Enhancing cross-domain protein and peptide interaction with retrained deep learning models

Abstract

Accumulating evidence has shown that protein–peptide interactions (PPepIs) are critical for understanding biological processes and developing peptide-based therapeutics. While deep learning-based protein–protein interaction (PPI) prediction showed promise, it suffers from poor generalization and overfitting problems. This study addresses these challenges by focusing training on short proteins containing much less redundant noninteracting sequence. To avoid artificial PPI, only the experimentally validated PPI pairs from STRING database were used to construct the PPI training dataset. We integrated protein sequence and structure information and presented a multilevel deep learning framework. Training on short-protein datasets yielded higher accuracy and computational efficiency compared with training on long-protein datasets. Moreover, we applied the model to delineate human protein and SARS-CoV-2 virus PPI networks. Notably, we screened PPepIs of current drug peptides with human proteins and SARS-CoV-2 viral proteins, identifying numerous potential side effect or new therapeutic targets. Together, our retrained model could be extensively applied to delineate PPepI network, contribute to peptide drug target identification and side effect analysis, and also provide ample resource for viral infection investigations.

Introduction

Proteins are well-structured organic macromolecules made up of 20 standard amino acids (aa), which serve as key executors of essential biological processes [1]. However, proteins rarely function independently; instead, they participate in protein–protein interactions (PPIs) to regulate and execute complex biological processes. The specificity of PPIs has been widely studied, particularly their applications in drug design (e.g. peptide-based therapies) and viral infection mechanisms (e.g. viral invasion and immune evasion). For instance, designing small molecules or peptides that bind to target proteins allows therapeutic interventions [2–5]. Additionally, the understanding of viral–host protein interactions (V-HPI) provides valuable information on viral infection and immune evasion mechanisms [6–8]. A number of studies have shown the important role of proteins in biological processes and diseases, including infectious diseases and various cancers [9–12]. Various computational studies have been performed by targeting proteins to manage diseases, such as prostate cancer, lung cancer, gastric cancer, and colon cancer [13–16]. Given the vital roles that PPI play in living systems, characterizing these interactions is essential to gain a deeper understanding of life and improve health outcomes.

Many researchers have focused on discovering on PPIs applying experimental methods, including yeast two hybrid screens [17], Tandem affinity purification [18], and mass spectrometric protein complex identification [19]. Although these experimental methods have significantly advanced the characterization of PPIs, they suffer from substantial drawbacks, including high cost, time consumption, and, more critically, low throughput [20–22], which are insufficient to meet the growing demands of studying thousands of proteins. However, with the accumulation of experimental data, the computational methods, particularly those relying on sequence data, have been employed for PPI prediction, increasing throughput while reducing costs and labor.

Traditional machine learning techniques, such as support vector machine [23–26], discriminative vector machine [27], random forest classifier [28], gradient boosting decision tree algorithm [29, 30], weighted sparse representation-based classifier [31], and XGBoost [32], have achieved accuracy rates up to 90%, mainly relying on sequence-based features. More recently, the rapid advancement of deep learning technologies has provided powerful solutions for modeling PPIs, often integrating both sequence and structural data, achieving accuracy rates of up to 95% and even 99%. Representing examples such as Topsy-Turvy [33], TAGPPI [34], D-SCRIPT [35], PIPR [36], DNN-PPI [37], Graph-Bert [38], GCN, and GAT [21], most of which rely on sequence data, with some incorporating structural information. These approaches have also shown great promise across a wide range of bioinformatics applications such as drug–target interaction prediction and functional protein analysis [39, 40]. Despite their high accuracy on training datasets, these models still exhibit limited generalization to new, unseen data [41].

