PepPharmaHub: A Cloud-Based Platform Integrating Multimodel Language Architectures with Curated Data Resources for Therapeutic Peptide Discovery

doi:10.21203/rs.3.rs-7645806/v1

Download PDF

Method Article

PepPharmaHub: A Cloud-Based Platform Integrating Multimodel Language Architectures with Curated Data Resources for Therapeutic Peptide Discovery

https://doi.org/10.21203/rs.3.rs-7645806/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

Therapeutic peptides represent a rapidly expanding class of drug candidates due to their diverse biological activities and high specificity. However, accurately predicting peptide functions directly from sequence information remains a major challenge in computational peptidomics. Current tools, typically standalone applications or functionally constrained web servers, lack the flexibility and scalability essential for modern peptide discovery workflows. Therefore, it is necessary to develop a cloud-based, no-code platform that enables customizable modeling and high-throughput functional screening of therapeutic peptides.

Results

PepPharmaHub (http://bioinmed.jflab.ac.cn:18090/peppharmahub/) provides a cloud-based, end-to-end platform that integrates advanced sequence-based language modeling with curated benchmark datasets and interactive visualization modules. The platform features a high-throughput screening module powered by a diverse set of 24 models targeting 20 therapeutic properties, alongside a customizable model training pipeline for user-defined screening tasks. Comprehensive benchmarking on 24 public datasets demonstrates that PepPharmaHub matches or surpasses state-of-the-art predictors, significantly improving the efficiency of large-scale peptide screening. Compared with existing public web servers, PepPharmaHub attains a higher, more tightly distributed accuracy on 3,475 newly reported bioactive peptides from 2023–2025 (20 independent tasks), indicating stronger generalization and practical utility.

Conclusions

PepPharmaHub enables accurate, high-throughput prediction of peptide functions through customizable deep learning models and a no-code interface. By outperforming existing tools across multiple benchmarks and supporting interpretable sequence analysis, the platform offers a practical solution for accelerating peptide-based drug discovery.

Therapeutic peptide discovery

Web server

Deep learning

BERT

Therapeutic peptides have emerged as a promising class of drug candidates owing to their high specificity and favorable safety profiles. These peptides exert therapeutic effects by mimicking or modulating the biological activities of endogenous peptides, proteins, or hormones[1], targeting pathways such as receptor agonism or antagonism, enzyme modulation, signal transduction regulation, and immune system modulation [2]. Over 40 types of functional peptides have been identified, including antihypertensive, antioxidant, immunomodulatory, anticancer, antibacterial, opioid-like, cholesterol-lowering, antithrombotic, or antidiabetic activities [3, 4]. These peptides have attracted significant attention due to their high target specificity, functional diversity, good biocompatibility, and low immunogenicity [5]. Consequently, peptide drugs are increasingly recognized as a pivotal direction for future drug development [6, 7].

Despite their therapeutic potential, the discovery and development of these molecules are hindered by the complexity of predicting functional activities from peptide sequences. The vast diversity of peptide sequences in biological systems complicates the functional annotation, and traditional experimental methods are often too resource-intensive to capture this complexity [8]. Advances in bioinformatics and artificial intelligence have enabled data-driven approaches, such as machine learning and deep learning to analyze the relationships between sequence features and experimentally determined activities, thereby improving functional prediction [9, 10].

Numerous computational predictors have been developed for specific therapeutic properties, including anticancer activity [11], angiotensin-converting enzyme (ACE) inhibition [12], antimicrobial properties [13], antioxidant [14], cell penetration [15], toxicity [16], dipeptidyl peptidase IV inhibition [17], and anti-coronavirus activity [18]. These predictors utilize a variety of modeling approaches, ranging from Simple Recurrent Neural Networks (SimpleRNN), Long Short-Term Memory Networks (LSTM), Bidirectional Long Short-Term Memory Networks (BiLSTM), Gated Recurrent Unit (GRU), and transformer-based pretrained models like BERT, as well as hybrid variants. Sequence characterization methods can be categorized into five main types: (i) encoding methods leveraging physicochemical descriptors of amino acid residues (e.g., hydrophobicity, polarity, volume) and one-hot sparse matrices [19]; (ii) deep feature extraction using attention mechanisms from pretrained models [20]; (iii) computation of physicochemical properties such as molecular weight, volume, and isoelectric point [21]; (iv) k-mer segmentation for fixed-length fragment representation [22]; (v) molecular fingerprint encoding based on chemical structures or simplified molecular input line entry system (SMILES) representations, generating molecular fingerprints like extended connectivity fingerprints (ECFP) to describe structural features of molecules [23].

However, this field faces three critical challenges. First, the absence of a unified modeling platform undermines reproducibility and limits the diversity, and timeliness of data analysis. Second, the insufficient computational resources, standard data integration, and pre-selected peptide libraries hinder the development of integrated algorithms and high-throughput screening capabilities. Third, poor model interpretability, with inadequate visualization of key residues and motifs, impedes mechanistic insights into peptide functionality.

