Deep learning-driven integrated pipeline for de novo design and synthesis of antimicrobial peptides Authors: Jiahui Liu, Yun Chen, Jian Tang, Xupu Xing, Jin-Shun Lin, Juping Sun, Xin-Hui Xing, Juan Li, Can Yang Zhang Published: 2026 ABSTRACT AbstractAntimicrobial peptides (AMPs) are promising alternatives to conventional antibiotics against bacterial infections. However, the discovery of AMPs is impeded by the limitations of biochemical screening and the difficulty computational approaches face in balancing efficacy with structural diversity. We proposed an integrated “generation-evaluation-validation” framework to facilitate de novo discovery of AMPs. First, we constructed a soft prompt-tuned ProtGPT2 to efficiently generate candidates AMPs with both novel structures and promising therapeutic potential. Secondly, we adopted a multiple-choice learning ensemble model that enables high-confidence evaluation of candidates via a dynamic voting network. Finally, antimicrobial experiments were used to validate the activity of top-ranked de novo AMPs by monitoring bacterial surface changes. Out of nine candidates, four exhibited potent strain-specific activity, while two demonstrated broad-spectrum efficacy. All tested AMPs exhibited strong biofilm inhibition, potent membrane disruption, and minimal hemolysis, indicating significant therapeutic potential. With strong generalizability and versatility beyond AMPs, the proposed framework’s modular design will facilitate adaptation to diverse peptide design tasks in the future. By integrating soft prompt tuning, multimodal ensemble learning, and experimental verification, this framework presents a practical and scalable strategy for rapid, resource-efficient de novo peptide discovery, particularly suited for applications where experimental throughput and cost are critical constraints.Subject terms: Computational biology and bioinformatics, Drug discovery, Microbiology Introduction In recent years, the proliferation of drug-resistant bacteria caused by the misuse and overuse of antibiotics has become a major threat to global public health security1. Antimicrobial resistance (AMR) is progressively diminishing the efficacy of existing antibiotics, jeopardizing modern medicine and posing severe risks to populations worldwide1–3. Projections indicate that a total of 39 million people will die directly from drug-resistant infections between 2025–2050, with annual deaths reaching 1.91 million by 2050—a staggering 70% increase over 2021 figures2,4. Against this alarming backdrop, the development of effective antimicrobial therapies has become an urgent global priority. Antimicrobial peptides (AMPs) represent a novel class of agents with growing interest as alternatives to conventional antibiotics, particularly in the face of rising antibiotic resistance, attributed to the low toxicity, broad-spectrum antimicrobial activity, low propensity for resistance development, and immunomodulatory properties5,6. Distinct from traditional small-molecule antibiotics that act on single molecular targets, AMPs exert bactericidal effects by physically disrupting the integrity of bacterial cell membranes, thereby causing leakage of intracellular contents and subsequent bacterial death5,7. Crucially, this multi-target mode of action renders it difficult for bacteria to develop resistance via single-gene mutations, thus offering a novel strategy to address the challenge of AMR8. The clinical success of AMPs demonstrates their therapeutic viability and provides a robust foundation for next-generation antimicrobial development9. However, conventional experimental approaches for AMP discovery are time-consuming, costly, and inefficient, therefore severely impeding research and development efficiency. In recent years, large-scale pre-trained generative models had emerged as a promising solution to this bottleneck10. Models such as GPT11 and LLaMA12 in natural language processing, alongside ProtGPT213 and ProGen14 in the domain of protein design, had demonstrated exceptional capabilities in capturing the properties of biological macromolecules through pre-training on extensive protein sequence corpora. This is precipitating a paradigm shift in biological sequence generation and unlocking new avenues for peptide and protein engineering. Theoretically, when these models are subjected to targeted fine-tuning using validated antimicrobial sequence datasets, they can specifically capture the unique physicochemical properties of AMPs15, generate novel AMP candidates with biological plausibility of sequences, and efficiently navigate the vast combinatorial space of sequences16–18. However, existing fine-tuning strategies suffer from notable limitations. Full fine-tuning not only entails exorbitant computational costs but also tends to induce overfitting on small-scale AMP datasets. On the other hand, unoptimized generative models lack directional guidance toward the functional characteristics of AMPs, resulting in generated sequences that may be structurally similar to AMPs but devoid of the expected biological activity. This “efficiency-specificity dilemma” represents a key bottleneck in adapting pre-trained models for AMP discovery. Further compounding this bottleneck is the problematic assessment of generated sequence quality. Current evaluation methods, including per-token log-likelihood scoring and similarity-based comparisons to known protein sequences, exhibit notable limitations19. The former, while capable of computing the conditional probability for each residue in a sequence, is hindered by sensitivity to sequence length, inability to capture semantic coherence, and misalignment with human judgment20. The latter, exemplified by its application in the GAN-based framework ProteinGAN, involves evaluating sequence identity with known target natural proteins, where generated sequences are prioritized based on their similarity to known homologs21. Both approaches risk overlooking promising sequences that diverge significantly from known proteins22. Further complicating matters, reliance on existing protein prediction tools presents challenges in both tool selection and usability: determining suitable evaluation metrics and selection criteria demands substantial effort, while mastering the computational environments, dependencies, and workflows of diverse tools requires significant resource investment. To address these challenges, we proposed an innovative integrated “generation-evaluation-validation” framework that combined soft-prompt tuning of ProtGPT2 with a dedicated AMP discriminator model (Fig. 1). Firstly, we applied soft-prompt tuning to ProtGPT2, with this optimized model designated as AMP-GPT for de novo generation of AMPs. This parameter-efficient tuning method not only avoided the overfitting risk of full fine-tuning but also retained the model’s powerful protein generation capability. Distinct from adapter-based methods like Low-Rank Adaptation (LoRA), which inject trainable matrices into the model’s internal layers to alter attention mechanisms, soft prompt tuning operates exclusively at the input level. This allows us to rigorously test whether the pre-trained model can be steered toward AMP generation solely through “external instruction” without manipulating its internal weight representations. Subsequently, the Evolutionary Scale Modeling 2 (ESM-2) model13, renowned for capturing deep biological features of sequences, was employed for feature extraction23. These features were then input into the multiple-choice learning (MCL-AMP) ensemble model, which comprises a multi-layer perceptron (MLP), a convolutional neural network (CNN), and attention-based bidirectional long short-term memory network (Bi-LSTM). Through a dynamic voting network that adaptively weighted expert predictions and a self-confidence hinge loss penalizing overconfident errors, MCL-AMP mitigated the inherent biases of individual models, enabling robust AMP potential evaluation. Finally, candidate peptides were selected and validated through wet-lab experiments covering antimicrobial activity, biosafety, and mechanism of action, forming an integrated pipeline from AI design to experimental