Method for acquiring an antibacterial peptide discriminator

Abstract

The application discloses an anti-bacterial peptide discriminator acquisition method, relates to the field of anti-bacterial peptide identification, and comprises the following steps: constructing multiple positive sample training sets with different similarities through a disclosed anti-bacterial peptide database, acquiring a positive sample target training set through each positive sample training set based on a roberta-base model, acquiring multiple negative sample training sets through a UniProt protein database, acquiring a negative sample target training set based on the negative sample training sets B, C, D and E and the roberta-base model, acquiring an optimal word segmentation strategy through the multiple negative sample training sets, determining an optimal model subject by performing anti-bacterial peptide identification and effect prediction on each pre-training model on a preselected negative sample training set based on the optimal word segmentation strategy, training the optimal model subject through the positive sample target training set and the negative sample target training set to obtain a target model subject, constructing an anti-bacterial peptide discriminator through the optimal word segmentation strategy, the target model subject and an output layer, and realizing accurate and intelligent identification of anti-bacterial peptides through the anti-bacterial peptide discriminator.

Description

Technical Field

[0001] This invention relates to the field of antimicrobial peptide recognition, and more particularly to a method for obtaining an antimicrobial peptide discriminator. Background Technology

[0002] Antimicrobial peptides are molecules produced by multicellular organisms as their first line of defense against external invasions. They exist in a wide range of organisms, from prokaryotes to humans, including mammals, amphibians, insects, microorganisms, and plants. Antimicrobial peptides are typically positively charged, composed of 10 to 50 amino acid residues, and possess broad activity, capable of directly killing bacteria, fungi, viruses, and even cancer cells. Machine learning-based antimicrobial peptide identification integrates mathematical theory, computer technology, and bioinformatics. It extracts information such as sequence patterns, structural features, and biochemical properties of antimicrobial peptides through various methods, and then uses machine learning algorithms to learn from existing data and experience. Through reasoning, induction, or model fitting, the learned patterns are applied to determine whether an unknown sequence is an antimicrobial peptide. The development of machine learning has provided new tools for antimicrobial peptide identification. Researchers have begun to combine deep learning methods to identify antimicrobial peptides. Inspired by natural language processing, Hamid et al. used a skip-word model to train static amino acid vectors on a protein database and used a recurrent neural network to obtain sequence-level information, then used the modeling results to identify antimicrobial peptides. Furthermore, researchers hope that models trained on large-scale peptide data can learn the implicit biochemical properties, secondary and tertiary structures, and functional activities within the training dataset. These trained models can then be used, much like language models, for various downstream peptide-related tasks, such as peptide function prediction, peptide structure prediction, and the generation of specific functional peptides. Such pre-trained models are called protein language models. Language models trained on large-scale unlabeled text corpora are a fundamental and widely used approach in natural language processing, capable of learning general grammatical and semantic information, which proves highly effective in downstream tasks. Proteins share similar characteristics with text, and protein language models trained on unlabeled protein data based on the idea of ​​pre-trained language models hold great promise in protein research. This invention collects antimicrobial peptides and non-antimicrobial peptides from relevant protein databases to construct an antimicrobial peptide dataset, providing a new approach for the intelligent identification of antimicrobial peptides. Summary of the Invention

[0003] To achieve accurate and intelligent identification of antimicrobial peptides, this invention proposes a method for obtaining an antimicrobial peptide discriminator. The antimicrobial peptide discriminator includes an input word embedding layer, a target model body, and an output layer. The input word embedding layer is used to obtain a fixed-dimensional vector corresponding to the peptide sequence using an optimal word segmentation strategy. The model body is used to extract features from the fixed-dimensional vector and output a latent space feature vector. The output layer is used to perform linear transformation, normalization, and activation function processing on the latent space feature vector output by the model body, and uses the processed latent space feature vector to predict the probability that the input peptide sequence is an antimicrobial peptide. The latent space feature vector represents the latent feature space modeled on the input peptide sequence. The acquisition method includes:

[0004] A positive sample dataset was constructed using a publicly available antimicrobial peptide database. pos Through the positive sample dataset pos Obtain the positive sample test set S test-pos With positive sample training set S train-pos Through S train-pos Obtain the first, second, and third positive sample training sets; based on the Roberta-base model, utilize S... train-pos The first positive sample training set, the second positive sample training set, and the third positive sample training set are used to obtain the positive sample target training set; the similarity between the positive sample training sets is different.

