A Foundation Model Identifies Broad-Spectrum Antimicrobial Peptides against Drug-Resistant Bacterial Infection

Antimicrobial resistance (AMR) is a major global health threat, projected to cause up to 10 million deaths annually by 2050. The antibiotic development pipeline is sparse, with few novel approvals since the late 1980s and high development costs and timelines. AMPs (10–50 amino acids) are promising alternatives due to membrane disruption mechanisms that can eradicate resistant bacteria. However, traditional experimental discovery of AMPs is slow and costly. The study addresses this gap by proposing a rapid, machine learning–driven framework to identify potent, broad-spectrum AMPs with reduced propensity for resistance.

Prior computational approaches for peptide discovery include virtual screening, de novo design, genetic/evolutionary algorithms, and deep generative models. Challenges include: limited and noisy labeled AMP datasets; feature extraction and generalization limits in screening models; unconstrained generation yielding hard-to-synthesize candidates and high experimental burden; and data scale disparity relative to NLP that hinders deep models from learning relevant biochemical properties. Recent successes of language models inspired a peptide language-model approach to overcome small data constraints via pretraining and fine-tuning.

The deepAMP framework consists of four components: (1) deepAMP-general (pretrained peptide language model) for rational peptide generation; (2) deepAMP-AOM (antimicrobial optimization model) fine-tuned to convert low- to high-activity AMPs; (3) deepAMP-POM (penetratin optimization model) re-fine-tuned to enhance broad-spectrum activity and membrane permeability; and (4) deepAMP-predict (virtual screening) to prioritize candidates prior to experiments. - Pretraining: Masked language model trained on 300,000 peptide sequences (UniProt, lengths 10–50) to learn peptide syntax and generate valid, diverse non-functional peptides. - Sequence degradation and paired data construction: To overcome scarcity of paired optimization data, high-activity AMPs are degraded using deepAMP-general by random masking (≤30%) and decoding to generate lower-activity variants, forming pairs. Datasets included: (i) 321 highly active AMPs (GRAMP3; E. coli MIC <2.5) degraded 100× to yield 24,928 pairs for AMP optimization; (ii) penetratin-related high-activity sequences (29 sequences, MIC <150 μM, length 20–29) degraded to generate ~1000 <low, high> activity pairs; and (iii) construction of penetratin training pairs for fine-tuning deepAMP-POM. - Fine-tuning: deepAMP-AOM trained on AMP pairs; deepAMP-POM initialized from AOM and fine-tuned on penetratin pairs using small learning rates (batch size 32, 200 epochs, initial LR 1e-5, decay 1e-3). A Temporin-AI optimization model (deepAMP-TOM) was similarly fine-tuned for benchmarking. - Virtual screening: SVM classifier (TREC features, 400-dim) trained on 6,760 positive (GRAM-P) and 6,760 negative (UniProt) peptides (test: 500 pos/500 neg) achieved AUC 0.96 (5-fold CV). Used to rank generated candidates (deepAMP-predict). - Computational evaluation: Iterative optimization rounds with fitness scoring for amphipathic α-helices; UMAP visualization of sequence feature space distributions; mutation (Hamming) distances to assess novelty (average 4 mutations from training set for top 28 predicted AMPs). - Experimental validation: Two design rounds from penetratin template. Round 1: 92 candidates; top hits T1–T8 selected for in vitro MIC/MBC assays against S. aureus, E. coli, K. pneumoniae, P. aeruginosa, and MRSA; hemolysis on human RBCs and MTT cytotoxicity on NHDF and keratinocytes; mechanism-of-action assays (PI staining/confocal and flow cytometry; SEM morphology; membrane depolarization DiSC(5,5); NPN outer membrane permeability). Round 2: 11 second-generation peptides derived from T15 with broader testing. In vivo: murine excision wound model infected with P. aeruginosa; treatment with T1-2, T1-5, T2-10 or levofloxacin at 10 mg/kg; CFU quantification and histology; safety assessments including blood parameters and organ histology. Serial passaging (25 days) assessed resistance development vs. ciprofloxacin.

