Designing Antimicrobial Peptides for E. coli

ISEF Category: Microbiology

Subcategory: Antimicrobials and Antibiotics  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Bacteria evolve fast, and some antibiotics are losing the race. Your project takes a different route, by designing tiny peptides that can damage bacterial cells. Think of each peptide like a custom lockpick for a microbial membrane. If your model works, you can go from database data to a real wet-lab test.

What Is It?

Antimicrobial peptides, or AMPs, are short chains of amino acids that can stop microbes from growing. Many AMPs attack bacterial membranes, the thin outer barrier that cells need to survive. If you picture a cell like a water balloon, some AMPs act like tiny needles that make the balloon leak.

Your project starts with a database of known AMPs, then uses protein language model embeddings from ESM2. Embeddings are numeric summaries of a sequence, kind of like a fingerprint that captures patterns in the peptide. You train a small classifier on those fingerprints, score new candidate sequences, and pick a few to test against E. coli K-12. Radial-diffusion assays in low-salt agar give you a visible readout of how well each peptide slows bacterial growth.

Why This Is a Good Topic

This makes a strong science fair topic because you can test a clear idea, compare real candidates, and measure real biological activity. You also connect computational design with wet-lab validation, which is exactly the kind of bridge judges like to see. The topic relates to antibiotic resistance, so your work has a real-world hook. You can learn sequence analysis, model training, assay design, and basic statistics in one project.

Research Questions

• How does the classifier score of a peptide relate to its measured inhibition zone against E. coli K-12?
• What is the effect of peptide net charge on radial-diffusion activity in low-salt agar?
• Does hydrophobicity predict which designed peptides show the largest inhibition zones?
• To what extent do ESM2-derived embeddings separate active from inactive AMP-like sequences?
• Which sequence length range within 8 to 25 amino acids gives the best activity for designed candidates?
• How does a designed peptide compare with a known reference AMP under the same assay conditions?

Basic Materials

• Computer with internet access and enough memory to run sequence analysis tools.
• Public AMP dataset from APD3 or a similar open database.
• Python with common data science libraries installed.
• Spreadsheet software for tracking sequences, scores, and assay results.
• Micropipettes and sterile tips.
• Sterile Petri dishes with low-salt agar.
• E. coli K-12 culture from an approved school or university lab source.
• Basic microbiology safety supplies, including gloves, lab coat, and disinfectant.
• Measuring ruler or digital calipers for inhibition zones.
• Incubator approved by your lab supervisor.

Advanced Materials

• Access to peptide synthesis quotes or ordering through a mentor or university account.
• University or teaching lab access for peptide handling and bacterial culture.
• ESM2-compatible computing access or a workstation with GPU support.
• Sequence analysis software for feature extraction and model comparison.
• HPLC or mass spectrometry access to verify peptide identity, if available.
• Spectrophotometer or plate reader for complementary growth measurements.
• Sterile filtration supplies for peptide solutions, if required by the lab protocol.
• Reference antimicrobial peptides for positive controls.
• Negative control peptides with similar length but no expected activity.
• Statistical software for model testing and effect size analysis.

Software & Tools

• Python: Runs data cleaning, feature extraction, model training, and result plots.
• Biopython: Helps you handle peptide sequences and calculate simple sequence features.
• scikit-learn: Trains and compares small classifiers on AMP labels and embeddings.
• pandas: Organizes database records, candidate scores, and assay outcomes.
• matplotlib: Plots inhibition zones, score distributions, and model comparisons.

Experiment Steps

1. Define the exact prediction target, such as active versus inactive AMP-like sequences, and choose the biological outcome you will test in the lab.
2. Assemble a clean training set from APD3, remove duplicates, and decide which sequence features you will score besides the ESM2 embeddings.
3. Train and compare a few small classifiers, then rank new peptide candidates by model score, diversity, and practical synthesis cost.
4. Plan your controls, including a known active peptide, a negative control, and a matching assay condition that keeps the comparison fair.
5. Design the wet-lab validation so your readout, measurement method, and replicate plan match the prediction question.
6. Decide how you will analyze agreement between model score and assay outcome, including effect size, error bars, and a simple statistical test.

Common Pitfalls

• Training on APD3 labels without removing near-duplicate sequences, which makes the model look better than it really is.
• Picking candidates only by high score, which can give you very similar peptides instead of a useful comparison set.
• Ignoring peptide solubility or charge, which can make a sequence look inactive because it never reaches the bacteria well.
• Comparing inhibition zones across plates with different agar depth or salt conditions, which distorts the readout.
• Skipping proper negative controls, which makes it hard to tell whether a small halo came from the peptide or from experimental noise.

What Makes This Competitive

A stronger project does more than rank peptides and measure halos. You would compare several model types, test whether the model learns real biology, and show why your top candidates differ from the rest. You could also link sequence features to activity, not just report active or inactive results. Clear controls, honest validation, and a clean analysis plan matter as much as the peptide itself.

Project Variations

• Test peptides against a second bacterial strain to see whether your design rules transfer beyond E. coli K-12.
• Compare ESM2-based scoring with a simpler feature set, such as charge, hydrophobicity, and length, to see which predicts activity better.
• Add a toxicity screen on a safe nonbacterial proxy assay or approved host-cell model, if your lab has one, to study selectivity as well as antibacterial action.

Learn More

• APD3, Antimicrobial Peptide Database: Search the APD3 database for known AMP sequences, annotations, and activity labels.
• PubMed: Search review articles on antimicrobial peptide design, machine learning, and membrane-active peptides.
• NIH PubChem: Look up amino acid properties and basic compound information for peptide components.
• Antimicrobial Peptides: Methods and Protocols: Search library catalogs or chapter listings for assay design ideas and background reading.
• MIT OpenCourseWare: Search biology and computation courses for sequence analysis, basic statistics, and machine learning concepts.