To address these issues, we have developed PepPharmaHub, a comprehensive web-based platform that integrates modules for model training, model deployment, virtual screening, and sequence characterization. PepPharmaHub features a user-friendly interface that eliminates the need for specialized programming skills, allowing researchers to visually monitor model training, filtering, and functional motif analysis. Additionally, the platform includes small-sample standard datasets and pre-selected peptide libraries. In summary, PepPharmaHub addresses critical challenges in therapeutic peptide discovery by improving the accuracy, efficiency, and interpretability of functional prediction. It provides a powerful, user-friendly tool for biomedical researchers and drug developers, accelerating the discovery and development of peptide-based therapeutics.

PepPharmaHub facilitates one-stop screening for therapeutic peptides

To streamline the development of therapeutic peptides, we developed PepPharmaHub, a unified web platform that integrates curated peptide databases, customizable modeling frameworks, high-throughput virtual screening, and result visualization into a cohesive, end-to-end workflow. The system is organized into five core functional modules, each contributing to a one-stop pipeline comprising data access, model training, virtual screening, visualization, and analysis (Fig. 1).

(1) Built-in centralized and structured peptide data resources

This database module provides foundational data support for model development and benchmarking. It integrates 24 public peptide datasets and two in-house sequence libraries encompassing over 2.3 million peptides with unknown activity. Additionally, it hosts 24 prediction datasets and characteristic sequence profiles for 20 classes of bioactive peptides (Supplementary Fig. 1), offering a rich and standardized resource for both model training and screening tasks.

(2) Deep Learning and Dual-Mode Screening

PepPharmaHub provide a flexible and extensible framework for constructing and deploying deep learning-based peptide prediction models. The training module supports the online development of task-specific models through two components: the RNN-Trainer, which offers four types of recurrent neural networks with customizable configurations, including 24 residue encoding schemes, 19 activation functions, adjustable network depth, early stopping, and checkpointing; and the BERT-Trainer, which enables fine-tuning of transformer-based architectures using both standard parameters such as batch size, epochs, and learning rate, and advanced parameters such as attention heads, dropout, and activation layers (Fig. 1b). These tools allow users to build biologically tailored models without requiring programming expertise. For downstream deployment, the Virtual Screening module (Fig. 1c) provides two complementary options: the Peptide-Predictor, which includes 24 optimized pre-trained models for functional classification across 20 peptide attributes; and the Model-Caller, which supports high-throughput batch predictions using user-defined models trained within the platform. This dual-mode screening framework distinguishes PepPharmaHub from conventional predictors by enabling both rapid inference using standardized models and customizable prediction via user-trained models.

(3) Web-based implementation for interactive visualization

PepPharmaHub features a web-based architecture that enables seamless user interaction and real-time task monitoring. The Visual Analysis module (Fig. 1d) allows users to interactively inspect modeling and screening workflows by associating each task with a unique identifier, through which dynamic logs, intermediate outputs, final results, and performance metrics can be accessed and downloaded. Visualization capabilities include sequence feature comparisons, classification overviews, and model evaluation statistics, supporting transparent and reproducible analysis. The platform is organized into six functional interfaces including: Home, Database, RNN-Trainer, BERT-Trainer, Peptide-Predictor, and Model-Caller (Supplementary Fig. 2 and Fig. 1e), which together support end-to-end task execution. Users can initiate analyses by uploading FASTA files and selecting relevant parameters, while monitoring progress and retrieving results for up to 15 days. The modular design allows individual components to be used independently or combined into integrated pipelines, thereby enhancing flexibility and usability for both expert and non-expert users in therapeutic peptide research.

PepPharmaHub demonstrates robust generalization and adaptability across open benchmark datasets

To evaluate the generalization performance and adaptability of PepPharmaHub across diverse peptide prediction tasks, we conducted systematic cross-validation and independent testing on 24 publicly available benchmark datasets (Table 1). Using the BiLSTM model with SDPZ27 encoding and the BERT model with k-mer = 1, cross-validation demonstrated robust performance: 17 BiLSTM-based models achieved average accuracies exceeding 0.8015, with Sensitivity, Specificity, MCC, and AUC ranging from 0.3000 to 1.0000, 0.8514 to 1.0000, 0.3191 to 0.9729, and 0.8000 to 0.9975, respectively (Supplementary Table 2). Similarly, 22 BERT-based models yielded average accuracies above 0.8242, with respective performance metrics spanning 0.6800-1.0000 (Sensitivity), 0.8438-1.0000 (Specificity), 0.6401-1.0000 (MCC), and 0.8741-1.0000 (AUC) (Supplementary Table 3). On independent test sets (Fig. 2a-x), PepPharmaHub-BiLSTM achieved superior predictive performance on two antioxidant, interleukin-6-inducing, and ACE-inhibitory datasets, showed comparable performance on 7 datasets, and relatively lower performance on the remaining 13 datasets. In contrast, PepPharmaHub-BERT outperformed on 11 datasets, was comparable to state-of-the-art models on 7 datasets, and underperformed on 6 tasks. Notably, BERT exhibited stronger feature extraction capabilities than BiLSTM on 17 datasets, though its performance was slightly inferior on antioxidant, ACE-inhibitory, anti-coronavirus, antiparasitic, and antibacterial tasks. Overall, the PepPharmaHub models matched or exceeded the performance of existing methods on 19 out of 24 datasets. These findings underscore the platform’s capacity to deliver competitive and generalizable predictive models across a wide range of peptide functional categories. Importantly, users can complete the entire evaluation pipeline by simply uploading datasets, selecting model architectures, and defining training parameters, demonstrating the platform’s accessibility and suitability for broad peptide screening applications without the need for complex modeling workflows.

