Multi-modal deep learning-based anticancer peptide comprehensive prediction and screening method
By fusing multiple features using a multimodal deep learning framework, the problem of low prediction accuracy in anticancer peptide screening was solved, achieving efficient and accurate anticancer peptide screening and cancer type-specific prediction, while reducing experimental costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HENAN UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2025-12-04
- Publication Date
- 2026-05-01
AI Technical Summary
Existing methods for screening anticancer peptides rely on single-modality features, resulting in low prediction accuracy, failure to fully integrate multi-source information, and lack of efficient information mining capabilities.
A multimodal deep learning framework is adopted, which combines binary classification model, multi-label classification model and activity regression model. By integrating basic sequence features, AAlndex physicochemical property features and ESM2 language model features, multimodal features are constructed to achieve high-accuracy screening of anticancer peptides and cancer type-specific prediction.
It significantly improved the screening efficiency and prediction accuracy of anticancer peptides, enabling accurate identification of anticancer peptides, cancer-specific classification, and quantitative analysis of drug efficacy, while reducing experimental verification costs.
Smart Images
Figure CN121963841A_ABST
Abstract
Description
A Comprehensive Prediction and Screening Method for Anticancer Peptides Based on Multimodal Deep Learning Technical Field
[0001] This invention relates to the field of anticancer peptide screening technology, specifically a comprehensive prediction and screening method for anticancer peptides based on multimodal deep learning. Background Technology
[0002] Anticancer peptides (ACPs) are a class of biologically active polypeptide molecules that can selectively kill tumor cells through various mechanisms (such as cell membrane disruption and apoptosis induction), making them important candidates for anticancer drug development. However, experimental screening of ACPs faces challenges such as high cost, long cycle time, and low throughput, and efficiently identifying highly active ACPs from massive peptide sequences remains difficult. To overcome the limitations of experimental screening, deep learning and machine learning are increasingly being used for ACP activity prediction to improve screening efficiency, such as support vector machines, random forests, and XGBoost. However, these models rely on basic sequence features such as amino acid composition and physicochemical properties for binary classification prediction. With the development of deep learning and protein language models, especially large-scale pre-trained models such as ESM, ProtTrans, and PeptideBERT, high-dimensional feature representations can be automatically learned from peptide sequences, providing new ideas for ACP prediction.
[0003] However, the above prediction methods rely only on single-modal features and fail to fully integrate multi-source information, resulting in insufficient high-level semantic mining and thus low prediction accuracy. Summary of the Invention
[0004] To address the issue of low prediction accuracy in existing technologies, this invention provides a comprehensive prediction and screening method for anticancer peptides based on multimodal deep learning. This method boasts high accuracy, strong generalization, and good interpretability, significantly improving the computational screening efficiency of candidate anticancer peptides and reducing experimental verification costs.
[0005] To achieve the above objectives, the specific solution adopted in this invention is as follows: a comprehensive prediction and screening method for anticancer peptides based on multimodal deep learning. This involves obtaining a set of sequences to be tested, inputting the set into a pre-constructed multimodal deep learning framework, which includes a binary classification model, a multi-label classification model, and an activity regression model. One model or a combination of the three models is selected to process the set of sequences to be tested. When the binary classification model is used, the set of sequences is screened based on a confidence threshold to obtain the confidence screening results. When the multi-label classification model is used, the type prediction results for each sequence in the set are obtained. When the activity regression model is used, multiple first activity values are obtained. When the set of sequences is processed sequentially by the binary classification model, the multi-label classification model, and the activity regression model, multiple second activity values are obtained, and candidate anticancer peptides are obtained based on these second activity values.
[0006] As an optimization of the above-mentioned method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning, the method for obtaining the test sequence set is as follows: analyze the bioindicators of existing sequence sets in the anticancer peptide database, obtain the indicator results, determine the sequence length, and generate the test sequence set based on the indicator results and the sequence length.
[0007] As another optimization scheme for the above-mentioned comprehensive prediction and screening method of anticancer peptides based on multimodal deep learning: the biological indicators include AAC, DPC and TPC; the indicator results are: AAC includes K, L, R, A and G, DPC includes KK, KL and RR, and TPC includes KKL, LKK and AKK.
[0008] As another optimization scheme for the above-mentioned multimodal deep learning-based comprehensive prediction and screening method for anticancer peptides, the sequence length is 18.