The state-of-the-art (SOTA) model, such as PPI-GNN has achieved 98.34 accuracy and 98.87 F1-score on human datasets. This exceptional performance is largely due to the way the negative pairs in its PPI dataset were constructed. Since experimentally verified noninteracting protein pairs are rarely documented, the negative samples were generated by randomly pairing proteins of different subcellular localizations into a new artificial data sample [21]. However, training and testing on this artificially generated dataset implies overfitting [42], as it will cause the model to pick up interaction patterns that are unique to the artificially generated noninteracting data or patterns influenced by differences in subcellular positions; and neither of these factors truly capture the biological variability of real-world negative pairs. In reality, noninteracting pairs may share similar sequence or interaction features, making it difficult for the model to differentiate them based only on randomly paired negative pairs from different subcellular positions [43]. Training exclusively on this simplified, nonrepresentative negative data could impair the model’s ability to generalize to true, unseen noninteracting proteins [43], especially in complex applications such as cross-species and peptide–drug interactions. This highlights the significance of obtaining a dataset of representative of real-world and addressable to cross-species interaction patterns [44].

Protein interactions primarily occur within small, specific domains [45]. High-molecular-weight proteins with long sequences often contain substantial redundant information, which is not involved in PPI and increases computational costs and the risk of overfitting, thereby impairing the generalizability of the model [46]. Additionally, models often fail to accurately identify key interaction regions when relying solely on sequence data, while the lack of structural information further limits prediction accuracy. Integrating both sequence and structural information from low-molecular-weight proteins (sequence length <300 aa) offers a promising and cost-effective strategy to enhance model performance [47]. This approach improves focus on critical interaction regions, reduces noise, and lowers computational costs. Low-molecular-weight proteins (hereby referred to as “short proteins”) are particularly beneficial because they typically exhibit fewer redundant sequences outside of their functional binding domains. Their simpler structures also provide valuable spatial features, enabling the model to better capture protein folding and interaction interfaces. By focusing on these short proteins, the model benefits from reduced computational and biological complexity, leading to improved prediction accuracy [48]. Furthermore, this approach accelerates both the training and inference processes, enhances model generalization, and is particularly effective for tasks that require a focus on key interaction domains, such as peptide–protein and drug–protein interactions [49].

In this work, we propose to train a deep learning model on short proteins only, i.e. proteins with lengths shorter than 300 aa only. (These proteins will be referred to as “short proteins” and proteins with lengths longer than 300 aa will be referred to as “long proteins” hereafter.) Short proteins were selected primarily because long proteins contain redundant regions in PPI, especially for interactions of peptide drugs or short virus proteins: in most cases, only a very short domain—typically 10–20 aa [50]—is responsible for the active interaction. They contain essential functional domains, and are particularly advantageous in understanding the key interacting regions that govern protein behavior. This is especially important in the context of peptide drug development, as it facilitates the identification of potential targets for drug design and drug repurposing [49]. Short proteins also consume less computational time and memory, which allows for more data samples to be fed for model training given the same computational resources. Therefore, in contrast with using long and short proteins in PPI training data, short proteins are more efficient and contain more essential information, which is particularly valuable for training models to understand critical interaction domains and predicting protein–peptide interactions (PPepI). We also apply the model trained on short proteins to predict and construct protein and peptide PPI network with human and viral proteins. The overarching goal of this study is to improve the generalization and reduce the overfitting of deep learning-based PPepI prediction by focusing on training with short proteins.

Materials and methods

Human sPPI dataset collection

Human small PPIs (sPPIs) obtained from the STRING database (version 12.0) were employed as the training dataset [51] (Table 1). We propose a rigorous curation strategy, utilizing the string network data (protein.links.full.v12.0.txt) as reference, Positive samples were defined as protein pairs with an experimental score >0, indicating experimental evidence for interaction, and with protein lengths <300 aa. This selection led to a total of 58 173 positive protein pairs. For the negative sample set, pairs were chosen where all four scores (experiment, database, experiments_transferred, and database_transferred) were 0, indicating no evidence of interaction. The length of proteins in these pairs was also limited to <300 aa, yielding 684 372 negative instances (positive : negative ratio 1:11.8).

Table 1.

