---

AlphaFold: A Complete Guide — Basic to Advanced

---

PART 1 — FOUNDATIONS: The Problem AlphaFold Solves

1.1 What is a Protein?

1.2 The Protein Folding Problem

"The three-dimensional structure of a protein is determined by its amino acid sequence." — Christian Anfinsen, Nobel Prize 1972

• Drug design requires knowing the precise binding pocket shape
• Understanding disease mutations requires knowing structural context
• Engineering new proteins for biotechnology requires structure prediction

1.3 Traditional Experimental Methods (Before AlphaFold)

---

PART 2 — HISTORICAL CONTEXT & CASP

2.1 The CASP Competition

• Measures what fraction of Cα atoms are within 1, 2, 4, and 8 Å of true position
• Score of 100 = perfect prediction; Score >90 = essentially experimental accuracy

2.2 The AlphaFold Timeline

---

PART 3 — ALPHAFOLD1 (2018)

1. If you know which pairs of residues are close in space, you can reconstruct the 3D structure
2. Co-evolutionary analysis: if two residues evolve together across species (when one mutates, the other compensates), they must be in contact in the structure
3. Used deep ResNets + dilated convolutions to predict a distance map (N×N matrix of predicted distances between all pairs of residues)
4. Then used gradient descent to fold a structure consistent with the predicted distances

---

PART 4 — ALPHAFOLD2: THE BREAKTHROUGH (2021)

4.1 The Complete Pipeline

---

4.2 Stage 1: Input Feature Generation

• UniRef90 (~90% sequence identity clusters of UniProt)
• BFD (Big Fantastic Database) — 65 billion residues
• MGnify — metagenomics sequences

Query:   MKTAYIAKQRQISFV...
Seq 2:   MKTAYI-KQRQIAFV...   (A→A mutation at position 14)
Seq 3:   MKTGYIAKQRQISFV...   (A→G at position 4)
         ...thousands more sequences...

• Residue pair features: relative distances, bond angles, orientation between residues
• Generated from both MSA statistical co-variation AND optional template structures from the PDB

• If known homologous structures exist in the PDB, their backbone coordinates are used as templates
• AlphaFold2 tends to ignore templates when MSA depth is sufficient

---

4.3 Stage 2: The Evoformer — The Heart of AlphaFold2

• MSA Representation (Nseq × Nres × 256): tracks evolutionary relationships across all sequences
• Pair Representation (Nres × Nres × 128): tracks geometric/spatial relationships between all residue pairs

MSA Track Updates:

1. Row-wise Gated Self-Attention — Each sequence in the MSA attends to all positions along that sequence. The pair representation is projected as a bias added to the attention logits → pair geometry influences MSA attention
2. Column-wise Gated Self-Attention — At each residue position, attention runs across all sequences in the MSA → correlates how different homologous sequences handle each position
3. MSA Transition — Standard feedforward MLP for non-linear mixing

Pair Track Updates:

1. Outer Product Mean — Takes the outer product of each pair of MSA column embeddings, averaged over sequences → this is how evolutionary co-variation in the MSA updates the pair representation (MSA → Pair information flow)
2. Triangle Multiplicative Update (outgoing) — For each edge (i,j) in the pair representation, update it using the path i→k→j for all intermediate k. Represents the "triangle inequality" of 3D geometry: if i-k and k-j distances are known, i-j distance is constrained. Formula: m_ij = sigmoid(gate_ij) × Linear(m_ik ⊕ m_kj)
3. Triangle Multiplicative Update (incoming) — Same concept but using paths k→i→j
4. Triangle Self-Attention (around starting node) — Attention where edge (i,j) attends to all (i,k) edges
5. Triangle Self-Attention (around ending node) — Edge (i,j) attends to all (k,j) edges
6. Pair Transition — Feedforward MLP

---

4.4 Stage 3: Structure Module

• A backbone frame (3D rotation + translation), defined by the N, Cα, C backbone atoms
• Side chain torsion angles (χ₁, χ₂, χ₃, χ₄) for side chain placement

• The single representation (first row of the MSA, projected) encoding sequence identity
• The pair representation (biases the attention)
• 3D points generated relative to each residue's current backbone frame

---

4.5 Training Objectives

---

PART 5 — CONFIDENCE METRICS

5.1 pLDDT — Per-Residue Local Confidence

5.2 PAE — Predicted Aligned Error

• Low PAE (dark green on the matrix) = model is confident in the relative position of these two residues
• High PAE (light/white) = model is uncertain about the relative positioning

• PAE blocks along the diagonal = confident domains
• High off-diagonal PAE between blocks = flexible linker connecting two independently-folded domains
• Confirms if a multi-domain prediction reflects a real rigid complex or merely two independently folding chains placed arbitrarily near each other

5.3 TM-score and GDT

• TM-score (Template Modeling score): ranges 0–1; >0.5 = same fold; >0.9 = essentially identical
• GDT_TS: fraction of Cα within 1/2/4/8 Å thresholds, averaged; AlphaFold2 achieved 92.4 at CASP14
• lDDT (Local Distance Difference Test): compares local inter-residue distances; used to assess model quality at the residue level

---

PART 6 — ALPHAFOLD-MULTIMER (2021)

• MSA processing modified: paired MSAs are constructed — sequences from the same organism are paired across the two chains being modeled
• New "entity" token type to distinguish chains
• Modified training sampling to focus on inter-chain contact learning
• Asymmetric unit (multiple chains) handled as a single prediction
• Uses ipTM (interface TM-score) as the key confidence metric for the predicted complex interface quality

---

PART 7 — ALPHAFOLD3 (2024)

7.1 What's New in AF3?

• Single proteins and multi-chain complexes
• Protein–DNA interactions (transcription factor binding, etc.)
• Protein–RNA interactions
• Protein–small molecule (ligand) docking — dramatically outperforming dedicated docking tools
• Protein–ion complexes
• Modified residues (phosphorylation, glycosylation, etc.)

