<html>
<head>
<meta http-equiv="content-type" content="text/html; charset=UTF-8">
</head>
<body>
<p>Last month, thanks to some of my colleagues*, I was introduced to
the breakthrough by AlphaFold2 announced November, 2020.</p>
<p>AlphaFold2 predicted protein structures with truly astonishing
accuracy, certified by the bi-annual double-blind competition,
CASP 14, founded by John Moult in the early 1990s. Its predictions
were based on the amino acid sequences of the target proteins,
using massive artificial intelligence machine learning from
sequence and structure databases. Predictions were made "blind",
without access to empirical structures of the targets, and were
judged later in 2020 when empirical structures became public. The
judges did not know who made which prediction.<br>
</p>
<p>AlphaFold2 was one of over 100 groups that submitted predictions
for over 100 target single-chain domains. In most cases,
AlphaFold2 made the best prediction, while the second best
prediction was far less accurate. This was particularly impressive
for "free modeling" targets, those for which no suitable homology
modeling templates were available.</p>
<p>I have briefly summarized the breakthrough here:</p>
<p><a class="moz-txt-link-freetext"
href="https://proteopedia.org/w/Theoretical_modeling#Ab_Initio_Models"
moz-do-not-send="true">https://proteopedia.org/w/Theoretical_modeling#Ab_Initio_Models</a></p>
<p>I have analyzed two free modeling cases in detail, with
interactive 3D comparisons. One (92 amino acids) is the ORF8
virulence factor from SARS-CoV-2. Among the free modeling targets,
it had the largest discrepancy between the best and 2nd best
predictions. The second is a phage RNA polymerase, the longest
free-modeling target domain (404 amino acids). See:</p>
<p><a class="moz-txt-link-freetext"
href="https://proteopedia.org/w/AlphaFold2_examples_from_CASP_14"
moz-do-not-send="true">https://proteopedia.org/w/AlphaFold2_examples_from_CASP_14</a></p>
<p>I welcome comments, corrections, additions, or feedback!</p>
<p>*Special thanks to my colleagues Roman Sloutsky, Nathaniel Kuzio,
Erik Nordquist, Can Ozden, Thomas Sawyer, Katie Wahlbeck, Jeanne
Hardy, and Scott Garman all at UMass Amherst, and Woody Sherman of
Silicon Therapeutics.</p>
<p>-Eric</p>
<p> </p>
<div style="-en-clipboard:true;">Eric Martz, Professor Emeritus,
Dept Microbiology (he/him/his) </div>
<div>U Mass, Amherst -- <a href="http://martz.molviz.org/"
moz-do-not-send="true">Martz.MolviZ.Org</a> </div>
<div>Guided Exploration of Macromolecules with Powerpoint-Ready
Animations: <a href="http://firstglance.jmol.org/"
moz-do-not-send="true">FirstGlance.Jmol.Org</a> </div>
<div>Protein 3D Structure Wiki: <a
href="http://proteopedia.org/w/User:Eric_Martz"
moz-do-not-send="true">Proteopedia.Org</a> </div>
<div>Education: Biochem in 3D at <a href="http://molviz.org/"
moz-do-not-send="true">MolviZ.Org</a> </div>
<div>Find Functional Patches in Proteins: <a
href="http://consurf.tau.ac.il/" moz-do-not-send="true">ConSurf.tau.ac.il</a>
</div>
<div>Atlas of Macromolecules: <a href="http://atlas.molviz.org/"
moz-do-not-send="true">Atlas.MolviZ.Org</a> </div>
<div>See multiple protein sequence alignments clearly: <a
href="http://msareveal.org/" moz-do-not-send="true">MSAReveal.Org</a>
</div>
<div>Pockets And Cavities Using Pseudoatoms in Proteins: <a
href="http://molviz.org/pacupp" moz-do-not-send="true">PACUPP</a>
</div>
<div>Interactive Molecules in Architectural Spaces: <a
href="http://molecularplayground.org/" moz-do-not-send="true">MolecularPlayground.Org</a>
</div>
<div>Syllabi: <a href="http://workshops.molviz.org/"
moz-do-not-send="true">Workshops.MolviZ.Org</a> </div>
</body>
</html>