An ensemble-based model comprising deep learning for predicting peptide-binding residues in proteins

Abstract

Protein–peptide interactions are fundamental to numerous cellular processes and are linked to diseases like cancer when disrupted. Understanding these interactions is critical for both functional genomics and drug discovery. Despite growing availability of protein–peptide complexes, experimental methods to study them remain resource-intensive and costly. While computational approaches offer a complementary solution, their predictive accuracy is often inadequate. To overcome these limitations, we present PepENS, an ensemble model combining deep learning and traditional machine learning techniques that integrates both structural and sequence-based features from primary protein sequences. By leveraging half-sphere exposure, position-specific scoring matrices from multiple-sequence alignments, and embeddings from a pre-trained protein language model, PepENS demonstrates superior performance compared to the state-of-the-art methods. The proposed model demonstrated strong performance, achieving a precision of 0.596 and an AUC of 0.860 on the Dataset 1 test set. On the Dataset 2 test set, it attained a precision of 0.539 and an AUC of 0.846. Notably, these results reflect improvements over state-of-the-art methods in terms of precision and AUC by 2.8% and 0.5%, respectively, on Dataset 1, and by 2.3% and 2.4%, respectively, on Dataset 2. The PepENS software and associated datasets are available at https://doi.org/10.6084/m9.figshare.28490012.v2.

Introduction

Protein–peptide interactions play a crucial role in numerous cellular processes, such as metabolism, gene expression, and DNA replication [1, 2]. While these interactions are vital for maintaining cellular health, they can also contribute to diseases like viral infections and cancer [3, 4]. Gaining a molecular-level understanding of these interactions could lead to significant advancements in therapies and diagnostic techniques. Notably, small peptides are responsible for mediating ∼40% of these important interactions [5]. Furthermore, peptide-based drugs are gaining increasing prominence due to their high specificity, low toxicity, and favorable safety profiles compared to traditional small-molecule therapeutics [6, 7].

The experimental methods for studying protein–peptide interactions, even with advancements in structural biology, face considerable challenges [8]. These methods can be expensive, time-consuming, and technically difficult due to the small size of peptides [9], their weak binding affinities [10], and their flexibility [11]. Conversely, computational approaches provide a complementary option but are hindered by challenges related to prediction accuracy and computational efficiency. These challenges largely stem from the limitations of existing algorithms in dealing with the complex nature of protein–peptide interactions.

Computational approaches for predicting protein–peptide interactions generally fall into two main categories: structure-based and sequence-based. Structure-based models, such as PepSite [12], SPRINTStr [13], and Peptimap [14], rely on structural features like accessible surface area (ASA), secondary structure (SS), and half-sphere exposure (HSE) to make predictions. On the other hand, sequence-based methods like SPRINT-Seq [15], PepBind [16], Visual [17], PepNN-Seq [18], PepBCL [19], Le et al. [20], E2EPep [21], and PepCA [22] use machine learning algorithms, drawing on factors such as embeddings, amino acid composition, physicochemical properties, and evolutionary features. Notably, PepBind [16] was the first to consider intrinsic disorder in its feature design, recognizing its importance in protein–peptide interactions. Some recent methods have combined the structure-based and sequence-based features to develop their models. For instance, SPPPred [23] used HSE, SS, ASA, position specific scoring matrices (PSSM), and physicochemical properties, and DP-site [24] employed PSSM, ASA, SS, local backbone angles, physicochemical properties, and amino acid occurrence to predict protein–peptide binding residues.

The advent of deep learning (DL) technologies has introduced a new paradigm to computational proteomics. Modern algorithms now enable the transformation of protein features into image-like formats, making them suitable for DL architectures like 2D convolutional neural networks (CNNs) [25, 26]. For instance, Wardah et al. [17] developed a CNN-based approach called Visual, which encodes protein sequences into image-like forms to predict peptide-binding residues. Additionally, transformer-based models have become powerful tools for sequence representation [27], often surpassing traditional algorithms by effectively capturing long-range interactions within sequences. For example, Abdin et al. [18] introduced PepNN-Seq, which utilizes a pre-trained contextualized language model called ProtBert [27] for embedding protein sequences. Wang et al. [19] employed ProtBert [27] within a contrastive learning framework; Le et al. [20] and Huang et al. [22] utilized the embedding from ProtT5 [27] and ESM-2 [28] pre-trained language models, respectively, and Hu et al. [21] fused the embeddings from ESM-2 [28] and ProtBert [27] to predict protein–peptide binding residues.

