A Dual-Track Computational Framework Integrating Deep Learning and Large Language Models for Rational Optimization and Target Validation of Antimicrobial Peptides

Objective: To address the growing challenge of antibiotic resistance, this study aims to develop a computational framework that integrates artificial intelligence (AI) with molecular simulations. The framework enables efficient genomic mining and assembly of novel antimicrobial peptides (AMPs), rational optimization of known AMPs, and elucidation of their potential mechanisms of action.

Methods: The framework incorporates two complementary AI strategies. First, a deep learning pipeline was applied to an extremophilic Bacillus sequence library to identify AMP with novel sequence features (designated 7, 4, and 2), followed by molecular dynamics (MD) simulations to select the most structurally stable tandem configurations. Second, a large language model (LLM) was used to direct modifications of the classical AMP cecropin B, generating a series of functionally enhanced variants, including "charge-enhanced" and "structure-optimized" mutants. Candidate peptides derived from both strategies were then systematically evaluated against a representative bacterial protein target panel using molecular docking and long-timescale MD simulations to assess binding stability and affinity.

Results: The candidate peptide (AMP-NJY-LYY-742) demonstrated excellent structural stability and exhibited a strong binding preference for the bacterial outer membrane LptD--LptE complex. LLM-optimized cecropin B mutants also showed target-specific functional improvements: charge-enhanced variants markedly increased binding affinity toward LptD--LptE, while structure-optimized variants enhanced interactions with other targets such as DppA.

Conclusion: This dual-track AI framework effectively integrates exploratory discovery and rational optimization of AMPs. Importantly, both independent approaches converged on the LptD--LptE complex as a key molecular target. The framework provides an efficient and reusable computational paradigm to accelerate AMP development and offers valuable candidate molecules and target information for subsequent experimental validation.

Dataset Construction and Model Training. To develop a robust predictive model for antimicrobial peptides (AMPs), we first constructed a balanced dataset comprising over 10,000 sequences. This was achieved by integrating positive samples (AMPs) from several authoritative databases (e.g., APD3, DADP, and DBAASP) and selecting negative samples (non-AMPs) from the UniProtKB/Swiss-Prot database. We employed a bidirectional long short-term memory (Bi-LSTM) network incorporating an attention mechanism as the core model architecture. Subsequently, an ensemble model was trained using a five-fold cross-validation strategy to ensure its predictive accuracy and generalization capability.

Screening and Filtering. Upon completion of training, the ensemble model was utilized for a high-throughput screening of a genomic sequence library derived from extremophilic Bacillus. Sequences predicted by the model as high-probability AMPs (prediction probability > 0.9) were subjected to a series of stringent bioinformatics filtering criteria. These criteria included, but were not limited to, the evaluation of their potential to form amphipathic helices (hydrophobic moment), the calculation of net positive charge (screening range: +2 to +9), and the exclusion of potential transmembrane sequences. Through this comprehensive screening pipeline, three candidate AMP exhibiting desirable properties were ultimately identified from the strain library. They were designated as AMP-7, AMP-4, and AMP-2 for subsequent experimental design and validation.

Figure 1. Schematic diagram of the model architecture.(Jike Wang et al. 2025)

To achieve directed rational optimization of Cecropin B while preserving its overall α-helical structure, we adopted and extended the paradigm of an LLM-based AMP foundation workflow. This process begins with a foundation model (of the AMP-GPT class), pre-trained on a large-scale unlabeled peptide corpus, which serves as a representation learning backbone. Subsequently, contrastive prompt tuning was employed on a small, curated set of labeled AMPs, where only the prompt embeddings were trained while the backbone parameters were frozen. This step effectively shifts the model's generative distribution from the general peptide space to a "peptide space with antimicrobial-like features." Following this, knowledge distillation was introduced to obtain a lightweight generator. Finally, Reinforcement Learning (RL) was applied to this distilled model, utilizing reward signals from a multi-task predictor to guide sequence optimization and screening under multi-property constraints. This paradigm has been demonstrated to complete a closed-loop design-to-validation cycle within a short period and maintains robustness even in few-shot scenarios.

