Our laboratory has constructed DRAMP, a comprehensive open-access database of AMPs dedicated to accelerating the clinical translation of AMPs. The version 4.0 contains 30,260 entries (8,001 new additions compared to the previous version), with core innovations and data integration including: the first introduction of a stability dataset, 110 newly added verified serum/protease stability and half-life data; the expansion of clinical entries to 96, integrating repetitive sequences and associating clinical trial numbers; 2,891 new hemolytic activity and 2,674 new cytotoxicity experimental data, systematically marking toxicity bottlenecks; the addition of an “Expanded AMPs” subset (179 entries) including experiments verified by Chinese journals not indexed in international databases; the number of experimentally verified structures (PDB) increased to 419, and the number of predicted structures reached 1,570. The database provides categorized downloads, covering 70.47% of unique sequences (18,580 entries not found in APD/DBAASP/CAMP). The database can be accessed at: http://dramp.cpu-bioinfor.org/ [52–55].

dbAMP is dedicated to addressing the increasingly severe AMR challenges in the post-pandemic era. This database has significantly expanded its data scale and functional depth through continuous updates. The latest version (dbAMP 3.0, 2025) integrates data from over 5,200 scientific papers, including 33,065 AMPs and 2,453 antimicrobial proteins, covering 3,534 source organisms, representing a nearly threefold increase from the initial version in 2019 (12,389 entries). The core breakthrough lies in the use of ESMFold to predict the three-dimensional structures of over 30,000 AMPs and to provide contact maps and secondary structure information, which greatly facilitates structural-function research. To accelerate clinical translation, the platform has added an AMP rational design pipeline, using multi-objective optimization algorithms (balancing antimicrobial activity, low toxicity, and stability) to assist in the development of new antibacterial peptides. At the same time, dbAMP integrates over 20 prediction tools and benchmark data set portals, becoming the core hub for AMP discovery, mechanism research, and clinical design. The database can be accessed at: https://awi.cuhk.edu.cn/dbAMP/ [50, 51].

DBAASP is currently the most comprehensive resource library for antimicrobial and cytotoxic peptides, containing over 15,700 entries, including > 14,500 monomers and nearly 400 homo- and hetero-multimers. Its core advantage lies in integrating multi-dimensional data, covering more than 2,700 ribosomal synthesis peptides, over 170 non-ribosomal peptides, and more than 12,000 synthetic peptides, including various modifications such as linear, cyclic, and disulfide bonds; providing experimental activity data for 4,200+ targets [such as minimum inhibitory concentration (MIC)/IC₅₀], as well as hemolysis/cytotoxicity indicators; generating 3,200+ peptide conformational trajectories through an automated MD process and correlating with 400+ PDB experimental structures, supporting online visualization and analysis; the newly added PAASS tool can predict the activity of peptides against specific pathogens, and expands the calculation of physicochemical properties. As a key platform for antibacterial peptide design, DBAASP v3 promotes the rational development of new anti-infective peptides through high-precision experimental data, dynamic structure simulation, and ML tools. The access address is: http://dbaasp.org [8, 56].

AVPdb is a specialized database dedicated to experimental verification of antiviral peptides (AVPs), containing 2,683 peptide sequences targeting over 60 medically significant viruses (such as influenza, HCV, HSV, etc.), including 624 chemically modified peptide segments. This database provides detailed experimental information for each peptide, including sequence, source, target virus, inhibitory activity, mechanism of action, cell line, detection method, and complete PubMed literature reference. Additionally, AVPdb integrates various analysis tools, such as BLAST alignment, physicochemical property calculation, and structure prediction, enabling users to browse, search, and download data. This resource aims to promote AVP research and drug development, and is publicly accessible at: http://crdd.osdd.net/servers/avpdb [67].

PlantPepDB is a comprehensive database focusing on plant-derived active peptides, which includes 3,848 peptide sequences from 443 plant species, covering 9 major functional categories, such as antimicrobial, anti-cancer, antiviral, anti-hypertension, immune regulation, and many other biological activities. This database integrates data from 11 existing databases and 835 research papers. Each peptide entry provides detailed sequence, source plant, functional description, experimental verification level (protein level, transcriptional level, prediction or homology inference), physicochemical properties, tertiary structure prediction, and dictionary of protein secondary structure (DSSP) secondary structure report. The database offers multiple search methods and integrates analysis tools such as BLAST, Smith-Waterman alignment, and peptide mapping, enabling users to perform function prediction, structure analysis, and target screening. The database also places particular emphasis on data deduplication and information integration, merging duplicate peptide entries from different sources into a single record with more comprehensive information. The database is publicly accessible at: http://www.nipgr.ac.in/PlantPepDB/ [65].