validation. It was important to note that AMP-GPT and MCL-AMP functioned as distinct modules within this pipeline and were trained independently without mutual gradient updates. The integrated pipeline employed here refers to a macroscopic “generation-evaluation-validation” workflow rather than an algorithmic reinforcement learning loop. By integrating parameter-efficient generative model tuning and multi-modal ensemble evaluation, this work provides an innovative and reliable solution for the efficient de novo discovery of new AMPs, and offers a referable, generalizable research paradigm for AI-driven peptide drug design. Endowed with inherent versatility, flexibility, and portability, the framework is expected to accelerate the development of novel anti-infective candidates amid the growing challenge of AMR. AI-driven de novo discovery pipeline of AMPs. Overview of the integrated “generation-evaluation-validation” framework, including candidate generation, computational evaluation, and wet-lab validation of selected AMPs. Created with BioRender.com. Results Generation of AMP candidates via soft prompt tuning We established an integrated pipeline for AI-driven AMP de novo discovery. To resolve the critical trade-off between computational efficiency and functional specificity in adapting pre-trained models for AMP discovery, we developed AMP-GPT—a soft-prompt tuned variant of ProtGPT2 (Fig. 2a). This parameter-efficient strategy introduced a learnable prompt matrix prepended to input embeddings, with only prompt parameters updated during training while the base model remains frozen. Compared to full fine-tuning, this approach retained ProtGPT2’s generative diversity, reduced computational costs by updating less than 1% of parameters, enabled flexible adaptation to diverse design objectives, and efficiently guided AMP synthesis while preserving broad protein-generation capabilities24. By choosing this strategy over internal-parameter-efficient methods (e.g., LoRA), we aimed to specifically evaluate the “promptability” of the protein language model by comparing the efficacy of external input guidance directly against the comprehensive internal weight adaptation (Fullfine-tuning). To evaluate the effectiveness of soft-prompt tuning, we generated 1000 peptide sequences for each of four sample groups (natural AMPs from test data, pretrained ProtGPT2, fully fine-tuned ProtGPT2, and AMP-GPT), and compared their structural, functional, and diversity metrics. Using modlAMP, we analyzed 6 key physicochemical properties, including amino acid distribution, isoelectric point, hydrophobicity, hydrophobic moment, charge, and charge density to quantify the similarity between generated sequences and natural AMPs. As shown in Fig. 2b–g, the pretrained ProtGPT2 model, which was designed for general protein generation, produced sequences that markedly deviated from the characteristics of AMPs. However, sequences generated by both fully fine-tuned and AMP-GPT exhibited physicochemical properties significantly closer to those of natural AMPs, highlighting the necessity of task-specific adaptation. Notably, AMP-GPT achieved performance comparable to full fine-tuning despite updating only a small number of parameters. To further compare sequence diversity generated by full fine-tuning and soft-prompt tuning, we calculated word frequencies via n-gram analysis (2-mer and 3-mer) and used Shifterator25 to evaluate the Shannon entropy divergence of k-mer distributions under these two tuning paradigms. Figure 2h, i revealed that 3-mers generated by soft-prompt tuning exhibited a broader distribution, particularly in high-ranking motifs, indicating their ability to produce more diverse 3-mer combinations. In contrast, full fine-tuning reduced sequence diversity due to the model’s acquisition of specific patterns, with a similar trend observed in 2-mer analysis (Fig. 2h, i). To further evaluate whether AMP-GPT learned the intrinsic statistical structure of natural AMPs, we performed a tripeptide-level entropy analysis comparing natural AMPs, AMP-GPT–generated peptides, and random sequences. The generated peptides closely reproduced the Φ-score shift patterns of natural AMPs, whereas random sequences showed dispersed non-AMP-like distributions, as reflected by a low Jensen–Shannon divergence with natural AMPs (JSD = 0.1425) and a substantially larger divergence for random sequences (JSD = 0.6594) (Supplementary Fig. S1). These results demonstrated that while full fine-tuning enhanced sequence pattern specificity, it might diminish the functional variability of generated AMPs (a critical factor influencing antimicrobial activity). Soft-prompt tuning preserved higher sequence diversity, maintaining a broader k-mer composition distribution while avoiding overfitting to narrow specific patterns. Collectively, we adopted a synergistic evaluation framework commonly utilized in AI-generated AMP research, integrating Shannon entropy for quantifying sequence diversity and analyses of amino acid composition alongside key physicochemical properties for validating biophysical rationality26–28. These complementary assessments collectively confirmed that AMP-GPT-derived sequences not only exhibited high diversity but also aligned with the intrinsic statistical and biophysical characteristics of natural AMPs. Therefore, we employed the AMP-GPT to generate 30,000 candidate AMP sequences, ensuring they not only conformed to the core structural and chemical characteristics of AMPs (e.g., amphipathic α-helices, strong positive charge) but also possessed sufficient diversity to cover potential functional variations. Deep learning-based generation of AMP candidates and related analyses. a Schematic of soft-prompt tuning (created with Microsoft PowerPoint). Comparisons of physicochemical properties in generated peptides, including b amino acid composition, c isoelectric point, d hydrophobicity, e hydrophobic moment, f charge, and g charge density. All panels correspond to natural AMPs from test data (Real-AMP), pretrained ProtGPT2 (ProtGPT2), fully fine-tuned ProtGPT2 (Finetune-ProtGPT2), and AMP-GPT (Prompt-ProtGPT2). h The Shannon entropy divergence of 2-mer composition distributions under the two tuning paradigms was shown, with full fine-tuning in purple and soft-prompt tuning in yellow. i The Shannon entropy divergence of 3-mer composition distributions under the two tuning paradigms was shown, with full fine-tuning in purple and soft-prompt tuning in yellow. Ensemble evaluation using the MCL-AMP model To reliably prioritize clinically viable AMPs from the generated candidates, we deployed the MCL-AMP ensemble discriminator model to evaluate AMPs properties and activity (Fig. 3a). Amino acid sequences represent the most fundamental feature of proteins, but effective feature extraction methods are required to enable the model to better capture the underlying information embedded in peptide sequences, thereby facilitating the construction of predictive models with superior performance. ESM-2 is a deep learning model based on the transformer architecture, which demonstrates exceptional performance in capturing evolutionary relationships and functional features of protein sequences, achieving state-of-the-art results in downstream tasks such as structure and function prediction. Here, we utilized the esm2_t33_650M_UR50D variant of the ESM-2 protein language model to generate sequence embeddings, which were then input into the MCL-AMP ensemble discriminator model. The MCL-AMP ensemble discriminator model used in this study comprised three expert networks: namely, MLP, CNN, and attention-based Bi-LSTM (Fig. 3b). The model incorporated a voting network that adaptively assigns confidence weights to each expert model, alongside a self-confidence hinge loss function29 to mitigate overconfidence in incorrect predictions. Through this design, MCL-AMP’s ensemble architecture balanced the strengths and limitations of each component, enabling comprehensive and reliable AMP identification. Specifically, the CNN was specialized in extracting local sequence patterns, detecting short-range amino acid interactions, and capturing key AMP-relevant local