Individual Final Project

AI-Designed Antimicrobial Peptide Cocktails Tested via Cell-Free Protein Synthesis

Sean Murphy · MAS.885 Spring 2026

Abstract

Antimicrobial resistance is on track to cause more than 10 million deaths per year by 2050, and the discovery pipeline for new small-molecule antibiotics has slowed sharply over the past two decades. Antimicrobial peptides (AMPs) are an attractive alternative class because they kill bacteria primarily by membrane disruption, a mechanism that is intrinsically harder for pathogens to evolve resistance against.

The broad objective of this project is to ask whether modern AI peptide-design tools, paired with cell-free protein synthesis (CFPS), can be combined into a fast prototyping platform that not only identifies single-peptide hits, but also surfaces synergistic peptide cocktails. The hypothesis is that AMPs produced by mechanistically distinct generative models (a length-conditioned latent-space diffusion model trained on natural AMPs and a CLIP-based target-binder generator) will exhibit synergy when co-expressed, with the most pronounced synergy appearing in cross-method pairs that combine membrane-active generalists with target-directed binders.

To test this, the specific aims are:

1. Generate candidates with AMP-Diffusion, PepPrCLIP, and a MadSBM fallback strategy
2. Order 15 codon-optimized linear gene fragments from Twist
3. Express each peptide in BL21 Star DE3 lysate with GamS nuclease inhibitor to protect linear templates
4. Score crude lysates against E. coli ATCC 25922 and B. subtilis ATCC 6633 in a two-organism antimicrobial assay, then run pairwise co-expression on top hits and score synergy by an FICI-style metric

Project Aims

Aim 1 · Experimental Aim (this project)

Determine whether AI-designed antimicrobial peptides exhibit synergistic activity when co-expressed. The platform integrates three generative peptide-design tools, 15 codon-optimized Twist linear gene fragments, a Ginkgo Bioworks BL21 Star DE3 CFPS lysate supplemented with NEBExpress GamS Nuclease Inhibitor, and a dual-organism optical-density growth-inhibition assay scored against scrambled-peptide and Cecropin B controls. A primary single-peptide screen identifies hits, and a follow-on pairwise co-expression screen at half-dose scores synergy using an FICI-style fractional inhibition index calculated separately for each organism.

Aim 2 · Development Aim

Close the design, build, test loop by feeding measured single-peptide and pairwise activity data back into the generative models as a fine-tuning signal, producing a second wave of candidates conditioned on both predicted activity and predicted synergy partners. Practically this means:

Aim 3 · Visionary Aim

A personalized, on-demand antibiotic-cocktail design platform: a clinician submits a patient’s pathogen genome (or a culture resistance profile), a generative pipeline designs a custom cocktail targeting that strain’s essential proteins and membrane chemistry, the cocktail is expressed locally in a freeze-dried CFPS reaction, and the peptide mixture is administered as a same-day therapeutic. The cocktail approach addresses resistance directly: pathogens that mutate around one membrane-active peptide are far less likely to escape three simultaneously, each with a distinct mechanism. If achieved, this could reduce time-to-treatment for drug-resistant infections from months to days.

Background

Key References

Novelty

Why It Matters

The WHO estimates that drug-resistant infections cause ~1.27 million deaths each year today, projected to climb past 10 million per year by 2050 unless the discovery pipeline accelerates. The pharmaceutical industry has largely exited small-molecule antibiotic development because the economic returns are poor, so the gap is increasingly being filled by academic and government-funded efforts to find chemically distinct alternatives. AMPs are a particularly promising chemical class because their primary action (disrupting bacterial membranes) is harder for pathogens to evolve away from than enzyme-target binding, and peptide synthesis can be reprogrammed faster than small-molecule chemistry. If the methods developed here generalize, the broader contribution is a faster, cheaper, more modular antibiotic discovery workflow that can be rerun for each emerging resistant strain.

Ethical Considerations

This project sits at the intersection of three ethical concerns:

Guiding principles: non-maleficence (do no harm, including by not enabling foreseeable misuse), responsibility (using the synthesis-screening infrastructure consciously and transparently), beneficence (adding to humanity’s antibiotic arsenal), and justice (cocktail-based on-demand antibiotic design must be paired with deployment strategies that reach low-income contexts).

