AlphaFold is the AI system developed by Google DeepMind that solved the 50-year grand challenge of protein structure prediction — determining a protein's precise 3D atomic structure from its 1D amino acid sequence with experimental accuracy — triggering a revolution in structural biology, drug discovery, and our fundamental understanding of life's molecular machinery.

What Is AlphaFold?

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="-apple-system,Segoe UI,Roboto,sans-serif">
  <rect x="0" y="0" width="760" height="470" fill="#0d1117"/>
  <text x="380" y="28" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">AlphaFold — AI Protein Structure Prediction</text>
  <text x="380" y="48" fill="#8b98a5" font-size="12" text-anchor="middle">amino acid sequence → 3D atomic coordinates with experimental-level accuracy</text>

  <!-- Pipeline -->
  <rect x="30" y="65" width="700" height="165" rx="6" fill="#080d14" stroke="#233043" stroke-width="1"/>
  <text x="380" y="84" fill="#e6edf3" font-size="11" font-weight="600" text-anchor="middle">AlphaFold2 Architecture Pipeline</text>

  <!-- Input: sequence -->
  <rect x="50" y="100" width="100" height="110" rx="5" fill="#0b1220" stroke="#60a5fa" stroke-width="1"/>
  <text x="100" y="118" fill="#93c5fd" font-size="9" font-weight="600" text-anchor="middle">Input</text>
  <text x="100" y="136" fill="#8b98a5" font-size="7.5" text-anchor="middle">amino acid seq</text>
  <text x="100" y="150" fill="#6b7684" font-size="7" text-anchor="middle">MVLSPADKTN...</text>
  <text x="100" y="168" fill="#8b98a5" font-size="7.5" text-anchor="middle">MSA search</text>
  <text x="100" y="182" fill="#8b98a5" font-size="7.5" text-anchor="middle">template search</text>
  <text x="100" y="198" fill="#6b7684" font-size="7" text-anchor="middle">(genetic databases)</text>

  <path d="M152,155 L175,155" fill="none" stroke="#3a4453" stroke-width="1"/>
  <polygon points="173,152 179,155 173,158" fill="#3a4453"/>

  <!-- Evoformer -->
  <rect x="182" y="100" width="140" height="110" rx="5" fill="#0b1220" stroke="#a78bfa" stroke-width="1.2"/>
  <text x="252" y="118" fill="#c4b5fd" font-size="9" font-weight="600" text-anchor="middle">Evoformer (×48)</text>
  <text x="252" y="138" fill="#8b98a5" font-size="7.5" text-anchor="middle">MSA representation</text>
  <text x="252" y="153" fill="#8b98a5" font-size="7.5" text-anchor="middle">(row + col attention)</text>
  <text x="252" y="170" fill="#8b98a5" font-size="7.5" text-anchor="middle">pair representation</text>
  <text x="252" y="185" fill="#8b98a5" font-size="7.5" text-anchor="middle">(triangular updates)</text>
  <text x="252" y="202" fill="#6b7684" font-size="7" text-anchor="middle">co-evolution signals</text>

  <path d="M324,155 L347,155" fill="none" stroke="#3a4453" stroke-width="1"/>
  <polygon points="345,152 351,155 345,158" fill="#3a4453"/>

  <!-- Structure module -->
  <rect x="354" y="100" width="140" height="110" rx="5" fill="#0b1220" stroke="#f59e0b" stroke-width="1.2"/>
  <text x="424" y="118" fill="#fbbf24" font-size="9" font-weight="600" text-anchor="middle">Structure Module</text>
  <text x="424" y="138" fill="#8b98a5" font-size="7.5" text-anchor="middle">IPA (Invariant Point</text>
  <text x="424" y="153" fill="#8b98a5" font-size="7.5" text-anchor="middle">Attention) × 8 layers</text>
  <text x="424" y="170" fill="#8b98a5" font-size="7.5" text-anchor="middle">SE(3)-equivariant</text>
  <text x="424" y="185" fill="#8b98a5" font-size="7.5" text-anchor="middle">backbone frames</text>
  <text x="424" y="202" fill="#6b7684" font-size="7" text-anchor="middle">iterative refinement</text>

