2. Introduction to Structure Prediction

This module provides the foundational concepts behind protein structure prediction—the theoretical framework that underlies tools like AlphaFold2, ESMFold, and other modern prediction methods.

Live Workshop Session

🎥 Live workshop recording — Protein structure prediction: history, challenges, and AI breakthroughs
📊 View slide deck

What is Structure Prediction?

At its core, protein structure prediction attempts to answer a deceptively simple question:

Given the linear amino acid sequence of a polypeptide, what is its 3D structure?

This is known as the protein folding problem, and it has been one of the grand challenges of molecular biology for over 50 years.

Why is this hard?

A protein sequence like:

MGDIQVQVNIDDNGKNFDYTYTVTTESELQKVLNELMDYIKKQGAKRVRISITARTKKEAEKFAAILIKVFAELGYNDINVTFDGDTVTVEGQLE

Must fold into a precise three-dimensional arrangement where: - Every atom has specific coordinates - The structure is thermodynamically stable - The fold enables biological function

The Theoretical Foundation

Anfinsen’s Thermodynamic Hypothesis (1960s)

Christian Anfinsen’s Nobel Prize-winning work established a fundamental principle:

The native (functional) structure of a protein is at a free energy minimum.

This means: - The sequence contains all the information needed to specify the structure - Given the right conditions, a protein will spontaneously fold to its native state - Structure is determined by physics, not by cellular machinery

NoteAnfinsen’s Experiment

Anfinsen showed that ribonuclease A, when completely unfolded (denatured), would spontaneously refold to its active form when denaturing conditions were removed. This proved that the amino acid sequence alone determines the final structure.

Levinthal’s Paradox

Shortly after Anfinsen’s work, Cyrus Levinthal raised a puzzling question:

How can proteins fold so quickly if they have so many possible conformations?

Consider a modest 101-residue protein: - Each residue has ~3 possible backbone conformations (φ/ψ angles) - Total possible conformations: 3^200 ≈ 10^95 states - If the protein sampled one conformation per picosecond, it would take longer than the age of the universe to find the right one

Yet proteins fold in milliseconds to seconds. How?

The Folding Funnel

The resolution to Levinthal’s paradox came from understanding that the energy landscape is not flat—it’s shaped like a funnel:

     ╱ ╲ ╱ ╲      ← High energy, many conformations (unfolded)
    ╱   ╲   ╲
   ╱     ╲   ╲
  ╱       ╲   ╲
 ╱    ↓    ╲   ╲   ← Energy decreases as structure forms
╱___________╲___╲
      ●          ← Native state (energy minimum)

Key insights:

  1. Not a random search: The protein doesn’t sample all conformations—it follows an energetic gradient
  2. Local interactions guide folding: Secondary structure (helices, sheets) forms first, providing scaffolding
  3. Funnel shape: Many paths lead to the same native state
  4. Rugged landscape: In reality, there are local minima (kinetic traps) that can slow folding
ImportantReality is More Complex

In practice, the folding landscape isn’t one clean funnel—it’s many interconnected funnels. This creates: - Folding intermediates: Partially folded states - Kinetic traps: Local minima that delay folding - Multiple conformations: Some proteins sample several states - Intrinsically disordered regions: Parts that never fully fold

This complexity is what makes structure prediction so challenging!

A Brief History: CASP and the Road to AlphaFold

Critical Assessment of protein Structure Prediction (CASP)

CASP is a biennial competition that has tracked progress in structure prediction since 1994. Organizers release sequences of proteins whose structures have been solved but not yet published, and teams submit blind predictions.

Key milestones:

Year CASP Major Development
1998 CASP3 Fragment-based assembly (Rosetta)
2014 CASP11 Co-evolutionary analysis + deep learning
2018 CASP13 AlphaFold1: Deep learning-guided minimization
2020 CASP14 AlphaFold2: Near-experimental accuracy
2022 CASP15 AlphaFold-Multimer, ESMFold emerge

The Pre-AlphaFold Era

Fragment-based methods (Rosetta, ~1998): - Break known structures into fragments - Assemble fragments guided by energy functions - Sample many conformations, select lowest energy

Co-evolutionary methods (~2014): - Key insight: Residues that contact each other in 3D co-evolve in sequence - Multiple sequence alignments (MSAs) reveal which residues co-vary - These “contact maps” constrain structure prediction

Early deep learning (~2018): - Neural networks predict contact maps from MSAs - Use contacts to guide physics-based minimization - AlphaFold1 used this approach

The AlphaFold2 Revolution (2020)

AlphaFold2 achieved what many thought was decades away:

  • GDT scores of ~90 on difficult targets (previously ~60 was state-of-the-art)
  • Approaching experimental accuracy for many proteins
  • End-to-end learning: Directly predicts coordinates, not intermediate features

We’ll explore AlphaFold2’s architecture in detail in the next module.