physicochemical features (e.g., α-helical segments, amphipathicity, charge distribution, and hydrophobicity; Fig. 3b). The attention-based Bi-LSTM models sequential context, captured long-range dependencies and key amino acid patterns (e.g., hydrophobicity gradients), and thereby recognized residue co-dependencies and secondary-structure propensities characteristic of AMPs30. The MLP further integrated these features to assess global properties such as charge distribution, consolidating multi-scale evaluations into a unified framework. Crucially, the dynamic voting network adaptively weighted predictions from each expert, resolved conflicts, and prioritized the most reliable outputs. This adaptive mechanism mitigated inherent biases of individual models (e.g., oversimplification of sequence complexity by the MLP, or myopic focus on local motifs by the CNN) while reducing implementation barriers and ensuring consistent evaluation standards. Architecture of MCL-AMP and convolutional neural network. a Composition of MCL-AMP (MLP, CNN, attention-based Bi-LSTM) (created with Draw.io). b Structure of the CNN (created with Microsoft PowerPoint). The performance of our model had been compared with various state-of-the-art models on the validation sets, the result as shown in Table 1. These results demonstrated that MCL-AMP outperformed state-of-the-art methods across critical metrics. It attained 88.89% sensitivity and 87.63% specificity, effectively balancing the identification of true AMPs and the avoidance of false positives. Notably, its AUC of 94.66% and MCC of 76.53% reflected superior discriminative power and classification robustness compared to alternatives. While AMPlify prioritized specificity, its lower sensitivity limited practical utility. To rigorously quantify the contribution of each component to the overall performance of the MCL-AMP model, a comprehensive ablation study was conducted. The results confirmed the superiority of the ensemble architecture over individual specialist models (Supplementary Table S1). They highlighted the critical roles of the dynamic voting network (Supplementary Table S2), the integration of all three expert networks (Supplementary Table S3), and the confident hinge loss function (Supplementary Table S4) in achieving state-of-the-art performance. These findings validated the necessity of heterogeneous model integration and adaptive weighting mechanisms in developing robust computational tools for AMP discovery, particularly in scenarios requiring high-precision prediction and good generalization across diverse peptide sequences. Through multi-branch collaborative architecture, the MCL-AMP model yielded a confidence score between 0 and 1, representing the predicted probability that a sequence exhibits antimicrobial activity. The final score represented a comprehensive, weighted assessment that integrated multiple feature representations directly learned from real AMPs, rather than being a simple black-box output. This approach transcended the limitations of hand-crafted, pre-defined feature thresholds9 (e.g., net charge > +2) by learning complex nonlinear relationships between features directly from empirical data, enabling more accurate identification of functional AMPs. Collectively, these results underscored that the MCL-AMP model, through its rational integration of heterogeneous networks and adaptive mechanisms, provided a robust and high-performance computational framework for AMP identification. From the 30,000 candidate sequences generated by the optimized AMP-GPT, after undergoing deep biological feature extraction by ESM-2 and evaluation by MCL-AMP, we selected the top 9 scoring peptides for further analysis and validation (sequences shown in Table 2). The selection of these top 9 candidate peptides was driven by a balanced consideration of resource efficiency, statistical rigor, and scientific representativeness. Given the prohibitive costs of peptide synthesis, full experimental validation of a large number of candidates was economically infeasible. Furthermore, this choice aligned with established practices in AI-driven AMP studies, such as studies by Cao et al.29 and Huang et al.30, which typically validated 5–15 candidates to balance novelty exploration with resource limitations. This sample size also ensured sufficient statistical power to detect meaningful biological activity while maintaining the ability to conduct in-depth structural and functional characterization of each candidate. Furthermore, to confirm that these AMP-GPT–generated peptides were not close variants of known AMPs, we conducted a BLASTP-based novelty assessment against the NCBI database. None of the 9 peptides displayed high-identity, high-coverage matches (>70% identity and >70% coverage), indicating that all candidates represent structurally novel sequences (Supplementary Table S5). Table 1Comparison with state-of-the-art methodsModelSn(%)Sp(%)Acc(%)MCC(%)AUC(%)MCL-AMP88.8987.6388.2676.5394.66CAMP(RF)49–5184.6972.7478.7157.8583.04CAMP(SVM)4984.7077.9481.7267.8585.35AMPlify78.9990.3384.1668.8490.75 Comparison with state-of-the-art methods Table 2The amino acid sequences of top-9 candidate peptidesNameSequenceNet charge (pH = 7)Aromaticity (%)Hydrophobic ratio (%)YR26YRGIRGKGWWKAAKAALKAVSSAAKH+711.5457.69GL18GLFKIFKKSVKHACKKKK+811.1144.44GI13GIKGALKSIHFAL+27.6969.23FF24FFPIIARLAAKVIPSLVRAVTKKC+58.3362.50GW20GWLNKIIKKAWRKFHEFSK+620.0045.00KI16KIAKRIWKILLRRLFR+712.5056.25IK15IKHWKKVWKHAKKVL+613.3346.67WF13WFIIWHWRKIPKK+430.7753.85AT18ATFKRCRRWKKALHACRR+811.1144.44 The amino acid sequences of top-9 candidate peptides Notably, high net charge, aromaticity, and hydrophobic ratio are key physicochemical factors determining the antimicrobial activity of AMPs8. Interestingly, all 9 candidate peptides in this study carried a positive charge with an average net charge of +5.80, significantly higher than the average net charge of existing AMPs (+3.32)10. Among them, 8 candidate AMPs had a net charge exceeding +4, confirming their strong positive charge characteristic (Table 2). The results also showed that the aromaticity (7.69%–30.77%) and hydrophobic ratio (44.44%–69.23%) of different candidate peptides varied significantly. All candidate peptides in this study had a hydrophobic ratio higher than the average value of existing natural and synthetic AMPs (41%), and 8 candidate peptides had an aromaticity greater than 10%31. These results pronounced cationic properties, coupled with appropriate aromaticity and optimal hydrophobic ratio, were functionally critical. Specifically, the positive charge facilitates electrostatic interactions with negatively charged microbial membrane components, aromaticity enhances membrane insertion, and hydrophobic ratio promotes subsequent membrane disruption32. Collectively, these features not only serve as key indicators of the candidate peptides’ excellent potential antimicrobial activity but also demonstrate the robustness of our screening framework in accurately identifying sequences with evolutionarily conserved structural hallmarks of functional AMPs. Computational structural characterization of top candidates To further investigate the structural basis of antimicrobial activity, we adopted AlphaFold3 to predict the three-dimensional structures of the 9 candidate peptides33,34. As illustrated in Fig. 4a, 8 out of 9 candidates adopted amphipathic α-helical conformations, a structural signature characteristic of natural AMPs. This conformational similarity provided initial evidence for the biological plausibility of our designed sequences. To quantitatively assess the structural properties relevant to antimicrobial function, helical wheel projections were utilized to analyze the spatial distribution of residues in the selected peptides. As shown in Fig. 4b, the generated sequences predominantly exhibited well-defined amphipathic α-helical structures, a key feature enabling membrane interaction and disruption in functional AMPs. Among the candidates, only YR26 displayed a relatively low hydrophobic moment with more dispersed hydrophobic and hydrophilic residues, deviating slightly from the canonical amphipathic pattern. The remaining 