  <path d="M496,155 L519,155" fill="none" stroke="#3a4453" stroke-width="1"/>
  <polygon points="517,152 523,155 517,158" fill="#3a4453"/>

  <!-- Output: 3D structure -->
  <rect x="526" y="100" width="180" height="110" rx="5" fill="#0b1220" stroke="#34d399" stroke-width="1.2"/>
  <text x="616" y="118" fill="#6ee7b7" font-size="9" font-weight="600" text-anchor="middle">3D Structure Output</text>
  <!-- Simplified protein backbone path -->
  <path d="M555,140 C570,130 585,150 600,135 C615,120 630,145 645,138 C660,132 675,155 685,145" fill="none" stroke="#34d399" stroke-width="2" stroke-linecap="round"/>
  <path d="M560,158 C575,168 590,148 610,162 C625,172 640,155 660,165 C672,172 680,160 690,170" fill="none" stroke="#60a5fa" stroke-width="1.5" stroke-linecap="round"/>
  <text x="616" y="190" fill="#8b98a5" font-size="7.5" text-anchor="middle">atom coordinates (x,y,z)</text>
  <text x="616" y="204" fill="#34d399" font-size="7.5" text-anchor="middle">pLDDT confidence per residue</text>

  <!-- Accuracy + Impact -->
  <rect x="30" y="242" width="345" height="135" rx="6" fill="#080d14" stroke="#233043" stroke-width="1"/>
  <text x="202" y="262" fill="#e6edf3" font-size="10" font-weight="600" text-anchor="middle">Accuracy &amp; Impact</text>

  <text x="50" y="286" fill="#34d399" font-size="9" font-weight="600">CASP14 (2020): GDT-TS = 92.4</text>
  <text x="50" y="302" fill="#8b98a5" font-size="8">solved the protein folding problem (50-year grand challenge)</text>

  <text x="50" y="322" fill="#60a5fa" font-size="9" font-weight="600">AlphaFold DB: 200M+ structures</text>
  <text x="50" y="338" fill="#8b98a5" font-size="8">every known protein in UniProt — free and open</text>

  <text x="50" y="358" fill="#a78bfa" font-size="9" font-weight="600">AlphaFold 3 (2024):</text>
  <text x="50" y="374" fill="#8b98a5" font-size="8">proteins + DNA + RNA + ligands + ions (diffusion-based)</text>

  <!-- Key innovations -->
  <rect x="390" y="242" width="340" height="135" rx="6" fill="#080d14" stroke="#233043" stroke-width="1"/>
  <text x="560" y="262" fill="#e6edf3" font-size="10" font-weight="600" text-anchor="middle">Key Technical Innovations</text>

  <text x="410" y="286" fill="#f59e0b" font-size="8.5" font-weight="600">Triangular attention:</text>
  <text x="410" y="302" fill="#8b98a5" font-size="8">if A close to B, and B close to C, then A~C</text>

  <text x="410" y="322" fill="#a78bfa" font-size="8.5" font-weight="600">IPA (Invariant Point Attention):</text>
  <text x="410" y="338" fill="#8b98a5" font-size="8">attention in 3D space, rotation/translation invariant</text>

  <text x="410" y="358" fill="#34d399" font-size="8.5" font-weight="600">Recycling (×3):</text>
  <text x="410" y="374" fill="#8b98a5" font-size="8">re-run Evoformer on own output for iterative refinement</text>

  <!-- Footer -->
  <rect x="30" y="390" width="700" height="40" rx="6" fill="#080d14" stroke="#233043" stroke-width="1"/>
  <text x="380" y="410" fill="#8b98a5" font-size="8.5" text-anchor="middle">Runtime: ~minutes per protein on a single GPU. Training: 128 TPUv3 for 11 days. Nobel Prize in Chemistry (2024).</text>
  <text x="380" y="424" fill="#6b7684" font-size="7.5" text-anchor="middle">Applications: drug discovery, enzyme engineering, antibody design, crop improvement, disease mechanism understanding</text>