PepPharmaHub outperforms greater generalization capabilities than existing web servers

To assess the real-world generalization capability of PepPharmaHub relative to existing peptide prediction servers, we deployed 24 optimal models comprising 18 BERT-based and 6 RNN-based architectures into the Peptide-Predictor module (http://bioinmed.jflab.ac.cn:18090/peppharmahub/PeptidePredictor/input.jsp) and evaluated their performance on a dataset of 3,475 newly reported bioactive peptide sequences collected between 2023 and 2025 (Table 1). Across 20 independent test tasks, Peptide-Predictor demonstrated a more concentrated and robust prediction accuracy distribution (ranging from 44.35% to 92.37%) compared to the broader and sparser accuracy range observed in existing web predictors (0% to 93.37%) (Fig. 3). Notably, the predictor achieved superior performance in 14 tasks (Fig. 3a-n), with prediction accuracies ranging from 52.38% to 88.89%, surpassing the best existing predictors by a margin of 0.38% to 75%. In particular, it attained accuracies of 81.3% for antimicrobial prediction (Fig. 3d), 85.71% for quorum sensing (Fig. 3f), and 88.89% for antiparasitic prediction (Fig. 3i). In neuropeptide and blood-brain barrier datasets (Fig. 3o–p), the Predictor showed comparable performance, with accuracies of 70.37% and 91.67%. While Matthews correlation coefficient (MCC) values for all 54 models varied broadly (− 0.3068 to 0.6064), 13 PepPharmaHub-based models exhibited relatively high MCCs in the range of 0.0085 to 0.5774, indicating meaningful discriminative power in many cases. The system performed relatively slightly lower in a subset of tasks, including antibacterial, antiviral, bitter and anticancer peptide prediction (Fig. 3q-t). Overall, these results suggest that Peptide-Predictor delivers enhanced functional prediction accuracy for novel peptide sequences compared to current publicly available web servers, highlighting its strong generalization capability and practical utility in peptide discovery workflow.

Rapid screening pipeline and feature mapping for bioactive peptides

To evaluate the interpretability and functional relevance of model predictions, we implemented a rapid screening and feature mapping pipeline using PepPharmaHub. A total of 2,321,342 unlabeled peptide sequences (length 2–20) were retrieved from Sequence Library-1 and screened using the 24 packaged predictive models in the Peptide-Predictor module (task ID: PRE20250616113524). The resulting predictions, visualized through the platform’s analysis interface, included workflow logs, data preprocessing, classification summaries, sequence profiles, and downloadable result packages (Fig. 4a). These outputs were stored in the Pre-selected Library (http://bioinmed.jflab.ac.cn:18090/peppharmahub/Database/SeqLogoPrediction.jsp?PreID=PRE20250616113524) for further inspection. Visualization of sequence feature distributions revealed substantial variation in the predicted abundance of different bioactive peptides—e.g., tumor T cell antigens (1,927,539), antioxidants ( 1,794,986), and quorum sensing peptides (1,674,051) were highly represented, while anti-coronavirus (7,837), anti-MRSA (89,058), and anticancer peptides (341,509) were comparatively rare (Fig. 4b). Amino acid composition analysis indicated both shared and unique residue preferences across peptide types; for instance, Leu was commonly enriched in 21 peptide classes but underrepresented in toxicity and antiparasitic peptides. Antioxidants featured residues such as Leu, Pro, Gly, and Ala, whereas anticancer peptides were enriched in Lys, Leu, Ala, Gly, and Ile. Notably, sequence composition analyses revealed that screened and experimental peptides exhibited concordant N-/C-terminal and key residue profiles (Fig. 4c), suggesting that the models effectively captured discriminative sequence features. To our knowledge, this constitutes one of the most extensive feature maps for bioactive peptide sequences, demonstrating that PepPharmaHub not only enables high-throughput screening but also enhances model transparency and interpretability through comprehensive feature analysis.

Self-trained framework for efficient screening of ACE-inhibitory peptides

To improve predictive accuracy for ACE-inhibitory peptides, we constructed a new benchmark dataset (ACEiPs) and employed the PepPharmaHub platform to develop 27 self-trained models using BiLSTM and fine-tuned BERT architectures via the RNN-Trainer and BERT-Trainer modules. Figure 5a demonstrates the whole screening workflow. Cross-validation on the ACEiPs benchmark dataset showed that 20 models achieved over 80% accuracy, with precision, specificity, sensitivity, MCC, and AUC ranging from 0.7902 to 0.9241, 0.7821 to 0.9250, 0.7904 to 0.9126, 0.6068 to 0.8384, and 0.8490 to 0.9538, respectively. Independent testing on ACEiPs-test confirmed consistent high performance across the same 20 models (Supplementary Table 4). Among all models, the RNN model (RNN20250624111715) encoded with VVSFZL37 demonstrated the best generalization performance, achieving accuracy, precision, specificity, sensitivity, and MCC scores of 0.9169, 0.9088, 0.9070, 0.9268, and 0.8340, respectively (Fig. 5b). When benchmarked against three established ACE prediction servers (mAHTPred[24], ACEiPP [12], and pLM4ACE[25], which includes Logistic Regression, SVM, and Multilayer Perceptron models), the RNN20250624111715 model consistently outperformed across all evaluation metrics, with improvements ranging from 0.0366 to 0.4479 in accuracy and up to 0.896 in MCC (Fig. 5c). Notably, while mAHTPred and ACEiPP showed moderate performance, pLM4ACE performed poorly with accuracies below 0.4845 and negative MCCs.

