Anti-cancer peptide semi-supervised prediction algorithm and system

Abstract

This invention relates to the interdisciplinary field of bioinformatics and artificial intelligence, and discloses a semi-supervised prediction algorithm for anticancer peptides. The algorithm includes the following steps: first, positive samples are constructed by obtaining experimentally validated anticancer peptide sequences from public databases, and negative samples are constructed by sampling natural peptide sequences, forming an initial labeled set; conserved and non-conserved regions of the anticancer peptides are identified, and the non-conserved regions are masked and fed into a Transformer encoder for masked language modeling pre-training; structural regularization constraints are applied to the attention matrix based on the peptide's three-dimensional structure contact map; pseudo-variant sequences are generated based on the model's predicted amino acid substitution probabilities to construct an enhanced dataset. Global embeddings of peptide sequences are extracted using a protein language model, a teacher-student decoder architecture is built, the training set is expanded using high-confidence pseudo-samples, iterative training is performed, and teacher parameters are updated using an exponential moving average. Finally, the anticancer activity prediction of the target peptide is completed through the student decoder, effectively suppressing the accumulation of label noise.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics and artificial intelligence, and in particular to a semi-supervised prediction algorithm and system for anticancer peptides. Background Technology

[0002] Cancer is a complex disease threatening human health, and determining the correct treatment plan for each patient is a formidable challenge. Given the difficulties in developing small molecule drugs and the increasing homogenization of large molecule drugs such as antibodies, the search for novel and effective anticancer therapies has become imperative. Anticancer peptides (ACPs) are a class of short amino acid sequences with the potential to selectively target cancer cells while minimizing damage to normal healthy cells. Due to their low toxicity and high specificity, they have become a promising class of therapeutic molecules. However, identifying anticancer peptides experimentally is time-consuming, laborious, and has low throughput, making virtual screening using intelligent models an attractive alternative. Although recent progress has been made, the number of available anticancer peptide samples in existing databases remains limited, which may lead to overfitting of predictive models to labeled information and insufficient generalization ability. Therefore, how to effectively expand the scale of training data without compromising the core function of anticancer peptides, given a limited number of labeled samples, thereby improving the generalization ability of predictive models to unknown peptides, is a pressing technical problem that needs to be solved in this field.

[0003] In recent years, numerous computational methods for predicting anticancer peptides have been proposed, primarily focusing on two aspects: peptide sequence representation learning and activity prediction. Regarding representation learning, early methods relied on hand-constructed features to build the input space, such as amino acid composition, physicochemical properties, and k-mer frequency matrices. Currently, natural language processing (NLP) techniques are increasingly popular, treating amino acids as words and peptide sequences as sentences to extract potential embedding representations. For example, Lv et al. (2021) used a pre-trained protein language model integrated with an LSTM network to derive sequence embeddings for anticancer peptides. Guntuboina et al. (2023) introduced a self-supervised BERT model, utilizing a masked residue reconstruction task to enhance the representation learning of anticancer peptide sequences. Li et al. (2023) employed a Transformer encoder to capture contextual semantic embeddings from anticancer peptide samples. In contrast, research on structural representation learning for peptides is relatively limited due to the scarcity of three-dimensional conformational data. A few studies (2024) have attempted to use tools such as AlphaFold to generate structures from sequences, and then transform the structures into residue-level or atomic-level graphs, thereby encoding spatial and positional information through geometric deep learning. In terms of activity prediction, traditional methods (2019) employ machine learning algorithms (such as random forests, XGBoost, and ensemble learning (2022)) to determine whether a peptide possesses anticancer activity. Thanks to the end-to-end capabilities of deep learning, recent advances have led to the design of various neural networks to simultaneously optimize the representation and prediction of anticancer peptides. For example, Rao et al. (2020) reformulated anticancer peptide prediction as a graph classification task and applied graph convolutional networks to classify peptides based on residue co-occurrence patterns. Srivathsan et al. (2024) used a contrastive learning strategy to align sequence modalities and structural modalities to a shared space, and then used a BERT decoder to predict the properties of anticancer peptides. Zhang et al. (2024) deployed a convolutional neural network incorporating a flow attention mechanism to capture potential patterns in anticancer peptide sequences.

[0004] Despite these advances, anticancer peptide modeling remains constrained by the scarcity of labeled data. For example, benchmark datasets such as CancerPPD contain only about 1,000 known anticancer peptide sequences, far from sufficient to train robust models. Obtaining structural data is even more challenging due to the difficulty of peptide crystallization. While some studies have attempted to enhance structural representations by incorporating AlphaFold-predicted 3D conformations, these strategies do not provide additional supervisory labels. Meanwhile, several works have considered expanding the feature space by sampling non-anticancer peptides from antimicrobial or antiviral peptides, but this does not amplify the positive signal specific to anticancer peptides. Against this backdrop, data augmentation offers a promising avenue for enriching supervisory information by generating pseudo-samples, an approach widely explored in both small and large molecule domains. However, unlike image or text data, conventional augmentation methods based on random perturbations (such as sequence mutations or structural noise additions) are highly unlikely to generate positive anticancer peptide instances, as even small perturbations can lead to significant functional alterations.

[0005] While existing technologies have made some progress in sequence representation learning and activity prediction of anticancer peptides (ACPs), the core bottleneck lies in the extreme scarcity of labeled data: conventional data augmentation methods (such as random mutation or noise addition) are likely to destroy activity due to the sensitivity of anticancer peptide function and cannot effectively generate positive pseudo-samples; at the same time, existing semi-supervised strategies are difficult to improve the model's ability to predict unknown peptides under limited labeling conditions. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a semi-supervised prediction algorithm and system for anticancer peptides, which can effectively suppress the accumulation of label noise during the iteration process.