DL algorithms, a specialized branch of machine learning, have demonstrated significant potential in tackling complex issues in protein science and structural biology [29–31]. These algorithms, modeled after human cognitive processes, use artificial neural networks to learn intricate data representations [32, 33]. Unlike traditional machine learning methods such as k-nearest neighbors and support vector machines, DL models excel at independently identifying patterns and features within data [34]. Initially gaining popularity in areas like medical imaging, speech recognition, computer vision, and natural language processing, these algorithms have achieved milestones like accurately predicting protein folding, proving particularly effective for large and complex datasets [35]. CNNs [36] have shown remarkable effectiveness in image classification tasks, indicating their potential for analyzing other types of spatial data, such as protein structures [37, 38]. Their capacity to maintain spatial hierarchies within data makes them particularly well-suited for use in proteomics. At the same time, progress in natural language processing has led to the creation of pre-trained contextualized language models tailored for protein biology, thereby expanding the range of computational tools available to the field [39, 40]. Given the data-heavy demands of modern biotechnological research, DL technologies are increasingly being applied to proteomics, making it a promising area for further exploration [41, 42].

Ensemble learning is a robust technique that enhances predictive performance by combining the strengths of multiple models, often through methods like averaging or voting [43]. This approach creates a model that outperforms individual models. However, training deep ensemble models is computationally intensive. By integrating DL with traditional machine learning techniques, ensembles can capitalize on the strengths of both, resulting in more accurate and reliable models [44]. For example, Gao et al. [45] used residual CNN with XGBoost, Random Forest, LightGBM, and extremely randomized trees to build a protein–protein interaction prediction ensemble model, which surpassed the performance of different individual DL frameworks, traditional classifier methods, and state-of-the-art methods. The ensemble model, comprising DL and traditional machine learning techniques, offers a promising prediction pipeline with a balanced approach that captures the strengths of both methodologies.

Driven by these technological advancements, we developed PepENS, an innovative model that seamlessly integrates sequence- and structure-based information with ensemble learning. Our approach represents a pioneering, consensus-based method by combining embeddings from ProtT5 (ProtT5-XL-UniRef50), a protein language model (pLM) based on transformers by Elnaggar et al. [27], with PSSM and HSE features to train an ensemble model consisting of EfficientNetB0, CatBoost, and Logistic Regression (LR), as shown in Fig. 1. To the best of our knowledge, this is the first time tabular data, including features derived from transformer-based attention mechanisms, are converted into images through DeepInsight technology [25] to enable the application of CNNs in protein–peptide binding prediction. The spatial relationships formed between the features by DeepInsight to produce images, and the feature extraction from those images by the CNN, play a key role in the performance of the ensemble model. Extensive evaluations demonstrate that PepENS establishes a new benchmark, surpassing existing methods such as the sequence-based predictors Le et al. [20], PepBCL, and PepNN-Seq, which utilize a pre-trained language model PepBind with intrinsic disorder features; SPRINT-Str, which focuses on structural features like ASA, SS, and HSE; and the SPPPred and DP-site methods that integrate both structural and sequence-based features. Superior predictive performance of PepENS redefines computational approaches and holds the potential to accelerate drug discovery, deepen our understanding of disease mechanisms, and inspire new computational strategies in bioinformatics. The key contributions of this work are:

Figure 1.

Architecture of the proposed PepENS model for predicting binding and non-binding residues in protein sequences. It has five primary modules: Data acquisition, Feature extraction, Sample extraction, Transformation by DeepInsight, and Classification.

• PepENS, a novel ensemble model that integrates large language models, PSSM profiles, and structure-based (HSE) features for protein–peptide binding prediction.

• Introduction of DeepInsight to convert tabular transformer-based features into images for the first time in this context, enabling the use of CNNs (e.g. pretrained EfficientNetB0) for binding site prediction.

• A hybrid ensemble architecture combining EfficientNetB0, CatBoost, and LR to capture both spatial and non-spatial patterns in the data.

• Comprehensive benchmarking demonstrating that PepENS consistently outperforms several state-of-the-art methods across diverse feature sets.

• Demonstrated potential for accelerating drug discovery, uncovering disease mechanisms, and guiding future research in computational biology and bioinformatics.

Materials and methods

The architecture of PepENS includes five primary modules: Data acquisition, Feature extraction, Sample extraction, Transformation by DeepInsight, and Classification. The following subsections describe each of the modules in detail.

Data acquisition

We utilized two widely recognized benchmark datasets in this study, denoted Datasets 1 and 2, to ensure a fair comparison between our proposed method and existing approaches. These datasets are employed by recent state-of-the-art methods for training and testing models, allowing for consistent evaluation and comparison [19]. The two datasets were originally sourced from the BioLiP database [46], and to ensure uniqueness, sequences with over 30% sequence identity were excluded using the “blastclust” tool from the BLAST package [47]. This resulted in a total of 1279 peptide-binding proteins, and a residue within these sequences is considered binding if any of its heavy atoms are found to be within 3.5 Å (angstroms) of a heavy atom in the peptide through laboratory experiments [15].

Dataset 1

Dataset 1 was originally introduced by Taherzadeh et al. [13] in their structure-based method, SPRINT-Str, where the test set (TE125) included 125 proteins. They selected proteins with at least 30 amino acids and three or more binding residues, randomly assigning 10% to the test set and the rest to the training set. TE125, therefore, contains 29 151 non-binding residues and 1719 binding residues. We followed the same procedure as Taherzadeh et al. [13] to construct our training set.