Structure Prediction Metrics

When evaluating predictions, several metrics are used:

Root Mean Squared Deviation (RMSD)

What it measures: Average distance between corresponding atoms after optimal superposition.

\[RMSD = \sqrt{\frac{1}{N}\sum_{i=1}^{N}d_i^2}\]

where \(d_i\) is the distance between atom \(i\) in the prediction and reference.

Interpretation: - < 2 Å: Excellent (atomic-level accuracy) - 2-4 Å: Good (correct fold, minor deviations) - 4-8 Å: Moderate (correct topology, some errors) - > 8 Å: Poor (likely wrong fold)

WarningRMSD Limitations
  • Length-dependent: Longer proteins tend to have higher RMSD
  • Sensitive to outliers: One badly placed domain can dominate
  • Requires superposition: Choice of alignment region matters

Local Distance Difference Test (lDDT)

What it measures: Local structural similarity without superposition.

For each residue, lDDT asks: “Are the distances to neighboring residues preserved?”

Key properties: - Range: 0 to 1 (or 0-100 as percentage) - Length-independent: Can compare proteins of different sizes - Per-residue: Identifies which regions are well-predicted

This is the metric used for pLDDT (predicted lDDT) in AlphaFold2!

Global Distance Test (GDT)

What it measures: Fraction of residues within various distance thresholds.

\[GDT_{TS} = \frac{1}{4}(P_1 + P_2 + P_4 + P_8)\]

where \(P_x\) is the percentage of residues within \(x\) Å of the reference.

Interpretation: - > 90: Excellent - 70-90: Good - 50-70: Moderate - < 50: Poor

GDT-TS was the primary metric in CASP until AlphaFold2 essentially “solved” it.

Template Modeling Score (TM-score)

What it measures: Global structural similarity, normalized by protein length.

Key properties: - Range: 0 to 1 - Length-independent: Enables comparison across protein sizes - Threshold: TM-score > 0.5 generally indicates same fold

TM-score Interpretation
> 0.5 Same fold
0.3-0.5 Possibly related
< 0.3 Likely different folds

Applications of Structure Prediction

1. Molecular Replacement in Crystallography

Problem: X-ray crystallography requires initial “phases” to solve structures.

Solution: Predicted structures can provide starting models for: - Phase determination - Completing partial experimental structures - Filling in loops or disordered regions

2. Interpreting Experimental Results

Example scenarios: - A mutation causes loss of function—why? The structure shows it’s in the active site. - A protein doesn’t express well—the structure reveals aggregation-prone regions. - Two proteins interact—structures suggest the binding interface.

3. Functional Prediction and Hypothesis Testing

From structure, you can infer: - Active site location and chemistry - Binding pocket characteristics - Likely interaction partners (from shape complementarity) - Potential allosteric sites

4. Starting Point for Protein Engineering

Critical for: - Rational design: Knowing where to put mutations - Directed evolution: Understanding which regions to diversify - De novo design: Validating whether designed sequences will fold

TipStructure Prediction in Your Research

In this bootcamp, you’ll use structure prediction to:

  1. Predict structures of proteins of interest
  2. Validate designs from tools like RFdiffusion
  3. Assess confidence to know which regions to trust
  4. Compare predictions from different methods

The Modern Prediction Landscape

Today’s structure prediction tools fall into two main categories:

MSA-Based Methods (e.g., AlphaFold2, OpenFold)

Strengths: - Highest accuracy for proteins with many homologs - Leverage evolutionary information - Multiple model ensemble provides uncertainty estimates

Limitations: - Slower (MSA generation takes time) - Less accurate for orphan proteins (few homologs) - Computationally expensive

Language Model-Based Methods (e.g., ESMFold)