[0007] This invention provides a semi-supervised prediction algorithm for anticancer peptides, comprising the following steps: Experimentally validated anticancer peptide sequences were obtained from publicly available peptide databases and preprocessed to form a positive sample dataset. A negative sample dataset was constructed by sampling from a natural peptide sequence library, and together with the positive sample dataset, they formed the initial labeled dataset. Each sequence in the positive sample dataset is divided into conservative and non-conservative regions, and the conservative regions are locked. In non-conservative regions, amino acid sites are selected for masking operations, and the masked sequence is input into the Transformer encoder for masked language modeling pre-training. During pre-training, a structure regularization constraint is applied to the attention matrix of the Transformer encoder using the contact map obtained from the three-dimensional structure prediction of the anticancer peptide. After pre-training, the Transformer encoder is used to predict the substitution probability of each amino acid at the mask position. The amino acid with the highest probability is selected to replace the wild-type residue, generating pseudo-variant sequences as pseudo samples, thus forming an augmented dataset. Based on the augmented dataset and the initial labeled dataset, a teacher-student semi-supervised learning framework is used to train the student decoder, and the trained student decoder is used to predict the anticancer activity of the test peptide sequence.

[0008] Specifically, the conservative region includes the global conservative region and the local conservative region; Global conservative region CR a It is composed of five amino acid residues: arginine, lysine, histidine, leucine, and glycine. Mutations of residues in this region are prohibited during data augmentation. The construction of locally conserved regions (CRᵦ) includes: calculating the Levenstein distance between the target sequence and other sequences in the positive sample dataset, selecting the k sequences with the smallest distance as the nearest neighbor set; performing the longest common subsequence alignment between the target sequence and each sequence in the nearest neighbor set, and calculating the conservation score of each site on the target sequence; marking sites with conservation scores higher than a preset threshold δ as locally conserved regions (CRᵦ), where residues are prohibited from mutation during data augmentation. The non-conservative region is CR a The area outside of CRᵦ.

[0009] Specifically, global conservative region CR a The five amino acids were identified as follows: arginine, lysine, and histidine are related to the cationic properties of anticancer peptides; leucine is related to the hydrophobic properties of anticancer peptides; glycine is related to the structural flexibility of anticancer peptides; and the above five amino acids have a statistically significant high frequency of occurrence in experimentally verified anticancer peptide sequences.

[0010] Specifically, structural regularization constraints are applied to the attention matrix of the Transformer encoder, including: The three-dimensional structure of each original anticancer peptide sequence was predicted using a structure prediction tool, and a binary contact diagram matrix was constructed based on the Euclidean distance between Cα atoms in the main chain. Extract the inter-amino acid attention weight matrix calculated by the Transformer encoder during the masked language modeling pre-training process; The attention weight matrix is ​​mapped using an activation function and then compared with the binary contact map matrix to calculate the similarity loss. The similarity loss and the cross-entropy loss from the masked language modeling pre-training are weighted and summed to form the enhancement loss, which is used to jointly optimize the Transformer encoder.

[0011] Specifically, the semi-supervised learning framework trains the student decoder, including: A pre-trained protein language model is used to encode and aggregate features of samples in the initial labeled dataset and pseudo samples in the augmented dataset to obtain fixed-dimensional peptide sequence global embedding vectors, which serve as inputs to the teacher decoder and student decoder. Construct teacher decoders and student decoders with identical structures; The teacher decoder is first pre-trained on the initial labeled dataset. After convergence, it predicts the pseudo samples in the augmented dataset and retains only the pseudo samples whose prediction confidence is higher than the positive threshold or lower than the negative threshold and their corresponding pseudo labels. These are then merged into the initial labeled dataset to form an extended training set. The student decoder is trained on an extended training set and the teacher decoder parameters are updated using an exponential moving average strategy. The pseudo-label screening, student model training and teacher parameter update steps are repeated until the model converges. The peptide sequence to be tested is then feature-encoded and aggregated by a pre-trained protein language model and input into the student decoder to output the predicted probability of its anti-cancer activity.

[0012] Specifically, the pre-trained protein language model is an ESM-2 model pre-trained on a large-scale protein sequence database; feature encoding and aggregation are performed by average pooling of the residue-level features output by ESM-2.

[0013] Specifically, the teacher decoder's rules for predicting and filtering pseudo-samples are as follows: For each pseudo-variant sequence, the teacher decoder outputs a prediction confidence score; If the prediction confidence is greater than or equal to the preset positive confidence threshold, the pseudo-sample is retained and assigned a positive pseudo-label. If the predicted score is less than or equal to the preset negative confidence threshold, the pseudo-sample is retained and assigned a negative pseudo-label. If the predicted score is between the positive confidence threshold and the negative confidence threshold, it is considered a low-confidence sample and is discarded.

[0014] Specifically, the exponential moving average strategy updates the teacher decoder parameters by weighting and summing the current student decoder parameters with the previous teacher decoder parameters to obtain the current teacher decoder parameters.

[0015] A semi-supervised prediction system for anticancer peptides, comprising: The dataset construction and preprocessing module is used to obtain experimentally validated anticancer peptide sequences from public peptide databases and preprocess them into positive sample datasets, and to sample and construct negative sample datasets from natural peptide sequence libraries to form the initial labeled dataset. The bio-inspired data augmentation module is used to divide each sequence in the positive sample dataset into conservative and non-conservative regions and lock the conservative regions. Mask pre-training is performed in the non-conservative regions, and pseudo-variant sequences are generated as pseudo-samples under the structural contact graph regularization constraint to form the augmented dataset. The semi-supervised learning framework module is used to train a student decoder based on the augmented dataset and the initial labeled dataset, and then use the trained student decoder to predict the anti-cancer activity of the peptide sequence to be tested.

[0016] The technical solution provided by the embodiments of the present invention has the following advantages compared with the prior art: The embodiments of the present invention divide the positive sample anticancer peptide sequence into conserved and non-conserved regions and lock the conserved regions, strictly limiting the masking enhancement operation to be performed within the non-conserved regions. This avoids interference with the core functional motifs that determine anticancer activity at the source, ensuring that the generated pseudo-variants maintain their biological activity basis from the beginning of sequence amplification. Secondly, in the above-mentioned masking language modeling pre-training process, a contact map obtained from the three-dimensional structure prediction of the anticancer peptide is introduced to apply structural regularization constraints to the attention matrix of the Transformer encoder, forcing… When recovering masked amino acids, the model not only follows the semantic coherence of the sequence context, but also satisfies the physical prior with higher attention weight between spatially adjacent residues. This ensures that the predicted high-probability substitution amino acids have additional structural consistency on the basis of sequence rationality, significantly improving the probability that pseudo-variants maintain their original folding conformation and functional properties. Finally, the pre-trained protein language model is used to process the samples in the initial labeled dataset, effectively filtering out noise that may be introduced by poor pseudo-samples and overcoming the model's over-reliance on limited original labeled information. This fundamentally achieves a significant improvement in the generalization ability to predict the anticancer activity of unknown peptides. Attached Figure Description

[0017] Figure 1 A flowchart of an anticancer peptide prediction algorithm based on a bio-inspired data augmentation and semi-supervised learning framework provided for an embodiment of the present invention; Figure 2 A schematic diagram of a model for an anticancer peptide prediction algorithm based on a bio-inspired data augmentation and semi-supervised learning framework, provided for an embodiment of the present invention; Figure 3 This is a comparison chart of the embodiments of the present invention with other comparative methods on four binary classification evaluation indicators: Area under the receiver operating characteristic curve (top left), Area under the precision-recall curve (top right), Precision (bottom left), and Balanced score (bottom right). Figure 4The graph shows the prediction performance of the embodiments of the present invention and the baseline method in different sequence length ranges, with the area under the receiver operating characteristic curve (top left), the area under the precision-recall curve (top right), precision (bottom left), and balance score (bottom right). Figure 5 The images show the t-SNE embedding visualization distribution of the embodiments of the present invention and the baseline method. The top left image shows the t-SNE embedding visualization distribution of the embodiments of the present invention, and the remaining sub-images show the t-SNE embedding visualization distribution of the comparison methods. Figure 6 The following is a statistical analysis graph of the embodiment of the present invention on an external independent validation set. The left graph (a) shows the prediction results for different sequence interval lengths, where each point represents the prediction probability of a sample. The right graph (b) shows the multi-sequence alignment results for the interval [41, ∞]. Detailed Implementation

[0018] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] The present invention will be described below through several specific embodiments. To keep the following description of the embodiments clear and concise, detailed descriptions of known functions and components may be omitted. When any component of an embodiment of the present invention appears in more than one drawing, the component may be represented by the same reference numerals in each drawing.

[0021] Figure 1 This is a flowchart illustrating an anticancer peptide prediction algorithm based on a bio-inspired data augmentation and semi-supervised learning framework, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a model for an anticancer peptide prediction algorithm based on a bio-inspired data augmentation and semi-supervised learning framework, provided in an embodiment of the present invention. Figure 3 This is a comparison chart of the embodiments of the present invention with other comparative methods on four binary classification evaluation indicators: Area under the receiver operating characteristic curve (top left), Area under the precision-recall curve (top right), Precision (bottom left), and Balanced Score (bottom right). Figure 4 The graph shows the prediction performance of the embodiments of the present invention and the baseline method in different sequence length ranges, with the area under the receiver operating characteristic curve (top left), the area under the precision-recall curve (top right), precision (bottom left), and balance score (bottom right). Figure 5 This is a visualization of the t-SNE embeddings of the embodiments of the present invention and the baseline method. The top left corner shows the visualization of the t-SNE embeddings of the embodiments of the present invention, and the remaining sub-figures show the visualization of the t-SNE embeddings of the comparison methods. Figure 6 The following is a statistical analysis graph of the embodiment of the present invention on an external independent validation set. The left graph (a) shows the prediction results for different sequence interval lengths, where each point represents the prediction probability of a sample. The right graph (b) shows the multi-sequence alignment results for the interval [41, ∞].

[0022] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a semi-supervised prediction algorithm for anticancer peptides, including the following steps: Dataset construction and preprocessing: experimentally validated anticancer peptide sequences are obtained from publicly available large-scale peptide databases, and after preprocessing, they are used as positive sample datasets; negative sample datasets are constructed by randomly sampling from human natural peptide sequence libraries, thereby forming an initial labeled dataset.

[0023] Bioinspired data augmentation: In terms of amino acid composition, anticancer peptides are typically rich in positively charged lysine, arginine, and histidine, as well as the highly hydrophobic leucine. These residues endow them with cationic properties and hydrophobicity, while random coil anticancer peptides are rich in glycine. Based on this, a globally conserved region composed of arginine (R), lysine (K), histidine (H), leucine (L), and glycine (G) was constructed. Based on Levenstein distance, nearest neighbors are identified, and conserved sites are selected after longest common subsequence alignment to construct locally conserved regions. ;Will and The regions outside the protected area are considered non-conservative regions. A masking operation is performed on the amino acid sites in the non-conservative regions, and the results are input into the Transformer encoder for pre-training to capture semantic knowledge. During the pre-training process, the attention matrix is ​​regularized using a contact map derived from the structure of the anticancer peptide to capture the inherent structural patterns in the conformation of the anticancer peptide. After training, the substitution probability distribution of each amino acid at the mask position is output by the prediction head, and the amino acid with the highest prediction probability is selected to replace the wild-type residues to generate pseudo-variant sequences, thereby obtaining the augmented dataset.