[0009] As an alternative optimization of the aforementioned multimodal deep learning-based comprehensive prediction and screening method for anticancer peptides, the construction method of the multimodal deep learning framework is as follows:
[0010] Obtain multimodal features based on existing databases;
[0011] Multimodal features were used to train binary classification models, multi-label classification models, and liveness regression models to obtain a multimodal deep learning framework.
[0012] As an optimization of the aforementioned multimodal deep learning-based comprehensive prediction and screening method for anticancer peptides, the following method for obtaining multimodal features is employed:
[0013] Multiple sequences were selected from an existing database and preprocessed to obtain positive sample data, negative sample data, multi-label data, and IC. 50data;
[0014] For positive sample data, negative sample data, multi-label data and IC 50 The data underwent feature extraction to obtain basic sequence features, AAlndex physicochemical property features, and ESM2 language model features;
[0015] Multimodal features are obtained by fusing basic sequence features, AAlndex physicochemical property features, and ESM2 language model features.
[0016] As an alternative optimization of the aforementioned method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning, the binary classification model is a fully connected neural network (DNN), which includes an input layer, a hidden layer, and an output layer.
[0017] As an alternative optimization of the aforementioned method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning, a multi-label classification model is obtained through a weighted ensemble of random forest, XGBoost, linear support vector machine, and logistic regression.
[0018] As an alternative optimization scheme for the above-mentioned multimodal deep learning-based comprehensive prediction and screening method for anticancer peptides, the activity regression model adopts the CatBoost algorithm.
[0019] As another optimization scheme for the above-mentioned comprehensive prediction and screening method of anticancer peptides based on multimodal deep learning, the process of the test sequence set being processed by a binary classification model, a multi-label classification model and an activity regression model is as follows: The test sequence set is input into the binary classification model, and the test sequence set is screened according to the confidence threshold to obtain a high-confidence sequence set with a confidence greater than the threshold. The obtained high-confidence sequence set is then input into the multi-label classification model, which predicts the type of each sequence in the high-confidence sequence set to obtain a specific sequence set. The specific sequence set is then input into the activity regression model to obtain the corresponding second activity value.
[0020] Compared with the prior art, the present invention has the following beneficial effects:
[0021] This invention provides a comprehensive prediction and screening method for anticancer peptides based on multimodal deep learning. The test sequence set is input into a multimodal deep learning framework for processing. First, a binary classification model processes the test sequence set, filtering it based on confidence thresholds to obtain a high-confidence sequence set. This high-confidence sequence set is then input into a multi-label classification model for identification, determining the cancer type specificity of each sequence in the high-confidence sequence set, and selecting anticancer peptides for the desired cancer type to form a specific sequence set. The specific sequence set is then input into an activity regression model to predict the activity of each sequence, obtaining the corresponding second activity value. The second activity values are then sorted in ascending order to obtain candidate peptides with the highest anticancer potential. This invention enables accurate identification of anticancer peptides, allows for specific classification for different cancer types, and achieves quantitative analysis of drug efficacy through activity regression, forming an end-to-end unified prediction process.
[0022] In this invention, multimodal features are formed by fusing basic sequence features, AAlndex physicochemical property features, and ESM2 language model features. This achieves information complementarity at the feature level, preserving the basic compositional information of the sample sequence while introducing deep semantic features and physicochemical property distribution features, further improving the expressive power and prediction accuracy of the multimodal deep learning framework. Specifically, by forming a unified high-dimensional representation vector through early feature-level fusion, information complementarity and deep feature expression are achieved, significantly enhancing the model's representational power and prediction accuracy.
[0023] This invention introduces a SHAP-based interpretable analysis method to quantitatively evaluate the contribution of the model at different modalities and feature levels. By calculating the marginal contribution value of each feature in the predicted output, the relative importance of basic sequence features, AAindex physicochemical property features, and ESM2 language model features in anticancer peptide sequence recognition, cancer type classification, and activity regression can be clarified. This process not only realizes the visualization and interpretability of the multimodal deep learning framework but also significantly enhances its scientific rigor and credibility.