Overview of dataset preparation for human PPI training, including criteria for sample selection, number of protein pairs, and post-filtering balancing

Dataset	Selection criteria	Number
Positive sample set	- “Experiment” score > 0 (indicating experimental support)	58 173
	- Protein length < 300 aa
Negative sample set	- All four scores (experiment, database, experiments_transferred, and database_transferred) = 0	684 372
	- Protein length < 300 aa
Unique sequences (positive)	- Clustering at 70% similarity using CD-HIT	8602
Unique sequences (negative)	- Clustering at 70% similarity using CD-HIT	8490
Post-filtering positive set	- After removing redundancies and interactions absent from the nonredundant sequence database	56 629
Post-filtering negative set	- After removing redundancies and interactions absent from the nonredundant sequence database	660 095
Balanced dataset	- 56 629 positive and 56 629 negative pairs selected for training	56 629 (both positive and negative sets)

Protein sequences in FASTA format were retrieved from the UniProt database [52]. Sequence redundancy was removed using CD-HIT (v4.8.1) with a 70% sequence identity threshold [53, 54], resulting in 8602 nonredundant sequences from positive interactions and 8490 from negative interactions (positive-to-negative ratio 1:0.99). Structural coordinates were acquired from the AlphaFold Protein Structure Database [55] for subsequent structural analyzes.

To ensure data quality and eliminate redundancy, the dataset was subjected to rigorous filtration, retaining unique interaction pairs. Following these procedures, the curated positive set comprised 56 629 nonredundant interactions [56], while the negative set contained 660 095 filtered pairs [57]. Interactions lacking representation in the nonredundant sequence database were systematically excluded. To address class imbalance, a random subset of 56 629 negative pairs [58] was selected through stratified sampling, establishing balanced positive/negative cohorts (1:1 ratio). This stratification protocol ensures mitigation of training bias and improves the reliability of the evaluation.

Yeast and Arabidopsis sPPI dataset collection

For the Yeast and Arabidopsis datasets, we followed a parallel curation pipeline analogous to the human sPPI collection. After redundancy removal and filtration, we curated 3000 balanced positive/negative interaction pairs for Yeast, and 500 balanced positive/negative interaction pairs for Arabidopsis thaliana. Each dataset was constructed at a strict 1:1 ratio of positive-to-negative pairs, ensuring class balance and improving the reliability of subsequent computational evaluations.

Alternative negative dataset

The exceptionally high performance of the previous model, PPI-GNN [21], raises concerns about potential overfitting to its training dataset, as the negative samples were generated by manually randomly pairing proteins originating from different subcellular localizations, causing the model to learn interaction patterns that are specific to the artificially generated negative pairs, or patterns that are tied to subcellular localization differences; neither of which accurately reflect the biological diversity of noninteracting pairs in real-world scenarios.

To investigate overfitting concern, we constructed a dataset with the same number of positive data as the model’s original dataset, sourced from our STRING database, and combined it with the original dataset’s negative protein pairs. As a control dataset, we used the same amount of positive and negative pairs as the original dataset (16 220: 5997) from our own, and compared their accuracy and F1-score along with the original dataset’s metrics. All datasets’ metrics were evaluated via an 8:2 train-test split on their respective datasets.

Retrained GNN model for PPI prediction

We retrained a model based on graph neural networks (GNN), as outlined by Jha et al. [21]. The model starts by building a molecular graph representation of proteins, where nodes correspond to amino acids, and edges represent the spatial relationships between residues, with a cutoff distance of 6 Å between atoms. The node features are derived from protein sequences using pretrained language models such as SeqVec [59] and ProtBert [60], along with additional physicochemical properties encoded as one-hot vectors. The graph is processed using both graph convolutional networks (GCN) and graph attention networks (GAT) to learn the protein’s structural and sequence information, with attention mechanisms in GAT focusing on the most relevant neighboring residues.

For GCN, the propagation rule for each layer is given by the following equation:

(1)

where is the adjacency matrix augmented with the identity matrix (to include self-loops), is the degree matrix, and is the trainable weight matrix in layer . The function represents an activation function such as ReLU.

For GAT, node feature aggregation is defined by the following attention mechanism:

(2)

where is the attention coefficient computed as:

(3)

Here, is the attention function and is the weight matrix that applies a linear transformation to the input feature matrix .

After processing through GCN or GAT layers, the features of the two proteins in an interaction pair are concatenated and passed through fully connected layers with LeakyReLU activation, followed by a sigmoid output layer that predicts the probability of the interaction.

