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

Model pre-training and fine-tuningWe follow the idea of pre-training and fine-tuning (network parameters in Supplementary Tables 1, 2), which has been shown to work well on tasks with small datasets. To better fit the task requirements, we divided the model into 3 parts, (1) the peptide generative model (deepAMP-general), (2) the antimicrobial peptide optimization model (deepAMP-AOM), and (3) the penetratin optimization model (deepAMP-POM). The model details are described in the following.Data sets of AMPsInitially, we aimed to obtain a training dataset de novo that was inexpensive, chemically diverse, and did not require sophisticated laboratory resources. This would allow for the development of a robust model with which new AMPs could be predicted and optimized, without the practical hurdles that can be associated with large-scale antibiotic screening efforts.We trained the three parts of the model separately with different datasets (Supplementary Table 3). (1) We constructed a pre-trained peptide dataset using the database Uniprot, which contains a large amount of protein and peptide data from different organisms and different sources, thus providing a diverse data resource for peptide research. From this, data with sequence lengths as close to peptides as possible (length range 10 ~ 50) are filtered, including 300,000 peptide sequences. (2) In order to meet the demand for antimicrobial peptide optimization, the model needs to use paired data. However, the optimized peptides in the available data differ significantly from the original peptides, in addition, there are rarely paired data samples with equivalent length and sequence similarity. For this reason, we designed an antimicrobial peptide sequence degradation method to degrade the high-activity samples into low-activity samples. We selected 321 antimicrobial peptides with relatively high activity (for E. coli MIC below 2.5) for degradation in the GRAMPA19 dataset, where each sequence was randomly masked with up to 30% of the length of the site and was sampled 100 times. Finally, we generated 24,928 pairs containing <low activity, high activity> antimicrobial peptide optimization data set. (3) Since penetratin has been shown to have good cell permeability and average antimicrobial activity, our ultimate goal is to optimize penetratin to achieve high activity antimicrobial effect. Nevertheless, there is currently no constructed dataset for penetratin. Here we use a data mining approach to collect data on penetratin, which includes both the original and the penetratin-optimised analogs. The data of 54 peptides against eight bacterial species were collected. After screening this part of the dataset, 29 penetratin sequences with high activity (MIC value lower than 150 μg/ml) and sequence length within 20 were selected. Then, sequence degradation was performed on these 29 sequences to generate an optimized dataset of 1009 pairs of <low activity, high activity> samples. The distribution of the predicted scores for the training set is shown in Supplementary Fig. 48.Generalized Peptide Generative Model (deepAMP-general)The framework of the generative model is composed of an encoder and a decoder, where we use the Transformer module based on a self-attention mechanism for the encoder and a dense network for the decoder. In the polypeptide generative model, the model uses a constructed dataset containing 300,000 peptides to train the generative model in an unsupervised manner. Since our final task requirement is to optimize the antimicrobial peptide sequences, which means only the amino acids at specific sites in the sequence need to be changed to alter the activity of the sequence, we use the masked language model (MLM) training through similar to the BERT model. Suppose the existing samples \({S}_{i}=\left\{{s}_{0},{s}_{1},\ldots,{s}_{n}\right\}\), we randomly mask no more than 30% of the tokens in the sequence, and the input samples of the masked model are \({M}_{i}=\left\{{m}_{0},{m}_{1},\ldots,{m}_{n}\right\}\), the training goal is to encode the input samples and then decode the masked token \({m}_{k}\) into the original one by the decoder. By decoding the token \({m}_{k}\) into the original token \({s}_{k}\), and the objective function of this part is:$$L\left(S\right)={\sum}_{i}{\sum}_{k}\log P\left({s}_{k} | {m}_{0},\ldots,{m}_{n}{{\rm{;}}}\, \theta \right),$$