To demonstrate scalability, the best-performing self-trained model was deployed via the Model-Caller module to screen 3,740,614 peptides from Sequence-Library 2. As shown in Fig. 5c, the predicted ACEiPs have similar residue composition and two-terminal features to experimental ACEiPs, i.e., they are all mainly composed of hydrophobic amino acids (Pro, Leu, Val, Gly, and Ala), the C-terminus tending to hydrophobic (Pro, Phe, and Leu), positively charged (Arg and Lys), and bulky aromatic (Tyr) amino acids, and the N-terminus favoring hydrophobic amino acids (Leu, Val, Ala, Gly, Ile, and Pro). Taken together, these findings validate the self-training framework implemented in PepPharmaHub as a robust and efficient approach for the identification of ACE-inhibitory peptides, demonstrating superior predictive performance, enhanced scalability, and greater biological relevance compared to existing web-based prediction tools.

In this study, we present PepPharmaHub, a comprehensive and scalable web platform that enables end-to-end modeling, screening, sequence feature interpretation, and data sharing for bioactive peptides. By integrating multiple natural language processing (NLP) architectures into a unified, user-friendly web framework including SimpleRNN, LSTM, GRU, BiLSTM, and BERT, the platform supports both platform-provided and self-trained prediction models, covering a wide range of peptide classification tasks without requiring programming expertise or local computational resources. Across 24 benchmark datasets, PepPharmaHub demonstrates superior or comparable performance to existing state-of-the-art models in terms of accuracy, generalization, and throughput. Compared to conventional web servers (such as ACEiPP[12], AnOxPP [14], TransImbAMP [26], BERT4Bitter[27], and SCMRSA[28]) PepPharmaHub overcomes limitations associated with fixed model architectures and static data, enabling continuous model retraining and task-specific customization through its RNN-Trainer, BERT-Trainer, and Model-Caller modules.

Beyond predictive performance, the platform facilitates sequence-level interpretability through feature mapping and comparative analysis between theoretical predictions and experimental peptide profiles. While strong concordance was observed for many peptide classes, notable inconsistencies were detected in specific sequence regions (e.g., N-termini of bitter, blood-brain barrier, and antifungal peptides; C-termini of DPP IV inhibitors), likely reflecting the limited size and diversity of current training datasets. These findings underscore the need for continued dataset expansion and iterative model refinement to improve sequence-level resolution.

Despite its strengths, the current version of PepPharmaHub has two primary limitations. First, computational throughput is constrained by limited GPU availability, which may affect real-time usability during peak demand. Second, the platform currently supports only RNN and BERT based models. While effective, other emerging architectures such as CNN BiLSTM attention hybrids [29], CNN BiLSTM SVM classifiers [30], and BiLSTM multi scale CNNs [31] may offer complementary benefits and warrant integration in future versions. In future, PepPharmaHub will expand computational resources, integrate broader peptide data sources, and support additional modeling paradigms. The platform's low-code interface and modular design aim to democratize access to deep learning-driven peptide discovery, empowering researchers across disciplines to deploy, refine, and interpret advanced models with minimal technical barriers. Ultimately, PepPharmaHub offers a sustainable and extensible solution to accelerate therapeutic peptide research and improve prediction-driven biological discovery.

PepPharmaHub is a comprehensive, user-friendly, and scalable platform that streamlines therapeutic peptide discovery by integrating curated datasets, advanced language modeling, and real-time visualization into a unified workflow. Through its dual-mode high-throughput screening and self-training capabilities, the platform supports both rapid prediction with platform-provided models and flexible customization using user-defined architectures. Benchmarking across 24 public datasets and 20 external test sets demonstrates that PepPharmaHub consistently achieves state-of-the-art or superior performance, while also enhancing interpretability through sequence feature mapping. By lowering technical barriers and enabling reproducible, interpretable, and high-throughput predictions, PepPharmaHub offers a powerful tool for accelerating data-driven peptide drug discovery.

Data collection

This study utilizes 20 types of open-source peptide sequence datasets, comprising a total of 24 datasets, including anti-hypertensive, antioxidant, dipeptidyl peptidase IV inhibition, bitter, umami, antimicrobial, antimalarial, quorum sensing, anticancer, anti-MRSA strains, tumor T cell antigens, blood-brain barrier, antiparasitic, neuropeptide, antibacterial, antifungal, antiviral, toxicity, anti-coronavirus and interleukin-6 inducing activity, to train and evaluate RNN and BERT modeling methods. Table 1 provides an overview of these open-source datasets. The data volume distribution is as follows: >2000 (11 datasets), 1000–2000 (5 datasets), < 1000 (8 datasets), including 15 balanced datasets with an equal number of positive and negative samples, and 9 imbalanced datasets. To ensure a fair comparison with existing methods, the training set, validation set, and independent test set used are consistent with those in the original paper, and the same number of cross-validation folds are applied. Specifically, 20 fair external independent tests were constructed by collecting data on 20 types of bioactive peptides, newly identified in 2023–2025 and not included in the aforementioned 24 datasets (Table 1), to assess the model's generalization ability. The ACEiPs dataset was constructed by deduplicating and balancing the ACEiPP dataset [12], AHTpin dataset [32], and newly reported ACE inhibitory peptides (104 positive samples and 98 negative samples), resulting in a total of 1181 positive and negative samples (Table 1). The ACEiPs dataset was randomly split into a benchmark dataset (ACEiPs_benchmark) and an independent test set (ACEiPs_test) in a 7:3 ratio.