Dataset 2

Dataset 2 was proposed by Zhao et al. [16] in their sequence-based method, PepBind. They randomly split the 1279 proteins into two equal parts for training and testing. The test set (TE639) includes 141 840 non-binding and 8490 binding residues across the 639 proteins. For training, they selected 20% of the 640 proteins to reduce training time. In this study, however, we considered all the 640 proteins for training, which had 149 103 non-binding residues and 8259 binding residues.

Feature extraction

The features selected in this study include representations from a pre-trained pLM, evolutionary relationships using a multiple sequence alignment (MSA) tool, and structural attributes related to residue solvent exposure. During feature extraction, these three types of features were obtained by analyzing the 1279 proteins with three tools: ProtT5 pLM [27] for embedding values, PSI-BLAST [47] for PSSM, and hsexpo [48] for HSE values. The following subsections provide a detailed explanation of each feature type.

Transformer embedding

Transformer models, which utilize advanced DL algorithms, have demonstrated significant potential in proteomics by leveraging vast and growing protein sequence databases. By applying natural language processing techniques, these models interpret protein sequences similarly to sentences, with each amino acid functioning as a word. This approach enables transfer learning, where insights gained from tasks with abundant data can be applied to related tasks with limited data, enhancing their applicability in diverse proteomic analyses [49, 50].

Elnaggar et al. [27] developed several pLMs, with ProtT5 being one of the most widely used for various tasks [51]. ProtT5 is built on the T5 architecture [52], which is similar to the original design for language translation [53], featuring encoder and decoder blocks. The encoder transforms the input sequence into an embedding space, while the decoder generates the output embedding. ProtT5, a model with 3 billion parameters, was trained on the Big Fantastic Database (2.1 billion proteins) [54] and fine-tuned on the UniRef50 database (45 million proteins) [55]. When extracting embeddings, only the encoder blocks are used, as they consistently outperformed the decoder in all tasks in the proposed work.

The ProtT5 model generates an output embedding in the form of a matrix with dimensions L × 1024, where L is the protein’s length and 1024 represents the values from the network’s last hidden layer. This matrix, created through the attention mechanism, captures intricate relationships between amino acid residues in the input protein sequence. It produces a comprehensive set of features that reflect important structural and functional information about the protein.

Half-sphere exposure values

Information about a protein’s surface is crucial for predicting protein–peptide binding sites, as peptides often bind to shallow surface regions [56]. HSE effectively measures solvent exposure to distinguish buried, partially buried, and exposed residues [57]. Widely used in binding prediction tasks [13, 23, 58, 59], HSE values in this study were obtained using the hsexpo program, which assesses how buried an amino acid is in a protein’s 3D structure. HSE is calculated by dividing a sphere of radius r_d = 13 Å around a residue’s Cα atom into two halves, providing HSE-up and HSE-down values. HSE-up corresponds to the number of Cα atoms in one half of the sphere that is oriented in the direction of the side chain, while HSE-down represents the number of Cα atoms in the opposite half, oriented away from the side chain [60]. The total number of Cα atoms within the sphere of a residue is also considered [48], known as the contact number. The result of hsexpo is a feature matrix of dimension L × 3 that gives the values for HSE-up, HSE-down, and the contact number for each residue in the protein sequence.

Position specific scoring matrices

The PSSM are created using MSA algorithms to identify homologous sequences in protein database and determine the frequency of amino acids at each position. PSSM has played a crucial role in predicting protein contacts and structures [61, 62], as well as in tasks like assessing the functional impacts of mutations [63] and rational protein design [64]. In this study, PSI-BLAST was used to generate the PSSM with an E-value threshold of 0.001 across three iterations. The PSSM, with dimensions L × 20 (where L is the sequence length and 20 represents the amino acids of the genetic code), shows the probabilities of substituting each amino acid in the sequence. A higher value indicates a greater likelihood of substitution by the corresponding amino acid. PSSM can be formulated as P = {P_ij : i = 1...L and j = 1...20}, where P_ij represents the linear probability for the jth amino acid at the ith position of the input sequence.

Sample extraction

In this work, there are two ways the samples in the protein sequences are extracted: window sizes 1 and 3. The extraction is with respect to the window sizes around the amino acid residues. For a given residue, the residues to its left and right can be referred as the upstream and downstream residues, respectively. Window size 1 represents the extraction based on the residue itself, while the window size 3 comprises the residue of concern (center residue) and its immediate upstream and downstream neighboring residues. In the case where there are not enough residues to form the window size 3 (at the start or end of the sequence), the mirroring technique [65] is used to make up for the insufficient residues. Residue at the start of the sequence that does not have an upstream residue requires left mirroring of the downstream residue to form the window, while the residue at the end of the sequence that does not have a downstream residue requires right mirroring of the upstream residue to form the window. Extracting samples with different window sizes with the corresponding features from protein sequences is depicted in Fig. 1. The window sizes can be formulated as

(1)

(2)

where R is an amino acid residue, S_$w$1 the window size 1, and S_$w$3 the window size 3.