Experimental Design

Peptide Panel

Construct Architecture

Each construct is a single linear cassette:

[5′ pad] → [T7 promoter] → [strong E. coli RBS] → [ATG] → [codon-optimized ORF] → [TAA TAA] → [T7 terminator] → [3′ pad]

All 15 constructs are ≥ 300 bp (Twist gene-fragment minimum), adapters off, codon-optimized for E. coli K-12, and screened against BsaI, BbsI, EcoRI, HindIII sites and ≥6 nt homopolymer runs.

Timeline

Techniques Used

Two Techniques in Depth

Cell-Free Reactions

The entire bench phase is built around BL21 Star DE3 cell-free protein synthesis. Each Twist linear gene fragment is resuspended to 10 ng/µL and added at 50–100 ng to a 20 µL reaction containing lysate, reaction buffer, and NEBExpress GamS Nuclease Inhibitor (0.6 µg per 20 µL reaction). GamS is essential: it sequesters the RecBCD exonuclease in the crude lysate, which would otherwise degrade the linear DNA template within minutes. The no-GamS validation run (below) confirms that without it, peptides are not produced at detectable levels.

Reactions incubate at 30 °C for 4 h, then 5 µL of crude lysate is transferred directly into a 100 µL bacterial test well. No peptide purification step is needed because the assay tolerates lysate background, controlled for by the scrambled-Cecropin negative control. Using CFPS rather than plasmid-based in vivo expression compresses the design-build-test loop from weeks to days and decouples the peptide chemistry from any host-toxicity confounder, since the host that would normally express the AMP is not part of the assay.

Setting up CFPS reactions at the bench.

Transferring template and reagents into reaction strips.

Incubator with rack ready for the 30 °C × 4 h CFPS step.

Spiking crude lysate into the bacterial test plate.

Protein Design (Generative Peptide Modeling)

Three AI tools were used in this project:

All peptide ORFs were reverse-translated using a high-expression E. coli K-12 codon table, screened against common restriction sites and homopolymer runs, then padded out to 300 bp for Twist.

Industry Council Partners

Results

What Has Been Validated

Two aspects of the project are validated so far:

1. Computational design pipeline. All three AI tools were stood up successfully and produced sequences that passed physicochemical and developability filters; the resulting 15 codon-optimized Twist constructs are documented in constructs/construct_details.csv and assembled in twist_orders/twist_order_batch1_FINAL.csv.
2. Bench workflow’s interpretive logic. Group A AMP peptides were expressed in BL21 Star DE3 lysate without GamS and assayed against E. coli ATCC 25922 as a negative-by-design control. The expectation was that the linear DNA would be chewed up by RecBCD before useful amounts of peptide could accumulate, so no antimicrobial activity should be visible above scrambled control. That expectation was met, which validates the necessity of GamS in subsequent runs and demonstrates that the assay correctly returns a null result when no functional peptide is produced.

No-GamS Validation Run, Quantitative Data

Wells A1 through F1 correspond to AMP-D-01 through AMP-D-06 spiked into E. coli ATCC 25922 in Mueller-Hinton broth at ~5 × 10⁵ CFU/mL starting density. OD600 read every 30 min for 7.5 h on the plate reader.

Figure 1. OD600 growth curves over 7.5 h for the no-GamS Group A validation run. Wells A1 through F1 correspond to AMP-D-01 through AMP-D-06 in CFPS reactions added to E. coli ATCC 25922 in Mueller-Hinton broth. All six wells show normal sigmoidal growth with no inhibition relative to one another, confirming that without GamS protection the linear DNA template is degraded by RecBCD before functional peptide accumulates.

The growth curves overlap heavily through lag and early-exponential phases and only diverge modestly in stationary phase, the normal pattern for untreated wells of the same strain. The key interpretive point is that no peptide well showed any reduction in E. coli growth relative to the others. The well that looks like an outlier (E1) is higher, not lower, which is the opposite of an antimicrobial signal. This is the expected null result. Without GamS, the linear template is degraded by RecBCD before functional peptide accumulates, so no antimicrobial activity is detectable, even though the bacteria are clearly viable and growing normally. The same plate reader, lysate prep, Twist fragments, and E. coli culture will be used for the +GamS run, so any growth suppression observed there can be attributed to peptide expression rather than to assay artifact.

Quantitative Expectations for the +GamS Phase

Published CFPS-AMP studies typically report 30–70% growth inhibition for active AMPs versus scrambled controls.

Challenges and Limitations

Supply List and Budget

References

Group Final Project