8 peptides maintained clear segregation of hydrophobic and hydrophilic domains, mirroring the structural attributes of well-characterized natural AMPs (Fig. 4b). These structural analyses collectively validated the efficacy of our AMP generation model by demonstrating its ability to design sequences with structural features consistent with known functional AMPs. The model not only produced sequence-diverse candidates but also generated peptides with conformationally relevant characteristics that were crucial for antimicrobial activity. Furthermore, the training process was monitored and showed stable convergence, as indicated by the consistently decreasing loss and perplexity values across 200 epochs (Supplementary Figs. S2–4). Computational structural features of candidate AMPs. a Predicted 3D structures (with secondary structural elements) of top-9 candidate AMPs (by AlphaFold3). b Helical wheel projections of the top-9 candidate AMPs. Experimental validation of antimicrobial and antibiofilm activities To further validate the functional characteristics of the screened AMPs, we chemically synthesized the top-scoring 9 predicted AMPs and conducted a systematic in vitro evaluation through assessments of antimicrobial activity, mechanism of action, antibiofilm capability, and safety. Firstly, we determined the MIC and minimum bactericidal concentration (MBC) values of those AMPs against one gram-positive bacteria (Staphylococcus aureus (S. aureus)) and three strains of gram-negative bacteria (Pseudomonas aeruginosa (P. aeruginosa), Escherichia coli (E. coli), and Klebsiella pneumoniae (K. pneumoniae)), as shown in Table 3. Table 3MIC of top-9 peptides against four standard bacterial strainsPeptidesMIC (μg/ml)S. aureus ATCC 25923P. aeruginosa PAO1E. coli wildK. pneumoniae ATCC 13883YR26>256>256128128GL18>256>256>256>256GI13>256>256>256>256FF24>256128>256128GW2025616328KI16128321616IK15>256>256>256>256WF13>256>256>256>256AT18>256>256>256>256Ciprofloxacin8884Kanamycin1111 MIC of top-9 peptides against four standard bacterial strains Among the 9 candidates, 4 exhibited high antibacterial activity against at least one bacterial strain. Specifically, the GW20 and KI16 peptides exhibited outstanding bacteriostatic effects, with their MIC values against all 4 tested strains being significantly low. Compared with the positive control drugs ciprofloxacin and kanamycin, the MIC values of GW20 and KI16 were slightly higher (Table 3). However, it should be noted that positive drugs have undergone long-term clinical optimization and structural modification, whereas the AMPs in this study were newly designed sequences. These results already showed a bacteriostatic trend close to that of positive drugs (e.g., the MIC of GW20 against KP was 8 μg/mL, which was in the same order of magnitude as the MIC of ciprofloxacin against this bacterium (4 μg/mL)). These results initially confirmed the success of the model’s screening and laid a foundation for further improving activity through structural optimization and chemical modification. To elucidate the underlying reasons for the divergent antimicrobial activities among the candidate peptides, we performed a systematic analysis of key physicochemical properties, including net charge, aromaticity, and hydrophobic ratio. This analysis revealed that compared with other poor antibacterial candidates such as GL13 and IK15, the superior antibacterial efficacy of GW20 and KI16 stemmed from their favorable combination of these characteristics. For example, GW20 (+6) and KI16 (+5) possessed a higher net positive charge than GL13 (+4) and IK15 (+3), both strengthening electrostatic binding to the negatively charged bacterial membrane surface. Additionally, GW20 had an optimal hydrophobic ratio of 45.00%, and KI16 featured a hydrophobic ratio of 56.25%. These values promoted insertion into the bacterial membrane’s hydrophobic core while avoiding excessive aggregation. Moreover, GW20 contained two tryptophan residues (10% of total residues), and KI16 had one tryptophan and one phenylalanine (12.50% aromaticity), which enhanced membrane insertion efficiency and stabilized peptide-membrane interactions, which features absent in GL13 and IK15. Collectively, the synergistic effects of high net charge, optimal hydrophobic ratio, and unique aromaticity in GW20 and KI16 contributed to their superior antimicrobial activity, highlighting that balanced physicochemical properties are critical for AMP functionality. The MBC results of GW20 and KI16 against the four tested bacteria were shown in Table 4. The results indicated that both peptides exerted bactericidal effects on all four bacteria, with most MBC values being 2 × MIC, suggesting that they can effectively kill bacteria on the basis of inhibiting bacterial growth, demonstrating strong antimicrobial activity. Subsequently, we detected the hemolytic effect of the peptides on rabbit red blood cells to evaluate the biological safety of GW20 and KI16. The result was shown in Supplementary Fig. S5a, for GW20, as the concentration decreased, the hemolysis rate decreased significantly. At concentrations of 64 μg/mL and lower, the hemolysis ratios were extremely low, close to the negative control level. This suggested that GW20 had a concentration-dependent hemolysis effect, with higher concentrations being more likely to cause RBC lysis, but at lower concentrations, it showed minimal hemolytic activity. For KI16, there were no significant differences among the different concentration groups when compared to the negative control (Supplementary Fig. S5b). This indicated that KI16 had very weak hemolytic activity on rabbit RBCs across the tested concentration range. Both the bactericidal activity and safety profiles of GW20 and KI16 were highly consistent with the high activity characteristics predicted by our model based on sequence features (such as amphipathic α-helical structure and hydrophobic/hydrophilic residue distribution), initially verifying the effective integrated pipeline from “sequence design” to “functional verification” of the model. While the model successfully identified potent AMPs, the observation of high-scoring but inactive candidates (e.g., YR26 and GL18) highlighted the complexity of peptide design. Structural analysis confirmed that these peptides possessed the typical amphipathic α-helical structure, suggesting that their lack of potency might stem from other factors, such as self-aggregation or stability issues, which were not fully captured by sequence-based predictors. This underscored that while the current framework effectively screened for structural plausibility, experimental validation remained essential to confirm final biological activity. From a statistical perspective, the validated hit rate of our framework reached approximately 44.4% (4 out of 9 top-ranked candidates), which reflects its superior efficiency compared to traditional wet-lab screening8. Specifically, the integrated pipeline compressed the vast combinatorial peptide sequence space into a high-quality candidate pool, minimizing labor-intensive and large-scale experimental trials. This capability enabled us to rapidly identify potent leads with broad-spectrum antimicrobial activity from 30,000 de novo generated candidates, completing the entire discovery process within a significantly shortened research and development cycle27,35. Collectively, these results demonstrate the framework’s strong potential as an efficient engine for accelerating AMP lead discovery. Table 4MBC of GW20 and KI16 against four standard bacterial strainsPeptidesMBC (μg/ml)S. aureus ATCC 25923P. aeruginosa PAO1E. coli wildK. pneumoniae ATCC 13883GW20>256326416KI16256643232Ciprofloxacin8884Kanamycin1111 MBC of GW20 and KI16 against four standard bacterial strains To evaluate the dynamic bactericidal effects of GW20 and KI16, we conducted a time-inhibition study on four bacterial strains. The time-inhibition growth curves illustrated the effect of two peptides on four different bacterial strains over a 12 h period (Fig. 5a–d for GW20) and (Fig. 5e–h for KI16). Against all four bacterial strains, the control groups showed a consistent upward trend in