Finally, the model computes the probability of a PPI using the sigmoid activation function as follows:

(4)

where represents the concatenated feature vectors of the two proteins, is the weight matrix for the fully connected layer, and is the bias term. The resulting output is a probability indicating the likelihood that the protein pair interacts.

The model was trained with a batch size of 64 and a learning rate of , using mean squared error as the loss function. Early stopping was applied if the loss did not decrease for 10 consecutive epochs, and random seed for training was fixed at 42. All other model parameters are provided in Supplementary Table S1.

Testing performance of short datasets

As the model with original parameters evaluated on our dataset demonstrated suboptimal performance, we propose to train the model using peptide and protein domain interaction task-specific datasets to improve prediction accuracy: interaction pairs consisting of short proteins only. To evaluate the performance of protein interaction prediction for short protein datasets, we conducted a series of experiments varying both dataset size and random seed, as these factors influence performance metrics differently during training. We created 40 distinct dataset sizes, ranging from 500 to 20 000 protein interaction pairs, which were randomly selected from the entire dataset via 18 different random seeds. For each dataset size, we trained the models across 4 separate RTX 3090 GPUs, resulting in a total of experiments per dataset size. Accuracy, F1-score, sensitivity, precision, specificity, and as well as the training time were calculated to evaluate each dataset’s performance.

Acquisition of peptide drugs and human/viral protein domains

Peptide drug dataset

FDA-approved polypeptide drugs data were obtained from THPdb database [61] containing >200 different peptides. For data quality and redundancy reduction, sequences with the same name were removed, leaving only unique peptide sequences. After applying this filtering, the dataset was reduced to 78 unique peptide sequences [62]. To obtain structural information for these peptides, we used the Phyre2 online platform [63], where structural models of the peptides were generated using homology modeling. This step ensured that each peptide in the dataset had its corresponding 3D structure, allowing for further analysis of peptide-drug interactions. The final dataset included 69 unique peptides with available structural data, since 9 peptides were excluded because their lengths were <30 aa, which prevented the system from performing structural simulation. This dataset is now ready for evaluation of models that predict peptide–drug interactions, as well as for further analysis of peptide-target binding and potential side effects.

Human and virus domain dataset

To facilitate the prediction of PPIs between human proteins and peptide drugs, We retrieved data on interacting structural domains from the 3DID database [64]. Data were filtered to include only rows beginning with “# =3D” [65], which provided the PDB ID and the starting and ending positions of the associated chain. These positions were used to define the interaction domains, which were extracted for further analysis. PDB IDs were used to retrieve corresponding structural data from the PDB database, focusing on human proteins (Homo sapiens) [66] and virus proteins [67]. The organism of each structure, the experimental method, and the resolution were recorded. The structural domains were then extracted based on the chain and position information.

Dataset construction and prediction workflow

Three datasets were constructed with these peptides and proteins: peptide drugs versus human protein domains, virus protein domains versus peptide drugs, and virus versus human protein domains. The first two datasets were constructed by matching the 2 sets pairwise, and the third dataset was formed by selecting 7 important virus protein domains from the organisms severe acute respiratory syndrome coronavirus 2 (SARS-CoV2) and SARS coronavirus Tor2, and then pairwise matching them with all the 160 000 human protein domains, matching all virus protein domains would result in an excessively large dataset. After embedding the peptide and protein’s structure data into PyTorch tensors, the datasets are sent to the model for prediction and the direct output score of the model is logged. This output score is a floating point number ranging from 0 to 1, and reflects the probability that the interaction exists between the two peptides/proteins. We then plot down the interaction activity scores, PPI graphs, and simulate some real interaction structure figures via AlphaFold 2 [68].

Results

Overfitting evaluation

Current deep learning methods, such as the SOTA PPI-GNN, achieve strong performance on their respective datasets but often struggle to generalize to new data. To enhance its applicability to PPepI tasks and mitigate potential overfitting, we retrained the PPI-GNN model using small proteins that contain key interaction domains while minimizing redundant noninteracting sequences. The model represents each protein as a molecular graph, where nodes correspond to amino acids and edges encode spatial relationships. Sequence features extracted from pretrained models like SeqVec and ProtBert are integrated with structural information using GCN and GAT (Fig. 1).