DCTPep is an open-source comprehensive database specifically designed for cancer therapeutic peptides. It contains a total of 6,214 entries, including 6,106 peptide sequences and 108 peptide drugs. This database not only encompasses traditional anti-cancer peptides (ACPs) but also integrates information on cell-penetrating peptides with targeting functions, cancer-targeting peptides, and peptide-drug conjugates. It provides detailed annotations such as activity, targets, physicochemical properties, and three-dimensional structures predicted by AlphaFold. The database covers data on over 60 targets and more than 400 cancer cell lines. Its purpose is to offer valuable resources for the design, screening, and mechanism-of-action research of ACPs. The database is publicly accessible at: http://dctpep.cpu-bioinfor.org/ [70].

The basic models mainly include DNN, convolutional neural networks (CNN), recurrent neural networks (RNN), their variants (LSTM, BiLSTM, GRU, etc.), and other hybrid models. CNN is adept at extracting local conserved motifs using multi-scale convolution and can effectively capture local sequence patterns, making it suitable for sequence encoding; DNN, which is a multi-layer fully connected network, is suitable for the fusion and classification of various manual features (such as amino acid composition, dipeptide composition, physical and chemical properties), and achieves this by globally optimizing the network parameters through minimizing the selected loss function. This is similar to CNN. Many works have implemented AMP prediction. For instance, Ma et al. [89] employed DL tools, including DNN, to mine and predict novel AMPs from human gut metagenomic data on a large scale. Subsequent experimental validation confirmed the antimicrobial activity of over a hundred of these predicted peptides, demonstrating the powerful application of DNN models in the field of AMP discovery [89]. RNN captures long-range dependencies between amino acids through a recurrent mechanism, but has a slower training speed and requires a large amount of data. It is often cascaded with CNN to form the CNN-BiLSTM architecture, and is supplemented by transfer learning, multi-task learning, or stacking integration to address the issues of insufficient data and model robustness [90]. DL variants (such as CNN, LSTM, DNN) simulate the processing of complex patterns by human brain neurons. In 2022, the research integrated LSTM, Attention, BERT, and other natural language processing (NLP) models. The prediction accuracy of AMP exceeded 91%, and the AUPRC reached 0.9244. Moreover, it does not require sequence homology or amino acid composition information and can discover new peptides with a homology lower than 40%, significantly reducing false positive (FP) [89]. TriNet further integrates sequence fingerprints, position-specific scoring matrix (PSSM) evolutionary information, and physicochemical properties with a three-layer network. It introduces the training-validation interaction (TVI) mechanism and supports hyperparameter tuning to enhance the generalization ability of ACP/AMP predictions [84].

Currently, such prediction models generally encounter the following common bottlenecks: Firstly, the models are highly dependent on known databases for training, which limits their generalization ability and makes it difficult for them to identify sequences with low homology and novel categories. Moreover, the training data is mostly concentrated on short peptides, resulting in poor prediction performance for long sequences. Secondly, most models only learn based on amino acid sequences and lack the integration of key physical and chemical information, such as protein spatial conformation, dynamic behavior, and interactions. Finally, the selected candidate molecules often have toxicity issues, and the experimental verification cost is high, which limits their efficient transformation from computational prediction to practical application [84, 89].

The role of the language model is to use large-scale unsupervised pre-training to transform amino acid sequences into semantic-rich context vectors, and then perform lightweight fine-tuning to achieve performance even surpassing that of traditional feature engineering. This includes BERT and its variants, pLMs, and peptide-specific lightweight models, etc. BERT and its variants obtain context-sensitive sequence embeddings through large-scale masked language modeling and encode and extract features for AMPs [49, 91], significantly reducing the need for labeled data and improving prediction performance, and can further fine-tune the input for other models to adapt to AMP classification, functional annotation, and even design tasks. pLMs are inspired by BERT and capture complex dependencies in protein sequences through self-supervised learning and multi-task pre-training, such as Diff-AMP [92] and AMP-BERT [93], etc. In 2023, Lobo et al. [94] used a concise and transferable “pre-training and fine-tuning” pipeline for rapid prediction of the antifungal activity of peptide segments, and ultimately found that the combination of SeqVec + k-Best feature selector + support vector classification (SVC) had the best effect. Peptide language models draw on the idea of pre-trained language models in NLP, are specifically designed for protein or peptide sequences, but need to be combined with structural biology and efficient algorithms to solve the key bottlenecks in the design of AMPs [88].

In recent years, the Transformer neural network architecture has demonstrated significant application value in the field of bioinformatics. Originally proposed by Vaswani et al. [95], its core self-attention mechanism effectively captures long-range dependencies in sequences, providing new technical approaches for protein and peptide sequence analysis.