[0005] Obtain the negative sample dataset from the UniProt protein database. neg via dataset neg Obtain the negative sample test set S test-neg The negative sample training sets are A, E, D, C, and B; the positive sample dataset represents a dataset composed of antimicrobial peptides, and the negative sample dataset represents a dataset composed of non-antimicrobial peptides; the negative sample target training set is obtained using the negative sample training sets B, C, D, and E based on the Robertta-base model.

[0006] Each preset word segmentation strategy was used to train a preset model on negative sample training sets A, B, and C in turn to verify the effectiveness of each preset word segmentation strategy and obtain the optimal word segmentation strategy.

[0007] Multiple pre-trained models are obtained, and based on the optimal word segmentation strategy, antimicrobial peptide identification and effect prediction are performed on each pre-trained model on a pre-selected negative sample training set to obtain the accuracy and F1-score of each pre-trained model. The optimal model body is determined by the accuracy and F1-score. The optimal model body is obtained by training the optimal model body on the positive sample target training set and the negative sample target training set.

[0008] Furthermore, the aforementioned via S train-pos Obtain the first positive sample training set, the second positive sample training set, and the third positive sample training set, specifically as follows:

[0009] Redundancy thresholds were set to 0.5, 0.7, and 0.9 respectively using CD-HIT for S. train-pos Redundancy removal was performed to obtain the first positive sample training set S with a similarity of less than 50%. train-pos-id50 The second positive sample training set S with a similarity of less than 70% train-pos-id70 The third positive sample training set S with a similarity of less than 90% train-pos-id90 .

[0010] Furthermore, the aforementioned through dataset neg Obtain the negative sample test set S test-neg The negative sample training sets E, D, C, and B are as follows:

[0011] From dataset neg According to dataset pos The length distribution of multiple non-antimicrobial peptide sequences was extracted as a non-antimicrobial peptide dataset A.

[0012] The non-antimicrobial peptide dataset A is randomly divided into a negative sample training set A and a negative sample test set S in an 8:2 ratio. test-neg ;

[0013] S test-neg From dataset neg After removing non-antimicrobial peptide sequences, multiple non-antimicrobial peptide sequences were randomly selected as the negative sample training set E.

[0014] Construct a negative sample training set D by drawing sequences from the negative sample training set E with a probability of 0.5;

[0015] Construct a negative sample training set C by drawing sequences from the negative sample training set D with a probability of 0.5;

[0016] Construct a negative sample training set B by drawing sequences from the negative sample training set C with a probability of 0.34.

[0017] Furthermore, the verification of the effect of the preset word segmentation strategy on the negative sample training sets A, B, and C specifically involves:

[0018] Using the positive sample training set S train-posAs the training set for antimicrobial peptides, negative sample training sets A, B, and C, as non-antimicrobial peptide training sets, were used to train bert-base-cased, bert-base-uncased, and Roberta-base peptides, respectively. This was achieved through S... test-pos With S test-neg Evaluations were conducted to obtain the Accuracy and F1-Score metrics for each model on negative sample training sets A, B, and C. The optimal word segmentation strategy was determined based on the Accuracy and F1-Score metrics.

[0019] Furthermore, the input word embedding layer is specifically used to treat each amino acid of the peptide sequence as an independent word according to the optimal word segmentation strategy, and then embed each amino acid to obtain a fixed-dimensional vector, which is input into the main body of the model. The vector includes the original information and positional relationship of each amino acid.

[0020] Furthermore, the specific method for obtaining the positive sample target training set is as follows:

[0021] respectively through S train-pos The first positive sample training set, the second positive sample training set, and the third positive sample training set are used to train the Robertta-base model to obtain the Accuracy and F1-Score corresponding to each training set. The positive sample target training set is obtained through the Accuracy and F1-Score.

[0022] Furthermore, the specific method for obtaining the negative sample target training set is as follows:

[0023] The Robertta-base model is trained using negative sample training sets B, C, D, and E respectively, and the accuracy and F1-score corresponding to each training set are obtained. The negative sample target training set is obtained through the accuracy and F1-score.

[0024] Furthermore, the pre-trained models include bert-base-cased, bert-large-cased, robotic-base, robotic-large, biobert-base-cased, ProtBERT, esm2-t12-35M-UR50D, and esm2-t30-150M-UR50D.

[0025] Furthermore, the preset models include: bert-base-cased, bert-base-uncased, and Roberta-base models; the publicly available antimicrobial peptide databases include: APD3, CAMP, and LAMP2.