[0024] This invention can be widely applied to the virtual screening and experimental validation of peptide drugs for different cancer types. Specifically, it is applicable to the screening and activity evaluation of anticancer peptides for various cancer types, including breast cancer, lung cancer, liver cancer, colorectal cancer, gastric cancer, prostate cancer, and leukemia. Through multi-task collaborative prediction and quantitative activity evaluation, potential candidate peptide molecules can be rapidly screened from large-scale sequence libraries, significantly improving drug development efficiency and providing intelligent technical support for the discovery and optimization of peptide anticancer drugs. Attached Figure Description
[0025] Figure 1 is a schematic diagram of the process of the present invention.
[0026] Figure 2 is a schematic diagram of anticancer peptide data analysis. 2A represents binary classification datasets, multi-label classification data, and IC. 50 A bar chart showing the number of regression data samples; 2B represents binary classification datasets, multi-label classification data, and IC (inductively coupled plasma) datasets. 50 Box plot of peptide sequence length (number of amino acids) distribution in regression data; 2C is a bar chart of the frequency of seven cancer type labels in multi-label data; 2D is IC 50 Histogram of the distribution of values after Z-score normalization.
[0027] Figure 3 is a multimodal feature contribution analysis diagram, showing the contributions of 3A as basic sequence features, AAindex physicochemical property features, and ESM2 language model features in binary classification, multi-label classification, and IC. 50 The relative contribution ratios in the three tasks are shown in the bar chart; 3B is a heatmap of the importance scores of multimodal features; 3C is a radar chart of the contribution of the three types of features under different tasks.
[0028] Figure 4 is a schematic diagram of the prediction performance of the binary classification model; 4A is a curve showing the change of the accuracy of the binary classification model on the validation set with the number of training rounds; 4B is a curve showing the change of the loss value of the binary classification model on the validation set with the number of training rounds; 4C is a confusion matrix diagram of the binary classification model on the test set; 4D is a bar chart of the comprehensive evaluation indicators of the binary classification model on the test set, including Kappa, MCC, F1, specificity (Spe), sensitivity (Sen), and accuracy (ACC).
[0029] Figure 5 is a schematic diagram of the prediction performance of the multi-label classification model; 5A is a heatmap of the four indicators of precision, recall, F1 score and accuracy for seven cancer types; 5B is a bar chart of the prediction accuracy for different cancer types; 5C is a graph showing the changes of F1, precision, recall and macro-average precision (RA) with the number of folds under five-fold cross-validation; 5D is a graph showing the changes of exact match rate (EM-Rate) and Hamming loss (H-Loss) with the number of folds under five-fold cross-validation.
[0030] Figure 6 is a schematic diagram illustrating the performance of the active regression model. 6A represents the IC50 value. 50 Predicted value vs. actual experimental measurement (Log(IC)) 50 6B is a scatter plot of the correlation between the predicted residuals and the predicted values; 6C is a bar chart of the mean squared error (MSE) of different cancer-specific regression models on the independent test set; 6D is a bar chart comparing the MSE of the active regression model on the training set (cross-validation) and the test set. Detailed Implementation
[0031] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. Parts not described or disclosed in detail in the following embodiments of the present invention should be understood as prior art known or should be known by those skilled in the art.
[0032] A comprehensive prediction and screening method for anticancer peptides based on multimodal deep learning includes the following steps:
[0033] Obtain the sequence set to be tested. Specifically, analyze the biomarkers of existing sequence sets in the anticancer peptide database to obtain the indicator results. The existing sequence set is the CancerPPD database. From this database, a set of sequences with recorded IC values is selected. 50 The anticancer peptide sequences were identified, forming an existing sequence set. Biomarkers included AAC (amino acid composition), DPC (dipeptide composition), and TPC (tripeptide composition). The results were as follows: AAC included K (lysine), L (leucine), R (arginine), A (alanine), and G (glycine); DPC included KK (two consecutive positively charged lysines), KL (one lysine and one leucine), and RR (two arginines); TPC included KKL (two lysines followed by one leucine), LKK (one leucine followed by two lysines), and AKK (one alanine followed by two lysines).
[0034] The sequence length was determined to be 18. Based on the indicator results and the sequence length, multiple test sequences were generated to form a test sequence set. The generation principle was to generate 18 peptides, that is, peptide sequences composed of 18 amino acids, with 18 positions filled with 20 kinds of amino acids. These 20 amino acids were randomly arranged and combined at the 18 positions. Combining the above indicator results (anticancer peptides appear frequently at each position, some two-amino acid combinations appear frequently, and three-amino acid combinations appear frequently), a batch of sequences was generated to form a test sequence set.