A sample can therefore be represented depending on the classifier as

(3)

where S can be categorized either as a positive sample (binding) or a negative sample (non-binding) based on the experimental label:

(4)

Transformation by DeepInsight

CNNs are inherently designed for image data, so feature vectors need to be converted into images for CNN models to process them. Sharma et al. [25] introduced the DeepInsight method, which transforms non-image data into images to utilize CNN’s feature extraction and classification capabilities. The steps involved in the transformation are shown in the Transformation by DeepInsight component of Fig. 1. The DeepInsight method begins by applying the t-SNE dimensionality reduction algorithm [66] to generate a 2D plane for the normalized feature vectors, which have values between 0 and 1. It is important to note that the dimensionality reduction takes place on the number of samples rather than the dimension of the feature vector, which gives the locations of the features on the 2D plane based on the samples in the training set. The convex hull algorithm then determines the smallest area containing all feature coordinates, allowing for the formation of a rectangular box that encompasses all the points. The box is rotated to align horizontally and vertically, and the Cartesian coordinates are converted to pixel coordinates. Finally, the values of the features are mapped to the corresponding pixel positions, thereby forming an image from the feature vector. During the transformation, if multiple features are mapped to the same pixel location, they are averaged and placed at that spot. As a result, if the image resolution or grid size is too small relative to the number of features, many features may overlap, leading to a less accurate image representation. It’s essential to choose an appropriate resolution based on both hardware capacity and the number of features needed for processing.

Classification

In this work, ensemble learning is used for the classification of residues in the protein sequences into the two categories: binding and non-binding. The following subsections describe the development of the classification module of the model architecture.

Dataset balancing

The benchmark datasets used in this study are significantly imbalanced, with fewer positive samples compared to negative ones. In the training sets of Datasets 1 and 2, the imbalance ratios are 1:17 and 1:18, respectively. Such imbalance can skew the classification process, so it’s crucial to address this for reliable model building. To handle the issue, random under-sampling [67, 68] was applied, which involved randomly selecting negative samples, making the dataset balanced.

Statistical measures

The proposed method was evaluated using the TE125 and TE639 test sets after being trained on their respective balanced training sets. Since the test sets are also highly imbalanced, appropriate evaluation metrics were selected to assess the classification performance. Five statistical measures were employed: sensitivity, specificity, precision, matthews correlation coefficient (MCC), and the area under the ROC curve (AUC), ensuring a comprehensive evaluation of the method’s effectiveness [69–71]. The statistical measures from sensitivity to MCC are defined as follows:

(5)

(6)

(7)

(8)

where TP, TN, FP, and FN are true positives, true negatives, false positives, and false negatives, respectively.

TP refers to the number of actual binding residues that are correctly predicted by the model, TN indicates the actual non-binding residues that are accurately classified by the model, FP represents the actual non-binding residues that are incorrectly predicted as binding, and FN is the number of actual binding residues that are incorrectly classified as non-binding by the model.

For the given model, sensitivity (Equation 5) measures the proportion of correctly predicted binding residues, while specificity (Equation 6) reflects the fraction of non-binding residues accurately predicted. Precision (Equation 7) evaluates how many of the predicted binding residues are correct, and MCC (Equation 8) assesses the model’s performance for both binding and non-binding residues. Sensitivity, specificity, and precision range from 0 to 1, with higher values indicating better predictions, while MCC ranges from −1 (strong negative correlation) to +1 (strong positive correlation). It is important to note that the metrics discussed earlier depend on the classifier’s probability threshold, meaning the values can change with different thresholds. This variability limits their reliability for model evaluation. To address this, we also include the AUC metric, which is calculated from classification probabilities and remains independent of threshold settings. AUC offers a more reliable evaluation of model performance by assessing its ability to distinguish between binding and non-binding residues. It takes on values from 0 to 1, with higher values indicating better separability.

Ensemble model

In the ensemble model of this work, we have used models from the domains of DL and traditional ML: EfficientNetB0, CatBoost, and LR. There are four individual predictors that make up the ensemble: 1 × EfficientNetB0, 2 × CatBoost, and 1 × LR. The prediction probabilities of these models are averaged to find the final prediction of the ensemble model.

EfficientNetB0: A CNN model designed to improve performance by balancing the network’s width, depth, and resolution. It outperforms previous CNNs in both accuracy and efficiency. The key aspect of its architecture is a scaling method that uses a compound coefficient to adjust the three dimensions uniformly. This process optimizes depth, width, and resolution for maximum accuracy given resource constraints. EfficientNet’s primary component is the mobile inverted bottleneck convolution with squeeze-and-excitation. EfficientNet models range from B0 to B7 [72], based on this scaling approach. In this study, EfficientNetB0 was employed to classify the binding and non-binding sites based on the images created by the DeepInsight method.

