Antibiofilm peptide classification method and system based on transfer learning and multi-source feature fusion

Abstract

The application provides a kind of antimicrobial peptide classification method and system based on transfer learning and multi-source feature fusion, relating to bioinformation analysis technical field, the method includes: obtaining the peptide sequence to be classified, respectively input multiple pre-training protein language model to extract semantic embedding features, and extract physicochemical characteristics and statistical characteristics, and obtain multi-source feature vector by fusion;Grouping perception based on feature source is carried out feature screening, and the first feature subset is obtained;The first feature subset is sorted according to importance, and the features meeting the threshold are selected to carry out cross product operation to construct interaction features and screening, and the second feature subset is obtained;The two feature subsets are spliced and fused, and then encoded by Transform, the obtained hidden layer representation is respectively input into the first classifier and the second classifier, and the output probabilities of the two are soft voted and integrated to determine the result.The application realizes the accurate classification of antimicrobial peptides under the conditions of small sample and imbalance.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics analysis technology, and in particular to a method and system for classifying anti-membrane peptides based on transfer learning and multi-source feature fusion. Background Technology

[0002] Anti-Biofilm peptides (ABPs) are a class of bioactive peptides that can inhibit or disrupt bacterial biofilm formation, and they have significant clinical application value in addressing the antibiotic resistance crisis and developing novel anti-infective drugs. With the rapid development of high-throughput sequencing technology, efficiently and accurately identifying candidate peptides with anti-biofilm activity from massive peptide sequences has become a key requirement in the field of drug discovery.

[0003] Currently, the identification of anti-biomembrane peptides mainly relies on wet experimental methods based on chemical synthesis and biological experiments, which are time-consuming and costly, making it difficult to meet the needs of large-scale screening. In terms of computational methods, only the hiABF method proposed by Tarki is currently available. This method constructs traditional feature vectors based on the primary structure, amino acid sequence, and physicochemical properties of the peptide sequence, and then uses classic machine learning algorithms such as support vector machines and random forests for prediction.

[0004] However, the distribution of positive and negative samples of anti-biomembrane peptides is inherently severely imbalanced. Wet experiments have verified that the number of peptides with anti-biomembrane activity is far less than that of inactive or weakly active peptides. This directly results in a very limited amount of labeled training data. Furthermore, obtaining new positive samples relies on expensive biochemical experiments, making high-quality target domain data suitable for model training extremely scarce. As a traditional supervised learning approach, the hiABF method struggles to effectively learn the discrimination boundary from the few positive samples in scenarios with small sample sizes and a very low proportion of positive samples, leading to significantly insufficient classification accuracy. Summary of the Invention

[0005] To address the aforementioned technical problems in the existing technology, this invention provides a method and system for classifying anti-biomembrane peptides based on transfer learning and multi-source feature fusion.

[0006] The first aspect of this invention provides a method for classifying anti-biomembrane peptides based on transfer learning and multi-source feature fusion, comprising: Obtain the peptide sequence to be classified, input the peptide sequence to be classified into multiple pre-trained protein language models to extract semantic embedding features, and extract the physicochemical and statistical features of the peptide sequence to be classified. Then, fuse the semantic embedding features with the physicochemical and statistical features to obtain a multi-source feature vector. Based on the feature source, the multi-source feature vectors are grouped and filtered to obtain the first feature subset; The first feature subset is sorted by importance, and features that meet a preset threshold are selected for cross-product operation to construct interactive features. The interactive features are then filtered to obtain the second feature subset. The first feature subset and the second feature subset are concatenated and fused, and then input into the Transformer encoder for deep semantic encoding to obtain the hidden layer representation; The hidden layer representation is input into the first classifier and the second classifier respectively, and the first probability output by the first classifier and the second probability output by the second classifier are integrated by soft voting. Based on the integrated probability, it is determined whether the peptide sequence to be classified is an anti-biomembrane peptide.