[0024] A semi-supervised learning framework is employed: a pre-trained protein language model is used to encode features from real samples and augmented pseudo-samples. The encoded residue-level features are aggregated to obtain fixed-dimensional global embedding vectors for peptide sequences. A teacher decoder and a student decoder with identical structures are constructed. The teacher decoder is pre-trained on the initial labeled dataset and, after convergence, is used to generate pseudo-labels for the augmented pseudo-samples. Only pseudo-samples with prediction confidence above a positive threshold or below a negative threshold, along with their pseudo-labels, are retained and merged into the initial labeled dataset to form an extended training set. The student decoder is trained on the extended training set and updates the teacher decoder parameters using an exponential moving average strategy. After convergence, it is used to predict the anticancer activity of the target peptide sequence. By using the above method, the mutation operation in the data augmentation process is strictly limited to non-conservative regions, thereby preserving the core functional framework of anticancer peptides while expanding the training samples.

[0025] By introducing a contact graph derived from AlphaFold2 predictions as an attention regularization constraint during mask pre-training, the Transformer encoder is forced to focus on spatially adjacent amino acids when predicting mask residues. This makes the generated pseudo-variant sequences consistent with the spatial folding tendency of natural anticancer peptides, thus improving the rationality of augmenting the dataset.

[0026] By adopting a teacher-student semi-supervised learning framework, a teacher model pre-trained on real labeled data is used to screen for high-confidence pseudo-labels on augmented pseudo-samples, retaining only confidence samples to expand the training set. This effectively suppresses the accumulation of label noise during the iteration process and significantly improves the generalization ability of the model under limited real data conditions.

[0027] In the semi-supervised iteration process, an exponential moving average strategy is used to update the teacher decoder parameters, so that the parameter update of the teacher model lags behind that of the student model, thereby maintaining the stability of the supervision signal during the pseudo-label generation process and further improving the prediction accuracy of the model on unknown oligopeptides and external independent datasets.

[0028] Furthermore, the positive sample dataset refers to experimentally validated anticancer peptides and proteins obtained from the publicly available large-scale peptide database CancerPPD2.0, serving as positive samples. The preprocessing operation involves preprocessing the positive samples to remove non-natural amino acid sequences beyond the standard 20 amino acids, as well as sequences longer than 80 amino acids. After filtering, the positive samples contain 5044 unique anticancer peptide sequences. The negative sample dataset refers to randomly selecting negative samples from the Human PeptideAtlas database, which contains over one million peptide sequences, using a negative sample sampling strategy. The final dataset has a positive-to-negative sample ratio of 1:5. The labeled peptide dataset is represented as follows:

[0029] in The total number of samples, For the input peptide sequence, This indicates whether the sample possesses anticancer activity. The goal is to learn a mapping function that outputs a probability value. The higher the value, the greater the likelihood of anti-cancer activity.

[0030] Furthermore, bio-inspired data augmentation refers to a reliable data augmentation strategy designed considering the biological complexity of peptides. This strategy ensures that key amino acid sites remain unchanged during mutation sampling, generating pseudo-samples that more closely resemble real anticancer peptides. These key amino acid sites are considered conserved regions, while the remaining sites on the amino acid sequence are considered non-conserved regions.

[0031] Furthermore, globally conservative regions Inspired by the biological knowledge behind anticancer peptide libraries, this study focuses on amino acids with high frequency and important functions in existing anticancer peptides, ultimately identifying a conserved region composed of arginine (R), lysine (K), histidine (H), leucine (L), and glycine (G). , represented as: in, Indicates the first A sequence of anticancer peptides, Indicates the first position in the sequence The types of amino acid residues at each site; Represents the global set of key amino acid residues, with values ​​ranging from 1 to 2. These correspond to arginine, lysine, histidine, leucine, and glycine, respectively.

[0032] For sequence The set of site indexes marked as globally conserved regions in the data is used to prevent mutations of residues within these regions during subsequent data augmentation to maintain the core functional backbone of the anticancer peptide.

[0033] Furthermore, in step (2), the locally conservative region The construction includes the following steps: Nearest neighbor sequence screening: Calculating target anticancer peptide sequences Choose the sequence with the smallest Levenstein distance from other sequences in the initial labeled dataset. The sequences form a nearest neighbor set. The calculation formula is: in, Indicates the sequence Convert to sequence Minimum number of single-character editing operations required This indicates that the data is truncated after sorting by distance in ascending order. The process involves analyzing individual samples to identify the target anticancer peptide sequence. In the positive sample space The nearest neighbor.

[0034] Site conservation statistics: targeting anticancer peptide sequences Set up with neighbors For each sequence in the target sequence, the longest common subsequence alignment is performed. The frequency of each residue at each point in the target sequence during the alignment is counted, and the conservation score at each point is calculated. : in, Represents a sequence and The longest common subsequence contains The set of site indexes This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0035] Conservative site markers: Conservatism scores Greater than the preset threshold The sites were marked as locally conserved sites and classified as... The region extraction formula is: in, A preset consensus threshold is used to control the stringency of screening for locally conserved sites. In summary, for target anticancer peptide sequences... Local conservative region , represented as:

[0036] Represents a sequence The set of site indexes marked as locally conserved regions is used to prevent mutations of residues within these regions during subsequent data augmentation, in order to maintain the local consistency between the target anticancer peptide sequence and its neighboring samples in the sequence feature space.

[0037] Furthermore, the specific steps for semantic knowledge capture are as follows: Constructing the masked sequence: For the sequence after the conserved region is locked, 2 to 3 amino acid sites are randomly selected in the non-conserved region to perform a masking operation, and the residues at the corresponding positions are replaced with mask markers; the masked sequence is converted into an input feature vector by one-hot encoding and sent to the Transformer encoder for forward propagation.