Figure 1.

Workflow for predicting protein-peptide interactions by deep learning: (A) categorization of training data by species, type, protein length, and application; (B) deep learning framework using GCN/GAT layers to predict interactions; (C) training strategy using short proteins to improve efficiency and focus on interaction sites.

To test the adaptability for PPepI tasks, we constructed alternative datasets including one with the original dataset’s negative samples swapped (Neg-swap proposed data) and another one where the numbers of positive and negative samples were exactly balanced with the original dataset (16 220: 5997). First, we performed the training on these dataset using PPI-GNN and conducted multiple evaluation tests on the original PPI-GNN dataset, and other two modified datasets. We plotted the average accuracy and F1-score in Fig. 2A. The metrics sensitivity, precision, specificity and AUROC were plotted in Supplementary Fig. S1. The results indicated that the performance of the neg-swap dataset surpasses that of the original dataset, with accuracy at 97.91% sensitivity even reaching 100%, whereas the performance in the new dataset in our study is much worse (accuracy only 86.02%). Given that the positive samples remain the same across the neg-swap dataset and our dataset, this suggests that the artificially generated negative samples in the original dataset contribute significantly to overfitting. The observed performance gap highlights the overfitting problem and the impact of negative sample selection on model generalization, reinforcing the need for careful dataset construction.

Figure 2.

Metrics and results of different experiment sets: (A) baseline model performance on different datasets including negative-swap and balanced data; (B) training time for Short and Long+Short protein datasets; (C–D) accuracy and F1 score under different training/testing conditions; (E–F) cross-species accuracy and F1 score comparison with SOTA models.

Model training with short proteins and evaluation

As the PPI and PPepI occur mainly in a specific protein domain, the training with long proteins with large redundant sequence beyond the PPI domain may cause lots of noise, we trained the deep learning model using interaction data that only contain short proteins (smaller than 300 aa). Initially, we evaluated the model by directly using pretrained parameters without any fine-tuning (Fig. 2A). The results demonstrated a suboptimal performance, which highlighted the need to retrain the model on domain-specific data to improve its prediction accuracy.

To validate our hypothesis that training on short proteins improves PPI/PPepI prediction, we conducted sets of experiments varying in dataset size, construction random seed, CUDA device. For each experiment, the model was trained on a short protein PPI dataset and a full-length protein (hereafter referred to as “long & short protein”) PPI dataset, respectively. Both sets were tested on short and long & short protein PPI test sets, and the accuracy, F1 Score, sensitivity, precision, and computational resources were logged, resulting in four sets of metric values per experiment. To reduce the fluctuations in the results due to randomness, we then took the average of all the metrics for each dataset size, with sample size of for each dataset size. The 40 different dataset sizes are ranging from 0.5k to 20.0k.

Our data demonstrated that the model training time on S (short protein) datasets are 1 times shorter than LS (long & short protein) datasets, and the accuracy and F1 score metrics of the S datasets are around 15%–20% higher than that of the LS datasets. Figure 2B shows the training time comparison of S VS LS datasets, and Fig. 2C and D namely shows the comparison of accuracy and F1-score. Computational memory of training and other related metrics are shown in Supplementary Fig. S2. These results demonstrate that short protein PPI datasets exhibit higher efficiency and superior performance across various metrics, including training time until convergence, regardless of the dataset sizes. This result confirmed our hypothesis that S datasets training is indeed more efficient, more accurate which may enable the model to capture better essential interacting information.

A notable performance gap between the LS datasets and the “new dataset” mentioned above is due to the positive and negative data ratio. The new dataset, which achieved an accuracy of 86%, was constructed with a positive-to-negative ratio of 16 220: 5997, ensuring consistency in class distribution with the original dataset of PPI-GNN for comparison. With the class distribution balanced at 1:1, LS datasets’ accuracy & F1-score metrics can only reach 76%. This highlights the clear advantage of using short proteins for training, as metrics can increase to 85%—demonstrating that they provide a more effective basis for model training.