In protein sequence representation learning, the evolutionary scale modeling (ESM)-2 model developed by Lin et al. [96] employs self-supervised pre-training strategies to learn semantically rich embeddings from large-scale protein sequence data. Similarly, the ProtT5 model proposed by Elnaggar et al. [97] utilizes advanced pre-training methods and has achieved excellent performance in various protein-related tasks.

In the field of drug discovery, the Transformer architecture also plays an important role. The PepMLM model developed by Chen et al. [98] employs masked language modeling and fine-tuning techniques to enable effective prediction and design of peptide sequence functions. This model can adapt to specific biomedical application scenarios with minimal labeled data, demonstrating good practicality and generalization value. However, such models often encounter the problem of a limited experimental verification scope during the experiment. The selection process may still have a relatively high FP rate. Leveraging features from larger-scale sequence datasets can improve AMP classification performance, but when it comes to model size, bigger isn't always better. Moreover, incorporating structural information can greatly enhance screening efficiency [98, 99].

The core idea of GNN in the prediction of AMPs is: treating the atoms/residues/fragments of the peptide as nodes, and establishing edges using chemical bonds, field atomic information, or spatial adjacency relationships to construct a graph structure, converting the sequence into a structural graph. GNN aggregates neighbor information through message passing to simultaneously capture sequence order and three-dimensional/topological features, featuring direct peptide graph representation, hierarchical graph representation, and alternative representations leveraging DL. It can incorporate physical-chemical properties (charge, hydrophobicity) as node or edge features, but the structural prediction error will gradually increase with each level or due to reasons such as over-smoothing [88, 100, 101]. For example, graph neural network models such as PNAbind can be used for high-throughput annotation of nucleic acid binding functions of unknown proteins, which can complete local prediction at the level of a single residue and global prediction at the level of the whole protein complex. As a structural and functional characterization tool, PNAbind is suitable for functional analysis of various proteins [102].

Although graph neural networks can effectively model the structural features and spatial relationships of peptide sequences in AMP prediction, their application still faces certain limitations: Firstly, the model performance is highly dependent on the prediction accuracy of the input structure, which requires a large number of high-quality experimental resolved structures. However, the AMP peptide sequences often provide enough information to construct complex graph structures, resulting in unstable performance and low accuracy. Secondly, it is difficult for graph models to integrate non-graphical information (such as evolutionary spectrum, chemical modification, etc.), which limits their ability to model multi-dimensional features. In addition, although its inference speed is fast, the training process has certain requirements on computational resources, which may limit its wide application in resource-limited environments. Therefore, the advantages of graph models in AMP prediction have not been fully exploited, and it is still necessary to combine multimodal information (such as language models) for optimization and fusion [88, 102].

The mixed and multimodal model achieves information complementation within a unified framework by jointly employing sequence embeddings, predictive structural graphs, evolutionary spectra, and physical-chemical properties, and aligning multi-source features through attention or contrastive learning. For example, APEX screened out 37,176 broad-spectrum active peptides from extinct species (11,035 of which only appeared in extinct species), verified through 69 synthetic experiments, 59% of which were active, and the in vivo efficacy was comparable to polymyxin B, proving that multimodal DL combined with molecular de-extinction is an efficient path for discovering new antibiotics [103]; UniproLcad is an AMP predictor that integrates a multi-pLM. It concatenates the sequence representations of three pre-trained models, ESM-2, ProtBert, and UniRep, and then uses BiLSTM, 1D-CNN, and Attention to refine the features, and finally uses multi-layer perceptron (MLP) for classification. It significantly outperforms existing methods [104].

The wide application of such models still faces a series of common problems. First of all, such models usually rely on large-scale labeled datasets for training. When the size of the dataset is insufficient and the source is single, it is difficult to truly reflect the robustness of the model in diverse scenarios. However, the heterogeneity, unbalanced length distribution, and potential redundancy of biological data may lead to the model learning the specific deviation of the dataset instead of the universal biological law. Secondly, multimodal fusion often introduces multiple deep subnetworks and complex interaction modules to form a highly nonlinear “black box” system. Although the prediction accuracy is improved, it is difficult to explain the decision basis from the biological perspective. Even with the introduction of attention mechanisms, most studies still stay at the visualization level and lack deep integration with experimental verification, which leads to doubts about the biological significance of attention weight. Thirdly, it is common practice to send the outputs of each modal encoder to the prediction layer after simple splicing or weighted summation. This late fusion method fails to fully model the complex intersection between modes, which limits the ability of the model to describe the internal mechanism of the biological system. In addition, most hybrid multimodal models do not take into account the dynamics and physicochemical details of biological systems, such as protein conformational changes, water molecule-mediated interactions, ligand flexibility, and other key factors, resulting in a gap between the prediction results and the real physiological environment [44, 103, 104].