- Model performance and optimization: - deepAMP-TOM outperformed random mutation, Baseline-T, and HydroAMP in temporin optimization across three iterations; deepAMP-GOM achieved top fitness score 0.594 and surpassed Baseline-G by iteration 3 in the Pg-AMP1 benchmark. - Top predicted AMPs exhibited chemical novelty with average mutation distance of 4 from training AMPs; distribution analyses (UMAP) showed diverse chemical space coverage. - Antibacterial activity (in vitro): - >90% of designed AMPs showed improved antibacterial activity compared to penetratin across Gram-positive (S. aureus) and Gram-negative bacteria (K. pneumoniae, P. aeruginosa, E. coli). - Multiple first-generation candidates (e.g., T1, T5, T6) showed MIC 15.63 μg/mL against K. pneumoniae; T1 and T5 showed strong activity against S. aureus (MIC ~11.7–15.63 μg/mL) and MRSA (MIC 15.63 μg/mL), up to 64-fold lower than penetratin (>1000 μg/mL). - Table data indicate potent activity across T1–T8 and T2-series; T2-9 was highlighted as the strongest with activity comparable to FDA-approved antibiotics. - MBC values for several peptides (e.g., T1–T8) matched MICs, indicating rapid bactericidal action; T1 and T8 fully eliminated bacteria at MIC. - Safety: - Several lead AMPs (e.g., T1–T6, T1, T2, T9) exhibited low RBC hemolysis (<20%). - Low cytotoxicity on NHDF and human keratinocytes with IC50 >50 μg/mL, exceeding MIC values (favorable therapeutic window). - Mechanism of action: - PI staining and flow cytometry indicated increased membrane permeability and bacterial death. - SEM showed membrane damage and surface wrinkling after treatment (e.g., T1-2, T2-9, T2-10). - DiSC(5,5) depolarization assay: fluorescence after AMP treatment (e.g., T1-2, T1-5, T2-10) increased; T2-10 produced 19.5× higher fluorescence than polymyxin B control, indicating strong membrane depolarization. - NPN uptake doubled vs. untreated and was comparable to polymyxin B, indicating outer membrane permeabilization. - Resistance development and biofilm: - Over 25 passages, S. aureus developed resistance to ciprofloxacin by passage 10, whereas T1-2, T1-5, and T2-10 did not induce detectable resistance after 25 passages. - These AMPs inhibited biofilm formation in S. aureus and E. coli (crystal violet assay). - In vivo efficacy: - In a mouse excision wound infection model (P. aeruginosa), treatment with T1-2, T1-5, and T2-10 (10 mg/kg) significantly reduced bacterial loads, comparable to levofloxacin, with supportive histology and safety assessments. - Virtual screening: - deepAMP-predict SVM achieved AUC 0.96, enabling effective prioritization of candidates prior to wet-lab testing.

The study addresses the urgent need for new therapeutics against drug-resistant bacteria by leveraging language-model pretraining and targeted fine-tuning to discover AMPs with broad-spectrum potency and low propensity for resistance. By constructing paired datasets via sequence degradation, deepAMP learns transformations from low- to high-activity peptides, improving optimization efficiency and generalization despite limited labeled data. Experimental validation demonstrates that the designed AMPs disrupt bacterial membranes (depolarization and outer membrane permeabilization), leading to rapid bactericidal effects, reduced biofilm formation, and low likelihood of resistance evolution compared with a standard antibiotic (ciprofloxacin). The in vivo wound model confirms therapeutic potential. Compared to prior virtual screening and de novo generative approaches, deepAMP balances diversity with synthesizability and integrates predictive screening to reduce experimental burden, underscoring its relevance for AMP and broader peptide therapeutic discovery.

This work presents deepAMP, a peptide language model–based framework that discovers broad-spectrum, potent AMPs with low resistance potential. More than 90% of designed peptides outperformed the penetratin template in MIC assays across Gram-positive and Gram-negative pathogens. Lead candidates (e.g., T1-2, T1-5, T2-10) showed strong membrane-disruptive mechanisms, minimal resistance development over serial passages, and effective bacterial clearance in a mouse wound infection model. The approach demonstrates how large-scale pretraining, paired-data fine-tuning via sequence degradation, and virtual screening accelerate AMP discovery. Future directions include integrating 3D structural information (e.g., AlphaFold-derived features), improving interpretability, scaling experimental validation to larger candidate sets, and evaluating efficacy in additional infection models (e.g., deep thigh and pneumonia).

- Current sequence-based models do not capture 3D structural/conformational information; incorporating structural data may improve accuracy. - Model interpretability is limited (black-box), hindering mechanistic feature attribution. - Experimental validation set size is relatively small (initially 29 high-activity penetratin-related sequences and limited lead testing), necessitating broader validation. - In vivo testing was limited to a wound infection model; additional models (e.g., deep tissue and lung infections) are needed to generalize efficacy and safety.