[0007] Furthermore, the multiple pre-trained protein language models include ESM-2, ProtT5, ProtBERT, and ProtT5-XL. The hidden states output by the encoders of each model are extracted and averaged along the sequence length dimension to obtain the sequence-level embedding vectors corresponding to each model. The physicochemical and statistical features include amino acid composition features, autocorrelation features, reduction-based dipeptide component features, and Moreau-Broto features. The four embedding vectors are concatenated in parallel with the physicochemical and statistical features to obtain the multi-source feature vector.

[0008] Furthermore, the group-aware feature selection adopts a group-aware feature sampling strategy: the multi-source feature vector is divided into several feature groups according to the feature source, and the sampling quota of each group is determined by weighting the square root of the dimension of each feature group at each decision tree split node. Candidate features are sampled from each group to ensure that each feature group has a competitive opportunity that matches its information content. The group-aware feature selection also adopts a random pruning strategy. After the optimal split is completed at each internal node of each decision tree, the split is terminated early by randomly deciding whether to mark the node as a leaf node with a preset pruning probability.

[0009] Furthermore, in the group-aware feature sampling strategy, the sampling quota for each feature group is calculated according to the following formula:

[0010] in, For the total number of features, For the first Group feature dimensions For the first Group feature dimensions This represents the total number of feature groups.

[0011] Furthermore, the Transformer encoder sequentially includes a linear projection layer, a position encoding layer, and stacked multi-head self-attention layers and feedforward network layers. The linear projection layer is used to map the spliced ​​and fused features to a dense representation with a preset hidden dimension. The position encoding layer is used to inject virtual position information into the mapped features. The multi-head self-attention layer is used to capture the dependencies between features in parallel from multiple representation subspaces. The soft voting integration adopts a probability averaging method. When the average probability is greater than a preset classification threshold, the peptide sequence is determined to be an anti-biomembrane peptide.

[0012] Furthermore, the Transformer encoder is trained using a two-stage transfer learning strategy: in the first stage, the Transformer encoder is pre-trained on the source domain antimicrobial peptide dataset using the cross-entropy loss function; in the second stage, the pre-trained Transformer encoder is fine-tuned on the target domain antibiofilm peptide dataset using the Focal loss function, which is used to reduce the loss weight of easily classified samples and increase the loss weight of difficult-to-classify samples.

[0013] A second aspect of the present invention provides an anti-biomembrane peptide classification system based on transfer learning and multi-source feature fusion, comprising: The feature extraction and fusion module is used to obtain the peptide sequence to be classified, input the peptide sequence to be classified into multiple pre-trained protein language models to extract semantic embedding features, extract the physicochemical features and statistical features of the peptide sequence to be classified, and fuse the semantic embedding features with the physicochemical features and statistical features to obtain a multi-source feature vector. The feature filtering module is used to perform group-aware feature filtering on the multi-source feature vectors based on the feature source to obtain a first feature subset; An interactive feature construction module is used to sort the first feature subset by importance, select features that meet a preset threshold for cross-product operation to construct interactive features, and filter the interactive features to obtain a second feature subset. The encoding module is used to concatenate and fuse the first feature subset and the second feature subset, input them into the Transformer encoder for deep semantic encoding, and obtain the hidden layer representation; The classification integration module is used to input the hidden layer representation into the first classifier and the second classifier respectively, and to perform soft voting integration on the first probability output by the first classifier and the second probability output by the second classifier, and to determine whether the peptide sequence to be classified is an anti-biomembrane peptide based on the integration probability.

[0014] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion as described in the first aspect of the present invention.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion as described in the first aspect of the present invention.

[0016] A fifth aspect of the present invention provides a computer program product comprising software code, wherein the program in the software code performs the steps of the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion as described in the first aspect of the present invention.

[0017] Compared with existing technologies, the present invention provides a method and system for classifying anti-biomembrane peptides based on transfer learning and multi-source feature fusion. Through the deep fusion of multi-source protein language model embedded features and physicochemical statistical features, two-stage differentiated random forest feature screening and cross-product interactive feature modeling, two-stage transfer learning strategy, and the synergistic effect of soft voting integration of dual classifiers, it effectively solves the technical problems of single feature source, lack of high-order interactive modeling, and insufficient generalization ability under small sample and imbalanced conditions in the field of computational recognition of anti-biomembrane peptides, and significantly improves the accuracy and robustness of anti-biomembrane peptide classification. Attached Figure Description

[0018] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0019] Figure 1 This is a flowchart of the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion provided in Embodiment 1 of the present invention.

