Short amyloids ≠ antimicrobial peptides? 🧬❌🦠

Experimental validation reveals that predicted antimicrobial amyloids lack real activity, highlighting a major gap in AMP datasets.

Keywords

amyloids, antimicrobial peptides, AMP, machine learning, AmpGram, negative dataset, bioinformatics

---

📌 Project highlights

• 🧬 Tests short amyloids predicted as AMPs
  
• 🤖 Uses AmpGram + 15 ML models
  
• 🧪 Validates experimentally on 10 bacterial strains
  
• ❌ Finds no antimicrobial activity
  
• 📊 Proposes high-quality negative dataset for ML

---

🎉 New paper!

👉 testing if amyloids can act like antimicrobial peptides

👉 Testing Antimicrobial Properties of Selected Short Amyloids

---

🎧 Audio summary

Machine learning said “these might be AMPs”…
experiments said “nope” 😅

👉 Here’s a quick breakdown 🎧:

---

🔬 What is this about?

Amyloids and antimicrobial peptides (AMPs) look surprisingly similar:

• both can disrupt membranes
  
• both form aggregates (β-sheet structures)
  
• both interact with the immune system

👉 So the question: Can short amyloids act as antimicrobial peptides?

---

⚙️ The core problem

AMP discovery relies heavily on machine learning

BUT:

👉 models lack true negative examples

• very few experimentally confirmed non-AMPs
  
• datasets often contain false negatives

👉 leads to over-optimistic predictions

---

🧠 What they did

🧩 Step 1: ML-based selection

• screened 509 amyloids (WALTZ-DB)
  
• used AmpGram (n-grams + random forest)
  
• selected top 10 candidates

👉 only ~6% predicted as AMPs

---

⚖️ Step 2: model comparison

• tested 15 additional AMP predictors

👉 result:

• highly inconsistent predictions
  
• only 1 peptide agreed across all models

---

🧪 Step 3: experimental validation

They tested:

• 🦠 10 bacterial strains (Gram+ & Gram−)
  
• 🧫 peptide concentrations up to 128 µg/mL
  
• 🧬 human cell toxicity

---

🔍 Key results

❌ No antimicrobial activity

• all peptides remained green (bacteria survived)
  
• MIC > 128 µg/mL → no effective killing

👉 even top ML predictions failed

---

🧬 No cytotoxicity either

• no harmful effects on human cells
  
• IC50 values relatively high

👉 peptides are biologically inactive (safe but useless as AMPs)

---

🧩 Some amyloids still aggregate

• 4/10 formed fibrils
  
• aggregation ≠ antimicrobial activity

👉 structural similarity ≠ functional similarity

---

⚠️ ML models struggle here

• short amyloids = edge cases
  
• predictions vary wildly across tools

👉 shows limits of current AMP predictors

---

💡 Key insight

👉 This is a negative result and that’s the point

These peptides:

• look like AMPs
  
• are predicted like AMPs
  
• behave like AMPs structurally

BUT:

👉 they are NOT AMPs

---

🚀 Why this matters

🧠 Better ML models

These sequences are:

👉 perfect “hard negatives”

• highly similar to AMPs
  
• experimentally validated as non-AMPs

👉 ideal for training robust models

---

⚠️ Dataset problem

Only ~24 confirmed non-AMPs exist (!!!)

👉 massive bottleneck in AMP prediction

---

📣 Cultural shift in science

The paper strongly argues:

👉 publish negative results

• prevents bias
  
• improves ML datasets
  
• accelerates discovery

---

💚 BioGenies perspective

This paper is 🔥 for one reason:

👉 it exposes a hidden failure mode in bio-ML

• models ≠ biology
  
• predictions ≠ function
  
• similarity ≠ activity

---

👉 And more importantly:

negative data = high-value data

This is exactly what:

• 🧠 interpretable ML
  
• 🧬 robust datasets
  
• 🔬 real-world validation

should look like