absorbance at 600 nm with time, indicating unimpeded bacterial growth. For GW20, treatment with 1/4 × MIC already exhibited a discernible growth-retarding effect compared to the control (Supplementary Fig. S6). When treated with 1/2 × MIC of GW20 and KI16, there was a noticeable slowdown in growth compared to the controls for each strain, yet the absorbance still increased over time, suggesting a moderate inhibitory effect. At 1 × MIC and 2 × MIC, the growth of most strains were almost completely suppressed, with the absorbance staying close to the initial values, clearly demonstrating a potent bacteriostatic effect. This concentration-dependent antimicrobial activity was consistent with the functional trend predicted by our model based on sequence physicochemical properties (such as hydrophobic moment and positive charge), verifying the accurate prediction of peptide activity by the screening system. Experimental validation of antimicrobial and antibiofilm activities. GW20 against a S. aureus, b P. aeruginosa, c E. coli, and d K. pneumoniae (created with GraphPad Prism 10.1.2.324). KI16 against e S. aureus, f P. aeruginosa, g E. coli, and h K. pneumoniae (created with GraphPad Prism 10.1.2.324). i SEM images showing morphological changes in S. aureus and E. coli after treatment with GW20 or KI16 at 4 × MIC, respectively. Scale bar: 1 μm. Effects on E. coli biofilm: inhibition of j adherence, k formation, and l disruption of established biofilm (created with GraphPad Prism 10.1.2.324). Further, a common mechanism of action of AMPs is to form transmembrane pores through specific interactions with bacterial cell membranes, ultimately leading to the leakage of bacterial contents and cell lysis. To verify the mechanisms of GW20 and KI16, we observed the morphological changes of E. coli and S. aureus treated with the two peptides using scanning electron microscopy (SEM). For mechanistic illustration under pronounced conditions, observations were first performed at 4 × MIC. The results showed that untreated bacteria exhibited intact rod-like (E. coli) or spherical (S. aureus) morphologies with smooth and well-defined cell membranes. In contrast, after treatment with GW20 and KI16, the bacterial surfaces showed obvious shrinkage, depression, and even rupture (Fig. 5i). To further investigate the concentration-dependent membrane damage, we performed a series of SEM observations on S. aureus treated with GW20 across a range of concentrations. As shown in Supplementary Fig. S7, even at 1/4 × MIC and 1/2 × MIC, concentrations that fall within its defined safety window, bacterial cells began to exhibit noticeable surface irregularities and early signs of membrane distortion. The extent of membrane damage, including shrinkage and depression, became progressively more severe with increasing peptide concentration, culminating in extensive rupture at 1 × MIC and 2 × MIC. This graded morphological response confirmed that the membrane-targeting action of GW20 was initiated at low, therapeutically relevant concentrations and was not an artifact of extreme, cytotoxic doses. Notably, the observed membrane disruption effect was highly consistent with the structural features predicted by AlphaFold3. The predicted α-helical conformation exhibited pronounced amphipathicity: its hydrophobic moiety inserted into the hydrophobic core of the bacterial lipid bilayer, while the hydrophilic moiety that enriched with positively charged residues, engaged in tight electrostatic interactions with the negatively charged bacterial membrane surface3. This targeted binding subsequently triggered conformational dynamics in the peptide, ultimately compromising membrane integrity through the formation of pores or disruption of lipid packing36. These results verified the cross-scale predictive ability of our model (from sequence to structure to function), with the amphipathic structure of AMPs generated through AMP-GPT directly laying the structural foundation for their membrane-damaging effects. In addition to direct bactericidal activity, the ability of AMPs to interfere with biofilms is another important indicator for evaluating their clinical application potential37. As a “protective community structure” of bacteria, biofilm formation involves four stages: initial adhesion, community proliferation, extracellular matrix secretion, and maturation38. Traditional antibiotics often exhibit limited efficacy due to their inability to penetrate dense extracellular matrices or ineffectiveness against drug-resistant bacteria within biofilms2. Therefore, we systematically investigated the effects of GW20 and KI16 on three key stages of E. coli biofilm formation, focusing on their ability to inhibit initial bacterial adhesion, suppress biofilm development, and disrupt established biofilms. In the biofilm adherence assay, bacteria in the control group showed strong adhesive capacity, providing a basis for biofilm formation (Fig. 5j). At 1 × MIC, ampicillin reduced adhesion but was far less effective than KI16 and GW20, with the adhesion rate in the ampicillin group was significantly higher (Fig. 5j). At 2 × MIC and 4 × MIC, the adhesion rates of the three agents were comparable, suggesting that their effects on adhesion were similar at high concentrations. These results confirmed that KI16 and GW20 were superior to ampicillin in interfering with initial bacterial adhesion, especially at lower concentrations, which was crucial for blocking the initiation of biofilm formation. In the biofilm formation inhibition assay, KI16 and GW20 exhibited significant inhibitory activity, which was often comparable to or even exceeded that of ampicillin (Fig. 5k). This indicated that both peptides could effectively disrupt the key processes in biofilm maturation, such as bacterial proliferation and extracellular matrix secretion, thereby inhibiting the development of biofilms. For pre-established mature biofilms, KI16 and GW20 also showed strong disruptive effects (Fig. 5l), which was of great significance for eliminating persistent biofilm-associated infections. Overall, the experimental results confirmed that the AMPs KI16 and GW20 exhibited potent anti-biofilm properties, as demonstrated by reduced adhesion, inhibited formation, and disruption of existing biofilms, with performance surpassing that of ampicillin. Combining wet experimental data and computational results, we proposed the mechanism underlying biofilm disruption by AMPs as follows. The positively charged regions of their amphipathic α-helical structure tightly bound to the negatively charged components on the bacterial membrane surface, thereby inhibiting bacterial adhesion39. Meanwhile, the hydrophobic regions could insert into the extracellular polysaccharide matrix, disrupting its structural integrity and interfering with bacterial community proliferation and matrix secretion to suppress biofilm maturation36. For mature biofilms, the amphipathic structure of AMPs effectively disassembled the biofilm through hydrophobic interactions and electrostatic interactions, releasing bacteria and enhancing antimicrobial efficacy. Overall, the experimental results confirmed that the AMPs KI16 and GW20 exhibited potent anti-biofilm properties, as demonstrated by reduced adhesion, inhibited formation, and disruption of existing biofilms, with performance surpassing that of ampicillin. In summary, the in vitro experimental results showed that the AMPs. KI16 and GW20 obtained through our model-driven screening system exhibited excellent performance in antimicrobial activity, antibiofilm properties, and biological safety, and their functional characteristics were highly consistent with the predictions of our model. These not only verified the effectiveness of our system throughout the whole process from sequence generation, feature screening to function prediction, but also provided solid experimental support for the efficient discovery of clinical-grade AMPs. Discussion This study established an AI-driven integrated pipeline system for AMP discovery, bridging computational design and experimental validation to address the key challenges in AMP de novo design. By integrating parameter-efficient soft