Overall, these four models have their own focuses: Basic models are lightweight and efficient, language models are semantically rich, graph models have precise structures, and multimodal fusion balances performance and generalization, collectively forming the current technical panorama of AMP prediction. DL has four main technical routes in AMP prediction: CNN/RNN basic models, language models, graph neural networks, and multimodal models.

The foundation of modern LLMs is the Transformer architecture, introduced by Vaswani et al. [95], which utilizes a self-attention mechanism to weigh the importance of different parts of the input sequence, effectively capturing global dependencies. This architecture has been adapted for proteins, leading to the development of specialized “pLMs”. These models are pre-trained on massive datasets of protein sequences (e.g., from UniRef or PDB) to learn the underlying “grammar” of protein sequences.

Several prominent pLMs have been developed and applied to AMP discovery. For instance, the ESM family of models, such as ESM-1b and ESM-2, developed by Lin et al. [96], uses self-supervised learning to generate rich embeddings that capture biological and structural properties from sequence alone [106]. Similarly, the ProtTrans family, including ProtT5 proposed by Elnaggar et al. [97], leverages Transformer architectures and has achieved state-of-the-art performance in various protein prediction tasks by providing powerful feature representations.

pLMs serve as powerful feature extractors for downstream tasks like AMP prediction. Their pre-trained knowledge can be leveraged through fine-tuning or by using their embeddings as input to simpler classifiers. This approach has proven highly effective, especially in data-sparse scenarios common in peptide research [105].

For example, HMD-AMP, developed by Yu et al. [106], utilizes the ESM-1b pLM combined with a deep forest classifier to accurately identify AMP functions across 11 target types, demonstrating robustness even with small sample sizes. Another model, AMP-BERT, proposed by Lee et al. [93], fine-tunes a BERT-based model specifically for AMP function prediction, showcasing the adaptability of the Transformer architecture. PeptideBERT, introduced by Guntuboina et al. [107], is a lightweight Transformer-based language model specifically fine-tuned for peptide property prediction, including antimicrobial activity. It demonstrates that even smaller, specialized models can achieve high performance by leveraging the power of the Transformer architecture [107]. Furthermore, Lobo et al. [94] successfully used a “pre-training and fine-tuning” pipeline with the SeqVec pLMs to rapidly and accurately predict antifungal peptides, outperforming traditional methods.

A more recent advancement is the integration of LLMs with other architectures to create hybrid models. PepGraphormer, developed by Lin et al. [108], fuses the semantic power of the ESM-2 language model with a graph attention network (GAT). This hybrid framework constructs a heterogeneous graph containing peptide and amino acid nodes, allowing it to capture both global semantic understanding and local structural relationships for high-precision AMP identification without requiring explicit 3D structures [108].

Beyond prediction, LLMs and deep generative models are increasingly used for the de novo design of novel AMPs. These models learn the distribution of functional peptide sequences and can generate new ones with desired properties. For instance, deepAMP, introduced by Li et al. [109], is a deep generation framework based on peptide language models. It uses a pre-training and fine-tuning strategy to optimize peptides from low-activity to high-activity and broad-spectrum candidates, with subsequent experimental validation confirming its efficiency [109]. Similarly, BroadAMP-GPT, proposed by Li et al. [110], combines CNN and LSTM architectures for AI-driven design and multi-model screening, achieving high experimental hit rates. In the same type, BatGPT-Chem proposed by Yang et al. [111], based on the learning of massive chemical data, establishes the mapping relationship between molecular properties and molecular structures represented by SMILES, and the SMILES code is used as a “special language” to carry out unified sequence modeling with natural language. The peptide molecular codes that meet the requirements of these properties are directly output, so as to complete the conversion from “property description” to “peptide structure generation” [111].

Other generative models, while not strictly LLMs, also play a crucial role. HydrAMP, a conditional variational autoencoder (cVAE) developed by Szymczak et al. [112], creates a continuous latent space for peptide sequences, enabling both the optimization of existing peptides and the generation of entirely new, highly potent AMPs. Multi-CGAN, introduced by Yu et al. [113], uses a conditional generative adversarial network to design peptides with multiple desired attributes, such as high activity, low toxicity, and specific secondary structures.

The application of LLMs represents a paradigm shift in AMP discovery, moving towards more intelligent and data-efficient design [5]. By treating sequences as a language, these models can uncover hidden patterns and design principles that are difficult to capture with traditional methods. However, challenges remain. As highlighted by several studies, the quality and scale of training data are paramount [47, 114]. Models trained on biased or limited datasets may have poor generalization. Furthermore, while powerful, LLMs are often “black boxes”, and enhancing their interpretability to reveal the key structural motifs driving activity is a critical area of future research [115]. The integration of LLMs with structural modeling and experimental validation will be key to realizing their full potential in developing next-generation antimicrobial therapeutics [6].