(1)
where \({{\rm{n}}}\) denotes the sequence length, k denotes the token position of the mask, \(\left\{0\le k\le n\right\}\), the number of \(k\) ranges from \((0,\lceil 30\%\ast n\rceil )\), and the conditional probability \(P\) is a generative model function with parameter θ.The AMPs optimization model (deepAMP-AOM)deepAMP-AOM is fine-tuned using deepAMP-general and initialized using the deepAMP-general parameters, in the fine-tuning stage, due to the difference in the amount of training sample data. To avoid overfitting the model and not learning effectively small data, it is often necessary to set a lower learning rate and increase the decay rate of weights compared to the pre-training phase. Finally, several training hyperparameters are applied to the antimicrobial peptide optimization model AOM: a batch size of 128, number of training rounds of 200, initial learning rate of \(6\times {10}^{-5}\), and weight decay of \(1\times {10}^{-3}\). The deepAMP-AOM model is different from the unsupervised training of deepAMP-general. In this step we utilize pairs of <low activity, high activity> training samples, the string of the input mask is the low activity sample, and the target string in decoding is the string of the mask position corresponding to the high activity sample. An optimized training set \( < \, {S}_{i},{T}_{i} \, > \) is used, where \({S}_{i}=\left\{{s}_{0},{s}_{1},\ldots,{s}_{n}\right\}\) is the low-activity antimicrobial peptide and \({T}_{i}=\left\{{t}_{0},{t}_{1},\ldots,{t}_{n}\right\}\) is the high-activity antimicrobial peptide. The model randomly masks no more than 30% of the token of low activity antimicrobial peptides, and the masked model input is \({M}_{i}=\left\{{m}_{0},{m}_{1},\ldots,{m}_{n}\right\}\). The goal of training is to encode the whole input sample, and after the decoder can decode the masked token \({m}_{k}\) back to the corresponding position of high activity antimicrobial peptide token \({t}_{k}\). The objective function is:$$L\left(S\right)={\sum}_{i}{\sum}_{k}\log P\left({t}_{k} | {m}_{0},\ldots,{m}_{n}{{\rm{;}}}\, \theta \right),$$
(2)
where \({{\rm{n}}}\) denotes the sequence length, k denotes the token position of the mask, \(\left\{0\le k\le n\right\}\), the number of \(k\) ranges from \((0,\lceil 30\%\ast n\rceil )\), and the conditional probability \(P\) is a generative model function with parameter \(\theta\).Penetratin optimized model (deepAMP-POM)The antimicrobial peptide contains penetratin, and with deepAMP-AOM as an intermediate layer, the model is already capable of the task of generating high-activity antimicrobial peptides, at which point we again fine-tune the parameters (Supplementary Data 5) in deepAMP-AOM to be used in the initialization of deepAMP-POM. In the training process we also classify penetratin sequences into two categories, high activity and low activity. The process and objective function are the same as deepAMP-AOM, but the data set of penetratin is small, so we need to use a smaller learning rate to learn. By adjusting the training process, we finally get a batch size of 32. The number of training rounds is 200, the initial learning rate is \(1\times {10}^{-5}\) and the weight decay is \(1\times {10}^{-3}\) .Temporin-Ali optimized model (deepAMP-TOM)For the optimization task of Temporin-Ali, we fine-tuned deepAMP-AOM on the Temporin-Ali analog dataset to achieve better optimization. In the training process, we also classify Temporin-Ali sequences into two categories, high activity and low activity. The process and objective function are the same as deepAMP-AOM, but the data set of Temporin-Ali is small, so we need to use a smaller learning rate to learn. By adjusting the training process, we get a batch size of 32. The number of training rounds is 200, the initial