prompt tuning for generation, multi-modal evaluation, and rigorous wet-lab validation, the framework overcame inefficiency of traditional experimental trial-and-error methods and the struggle of computational approaches to balance functionality and structural diversity. The core contributions of this work were threefold: firstly, soft prompt tuning effectively balanced generative diversity and functional specificity, offering a cost-effective strategy to repurpose large pre-trained models for targeted peptide design; secondly, the MCL-AMP ensemble model, with its dynamic weighting mechanism, reduced false positives in virtual screening and enhanced the translatability of computational predictions to experiments; thirdly, the systematic validation of top-ranked candidates (e.g., GW20 and KI16) confirmed that AI-designed AMPs could approach clinical antibiotics in antimicrobial efficacy while outperforming them in antibiofilm activity, validating the reliability and efficiency of the integrated framework. Notably, the framework prioritizes efficient reduction of the vast peptide sequence space over achieving a near-perfect hit rate, reflecting the biological complexity of AMP function and the inherent sparsity of functional AMPs in sequence space, and its value lies in enriching promising candidates for experimental validation rather than flawless prediction of individual sequences. While GW20 and KI16 had higher MICs against planktonic bacteria than conventional antibiotics such as ampicillin, they demonstrated a valuable therapeutic profile with potent multi-stage anti-biofilm activity. They significantly inhibited initial adhesion at 1 × MIC and effectively disrupt mature biofilms, directly addressing a critical clinical challenge. Combined with low hemolysis, these properties supported their development as topical anti-biofilm agents for chronic wounds or implant-related infections, where both biofilm eradication and planktonic control were required. Beyond AMP discovery, the “generative–evaluation–validation” pipeline, with its modular architecture and model-agnostic nature, boasts exceptional methodological versatility and broad portability. It is designed as a universal research paradigm that transcends AMP discovery, enabling seamless transfer to diverse bioactive peptide and protein design tasks through targeted manual adjustments of core modules, without overhauling the pipeline’s fundamental structure. For example, in anticancer peptide (ACP) design, the ProtGPT2-based generative module can be fine-tuned with ACP-specific sequence datasets to steer generation toward cancer-cell-targeting sequences, while the MCL-AMP evaluation framework can be reconfigured to incorporate expert models trained on anticancer activity labels, hemolytic toxicity predictors, and cancer-cell-line-specific cytotoxicity data. Similarly, the multi-dimensional validation strategy, combining in silico predictions with functional and structural assays, can be adapted to include tumor spheroid inhibition assays and cancer-target binding affinity measurements. For enzyme engineering, the sequence generation module can be guided by active-site conservation constraints or structure-conditioned sampling to propose catalytically competent variants, while the evaluation stage can integrate structure–function predictors (e.g., catalytic efficiency models, substrate docking scores, and thermostability predictors) alongside molecular dynamics simulations for mechanistic assessment. This adaptability, from generative modeling to task-specific multi-expert evaluation, ensures the framework can be rapidly adjusted to diverse molecular design objectives with minimal retraining effort. Looking forward, several directions can further strengthen the robustness and biological relevance of the proposed framework. First, incorporating multi-objective optimization, including antimicrobial activity, structural stability, toxicity, solubility, and amphipathicity, will better capture the multifaceted requirements of functional peptide design. Such optimization could be achieved through reinforcement learning or evolutionary strategies that actively balance these competing attributes. Second, our structural analyses using AlphaFold3 and helical wheel projections highlight the importance of integrating structural filters into the generative pipeline. Future models could incorporate predicted secondary structure confidence, amphipathic moment, and membrane-binding propensity directly into the scoring function, thereby reducing the mismatch between high in silico scores and poor wet-lab performance. Additionally, failure examples identified in this study provide valuable negative feedback that can be used to refine both the generator and the predictor via adversarial or contrastive learning. Third, reinforcement learning guided by real-time experimental feedback represents a promising avenue to narrow the design–validation gap. Adaptive reward functions can be crafted to favor peptides with experimentally confirmed membrane disruption, synergistic activity, or strain-specific effects. Furthermore, expanding the framework to other therapeutic peptide classes, such as antiviral, anticancer, anti-inflammatory, or quorum-sensing–interfering peptides, will broaden its applicability across biomedical domains40,41. Combining generative models with synthetic biology platforms, such as automated peptide synthesis and microfluidic high-throughput screening, may also accelerate integrated pipeline design-to-validation cycles. Overall, this study highlights the transformative potential of combining AI-driven generative modeling with biochemical validation and structural analysis. By offering a scalable and adaptable pipeline for peptide discovery, our framework contributes to ongoing global efforts against antimicrobial resistance and lays a foundation for next-generation computational peptide engineering. Methods Dataset collection and preprocessing Most AMPs typically range in length from 10 to 50 amino acids42, with peptides exceeding 30 amino acids being more challenging to synthesize. Therefore, we curated a dataset of AMPs with lengths ranging from 10 to 30 amino acids, sourced from four publicly available databases: ADAM43, CAMPR444, DRAMP 2.045. Due to the limited incentive to experimentally verify non-AMPs, repositories of peptides confirmed to lack antimicrobial activity remain scarce. The impact of negative data sampling on model performance and benchmarking has been thoroughly investigated by Sidorczuk et al.46, who demonstrated that models exhibit improved accuracy when their negative dataset is generated using AmpGram47, a sampling method applied to UniProt dataset48. So we employed AmpGram to produce peptide sequences of 10–30 amino acids, thereby constructing a non-AMP database for this study. With the AMPs and non-AMPs collected, we implemented the following preprocessing steps: Remove unnatural amino acids and repetitive sequences. Remove unnatural amino acids and repetitive sequences. Cluster database at high identity with tolerance was adopted to remove redundancy and homologs. For the AMP dataset, a threshold of 0.8 was applied, whereas for the non-AMP dataset, a more stringent threshold of 0.4 was used to prevent overrepresentation of highly similar sequences. Cluster database at high identity with tolerance was adopted to remove redundancy and homologs. For the AMP dataset, a threshold of 0.8 was applied, whereas for the non-AMP dataset, a more stringent threshold of 0.4 was used to prevent overrepresentation of highly similar sequences. Remove redundant non-AMP samples to balance the dataset. Remove redundant non-AMP samples to balance the dataset. The resulting dataset consisted of 4134 AMPs and 4134 non-AMPs, and the data was then split into training, validation, and testing sets in a ratio of 70:15:15. To address the dual objectives of generating structurally novel peptides and accurately evaluating their functional potential, we designed a hierarchical two-stage computational pipeline. This pipeline consists of a generative model for candidate exploration and a discriminative model for activity-based prioritization. The