[0038] Contextual semantic reconstruction: The Transformer encoder obtains three vectors after linearly transforming each amino acid tag in the input feature vector: (Query) (key) and (Value); any two amino acids in the sequence context and Attention scores between Calculate using the following formula: in, For the first A query vector for each amino acid. For the first The bond vector of an amino acid. The dimension of the bond vector is given; the Softmax function normalizes the similarity scores into a probability distribution, thus yielding an amino acid interaction matrix. , Given the sequence length; the attention weights are summed with the value vector to obtain the context-aware embedding representation for each mask position. : in, For the first A vector of values ​​for each amino acid.

[0039] Masked residue prediction: embedding representation The data is fed into a multilayer perceptron and outputs the substitution probability distribution of 20 amino acids at the mask positions. : in, This indicates the sequence after the mask position has been excluded. The mask represents the original amino acid residues at the mask location; the multilayer perceptron consists of a fully connected layer and a softmax output layer, completing the mapping from the embedding representation to the amino acid category probability; the training of the mask language modeling target adopts the cross-entropy loss between the predicted amino acid type and the real amino acid at the mask location. : in, For the training sample set, This represents the expectation of random mask positions; by minimizing the cross-entropy loss, the model is forced to learn the contextual dependencies between amino acid residues in the sequence, thereby capturing the semantic knowledge of the anticancer peptide in the embedding representation.

[0040] Furthermore, the specific steps for capturing structural patterns are as follows: Structure contact diagram acquisition: The AlphaFold2 tool was used to predict the three-dimensional structure of each original anticancer peptide sequence and extract the main chain. The three-dimensional spatial coordinates of an atom; for any two amino acid residues in the sequence and Calculate its The Euclidean distance between atoms; when the Euclidean distance is less than a preset contact threshold, the residue pair is determined to have a contact relationship in space and marked as 1; otherwise, it is marked as 0. This constructs a binary contact map matrix. , The sequence length is 0. The preset contact threshold is 8 Å, which corresponds to the maximum spatial distance between residues in the protein structure where there is a direct interaction.

[0041] Attention matrix acquisition: Extracting the calculated attention matrix during the context semantic reconstruction process. , where matrix elements Indicates the first in the sequence The amino acid for the first Attention weights for each amino acid.

[0042] Structural regularization joint optimization: Integrating the attention matrix Mapped to the [0,1] interval by the Sigmoid function, and then compared with the binary contact map matrix. Element-wise mean squared error is calculated to construct the similarity loss between the contact map and the attention matrix. : in, It is the Sigmoid activation function. Indicates the training sample set Calculate the expected value; minimize the similarity loss. To guide amino acid probability The prediction. The structure regularization loss. With masked language modeling loss The weighted summation results in the following enhanced loss. :

[0043] in, It is a hyperparameter used to balance the contribution of the regularization term. By minimizing the enhancement loss, the model is forced to pay attention to spatially adjacent amino acids when reconstructing the masked sequence markers, thereby generating substitutions consistent with the peptide structure organization.

[0044] Furthermore, the specific steps for constructing the augmented dataset are as follows: After the Transformer encoder is trained, the masked sequence is input into the model for forward propagation; the multilayer perceptron in the prediction head outputs the substitution probability distribution of 20 amino acids at each masked position; for each selected masked site, the amino acid with the highest probability value in the substitution probability distribution is selected to replace the wild-type residue; each original mother sequence is generated. strip pseudovariant ( The mutation sites are randomly selected from non-conserved regions, and the number of mutation sites in each pseudovariant sequence is controlled to be between 2 and 3. The pseudovariant sequences generated from the initial positive sample sequences through the above steps are collected to form the augmented dataset. It is used for downstream anticancer peptide prediction tasks.

[0045] Furthermore, the specific steps of the semi-supervised learning framework are as follows: Feature Encoding and Embedding: A Transformer-based protein language model, ESM-2, pre-trained on the UniRef database, is used to encode and aggregate features from real samples in the initial labeled dataset and pseudo samples in the augmented dataset, mapping each amino acid residue to a fixed-dimensional feature vector. Average pooling is applied to all amino acid sites to obtain the global embedding vector. : in, This represents a pre-trained protein language model. This refers to the amino acid residue arrangement of the input peptide sequence. The sequence length is given; based on the extracted embedding vectors, a teacher-student framework is constructed, where both are implemented by a multilayer perceptron decoder with the same architecture.

[0046] Teacher Decoder Pre-training: Teacher Decoder First, in the initial labeled dataset Training is performed on the above, using binary cross-entropy loss. Optimize model parameters: in, Teacher decoder The output logit value, For real labels, The number of samples in the initial labeled dataset; Pseudo-label generation and filtering: utilizing a pre-trained converged teacher decoder Augmented dataset The pseudo-samples in the dataset are used for prediction. Only pseudo-samples with prediction confidence levels higher than the positive threshold or lower than the negative threshold, along with their pseudo-labels, are retained. The selection rules are as follows: in and These represent the confidence thresholds for positive and negative results, respectively. The selected high-confidence pseudo-samples and their pseudo-labels are then merged with the initial labeled dataset to form an expanded training set. :

[0047] Student Decoder Training and Parameter Update: Student Decoder Expanding the training set The model is trained on top of the real samples and pseudo-labeled samples simultaneously. Instead of updating the teacher model's parameters through backpropagation at each step, an exponential moving average strategy is used to transfer the student decoder parameters to the teacher decoder. The update formula is as follows: in The coefficient of the exponential moving average. For the first Step student parameters, For the first The teacher's parameters are updated step by step; through an exponential moving average strategy, the teacher decoder's parameter updates lag behind the student decoder, providing a more stable supervision signal for pseudo-label generation, while the student model gradually improves its performance through training on an expanded dataset; the process of pseudo-label generation and screening, student decoder training and exponential moving average parameter updates is repeated until the model converges; the peptide sequence to be tested is input into the finally trained student decoder, which outputs the predicted probability that it has anti-cancer activity.