[0020] Figure 2 This is an architecture diagram of the anti-biomembrane peptide classification system based on transfer learning and multi-source feature fusion provided in Embodiment 2 of the present invention. Detailed Implementation

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms "comprising," "having," and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0025] Example 1 like Figure 1 This embodiment provides a method for classifying anti-membrane peptides based on transfer learning and multi-source feature fusion, including: S101, obtaining the peptide sequence to be classified, inputting the peptide sequence to be classified into multiple pre-trained protein language models to extract semantic embedding features, and extracting the physicochemical and statistical features of the peptide sequence to be classified, and fusing the semantic embedding features with the physicochemical and statistical features to obtain a multi-source feature vector.

[0026] Specifically, firstly, FASTA format files of source and target domain peptide sequences are obtained from an antimicrobial peptide database. Then, the FASTA sequences are input into the following four pre-trained protein language models for embedding feature extraction: (1) ESM-2 (650M parameters): Extract the hidden state output by the 33rd layer Transformer encoder, perform mean pooling on the sequence length dimension, and obtain a 1280-dimensional sequence-level embedding vector.

[0027] (2) ProtT5 (3B parameters): Extract the hidden state of the last layer of the encoder, and obtain a 1024-dimensional embedding vector after average pooling.

[0028] (3) ProtBERT (420M parameters): Extract the hidden state of the last layer of the encoder, and obtain a 1024-dimensional embedding vector after average pooling.

[0029] (4) ProtT5-XL (11B parameters): Extract the hidden state of the last layer of the encoder, and obtain a 1024-dimensional embedding vector after average pooling.

[0030] The four sets of embedding features mentioned above originate from different training datasets, comprehensively reflecting the sequence evolution information and three-dimensional structural features of proteins, together forming a multi-dimensional, highly generalizable feature representation system. Furthermore, physicochemical and statistical features (amino acid composition features AAC, autocorrelation features AC, reduction-based dipeptide component features RAAC, Moreau-Broto features, etc.) are calculated independently and then concatenated to form a 152-dimensional feature representation. Finally, the four sets of embedding vectors are concatenated in parallel with the physicochemical and statistical features to obtain a 4656-dimensional multi-source fusion feature vector.

[0031] Furthermore, row standardization is applied to the obtained multi-source fusion feature vectors to eliminate dimensional differences between feature dimensions. For each sample's feature vector... Its standardized characteristics are:

[0032] in, The mean of this feature. The standard deviation of this feature. It is a non-zero, minimal numerical stability factor used to prevent... When the vector approaches zero, computational overflow occurs. After standardization, a standardized multi-source fusion feature vector is obtained, which serves as the input for subsequent feature selection.

[0033] S102. Based on the feature source, perform group-aware feature filtering on the multi-source feature vectors to obtain the first feature subset. .

[0034] Because traditional random forests uniformly and randomly sample candidate features from all features at each split node, small-dimensional feature groups are systematically ignored in most nodes. To address this, this step modifies the traditional random forest algorithm in two ways: group-aware feature sampling and random pruning, to adapt to multi-source heterogeneous features and prevent overlearning, thus constituting the coarse-screening stage.

[0035] Improvement direction one: Group-based perceptual feature sampling. Divide the 4656-dimensional features according to their source in S101. There are three feature groups: ESM-2 (1280 dimensions), ProtT5 (1024 dimensions), ProtBERT (1024 dimensions), ProtT5-XL (1024 dimensions), and physicochemical statistics (152 dimensions). At each decision tree split node, instead of uniformly random sampling from all features, the sampling quota for each feature group is determined by weighting the square root of its dimensions, and candidate features are sampled from each group accordingly. The sampling quota for each feature group is calculated using the following formula:

[0036] in, Total number of features ( ), For the first Group feature dimensions For the first Group feature dimensions The total number of feature groups is G=5. This weighting strategy increases the average number of candidates per node in the physicochemical statistical feature group from about 2.2 to about 6, ensuring that small-source features have a competitive opportunity that matches their information content.