generative model, AMP-GPT, is trained exclusively on natural AMP sequences to learn their underlying grammatical and structural patterns, enabling the creation of novel yet plausible candidates. The discriminative model, MCL-AMP, is trained on a balanced dataset of AMPs and non-AMPs to learn the discriminative features that correlate with antimicrobial activity. It acts as a functional filter and ranker for the generated candidates. This division of labor ensures both exploration of the novel sequence space and precision in functional evaluation. AMP generation model The goal of the generation stage is to explore a broad and structurally plausible space of novel peptide sequences that resemble natural AMPs. To this end, we fine-tuned a protein language model under two distinct parameter-update paradigms. The first two methods were trained for 200 epochs using the AdamW optimizer with a consistent batch size of 16. For fair comparison, the learning rate was fixed at 1 × 10−4 for both fine-tuning and soft-prompt tuning configurations. The soft-prompt tuning framework was implemented with a 20-token prompt length to balance computational efficiency and task-specific guidance. This setup ensured identical training conditions across methods, isolating the impact of parameter update strategies (full fine-tuning vs. soft-prompt tuning) on generation quality. The soft prompt was initialized by sampling values uniformly random from range [−0.5, 0.5]. This approach prevents initial parameter magnitudes from dominating the early optimization dynamics while providing a neutral starting point for task-specific adaptation. MCL-AMP To evaluate and rank the generated candidates based on their predicted antimicrobial activity, we developed MCL-AMP, a novel discriminative model. Unlike the generative model, which learns from positive examples only, MCL-AMP is trained to distinguish AMPs from non-AMPs, thereby providing critical “negative knowledge” and a fine-grained activity score for prioritization. MCL-AMP is a novel model for an AMP evaluator, which combines three specialists using a voting network. Attention-based Bi-LSTM is ideal for capturing long-range dependencies and sequential relationships in peptide sequences. CNNs can effectively capture short-range interactions between amino acids, such as motifs or repeating patterns that are characteristic of AMPs, and the MLP serves as a versatile model for global feature extraction and general pattern recognition. The voting network will receive low-level feature outputs from each expert model, and it learns to generate confidence scores based on these features. The outputs of the specialists are then aggregated according to these confidence scores. Assume we have M specialists, and for each input sample xi, the prediction from the mth expert model is Pθm(y|xi), the voting network computes confidence scores for each expert model based on the low-level feature outputs from all specialists. The voting network takes the concatenation of the low-level outputs from all experts as input and produces confidence scores for each expert, represented by wi1,wi2,…,wiM. These weights are computed using the softmax function over the logits, ensuring that the sum of all weights equals 1, as shown in Eq. (1): Here, logitsm represents the raw output for the m-th expert, which is passed through the softmax function to obtain the weights wim. These weights represent the confidence of each expert model for the input sample xi. The aggregated prediction, Popt(c|xi), is calculated as the weighted sum of the predictions from all experts, as shown in Eq. (2): where Pθm(y=c|xi) is the predicted probability for class c from the m-th expert model for input xi. The final prediction Popt(c|xi) is the weighted sum of all experts’ predictions, based on the confidence weights determined by the voting network. To optimize the performance of the MCL-AMP, we incorporate confident hinge loss21 to tackle overconfidence and enhance the model’s ability to correctly classify AMPs. Traditional loss functions lead to overconfidence, where the model assigns excessively high probabilities to incorrect predictions. Confident hinge loss introduces a margin-based mechanism that penalizes incorrect classifications when the model’s prediction is too close to the correct class, thus forcing the model to increase its confidence in the correct class. The loss function is designed to ensure that the correct class has a higher probability than any incorrect class by a fixed margin. This prevents the model from assigning high confidence to incorrect predictions, improving the overall classification performance. The confident hinge loss is defined as shown in Eq. (3): where vim is the indicator variable for the m-th expert, representing the confidence of the expert for the i. Pθm(y|xi) is the prediction probability from the m-th expert for the i-th example. The margin-based term max(Popt(c|xi)−Popt(yi|xi)+β,0) enforces the margin-based constraint, where β controls the confidence margin and α is a hyperparameter controlling the weight of the margin-based loss. lvi,wi is the loss related to the voting network. Evaluation metrics for the AMP classification model The performance of the model is evaluated based on five metrics: sensitivity (Sn), specificity (Sp), accuracy (Acc), Matthews correlation coefficient (MCC) and the area under the ROC curve. The closer these values are to 1, the better the model’s performance. The formulas for calculating these four metrics are provided in Eq. (4), The area under the ROC curve was calculated using the SK-Learn metrics package. Where TP represents the count of positive samples that are correctly identified, TN denotes the count of negative samples accurately classified, FP indicates the count of negative samples misclassified as positive, and FN indicates to the count of positive samples mistakenly labeled as negative. Materials for wet-laboratory validation Bacteria: wild-type E. coli, S. aureus ATCC25923, and P. aeruginosa PAO1 are provided by the Institution of Institute of Biopharmaceutical and Health Engineering, K. pneumoniae ATCC13883 was purchased from Ningbo Mingzhou Co., Ltd. Candidate peptides with purity >95% were custom synthesized by Shanghai China Peptides Co., Ltd. The sources, identifiers of bacterial strain, the media and chemicals, and equipment used in this study are provided in Supplementary Table S6 and Supplementary Table S7. During wet-laboratory validation, no unexpected or unusually high safety hazards were encountered. MIC and MBC test Bacteria were cultured overnight in LB at 37 °C and 120 rpm to log phase, dilute the bacterial culture to 1 × 105 CFU/mL. The peptides were dissolved into PBS with an initial concentration of 1024 μg/mL, followed by 2-fold serial dilutions, then, 50 μL of each peptide solution was added to 50 μL of the bacterial suspension in a 96-well plate. The final test concentration of peptides was 1–512 μg/mL, three replicates of each peptide concentration were tested against each bacterial strain. Plates were incubated at 37 °C for 16 h. The value of MIC was defined as the lowest peptide concentration at which no visible bacterial growth could be detected. Each test included a negative control consisting of PBS with bacterial inoculums and positive controls such as Ciprofloxacin and kanamycin. For the MBC test, samples are taken from the wells in which no growth is observed and where the MIC has been reached. These samples are then sub-cultured onto agar plates and incubated at 37 °C. The lowest concentration at which no colonies develop is defined as the MBC. Time-inhibition test Bacteria were cultured overnight in LB at 37 °C and 120 rpm to log phase, dilute the bacterial culture to 1 × 105 CFU/mL, mix peptides with bacterial solution to obtain a mixture containing 1/4 × MIC (for GW20 against S. aureus only), 1/2 × MIC, 1 × MIC, and 2 × MIC peptide concentrations, and culture at 37 °C. Measure OD600 every 2 h for a total of 12 h. Repeat the experiment three times, sterile PBS as blank control. Hemolysis test Red blood cells (RBCs) from rabbits were used for this test. Hemolytic activity of the peptides was