Highly accurate protein structure prediction with AlphaFold
John Jumper, Richard Evans, Alexander Pritzel et al.
2021
Proteins are essential to life, and understanding their structure is key to understanding their function. We describe AlphaFold 2, a system that achieves around 92.4 GDT score on CASP14 — significantly outperforming all other methods.
Reading Level:
AlphaFold 2: Architecture Overview
AlphaFold 2 combines evolutionary information (Multiple Sequence Alignments) with geometric reasoning through novel architectural components.
Evoformer Block
Processes a pair representation $z_{ij}$ and MSA representation $m_{si}$ jointly:
Row-wise gated self-attention with pair bias: $$a_{qk} = \text{softmax}\left(\frac{1}{\sqrt{c}}(\mathbf{q}_q^T \mathbf{k}k + b{qk})\right)$$
Outer product mean updates pair from MSA: $$z_{ij} \leftarrow z_{ij} + \text{LayerNorm}\left(\frac{1}{s}\sum_s \mathbf{o}{si} \otimes \mathbf{o}{sj}\right)$$
Triangle multiplicative update enforces geometric consistency: $$z_{ij} \leftarrow \text{LayerNorm}\left(\sigma(g_{ij}) \odot \sum_k z_{ik} \odot z_{jk}\right)$$
Structure Module
Operates on invariant point attention (IPA) in 3D space:
$$h_i^{(l+1)} = \text{IPA}(h_i^{(l)}, z_{ij}, T_i^{(l)})$$
Where $T_i = (R_i, \mathbf{t}_i)$ represents backbone frame as rotation + translation.
Results
- CASP14: 92.4 GDT (previous best: 68.0)
- 98.5% of residues within 2Å of experimental structure (TM-score > 0.9)
- Predicted structures for 214M proteins in UniRef90
XP Reward
+225
Earned after 10 seconds of reading
Key EquationsClick for Python code
Citation Graph
References (3)
Protein structure prediction using multiple deep neural networks
Senior et al. · 2020MSA Transformer
Rao et al. · 2021End-to-end differentiable learning of protein structure
Ingraham et al. · 2019