CatBoost: It is a very recent member of the family of Gradient Boosted Decision Trees which is quite enhanced that operates on binary decision trees [73, 74]. It introduces improvements in handling categorical data, boosting accuracy, and reducing overfitting, while maintaining fast performance. Unlike other models, CatBoost combines predictions from multiple models to create more stable and generalizable results, reducing overfitting by averaging errors across models. It uses symmetric trees, which ensures uniform splitting at the same depth for faster computation and better overfitting control. CatBoost also applies a strategic splitting method, selecting splits with minimal penalties to optimize decisions. Moreover, the model uses perfect hashing to store categorical feature values efficiently, reducing memory consumption. This approach minimizes the need for tuning many hyperparameters. The use of CatBoost in bioinformatics has been gaining traction due to these advantages such as the ability to handle categorical features effectively and its robust performance in complex classification tasks [75–78]. Based on the above reasons, we have employed two CatBoost models in our ensemble model.

Logisitic regression: It is a discriminative classifier that maps a function from dataset features to target classes, estimating the probability that a new instance belongs to a specific class. It is highly computationally efficient, uses minimal memory, not prone to overfitting, and can be easily trained and deployed on smaller datasets. LR offers a simple, linear framework that complements more advanced, non-linear models in ensemble learning, thereby reducing error rates [79]. Its inclusion helps to stabilize predictions and therefore enhances both the consistency and generalization capability of the ensemble model. In addition, it serves as a regularizing element, preventing the ensemble from becoming overly dependent on complex patterns that may only perform well on the training data [80]. Given the simplicity and robust performances of LR [81], and its complementarity with complex models, we have incorporated a L2-regularized LR into our ensemble model.

Results

Experimental details

Before training the individual classifiers, the Datasets 1 and 2 were balanced to bring down the imbalance ratio to 1:1.5. The AUC metric was used for evaluating the performances during the hyperparameter tuning and the model training stage. As part of our benchmarking, we trained multiple tree-based classifiers, including Random Forest, XGBoost, and CatBoost, and observed that CatBoost consistently achieved the highest performance. The comparative results are presented in the Supplementary Fig. S1. For the two CatBoost classifiers used in the final ensemble, each was trained on training data with different distribution of S_$w$3 samples from the undersampling procedure. The CatBoost algorithm, from the open-source software library developed by Yandex [82], was utilized, and its hyperparameters—depth, iterations, and learning rate—were tuned using the Hyperopt optimization algorithm [83] with the logloss loss function. EfficientNetB0 on the TensorFlow framework [84] was utilized with input image size of 130 × 130 pixels obtained from the S_$w$1 samples. The S_$w$1 samples were chosen since they have the least feature dimension, which would result in less feature overlaps during the transformation by DeepInsight. Transfer learning was then applied, whereby the weights of the model were loaded from ImageNet and fine-tuned. The output of the model was flattened, and a dropout of 0.2 was applied to avoid overfitting before being fed to the single-neuron output for classification. The training set was divided into 80% for training and 20% for validation. The hyperparameters—learning rate, momentum, and weight decay—were optimized using the gradient descent algorithm with the binary cross-entropy loss function. The network was optimized using Hyperband tuning algorithm in the Keras Tuner library [85]. To employ the LR algorithm, scikit-learn library was used and trained on the S_$w$3 samples. The Hyperopt algorithm [83] was employed to optimize the hyperparameters of inverse regularization strength, maximum number of iterations, penalty type, and solver selection. For the CatBoost and LR classifiers, the window sizes S_$w$1, S_$w$3, and S_$w$5 (window size 5) were evaluated and are provided in the Supplementary Fig. S2. The hyperparameters of each classifier used to build the ensemble model for Datasets 1 and 2 are provided in the Supplementary Table S1.

Performance comparison with existing methods

To demonstrate the effectiveness of our PepENS model, we compared its performance against fourteen existing approaches: Pepsite [12], Peptimap [14], SPRINT-Seq [15], SPRINT-Str [13], PepBind [16], Visual [17], PepNN-Seq [18], PepBCL [19], SPPPred [23], PepCNN [86], Le et al. [20], DP-site [24], E2EPep [21], and PepCA [22]. We had also intended to compare our approach with a not very recent method called DeepProSite [87], which has demonstrated remarkable performance. However, our attempts to reproduce the method were unsuccessful, and despite reaching out to the author, we received no response. The evaluation metrics used included sensitivity, specificity, precision, MCC, and AUC. Sensitivity reflects the true positive rate, specificity measures the true negative rate, precision indicates the positive predictive value, MCC assesses the correlation between predicted and experimental labels, and AUC evaluates the overall classification capability of the model. Unlike the other metrics, AUC is independent of the probability threshold, offering greater reliability for model performance assessment [88, 89].