7.2 Architecture Changes

1. Starts with a cloud of randomly noised atom positions
2. Iteratively denoises the cloud, converging on the most likely structure
3. Operates directly on raw XYZ coordinates of all atoms — no rotational frames required
4. Multiple samples can be drawn from the same input (stochastic sampling), allowing AF3 to explore conformational diversity

7.3 AF3 Performance Benchmarks (Abramson et al., Nature 2024)

• Protein–ligand docking: PoseBusters benchmark — AF3 achieves 76% success vs ~50% for best docking tools (Glide, Gold)
• Protein–nucleic acid: substantially better than RoseTTAFold2NA and other specialist tools
• Antibody–antigen modeling: best-performing single model on H3 loop prediction
• Standard protein structure: broadly comparable to AF2, with some improvements on multimers

---

PART 8 — THE ALPHAFOLD DATABASE

• >200 million predicted protein structures (covering essentially every known UniProt sequence)
• Structures for all 20,226 human proteins
• Proteomes of >48 major organisms
• Downloadable in PDB and mmCIF format

1. Go to alphafold.ebi.ac.uk
2. Search by protein name, UniProt ID, or gene name
3. View 3D structure in browser (Mol* viewer) colored by pLDDT
4. Download structure + PAE matrix
5. Interpret confidence carefully before drawing biological conclusions

---

PART 9 — RUNNING ALPHAFOLD YOURSELF

9.1 ColabFold — The Practical Route

• Uses MMseqs2 for fast MSA generation (vs slow database searches in original AF2)
• ~40-60× faster than original AF2
• Supports monomers, multimers, and custom MSAs
• Free to use via Google Colab

1. Open ColabFold notebook in Google Colab
2. Paste your amino acid sequence(s)
3. Configure: num_recycles (3 default), num_models (1-5)
4. Run all cells (~10-30 min for typical protein)
5. Download: .pdb structures, PAE plots, pLDDT plots

9.2 Input Tips

• Sequence length: AF2 handles up to ~2500 residues well; longer sequences require more GPU memory
• MSA depth matters: Proteins with few homologs (orphan proteins, viral proteins unique to one clade) tend to have lower confidence
• Multimer modeling: Provide comma-separated chains; ColabFold handles chain pairing automatically
• Templates: Can be disabled to test truly ab initio prediction

9.3 Interpreting Your Output

Step 1: Check global pLDDT — overall model quality
Step 2: Identify high-confidence regions (pLDDT >70) — trust these
Step 3: Identify low-confidence regions (<50) — may be IDRs
Step 4: Check PAE matrix — confirms domain architecture and 
        whether inter-domain arrangements are meaningful
Step 5: If modeling a complex, check ipTM score (>0.75 = confident interface)
Step 6: Compare to evolutionary data — are predicted contacts consistent 
        with your functional knowledge?

---

PART 10 — LIMITATIONS & CRITICAL CAVEATS

---

PART 11 — APPLICATIONS IN RESEARCH & MEDICINE

Drug Discovery

• Target identification: AF3 predicts binding pockets for previously "undruggable" proteins
• Virtual screening: Docking small molecules into AF-predicted structures (though caution needed for induced-fit scenarios)
• Cryptic site discovery: Low-pLDDT regions sometimes reveal hidden binding pockets in MD simulations seeded from AF structures

Vaccine Design

• Predict surface-exposed epitopes on pathogen proteins (viral spike proteins, bacterial surface antigens)
• Design antigens that present conserved epitopes in native-like conformations

Protein Engineering & Synthetic Biology

• Design new enzymes by combining AlphaFold structures with RFdiffusion (David Baker lab) for de novo protein design
• Predict the effect of mutations on protein stability

Basic Science

• Completed the structural coverage of the human proteome
• Enabled AlphaFill — transplants ligands/cofactors into AF models
• Structural annotation of metagenomics sequences (millions of uncultured organism proteins)
• Understanding disease mutations: map variant to structure to understand mechanism

---

PART 12 — COMPETING METHODS

---

Summary: The AlphaFold Conceptual Stack

BIOLOGY LAYER
└── Protein sequence (primary structure) encodes 3D shape
    └── Co-evolution in related species reveals structural contacts

MACHINE LEARNING LAYER
└── MSA → Evoformer (transformer with triangle geometry updates)
    └── Pair representation encodes geometric hypothesis
        └── Structure module / Diffusion → explicit 3D coordinates
            └── Recycling → iterative refinement

OUTPUT LAYER
└── 3D atomic coordinates (.pdb / .cif)
    └── Per-residue pLDDT (confidence)
        └── PAE matrix (inter-residue confidence)

---

Key References

• Jumper J, et al. (2021). "Highly accurate protein structure prediction with AlphaFold." Nature 596:583–589
• Abramson J, et al. (2024). "Accurate structure prediction of biomolecular interactions with AlphaFold 3." Nature 630:493–500
• Mirdita M, et al. (2022). "ColabFold: making protein folding accessible to all." Nature Methods
• Chen L, et al. (2024). "AI-Driven Deep Learning Techniques in Protein Structure Prediction." Int J Mol Sci 25:8426 [PMID: 39125995]
• Krokidis MG, et al. (2025). "AlphaFold3: An Overview of Applications and Performance Insights." Int J Mol Sci 26:3671 [PMID: 40332289]
• EMBL-EBI AlphaFold Training Course — free, comprehensive, developed with Google DeepMind
• AlphaFold Database — 200M+ structures, free accessThis guide covers AlphaFold from the ground up across 12 structured parts. Here's a quick navigation map: