Efficient prediction of anticancer peptides through deep learning

Anticancer peptide prediction

Several techniques have been developed for predicting the anticancer activity of peptides. These techniques can be broadly divided into two main approaches: structure-based methods and sequence-based methods (Dukka, 2013).

Sequence-based methods focus on analyzing the primary sequence of peptides to extract features and patterns associated with anticancer activity. These methods often rely on physicochemical properties, amino acid compositions, or sequence motifs as input (Lv et al., 2021). Machine learning algorithms such as the support vector machine (SVM), random forest, and naïve Bayes have been used to classify peptides based on their anticancer properties (Yu et al., 2020). One popular sequence-based approach is the position-specific scoring matrix (PSSM), which represents the probability of observing a specific amino acid at a given position in a sequence (Vanhoof et al., 1995). PSSM profiles capture evolutionary information and have been successfully applied in predicting anticancer peptides.

Structure-based methods, in contrast, incorporate the three-dimensional structure of peptides into the prediction process. These methods typically involve molecular docking, molecular dynamics simulations, or quantitative structure–activity relationship (QSAR) modeling. Structure-based methods consider factors such as binding affinity, interaction energy, and spatial arrangement of atoms to predict anticancer activity (Farhadi & Hashemian, 2018).

While both sequence-based and structure-based methods have shown promise, they have limitations. Sequence-based methods often struggle to capture the intricate relationships between peptide sequences and anticancer activity, especially when facing diverse and complex sequence patterns. Structure-based methods, on the other hand, require knowledge of the three-dimensional structure, which may not always be available for peptides.

To overcome these limitations, recent studies have explored the integration of multiple data sources and the use of hybrid approaches. For instance, hybrid methods combine sequence-based features with structural information, incorporating the best of both approaches. Furthermore, advanced machine learning algorithms, such as deep learning models, have been employed to improve the accuracy of anticancer peptide prediction (Hosen et al., 2022). However, despite the progress made in anticancer peptide prediction techniques, there are still challenges that need to be addressed. The imbalanced nature of anticancer peptide datasets, the limited availability of experimentally validated peptides, and the need for robust feature representation are areas that require further investigation.

Deep learning approaches in peptide prediction

Deep learning approaches have emerged as powerful tools for peptide prediction, offering the ability to capture complex patterns and representations from data. These methods utilize neural networks with multiple layers to automatically learn features and make accurate predictions.

Recurrent neural networks (RNNs) have been widely employed in peptide prediction tasks. RNNs were developed to deal with sequential data, and have shown effectiveness in modeling peptide sequences. Long short-term memory (LSTM) networks, a type of RNN, have been particularly successful in capturing long-range dependencies and modeling peptide sequence data (Angermueller et al., 2016). LSTM models have demonstrated promising results in predicting antimicrobial peptides (Sønderby et al., 2015) and anticancer peptides (Yang et al., 2023).

CNNs have also gained traction in peptide prediction. CNNs excel at capturing local patterns and features through the use of convolutional layers. These layers apply filters to local regions of the input, letting the network learn hierarchical representations of the data. CNN models have been employed to predict peptide activities such as antimicrobial activity (Hu, Hesham & Zou, 2022) and bioactivity against specific targets (Koutsoukas et al., 2011). Furthermore, 2D CNNs have been utilized to extract features from peptide structures, enabling accurate prediction of protein-peptide binding affinity (Guo et al., 2018).

Attention mechanisms have been incorporated into deep learning models for peptide prediction to improve their performance. Attention mechanisms enable the network to concentrate on relevant fragments of the input while disregarding irrelevant information. By applying attention mechanisms to peptide sequences or structures, the models can selectively attend to critical features, enhancing prediction accuracy (Loo et al., 2014). Attention-based models have been employed in tasks such as antimicrobial peptide prediction (Yan et al., 2020) and protein-peptide binding affinity prediction (Motmaen et al., 2023).

Hybrid models that combine deep learning with traditional machine learning algorithms have been explored. These models leverage the strengths of both approaches to improve prediction performance. For example, hybrid models combining CNNs with SVMs have been used to predict peptide bioactivity, and they outperformed individual methods (Sanders et al., 2011). Transfer learning, where deep learning models pre-trained on large-scale datasets are fine-tuned for specific peptide prediction tasks, has shown promise in improving prediction performance when training data are limited (Fenoy, Edera & Stegmayer, 2022).

Despite the advancements, challenges persist in the application of deep learning to peptide prediction. The availability of large and diverse annotated datasets is crucial for training deep learning models effectively. Additionally, the interpretability of deep learning models remains a challenge, making it difficult to understand the underlying reasons for their predictions (Koh & Liang, 2017).

Table 1 provides a comparison of the strengths, limitations, and gaps in the literature for different techniques used in the prediction of anticancer peptides. The table highlights the key aspects of each technique, such as proven performance and interpretability, as well as limitations such as reliance on handcrafted features and difficulty in handling sequence variability.

Deep learning approaches such as RNNs, CNNs, and attention mechanisms have shown promise in peptide prediction tasks by effectively capturing complex patterns and representations. Hybrid models and transfer learning techniques can further enhance prediction performance. However, addressing challenges related to data availability and interpretability is crucial for the continued advancement of deep learning in peptide prediction.

Dataset description and preprocessing

In this section, a comprehensive description is provided of the dataset used in this research and the preprocessing undertaken to ensure its quality and suitability for training the 2D CNN model.