[0026] Compared with the prior art, the present invention has at least the following beneficial effects:

[0027] (1) This invention constructs multiple positive sample training sets with different similarities using a publicly available antimicrobial peptide database, and obtains a positive sample target training set based on the Robertta-base model using each positive sample training set. It also obtains multiple negative sample training sets using the UniProt protein database, and obtains a negative sample target training set using negative sample training sets B, C, D, and E based on the Robertta-base model. The optimal word segmentation strategy is obtained through multiple negative sample training sets. Based on the optimal word segmentation strategy, antimicrobial peptide identification and effect prediction are performed on each pre-trained model on the pre-selected negative sample training set to obtain the accuracy and F1-Score corresponding to each pre-trained model. The optimal model body is determined by the accuracy and F1-Score. The optimal model body is trained using the positive sample target training set and the negative sample target training set to obtain the target model body. An antimicrobial peptide discriminator is constructed using the optimal word segmentation strategy, the target model body, and the output layer. The antimicrobial peptide discriminator achieves accurate and intelligent identification of antimicrobial peptides.

[0028] (2) To verify the negative impact of filtering similarity data on the antimicrobial peptide discriminator or its accuracy, this invention constructed three datasets S. train-pos-id50 S train-pos-id70 and S train-pos-id90 The values ​​represent the number of antimicrobial peptides with similarity less than 50%, less than 70%, and less than 90%, respectively, used to construct the positive sample training set. The Roberta-base model is trained using these three datasets to obtain the accuracy and F1-score for each training set. The positive sample target training set is obtained using the accuracy and F1-score. The main body of the antimicrobial peptide discriminator model is trained using the positive sample target training set, thereby improving the recognition accuracy of the antimicrobial peptide discriminator.

[0029] (3) To verify the effect of the positive and negative sample ratio on the antimicrobial peptide discriminator, this invention constructs a relative S train-pos Training sets B, C, and D of negative samples of different sizes are used to train a robotic-base model, obtaining the accuracy and F1-score for each training set. The target training set of negative samples is obtained through the accuracy and F1-score, and the main body of the antimicrobial peptide discriminator is trained through the target training set of negative samples, which further improves the recognition accuracy of the antimicrobial peptide discriminator. Attached Figure Description

[0030] Figure 1 This is a flowchart of a method for obtaining an antimicrobial peptide discriminator;

[0031] Figure 2 This is a structural diagram of an antimicrobial peptide discriminator. Detailed Implementation

[0032] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0033] Example 1

[0034] To achieve accurate and intelligent identification of antimicrobial peptides (AMPs), such as... Figure 1 As shown, this invention proposes a method for obtaining an antimicrobial peptide discriminator, which includes an input word embedding layer, a target model body, and an output layer. The input word embedding layer is used to obtain a fixed-dimensional vector corresponding to the peptide sequence using an optimal word segmentation strategy. The model body is used to extract features from the fixed-dimensional vector and output a latent space feature vector. The output layer is used to perform linear transformation, normalization, and activation function processing on the latent space feature vector output by the model body, and uses the processed latent space feature vector to predict the probability that the input peptide sequence is an antimicrobial peptide. The latent space feature vector represents the latent feature space for modeling the input peptide sequence.

[0035] It should be noted that the identification or discrimination of antimicrobial peptides is a binary classification task, that is, given an amino acid sequence, determining whether the sequence is an antimicrobial peptide or a non-antimicrobial peptide. Therefore, the formal definition of antimicrobial peptide identification is: given a sequence containing n characters: seq={w1,w2,...,w n}, where w i ∈W={A,C,D,E,F,G,H,I,K,L,M,N,P,Q,R,S,T,V,W,Y}, corresponding to alanine, cysteine, aspartic acid, glutamic acid, phenylalanine, glycine, histidine, isoleucine, lysine, leucine, methionine, aspartic acid, proline, glutamine, arginine, serine, threonine, valine, tryptophan, and tyrosine, respectively. The goal of antimicrobial peptide identification is to determine whether the amino acid sequence seq containing N characters is an antimicrobial peptide or a non-antimicrobial peptide. The acquisition method includes:

[0036] A positive sample dataset was constructed using a publicly available antimicrobial peptide database. pos Through the positive sample dataset pos Obtain the positive sample test set S test-pos With positive sample training set S train-pos (positive sample dataset) pos The dataset is randomly divided into a positive sample training set S at a ratio of 8:2. train-posand positive sample test set S test-pos ), through S train-pos Obtain the first, second, and third positive sample training sets; based on the Roberta-base model, utilize S... train-pos The first positive sample training set, the second positive sample training set, and the third positive sample training set are used to obtain the positive sample target training set; the similarity between the positive sample training sets is different.