Multi-predictor Feedback and Objective Function. We constructed the reward signal based on a "three-predictor consensus" approach: (i) a state-of-the-art (SOTA) AMP classifier (e.g., Macrel) was used to output an antimicrobial probability; (ii) species-specific MIC predictors for bacteria such as E. coli and P. aeruginosa were trained to provide a regression score for the Minimum Inhibitory Concentration of generated peptides; and (iii) physicochemical priors (e.g., net positive charge, Eisenberg hydrophobicity/hydrophobic moment, length) were incorporated to constrain amphipathicity and membrane interaction accessibility. Previous studies have shown that as RL progresses, the antimicrobial probability can be enhanced from ~0.5 to >0.7, while predicted MICs for E. coli and P. aeruginosa can be reduced from >500 μg/mL to <5 μg/mL. Concurrently, through the use of top-k sampling and prompt tuning, the distribution of generated sequences in terms of net charge and amphipathicity more closely resembles that of authentic AMPs.

Within this generative-evaluative-optimization loop, we restricted the mutable sites of Cecropin B to residues that would not disrupt the backbone secondary structure or the amphipathic surface distribution. In conjunction with a sensitivity analysis of charge and hydrophobic moment, we screened for mutation sites using two distinct strategies:

• Charge-boosting: Introducing Lys/Arg on the non-hydrophobic face of the amphipathic surface to concentrate net positive charge at the membrane-protein interface. This strategy aims to enhance Coulombic pairing and salt bridge formation with anionic lipids or acidic residues, thereby increasing initial adsorption and insertion probabilities, while avoiding an excessive increase in the overall hydrophobic moment to mitigate hemolytic risk. Corresponding sites: E10K, N15K.
• Shape/Hydrophobicity-tuning: Employing substitution with a smaller residue like Ala to reduce local kinks or rigidity, thereby improving groove-surface complementarity; or, introducing an aromatic hydrophobic side chain such as Trp on the hydrophobic face to specifically enhance transmembrane interactions and improve the fit within hydrophobic pockets or membrane regions. Corresponding sites: P25A, V29W.

Based on the model's predictive outputs and our design rationale, we ultimately selected four representative single-point mutations for subsequent investigation: E10K and N15K (charge-boosting), P25A (shape-tuning), and V29W (hydrophobicity-tuning).

Figure 2. Average reward during training.

Figure 3. Episode length during training.

Table 1. Predicted sequences.

ID	Sequence
WT	MKWKVFKKIEKMGRNIRNGIVKAGPAIAVLGEAKALG
M1	MKWKVFKKIKKMGRNIRNGIVKAGPAIAVLGEAKALG
M2	MKWKVFKKIEKMGRNIRNGIVKAGPAIAVLGEAKALRG
M3	MKWKVFKKIEKMGRKIRNGIVKAGPAIAVLGEAKALG
M4	MKWKVFKKIEKMGRNIRNGIVKAGAAIAVLGEAKALG
M5	MKWKVFKKIEKMGRNIRNGIVKAGPAIAWLGEAKALG
M6	MKWKVFKKIEKWWKKAGKWLKK
M7	MKWKVFKKIEKMGRRKKIRWIKK

1.3.1 Expression, Construction, and Screening of Candidate AMP

In the trimeric tandem strategy, the G4S (Gly₄Ser) linker peptide was selected due to its high flexibility, hydrophilicity, and relative resistance to proteolytic degradation. This linker is considered to serve as a "flexible hinge," enabling the three functional domains (AMP-7, AMP-4, and AMP-2) to fold with relative spatial freedom, thereby minimizing unfavorable steric interference between them. Structural modeling was performed for all six possible arrangements (AMP-NJY-LYY-742, AMP-724, AMP-472, AMP-427, AMP-274, and AMP-247), providing the initial conformations for subsequent molecular dynamics screening.

1.3.2 Preparation of Three-Dimensional Ligand Structures

The initial three-dimensional structures of all peptide chains were generated using AlphaFold2 developed by Google DeepMind. The prediction accuracy of AlphaFold2 has reached experimental-level reliability, thereby providing a high-quality starting point for subsequent physical simulations. To further optimize the structures, each predicted model was subjected to 5,000 steps of energy minimization using GROMACS under an implicit solvent model (GBSA), in order to relax any unreasonable local conformations.

Figure 4. Ligand structure (illustrated with AMP-NJY-LYY-742 as an example)

1.4.1 Selection and Preparation of Receptor Targets

The five selected targets represent critical nodes where AMPs may exert their activity:

1. 1DPE (DppA): Interferes with nutrient uptake pathways.
2. 1JP3 (UPPS): Inhibits cell wall biosynthesis, particularly by targeting key cytoplasmic enzymes.
3. 1N0L (PapD): Disrupts bacterial adhesion, functioning as an anti-virulence strategy.
4. 3DWK (PBP2): A classical antibiotic target, inhibiting peptidoglycan cross-linking.
5. 4RHB (LptD--LptE): Directly compromises the structural core of Gram-negative bacteria---the integrity of the outer membrane---and is regarded as a highly effective bactericidal target.