Additionally, two large peptide sequence datasets with unknown activities were constructed to assess the model’s ability to pre-screen features. The first dataset, named Sequence-Library 1, contains a total of 3,768,340 peptide sequences retrieved from the UniProt database using the search term ‘(length:[2 TO 50])’. The second dataset, named Sequence-Library 2, consists of 2,321,342 non-redundant peptide sequences of length 2–20 from 21,249 food-derived proteins, downloaded from the ACEiPP database [12].

Recurrent neural network modelling module

The peptide sequence, based on the single-letter codes of the 20 natural amino acids (R, K, N, D, Q, E, H, P, Y, W, S, T, G, A, M, C, F, L, V, I), can be represented as:

Peptide = [A₁, A₂, ..., A_i] (1)

where A_i is the i-th amino acid in the peptide sequence, i.e., the positional index of the amino acid residue. And then, peptide sequences were transformed into feature vectors using One-hot coding and 23 sets of amino acid descriptors (AADs) to represent the types and physicochemical properties of sequence residues (Table S1). By referencing these AADs, peptide sequences can be transformed from non-numeric sequence data into numeric feature vectors, which are then input into recurrent neural networks for model training. The AADs coding is defined as follows:

$$\:\text{A}\text{A}\text{D}\text{s}=\left[\begin{array}{cccc}{\text{V}}_{1}^{1}&\:{\text{V}}_{2}^{1}&\:\cdots\:&\:{\text{V}}_{n}^{1}\\\:{\text{V}}_{1}^{2}&\:{\text{V}}_{2}^{2}&\:\cdots\:&\:{\text{V}}_{\text{n}}^{2}\\\:⋮&\:⋮&\:⋮&\:⋮\\\:{\text{V}}_{1}^{20}&\:{\text{V}}_{2}^{20}&\:\cdots\:&\:{\text{V}}_{\text{n}}^{20}\end{array}\right]$$

where V represents the feature variables of the 20 natural amino acids, n is the number of variables per residue, and the AADs coding matrix has dimensions of 20×n.

The RNN-Trainer module was developed based on the TensorFlow framework and incorporates four recurrent neural networks: SimpleRNN, LSTM, GRU, and BiLSTM. For each of these networks, 24 amino acid encoding schemes (Table S1), 19 activation functions, N-fold cross-validation (N = 5, 10, 15, 20), and other training parameters (network layers, neurons, learning rate, dropout, nEpochs, early stopping and checkpoint) are provided.

Table 1
Open-source benchmark and test datasets collected from various sources in the literature and platform
Bioactivity	Dataset reference	Training dataset		Test dataset		Newly reported peptides
Bioactivity	Dataset reference	Positive	Negative	Positive	Negative	Positive	Negative
Antihypertensive activity	mAHTPred [24]	913	913	386	386	102	93
	ACEiPP[12]	730	730	313	313
	ACEiPs (this study)	826	826	355	355
Antioxidant activity	AnOxPP [14]	848	848	212	212	112	42
Antioxidant activity	AnOxPePred-FRS [33]	530	593	146	135	112	42
DPP IV inhibitory activity	iDPPIV-SCM [34]	532	532	133	133	109	75
Bitter	BERT4Bitter [27]	256	256	64	64	179	184
Umami	iUMAMI-SCM [35]	112	241	28	61	244	19
Antimicrobial activity	TransImbAMP [26]	3876	9552	2584	6369	376	9
Antimalarial activity	iAMAP-SCM (Main dataset) [36]	111	1708	28	427	17	4
Antimalarial activity	iAMAP-SCM (Alternative dataset) [36]	111	542	28	135	17	4
Quorum sensing activity	QSPred-FL [37]	200	200	20	20	7	0
Anticancer activity	AntiCP 2.0 (Main dataset) [38]	689	689	172	172	570	71
Anticancer activity	AntiCP 2.0 (Alternative dataset) [38]	776	776	194	194	570	71
Anti-MRSA strains activity	SCMRSA [28]	118	678	30	169	12	1
Tumor T cell antigens	iTTCA-Hybrid [39]	470	318	122	75	122	75
Blood-Brain Barrier	BBPpred [40]	100	100	19	19	45	3
Antiparasitic activity	PredAPP [41]	255	1863	46	46	23	4
Neuropeptide	NeuroPred-CLQ [42]	1940	1940	485	485	9	18
Antibacterial activity	starPep_AB [43]	6583	6583	1695	1695	226	27
Antifungal activity	starPep_AF [43]	778	778	215	215	117	21
Antiviral activity	starPep_AV [43]	2321	2321	623	623	31	467
Toxicity	ATSE [44]	1663	1621	290	290	22	6
Anti-coronavirus activity	FEOpti-ACVP [18]	125	1587	32	397	63	38
Interleukin-6 inducing activity	StackIL6[47]	292	2393	73	597	2	1