[0037] The publicly available antimicrobial peptide database includes: APD3, CAMP, and LAMP2.

[0038] In this embodiment, antimicrobial peptide data are collected from three publicly available antimicrobial peptide databases: APD3, CAMP, and LAMP2, to construct a positive sample dataset. The antimicrobial peptide sequences collected from these three databases are fused to obtain a larger antimicrobial peptide dataset. After deduplication, 24,001 antimicrobial peptides are obtained. Based on the length characteristics of the antimicrobial peptides, these peptides are length-selected, retaining those with 5 to 300 amino acids. Finally, 23,966 antimicrobial peptides are obtained as the positive sample dataset of this invention. pos The S test-pos It is not used in the training of any model, but is reserved as a dataset for testing the model's performance.

[0039] The passage through S train-pos Obtain the first positive sample training set, the second positive sample training set, and the third positive sample training set, specifically as follows:

[0040] Using the similarity filtering tool provided by CD-HIT, the redundancy thresholds were set to 0.5, 0.7, and 0.9 for S. train-pos Redundancy removal was performed to obtain the first positive sample training set S with a similarity of less than 50%. train-pos-id50 The second positive sample training set S with a similarity of less than 70% train-pos-id70 The third positive sample training set S with a similarity of less than 90% train-pos-id90 .

[0041] In this embodiment, S train-pos-id50 S train-pos-id70 S train-pos-id90 The samples contained 2441, 4958, and 8465 positive samples, respectively.

[0042] The specific method for obtaining the positive sample target training set is as follows:

[0043] respectively through S train-posThe Robertta-base model is trained using the first, second, and third positive sample training sets to obtain the accuracy and F1-score for each training set. The positive sample target training set is then obtained using the accuracy and F1-score. The accuracy and F1-score for each training set are shown in Table 1 below.

[0044]

[0045] Overall, the performance of the antimicrobial peptide discriminator continuously improves as the sequence redundancy threshold (i.e., similarity threshold) increases. The most direct impact of changing the sequence redundancy threshold is a reduction in the number of sequences used for training. Reduced training data may lead to underfitting of the antimicrobial peptide discriminator. Simultaneously, sequences with high similarity allow the discriminator to learn more about the models, implicit biochemical properties, and secondary and tertiary structures of antimicrobial peptide sequences, enabling it to acquire more information beneficial for discrimination when encountering new sequences. Therefore, positive sample data should be retained as much as possible during antimicrobial peptide training. Thus, in this embodiment, S is selected. train-pos The target training set is the positive sample.

[0046] Obtain the negative sample dataset from the UniProt protein database. neg via dataset neg Obtain the negative sample test set S test-neg The negative sample training sets are A, E, D, C, and B; the positive sample dataset represents a dataset composed of antimicrobial peptides, and the negative sample dataset represents a dataset composed of non-antimicrobial peptides; the negative sample target training set is obtained using the negative sample training sets B, C, D, and E based on the Robertta-base model.

[0047] In this embodiment, sequences were obtained from UniProt to construct a non-antimicrobial peptide dataset. The method for obtaining negative samples is as follows: First, the filtering option "Subcellular Location" was set to cytoplasm, and sequences with any of the following annotations were removed: antimicrobial, antibiotic, antiviral, antifungal, effector, and excreted. Based on the sequences obtained from the above filtering, duplicates were removed from the AMP positive sample sequences, and only sequences with 5 to 300 amino acids were retained. Finally, 3,626,014 sequences were obtained as the negative sample dataset of this invention. neg .

[0048] The dataset neg Obtain the negative sample test set S test-negThe negative sample training sets E, D, C, and B are as follows:

[0049] From dataset neg According to dataset pos The length distribution of the non-antimicrobial peptides was used to extract multiple (23966) non-antimicrobial peptide sequences as non-antimicrobial peptide dataset A;

[0050] The non-antimicrobial peptide dataset A is randomly divided into a negative sample training set A and a negative sample test set S in an 8:2 ratio. test-neg ;

[0051] S test-neg From dataset neg After removing the non-antimicrobial peptide sequences, multiple (239660) non-antimicrobial peptide sequences were randomly selected as the negative sample training set E;

[0052] Construct a negative sample training set D by drawing sequences from the negative sample training set E with a probability of 0.5;

[0053] Construct a negative sample training set C by drawing sequences from the negative sample training set D with a probability of 0.5;

[0054] Construct a negative sample training set B by drawing sequences from the negative sample training set C with a probability of 0.34.

