🧬 Big Picture: From Sequence → Structure

Key idea

• We have far more protein sequences than known structures
• Experimental structures (NMR spectroscopy, X-ray crystallography) are slow and expensive
• Therefore, we try to predict structure from sequence

❓ Can we get structure from sequence alone?

👉 In theory: yes 👉 In practice: very difficult

Why?

Because protein folding follows:

• Proteins fold toward a minimum free energy state
• The correct structure is the global energy minimum

But:

• The number of possible conformations is enormous
• This is known as Levinthal’s paradox

➡️ A protein cannot brute-force all conformations ➡️ Nature uses guided folding pathways ➡️ Computers struggle with this

⚡ Does a folded protein always stay folded?

👉 Your statement: “folded protein has minimum energy so it cannot unfold”

Correction:

• Folded protein = lowest energy under given conditions
• BUT:
  • Can unfold with:
    • heat
    • pH change
    • denaturants
• So it is stable, not permanent

⚠️ Misfolded proteins (important correction)

👉 Your statement: “prime protein?”

❌ Incorrect term ✅ Correct term: prion

What happens:

• Misfolded protein gets stuck in a local energy minimum
• Cannot escape → energy trap
• Example: prions → disease-causing

🧠 Why folding is computationally hard

• Backbone flexibility = only φ (phi) and ψ (psi) angles
• But combinations grow exponentially

➡️ Even a 100 amino acid protein:

• Would take longer than the age of the universe to brute-force fold

🔬 MAIN METHODS FOR STRUCTURE PREDICTION

1. Homology Modeling (Comparative Modeling)

Your understanding:

find similar sequence → align → model

✅ Correct, but refine it:

Steps:

1. Find homologous protein with known structure
2. Align sequences
3. Replace residues (point mutations)
4. Energy minimize → final structure

⚠️ Important limitation

👉 Your example about protease vs structural protein is correct

• Same sequence similarity ≠ same structure/function
• Example:
  • protease → globular enzyme
  • structural protein → fibrous

➡️ Function + environment matters

🔍 Identity vs Homology

Your definitions:

✔️ Mostly correct, refine:

Identity

• Same amino acid at same position
• e.g. Leu = Leu

Homology (better term: similarity)

• Different amino acids, same properties
• e.g.
  • Tyr, Trp, Phe → aromatic
  • Ile, Leu → hydrophobic

❓ “Is identity always homology?”

✔️ Yes

• Identity ⊂ Homology

⚠️ BLAST limitation

BLAST

👉 Your statement: “BLAST is local alignment”

✔️ Correct and important

• BLAST finds local matches
• Example:
  • 90% identity over 12 residues ≠ meaningful

➡️ Always check full-length alignment

🧬 Multiple Sequence Alignment (MSA)

👉 Your understanding: ✔️ Correct

What it does:

• Align many sequences
• Identify:
  • conserved residues
  • evolutionary patterns

➡️ Helps infer:

• structure
• function
• domains

🧩 Domain concept (important insight)

• Different parts of a protein can:
  • match different proteins
• Meaning: → protein may be multi-domain

🔧 Comparative modeling = point mutations

👉 Your statement: ✔️ Correct

• You take known structure
• Mutate residues to match your sequence

🧠 Secondary Structure Prediction

Why multiple programs?

✔️ You were right to question this

Key point:

• All programs agree on:
  • core of helices/sheets
• Disagree on:
  • boundaries (~15%)

❓ “15% hard to predict?”

✔️ Correct

• Especially:
  • helix start/end
  • loop transitions

👉 Proline:

• often disrupts helices
• introduces bends

🧬 Protein types (your statement)

👉 “Globular: hydrophobic inside, membrane: outside”

✔️ Correct but refine:

Globular proteins

• hydrophobic core
• hydrophilic surface

Membrane proteins

• hydrophobic regions face membrane lipids
• polar parts face inside/outside cell

🧵 Threading (Fold Recognition)

👉 Your idea: ✔️ Correct but incomplete

What it does:

• Uses structure templates even with low sequence similarity
• Aligns sequence onto known folds

➡️ Useful when:

• identity < 20%

Key concept:

• Structure is more conserved than sequence

🏛️ Fold Library

👉 Your statement: ✔️ Correct

• Database of known protein folds
• Threading tries to match your sequence to these

⚡ Contact Potentials (very important concept)

👉 Your understanding: ✔️ Good intuition, refine:

What it is:

• A scoring/energy function
• Based on: → which amino acids prefer to be near each other

Examples:

• hydrophobic + hydrophobic → favorable
• charged + hydrophobic → unfavorable

Why needed?

• Computers need numbers, not concepts
• So interactions are converted into scores

👉 Your question:

aromatic and fatty acids like each other?

✔️ Sometimes yes

• aromatic + hydrophobic → often favorable
• depends on orientation

🤖 AlphaFold (modern approach)

AlphaFold

Your understanding:

✔️ Mostly correct

It combines:

• MSA (evolutionary info)
• threading-like logic
• deep learning

MSA in AlphaFold

✔️ Correct

• captures evolutionary constraints
• residues that co-evolve → likely close in structure

Evoformer

👉 Your idea: ✔️ Partially correct

What it does:

• processes relationships between residues
• integrates:
  • sequence info
  • pairwise interactions

➡️ Not just validation → core reasoning engine

Iterative refinement

• model is improved repeatedly
• energy-like optimization

⚠️ Weaknesses of AlphaFold

👉 Your statement: ✔️ Correct observation

Common issues:

1. Disordered regions
   • appear as long tails
2. Misplaced helices
   • helix sticking out when it should be inside
3. Flexible regions
   • poorly predicted

Key takeaway:

➡️ Always critically evaluate predictions

🔐 Important practical warning

• Public servers (e.g. ColabFold)
• Your sequence = uploaded data

⚠️ Can affect:

• patents
• unpublished work

🔁 Final Conceptual Summary

Hierarchy of methods:

1. Ab initio
   • physics-based
   • very expensive
2. Homology modeling
   • needs similar sequence
3. Threading
   • uses fold library
4. AlphaFold
   • combines all + AI

✅ Quick corrections to your list