AlphaFold: Difference between revisions

Line 2:

See [[Theoretical_models#2020:_CASP_14]] for more about the initial demonstration at CASP14, and the reactions to it. [[AlphaFold2_examples_from_CASP_14]] describes a detailed analysis of two of the CASP14 predictions.

==AlphaFold published in 2021==

AlphaFold was published in July, 2021<ref name="af2021">PMID: 34265844</ref>. Methods were described in considerable detail. The source code, trained weights, and inference script were made available under an open-source license. Structure prediction required about one GPU (Graphics Processing Unit) minute per model of about 384 amino acids.

Impressively, AlphaFold had remarkable success predicting a set of 10,795 protein chain structures (filtered for high reliability, lengths restricted to 80-1,400 residues) published in the [[PDB]] after AlphaFold's training set<ref>The training set cutoff was 2018/04/30. The test set was obtained between then and 2021/02/15.</ref>. Overall alpha carbon accuracy had a median of 1.46 Å [[Calculating GDT TS|RMSD]] at 95% coverage. The majority of chain structures were predicted with full-chain alpha carbon RMSD values <2 Å. About 25% were predicted with RMSD >4 Å.

Importantly, each prediction comes with a confidence score that reliably predicts the accuracy of the predicted structure.

Accurate prediction of sidechains required accurate prediction of the main chain. Accurate prediction required a multiple sequence alignment depth >~30 sequences, with a depth of ~100 sequences being adequate.

==References==

AlphaFold: Difference between revisions

Navigation menu

Page actions

Page actions

Personal tools

Navigation

Search

Tools

@@ Line 2: / Line 2: @@
 See [[Theoretical_models#2020:_CASP_14]] for more about the initial demonstration at CASP14, and the reactions to it. [[AlphaFold2_examples_from_CASP_14]] describes a detailed analysis of two of the CASP14 predictions.
+==AlphaFold published in 2021==
+AlphaFold was published in July, 2021<ref name="af2021">PMID: 34265844</ref>. Methods were described in considerable detail. The source code, trained weights, and inference script were made available under an open-source license. Structure prediction required about one GPU (Graphics Processing Unit) minute per model of about 384 amino acids.
+Impressively, AlphaFold had remarkable success predicting a set of 10,795 protein chain structures (filtered for high reliability, lengths restricted to 80-1,400 residues) published in the [[PDB]] after AlphaFold's training set<ref>The training set cutoff was 2018/04/30. The test set was obtained between then and 2021/02/15.</ref>. Overall alpha carbon accuracy had a median of 1.46 Å [[Calculating GDT TS|RMSD]] at 95% coverage. The majority of chain structures were predicted with full-chain alpha carbon RMSD values <2 Å. About 25% were predicted with RMSD >4 Å.
+Importantly, each prediction comes with a confidence score that reliably predicts the accuracy of the predicted structure.
+Accurate prediction of sidechains required accurate prediction of the main chain. Accurate prediction required a multiple sequence alignment depth >~30 sequences, with a depth of ~100 sequences being adequate.
 ==References==
 <references />