Since no benchmark dataset exists for short proteins or PPepIs, we evaluated generalization through cross-species tests on Human, Yeast, and Arabidopsis (Fig. 2E and F). The retrained PPI-GNN consistently outperformed SOTA methods (D-SCRIPT [35] and PIPR [36]) across all three datasets. On average, it achieved 10% higher precision and a higher or comparable F1 score relative to these baselines. For Graph-BERT [69], although we reconfigured the training procedure to adapt it from its original design for node classification to the pairwise prediction task, its performance remained suboptimal and markedly lower than PPI-GNN and other SOTA methods. These results indicate that the improvement is not dataset-specific but reflects robust cross-species generalization.

Prediction of interaction between protein and peptide drug

With the model retrained on short proteins, we explored the potential therapeutic applications of peptide drugs by performing PPepI tests on 69 FDA-approved polypeptide drugs from THPdb [61] and 160 000 human protein domains (Fig. 3A and B). We set the interaction threshold at 0.93, which offers high precision with sufficient recall to ensure reliable yet informative interaction discovery (Supplementary Fig. S3C and D). This analysis identified 224 significant interactions between peptide drugs and human proteins. We further analyzed the disease-related genes extracted from the GWAS Catelog to screen the potential new drug targets of FDA-approved polypeptide drugs. This approach provides a more comprehensive approach to screen potential peptide drug targets of the human proteins involved in various diseases.

Figure 3.

The predicted interactions of peptide drugs and human protein domain. (A) overview of the analysis pipeline; (B) interaction activity scores ranked; (C) PPI network highlighting top hubs; (D) human protein domain hubs（Top 10） and interacting peptides; (E) peptide hubs and interacting human protein domains; (F) representative 3D docking model between a peptide drug and human protein domain.

We constructed a peptide drug–human protein interaction network to systematically explore the therapeutic potential of peptide drugs. To eliminate promiscuous binding proteins that interact with a wide range of targets, we filtered out peptide drugs with excessive interactions. Only those interacting with up to eight potential target proteins were retained to ensure specificity and reduce the likelihood of nonspecific or biologically irrelevant interactions (Fig. 3C–E). This network highlights the potential for multitarget interactions, where a single peptide drug could be repurposed for multiple uses. For example, the polypeptide drug Tenecteplase (THPdb ID: Th1028), a modified form of recombinant human tissue plasminogen activator, is used for the emergency treatment of myocardial infarction and pulmonary embolism [70]. However, no target interaction data for Tenecteplase is available [61, 70]. Our analysis revealed that Tenecteplase interacts with Cadherin-3 (4ZMP, chain A), the risk-associated protein of ulcerative colitis [71]. This suggests a potential therapeutic effect on ulcerative colitis. Additionally, our findings highlight possible side effects of peptide drugs, as they interact with multiple proteins beyond their intended targets. This underscores the need for further investigation into their off-target effects and safety profiles. Ultimately, the data provide a valuable resource for further investigation of the repurposing of peptide drugs and their broader application in treating various diseases and side effect prediction.

Furthermore, we also took advantage of the tool AlphaFold 2: Multimer and obtained simulations of the 3D structural model of the interacting peptides and protein domains. As shown in Fig. 3F, we showed the polypeptide drug Botulinum Toxin Type B (THPdb ID: Th1037) docking with human protein SB27 T cell receptor alpha chain (3KXF, K-chain, residues 187266) with predicted template modeling score (pTM) = 0.4 and interface predicted template modeling (ipTM) = 0.71, which is relatively high. Additionally, the per-residue confidence scores (pLDDT) values for the modeled structure were consistently high (>70), further supporting the reliability of the predictions. These results of interaction spatial structure simulation not only deliver a human-friendly binding visualization, but also provide side-proof for the confidence of true interaction, and offer valuable insights for further structural and functional studies.

