Alphafold2

(OpenFold redirects here)

AlphaFold2 is a protein structure prediction method from 2020, and is the first to make extensive use of the Transformer architecture, consisting of 97 million parameters. AlphaMissense is a derivative of this method.

Architecture of AlphaFold2 from Jumper et al. (1)

Architectural and ML contributions

Evoformer
Invariant point attention
Frame aligned point error
Self-distillation for protein structures
Predicted aligned error
Triangular update

Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583–9. Available from: https://doi.org/10.1038/s41586-021-03819-2

Quartz 4

Explorer

Alphafold2

Architectural and ML contributions

AlphaFold2 recapitulates interaction biases from PDB when modeling antibodies

Antibody structure prediction improved with AlphaFold2 features

Dropout improves AF2 multimer prediction, including for antibody-antigen complexes

AF2 RMSD to ground truth correlated with RMSD between conformations

AlphaFold2 and RosettaFold sometimes produce different conformations

AlphaFold2 generalizes to unseen secondary structures and topologies

AlphaFold2 outperforms ESMFold at distinguishing designable and undesignable protein backbones

AlphaFold2 predicts Apo conformations with poorer RMSD than holo conformations

AlphaFold2 predicts ground states more often

AlphaFold2 predicts the holo form of proteins in most cases

AlphaFold3 performs comparably to AlphaFold2 when predicting multiple conformations of fold-switching proteins

Alternate conformations can be sampled by inverting AlphaFold2 and backpropagating along specific user-picked collective variables

Antibodies can be humanized using AlphaFold2 models

Conformationally stabilizing mutations can be identified from sequence clusters that are predicted to fold into that conformation when passed to AlphaFold2

Fine-tuning AlphaFold for MHC prediction by adding Evoformer layers improves RMSD and pLDDT

IPA-based structure prediction neural networks train sequentially along three spatial axes

Monomeric AF2 is better able to distinguish positive and negative PPIs than AF2-multimer

Protein folding neural networks can be subject to adversarial attacks

Protein folding neural networks make local optimizations in the absence of coevolutionary information

ProteinMPNN and AlphaFold structure update module can be combined into a VAE

Scrambled MSAs outperform single-sequence-prediction with AlphaFold

Structure prediction neural networks trained on the same structural dataset are likely to predict the same conformation

Structure prediction with AlphaFold improved when starting from an initial guess, rather than default initialization

The AlphaFold and ESMFold databases can be mined for alternate conformations

The confidence metrics of AlphaFold2 are better calibrated than those of AlphaFold3

ddG predictions calculated from AlphaFold structures are equally accurate as those derived from experimental structures