Strengths: - Very fast (no MSA needed) - Work on designed/synthetic proteins - Simpler pipeline

Limitations: - Generally lower accuracy than MSA-based methods - No ensemble diversity from single prediction - May miss evolutionary insights

Next-Generation Methods (e.g., AlphaFold3, Chai-1, Boltz-2)

The field is rapidly advancing with: - Multi-modal predictions: Proteins + ligands + nucleic acids - Diffusion-based approaches: New generative paradigms - Improved confidence estimation: Better uncertainty quantification

Key Takeaways

  1. Anfinsen’s hypothesis established that sequence determines structure
  2. Levinthal’s paradox highlighted the computational challenge
  3. The folding funnel explains how proteins fold efficiently
  4. CASP has tracked 25+ years of progress
  5. AlphaFold2 achieved a major breakthrough in 2020
  6. Multiple metrics (RMSD, lDDT, GDT, TM-score) capture different aspects of accuracy
  7. Structure enables function: Predictions have many practical applications

Looking Ahead

In the next modules, you’ll:

  1. Learn AlphaFold2’s architecture and how it achieves high accuracy
  2. Run your own predictions using ColabFold/LocalColabFold
  3. Compare methods (AlphaFold2 vs ESMFold) to understand trade-offs
  4. Visualize confidence and interpret prediction quality

Understanding these foundational concepts will help you use structure prediction tools effectively and interpret their outputs critically.

Hands-On Exercise

This module is primarily conceptual, but let’s reinforce these ideas with some exploration.

Part 1: Explore CASP Results

Goal: Understand how the field has progressed by looking at real CASP data.

  1. Visit the CASP website: predictioncenter.org

  2. Compare CASP rounds:

    • Look at results from CASP11 (2014) vs CASP14 (2020)
    • Find the GDT-TS scores for the top predictors
    • Notice how dramatically scores improved with AlphaFold2

  3. Questions to consider:

    • What was the typical GDT-TS for “hard” targets before AlphaFold2?
    • How did AlphaFold2’s scores compare to the rest of the field in CASP14?
    • Why do you think some targets are labeled “hard” vs “easy”?

Part 2: Metric Comparison Exercise

Goal: Develop intuition for different structure comparison metrics.

Using PyMOL, let’s compare two related structures:

# Fetch two related kinase structures
fetch 1ATP   # PKA with ATP bound
fetch 1J3H   # PKA in different conformation

# Align them
align 1J3H, 1ATP

# Note the RMSD printed in the console

Now think about:

  1. RMSD question: The RMSD might be 2-4 Å. Does this mean the structures are different, or is this expected variation?

  2. Per-residue analysis:

    # Color by RMSD per residue (after alignment)
# Red = high deviation, blue = low

    Which regions show the most difference? (Usually loops and termini)

  3. Why lDDT matters: If one domain moves relative to another (like in a kinase), RMSD after superposition will be high, but lDDT for each domain individually would still be good. This is why AlphaFold2 uses pLDDT as its confidence metric.

Part 3: Discussion Questions

Work through these questions with a partner or write brief notes:

  1. Anfinsen’s hypothesis:
    • What evidence supports the idea that sequence determines structure?
    • Can you think of exceptions? (Hint: chaperones, prions, intrinsically disordered proteins)
  2. Levinthal’s paradox:
    • If a protein has 100 residues and samples conformations at 10^12 per second, how long would it take to sample all 3^200 conformations?
    • Why doesn’t this happen in reality?
  3. Practical implications:
    • You predict a structure with AlphaFold2 and get average pLDDT of 75. What does this mean?
    • One region has pLDDT < 50. Should you trust it? What might it indicate biologically?
  4. Choosing metrics:
    • You’re comparing two predictions of the same protein. When would you use RMSD vs TM-score?
    • Why is GDT-TS preferred over raw RMSD in CASP?

Part 4: Prepare for AlphaFold2

To get ready for the next module:

  1. Review your HPC access: Make sure you can log into your cluster
  2. Check for GPU availability: Run nvidia-smi on a GPU node
  3. Prepare a test sequence: Find the FASTA sequence for a protein you’re interested in
TipSequence Resources
  • UniProt: uniprot.org - Search for proteins and download sequences
  • RCSB PDB: rcsb.org - Get sequences for proteins with known structures

Reflection

After this module, you should be able to: