PROTEIN STRUCTURE PREDICTION

NovaFold and NovaFold AI are both services that run through the Protean 3D interface.

NovaFold uses the I-TASSER algorithm developed by Professor Yang Zhang of the Department of Computational Medicine and Bioinformatics at the University of Michigan. This protein structure prediction algorithm utilizes a combination of “threading” and “ab initio folding” in predicting protein structure. Threading attempts to match portions of the query sequence to template sequences. The template sequences, and their experimentally solved structures, are part of the RCSB Protein Data Bank (PDB). Ab initio folding uses biophysical properties of the query sequence and simulations to determine the likely structure(s) of the protein.

NovaFold AI uses AlphaFold 2, an artificial intelligence-based algorithm developed by Google’s DeepMind, to predict distance and create dihedral maps using deep multiple sequence alignments as input. This algorithm does not directly use templates to place atomic coordinates but can use templates as part of the map building process.

The “gold standard” for determining the accuracy of protein structure prediction algorithms is the CASP (Critical Assessment of Protein Structure Prediction) challenge, organized by the Protein Structure Prediction Center. Each CASP experiment is a biennial, world-wide evaluation of structure prediction methods, with approximately 100 participating laboratories entering one or more different algorithms into competition. Each algorithm must predict dozens of structures based on their protein sequences alone. The predictions are then mathematically compared to each structure as determined through x-ray crystallography.

In CASP competitions from 2006-2018, the I-TASSER (also called “Zhang-Server”) algorithm used by NovaFold  was ranked:

CASP7 – 1st
CASP8 – 1st
CASP9 – 1st (tie)
CASP10 – 1st
CASP11 -1st
CASP12 – 3rd
CASP13 – 2nd

The CASP14 experiments in 2020 tested 146 algorithms. AlphaFold 2, the algorithm used in NovaFold AI, made its debut in this competition, and earned the first-place ranking. In fact, AlphaFold 2 was an astonishing 2.65x more accurate at predicting protein structures than the second-ranked algorithm. I-TASSER still put in a strong showing, coming in at 5th place.

Results of a completed NovaFold AI prediction can be opened and viewed just like any other structure in Protean 3D. By contrast, NovaFold predictions are shown in a specialized Report view in Protean 3D. After running a NovaFold prediction, the Report view opens automatically. The Report uses interactive images and tables to show which templates were used in the prediction and the models that were the best match to the query. Scroll to the prediction model of interest and click the link “Open model in new document” to open it as a new Protean 3D document for further analysis.

While NovaFold is largely based on I-TASSER, we’ve enhanced NovaFold protein structure predictions in several ways:

• NovaFold predictions include improved bond geometries and reduced atomic clashes, creating energetically minimized structure models.
• NovaFold can also search custom template libraries for a more specialized approach to homology modeling – contact support@dnastar.com or request a webinar for information on building a custom library.
• NovaFold is fully integrated with Protean 3D, creating a seamless experience for visualizing and analyzing your predicted protein models.

All of these enhancements build on the foundational I-TASSER algorithms in NovaFold.