measured by first mixing the RBCs with the peptide solutions (n = 5). The mixtures were then incubated at 37 °C for 1 h. Afterwards, the samples were centrifuged at 1500 rpm for 8 min. The optical density of the supernatant was detected at 570 nm using a microplate reader. RBCs in PBS were used as the negative control, and 0.5% (w/v) TritonX-100-treated RBCs were used as the positive control. The percentage of hemolysis was calculated as Eq. (5). Inhibition of bacteria adhesion The logarithmic phase E. coli was diluted to 1 × 108 CFU/mL with fresh LB broth. A 200 µL aliquot of the bacterial suspension, containing varying concentrations of AMPs, ranging from 512 µg/mL to 2 µg/mL, was added to each well of a 96-well plate. The plate was incubated at 37 °C for 6 h to allow bacterial adhesion to the wells. Subsequently, the solution was carefully aspirated from the wells, and the wells were washed with 200 µL of PBS to remove non-adherent cells. Following this, 200 µL of methanol was added to fix the bacteria for 15 min after which the methanol solution was aspirated and the plate was left to air-dry for 15 min. The bacteria were then stained with 100 µL of 10% crystal violet for 5 min. Excess crystal violet was gently washed away with ultrapure water, and the plate was air-dried. Finally, 200 µL of 33% glacial acetic acid was used to dissolve the crystal violet stain, and the absorbance was recorded at 600 nm with a microplate reader. The positive control consisted of bacterial suspension mixed with sterile water, while the negative control consisted of medium mixed with sterile water. Calculate the effects of peptides and antibiotics on initial bacterial adhesion. Inhibition of biofilm formation The logarithmic phase E. coli was diluted to 1 × 108 CFU/mL with fresh LB broth. A 200 µL aliquot of the bacterial suspension, containing varying concentrations of AMPs, ranging from 512 µg/mL to 2 µg/mL, was added to each well of a 96-well plate. And it was incubated at 37 °C for 6 h to allow bacterial adhesion to the wells. Then, the solution was carefully aspirated from the wells, and the wells were washed with 200 µL of PBS to remove non-adherent cells. 200 µL of methanol was added to fix the bacteria for 15 min, after which the methanol solution was aspirated and the plate was left to air-dry for 15 min. The bacteria were then stained with 100 µL of 10% crystal violet for 5 min. Excess crystal violet was gently washed away with ultrapure water, and the plate was air-dried. Finally, 200 µL of 33% glacial acetic acid was used to dissolve the crystal violet stain, and the absorbance was recorded at 600 nm with a microplate reader. The positive control consisted of bacterial suspension mixed with sterile water, while the negative control consisted of medium mixed with sterile water. Calculate the effects of peptides and antibiotics on initial bacterial adhesion. Destroy the formed biofilm The logarithmic phase E. coli was diluted to 1 × 105 CFU/mL with fresh LB broth. A 200 µL aliquot of the bacterial suspension, containing varying concentrations of AMPs, ranging from 512 µg/mL to 2 µg/mL, was added to each well of a 96-well plate, and the plate was incubated at 37 °C for 24 h. The positive control consisted of bacterial suspension mixed with sterile water, while the negative control consisted of medium mixed with sterile water. The plate was then processed according to the procedure described in “Inhibition of biofilm formation” to assess the inhibitory effects of GW20 and KI16 on biofilm formation. SEM measurement Bacterial suspensions with a concentration of 1 × 108 CFU/mL were then treated with GW20 and KI16 at final concentrations of 4 × MIC μg/mL at 37 °C for 6 h. For GW20, additional treatments were performed at 1/4 × MIC, 1/2 × MIC, 1 × MIC, and 2 × MIC to assess concentration-dependent morphological changes in S. aureus. Untreated bacterial cells were set as the control group. The bacterial pellet was obtained by centrifuging at 3000 rpm. for 10 min. The pellet was then fixed overnight at 4 °C with 2.5% glutaraldehyde solution. The pellet was washed twice with PBS, and a graded dehydration process was carried out using ethanol solutions of 10%, 30%, 50%, 70%, 90%, and 100% ethanol, with each step lasting 15 min. After the final dehydration step with 100% ethanol, the bacterial pellet was diluted with 100% ethanol. A 10 μL aliquot of the suspension was then deposited onto a clean silicon wafer, then examined using a SEM to observe the surface of the silicon wafer. Statistical analysis All quantitative results in this chapter are expressed as mean ± standard deviation, and all data are obtained using Prism perform one-way ANOVA using software and compare the differences between two experimental groups using Tukey test, P > 0.05 There is no significant difference between the two groups, P < 0.05 indicates a significant difference between the two groups, where P > 0.05 Marked as (ns), P < 0.05 is marked as (*), P < 0.01 is marked as (**), and P < 0.001 is marked as (***). Supplementary information Revised Supplementary Material Clean Acknowledgements This work was carried out at the Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School. The research platform is acknowledged for providing the public AI research platform. This work was supported by Key Research and Development Program of the Ministry of Science and Technology [2023YFA0913600 (2023YFA0913602)], National Natural Science Foundation of China (22278242), Guangdong Innovative and Entrepreneurial Research Team Program (2023ZT10C040), Pearl River Talent Program (2021QN02Y225), Shenzhen Medical Research Fund (SMRF: B2302009), Shenzhen Technology and Innovation Commission (KCXFZ20230731094459001, KCXFZ20240903093102004), Fund from Shenzhen International Graduate School at Tsinghua University (HW2023009, JC2021011). Author contributions J.L. and Y.C.: Conceptualization, data curation, formal analysis, investigation, methodology, experiment verification, writing—original draft. J.T., J.S., and X.X.: Data curation, formal analysis, experiment verification, resources. X.H.X., J.L., and J.S.L.: Conceptualization, resources, supervision, writing—review & editing. C.Y.Z.: Conceptualization, funding acquisition, project administration, supervision, writing—review & editing. Data availability All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. The AMP sequences used in this study were collected from publicly available databases, including ADAM, CAMPR4, and DRAMP 2.0. Negative (non-AMP) sequences were generated using the AmpGram sampling strategy based on UniProt protein sequences, following previously established protocols. After preprocessing, including length filtering, redundancy removal, and dataset balancing, a curated dataset of AMPs and non-AMPs was constructed for model training and evaluation. The processed datasets supporting the findings of this study are available from the corresponding author upon reasonable request. Code availability The code used in the study is publicly available from the GitHub repository: https://github.com/jiahuu/AMP-generation#. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Jiahui Liu, Yun Chen. Supplementary information The online version contains supplementary material available at 10.1038/s44386-026-00045-6. References Associated Data Supplementary Materials Revised Supplementary Material Clean Data Availability Statement All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. The AMP sequences used in this study were collected from publicly available databases, including ADAM, CAMPR4, and DRAMP 2.0. Negative (non-AMP) sequences were generated using the AmpGram sampling strategy based on UniProt protein sequences, following previously established protocols. After preprocessing, including length filtering, redundancy removal, and dataset balancing, a curated dataset of AMPs and non-AMPs was constructed for model training and evaluation. The processed datasets supporting the findings of this study are available from the corresponding author upon reasonable request. The code used in the study is publicly available from the GitHub repository: https://github.com/jiahuu/AMP-generation#.