learning rate is \(1\times {10}^{-5}\) and the weight decay is \(1\times {10}^{-3}\).Guavanin antimicrobial peptide optimized model (deepAMP-GOM)For the optimization task of Pg-AMP1 fragment, we fine-tuned deepAMP-AOM on the guavanin antimicrobial peptide analog dataset22 to achieve better optimization. In the training process, we also classify guavanin antimicrobial peptide sequences into two categories, high activity and low activity. The process and objective function are the same as deepAMP-AOM. By adjusting the training process, we get a batch size of 32. The number of training rounds is 200, the initial learning rate is \(1\times {10}^{-5}\) and the weight decay is \(1\times {10}^{-3}\).DeepAMP-based Virtual ScreeningEvaluating the optimization capability of a model through virtual screening is a general method, and we used the Support Vector Machine (SVM) classifier to evaluate the antimicrobial activity of the data generated by the optimization model in advance of the wet experiment. The model is trained to distinguish the activity of antimicrobial peptides based on transforming amino acid strings into a 400-dimensional feature vector representation using the RECM feature extraction method. We collected GRAMPA19 data as positive samples with antimicrobial activity and Uniprot peptide data as negative samples without antimicrobial activity, resulting in an antimicrobial peptide identification dataset of 6,760 positive and 6760 negative samples. Among them, 500 positive samples and 500 negative samples are assigned to the test set, and the other is the training set. The final AUC of the model is 0.96 using five-fold cross-validation, which indicates that the activity of peptide sequences can be evaluated to some extent by virtual screening.Sequence degradationSince optimization requires paired samples for learning, however, such data is lacking in the existing dataset. The deepAMP-general we trained is a peptide generative model, in which most of the trained peptides are inactive. Therefore, we resorted to deepAMP-general to degrade the activity of highly active sequences. Specifically, the steps of degenerating a highly active sequence are (1) random masking of the highly active sequence, the number of masks ranges from [1, length*30%). For example, for a sequence of length 10, the number of masks is at least 1 and at most 3, the mask positions are randomly combined among ten positions; (2) the sequence obtained after masking will be mapped to the hidden space of peptide by the encoder of deepAMP-general. Finally, the amino acid symbols at the mask positions will be parsed by the decoder of deepAMP-general. In this study, the process will be repeated more than 100 times for each sequence.Fitness functionThis function is used to evaluate the amphipathic α-helices of the peptide and is designed based on the ratio of Eisenberg’s hydrophobic scale32 and the sum of exponential α-helix propensity in Pace–Schols scale33, as in Eq. 3:$${Fitness}=\frac{\sqrt{{\left[{\sum }_{i=1}^{I}{H}_{i}\times \cos \left({\delta }_{i}\right)\right]}^{2}+{\left[{\sum }_{i=1}^{I}{H}_{i}\times \sin \left({\delta }_{i}\right)\right]}^{2}}}{{\sum }_{i=1}^{I}{e}^{H{x}_{i}}},$$
(3)
where \(\delta\) represents the angle between the amino acid side chains (100° for α-helix, on average); \(i\) the residue number in the position \(i\) from the sequence; \({H}_{i}\) the ith amino acid’s hydrophobicity on a hydrophobicity scale; \({{Hx}}_{i}\) the ith amino acid’s helix propensity in Pace–Schols scale; and \(I\) the total number of residues present in the sequence.Peptide novelty calculationTo calculate the number of mutation sites we use the Hamming Distance, which represents the number of amino acids at different positions between two polypeptide sequences (for unequal sequences we use left-alignment for comparison). Assuming that there are two equal length peptide sequences \({{\rm{A}}}\) and \({{\rm{B}}}\) which are of length \(L\), the number of mutation sites (Mutation Count) can be expressed as:$$M{{\rm{utation\; Count}}}={\sum }_{i=1}^{L}\delta \left({A}_{i},\,{B}_{i}\right)+\left|{len}(A)-{len}(B)\right|,$$