[0048] This application also includes an anticancer peptide prediction system based on a bio-inspired data augmentation and semi-supervised learning framework, using an anticancer peptide prediction algorithm based on a bio-inspired data augmentation and semi-supervised learning framework, including: The dataset construction and preprocessing module obtains experimentally validated anticancer peptide sequences from publicly available large-scale peptide databases, preprocesses them, and uses them as positive sample datasets; it also constructs negative sample datasets by randomly sampling from human natural peptide sequence libraries, thus forming the initial labeled dataset.

[0049] The bio-inspired data augmentation module constructs a globally conserved region composed of arginine (R), lysine (K), histidine (H), leucine (L), and glycine (G). Based on Levenstein distance, nearest neighbors are identified, and conserved sites are selected after longest common subsequence alignment to construct locally conserved regions. ;Will and The regions outside the protected area are considered non-conservative regions. A masking operation is performed on the amino acid sites in the non-conservative regions, and the results are input into the Transformer encoder for pre-training to capture semantic knowledge. During the pre-training process, the attention matrix is ​​regularized using a contact map derived from the structure of the anticancer peptide to capture the inherent structural patterns in the conformation of the anticancer peptide. After training, the substitution probability distribution of each amino acid at the mask position is output by the prediction head, and the amino acid with the highest prediction probability is selected to replace the wild-type residues to generate pseudo-variant sequences, thereby obtaining the augmented dataset.

[0050] The semi-supervised learning framework module uses a pre-trained protein language model to encode features of real samples and augmented pseudo-samples. The encoded residue-level features are aggregated to obtain a fixed-dimensional global embedding vector for the peptide sequence. A teacher decoder and a student decoder with identical structures are then constructed. The teacher decoder is pre-trained on the initial labeled dataset and, after convergence, is used to generate pseudo-labels for the augmented pseudo-samples. Only pseudo-samples with prediction confidence above a positive threshold or below a negative threshold, along with their pseudo-labels, are retained and merged into the initial labeled dataset to form an extended training set. The student decoder is trained on the extended training set and updates the teacher decoder parameters using an exponential moving average strategy. After iteration until convergence, it is used to predict the anticancer activity of the peptide sequence to be tested.

[0051] Experimental Parameter Settings and Performance Evaluation: To verify the effectiveness of this invention and evaluate model performance, the dataset was first randomly divided into a training set (80%) and a test set (20%), with no overlap. Subsequently, a five-fold cross-validation scheme was implemented on the training set, dividing it into five equal subsets. In each iteration, four subsets were used for model training, and the remaining subset was used for validation, ensuring that each subset was used as a validation set at least once. The model that achieves the best validation performance in each fold will then be evaluated on the test set. For binary classification prediction of anticancer peptides, performance was measured by four standard metrics: area under the curve (AUC), precision-recall AUC, accuracy, and F1 score.

[0052] AugACP was compared and evaluated with six established baseline methods for predicting anticancer peptides (AACFlow, ACP-MS, ACPred, ACP-EPC, Multi-Peptide, and ACPScanner, covering both sequence-based and structure-based methods). Experiments were conducted using Python and PyTorch. Model training employed the Adam optimizer with a learning rate of 0.0001, a batch size of 128, and 150 training epochs. The hyperparameter settings are shown in Table 1.

[0053] Table 1. Experimental hyperparameter settings: The performance comparison results of this invention with various baseline methods on the test set are attached. Figure 3 As shown, observations reveal that AugACP outperforms all baseline methods in AUC, AUPR, and F1 score, with mean values ​​of 0.9616, 0.9084, and 0.8475, respectively. While AugACP maintains only a slight lead in AUC, its superior performance in F1 score and AUPR is particularly noteworthy—given the imbalanced nature of our dataset (five times the number of negative samples compared to positive samples). A relatively high variance was also observed in five-fold cross-validation, potentially attributed to bias introduced by pseudo-labels used in augmented mean teacher learning. Furthermore, ACP-MS and ACPred significantly underperformed other baseline methods. Both rely on traditional machine learning algorithms, which may be less effective than deep learning models in capturing complex peptide representations. Meanwhile, baseline methods relying solely on sequence data showed little difference in performance compared to methods incorporating structural information. This is likely because the peptide structures used were predicted by AlphaFold2 rather than derived from experimental determination, and anticancer peptide samples typically contain fewer amino acids, thus diminishing the contribution of structural information. In summary, these results highlight AugACP's generalization ability and demonstrate its powerful efficacy in identifying anticancer peptides that extend beyond the training distribution.

[0054] To evaluate the necessity of each component in the model framework, a series of ablation comparison experiments were conducted between AugACP and its variants. The specific settings for each variant are as follows:

[0055] AugACP without data augmentation (w / o DA): Removes the data augmentation process and trains the predictor using only the original labeled dataset.

[0056] AugACP non-conservative region a (w / o) ): During data augmentation, conserved region locking consisting of five amino acids (R, K, H, L, G) is removed.

[0057] AugACP non-conservative region b (w / o) ): During data augmentation, conservative region locks identified by comparison of nearest neighbor and longest common subsequence are removed.

[0058] AugACP without semantic knowledge (w / o SK): Replaces semantic knowledge-guided enhancements with random site mutations, no longer distinguishing between conserved and non-conserved regions.

[0059] AugACP Contactless Graph (w / o CM): Removes the regularization loss constructed from the contact graph in mask language pre-training for data augmentation. .

[0060] AugACP without ESM-2 features (w / o ESM): ESM-2 representation extraction is omitted and replaced with BLOSUM62 matrix encoding.