The results for the TE125 and TE639 test sets are presented in Tables 1 and 2, respectively. A threshold of 0.875 was applied in Table 1, while a threshold of 0.880 was used in Table 2. Since these test sets were also utilized by previous methods, their results, as shown in the tables, are directly taken from their studies. From the comparison, it is evident that PepENS, our proposed method, outperforms all other previously established approaches on both the TE125 and TE639 test sets. In addition, we have included ROC and precision-recall curves for both test sets (Figs 2 and 3, respectively) to provide a more comprehensive evaluation of the model’s performance.

Table 1.

Performances of the proposed PepENS model and the existing methods on the TE125 test set

Method	Sensitivity	Specificity	Precision	MCC	AUC
Pepsite [12]	0.180	0.970	−	0.200	0.610
Peptimap [14]	0.320	0.950	−	0.270	0.630
SPRINT-Seq [15]	0.210	0.960	−	0.200	0.680
SPRINT-Str [13]	0.240	0.980	−	0.290	0.780
PepBind [16]	0.344	−	0.469	0.372	0.793
Visual [17]	0.670	0.680	−	0.170	0.730
PepNN-Seq [18]	−	−	−	0.278	0.805
PepBCL [19]	0.315	0.984	0.540	0.385	0.815
SPPPred [23]	0.315	0.959	−	0.230	0.710
PepCNN [86]	0.254	0.988	0.550	0.350	0.843
Le et al. [20]	0.340	0.985	0.564	0.413	0.856
DP-site [24]	0.770	0.799	0.580	0.381	0.829
E2EPep [21]	0.381	−	0.525	0.420	0.851
PepCA [22]	0.386	0.967	0.405	0.360	0.848
PepENS (ours)	0.251	0.990	0.596	0.365	0.860
PepENS* (ours)	0.774	0.776	0.169	0.289	0.860

The highest values in each column are highlighted in bold.

*Indicates PepENS evaluated using a probability threshold optimized to balance sensitivity and specificity.

Table 2.

Performances of the proposed PepENS model and the existing methods on the TE639 test set

Method	Sensitivity	Specificity	Precision	MCC	AUC
PepBind [16]	0.317	−	0.450	0.348	0.767
PepNN-Seq [18]	−	−	−	0.251	0.792
PepBCL [19]	0.252	0.983	0.470	0.312	0.804
PepCNN [86]	0.217	0.986	0.479	0.297	0.826
Le et al. [20]	0.390	0.958	0.527	0.334	0.823
E2EPep [21]	0.331	−	0.438	0.349	0.840
PepCA [22]	0.399	0.945	0.343	0.302	0.826
PepENS (ours)	0.187	0.990	0.539	0.296	0.846
PepENS* (ours)	0.767	0.761	0.161	0.275	0.846

The highest values in each column are highlighted in bold.

*Indicates PepENS evaluated using a probability threshold optimized to balance sensitivity and specificity.

Figure 2.

ROC and precision-recall curves for the TE125 test set.

Figure 3.

ROC and precision-recall curves for the TE639 test set.

For the TE125 test set (Table 1), PepENS achieves a sensitivity of 0.251, specificity of 0.990, precision of 0.596, MCC of 0.365, and AUC of 0.860. Compared to existing methods, including Le et al.’s approach [20], which was the best-performing method prior, PepENS shows improvements in specificity, precision, and AUC. Specifically, specificity improved by 0.2%, precision by 2.8%, and AUC by 0.5%, with the most significant enhancement observed in the precision metric, followed by AUC and specificity.

The performance results on the TE639 test set are detailed in Table 2, where our method achieved sensitivity, specificity, precision, MCC, and AUC values of 0.187, 0.990, 0.539, 0.296, and 0.846, respectively. Similar patterns were observed as in TE125, where our method outperformed the existing approaches in terms of specificity, precision, and AUC. The improvements were most notable in the precision metric with a 2.3% increase, followed by AUC with a 0.7% gain and specificity with a 0.4% enhancement.

As indicated by the results in Tables 1 and 2, our method outperformed other approaches in the majority of metrics for both test sets, particularly in AUC, which is a crucial measure of the classifiers’ overall discriminatory power. These enhancements highlight the significance of utilizing transformer embeddings, HSE values, and PSSM, as well as incorporating these features into ensemble learning models that combine DL with traditional machine learning algorithms.

Examination of binding and non-binding predictions

To scrutinize the prediction capability of the proposed model, we randomly selected a protein sequence from the TE125 test set to have a closer look at the predicted and the actual label of the amino acid residues. The protein sequence (pdbID 2ykaA) is shown in Fig. 4 in terms of its 3D structure and primary sequence. The experimental and predicted binding residues are depicted in orange and magenta colors, respectively, while the non-binding residues are shown in gray. As seen in the figure, the prediction of binding and non-binding residues by the proposed method closely matches that of the experimental labels.

Figure 4.

Visualization of the protein pdbID 2ykaA showing the experimental (in orange) and predicted (in magenta) binding residues of the amino acid residues using (A) 3D structure and (B) primary sequence. The non-binding residues are shown in gray. The 3D structure was obtained using the PyMOL software [90].