[0055] It should be noted that, in order to investigate the impact of the length distribution, quantity, and positive-to-negative sample ratio of the negative sample dataset on the antimicrobial peptide discriminator, this invention uses a dataset. neg Multiple negative sample training sets were constructed. First, starting from the dataset... neg According to dataset pos 23,966 non-antimicrobial peptide sequences were extracted from the length distribution to form a non-antimicrobial peptide dataset A. The non-antimicrobial peptide dataset A was then randomly divided into a negative sample training set A and a negative sample test set S in an 8:2 ratio. test-neg S test-neg It is not used in model training but is reserved as a dataset for testing model performance. To construct other datasets, S... test-neg From dataset negAfter removing non-antimicrobial peptide sequences, 239,660 non-antimicrobial peptide sequences were randomly selected as the negative sample training set E. Sequences were drawn from negative sample training set E with a probability of 0.5 to construct negative sample training set D. Sequences were drawn from negative sample training set D with a probability of 0.5 to construct negative sample training set C. Sequences were drawn from negative sample training set C with a probability of 0.34 to construct negative sample training set B. The length distributions of the five negative sample datasets are as follows: negative sample training set A is similar to the positive sample dataset in both length distribution and quantity; negative sample training set B is similar to the positive sample dataset in quantity but different in length distribution; negative sample training sets C, D, and E are different from the positive sample dataset in both quantity and length distribution. This invention constructs a relative S... train-pos Training sets B, C, and D of negative samples of different sizes are used to train a robotic-base model, obtaining the accuracy and F1-score for each training set. The target training set of negative samples is obtained through the accuracy and F1-score, and the main body of the antimicrobial peptide discriminator is trained through the target training set of negative samples, which further improves the recognition accuracy of the antimicrobial peptide discriminator.

[0056] The specific method for obtaining the negative sample target training set is as follows:

[0057] Roberta-base models were trained using negative sample training sets B, C, D, and E respectively, yielding the accuracy and F1-score for each training set. The negative sample target training set was then obtained using the accuracy and F1-score. The accuracy and F1-score for each training set are shown in Table 2 below.

[0058]

[0059] Experimental results show that increasing the number of negative samples does not consistently improve the accuracy of the antimicrobial peptide discriminator. When the ratio of positive to negative samples exceeds a certain threshold, a large increase in negative samples may lead to catastrophic amnesia in the model, resulting in a decrease in the accuracy of the antimicrobial peptide discriminator and an increase in both false positives and false negatives. Therefore, in this embodiment, the negative sample training set C is selected as the negative sample target training set.

[0060] Each preset word segmentation strategy was used to train a preset model on negative sample training sets A, B, and C in turn to verify the effectiveness of each preset word segmentation strategy and obtain the optimal word segmentation strategy (character segmentation).

[0061] The preset models include: bert-base-cased, bert-base-uncased, and Roberta-base models; in this embodiment, the preset word segmentation strategy includes character segmentation and a word segmenter.

[0062] The verification of the effect of the preset word segmentation strategy on the negative sample training sets A, B, and C specifically involves:

[0063] Using the positive sample training set S train-pos As the training set for antimicrobial peptides, negative sample training sets A, B, and C, as non-antimicrobial peptide training sets, were used to train bert-base-cased, bert-base-uncased, and Roberta-base peptides, respectively. This was achieved through S... test-pos With S test-neg Evaluations were conducted to obtain the Accuracy and F1-Score metrics for each model on negative sample training sets A, B, and C. The optimal word segmentation strategy was determined based on the Accuracy and F1-Score metrics.