[0061] AugACP without Mean Teacher update (w / o MTU): Excludes the exponential moving average update in the Mean Teacher paradigm, and the student decoder and teacher decoder no longer share parameters.

[0062] The ablation results of AugACP and its variants on the test set are listed in Table 2.

[0063] Table 2 Ablation Experiment Results: Observations revealed that removing the data augmentation step (w / o DA) significantly degraded the performance of AugACP across all metrics. This highlights the importance of additional training signals—even when using pseudo-labeled samples. Secondly, the variant AugACP (w / o ) and (w / o The results show that both semantically aware site selection strategies are crucial for helping the model select higher-quality augmentation samples. AugACP (w / o SK) results demonstrate that semantically knowledge-guided augmentation is more effective than random mutation. When this component was removed, the accuracy and F1 score decreased from 0.9510 and 0.8475 to 0.9314 and 0.8178, respectively. Furthermore, the performance degradation of AugACP (w / o ESM) confirms that extracting peptide hidden features through an information-rich protein language model can improve the accuracy of anticancer peptide prediction. In addition, the Mean Teacher paradigm module also plays a key role in our framework. The performance of AugACP (w / o MTU) confirms that a significant performance degradation occurred when the exponential moving average update was removed and the student and teacher decoders no longer shared parameters. In summary, each module in AugACP is indispensable for effective anticancer peptide prediction, and removing any component leads to a decrease in prediction performance.

[0064] To investigate the length specificity of the model, i.e., the impact of sequence length on predictive performance, this invention divided the anticancer peptide samples into five intervals (0–10, 11–20, 21–30, 31–40, and >40) based on sequence length, and tested AugACP and other methods accordingly. The results are attached. Figure 4 As shown.

[0065] The results show that AUC, AUPR, accuracy, and F1 score all increase with increasing sequence length. One possible explanation for this trend is that longer peptides exhibit clearer patterns and are easier for the model to recognize; while shorter peptides contain less sequence information, making modeling more difficult. Notably, the figure also shows that the prediction accuracy is significantly lower for oligopeptides with 1–10 amino acids; while for peptides with 11–40 amino acids, the prediction performance of all methods is relatively consistent. In particular, the AACFlow method, whose convolutional neural network seems to be weak in capturing amino acid correlations in short peptides, resulting in its lowest performance. We also analyzed the number of samples in each length range and found no significant linear relationship between the number of samples and prediction performance. These results indicate that sequence length is an important factor affecting the prediction of anticancer peptides, rather than simply relying on the expansion of the number of samples.

[0066] Given that the effectiveness of deep learning models largely depends on the quality of representation learning, we further explored its impact on prediction performance. This invention visualizes the peptide embedding representations learned during training (ACPred is excluded because it is based on traditional support vector machines rather than deep learning). Taking the test set as an example, the embedding representations are projected into three-dimensional space using t-SNE, and the results are shown in the appendix. Figure 5 As shown.

[0067] Visualization results demonstrate that several methods can separate the majority of non-anticancer peptides from the minority of anticancer peptides in the embedding space. Notably, AugACP and AACFlow exhibit clearer and more structured distribution patterns, indicating their ability to capture more complex patterns in peptide data. To quantitatively assess this observation, the silhouette coefficient, a metric used to measure intra-cluster compactness and inter-cluster separation, was further calculated. The results show that AugACP achieves superior discrimination ability compared to baseline methods (profile coefficient = 0.281). In summary, this finding confirms the effectiveness of AugACP in learning information-rich representations and highlights its advantage in distinguishing between anticancer and non-anticancer peptides.

[0068] To test whether AugACP can identify novel anticancer peptides beyond the distribution range of the original data, this invention conducted external validation using two independent databases (SATPdb and APD6). These two databases contain hundreds of therapeutic peptides with antitumor activity, none of which are included in the dataset described above. From these resources, 156 external anticancer peptide samples were collected, covering different sequence length ranges: 53 were between 10 and 20 amino acids, 91 were between 21 and 40 amino acids, and 12 were longer than 40 amino acids. The prediction results for the three length ranges are attached. Figure 6 As shown in (a).

[0069] Since the external dataset contains only positive samples, a higher prediction probability indicates better accuracy. Overall, AugACP performed well on the external dataset, with an average prediction accuracy of 0.9559, but a small number of samples still received relatively low confidence predictions.

[0070] This invention further visualized samples longer than 40 amino acids and found no obvious common patterns in the sequence alignment results, such as... Figure 6 As shown in (b).

[0071] Observations indicate that AugACP possesses strong generalization ability, capable of recognizing anticancer peptides with diverse amino acid combinations. Nevertheless, the AugACP model failed to predict two anticancer peptide cases: sequence 1 ("PQAAIVASG", named emericellipsin A) and sequence 2 ("PVIRHYYGEYS TAALKHPPEDIAEETH", named Boma-CATH). Previous in vitro studies showed that emericellipsin A exhibited selective cytotoxicity against HepG2 and HeLa cell lines, while Boma-CATH inhibited the proliferation, migration, and invasion of A375 cells through a dual antitumor mechanism (pyroptosis induction and metastasis inhibition). These cases reveal that some anticancer peptides remain challenging for models, especially when their sequence patterns differ significantly from those commonly found in training data. However, overall, these case studies demonstrate the potential of AugACP in recognizing novel anticancer peptide candidates.