[0035] The test sequence set is input into a pre-constructed multimodal deep learning framework, which includes a binary classification model, a multi-label classification model, and an activity regression model. One model or a combination of the three models is selected to process the test sequence set as needed. In this invention, the multimodal deep learning framework is a three-level prediction system. The first layer is a binary classification model to achieve binary classification identification of anticancer peptide sequences; the second layer is a multi-label classification model to achieve identification of different cancer types; and the third layer is an activity regression model to predict the activity value (IC) of the anticancer peptide sequences. 50 value).
[0036] The construction method of a multimodal deep learning framework is as follows:
[0037] Multimodal features were obtained based on existing databases; specifically, multiple sequences were selected from existing databases (CancerPPD database and UniProt database) and sequentially filtered, deduplicated, and normalized to obtain positive sample data, negative sample data, multi-label data, and IC data. 50 The data, including positive and negative sample data, is collectively referred to as a binary classification dataset. This includes positive sample data, negative sample data, multi-label data, and IC (Individual Classification) data. 50 Feature extraction was performed on the data to obtain basic sequence features, AAlndex physicochemical property features, and ESM2 language model features. Basic sequence features included amino acid composition, dipeptide frequency, and physicochemical property distribution, totaling approximately 436 dimensions. AAlndex physicochemical property features, with 566 indicators expanded to approximately 5094 dimensions through 9 statistical descriptions, covered biochemical parameters such as hydrophobicity, charge, and polarity. The ESM2 language model features comprised 1280 dimensions. Multimodal features were obtained by fusing these features. Based on the concept of multimodal deep learning, the multi-source heterogeneous information of the anticancer peptide sequences was fused. Multimodal feature representation was formed by jointly encoding the basic sequence features, AAlndex physicochemical property features, and ESM2 language model features. This mechanism achieves information complementarity at the feature level, preserving the basic compositional information of the sequence while introducing deep semantic features and physicochemical property distribution features, thereby significantly improving the framework's expressive power and prediction accuracy.
[0038] A multimodal deep learning framework is obtained by training a binary classification model, a multi-label classification model, and an activity regression model using multimodal features. The binary classification model is a fully connected neural network (DNN), consisting of an input layer, hidden layers, and an output layer. It is trained using a binary cross-entropy loss function and the Adam optimizer, combining batch normalization and Dropout regularization. The multi-label classification model is obtained through a weighted ensemble of random forest, XGBoost, linear support vector machine, and logistic regression. A one-vs-Rest strategy is used to handle the multi-label problem, and sample weighting is performed based on the label distribution. The activity regression model employs the CatBoost algorithm, specifically the CatBoost gradient boosting tree algorithm. Bayesian optimization is used to automatically adjust the tree depth, number of iterations, learning rate, and regularization parameters, constructing a pipeline structure of "feature selection—standardization—PCA—regression." This invention introduces an adaptive confidence threshold screening strategy at the output stage of the binary classification model. Only when the output probability of the anticancer peptide identification task is higher than a preset threshold of 0.95 can it proceed to the subsequent multi-label classification model and activity regression model, thereby reducing the propagation of false positives.
[0039] The multi-label classification model employs four algorithms: Random Forest, XGBoost, Linear Support Vector Machine, and Logistic Regression. During training, the model learns by optimizing the binary cross-entropy loss function, while weighting strategies are crucial for handling label imbalance. Weighting can be adjusted in terms of label weighting, sample weighting, and loss function weighting, often using the inverse of label frequency as weights to make the model focus more on less frequent labels. Evaluation uses metrics such as precision, recall, and F1-score, and calculates a weighted average to obtain overall performance. A reasonable weighting strategy can significantly improve the multi-label classification model's ability to predict rare labels.
[0040] When using a binary classification model to process the test sequence set, the test sequence set is filtered based on a confidence threshold of 0.95 to obtain the confidence screening results. The test sequence set is then input into the binary classification model, which outputs the probability of each anticancer peptide sequence in the test sequence set.
[0041] When a multi-label classification model is used to process the test sequence set, the type prediction results of each sequence in the test sequence set are obtained; when an activity regression model is used to process the test sequence set, multiple first activity values are obtained, and the number of first activity values corresponds one-to-one with the number of sequences in the test sequence set.