BERT pre-training fine-tuning module

A total of 556,603 protein sequences were downloaded from UniProt as the peptide-related pre-training corpus. UniProt, which integrates data from SWISS-PROT, TrEMBL, and UniParc, is the largest and most comprehensive protein database, providing ample data for model pre-training. Based on the extensive discussion in the literature regarding the relationship between peptide sequences and activity, amino acid residues and specific motifs (dipeptides and tripeptides) are key determinants of the activity of therapeutic peptides [45, 46]. We divide proteins into motifs, where every k (k = 1, 2, 3) residues in the sequence are grouped into k-mers. When fewer than k amino acids remain at the end of the sequence, the remaining amino acids are grouped together [13]. The computational formula of the BERT model is as follows:

MultiHead (Q, K, V) = Concat (head₁, …, head_h) (3)

Head_i = Attention ($\:{\text{Q}\text{W}}_{\text{i}}^{\text{Q}}$, $\:{\text{K}\text{W}}_{\text{i}}^{\text{K}}$, $\:{\text{V}\text{W}}_{\text{i}}^{\text{V}}$) (4)

Attention (Q, K, V) = softmax($\:\frac{{QK}^{T}}{\sqrt{{d}_{k}}}$)V (5)

where d_k is the dimension of the Key, $\:{\text{W}}_{\text{i}}^{\text{Q}}$, $\:{\text{W}}_{\text{i}}^{\text{K}}$, $\:{\text{W}}_{\text{i}}^{\text{V}}$ and $\:{\text{W}}^{\text{O}}$ are the parameter matrices.

Three pre-trained k-mer models were trained based on the TensorFlow framework for model fine-tuning. The developed BERT-Trainer module provides 7 standard configuration parameters: Batch size, Evaluate, Train epochs, Warmup proportion, Learning rate, Classification, and Cross-validation (supporting 5-, 10-, 15-, and 20-fold), along with 11 advanced configuration parameters: Attention probability, Dropout probability, 5 Hidden layer activation functions, Hidden layer dropout probability, Hidden layer size, Initializer range, Intermediate layer size, Maximum position embeddings, Number of attention heads, Number of hidden layers, Type vocabulary size, and 3 Vocabulary sizes (k-mers). Furthermore, Early stopping and Checkpoint parameters are designed to monitor the model's performance and systematically save its state during training.

Evaluation criteria

The prediction ability of the models in n-fold cross-validation and external testing is evaluated using six evaluation parameters: Precision, Sensitivity, Specificity, Accuracy, Matthew’s correlation coefficient (MCC), and the area under the receiver operating characteristics curves (AUC). Their definitions are as follows:

$$\:Precision=\frac{TP}{TP+FP}$$

$$\:Sensitivity=\frac{TP}{TP+FN}$$

$$\:Specificity=\frac{TN}{TN+FP}$$

$$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

$$\:MCC=\frac{TP\times\:TN-FP\times\:FN}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}$$

where TP, FP, TN, and FN represent the number of true positive samples, false positive samples, true negative samples, and false negative samples, respectively.

Author contributions

DY.Q. and Z.W. conceptualized and supervised the project. DY.Q. constructed PepPharmaHub web server, built the project website, and created online tutorial. DY.Q., Z.W., K.S. and H.F. wrote the manuscript and curated all figures. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Science and Technology Innovation Key R&D Program of Chongqing [CSTB2023TIAD-STX0001], National Natural Science Foundation of China [82470220, 32470681], National Key R&D Program of China [2022YFA1103300], Chongqing Municipal Science and Health Joint Medical Research Project [2025ZDXM004], Chongqing Municipal PhD Fast-Track Program [CSTB2024NSCQ-BSX0019], Youth Talent Development Program from Second Affiliated Hospital, Army Medical University [2022YQB014].

Data availability

The 25 benchmark datasets (Table 1), and 19 independent test sets are available in the Dataset Library of PepPharmaHub (http://bioinmed.jflab.ac.cn:18090/peppharmahub/
Database/datasetLibrary.jsp). Sequence Library-1 and Sequence Library-2 can be downloaded at http://bioinmed.jflab.ac.cn:18090/peppharmahub/Database/seqLibrary
.jsp. Screening results for 23 peptide types are accessible via the Pre-selected Library, with data downloads and sequence profile visualization at http://bioinmed.
jflab.ac.cn:18090/peppharmahub/Database/SeqLogoPrediction.jsp?PreID=PRE20241227191128. All model and task IDs mentioned in this study can be retrieved via the Home page search function (http://bioinmed.jflab.ac.cn:18090/pepnlp/index.jsp). Source data are provided with this paper.

Code availability

The developed PepPharmaHub platform is now available and can be accessed at: http://bioinmed.jflab.ac.cn:18090/pepnlp/index.jsp. Future updates and new versions will also be released through this link.

Competing interests

The authors declare no competing interests.