[0072] The above inventions are merely a few specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A semi-supervised prediction algorithm for anticancer peptides, characterized in that, Includes the following steps: Experimentally validated anticancer peptide sequences were obtained from publicly available peptide databases and used as positive sample datasets after preprocessing. A negative sample dataset was constructed by sampling from a natural peptide sequence library, and together with the positive sample dataset, it formed the initial labeled dataset; Each sequence in the positive sample dataset is divided into a conservative region and a non-conservative region, and the conservative region is locked. Amino acid sites are selected in the non-conservative region to perform masking operations, and the masked sequence is input into the Transformer encoder for masking language modeling pre-training; During pre-training, a structural regularization constraint is applied to the attention matrix of the Transformer encoder using a contact map obtained from the three-dimensional structure prediction of the anticancer peptide. After pre-training, the Transformer encoder is used to predict the substitution probability of each amino acid at the mask position. The amino acid with the highest probability is selected to replace the wild-type residue, generating a pseudo-variant sequence as a pseudo sample, thereby forming an augmented dataset. Based on the augmented dataset and the initial labeled dataset, a teacher-student semi-supervised learning framework is used to train the student decoder, and the trained student decoder is used to predict the anticancer activity of the test peptide sequence.

2. The semi-supervised prediction algorithm for anticancer peptides as described in claim 1, characterized in that, The conservative region includes a global conservative region and a local conservative region; The global conservative region CR a It is composed of five amino acid residues: arginine, lysine, histidine, leucine, and glycine. Mutations of residues in this region are prohibited during data augmentation. The construction of the locally conserved region CRᵦ includes: calculating the Levenstein distance between the target sequence and other sequences in the positive sample dataset, selecting the k sequences with the smallest distance as the nearest neighbor set; performing the longest common subsequence alignment between the target sequence and each sequence in the nearest neighbor set, and calculating the conservation score of each site on the target sequence; marking sites with conservation scores higher than a preset threshold δ as locally conserved regions CRᵦ, where residues are prohibited from mutation during data augmentation. The non-conservative region is CR. a The area outside of CRᵦ.

3. The semi-supervised prediction algorithm for anticancer peptides as described in claim 2, characterized in that, The global conservative region CR a The five amino acids were identified as follows: arginine, lysine, and histidine are related to the cationic properties of anticancer peptides; leucine is related to the hydrophobic properties of anticancer peptides; glycine is related to the structural flexibility of anticancer peptides; and the above five amino acids have a statistically significant high frequency of occurrence in experimentally verified anticancer peptide sequences.

4. The semi-supervised prediction algorithm for anticancer peptides as described in claim 1, characterized in that, The application of structural regularization constraints to the attention matrix of the Transformer encoder includes: The three-dimensional structure of each original anticancer peptide sequence was predicted using a structure prediction tool, and a binary contact diagram matrix was constructed based on the Euclidean distance between Cα atoms in the main chain. Extract the inter-amino acid attention weight matrix calculated by the Transformer encoder during the masked language modeling pre-training process; The attention weight matrix is ​​mapped using an activation function, and then the similarity loss is calculated between it and the binary contact map matrix. The similarity loss is weighted and summed with the cross-entropy loss from the masked language modeling pre-training to form an enhancement loss, which is used to jointly optimize the Transformer encoder.

5. The semi-supervised prediction algorithm for anticancer peptides as described in claim 1, characterized in that, The semi-supervised learning framework trains the student decoder, including: A pre-trained protein language model is used to encode and aggregate features of samples in the initial labeled dataset and pseudo samples in the augmented dataset to obtain a fixed-dimensional peptide sequence global embedding vector, which serves as the input to the teacher decoder and the student decoder. Construct teacher decoders and student decoders with identical structures; The teacher decoder is first pre-trained on the initial labeled dataset. After convergence, it predicts the pseudo samples in the augmented dataset and retains only the pseudo samples whose prediction confidence is higher than the positive threshold or lower than the negative threshold and their corresponding pseudo labels. These pseudo samples are then merged into the initial labeled dataset to form an extended training set. The student decoder is trained on the extended training set and the teacher decoder parameters are updated using an exponential moving average strategy. The pseudo-label screening, student model training and teacher parameter update steps are repeated until the model converges. After the peptide sequence to be tested is feature-encoded and aggregated by the pre-trained protein language model, it is input into the student decoder and the predicted probability of its anti-cancer activity is output.

6. The semi-supervised prediction algorithm for anticancer peptides as described in claim 5, characterized in that, The pre-trained protein language model is an ESM-2 model pre-trained on a large-scale protein sequence database; the feature encoding and aggregation are average pooling of the residue-level features output by ESM-2.

7. The semi-supervised prediction algorithm for anticancer peptides as described in claim 5, characterized in that, The teacher decoder's rules for predicting and filtering pseudo-samples are as follows: For each pseudo-variant sequence, the teacher decoder outputs a prediction confidence score; If the prediction confidence is greater than or equal to the preset positive confidence threshold, the pseudo-sample is retained and assigned a positive pseudo-label. If the predicted score is less than or equal to the preset negative confidence threshold, the pseudo-sample is retained and assigned a negative pseudo-label. If the predicted score is between the positive confidence threshold and the negative confidence threshold, it is considered a low-confidence sample and is discarded.

8. The semi-supervised prediction algorithm for anticancer peptides as described in claim 5, characterized in that, The exponential moving average strategy updates the teacher decoder parameters by weighting and summing the current student decoder parameters with the previous teacher decoder parameters to obtain the current teacher decoder parameters.

9. A semi-supervised prediction system for anticancer peptides, characterized in that, include: The dataset construction and preprocessing module is used to obtain experimentally validated anticancer peptide sequences from public peptide databases and preprocess them into positive sample datasets, and to sample and construct negative sample datasets from natural peptide sequence libraries to form the initial labeled dataset. The bio-inspired data augmentation module is used to divide each sequence in the positive sample dataset into conservative and non-conservative regions and lock the conservative regions. Mask pre-training is performed in the non-conservative regions, and pseudo-variant sequences are generated as pseudo-samples under the structural contact graph regularization constraint to form the augmented dataset. A semi-supervised learning framework module is used to train a student decoder based on the augmented dataset and the initial labeled dataset, and to use the trained student decoder to predict the anticancer activity of the peptide sequence to be tested.

Images