[0064] The basic character set of peptide sequences is contained within the 26 letters of the English alphabet, allowing models pre-trained on protein data to process antimicrobial peptides, as well as models pre-trained on a general English corpus. Both sets of vocabularies typically contain single characters and words trained on the corpus using specific word segmentation algorithms. Character-based word segmentation treats each amino acid in the peptide sequence as an independent word, embedding each amino acid (including positional embedding) to obtain a fixed-dimensional vector, which is then input into the neural network. This preserves the original information and positional relationships of each amino acid, extracting structural and dependency information during the main model's modeling process. Using the pre-trained model's word segmenter, each peptide sequence is treated as an oligomer in the protein space, and then embedded, fully utilizing the overall information of the amino acid combination. To verify the effectiveness of different word segmentation strategies for antimicrobial peptide recognizers, this embodiment uses BERT and Roberta pre-trained models on negative sample datasets A, B, and C to validate the effects of different word segmentation strategies. Both BERT and Roberta offer pre-trained models with different parameter sets (Base and Large) to suit various application scenarios. BERT also provides cased and uncased pre-trained models; cased indicates case sensitivity (preserving capitalization information), while uncased indicates case insensitivity (not preserving capitalization information). Roberta recently offers a case-sensitive version. This invention uses bert-base-cased, bert-base-uncased, and Roberta-base as the main components of the discriminator model to verify the impact of different word segmentation strategies on discriminator performance. Taking the antimicrobial peptide sequence "FKRIVQRIKDALR" as an example, the word segmentation results obtained by character segmentation are [F,K,R,I,V,Q,R,I,K,D,A,L,R], the segmentation results by the bert-base-uncased segmenter are [fk,ri,v,q,rik,dal,r], the segmentation results by the bert-base-cased segmenter are [FK,RI,V,Q,RI,K,DA,LR], and the result by the Roberta segmenter is [FK,R,IV,Q,RI,K,D,AL,R]. During training, the positive sample training set is used as the antimicrobial peptide training set, and the negative sample training sets A, B, and C are used as the non-antimicrobial peptide training datasets. test-pos and S test-neg The performance of the discriminator was evaluated. To fully utilize the device's computing power and improve training efficiency, the batch size was set to 128 during the training of all discriminators. The accuracy and F1-Score of the model on the training set were used as the scoring criteria. The experimental results are shown in Table 3 below:

[0065]

[0066]

[0067] Experimental results show that character segmentation generally outperforms word segmentation in the antimicrobial peptide recognition task. Compared to word segmentation, character segmentation preserves information about each amino acid and their relative positions, providing more complete information for subsequent modeling. Word segmentation may treat multiple amino acids as a single word, resulting in less comprehensive information than a single amino acid. Therefore, the discriminator based on character segmentation performs better. The discriminator trained on negative sample training set A outperforms those trained on negative sample training sets B and C, indicating that the length distribution of the training dataset affects discriminator performance. Furthermore, after training on the training set, the discriminator can fully utilize the antimicrobial peptide sequence features and secondary / tertiary structural information learned during training when distinguishing unknown sequences, without being affected by the length distribution of the training set. Therefore, character segmentation is the optimal segmentation strategy in this invention.

[0068] The first part of a natural language processing model architecture is usually the embedding layer. The embedding layer transforms discrete amino acid symbols into a fixed-dimensional, continuously numerical vector to facilitate computer processing and mathematical calculations. From a biological perspective, different amino acids have different biochemical properties. However, computers can only recognize an amino acid as a symbol and cannot associate it with its specific sequence pattern, structural features, and biochemical properties. By embedding amino acids (including positional embedding) into a fixed-dimensional, continuous vector, the numerical values ​​of different dimensions represent numerical measures of the amino acid's characteristics in certain aspects. This allows computers to quantify the features of amino acids, and amino acids with similar properties are closer together in this metric space. Ultimately, amino acid-related calculations and optimizations can be performed numerically.

[0069] The input word embedding layer is specifically used to treat each amino acid of the peptide sequence as an independent word according to the optimal word segmentation strategy, and then embed each amino acid to obtain a fixed-dimensional vector, which is input into the main body of the model. The vector includes the original information and positional relationship of each amino acid.

[0070] Multiple pre-trained models are obtained, and based on the optimal word segmentation strategy, antimicrobial peptide identification and effect prediction are performed on each pre-trained model on a pre-selected negative sample training set (negative sample training set A) to obtain the accuracy and F1-score corresponding to each pre-trained model. The optimal model subject is determined by the accuracy and F1-score. The optimal model subject is trained by the positive sample target training set and the negative sample target training set to obtain the target model subject.

[0071] The pre-trained models include bert-base-cased, bert-large-cased, robotic-base, robotic-large, biobert-base-cased, ProtBERT, esm2-t12-35M-UR50D, and esm2-t30-150M-UR50D.

[0072] It should be noted that different batch sizes can affect model performance; batch sizes are typically chosen as powers of 2 or multiples of 8. This invention first uses a positive sample dataset and a negative sample training set A with batch sizes of 16, 32, 64, and 128 to train the TkrBertBaseCasedAMP model to obtain the optimal batch size for the dataset, aiming to improve model performance in S... test-pos and S test-neg The accuracy and F1-Score were used as evaluation criteria, and the experimental results are shown in Table 4 below:

[0073]

[0074] Experimental results show that the model trained with a batch size of 32 performs best on the test set on this dataset. Too small a batch size leads to increased training time, severe gradient oscillations, and poor convergence, while too large a batch size reduces the model's generalization ability and causes memory or GPU memory overflow. Therefore, the batch size for the main body of the model in this invention is set to 32.