Figure 4B shows the label of each residue by our method and the wet-lab experiment. The numbers of FP and FN are 2 and 5, respectively, while the numbers of TP and TN are 8 and 109, respectively. These results yield a specificity of 0.982 and precision of 0.8. The values indicate that the precision of our method on this protein sequence is higher than the overall value obtained on the TE125 test set, while the specificity measure is reasonably similar. Nonetheless, this shows that the proposed method has a high capability to predict binding residues that are actually correct, and considering that our method achieved the highest precision and specificity compared to other methods, extrapolating this information to other protein sequences indicates that our method does not deviate much from the binding regions of the experiment and it also avoids falsely predicting binding residues at arbitrary positions. Given that fewer residues are involved in peptide binding than non-binding, our method can tremendously aid in narrowing down the regions for experimental verification, and this will greatly reduce the time, cost, and effort needed to understand protein–peptide binding sites in unannotated protein sequences.

Ablation study

To assess the contribution of each classifier, experiments were conducted to analyze their impact on the overall performance of the ensemble model. The AUC scores for various ensemble configurations were calculated on the TE125 test set by excluding classifiers one at a time and comparing the results with the full proposed ensemble. The results, displayed in Fig. 5, indicate that the highest AUC was achieved with the complete ensemble, comprising CatBoost 1, CatBoost 2, LR, and EfficientNetB0. The lowest performance occurred when EfficientNetB0 was omitted, underscoring the significance of the DL model to extract patterns from the images generated using the DeepInsight method. The absence of the CatBoost classifiers resulted in the second-largest reduction in performance, highlighting their high predictive capability, while removing the LR model had the least impact. These findings demonstrate that the inclusion of various classifiers enhances the ensemble model’s effectiveness. Additionally, Fig. 6 presents the AUC scores of each individual classifier on the test sets, providing further insight into their standalone predictive strengths.

Figure 5.

TE125 test set AUC scores of the ensemble model with individual classifiers removed.

Figure 6.

AUC performance of individual classifiers in the ensemble model on the test sets.

Beyond this, we employed SHAP (SHapley Additive exPlanations) [91] to interpret the contribution of individual features in each classifier within our ensemble model, using the training set from Dataset 1. Figures 7, 8, and 9 present SHAP summary plots for the LR, CatBoost1, and CatBoost2 models, respectively, illustrating how each feature influenced model predictions across the training data. In these plots, UR, CR, and DR denote upstream, center, and downstream residues, respectively. For EfficientNetB0, due to the high computational cost of SHAP calculations with DL models, we randomly selected 300 training samples and calculated the mean absolute SHAP values. Figure 10A illustrates the pixel intensities of these SHAP values on the image transformer generated by DeepInsight, indicating how much the pixels contributed to the model’s output, and Fig. 10B presents a bar chart of the features corresponding to the top pixels by importance. Each of the SHAP visualizations of the classifiers ranks the top 35 features according to their importance, displayed from top to bottom. In the SHAP summary plots, red points indicate high SHAP values (strong positive influence on the prediction), while blue points represent low SHAP values (negative influence). Positive SHAP values correspond to the binding class, whereas negative values indicate the non-binding class. The bar chart for EfficientNetB0 provides a global view of feature importance, reflecting how much each feature contributes to the model’s output on average. Across all classifiers, the HSE-derived features consistently emerged as the most influential, followed by the ProtT5 transformer-based embeddings and then the PSSM-derived features. This pattern highlights the importance of structural features (HSE) in protein–peptide binding prediction and underscores the complementary value of sequence-derived embeddings and evolutionary profiles in enhancing model performance.

Figure 7.

SHAP summary plot for the LR classifier.

Figure 8.

SHAP summary plot for the CatBoost1 classifier.

Figure 9.

SHAP summary plot for the CatBoost2 classifier.

Figure 10.

Feature ranking for the EfficientNetB0 model. (A) Mean absolute SHAP values of each pixel. (B) Bar chart of the features corresponding to the top pixels by importance.

Discussion

The study of protein–peptide interactions is significant because peptides, typically consisting of 5 to 15 residues, have low toxicity and small interaction interfaces, making them ideal for therapeutic applications and drug discovery [92]. Peptide-based inhibitors are also being explored for the treatment of conditions like diabetes, cancer, and autoimmune diseases [93]. While peptides were previously overlooked due to their short half-life and poor absorption [94], advancements in synthetic techniques have enabled modifications to enhance their biophysical and biochemical properties, renewing interest in them as potential drug candidates [95]. Other advantages of peptides include the ability to penetrate deep tissues, efficient cellular internalization, and reduced immunogenicity [96].