(4)
where \(\delta \left({A}_{i},\,{B}_{i}\right)\) denotes that at position \(i\). Whether the amino acids of peptide sequences \({{\rm{A}}}\) and \({{\rm{B}}}\) are different. If different, \(\delta \left({A}_{i},\,{B}_{i}\right)\) is equal to 1, and if the same \(\delta \left({A}_{i},\,{B}_{i}\right)\) is equal to 0, \({len}(A)\) denotes the length of the peptide sequence \(A\), and \({len}(B)\) denotes the length of the peptide sequence \(B\). The number of mutation sites is derived from the sum of all positions plus the difference in sequence lengths.Peptide synthesisStandard Fmoc solid-phase peptide synthesis was used to synthesize all peptides via a CSBio 163 peptide synthesizer. The resin-bound peptide underwent cleavage using a cocktail of TFA/thioanisole/H2O (95:2.5:2.5) for 3 h. Collecting the filtrate and the resin mixture was washed with excess TFA. The crude peptide was obtained by concentrating and precipitating the filtrate with cold ether. Subsequently, crude peptide purification was performed through reverse phase HPLC (Elite P230A/P) using a semi-preparative C18 column. After that, all peptides were lyophilized and analyzed via analytical HPLC and ESI (+) mass.Minimal inhibition concentration (MIC)The bacteria S. aureus (ATCC 25904), E. coli (ATCC BAA-1025), K. pneumoniae (ATCC27736), P. aeruginosa (ATCC 47085) and MRSA (ATCC BAA-1717) were kindly donated by Prof. Xinxin Feng’s group. 100 µL of 5 × 105 CFU/mL bacteria in Mueller Hinton Broth (MHB, Hopebio, HB6231-1) or MHB (bacteria free) was added to a 96-well. 20 mg/mL peptide stock solution was added to the first well and given serial twofold dilutions to obtain final concentrations of 0.97, 1.95, 3.90, 7.81, 15.63, 31.25, 62.5, 125, 250, 500, and 1000 µg/mL. Plates were incubated at 37 °C for 20 h. Bacteria growth was monitored by measuring the OD600 using a microplate reader (Molecular Devices, Spectra Max M5). The MIC value was the minimum corresponding concentration at which the OD600 value was equal to the negative control (bacteria-free MHB). Then, 20 μL of the bacterial suspension or diluent was evenly spread on an LB plate (BD Diagnostics, 221994) and incubated at 37 °C overnight, followed by colony counting. After MIC assay, 20 µL of samples under greater than or equal to MIC concentrations were sub-cultured overnight at 37 °C on LB Agar plates. The MBC value was the minimum concentration corresponding to no bacteria growth.Cell ViabilityAn MTT assay was employed to assess in vitro cytotoxicity of AMPs. NHDF and SHED cells were seeded a plated in 96 well plates at a density of 5000 cells/well and allow to adhere overnight at 37 °C, 5% CO2. The culture medium DMEM with 10% FBS and 1% P/S was replaced with culture medium containing 1, 2.5, 5, 10, 25, 50 µg/mL peptides. Blank medium was used as a negative control. After 24 h incubation period, cells were washed and added 10 µL of MTT solution and samples incubated for 3 h to form formazan. After that, the culture medium was replased with 100 µL DMSO for absolutely dissolved formazan. Absorbance was recorded at 570 nm using a microplate reader (Molecular Devices, Spectra Max M5). The percent viability was calculated as follows: (Absorbance peptide-treated cells-Absorbance blank) / (Absorbance untreated cells-Absorbance blank) ×100. GraphPad Prism 9 software was used to fit cytotoxicity curves and IC50 values were calculated with non-linear regression model.Hemolytic AssayHuman whole red blood cells were centrifuged to obtain RBCs that were then washed and resuspend in PBS to 0.25% (v/v). In a 96-well plate, 75 µL of the RBCs solution was mixed with an equal volume of 2 × peptide dissolved in PBS to initiate the assay. Blank or 1% Triton – X100 containing PBS were used as negative and positive controls, respectively. Samples were incubated for 1 h with gentle shaking, followed by centrifugation at 2200 g for 10 min at 4 °C.100 µL of the supernatant from each well was added to an empty 96-well plate. The hemolysis was detected by absorbance at 415 nm using a microplate reader (Molecular Devices, Spectra Max M5), and calculated using the following equation:$$ (({Absorbance}_{{treated\; hRBC}}{-}{Absorbance}_{{untreated\; hRBC}})\\ \quad/({{Absorbance}}_{{triton}-X100{treated\; hRBC}}{-}{{Absorbance}}_{{untreated\; hRBC}}))\times 100.