Rossino, G., et al., Peptides as Therapeutic Agents: Challenges and Opportunities in the Green Transition Era. Molecules, 2023. 28(20).
Li, C.M., et al., Novel Peptide Therapeutic Approaches for Cancer Treatment. Cells, 2021. 10(11).
D’Aloisio, V., et al., PepTherDia: database and structural composition analysis of approved peptide therapeutics and diagnostics. Drug Discovery Today, 2021. 26(6): p. 1409-1419.
Qin, D., et al., DFBP: a comprehensive database of food-derived bioactive peptides for peptidomics research. Bioinformatics, 2022. 38(12): p. 3275-3280.
Drucker, D.J., Advances in oral peptide therapeutics. Nature Reviews Drug Discovery, 2019. 19(4): p. 277-289.
Muttenthaler, M., et al., Trends in peptide drug discovery. Nature Reviews Drug Discovery, 2021. 20(4): p. 309-325.
Sharma, K., et al., Peptide-based drug discovery: Current status and recent advances. Drug Discovery Today, 2023. 28(2).
Mann, M., et al., Artificial intelligence for proteomics and biomarker discovery. Cell Systems, 2021. 12(8): p. 759-770.
Jia, W., et al., Exploring novel ANGICon-EIPs through ameliorated peptidomics techniques: Can deep learning strategies as a core breakthrough in peptide structure and function prediction? Food Research International, 2023. 174.
Wu, X., et al., Deep learning for advancing peptide drug development: Tools and methods in structure prediction and design. European Journal of Medicinal Chemistry, 2024. 268.
Ghafoor, H., et al., CAPTURE: Comprehensive anti-cancer peptide predictor with a unique amino acid sequence encoder. Computers in Biology and Medicine, 2024. 176.
Qin, D., et al., Sequence–Activity Relationship of Angiotensin-Converting Enzyme Inhibitory Peptides Derived from Food Proteins, Based on a New Deep Learning Model. Foods, 2024. 13(22).
Zhang, Y., et al., A novel antibacterial peptide recognition algorithm based on BERT. Briefings in Bioinformatics, 2021. 22(6).
Qin, D., et al., Prediction of antioxidant peptides using a quantitative structure−activity relationship predictor (AnOxPP) based on bidirectional long short-term memory neural network and interpretable amino acid descriptors. Computers in Biology and Medicine, 2023. 154.
Naseem, A., et al., BBB-PEP-prediction: improved computational model for identification of blood–brain barrier peptides using blending position relative composition specific features and ensemble modeling. Journal of Cheminformatics, 2023. 15(1).
Wang, J.-H. and T.-Y. Sung, ToxTeller: Predicting Peptide Toxicity Using Four Different Machine Learning Approaches. ACS Omega, 2024. 9(29): p. 32116-32123.
Wang, D., et al., StructuralDPPIV: a novel deep learning model based on atom structure for predicting dipeptidyl peptidase-IV inhibitory peptides. Bioinformatics, 2024. 40(2).
Jiang, J., et al., FEOpti-ACVP: identification of novel anti-coronavirus peptide sequences based on feature engineering and optimization. Briefings in Bioinformatics, 2024. 25(2).
Zhou, P., et al., Systematic Comparison and Comprehensive Evaluation of 80 Amino Acid Descriptors in Peptide QSAR Modeling. Journal of Chemical Information and Modeling, 2021. 61(4): p. 1718-1731.
Zhang, J., et al., Umami-BERT: An interpretable BERT-based model for umami peptides prediction. Food Research International, 2023. 172.
Klepach, A., et al., Characterization and impact of peptide physicochemical properties on oral and subcutaneous delivery. Advanced Drug Delivery Reviews, 2022. 186.
Kurata, H., S. Tsukiyama, and B. Manavalan, iACVP: markedly enhanced identification of anti-coronavirus peptides using a dataset-specific word2vec model. Briefings in Bioinformatics, 2022. 23(4).
Minkiewicz, P., A. Iwaniak, and M. Darewicz, Annotation of Peptide Structures Using SMILES and Other Chemical Codes–Practical Solutions. Molecules, 2017. 22(12).
Manavalan, B., et al., mAHTPred: a sequence-based meta-predictor for improving the prediction of anti-hypertensive peptides using effective feature representation. Bioinformatics, 2019. 35(16): p. 2757-2765.
Du, Z., et al., pLM4ACE: A protein language model based predictor for antihypertensive peptide screening. Food Chemistry, 2024. 431.
Pang, Y., et al., Integrating transformer and imbalanced multi-label learning to identify antimicrobial peptides and their functional activities. Bioinformatics, 2022. 38(24): p. 5368-5374.
Charoenkwan, P., et al., BERT4Bitter: a bidirectional encoder representations from transformers (BERT)-based model for improving the prediction of bitter peptides. Bioinformatics, 2021. 37(17): p. 2556-2562.
Charoenkwan, P., et al., SCMRSA: a New Approach for Identifying and Analyzing Anti-MRSA Peptides Using Estimated Propensity Scores of Dipeptides. ACS Omega, 2022. 7(36): p. 32653-32664.
Xing, W., et al., iAMP-Attenpred: a novel antimicrobial peptide predictor based on BERT feature extraction method and CNN-BiLSTM-Attention combination model. Briefings in Bioinformatics, 2024. 25(1).
Xiao, X., et al., iAMP-CA2L: a new CNN-BiLSTM-SVM classifier based on cellular automata image for identifying antimicrobial peptides and their functional types. Briefings in Bioinformatics, 2021. 22(6).
Qi, Y., P. Zheng, and G. Huang, DeepLBCEPred: A Bi-LSTM and multi-scale CNN-based deep learning method for predicting linear B-cell epitopes. Frontiers in Microbiology, 2023. 14.
Kumar, R., et al., An in silico platform for predicting, screening and designing of antihypertensive peptides. Scientific Reports, 2015. 5(1).
Olsen, T.H., et al., AnOxPePred: using deep learning for the prediction of antioxidative properties of peptides. Scientific Reports, 2020. 10(1).
Charoenkwan, P., et al., iDPPIV-SCM: A Sequence-Based Predictor for Identifying and Analyzing Dipeptidyl Peptidase IV (DPP-IV) Inhibitory Peptides Using a Scoring Card Method. Journal of Proteome Research, 2020. 19(10): p. 4125-4136.
Charoenkwan, P., et al., iUmami-SCM: A Novel Sequence-Based Predictor for Prediction and Analysis of Umami Peptides Using a Scoring Card Method with Propensity Scores of Dipeptides. Journal of Chemical Information and Modeling, 2020. 60(12): p. 6666-6678.
Charoenkwan, P., et al., iAMAP-SCM: A Novel Computational Tool for Large-Scale Identification of Antimalarial Peptides Using Estimated Propensity Scores of Dipeptides. ACS Omega, 2022. 7(45): p. 41082-41095.
Wei, L., et al., Comparative analysis and prediction of quorum-sensing peptides using feature representation learning and machine learning algorithms. Briefings in Bioinformatics, 2018.
Agrawal, P., et al., AntiCP 2.0: an updated model for predicting anticancer peptides. Briefings in Bioinformatics, 2021. 22(3).
Charoenkwan, P., et al., iTTCA-Hybrid: Improved and robust identification of tumor T cell antigens by utilizing hybrid feature representation. Analytical Biochemistry, 2020. 599.
Dai, R., et al., BBPpred: Sequence-Based Prediction of Blood-Brain Barrier Peptides with Feature Representation Learning and Logistic Regression. Journal of Chemical Information and Modeling, 2021. 61(1): p. 525-534.
Zhang, W., et al., PredAPP: Predicting Anti-Parasitic Peptides with Undersampling and Ensemble Approaches. Interdisciplinary Sciences: Computational Life Sciences, 2021. 14(1): p. 258-268.
Bin, Y., et al., Prediction of Neuropeptides from Sequence Information Using Ensemble Classifier and Hybrid Features. Journal of Proteome Research, 2020. 19(9): p. 3732-3740.
Pinacho-Castellanos, S.A., et al., Alignment-Free Antimicrobial Peptide Predictors: Improving Performance by a Thorough Analysis of the Largest Available Data Set. Journal of Chemical Information and Modeling, 2021. 61(6): p. 3141-3157.
Wei, L., et al., ATSE: a peptide toxicity predictor by exploiting structural and evolutionary information based on graph neural network and attention mechanism. Briefings in Bioinformatics, 2021. 22(5).
Wu, J., R.E. Aluko, and S. Nakai, Structural Requirements of Angiotensin I-Converting Enzyme Inhibitory Peptides: Quantitative Structure−Activity Relationship Study of Di- and Tripeptides. Journal of Agricultural and Food Chemistry, 2006. 54(3): p. 732-738.
Du, A. and W. Jia, New insights into the bioaccessibility and metabolic fates of short-chain bioactive peptides in goat milk using the INFOGEST static digestion model and an improved data acquisition strategy. Food Research International, 2023. 169.
Charoenkwan P, Chiangjong W, Nantasenamat C, Hasan MM, Manavalan B, Shoombuatong W. StackIL6: a stacking ensemble model for improving the prediction of IL-6 inducing peptides. Brief Bioinform. 2021;22(6):bbab172.