[0037] The second improvement direction is random pruning. To prevent individual decision trees in a random forest from overfitting to noisy patterns in the training data, a random pruning mechanism is introduced during the growth of each decision tree: after each internal node completes its optimal split, a pruning probability is set. A random decision is made as to whether to mark the node as a leaf node and terminate the split prematurely. This random pruning operation is independent of the maximum depth limit. Minimum number of samples for nodes Even if the node depth is not reached And the number of samples in the current node exceeds The node may still terminate prematurely due to random pruning.

[0038] node The impurity of Gini is defined as:

[0039] in, Number of categories ( ), For nodes Medium category The percentage of samples. Nodes The reduction in Gini impurity (i.e., splitting gain) before and after splitting is:

[0040] in, , , These represent the number of samples for the current node, its left child node, and its right child node, respectively.

[0041] In this step, build Decision trees, each with a maximum depth limit of 1. Minimum number of samples per node Random sampling is performed at each internal node. Optimal splitting search is performed on each candidate feature. The probability of random pruning is taken as... This means that on average, each node has a 10% probability of being randomly pruned and prematurely terminated.

[0042] feature The global importance in the coarse-grained random forest is:

[0043] in, For the total number of decision trees, For standard Gini gain, For indicator functions, when node Use features The value is 1 when a node splits, and 0 otherwise. Due to random pruning, only the split gain of non-pruned nodes is accumulated and included in the importance.

[0044] By global importance All features are sorted in descending order, and an adaptive truncation strategy is used to determine the number of features to retain: Before calculation Cumulative importance contribution rate of each feature:

[0045] Take the satisfaction The smallest As the retained feature number. Take the first one. These features constitute the first feature subset. (In this embodiment) First feature subset The vector representation of it is denoted as This is used for subsequent splicing and fusion operations.

[0046] S103. Sort the first feature subset by importance, select features that meet a preset threshold, perform cross-product operation to construct interactive features, and filter the interactive features to obtain the second feature subset. .

[0047] This step is the fine screening stage, which differs fundamentally from the coarse screening stage S102 in terms of algorithm strategy. S102 processes various raw features with the aim of mining meaningful main effect features; S103, in order to fully explore the correlation between features, constructs cross-product features by multiplying the most important features in S102 pairwise, with the goal of deeply mining the optimal interaction features in the cross-product space. The two complement each other and work synergistically.

[0048] Specifically, the first feature subset output from S102 Among them, the features with the highest importance are selected first. For any two distinct features, ... and ( Construct the cross-product feature using the following formula:

[0049] in, and The respective The first of the features The and the first Several features were obtained. In this embodiment, a total of [number] features were obtained. There are several cross-product features. Since the cross-product features have no natural grouping and their number is controllable, the candidate sampling of random forest is changed to exhaustive search, without grouping constraints.

[0050] Tree structure configuration for a finely screened random forest (S103 RF): Construction Maximum depth of the tree The decision tree with the minimum number of samples per node The standard Gini gain is used to evaluate the split quality. A random pruning mechanism is also introduced, whereby, after each internal node completes the optimal split, a random pruning is performed with probability... The decision to terminate early is made randomly.

[0051] feature The global importance in a finely screened random forest is:

[0052] in, This is the standard Gini gain. (All) After the training of each tree is completed, press Sort in descending order, take the first few. The most discriminative cross-product features constitute the second feature subset. Second feature subset The vector representation of it is denoted as This is used for subsequent splicing and fusion operations.

[0053] In addition, when each decision tree in the finely screened random forest splits, the source combination type of each cross-product feature pair is explicitly recorded (homogeneous PLM intra-cross, cross-PLM cross, PLM-physicochemical cross), providing a basis for tracing the interaction type for subsequent biological interpretability analysis.

[0054] S104. The first feature subset and the second feature subset are concatenated and fused, and then input into the Transformer encoder for deep semantic encoding to obtain the hidden layer representation.