[0075] This embodiment uses byte segmentation to validate the antimicrobial peptide recognition and prediction performance of different pre-trained models on the negative sample training set A. The batch size was set to 32 during the training of all models. The accuracy and F1-Score of the models on the training set were used as the scoring criteria. The experimental results are shown in Table 5 below:

[0076]

[0077]

[0078] When the batch size is 32, both bert-base-cased and Robert-base show improved performance. This phenomenon can be explained by the fact that training with smaller batches tends to converge to a flat minima, which varies only slightly within the minima's neighborhood, while larger batches converge to a sharper minima, which varies significantly. On the other hand, for models with the same architecture but different parameter counts, models with larger parameter counts are more difficult to train. Training models with large parameter counts requires significant computational power and memory, resulting in a limited batch size and longer forward and backward propagation times during training. After one round of training, ByteBertLargeCasedAMP and ByteRobertaLargeAMP performed worse than ByteBertBaseCasedAMP and ByteRobertaBaseAMP. Analysis of the training logs revealed underfitting in ByteBertLargeCasedAMP and ByteRobertaLargeAMP (Byte stands for byte segmentation, AMP stands for antimicrobial peptide). Therefore, this invention selects Roberta-base as the optimal model body. It should also be noted that, as... Figure 2 As shown, the main body of the optimal model in this embodiment includes, in sequence:

[0079] The model consists of an input word embedding layer, a main model, and an output layer. The main model includes position embedding, multi-head attention, summation and normalization (LayerNorm is used for normalization in this embodiment), and a fully connected neural network. The output layer includes a linear transformation, normalization, activation function, and Dropout connected in sequence. Multi-head attention is a mechanism proposed to enhance self-attention modeling capabilities to obtain different types of information; the activation function serves to delinearize; and Dropout is a mechanism to prevent overfitting.

[0080] This invention constructs multiple positive sample training sets with different similarities using a publicly available antimicrobial peptide database. Based on a Robertta-base model, it obtains a positive sample target training set from each positive sample training set. It also obtains multiple negative sample training sets from the UniProt protein database and uses negative sample training sets B, C, D, and E to obtain a negative sample target training set based on the Robertta-base model. The optimal word segmentation strategy is obtained from these multiple negative sample training sets. Based on this optimal strategy, antimicrobial peptide identification and performance prediction are performed on each pre-trained model on the pre-selected negative sample training sets to obtain the accuracy and F1-Score for each pre-trained model. The optimal model body is determined using the accuracy and F1-Score. The optimal model body is trained using the positive and negative sample target training sets to obtain the target model body. An antimicrobial peptide discriminator is constructed using the optimal word segmentation strategy, the target model body, and the output layer. This discriminator enables accurate and intelligent identification of antimicrobial peptides.

[0081] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0082] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0083] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0084] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A method for obtaining an antimicrobial peptide discriminator, characterized in that, The antimicrobial peptide discriminator includes an input word embedding layer, a target model body, and an output layer. The input word embedding layer is used to obtain a fixed-dimensional vector corresponding to the peptide sequence using the optimal word segmentation strategy. The model body is used to extract features from the fixed-dimensional vector and output a latent space feature vector. The output layer is used to perform linear transformation, normalization, and activation function processing on the latent space feature vector output by the model body, and to use the processed latent space feature vector to predict the probability that the input peptide sequence is an antimicrobial peptide. The latent space feature vector represents the latent feature space for modeling the input peptide sequence. The acquisition method includes: A positive sample dataset was constructed using a publicly available antimicrobial peptide database. pos Through the positive sample dataset pos Obtain the positive sample test set S test-pos With positive sample training set S train-pos Through S train-pos Obtain the first, second, and third positive sample training sets; based on the Roberta-base model, utilize S... train-pos The first positive sample training set, the second positive sample training set, and the third positive sample training set are used to obtain the positive sample target training set; the similarity between the positive sample training sets is different. Obtain the negative sample dataset from the UniProt protein database. neg Through dataset neg Obtain the negative sample test set S test-neg The negative sample training sets are A, E, D, C, and B; the positive sample dataset represents a dataset composed of antimicrobial peptides, and the negative sample dataset represents a dataset composed of non-antimicrobial peptides; the negative sample target training set is obtained using the negative sample training sets B, C, D, and E based on the Robertta-base model. Each preset word segmentation strategy was used to train a preset model on negative sample training sets A, B, and C in turn to verify the effectiveness of each preset word segmentation strategy and obtain the optimal word segmentation strategy. Multiple pre-trained models are obtained, and based on the optimal word segmentation strategy, antimicrobial peptide identification and effect prediction are performed on each pre-trained model on a pre-selected negative sample training set to obtain the accuracy and F1-score of each pre-trained model. The optimal model body is determined by the accuracy and F1-score. The optimal model body is obtained by training the optimal model body on the positive sample target training set and the negative sample target training set.