$$Circular Dichroism (CD) SpectroscopyDOPC and DOPS were mixed equally and dried to a film with argon and subsequently lyophilized overnight. After that, suitable PBS was added and the liposome with a size of 100 nm (2.5 mM) was formed by ultrasound, vortex, and extrusion techniques. Peptide solutions (50 μM) were used to yield a peptide/lipid ratio of 1:50. CD wavelength spectra of the peptides were measured from 260 to 200 nm using a 1 mm path length quartz cell at 37 °C. Mean residue ellipticity, [θ], was calculated using the following equation: [θ] = (θobs/10lc)/r, where θobs is the observed ellipticity in millidegrees, l is the length of the cell (cm), c is the concentration (M), and r is the number of residues.Morphology study by SEM imagingA suspension of bacteria in the exponential growth phase was incubated with AMPs (1 × MIC) for 12 h. A bacterial suspension without any treatment was used as the negative control. After centrifuging (1500 g, 12 min) and removing the supernatants, bacteria were fixed at 4 °C with PBS containing 2.5% glutaraldehyde (Sinopharm, 30092436). Samples were washed three times with PBS and then dehydrated using a series of ethanol solutions (30, 50, 70, 90, and 100%). The samples were sputter-coated with gold for observation using a Hitachi S-4800 field emission scanning electron microscope.Effects of AMPs on bacterial membrane permeabilityBacteria were collected and washed three times with PBS. Then bacteria were resuspended to a working concentration (OD600 = 0.1) and treated with AMPs at 1 × MIC for 5 h at 37 °C. After staining with 20 μg/mL PI for 15 min, the samples were transferred to a glass slide for confocal imaging (FV1200, Olympus) or detected red fluorescence of PI on FL2-PE via flow cytometry (Accuri C6 Plus, Becton Dickinson). Seven pictures were taken for each treatment group, and the percentage of PI-stained bacteria was calculated using the following equation:$${{\rm{PI}}}-{{\rm{stained}}}\; {{\rm{bacteria}}}\%\,=\, {{{\rm{Number}}}\; {{\rm{of}}}\; {{\rm{bacteria}}}}_{{{\rm{Stained}}}\; {{\rm{with}}}\; {{\rm{PI}}}}\\ /{{{\rm{Number}}}\; {{\rm{of}}}\; {{\rm{bacteria}}}}_{{{\rm{Stained}}}\; {{\rm{with}}}\; {{\rm{DAPI}}}}\times 100.$$
(5)
Molecular dynamics simulations of peptide T1-2 with a model membraneWe used the original PDB ID: 1OMQ as a starting point for the peptide structure and mutated it to the current sequence (T1-2). The bacterial membrane was modeled using a composition of 60% POPC, 20% POPG, and 20% cholesterol, which mimics the key components of the membrane. At the beginning of the simulation, the peptide was placed onto the membrane with its helical axis parallel to the membrane plane. The simulations were performed for 500 ns using NAMD v2.14, with a 1.2 nm cut-off for interactions and 150 mM NaCl ion concentration. The CHARMM36m force field was used throughout the simulations34,35. The lipid order parameter (Szz) was measure as \({S}_{{zz}}= < \frac{{3\cos }^{2}\theta -1}{2} > \), where θ is the angle between the \({C}_{a-1}-{C}_{a+1}\) vector and the z-axis.Membrane depolarization assayThe cytoplasmic membrane depolarization activity of the peptides was determined by measurements of fluorescence of the membrane-potential-sensitive dye DiSC3(5). Briefly, E. coli was centrifuged and washed three times with washing buffer (20 mM glucose, 5 mM HEPES, pH 7.2). Then resuspended to a working concentration (OD600 = 0.05) OD600 in the working buffer (20 mM glucose, 5 mM HEPES, 0.1 M KCl, pH 7.2). Thereafter, 100 μL bacteria were incubated with 20 nmol/L DiSC3(5) (53213-94-8, Aladdin) until a stable reduction of fluorescence was achieved (about 15 min). Samples were treated with 1 × MIC AMPs and membrane depolarization was then monitored by observing the change in the fluorescence emission intensity of DiSC3(5) (λex = 622 nm, λem = 670 nm).Permeabilization of