[0055] First, the first feature subset vector representation With the second feature subset vector representation Element-wise concatenation along the feature dimension yields the final classification feature vector:

[0056] Here, ∥∥ represents element-wise concatenation along the feature dimension. The first feature subset is 1020-dimensional, and the second feature subset is 30-dimensional. The two are concatenated element-wise along the feature dimension to obtain a final classification feature vector of 1050 dimensions. L2 normalization is used to eliminate scale differences from different feature sources. This hybrid feature vector has both global biological semantics (deep evolutionary signals from PLM embedding) and local high-order interaction characteristics (explicit modeling of key physicochemical mechanisms related to biofilm activity, such as surface charge distribution, amphiphilic helical structure, and hydrophobic torque, from cross-product features).

[0057] Subsequently, the 1050-dimensional final classification feature vector is input into the pre-trained Transformer encoder for deep semantic encoding. The encoding process includes the following three steps: Step 1: Linear Projection. Transform the 1050-dimensional features using a linear transformation. Mapped to a 128-dimensional hidden space, where This compresses high-dimensional sparse features into low-dimensional dense representations.

[0058] Step 2: Sine Position Encoding. Virtual position information is injected into the feature vector using sinusoidal position encoding. The position encoding formula is:

[0059]

[0060] in, For virtual location indexes, a fixed value is used in this embodiment. This means that the feature vector as a whole is treated as a single position, without needing to encode relative order relationships between multiple positions; To pre-define the hidden dimension, in this embodiment... ; For dimensional indexing, .

[0061] Step 3: Transformer encoder stacking processing. Two TransformerEncoderLayer layers are stacked, each layer containing: ① Multi-head self-attention sub-layer (4 attention heads, each head dimension) ), calculated using the following formula:

[0062] in, This represents the query vector, used to represent the query information corresponding to the currently viewed location; This represents the key vector, used to represent the feature identifiers at each position in the sequence; Value vectors are used to represent the actual content information at each location; scaling factor. This is used to prevent the gradient of the dot product result from vanishing as the dimension increases; ② Feedforward network sub-layer: The first linear layer maps 128 dimensions to 256 dimensions, introduces nonlinearity through the GELU activation function, and the second linear layer compresses 256 dimensions back to 128 dimensions; ③ Layer normalization and residual connections (Pre-LN structure). The encoded output is a 128-dimensional hidden layer representation.

[0063] The Transformer encoder is trained using a two-stage transfer learning strategy: The first stage involves pre-training the Transformer encoder on the source domain antimicrobial peptide dataset using the cross-entropy loss function (40 epochs), enabling the model to learn discriminative representations of general antimicrobial peptide sequences; the second stage involves fine-tuning the pre-trained Transformer encoder on the target domain antibiofilm peptide dataset using the Focal loss function (25 epochs). The Focal loss function is calculated as follows:

[0064] in This represents the model's predicted probability of the true class. The focus parameter is used to reduce the loss weight of easily classified samples and increase the loss weight of difficult-to-classify samples in order to overcome the problems of small sample size in the target domain and class imbalance.

[0065] S105. Input the hidden layer representation into the first classifier and the second classifier respectively, and perform soft voting integration on the first probability output by the first classifier and the second probability output by the second classifier. Determine whether the peptide sequence to be classified is an anti-biomembrane peptide based on the integrated probability.

[0066] In this embodiment, the first classifier uses a Transformer classifier head, the structure of which is as follows: a first linear layer maps the 128-dimensional hidden layer representation to 64 dimensions, introduces a non-linear transformation through the GELU activation function, performs regularization with a dropout rate of 0.2 through a Dropout layer to prevent overfitting, a second linear layer maps the 64 dimensions to 2-dimensional class scores, and normalizes the scores to class probabilities through the Softmax function, outputting the probability predicted by the Transformer model. .

[0067] The second classifier uses an RBF kernel support vector machine, trained on a Borderline-SMOTE oversampled training set, with class_weight='balanced' and probability=True to output probabilistic prediction results. .