2. The method for obtaining an antimicrobial peptide discriminator according to claim 1, characterized in that, The passage through S train-pos Obtain the first positive sample training set, the second positive sample training set, and the third positive sample training set, specifically as follows: Redundancy thresholds were set to 0.5, 0.7, and 0.9 respectively using CD-HIT for S. train-po Redundancy removal was performed on sample s to obtain the first positive sample training set S with similarity less than 50%. train-pos-id50 The second positive sample training set S with a similarity of less than 70% train-pos-id70 The third positive sample training set S with a similarity of less than 90% train-pos-id90 .

3. The method for obtaining an antimicrobial peptide discriminator according to claim 1, characterized in that, The dataset neg Obtain the negative sample test set S test-neg Compared with negative sample training sets E, D, C, and B, specifically: From dataset neg According to dataset pos The length distribution of multiple non-antimicrobial peptide sequences was extracted as a non-antimicrobial peptide dataset A. The non-antimicrobial peptide dataset A is randomly divided into a negative sample training set A and a negative sample test set S in an 8:2 ratio. test-neg ; S test-neg From dataset neg After removing non-antimicrobial peptide sequences, multiple non-antimicrobial peptide sequences were randomly selected as the negative sample training set E. Construct a negative sample training set D by drawing sequences from the negative sample training set E with a probability of 0.5; Construct a negative sample training set C by drawing sequences from the negative sample training set D with a probability of 0.5; Construct a negative sample training set B by drawing sequences from the negative sample training set C with a probability of 0.

34.

4. The method for obtaining an antimicrobial peptide discriminator according to claim 1, characterized in that, The verification of the effect of the preset word segmentation strategy on the negative sample training sets A, B, and C specifically involves: Using the positive sample training set S train-pos As the training set for antimicrobial peptides, negative sample training sets A, B, and C, as non-antimicrobial peptide training sets, were used to train bert-base-cased, bert-base-uncased, and Roberta-base peptides, respectively. This was achieved through S... test-pos With S test-neg Evaluations were conducted to obtain the Accuracy and F1-Score metrics for each model on negative sample training sets A, B, and C. The optimal word segmentation strategy was determined based on the Accuracy and F1-Score metrics.

5. The method for obtaining an antimicrobial peptide discriminator according to claim 4, characterized in that, The input word embedding layer is specifically used to treat each amino acid of the peptide sequence as an independent word according to the optimal word segmentation strategy, and then embed each amino acid to obtain a fixed-dimensional vector, which is input into the main body of the model. The vector includes the original information and positional relationship of each amino acid.

6. The method for obtaining an antimicrobial peptide discriminator according to claim 1, characterized in that, The specific method for obtaining the positive sample target training set is as follows: respectively through S train-pos The first positive sample training set, the second positive sample training set, and the third positive sample training set are used to train the Robertta-base model to obtain the Accuracy and F1-Score corresponding to each training set. The positive sample target training set is obtained through the Accuracy and F1-Score.

7. The method for obtaining an antimicrobial peptide discriminator according to claim 1, characterized in that, The specific method for obtaining the negative sample target training set is as follows: The Roberta-base model is trained using negative sample training sets B, C, D, and E respectively, and the accuracy and F1-score corresponding to each training set are obtained. The negative sample target training set is obtained through the accuracy and F1-score.

8. The method for obtaining an antimicrobial peptide discriminator according to claim 1, characterized in that, The pre-trained models include bert-base-cased, bert-large-cased, robotic-base, robotic-large, biobert-base-cased, ProtBERT, esm2-t12-35M-UR50D, and esm2-t30-150M-UR50D.

9. The method for obtaining an antimicrobial peptide discriminator according to claim 1, characterized in that, The preset models include: bert-base-cased, bert-base-uncased, and Roberta-base models; the publicly available antimicrobial peptide databases include: APD3, CAMP, and LAMP2.

Images