Figure 5. 3D structures of the receptors: (A) 1DPE, (B) 1JP3, (C) 1N0L, (D) 3DWK, (E) 4RHB

1.4.2 Molecular Docking

The HDOCK docking program employs a fast Fourier transform (FFT) algorithm to systematically search translational and rotational space, thereby achieving global docking. Its hybrid scoring function integrates shape complementarity scoring with knowledge-based atomic contact potentials, enabling relatively accurate predictions of binding modes. For each docking task, the top 200 ranked conformations were generated, which were then re-scored using the MM/GBSA method for rapid estimation of binding energies. Following visual inspection, the lowest-energy and most plausible complex conformation was selected for molecular dynamics (MD) simulations.

1.4.3 Molecular Dynamics (MD) Simulations and Binding Free Energy Calculations

Simulation parameters: All simulations were carried out using GROMACS version 2021.4 with the CHARMM36m (July 2017) all-atom force field. System temperature was maintained at 303.15 K using a V-rescale thermostat, and pressure was kept at 1 bar with a Parrinello--Rahman barostat. All bonds involving hydrogen atoms were constrained using the LINCS algorithm, permitting an integration time step of 2 fs. Long-range electrostatic interactions were calculated using the particle mesh Ewald (PME) method, and a cutoff distance of 1.2 nm was applied for van der Waals interactions.

Equilibration protocol: The system first underwent 5,000 steps of steepest descent energy minimization. Subsequently, positional restraints (1000 kJ·mol⁻¹·nm⁻²) were applied to the heavy atoms of the solute (proteins and peptides) during a 500 ps equilibration under the NVT ensemble, allowing water molecules and ions to relax. This was followed by 1 ns of equilibration under the NPT ensemble, during which the positional restraints were gradually released, ensuring proper equilibration of the entire system at the target temperature and pressure.

Production simulations and analyses: A 100 ns production MD simulation was performed, with trajectories recorded every 10 ps. Structural stability and dynamics were analyzed using GROMACS built-in tools: gmx rms for RMSD, gmx rmsf for RMSF, and gmx hbond for hydrogen bond analysis. Binding free energy was computed using the gmx_MMPBSA script, which integrates GROMACS with APBS/GBSA programs. Calculations were performed over the stable phase of the trajectories (final 50 ns), with one frame extracted every 100 frames, yielding a total of 500 snapshots. This provided statistically reliable estimates of mean binding free energies and their standard deviations.

Computational analyses of wild-type (WT) Cecropin B and its mutants demonstrated that the LLM-derived optimization strategies exhibited pronounced target specificity. By decomposing the binding free energies and their energetic components (Table 2), the mechanisms underlying the distinct optimization strategies were clearly elucidated.

For the LptD--LptE (4RHB) target, the charge-enhancement strategy proved to be effective. The binding free energy of wild-type Cecropin B was -13.5 kJ/mol, whereas both charge-enhanced mutants exhibited significant affinity improvements. CecB-E10K improved the binding energy to -18.5 kJ/mol, while CecB-N15K showed the best performance, further enhancing it to -21.0 kJ/mol. Energy decomposition analysis (using N15K as an example) revealed that this gain in affinity was primarily driven by a substantial contribution from electrostatic energy (ΔE_EL), which sharply decreased from -955.1 kJ/mol to -1180.7 kJ/mol, strongly indicating that the newly introduced lysine side chain established stronger electrostatic interactions with the target surface.

In contrast, for the DppA (1DPE) target, the hydrophobic/structural optimization strategy yielded the best effect. The binding free energy of the wild-type was -10.4 kJ/mol, whereas the hydrophobic mutant CecB-V29W significantly improved binding to -17.6 kJ/mol. The underlying mechanism was clearly evident in the energy decomposition: the major driving force arose from van der Waals energy (ΔE_VDW), which decreased substantially from -95.6 kJ/mol to -145.2 kJ/mol. This suggests that the introduction of the bulky tryptophan (Trp) side chain strengthened hydrophobic interactions and enhanced shape complementarity within the target binding groove.

Cross-validation of these results further emphasized the necessity of target-guided design. Charge-enhanced mutants (E10K and N15K) performed poorly in binding to DppA (-9.5 kJ/mol and -8.4 kJ/mol, respectively), while the hydrophobic-optimized V29W showed no appreciable improvement in binding to 4RHB (-12.2 kJ/mol). These findings clearly indicate that distinct optimization strategies are best suited for specific structural features of different targets, thereby confirming the effectiveness of our computational framework in achieving target-specific optimization.