V-HPI network and potential peptide drug targeting viral proteins

V-HPI network prediction

To unravel the complex mechanisms of viral infection and immune evasion, we constructed a comprehensive V-HPI network, focusing specifically on interactions between viral proteins — particularly 7 spike protein domains of SARS-CoV-2 and SARS Tor2 — and host human protein domains. The interaction scores of these virus protein domains on all the 160 000 human protein domains in descending order are shown in Fig. 4A. We set the threshold at 0.95 (Supplementary Fig. S3), prioritizing precision to minimize false positives in cross-species virus–host predictions, yielding 293 significant interactions. Importantly, relaxing the threshold to 0.93 yields a denser network but does not alter the main biological conclusions, supporting the robustness of our findings across nearby thresholds.

Figure 4.

Predicted interactions between virus protein domains and peptides or human protein domains, showing activity scores (A), PPI networks with top 10 hubs (B, C), representative interacting pairs (D–G), and 3D structural models of selected docking sites (H, I).

The constructed V-HPI network, as shown in Fig. 4C–E allow us to examine how viral proteins interact with host cell proteins, especially immune system-related proteins, providing valuable insights into the viral entry process and its impact on immune response modulation. We analyzed the SARS-CoV-2-derived domain (7LME B chain), which is part of a known virus–host protein complex. This domain exhibits the potential to interact with multiple human protein domains, suggesting a key role in virus-host interactions. Within the human protein complex 3AZJ, the B chain corresponds to the Histone H4 domain, which has been observed interacting with multiple viral spike protein domains in our predictions. This suggests a potential role for core histones in mediating virus–host interactions. Notably, previous studies have demonstrated that viral proteins can directly bind to human core histones, potentially modulating chromatin structure and host gene expression. For instance, the human cytomegalovirus (hCMV) IE1 protein contains two distinct histone-interacting regions with differential binding specificities for H2A-H2B and H3-H4 [72], highlighting a conserved mechanism by which viruses target host chromatin components. These findings support the reliability of our prediction results and highlight their potential applicability in understanding virus–host interactions.

Potential peptide drug targeting virus proteins

We further investigated the potential of peptide drugs in targeting viral proteins by conducting interaction tests on peptide drugs and SARS-CoV-2 spike protein domains. The results revealed that several peptide drugs, including Aprotinin (THPdb ID:Th1158) could interact with viral protein domains. Aprotinin is a natural proteinase inhibitor derived from bovine lung and is commonly used to reduce the risk of perioperative blood loss and the need for blood transfusions in high-risk patients undergoing cardiopulmonary bypass during coronary artery bypass graft surgery [70]. While no protein interaction data for Aprotinin is available in databases such as DrugBank and THPdb [61, 70], our analysis reveals that Aprotinin has the potential to interact with viral protein domains involved in host invasion. This finding suggests that Aprotinin may exert effects on viral infections, warranting further investigation into its therapeutic potential against virus. In addition to Aprotinin, other peptide drugs identified in our study were found to bind to key regions of viral proteins that play crucial roles in host invasion. These interactions suggest that peptide drugs may disrupt critical viral–host interactions essential for viral replication and immune evasion. Our findings demonstrate the feasibility of leveraging peptide-based therapies to target viral proteins, offering a promising avenue for the development of novel antiviral drugs.

Finally, we represent two simulated 3D structural model of these interacting virus protein domains, human protein domains, and peptides in Fig. 4H and I. In Fig. 4H, with ipTM = 0.29, pTM = 0.74, and pLDDT , peptide Susoctocog alfa (THPdb ID: Th1238) is docked with the SARS-CoV-2-derived domain (7LME, B chain, residues 187266). Figure 4I shows the human chloride intracellular channel protein 1 (CLIC1)-derived domain (3TGZ, A-chain, residues 2387) interacting with another SARS-CoV-2-derived domain (7LME, B chain, residues 29305), with pTM = 0.15, ipTM = 0.71, and pLDDT .

