Describe protein structure prediction (AlphaFold)

WHY does this problem exist?

WHY it matters: Structure ⇒ function. Enzymes, antibodies, receptors, and drug targets all work through shape. Knowing the shape lets us design drugs, understand disease mutations, and engineer proteins — without slow, expensive experiments (X-ray crystallography, cryo-EM, NMR).

WHAT AlphaFold uses as input

HOW AlphaFold works (conceptual pipeline)

Figure — Describe protein structure prediction (AlphaFold)

Step 1 — Search databases for homologous sequences ⇒ build the MSA, and find template structures (known related structures).

Step 2 — Evoformer (the core neural network). It maintains two representations and lets them "talk":

the MSA representation (evolutionary/coevolution signal),
the pair representation — a matrix $z_{ij}$ encoding the relationship between residues $i$ and $j$ .

The Evoformer uses attention to iteratively refine both, enforcing geometric consistency (e.g. the triangle inequality on distances between residues $i,j,k$ ).

Step 3 — Structure Module. Turns the refined pair representation into actual 3D coordinates. Each residue is treated as a rigid frame (a position + rotation), and the network predicts how to move each frame — an operation invariant to how you rotate/translate the whole protein.

Step 4 — Recycling. The output is fed back as input several times, sharpening the prediction.

Step 5 — Confidence score (pLDDT). For every residue it outputs a pLDDT (0–100) predicting how accurate that region is. High pLDDT ⇒ trustworthy; low pLDDT often marks flexible or disordered regions.

Formula Turning predicted distances into a structure (the core geometric logic)

WHY: AlphaFold predicts relationships between residues, then reconstructs coordinates. A clean way to see this is the classic distance-geometry step: if you know all pairwise distances $d_{ij}$ , you can recover 3D coordinates.

Derivation from scratch. Let residue $i$ have coordinate vector $\mathbf{x}_i$ . Place the first atom at the origin, $\mathbf{x}_1=\mathbf 0$ . The squared distance is $d_{ij}^2=\lVert \mathbf{x}_i-\mathbf{x}_j\rVert^2=\lVert\mathbf x_i\rVert^2+\lVert\mathbf x_j\rVert^2-2\,\mathbf x_i\!\cdot\!\mathbf x_j.$ Using $\mathbf x_1=\mathbf 0$ gives $\lVert\mathbf x_i\rVert^2=d_{1i}^2$ . Substitute: $\mathbf x_i\!\cdot\!\mathbf x_j=\tfrac12\big(d_{1i}^2+d_{1j}^2-d_{ij}^2\big).$ This builds the Gram matrix $G_{ij}=\mathbf x_i\!\cdot\!\mathbf x_j$ from distances alone. Diagonalising $G=U\Lambda U^\top$ and keeping the 3 largest eigenvalues gives coordinates $X=U_{3}\,\Lambda_{3}^{1/2},$ i.e. the 3D positions. AlphaFold learns a far more powerful version, but this shows why "predict distances/relations $\to$ get coordinates" is mathematically sound.

Worked examples

Worked example Example 1 — Reconstruct a triangle of 3 residues from distances

Given $d_{12}=3,\; d_{13}=4,\; d_{23}=5$ . Find 2D coordinates.

Put $\mathbf x_1=(0,0)$ . Why? Kills translation freedom.
Put $\mathbf x_2=(3,0)$ since $d_{12}=3$ . Why? Kills rotation freedom; place along an axis.
Let $\mathbf x_3=(x,y)$ . Then $x^2+y^2=d_{13}^2=16$ and $(x-3)^2+y^2=d_{23}^2=25$ .
Subtract: $x^2-(x-3)^2 = 16-25 \Rightarrow 6x-9=-9 \Rightarrow x=0$ . Why? Eliminates $y$ .
So $y^2=16 \Rightarrow y=4$ . Coordinates $(0,4)$ .
Check: $3^2+4^2=5^2$ ✓ — it's a right triangle, exactly consistent with the distances.