[0068] Leveraging the fundamental difference in decision boundaries between Transformer and SVM (global nonlinearity of neural networks vs. local boundary maximization of kernel methods), complementary classification confidences are extracted from the same hidden representation. The classification probabilities of both are then ensembled using soft voting; the ensemble probability is calculated using the following formula:

[0069] The decision rule is: when The peptide sequence to be classified is determined to be a peptide with anti-biomembrane activity; otherwise, it is determined to be a non-anti-biomembrane peptide. Soft voting is used to reduce the decision variance of a single classifier and improve classification robustness. When the outputs of two classifiers are consistent, the confidence level is strengthened; when the outputs are inconsistent, the average value is near the boundary, and a threshold is used to make the final decision.

[0070] Furthermore, the model's classification performance was comprehensively evaluated using five metrics: Matthews Correlation Coefficient (MCC), Accuracy (ACC), F1 score (F1), Sensitivity (Sn), and Specificity (SP). (1) Matthews correlation coefficient (MCC):

[0071] It is used to measure the consistency between the model's predictions and the true labels, and has strong robustness to class imbalance.

[0072] (2) Accuracy (ACC):

[0073] This indicates the proportion of samples that the model predicts correctly overall.

[0074] (3) F1 score:

[0075] It is the harmonic mean of precision and recall, which comprehensively reflects the balanced performance of the classifier.

[0076] (4) Sensitivity (Sn):

[0077] This indicates the proportion of peptides that the model correctly identifies as having anti-biomembrane activity.

[0078] (5) Specificity (SP):

[0079] This indicates the proportion of non-anti-biomembrane peptides correctly identified by the model.

[0080] Among the above metrics, TP (True Positive) represents the number of positive samples correctly predicted as positive by the model; TN (True Negative) represents the number of negative samples correctly predicted as negative by the model; FP (False Positive) represents the number of negative samples incorrectly predicted as positive by the model; and FN (False Negative) represents the number of positive samples incorrectly predicted as negative by the model.

[0081] In one specific embodiment of the present invention, the target domain dataset contains 1331 samples (308 positive classes / 1023 negative classes). Five-fold hierarchical cross-validation is used for performance evaluation. The classification performance of the method of the present invention is compared with that of the ESM-2 baseline method. The results are shown in the table below: Table 1. Comparison of classification performance between the method of this invention and the ESM-2 baseline method.

[0082] The confusion matrix summary results are: TN=990, FP=33, FN=67, TP=241.

[0083] Ablation experiments show that the performance is optimal when the number of cross-product features N=30, and the mean decreases when N=45; the soft voting weight of 0.5:0.5 is close to optimal; the variance between folds ( ) is much larger than the variance between seeds ( This indicates that data partitioning has a significantly greater impact on performance than random initialization.

[0084] Example 2 like Figure 2 This embodiment provides an anti-biomembrane peptide classification system based on transfer learning and multi-source feature fusion, including: The 201 feature extraction and fusion module is used to obtain the peptide sequence to be classified, input the peptide sequence to be classified into multiple pre-trained protein language models to extract semantic embedding features, extract the physicochemical features and statistical features of the peptide sequence to be classified, and fuse the semantic embedding features with the physicochemical features and statistical features to obtain a multi-source feature vector. 202 Feature filtering module, used to perform group-aware feature filtering on the multi-source feature vectors based on feature sources to obtain a first feature subset; 203 Interactive feature construction module, used to sort the first feature subset by importance, select features that meet a preset threshold for cross-product operation to construct interactive features, and filter the interactive features to obtain a second feature subset; The 204 encoding module is used to concatenate and fuse the first feature subset and the second feature subset, and input them into the Transformer encoder for deep semantic encoding to obtain the hidden layer representation; The 205 classification integration module is used to input the hidden layer representation into the first classifier and the second classifier respectively, and to perform soft voting integration on the first probability output by the first classifier and the second probability output by the second classifier, and to determine whether the peptide sequence to be classified is an anti-biomembrane peptide based on the integration probability.

[0085] Example 3 Embodiment 3 of the present invention provides an electronic device.