[0042] When the test sequence set is processed sequentially through a binary classification model, a multi-label classification model, and an activity regression model, multiple second activity values are obtained. Based on these second activity values, the anticancer peptide sequences in the specific sequence set are sorted in ascending order to obtain the anticancer peptides with the highest anticancer potential as candidate anticancer peptides. Specifically, the test sequence set is input into the binary classification model, where it is filtered according to a confidence threshold, resulting in a high-confidence sequence set with a confidence level greater than the threshold. This high-confidence sequence set is then input into the multi-label classification model, which predicts the type of each sequence in the high-confidence sequence set, determining the cancer type specificity of each anticancer peptide, thus obtaining a specific sequence set. The specific sequence set is then input into the activity regression model to obtain the corresponding second activity values, with the number of second activity values corresponding one-to-one with the number of sequences in the specific sequence set. The multi-label classification model is used to predict the cancer type of the sequence, and the final output is the confidence level of the sequence and the cancer type. For example, the sequence KKAAKKQQQ has a 90% confidence level against breast cancer.
[0043] This is equivalent to the test sequence set undergoing a three-stage funnel-shaped screening process. First, the first model eliminates sequences with low prediction confidence that are not anticancer peptides. The second stage filters out other types based on specific categories, such as anti-breast cancer peptides. Finally, the activity value is predicted. For example, if the test sequence set contains 100 peptide sequences, a binary classification model identifies 8 as anticancer peptides. These 8 anticancer peptides are then processed by a multi-label classification model, which outputs the probability that each anticancer peptide is anti-breast cancer peptide, thus identifying 5 anti-breast cancer peptide sequences and their probabilities. An activity regression model then predicts the activity values of these 5 anti-breast cancer peptide sequences and ranks them to determine candidate anticancer peptides. The above example is for illustrative purposes only and does not represent real-world data.
[0044] This invention introduces a SHAP-based interpretable analysis method to quantitatively evaluate the contribution of the model at different modalities and feature levels. By calculating the marginal contribution value of each feature in the predicted output, the relative importance of basic sequence features, AAindex physicochemical property features, and ESM2 language model features in anticancer peptide sequence identification, cancer type classification, and activity regression can be clarified. This process not only realizes the visualization and interpretability of the multimodal deep learning framework but also significantly enhances its scientific rigor and credibility.
[0045] In the construction of a three-stage model, the SHAP interpretability analysis method can be introduced after model training to analyze the predictive decisions of each model. Through SHAP analysis, the contribution of each feature (such as amino acid composition, dipeptide composition, etc.) to the prediction of each model can be evaluated, revealing the model's decision-making process, improving model transparency, and providing a basis for subsequent feature selection and model optimization. The SHAP method makes complex classification and regression tasks clearer in terms of feature importance assessment and decision-making process understanding, thereby enhancing the interpretability of the model.
[0046] Example
[0047] A comprehensive prediction and screening method for anticancer peptides based on multimodal deep learning was applied to a large-scale natural peptide sequence screening task. End-to-end screening was performed on 260,000 natural peptide sequences, ultimately obtaining 12 candidate sequences predicted as high-confidence anti-breast cancer peptides. The method includes the following steps:
[0048] First, a multimodal deep learning framework is constructed. Multiple sequences are selected from the CancerPPD and UniProt databases and sequentially filtered, deduplicated, and normalized to obtain positive sample data, negative sample data, multi-label data, and IC data. 50 The data, including positive and negative sample data collectively referred to as binary classification data, consists of 5712 sequences; the multi-label data contains 389 sequences; IC 50The data contains 453 sequences, as shown in Figure 2A. The sequence length distribution of each dataset is presented as a box plot, as shown in Figure 2B. Figure 2C shows the frequency statistics of the seven cancer type labels in the multi-label data. Figure 2D shows the IC after Z-score normalization. 50 The value distribution shows a right-skewed characteristic, with most values concentrated in the range of -1 to 1, which is consistent with the expected pattern of effective normalization of the original wide dynamic range data.
[0049] For positive sample data, negative sample data, multi-label data and IC 50 The data were subjected to feature extraction to obtain basic sequence features (436 dimensions), AAlndex physicochemical property features (5094 dimensions), and ESM2 language model features (1280 dimensions). The basic sequence features, AAlndex physicochemical property features, and ESM2 language model features were then fused to obtain multimodal features.
[0050] Multimodal features were used to train binary classification models, multi-label classification models, and liveness regression models to obtain a multimodal deep learning framework.