Common mistakes (steel-manned)

Recall Feynman: explain it to a 12-year-old

A protein is like a really long piece of string with different-coloured beads. In your body this string automatically scrunches up into one special shape, and that shape is like a key that fits a lock to do a job. Figuring out the shape used to take scientists years in the lab. AlphaFold is a very smart computer that guessed the shape by studying millions of similar strings from many animals and plants. It noticed: "whenever this bead changes, that far-away bead also changes — so they must sit next to each other when the string folds!" Using thousands of such clues it draws the finished folded shape, and even tells you which parts of its drawing it's sure about.

Recall flashcards

What determines a protein's function?

Its folded 3D structure (shape determines function).

State Levinthal's paradox.

A protein has astronomically many possible conformations, so random search couldn't fold it in reasonable time — yet it folds in ms, implying folding is guided (energy funnel), not random.

What is an MSA in AlphaFold?

A Multiple Sequence Alignment: the target plus many evolutionarily related sequences aligned column-by-column.

Why does coevolution reveal 3D contacts?

Residues in contact mutate together — a destabilising mutation in one is compensated by a change in its neighbour, so correlated columns imply spatial closeness.

What are the two representations the Evoformer refines?

The MSA representation and the pair representation (

z_{ij}

What does the Structure Module output?

3D coordinates by treating each residue as a rigid frame (position + rotation) invariant to global rotation/translation.

What is recycling in AlphaFold?

Feeding the network's output back as input several times to iteratively sharpen the prediction.

What does pLDDT measure?

Per-residue predicted confidence (0–100) of local structural accuracy; low values often mark flexible/disordered regions.

Why can you get 3D coordinates from pairwise distances?

Distances give the Gram matrix; its top 3 eigenvalues/vectors yield coordinates (shape fixed up to rotation/translation).

Which proteins are hardest for AlphaFold and why?

Orphan proteins with few homologues (shallow MSA) — little coevolution signal to infer contacts.

Does AlphaFold simulate the folding pathway?

No — it predicts the final structure directly via learned patterns and geometric inference, not physics-based folding dynamics.

Does AlphaFold predict function?

No — it predicts structure; function is inferred afterward.

Connections

Protein Structure Levels (Primary–Quaternary)
Multiple Sequence Alignment
Coevolution and Contact Prediction
Deep Learning and Attention Mechanisms
X-ray Crystallography and Cryo-EM
Distance Geometry and Eigen-decomposition
Structure-Based Drug Design
Intrinsically Disordered Proteins

Concept Map

Hinglish (regional understanding)

Intuition Hinglish mein samjho

Dekho, protein ek lambi chain hoti hai amino acids ki — bilkul beads-on-a-string jaisi. Par yeh chain apne aap ek specific 3D shape mein fold ho jaati hai, aur yahi shape decide karta hai ki protein ka kaam kya hoga (enzyme, receptor, antibody, sab). Problem yeh hai ki possible shapes itni zyada hoti hain (Levinthal's paradox) ki brute-force try karna impossible hai. AlphaFold ek deep-learning model hai jo sequence se seedha 3D shape predict kar deta hai — bina years lagaye lab experiments mein.

AlphaFold ka asli jaadu hai MSA (Multiple Sequence Alignment). Woh alag-alag species se milte hue similar proteins uthata hai aur dekhta hai ki kaun-se residues ek saath mutate hote hain. Agar residue $i$ aur $j$ hamesha saath mein badalte hain, iska matlab woh fold hone par 3D mein paas-paas hote hain (coevolution). Yeh contacts hi structure ka skeleton dete hain. Evoformer network attention se in signals ko refine karta hai, phir Structure Module har residue ko ek rigid frame maan kar actual 3D coordinates nikal deta hai.

Ek important cheez: AlphaFold physics ka movie nahi chalata (folding step-by-step simulate nahi karta) — woh pattern seekh kar seedha final shape ka guess deta hai. Aur woh har residue ke liye pLDDT confidence score deta hai: high score = bharosa karo, low score = shayad woh region flexible ya disordered hai. Jahan homologues bahut kam hain (orphan proteins), wahan MSA patla hota hai aur prediction weak — yeh model ki limitation hai. Isliye AlphaFold structure predict karta hai, function nahi; function humein shape dekh kar alag se infer karna padta hai.

Test yourself — Bioinformatics & Computational Biology

Connections