[0086] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. The processor includes, but is not limited to, at least one of a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a tensor processor (TPU), or an artificial intelligence acceleration chip. The program is used to execute the steps in the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion as described in Embodiment 1 of the present invention.

[0087] In one specific embodiment, the electronic device includes one or more central processing units, one or more graphics processing units, main memory, and auxiliary memory. Wherein: The auxiliary storage is used to persistently store pre-trained protein language model weight files (including ESM-2, ProtT5, ProtBERT, and ProtT5-XL), physicochemical and statistical feature calculation programs, transfer learning backbone network parameters, trained hybrid deep learning classification model parameters, and protein sequence data to be predicted.

[0088] The main memory is used to temporarily cache the resampled and length-normalized sequence data output by the sequence preprocessing module, the four sets of high-dimensional semantic embedding vectors output by the embedding feature extraction module, the individualized descriptors output by the physicochemical and statistical feature calculation module, the parallel concatenated joint feature vectors output by the feature fusion module, and the intermediate calculation results of the classification decision module.

[0089] The central processing unit is used to execute the logic control of the sequence preprocessing module (including resampling algorithm, sequence similarity calculation and length statistics), coordinate the data flow and pipeline scheduling between modules, and schedule program instructions and feature data in the auxiliary memory through the main memory.

[0090] The graphics processor is used to accelerate forward inference computation of protein language models (ESM-2, ProtT5, ProtBERT and ProtT5-XL) in the embedded feature extraction module, as well as parallel acceleration computation of Transformer multi-head self-attention mechanism in hybrid deep learning classifiers. It efficiently processes hierarchical local patterns and long-range dependency modeling of anti-biomembrane peptide sequences through high parallel computing capabilities.

[0091] The medium (auxiliary storage) for storing program code includes: hard disk drive, solid-state drive, electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), optical storage device, magnetic storage device, flash memory, magnetic disk or optical disk and / or combination of the above devices, and the main storage medium is random access memory (SRAM).

[0092] The detailed steps are the same as those of the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion provided in Example 1, and will not be repeated here.

[0093] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0094] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion as described in Embodiment 1 of the present invention.

[0095] The detailed steps are the same as those of the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion provided in Example 1, and will not be repeated here.

[0096] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0097] A computer program product includes software code, wherein the program in the software code performs the steps of the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion as described in Embodiment 1 of the present invention.

[0098] The detailed steps are the same as those of the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion provided in Example 1, and will not be repeated here.

[0099] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages. For example, in one implementation, the methods and systems can be developed based on deep learning frameworks (such as TensorFlow, PyTorch, etc.) and using the Python language. Those skilled in the art will understand that other suitable programming languages ​​or tools can also be used for implementation without departing from the core ideas of the present invention.

[0100] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for classifying anti-membrane peptides based on transfer learning and multi-source feature fusion, characterized in that, include: Obtain the peptide sequence to be classified, input the peptide sequence to be classified into multiple pre-trained protein language models to extract semantic embedding features, and extract the physicochemical and statistical features of the peptide sequence to be classified. Then, fuse the semantic embedding features with the physicochemical and statistical features to obtain a multi-source feature vector. Based on the feature source, the multi-source feature vectors are grouped and feature-filtered to obtain a first feature subset. The grouped feature filtering employs a grouped feature sampling strategy: the multi-source feature vectors are divided into several feature groups according to their feature source; at each decision tree split node, the sampling quota for each group is determined by weighting the square root of the dimension of each feature group; candidate features are sampled from each group to ensure that each feature group has a competitive opportunity commensurate with its information content. The grouped feature filtering also employs a random pruning strategy: after optimal splitting at each internal node of each decision tree, a preset pruning probability is used to randomly determine whether to mark the node as a leaf node and terminate the split early. In the grouped feature sampling strategy, the sampling quota for each feature group is calculated using the following formula: in, For the total number of features, For the first Group feature dimensions For the first Group feature dimensions The total number of feature groups; The first feature subset is sorted by importance, and features that meet a preset threshold are selected for cross-product operation to construct interactive features. The interactive features are then filtered to obtain the second feature subset. The first feature subset and the second feature subset are concatenated and fused, and then input into the Transformer encoder for deep semantic encoding to obtain the hidden layer representation; The hidden layer representation is input into the first classifier and the second classifier respectively, and the first probability output by the first classifier and the second probability output by the second classifier are integrated by soft voting. Based on the integrated probability, it is determined whether the peptide sequence to be classified is an anti-biomembrane peptide.