outer membranesThe membrane permeability of the peptides was determined by using the NPN uptake assay. E. coli was grown to an OD600 of 0.5, centrifuged (9168 g at 4 °C for 10 min), washed, and re-suspended in buffer (5 mM HEPES, 5 mM glucose, pH 7.4) to a working concentration (OD600 = 0.05). Then, samples were treated with 10 μM NPN solution (P110559, Aladdin) in a 96-well plate. The background fluorescence was recorded at λex = 350 nm and λem = 420 nm. 100 μl peptide solution was added to the 96-well plate and reached 1 × MIC as a final concentration. Fluorescence was recorded as a function of time until no further increase in fluorescence was observed.Bacterial resistance development assaysFor serial passage evolution, S. aureus was grown overnight in 3 mL MHB and diluted 1/10,000 by fresh MHB. 100 μL of bacteria were added to a 96-well plate, in the presence of varying concentrations of AMPs or ciprofloxacin (C861180, Maclin) at two-fold serial dilutions. Plates were incubated at 37 °C without shaking for 24 h, at which time they were read at 600 nm using a microplate reader (Molecular Devices, Spectra Max M5). After 24 h, bacteria that grew in the presence of the highest concentration of AMPs or ciprofloxacin were diluted 1/10,000 by fresh MHB, and once again introduced to varying concentrations of AMPs or ciprofloxacin at two-fold serial dilutions. This procedure was performed every 24 h over 25 days.Antibiofilm assays100 μL 106 CFU/mL bacteria suspension in MHB was treated with 1 × peptide solutions in a 96-well plate and incubated at 37 °C for 72 h. Then, the culturing medium in the wells was discarded and washed with PBS 3 times. Then fixed with 100 µl methanol (C861180, Sinopharm) for 30 min. The biofilm was air-dried and subsequently stained with 0.1% (w/v) crystal violet (C861180, Solarbio) for 30 min and then further washed with PBS. Thereafter the crystal violet was solubilized in 100 µL 30% acetic acid and the absorbance at 595 nm was measured with a microplate reader (Molecular Devices, Spectra Max M5).Wound Mouse ModelAnimal experiments were carried out complying with National Institutes of Health (NIH) guidelines for the Care and Use of Laboratory Animals, and the study protocol was approved by the Institutional Animal Care and Use Committee of Hunan University (HNU-IACUC-2021-102) under Dr. Junfeng Shi. Female ICR mice (6 to 8 weeks) were obtained from the SJA Laboratory Animal Co Ltd (Hunan, China). The mice were maintained under controlled conditions of temperature 22 °C ± 2 °C, relative humidity 50 ± 10%, and light–dark cycle 12 h.Forty female ICR mice (6 to 8 weeks, each weighing 28 to 35 g) were anesthetized with an intraperitoneal injection of chloral hydrate (50 mg/kg). A 1.0 cm2 open excision wound was created on the dorsal side of the skin of each mouse, reaching the depth of the loose subcutaneous tissue. P. aeruginosa suspensions (107 CFU) were inoculated on the excision wound area to establish the wound infection model. Mice were separated into five groups (PBS, T1-2, T1-5, T2-10, and levofloxacin), N = 8 mice per group. The treatment dose and frequency were 10 mg/kg and twice a day for 3 days, respectively. Treatments were started 1 h after infection by applying 60 μL of compound solutions or PBS to the wound area. On day 2, the bacterial load on the wound surface of the survived mice was determined by wiping the full wound surface with a sterile swab to transfer the bacteria to PBS and plating the PBS agar plate (supplemented with 20 μg/ml of ampicillin) for CFU determination. On day 3, mice were euthanized. The wound area and organ (liver, spleen and kidney) were collected for H&E staining analysis. Blood was withdrawn for evaluation of RBC parameters after treated with PBS, T1-2, T1-5, T2-10 and levofloxacin for 24 h by intraperitoneal injection. The statistical significance of differences in surface bacterial load was determined by the one-way ANOVA with Tukey’s multiple comparison test…Reporting summaryFurther information on research design is available in the Nature Portfolio Reporting Summary linked to this article.