Cell-free biosynthesis combined with deep learning accelerates de novo-development of antimicrobial peptides
Amir PandiiD, David Adam, Amir ZareiD, Van Tuan TrinhiD, Stefan L. SchaeferiD, Marie Burt, Björn KlabundeiD, Elizaveta BobkovaiD, Manish KushwahaiD, Yeganeh ForoughijabbariiD, Peter R. BrauniD, Christoph K. SpahniD, Christian PreußeriD, Elke Pogge von StrandmanniD, Helge B. BodeiD, Heiner von ButtlariD, Wilhelm BertramsiD, Anna Lena JungiD, Frank AbendrothiD, Bernd SchmeckiD, Gerhard HummeriD, Olalla VázqueziD, Tobias J. ErbiD
Abstract
Bioactive peptides are key molecules in health and medicine. Deep learning holds a big promise for the discovery and design of bioactive peptides. Yet, suitable experimental approaches are required to validate candidates in high throughput and at low cost. Here, we established a cell- free protein synthesis (CFPS) pipeline for the rapid and inexpensive production of antimicrobial peptides (AMPs) directly from DNA templates. To validate our platform, we used deep learning to design thousands of AMPs de novo. Using computational methods, we prioritized 500 candidates that we produced and screened with our CFPS pipeline. We identified 30 functional AMPs, which we characterized further through molecular dynamics simulations, antimicrobial activity and toxicity. Notably, six de novo-AMPs feature broad-spectrum activity against multidrug-resistant pathogens and do not develop bacterial resistance. Our work demonstrates the potential of CFPS for production and testing of bioactive peptides within less than 24 hours and <10$ per screen.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 173
- FWCI
- 31.50
- Field percentile
- top 1%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (23 Dec 2022) · CC BY 4.0
Autoencoder design plus a roughly 24-hour cell-free screen tested 500 candidates over five rounds, finding 30 functional AMPs; six were broad-spectrum with no resistance emerging in E. coli.
Deep learning can propose many antimicrobial peptide candidates, but chemical synthesis and cell-based expression limit how many can be tested, and potent peptides may kill the producer strain. Cell-free expression from linear DNA avoids cloning and purification, allowing screening at under 10$ per individual assay.
Key results
ShownFive design-predict-build-test cycles of cell-free screening turned 500 candidates into 30 functional AMPs, with the success rate rising from 0% in the first round to 12.7% in the fifth.
About 500,000 sequences were sampled and about 50,000 remained after filtering; 500 were screened and 30 were functional. Success rose from 0% to 12.7% between round one and round five. No TIR bias was found between the 500 and the 30. · De novo AMP design using deep learning; Cell-free biosynthesis enables rapid screening…
ShownThe functional AMPs are unrelated to natural sequences but follow typical AMP design principles, with mostly helical predicted folds and a more aromatic hydrophobic core.
BLAST gave no significant similarity to UniProt (~240 million entries) or to training AMPs. AlphaFold predicted a helix for 27 of 30 sequences. Phenylalanine was overrepresented and leucine underrepresented relative to natural AMPs. · Functional AMPs are unique but share properties…
SuggestedMolecular dynamics simulations suggest all 30 designed AMPs prefer a bacterial inner-membrane model over a human plasma-membrane model, with the degree of preference varying by peptide.
GROMACS 2020.3 with CHARMM36m, 1 μs per replicate at three lateral tensions. IM binding was stable and complete within at most 200 ns, whereas PM binding was mostly transient and peptides tended to unfold. · De novo AMPs prefer bacterial over human…; Molecular dynamics (MD) simulations
ShownOf the 22 AMPs that could be chemically synthesized, many were potent against E. coli and B. subtilis and showed low hemolysis and cytotoxicity, so sixteen were advanced.
Twenty AMPs had MIC ≤6 μM on B. subtilis and fifteen had MIC ≤25 μM on E. coli. Fifteen had HC50 >100 μM (thirteen >250 μM) on human red blood cells; thirteen had CC50 >100 μM (six >250 μM) on HCT116 cells. · De novo AMPs show favorable MIC to toxicity…
ShownSix AMPs (#3, #5, #13, #15, #16, #27) were broad-spectrum with a favorable therapeutic window, and no resistance emerged in E. coli over 21 days of serial passage, whereas it did for imipenem.