2. The method according to claim 1, characterized in that, The multiple pre-trained protein language models include ESM-2, ProtT5, ProtBERT, and ProtT5-XL. The hidden states output by the encoders of each model are extracted and averaged along the sequence length dimension to obtain the sequence-level embedding vectors corresponding to each model. The physicochemical and statistical features include amino acid composition features, autocorrelation features, reduction-based dipeptide component features, and Moreau-Broto features. The four embedding vectors are concatenated in parallel with the physicochemical and statistical features to obtain the multi-source feature vector.

3. The method according to claim 1, characterized in that, The Transformer encoder sequentially includes a linear projection layer, a position encoding layer, and stacked multi-head self-attention layers and feedforward network layers. The linear projection layer is used to map the spliced ​​and fused features to a dense representation with a preset hidden dimension. The position encoding layer is used to inject virtual position information into the mapped features. The multi-head self-attention layer is used to capture the dependencies between features in parallel from multiple representation subspaces. The soft voting integration adopts a probability averaging method. When the average probability is greater than a preset classification threshold, the peptide sequence is determined to be an anti-biomembrane peptide.

4. The method according to claim 1, characterized in that, The Transformer encoder is trained using a two-stage transfer learning strategy: in the first stage, the Transformer encoder is pre-trained on the source domain antimicrobial peptide dataset using the cross-entropy loss function; in the second stage, the pre-trained Transformer encoder is fine-tuned on the target domain antibiofilm peptide dataset using the Focal loss function, which is used to reduce the loss weight of easily classified samples and increase the loss weight of difficult-to-classify samples.

5. A classification system for anti-biomembrane peptides based on transfer learning and multi-source feature fusion, characterized in that, include: The feature extraction and fusion module is used to obtain the peptide sequence to be classified, input the peptide sequence to be classified into multiple pre-trained protein language models to extract semantic embedding features, extract the physicochemical features and statistical features of the peptide sequence to be classified, and fuse the semantic embedding features with the physicochemical features and statistical features to obtain a multi-source feature vector. The feature filtering module is used to perform group-aware feature filtering on the multi-source feature vectors based on their feature sources to obtain a first feature subset. The group-aware feature filtering employs a group-aware feature sampling strategy: the multi-source feature vectors are divided into several feature groups according to their feature sources; at each decision tree splitting node, the sampling quota for each group is determined by weighting the square root of the dimension of each feature group; candidate features are sampled from each group to ensure that each feature group has a competitive opportunity commensurate with its information content. The group-aware feature filtering also employs a random pruning strategy: after each internal node of each decision tree completes its optimal split, a preset pruning probability is used to randomly determine whether to mark the node as a leaf node and terminate the split early. In the group-aware feature sampling strategy, the sampling quota for each feature group is calculated using the following formula: in, For the total number of features, For the first Group feature dimensions For the first Group feature dimensions The total number of feature groups; An interactive feature construction module is used to sort the first feature subset by importance, select features that meet a preset threshold for cross-product operation to construct interactive features, and filter the interactive features to obtain a second feature subset. The encoding module is used to concatenate and fuse the first feature subset and the second feature subset, input them into the Transformer encoder for deep semantic encoding, and obtain the hidden layer representation; The classification integration module is used to input the hidden layer representation into the first classifier and the second classifier respectively, and to perform soft voting integration on the first probability output by the first classifier and the second probability output by the second classifier, and to determine whether the peptide sequence to be classified is an anti-biomembrane peptide based on the integration probability.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the anti-biomembrane peptide classification method based on transfer learning and multi-source feature fusion as described in any one of claims 1 to 4.

8. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps of the anti-biomembrane peptide classification method according to any one of claims 1 to 4.

Images