[0051] Acquisition of the test sequence set: A set of sequences with recording IC was selected from 260,000 natural peptide sequences in the CancerPPD database. 50 The anticancer peptide sequences were analyzed, forming an existing sequence set. Biomarkers included AAC (amino acid composition), DPC (dipeptide composition), and TPC (tripeptide composition). The results were as follows: AAC included K (lysine), L (leucine), R (arginine), A (alanine), and G (glycine); DPC included KK (two consecutive positively charged lysines), KL, and RR (two arginines); and TPC included KKL (two lysines followed by a leucine), LKK (a leucine followed by two lysines), and AKK (an alanine followed by two lysines). The sequence length was determined to be 18. Based on the marker results and sequence length, multiple test sequences were generated to form a test sequence set.
[0052] Prediction and screening process: The test sequence set is input into a multimodal deep learning framework. Specifically, the test sequence set is input into a binary classification model. Based on a confidence threshold, the test sequence set is screened to obtain a high-confidence sequence set with a confidence level greater than the threshold. This high-confidence sequence set is then input into a multi-label classification model. The multi-label classification model predicts the type of each sequence in the high-confidence sequence set, determining the cancer type specificity of each anticancer peptide in the high-confidence sequence set, thus obtaining a specific sequence set. The specific sequence set is then input into an activity regression model to obtain the corresponding second activity value. Based on the second activity value, the anticancer peptide sequences in the specific sequence set are sorted in ascending order to obtain the anticancer peptides with the highest anticancer potential as candidate anticancer peptides.
[0053] The contributions of multimodal features were analyzed, as shown in Figure 3A, where the relative contributions of each task-specific feature are: AAlndex physicochemical properties dominate in the binary classification model (60%), while ESM2 language model features contribute the most in the activity regression model (56%); Figure 3B shows a heatmap of the quantitative feature importance scores; and Figure 3C shows a radar chart of task-specific contribution patterns, highlighting the complementary characteristics of features between different target tasks. "Task specificity" refers to the optimization process at each stage based on its specific prediction goals and needs. The first stage is a binary classification task, aiming to determine whether a sample belongs to a specific category, with the model focusing on improving classification accuracy. The second stage is a multi-label classification task, predicting whether each sample contains multiple labels (e.g., different types of cancer), requiring handling the interdependence and label imbalance between labels. The third stage is an activity prediction task, aiming to predict the numerical attributes of samples (e.g., activity value or IC50). 50 This is a regression task, focusing on optimizing prediction error and accuracy. Each stage of the task has different objectives, so the model will be specifically optimized and adjusted according to the characteristics of the task at each stage.
[0054] The prediction performance of the binary classification model is shown in Figure 4A, with the accuracy curve stabilizing at 0.958. As shown in Figure 4B, the loss value decreases to 0.48. As shown in Figure 4C, the test set confusion matrix (sample size = 1143) achieves an accuracy of 95.28%. As shown in Figure 4D, the overall performance metrics (MCC = 0.906, Kappa = 0.905, F1 = 0.952) confirm the robustness of the binary classification model.
[0055] The prediction performance of the multi-label classification model is shown in Figure 5A, comparing classification metrics (precision, recall, F1 score, and accuracy) for seven cancer types. Skin cancer performed best (all metrics reached approximately 0.98). Figure 5B shows a comparison of accuracy across different cancer types, with performance ranging from 0.75 to 0.95. Figure 5C demonstrates the model's stability through five-fold cross-validation, with core metrics consistently remaining within the 0.875-0.950 range. Figure 5D shows that the exact match rate improved from 0.3 to 0.5, while the Hamming loss remained stable at each validation fold (approximately 0.2).
[0056] The predictive performance of the active regression model is shown in Figure 6A, IC 50 The scatter plot of predicted values and experimental measurements shows a significant correlation (coefficient of determination R). 2=0.473), and a fitted regression line is attached. As shown in Figure 6B, the residual plot shows that the data points are randomly distributed around the zero value, with no obvious systematic bias, confirming the reliability of the model. As shown in Figure 6C, the mean squared error heatmap of the five cancer-specific models and the general CatBoost algorithm shows that the mean squared error values on the independent test set range from 0.010 to 0.123. As shown in Figure 6D, the performance comparison between the training set (mean squared error of five-fold cross-validation = 0.48) and the test set (mean squared error = 0.387) shows that the active regression model does not exhibit overfitting while maintaining generalization ability.
[0058] Comparison of selected peptide prediction and experimental activity in 4T1 breast cancer cells
[0059] In vitro cell experiments verified the IC50 values of the 12 candidate anticancer peptides. 50 The values ranged from 0.88 to 36.83 μg / mL, all exhibiting significant anticancer activity (Table 1), verifying the effectiveness and practicality of the method of the present invention in the screening of anticancer peptides and prediction of efficacy.
[0060] In summary, the multimodal deep learning framework of this invention achieves an ACP binary classification accuracy of approximately 95.28% and an AUC of 0.991; the multi-label classification model has a macro-average F1 score of approximately 0.881 for cancer type classification, and the activity regression model has a Spearman correlation coefficient of approximately 0.774. Using this multimodal deep learning framework, 12 anti-breast cancer peptides were successfully screened from 260,000 candidate sequences, and their IC50 was verified in vitro. 50 The values range from 0.88 to 36.83 μg / mL. Compared with existing technologies, this invention is the first to achieve the integrated fusion of three tasks: anticancer peptide identification, cancer type-specific prediction, and activity quantification, solving the problems of single-task and low scalability of traditional models. This multimodal deep learning framework has end-to-end automatic screening capabilities, strong interpretability, traceable results, and can provide direct reference for experimental verification, significantly reducing drug development costs and time.
[0061] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A comprehensive prediction and screening method for anticancer peptides based on multimodal deep learning, characterized in that: The test sequence set is obtained and input into a pre-constructed multimodal deep learning framework. This framework includes a binary classification model, a multi-label classification model, and an activity regression model. One model or a combination of the three models is selected to process the test sequence set. When using a binary classification model, the test sequence set is filtered based on a confidence threshold to obtain the confidence filtering results. When using a multi-label classification model, the type prediction results for each sequence in the test sequence set are obtained. When using an activity regression model, multiple first activity values are obtained. When the test sequence set is processed sequentially through a binary classification model, a multi-label classification model, and an activity regression model, multiple second activity values are obtained, and candidate anticancer peptides are obtained based on the second activity values.
2. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 1, characterized in that: The method for obtaining the test sequence set is as follows: analyze the biomarkers of the existing sequence sets in the anticancer peptide database, obtain the indicator results, determine the sequence length, and generate the test sequence set based on the indicator results and sequence length.
3. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 2, characterized in that: The biomarkers include amino acid composition, dipeptide composition, and tripeptide composition; the results are as follows: amino acid composition includes K, L, R, A, and G; dipeptide composition includes KK, KL, and RR; and tripeptide composition includes KKL, LKK, and AKK.
4. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 2, characterized in that: The sequence length is 18.
5. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 1, characterized in that: The construction method of the multimodal deep learning framework is as follows: obtain multimodal features based on existing databases; use multimodal features to train binary classification models, multi-label classification models and liveness regression models to obtain the multimodal deep learning framework.
6. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 5, characterized in that: Methods for obtaining multimodal features: Select multiple sequences from an existing database and preprocess them to obtain positive sample data, negative sample data, multi-label data, and IC. 50 Data; data on positive samples, negative samples, multi-label data, and IC. 50 The data underwent feature extraction to obtain basic sequence features, AAlndex physicochemical property features, and ESM2 language model features; Multimodal features are obtained by fusing basic sequence features, AAlndex physicochemical property features, and ESM2 language model features.
7. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 1, characterized in that: The binary classification model is a fully connected neural network (DNN), which includes an input layer, hidden layers, and an output layer.
8. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 1, characterized in that: The multi-label classification model is obtained through a weighted ensemble of random forest, XGBoost, linear support vector machine and logistic regression.
9. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 1, characterized in that: The activity regression model uses the CatBoost algorithm.
10. The method for comprehensive prediction and screening of anticancer peptides based on multimodal deep learning as described in claim 1, characterized in that: The process of processing the test sequence set through a binary classification model, a multi-label classification model, and an activity regression model is as follows: The test sequence set is input into the binary classification model, and the test sequence set is filtered according to the confidence threshold to obtain a high-confidence sequence set with a confidence level greater than the threshold. The obtained high-confidence sequence set is then input into the multi-label classification model, which predicts the type of each sequence in the high-confidence sequence set to obtain a specific sequence set. The specific sequence set is then input into the activity regression model to obtain the corresponding second activity value.