No competing interests reported.

SupplementaryMaterial.docx

Download PDF

Editorial decision: Revision requested
27 Nov, 2025
Reviews received at journal
02 Nov, 2025
Reviews received at journal
17 Oct, 2025
Reviews received at journal
16 Oct, 2025
Reviews received at journal
06 Oct, 2025
Reviewers agreed at journal
26 Sep, 2025
Reviewers agreed at journal
26 Sep, 2025
Reviewers agreed at journal
26 Sep, 2025
Reviewers agreed at journal
26 Sep, 2025
Reviewers invited by journal
26 Sep, 2025
Editor assigned by journal
19 Sep, 2025
Submission checks completed at journal
18 Sep, 2025
First submitted to journal
18 Sep, 2025

You are reading this latest preprint version

PepPharmaHub: A Cloud-Based Platform Integrating Multimodel Language Architectures with Curated Data Resources for Therapeutic Peptide Discovery

Status:

Version 1

Abstract

Background

Results

Conclusions

Figures

Background

Results

PepPharmaHub facilitates one-stop screening for therapeutic peptides

(1) Built-in centralized and structured peptide data resources

(2) Deep Learning and Dual-Mode Screening

(3) Web-based implementation for interactive visualization

PepPharmaHub demonstrates robust generalization and adaptability across open benchmark datasets

PepPharmaHub outperforms greater generalization capabilities than existing web servers

Rapid screening pipeline and feature mapping for bioactive peptides

Self-trained framework for efficient screening of ACE-inhibitory peptides

Discussion

Conclusions

Methods

Data collection

Recurrent neural network modelling module

BERT pre-training fine-tuning module

Evaluation criteria

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1