Describe protein structure prediction (AlphaFold)

6.4.8 · Biology › Bioinformatics & Computational Biology

YEH problem exist kyun karta hai?

Yeh matter kyun karta hai: Structure ⇒ function. Enzymes, antibodies, receptors, aur drug targets sab shape ke through kaam karte hain. Shape jaanna hume drugs design karne, disease mutations samajhne, aur proteins engineer karne deta hai — bina slow, expensive experiments (X-ray crystallography, cryo-EM, NMR) ke.

AlphaFold input mein KYA use karta hai

AlphaFold HOW kaam karta hai (conceptual pipeline)

Figure — Describe protein structure prediction (AlphaFold)

Step 1 — Databases search karo homologous sequences ke liye ⇒ MSA banao, aur template structures (known related structures) dhundho.

Step 2 — Evoformer (core neural network). Yeh do representations maintain karta hai aur unhe "baat karne" deta hai:

MSA representation (evolutionary/coevolution signal),
pair representation — ek matrix $z_{ij}$ jo residues $i$ aur $j$ ke beech relationship encode karta hai.

Evoformer attention use karta hai dono ko iteratively refine karne ke liye, geometric consistency enforce karta hai (e.g. residues $i, j, k$ ke beech distances par triangle inequality).

Step 3 — Structure Module. Refined pair representation ko actual 3D coordinates mein convert karta hai. Har residue ko ek rigid frame (ek position + rotation) ki tarah treat kiya jaata hai, aur network predict karta hai har frame ko kaise move karein — ek operation jo puri protein ko rotate/translate karne par invariant rehta hai.

Step 4 — Recycling. Output ko kai baar input ki tarah wapas feed kiya jaata hai, prediction sharpen karne ke liye.

Step 5 — Confidence score (pLDDT). Har residue ke liye yeh ek pLDDT (0–100) output karta hai jo predict karta hai ki woh region kitna accurate hai. High pLDDT ⇒ trustworthy; low pLDDT aksar flexible ya disordered regions mark karta hai.

Formula Predicted distances ko structure mein convert karna (core geometric logic)

KYU: AlphaFold residues ke beech relationships predict karta hai, phir coordinates reconstruct karta hai. Ise samajhne ka ek clean tarika classic distance-geometry step hai: agar aap saare pairwise distances $d_{ij}$ jaante ho, toh aap 3D coordinates recover kar sakte ho.

Scratch se derivation. Maano residue $i$ ka coordinate vector $x_{i}$ hai. Pehle atom ko origin par rakho, $x_{1} = 0$ . Squared distance hai $d_{ij}^{2} = ∥ x_{i} - x_{j} ∥^{2} = ∥ x_{i} ∥^{2} + ∥ x_{j} ∥^{2} - 2 x_{i} \cdot x_{j} .$ $x_{1} = 0$ use karne par $∥ x_{i} ∥^{2} = d_{1 i}^{2}$ milta hai. Substitute karo: $x_{i} \cdot x_{j} = \frac{1}{2} (d_{1 i}^{2} + d_{1 j}^{2} - d_{ij}^{2}) .$ Yeh sirf distances se Gram matrix $G_{ij} = x_{i} \cdot x_{j}$ banata hai. $G = U Λ U^{⊤}$ diagonalise karo aur 3 largest eigenvalues rakho toh coordinates milte hain $X = U_{3} Λ_{3}^{1/2},$ yaani 3D positions. AlphaFold iska kaafi zyada powerful version seekhta hai, lekin yeh dikhata hai kyun "distances/relations predict karo $\to$ coordinates pao" mathematically sound hai.

Worked examples

Common mistakes (steel-manned)

Recall Feynman: ek 12-saal ke bachche ko explain karo

Ek protein ek bohot lamba dhaaga hai jisme alag-alag rang ke beads hain. Tumhare body mein yeh dhaaga automatically ek khaas shape mein simatt jaata hai, aur woh shape ek key ki tarah hoti hai jo ek lock mein fit hoti hai koi kaam karne ke liye. Yeh shape pata karne mein scientists ko lab mein saalon lagte the. AlphaFold ek bahut smart computer hai jo kaafi saare similar dhaagon ko kai janwaron aur paudhon se dekh kar shape guess karta hai. Usne notice kiya: "jab bhi yeh bead badalti hai, woh door wali bead bhi badalti hai — toh jab dhaaga fold hota hai woh ek doosre ke paas honi chahiye!" Aise hazaron clues use karke woh complete folded shape draw karta hai, aur yeh bhi batata hai ki uski drawing ke kaun se parts par use confidence hai.

Recall flashcards

Protein ka function kya decide karta hai?

Uski folded 3D structure (shape function determine karti hai).

Levinthal's paradox state karo.

Ek protein mein astronomically saari possible conformations hoti hain, isliye random search se reasonable time mein fold nahi ho sakta — phir bhi ms mein fold hoti hai, iska matlab folding guided hai (energy funnel), random nahi.

AlphaFold mein MSA kya hai?

Ek Multiple Sequence Alignment: target aur kaafi evolutionarily related sequences column-by-column aligned.

Coevolution 3D contacts kyun reveal karta hai?

Contact mein residues saath mutate hote hain — ek mein destabilising mutation doosre mein change se compensate hoti hai, isliye correlated columns spatial closeness imply karte hain.

Evoformer jo do representations refine karta hai woh kaun si hain?

MSA representation aur pair representation (

z_{ij}

Structure Module kya output karta hai?

3D coordinates, har residue ko ek rigid frame (position + rotation) treat karke jo global rotation/translation ke liye invariant hai.

AlphaFold mein recycling kya hai?

Network ka output kai baar input ki tarah wapas feed karna taaki prediction iteratively sharpen ho.

pLDDT kya measure karta hai?

Per-residue predicted confidence (0–100) local structural accuracy ka; low values aksar flexible/disordered regions mark karti hain.

Pairwise distances se 3D coordinates kyun mil sakte hain?

Distances Gram matrix deti hain; uske top 3 eigenvalues/vectors coordinates yield karte hain (shape rotation/translation tak fixed).

AlphaFold ke liye kaun si proteins sabse mushkil hain aur kyun?

Orphan proteins jinke paas kam homologues hain (shallow MSA) — contacts infer karne ke liye kam coevolution signal hota hai.

Kya AlphaFold folding pathway simulate karta hai?

Nahi — yeh final structure seedha learned patterns aur geometric inference se predict karta hai, physics-based folding dynamics se nahi.

Kya AlphaFold function predict karta hai?

Nahi — yeh structure predict karta hai; function baad mein infer kiya jaata hai.

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