Among the sixteen AMPs tested, fifteen had MIC ≤25 μM against E. faecium (thirteen ≤6.3 μM); the six were also active against Y. pestis and B. anthracis. Imipenem MIC rose up to 8-fold and passed the EUCAST breakpoint, while AMP MICs did not rise significantly. Only AMP #16 raised OMV release, 4-fold. · De novo AMPs show broad-band activities; No resistance was developed against de novo AMPs
Methods, data and reproducibility
Two CNN-RNN hybrid VAEs, pretrained on UniProt peptides and fine-tuned on known AMPs, generated sequences that CNN/RNN MIC regressors ranked. Linear DNA templates were expressed in a PURE cell-free system and the products tested directly on E. coli and B. subtilis cultures.
- UniProt peptide sequences shorter than 49 aa for VAE pretraining (half used) · n = 3,104,952 unique sequences; ~1.5 million used
- GRAMPA-derived AMP sequences with MIC data (filtered) for VAE transfer learning and regressor training · n = 5,319 unique AMP sequences
- UniProtKB non-AMP sequences used as negative examples for the regressors · n = 10,612 unique sequences
- De novo AMP candidates screened by cell-free expression on growth inhibition · E. coli MG1655, B. subtilis PY79 · n = 500 candidates; 30 functional
- Chemically synthesized AMPs tested for MIC, hemolysis (donor red blood cells) and cytotoxicity (HCT116) · Homo sapiens (cells); bacteria · n = 22 of 30 synthesized; 16 advanced
- MIC panel on ESKAPE and other pathogens, plus E. coli serial passaging · multiple bacterial species · n = duplicates (n = 2) for ESKAPE; triplicates (n = 3) for others
- MD simulations of AMPs on bacterial inner-membrane and human plasma-membrane models · n = 30 AMPs, 1 μs each (AMP #2 also without disulfide bond)
- Methods & tools
- Variational autoencoder (CNN-RNN hybrid, VAE_v1 and VAE_v2), CNN and LSTM MIC regressors, Keras with TensorFlow, AlphaFold, GROMACS 2020.3, CHARMM36m and CHARMM-GUI, BLAST, PUREfrex2.0 cell-free system, GraphPad Prism 9, RBS calculator and Kinfold
- Data & accessions
- https://github.com/amirpandi/Deep_cAMP
- Software & environment
- Keras 1.0 with TensorFlow 2.0 backend, Python 3.9, Google Colab pro; GROMACS 2020.3 with CHARMM36m force field.
- Assay conditions & compute
- Deep learning ran in the Google Colab pro environment; MD replicates were 1 μs each; GPU/CPU hardware and runtimes are not stated.
- Reproducibility gaps
- Raw growth curves and OMV, propidium iodide and MIC data are in a Source Data file or supplementary tables, with no repository accession given.
Limitations (3)
- Disulfide bonds for AMPs #2, #8 and #14 were assigned from 1-2 μs water-only simulations. The authors flag possible structural bias for AMP #2, so they also simulated it without an imposed bond. · stated by the authors
- The primary screen scored only growth suppression of E. coli and B. subtilis, so peptides active mainly against other species could be missed. Also, only 22 of the 30 hits were synthesized, which restricts the toxicity and MIC characterization to that subset. · noted in this summary
- Hemolysis and cytotoxicity rest on n = 2 experiments (one donor's blood, one cell line), ESKAPE MICs on duplicates, and resistance was tested only in E. coli. In vivo efficacy was not assessed. · noted in this summary
Who should read:Read this if you want a worked example of coupling generative peptide design with cell-free expression for medium-throughput antimicrobial screening and downstream validation.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Exploring the repository of de novo designed bifunctional antimicrobial peptides through deep learning
Ruihan Dong, Rongrong Liu, … Cheng Zhu·eLifePeer-reviewed·Mar 13, 2025
- ProT-Diff: A Modularized and Efficient Approach to De Novo Generation of Antimicrobial Peptide Sequences through Integration of Protein Language Model and Diffusion Model
Xue‐Fei Wang, Jing‐Ya Tang, … Dian‐Bing Wang·Advanced SciencePeer-reviewed·Sep 25, 2024
- Painting Peptides with Antimicrobial Potency through Deep Reinforcement Learning
Ruihan Dong, Qiushi Cao, Chen Song·Advanced SciencePeer-reviewed·Sep 12, 2025
- A deep reinforcement learning platform for antibiotic discovery
Hanqun Cao, Marcelo D. T. Torres, … Pheng‐Ann Heng·bioRxiv·Sep 23, 2025
- AIPAMPDS: an AI platform for antimicrobial peptide design and screening
Yanru Li, Liming Gao, … Cong Pian·bioRxiv·Mar 24, 2025
- Deep learning regression model for antimicrobial peptide design
Jacob J. Witten, Zack Witten·bioRxiv·Jul 12, 2019