  <text x="380" y="452" fill="#6b7684" font-size="11" text-anchor="middle">AlphaFold proved that AI can solve fundamental scientific problems — not just pattern matching, but physical reasoning.</text>
</svg>

• Definition: A deep learning system that predicts the three-dimensional folded structure of a protein from its amino acid sequence — solving what was previously an extraordinarily expensive, months-long experimental process.
• AlphaFold 2 (2020): Published at CASP14 competition, achieving median backbone accuracy of 0.96 Å RMSD — within the margin of experimental error for most proteins.
• AlphaFold DB: Google DeepMind and EMBL-EBI released predicted structures for 200M+ proteins covering the entire known proteome of life on Earth.
• Impact: Nature called AlphaFold 2 the "most important scientific achievement in decades." The 2024 Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper for AlphaFold.

Why AlphaFold Matters

• Eliminates Bottleneck: Before AlphaFold, determining a single protein structure via X-ray crystallography or Cryo-EM cost $100K–$1M and took months to years. AlphaFold predicts structures in minutes at near-zero marginal cost.
• Drug Target Identification: Understanding protein 3D structure reveals binding pockets — sites where drug molecules can bind and modulate protein function. AlphaFold opens thousands of previously "undruggable" targets.
• Enzyme Engineering: Design novel enzymes for industrial biotechnology, carbon capture, and sustainable chemistry by understanding and modifying active site geometry.
• Disease Understanding: Structural predictions reveal how genetic mutations disrupt protein folding, explaining disease mechanisms for Alzheimer's, Parkinson's, and rare genetic disorders.
• Antibiotic Resistance: Map bacterial protein structures to identify novel targets for next-generation antibiotics resistant to existing resistance mechanisms.

The Protein Folding Problem

Proteins are chains of amino acids (typically 100–1,000 residues) that spontaneously fold into precise 3D structures determined by their sequence. The folded structure determines function:

• Enzymes: Active site geometry determines what reactions they catalyze.
• Receptors: Binding pocket shape determines what molecules activate them.
• Structural proteins: Shape determines mechanical properties.

Anfinsen's dogma (1972): The 3D structure is fully determined by the amino acid sequence. Yet computing this fold was intractable — Levinthal's paradox showed even 100-residue proteins have more conformational states than atoms in the universe.

AlphaFold 2 Architecture

Evoformer:

• A novel attention architecture that jointly processes two representations:

1. Multiple Sequence Alignment (MSA) representation: evolutionary co-variation signals from homologous sequences across species. 2. Pair representation: predicted spatial relationships between every pair of residues.

• Attention flows bidirectionally between MSA and pair representations — capturing the relationship between evolutionary conservation and geometric constraints.
• 48 Evoformer blocks with ~86M parameters total.

Structure Module:

• Takes Evoformer output and iteratively refines 3D atomic coordinates using SE(3)-equivariant networks (invariant point attention).
• Outputs backbone and sidechain atom coordinates with confidence per-residue (pLDDT score).

Training Data:

• PDB (Protein Data Bank): 170,000+ experimentally determined structures.
• UniRef90: 270M protein sequences for MSA generation.
• Self-distillation on predicted structures of 350,000 unannotated sequences.

Confidence Scoring

• pLDDT (predicted Local Distance Difference Test): Per-residue confidence score 0–100. >90 = very high confidence; 70–90 = confident; <50 = disordered/flexible regions.
• PAE (Predicted Aligned Error): Confidence in relative position between residue pairs — identifies domain boundaries and multimer interfaces.

AlphaFold 3 (2024)

• Extended to predict structures of protein-DNA, protein-RNA, protein-small molecule, and protein-ion complexes.
• Uses a diffusion-based structure generation module replacing the invariant point attention module.
• Critical for drug design: predicts how drugs bind within protein pockets at atomic precision.
• AlphaFold Server: Free access for non-commercial research.

Ecosystem & Follow-On Models

AlphaFold is the proof-of-concept that AI can solve fundamental scientific challenges thought to require decades of experimental work — its success is catalyzing AI applications across genomics, protein engineering, and drug discovery, demonstrating that biology's deepest secrets are now accessible through data and computation.