Two-dimensional convolutional neural network architecture

This section presents a comprehensive description of the architecture of the 2D CNN model utilized for enhancing the prediction of anticancer peptides. The 2D CNN architecture comprises convolutional, pooling, and fully connected layers that work collectively to learn meaningful representations from peptide sequences. The key components are shown in Table 4.

The input layer receives the encoded peptide sequences and their associated features. The shape of the input is determined by the sequence length and the chosen encoding scheme. Convolutional layers apply a set of filters or kernels across the peptide sequences to capture local patterns and extract pertinent features. Each filter performs a convolution operation, sliding across the input and producing a feature map. An activation function such as the rectified linear unit (ReLU) is applied to introduce non-linearity and enhance the model’s ability to capture complex relationships between input and output. Max pooling layers downsample the feature maps generated by the convolutional layers, reducing the dimensionality and retaining the most salient features. This helps to decrease the model’s computational complexity while preserving important information. The flattened layer converts the pooled feature maps into a flattened vector, which serves as input for the fully connected layers. Fully connected dense layers are responsible for learning high-level representations and making predictions. The number of neurons in dense layers can be adjusted based on the complexity of the problem. An activation function such as the ReLU is typically applied to dense layers to introduce non-linearity and enhance the model’s expressive power. The output layer generates the final predictions, typically represented as the probability of each peptide sequence belonging to a particular class (e.g., active or inactive).

The specific configuration of the 2D CNN architecture, including the number of filters, kernel size, pooling size, and the number of neurons in dense layers, may vary depending on the complexity of the dataset and the existing computational resources.

Feature extraction and representation

These processes play a vital role in the performance of machine learning models, including convolutional neural networks, for peptide prediction tasks. This section describes feature extraction techniques employed to transform raw peptide sequences into meaningful representations that capture essential characteristics and patterns.

Training and optimization

This process plays a key role in the development of an accurate and robust predictive model for anticancer peptide prediction. In this section, we describe the steps involved in training the 2D CNN model and optimizing its performance.

Evaluation metrics

Evaluation of the trained model is essential to assess its performance and effectiveness in predicting anticancer peptide activity. The evaluation metrics to measure the model’s performance are described in this section.

Dataset description and preprocessing

The cornerstone of our experiments lies in a meticulously curated dataset comprising diverse peptide sequences annotated with anticancer activity labels, as shown in Table 9. This dataset amalgamates peptides from various sources, including public databases, experimental studies, and curated datasets.

Each peptide sequence is accompanied by a binary label indicating whether it possesses active or inactive anticancer properties. Supplementary metadata, such as peptide length, origin, and chemical properties, enrich our understanding of peptide characteristics, as shown in Table 10.

Dataset split and cross-validation

To ensure a robust evaluation of the proposed model, the dataset was divided randomly into training, validation, and testing sets. The dataset’s random split still preserved the distribution of positive and negative samples in each set. The training set is utilized to train the model, the validation set is used to fine-tune hyperparameters and monitor performance, and the test set is utilized for the final evaluation, as shown in Table 11.

Cross-validation further validates the model’s performance and assesses its generalization ability. K-fold cross-validation splits the dataset into k subsets or folds.

The model is trained and tested k times, each time with a different fold serving as the validation set, and the remaining folds serving as the training set. The performance metrics then average the k iterations to obtain robust and unbiased estimates of the model’s performance, as shown in Fig. 1.

The experiments used an 80:10:10 split for training, validation, and testing sets, respectively. Five-fold cross-validation ensures a comprehensive assessment of the model’s performance.

This split and cross-validation strategy enables us to evaluate the model’s performance on different subsets of the data, assess its generalizability, and ensure reliable and unbiased performance estimates, as shown in Fig. 2. It gives us confidence in the model’s effectiveness and its potential for real-world applications.

Model performance evaluation

We evaluated the performance of the proposed 2D CNN architecture using accuracy, precision, F1-score, recall, and AUC-ROC, as shown in Table 12.

These metrics provide insights into the classification accuracy, the model’s ability to correctly identify positive samples, and its overall performance in distinguishing between positive and negative samples, as shown in Fig. 3.

Comparison with existing methods

The performance of the proposed deep learning model for the prediction of anticancer peptides was compared with existing methods in the literature to assess its effectiveness and superiority over the state-of-the-art approaches.

The proposed method demonstrated superior performance in accuracy, precision, recall, F1-score, and AUC-ROC, as shown in Table 13. The accuracy of our proposed method reached 0.87, indicating it correctly predicted the anticancer activity for a significant proportion of the samples. A precision of 0.85 demonstrates that our method effectively identified true positive samples out of all samples predicted to be positive. The recall value of 0.89 shows that our method accurately captured a large portion of true positive samples out of all positive samples. The F1-score of 0.87, which combines precision and recall, reflects the overall effectiveness of our method in achieving a balanced performance. Furthermore, the AUC-ROC of 0.91 indicates an excellent ability to discriminate between positive and negative samples.

The significant improvement in performance compared to existing methods can be attributed to the enhanced ability of our model to capture intricate spatial patterns and dependencies within peptide sequences using a two-dimensional convolutional neural network. The 2D CNN architecture effectively extracted informative features and learned complex relationships, leading to improved prediction accuracy.

In addition to the quantitative evaluation, qualitative analysis and interpretability of the results further support the superiority of our method. The model’s interpretability provides insights into the important features and patterns that contribute to the prediction of anticancer peptide activity, helping to understand the underlying biological mechanisms.