Understanding the structure of protein–peptide complexes is crucial for designing peptide-based drugs [97]. However, studying these complexes presents distinct challenges compared to other molecular interactions like protein–protein and protein–ligand interactions. In protein–protein interactions, complexes form based on well-defined 3D structures, while protein–ligand interactions involve small molecules binding to deeply buried regions within proteins. In contrast, peptides often have unstable structures and typically bind with low affinity to large, shallow pockets on protein surfaces [56]. These factors, coupled with the limitations of experimental techniques like X-ray crystallography and nuclear magnetic resonance, underscore the need for advanced computational methods to study these complexes.

We have demonstrated, through the use of protein 3D structure and primary sequence, that the prediction of PepENS closely aligns with the experimental labels. Achieving a low false positive rate is desired, as only 5.4% of the residues are involved in binding, while the remaining 94.6% are not [19]. Close prediction of residues is essential for experimental processes, allowing researchers to focus on likely binding sites in proteins that have not been experimentally confirmed. The high specificity and precision of our method accelerates the identification and understanding of binding residues, ultimately speeding up the development of drugs for disease treatment.

A key component in the PepENS model is the inclusion of DeepInsight technology, which facilitates the use of a 2D CNN model in our ensemble model. This approach has since inspired several other methods for tabular-to-image conversion, including REFINED [98], TINTO [99], IGTD [100], and NCTD [101], and is among the more frequently cited and adopted techniques reported in the literature [102]. Its applicability has extended to diverse data modalities, including those employed in classical clinical prediction models, genomics, time series analysis, and natural language data [103]. DeepInsight offers several notable advantages, such as being adaptable to data from any domain, ability to uncover patterns by the establishment of spatial relationships between features, which may be difficult to detect in the raw feature space [104] and is also compatible with Vision Transformers (ViT), a relatively recent advancement in DL for image-based tasks [105]. Furthermore, it supports the use of different manifold learning techniques, such as UMAP (Uniform Manifold Approximation and Projection), t-SNE, and others. In this study, we selected the t-SNE method to perform the transformation to position the features on the Cartesian plane. Additionally, DeepInsight offers a number of discretization methods for converting data into pixel coordinates. Despite these advantages, DeepInsight faces a limitation: features may overlap in the same pixel due to the restricted pixel frame size of the image. To minimize distortion, we optimized the resolution by experimenting with various image sizes, as illustrated in Fig. 11. Furthermore, the perplexity parameter in the t-SNE algorithm plays a crucial role in generating accurate plots that reflect the original cluster sizes in high-dimensional space [106]. Larger perplexity values would be preferable for datasets with more samples. As a result, we manually tuned the perplexity by testing several values (shown in Fig. 12) and selecting the one with the highest AUC score. The choice of the perplexity value not only minimizes overlap but also enables the formation of meaningful spatial relationships between features, allowing CNN to effectively extract critical patterns.

Figure 11.

AUC scores on the TE125 test set evaluated using EfficientNetB0 across various image sizes generated by the DeepInsight technique. The model’s parameters were held constant during training for all image sizes. The best AUC performance was achieved with images sized 130 × 130 pixels.

Figure 12.

AUC values on the TE125 test set obtained using EfficientNetB0 trained on 130 × 130 pixel images, with varying perplexity settings in the t-SNE algorithm of the DeepInsight technique. The highest AUC was achieved at a perplexity value of 20.

The results above show that PepENS is capable of accurately predicting binding and non-binding residues in protein sequences. While achieving sensitivity and specificity values of 1 is ideal, this balance is rarely observed in clinical and computational biology research, as improving one metric often results in the decrease of the other [107]. Therefore, AUC is primarily used to evaluate binary classification models and compare their performance across datasets. In both datasets tested, the proposed method outperformed previous state-of-the-art approaches on most metrics, including AUC, demonstrating the utility of combining pLM embeddings, PSSM, and HSE features in an ensemble model composed of traditional and DL models. These results highlight the potential of our method for identifying binding sites in a wide range of proteins, which is vital for fundamental biological research and practical applications.

Conclusion

In this study, we introduced PepENS, an ensemble prediction model that integrates both DL and traditional machine learning techniques to predict protein–peptide binding residues. The model draws upon sequence-based features derived from a pre-trained pLM and an MSA tool. Additionally, it incorporates a structural feature known as HSE. By utilizing these diverse properties of protein sequences and leveraging the predictions of DL that capture complex patterns and boost the performance of the overall ensemble model, PepENS effectively distinguishes between binding and non-binding residues. As demonstrated on the two datasets, PepENS outperformed the existing state-of-the-art methods that also utilize primary protein sequence.

Despite the impressive performance of our method, there are still avenues for further enhancing the predictive power of the model. This can be achieved by employing a more deeper network, such as a higher-range model from the EfficientNet family of models, employing more sophisticated tools for converting feature vectors to images, such as Multi-representation DeepInsight technology [108], which is an improved method of DeepInsight, and exploring longer window sizes that can potentially include more information for samples.

Supplementary data

Supplementary data is available at NAR Genomics & Bioinformatics online.

Conflict of interest

None declared.

Data availability

The datasets and software codes used in this paper can be downloaded using the link: https://doi.org/10.6084/m9.figshare.28490012.v2