Discussion

While AI-based PPI prediction has been extensively explored, its accuracy, generalization capability, and overfitting problems still need to be significantly improved for biological and biomedical applications [21, 22]. In this study, we explored the prediction of PPepIs with a focus on using short proteins (<300 aa) and their potential applications in peptide drug repurposing and viral–host interactions. Integrating structural and sequence features of short proteins improved PPepIs prediction performance, leading to a 15%–20% average increase in accuracy and F1-score compared with models trained on longer proteins. Using short proteins highlighted the critical domains responsible for most PPepIs, reduced noise from irrelevant regions [45, 46], and halved training time, thereby providing a more efficient and accurate approach to capture key interaction sites and explore protein functions in diverse biological contexts. Cross-species evaluation on Human, Yeast, and Arabidopsis datasets showed that our retrained model consistently outperforms SOTA methods (D-SCRIPT [35] and PIPR [36]) while adapted Graph-BERT [69] performs suboptimally, demonstrating that our approach not only improves prediction accuracy but also generalizes robustly across species.

Owing to the high accuracy and generalization capabilities of our model, we applied it to identify human proteins that interact with the SARS-CoV-2 spike protein and constructed a V-HPI network. Focusing on spike-mediated interactions, our analysis revealed 293 high-confidence PPIs between viral and human proteins. These findings provide insights into the molecular mechanisms of viral immune evasion. Understanding how viral proteins engage host targets is critical for developing therapeutic strategies aimed at disrupting such interactions and preventing infection. Notably, we identified an interaction between the SARS-CoV-2 spike protein and human Histone H4, a potential drug target involved in host chromatin regulation. This predicted V-HPI network serves as a valuable resource for studying viral pathogenesis and host immune responses.

Importantly, our model enables both the repurposing of existing peptide drugs and the evaluation of their potential off-target effects. By predicting interactions between FDA-approved peptide therapeutics and human protein domains, we demonstrated that many of these drugs may possess novel therapeutic potential, including antiviral activity. Screening 69 FDA-approved peptide drugs against 160 000 human protein domains, we identified 224 high-confidence PPepIs. For example, Aprotinin, originally approved as a protease inhibitor, was predicted to strongly interact with the SARS-CoV-2 spike protein, highlighting its potential in antiviral intervention. At the same time, our analysis underscores the necessity of assessing possible side effects arising from the multitarget nature of peptide drugs. For instance, Tenecteplase, used in the treatment of myocardial infarction and pulmonary embolism, was found to interact with Cadherin-3 (PDB ID: 4ZMP, chain A), a protein implicated in ulcerative colitis, as well as several other human proteins. Such off-target interactions may interfere with normal protein functions, raising concerns about unintended physiological effects. Understanding these interactions is critical for ensuring both the safety and efficacy of peptide-based therapeutic strategies.

Conclusion

Our results highlight the importance of careful dataset design, particularly regarding the quality of negative samples and the length of protein sequences, for training reliable PPepI prediction models. This study connects performance improvements in deep learning with actionable biological insights, providing a scalable framework for therapeutic discovery and virological research. While our current approach does not explicitly address multidomain proteins, which may limit the ability to capture inter-domain interactions [73], we consider domain–domain interactions to be fundamentally pairwise. Future work may further enhance the model by incorporating higher-order structural dependencies.

Key Points

• A high-quality dataset was constructed using only experimentally validated protein–protein interaction (PPIs) from the STRING database to avoid artificial interactions.

• The retrained model achieved higher accuracy and computational efficiency in PPI/protein-peptide interactions (PepPI) prediction tasks.

• The model was successfully applied to predict PPepIs between peptide drugs, human proteins, and SARS-CoV-2 proteins, identifying potential therapeutic targets and side effects.

Author contributions

X.C. was responsible for retraining the model, manuscript writing, figure creation, and model testing. J.L. conducted the literature review, curated, and preprocessed the datasets, and assisted with figure creation and manuscript writing. F.M. provided guidance on model training and testing, and contributed to the manuscript. B.Y. contributed to figure creation and provided revisions to the manuscript. Y.Z. supervised the project, offering critical feedback, and guidance throughout. All authors reviewed and approved the final manuscript

Conflict of interest: None declared.

Funding

This work was funded by The Academician Expert Workstation in Yunnan Province (grant no. 202405AF140107) and the National Natural Science Foundation of China (grant No. 32330104).

Data availability

The codes and datasets (and their metadata) in this work are available on Github (https://github.com/XinCao02/Cross-PPepI-DL).