted masking and combine it with the compensation weight matrix Generate masked feature matrix : ； Where, is the compensation weight matrix; is a unit vector; is a unit vector Subtract the dynamic gate vector , characterize the generated inversion mask; Combined masked feature matrix And the fusion feature matrix Generate discriminative features.

7. The method for evaluating exosome targeting molecules according to claim 1, wherein: S203 generates an optimized targeting effectiveness probability classifier based on the fused feature vector, further comprising: S2031. Constructing a targeting efficacy probability classifier based on the fused feature vector; S2032. Adopt the adaptive class boundary margin mechanism to dynamically adjust the classification interval of the target effectiveness probability classifier to generate an optimized target effectiveness probability classifier.

8. The method for evaluating exosome targeting molecules according to claim 1, wherein: S3 obtains the trained exosome molecular structure evaluation model, which also includes: S301. Dynamically iteratively update all trainable parameters of the exosome molecular structure evaluation model based on mutual information loss and adaptive boundary loss; S302. Obtain the number of iterations based on the adaptive boundary loss and classification accuracy of the validation set to generate a trained exosome molecular structure evaluation model.

9. The method for evaluating exosome targeting molecules according to claim 1, wherein: Parameter updates are performed on the exosome molecular structure assessment model, including: S3011. Dynamically calculate mutual information loss based on mutual information loss and adaptive boundary loss Dynamic weight and adaptive boundary loss Dynamic weight ; S3012. Combined mutual information loss Dynamic weight and adaptive boundary loss Dynamic weight Generate fused total gradient ; S3013. Adam optimizer with gradient clipping and fused total gradient Iteratively update the parameters of the exosome molecular structure evaluation model.

10. The method for evaluating exosome targeting molecules according to claim 1, wherein: Calculate mutual information loss Dynamic weight and adaptive boundary loss Dynamic weight , also includes: ； ； Where, Mutual information loss Dynamic weight of is the adaptive boundary loss Dynamic weight of is the Frobenius norm; represents the mutual information loss Relative to model parameters gradient; represents the adaptive boundary loss Relative to model parameters gradient; is a natural constant; Control mutual information loss for the decay coefficient Contribution decays in the later stages of training; is the current training cycle; is the total number of training cycles.

Images