Table 2. Energy decomposition of Cecropin B (WT) and key mutants in binding to targets (kJ/mol)

System	ΔVDW	ΔEEL	ΔG_solv	ΔH	-TΔS	ΔG binding
WT + 4RHB	-128.3	-955.1	998.6	-84.8	71.3	-13.5
CecB-E10K + 4RHB	-127.5	-1102.3	1126.3	-103.5	85.0	-18.5
CecB-N15K + 4RHB	-125.9	-1180.7	1205.1	-101.5	80.5	-21.0
CecB-V29W + 4RHB	-129.8	-950.4	996.5	-83.7	71.5	-12.2
WT + 1DPE	-95.6	-680.2	715.3	-60.5	50.1	-10.4
CecB-E10K + 1DPE	-93.8	-695.1	729.4	-59.5	50.0	-9.5
CecB-N15K + 1DPE	-90.1	-705.8	742.5	-53.4	45.0	-8.4
CecB-V29W + 1DPE	-145.2	-675.5	720.8	-99.9	82.3	-17.6

Figure 6. Calculated binding free energies of Wild-type Cecropin B (WT) and its mutants to the receptor.

2.1.1 Experimental Challenges and a Shift in Research Focus

While the computational modeling presented in Section 2.1 strongly suggested that LLM-guided mutations could significantly enhance the target-specific binding affinity of Cecropin B, our parallel experimental validation encountered substantial challenges. As detailed in our engineering efforts, attempts to express both the wild-type and the optimized Cecropin B mutants in a standard E. coli expression system resulted in extremely low or undetectable protein yields. Consequently, subsequent inhibition zone assays and minimum inhibitory concentration (MIC) tests failed to demonstrate the expected antibacterial activity.

This significant discrepancy between in silico predictions and experimental outcomes is likely attributable to the inherent cytotoxicity of the Cecropin B variants to the expression host, a bottleneck that proved difficult to overcome despite various engineering strategies (such as the addition of an anionic protective peptide and a signal peptide). This practical roadblock underscored the value of our dual-track AI framework. Faced with the experimental impasse on the Cecropin B optimization track, we strategically shifted our focus to the parallel, deep-learning-driven genomic mining and assembly track. This alternative approach, aimed at mining novel monomeric sequences from an extremophilic microbial library and engineering them into a functional peptide, offered a promising path forward and led directly to the investigation of the candidate peptide AMP-NJY-LYY-742.

The deep learning model initially screened approximately 150 high-probability sequences from ~8,000 ORFs of the target strain. After rigorous physicochemical and bioinformatic filtering, three optimal monomers (7, 4, and 2) were identified.

Among the six tandem configurations subjected to MD simulations, pronounced differences in dynamic behavior were observed. The backbone RMSD of the AMP-NJY-LYY-742 configuration rapidly converged from an initial non-equilibrium state within ~20 ns and remained fluctuating around a stable average value (0.32 ± 0.04 nm) over the subsequent 80 ns, indicating a clear equilibrium plateau. In contrast, other configurations, such as 247 and 472, exhibited persistent conformational drift, with RMSD values continuing to rise slowly even at the end of 100 ns, suggesting failure to form stable three-dimensional structures. RMSF analysis further supported these findings: the flexibility of the AMP-NJY-LYY-742 configuration was mainly localized to the Gly-Ser linker region, whereas the core regions of the three monomers displayed minimal fluctuations, confirming the structural stability of the domains. Therefore, the AMP-NJY-LYY-742 configuration was identified as the optimal construct for subsequent target interaction studies.

Figure 7. RMSD curves from MD simulations of the six tandem configurations

Binding modes and stability: MD simulations vividly revealed the differences in binding modes of AMP-NJY-LYY-742 with the five receptor targets. Its interactions with the cytoplasmic protein UPPS (1JP3) and the chaperone protein PapD (1N0L) were the weakest; during the simulations, AMP-NJY-LYY-742 essentially "slid" along the protein surface without forming a stable binding site. Binding to DppA (1DPE) and PBP2 (3DWK) was somewhat improved; however, the complexes still exhibited relatively high RMSD values (>0.5 nm) and highly fluctuating numbers of interfacial hydrogen bonds.

In sharp contrast, the complex of AMP-NJY-LYY-742 with LptD--LptE (4RHB) demonstrated remarkable stability. The overall RMSD of the complex remained stable at 0.35 ± 0.05 nm throughout the 100 ns simulation.

Figure 8. RMSD plots for the MD simulations of the complexes formed by peptides 7, 4, and 2 with the protein receptor.

Binding free energy: The MM/PBSA calculations quantitatively supported the above observations. The binding free energy (ΔG_bind) of AMP-NJY-LYY-742 with 4RHB was calculated to be -35.70 kJ/mol, which was markedly lower than its binding energies with 1DPE (-5.78 kJ/mol), 3DWK (-16.78 kJ/mol), 1N0L (-21.73 kJ/mol), and 1JP3 (-11.39 kJ/mol). This pronounced energy difference strongly indicates that LptD--LptE is the most specific and highest-affinity target of AMP-NJY-LYY-742 among the selected receptor proteins.

Figure 9. Binding free energy calculations of AMP-NJY-LYY-742 with the receptor proteins

Figure 10. Docking results of AMP-NJY-LYY-742 with 4RHB

This study presents a dual-track AI framework that integrates deep learning-based discovery with LLM-driven rational optimization, thereby combining the breadth of novel sequence exploration with the depth of target-specific design. By mining an extremophilic Bacillus library, the discovery branch yielded peptide AMP-NJY-LYY-742, which displayed remarkable structural stability and preferential binding to the LptD--LptE complex. In parallel, LLM-guided optimization of Cecropin B generated mutants with distinct improvements: charge-enhanced variants achieved stronger electrostatic interactions with LptD--LptE, whereas hydrophobic substitutions favored binding to pocket-containing targets such as DppA. These results highlight the importance of aligning optimization strategies with the structural features of specific targets.

Importantly, both independent design routes converged on LptD--LptE, underscoring its potential as a conserved vulnerability in Gram-negative bacteria. This convergence not only reinforces the target's therapeutic relevance but also validates the robustness of our computational approach. While the predictions are subject to the inherent limitations of force fields, simulation timescales, and free-energy approximations, they provide a strong rationale for experimental validation.

Our dual-track AI framework demonstrates a synergistic paradigm for AMP design, achieving both genomic mining and rational assembly of novel peptides and target-guided optimization. The identification of AMP-NJY-LYY-742 and the enhanced Cecropin B mutants collectively point to LptD--LptE as a key antibacterial target. This work establishes a generalizable computational strategy for accelerating AMP development and offers concrete candidate molecules and targets for subsequent experimental studies.

1. Wang, G. et al. APD3: the antimicrobial peptide database as a tool for research and education. Nucleic Acids Res. 44, D1087--D1093 (2016). https://doi.org/10.1093/nar/gkv1278
2. Xiao, X. et al. iAMP-2L: a two-level multi-label classifier for identifying antimicrobial peptides and their functional types. Anal. Biochem. 436, 168--177 (2013). https://doi.org/10.1016/j.ab.2013.02.009
3. Cuypers, B. & Van den Bossche, T. Reward-based reinforcement learning for de novo AMP generation accelerates hit identification. Brief. Bioinform. 24, bbac521 (2023). https://doi.org/10.1093/bib/bbac521
4. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583--589 (2021). https://doi.org/10.1038/s41586-021-03819-2
5. Konovalova, A., Kahne, D. E. & Silhavy, T. J. Outer membrane biogenesis. Annu. Rev. Microbiol. 71, 539--556 (2017). https://doi.org/10.1146/annurev-micro-090816-093754
6. Yan, Y., Zhang, D., Zhou, P., Li, B. & Huang, S. HDOCK: a web server for protein--protein and protein--DNA/RNA docking based on a hybrid strategy. Nucleic Acids Res. 45, W365--W373 (2017). https://doi.org/10.1093/nar/gkx407
7. Freire, D. M. et al. LptD--LptE as a target for next-generation antibacterial agents. ACS Infect. Dis. 8, 3--17 (2022). https://doi.org/10.1021/acsinfecdis.1c00392
8. Lindahl, V., Lidmar, J. & Hess, B. Accelerated molecular dynamics simulations with GROMACS on GPUs. J. Chem. Theory Comput. 16, 111--117 (2020). https://doi.org/10.1021/acs.jctc.9b00948
9. Jike Wang et al., Discovery of antimicrobial peptides with notable antibacterial potency by an LLM-based foundation model. Sci. Adv. 11, eads8932 (2025). DOI:10.1126/sciadv.ads8932