Scientific Summary: Cyrus NextGen Antibody structure prediction beats Schrodinger and others in a large independent benchmark
Over the last few years Cyrus has worked on problems on a range of proteins from enzymes to non-antibody biologics to vaccines to some antibody work. However most biologic drugs are monoclonal antibodies or variants thereof, and historically this has been a small share of our work, and an area where there are many very strong commercial offerings. Our antibody structure prediction tool, based on previous methods in Rosetta, has been a good competitor in that space, but has faced stiff competition.
In other areas of protein structure prediction using structural homology, Rosetta and Cyrus have been the leaders in many independent benchmarks by academics and in blind tests by industry users. In antibody structure that has not been the case, given very good algorithms from Schrodinger and CCG.
Over the last two years scientists at Cyrus, led by our CSO Dr. Yifan Song, have built a new method for antibody prediction based on Rosetta algorithms historically used for general protein homology, but not for antibodies — Cyrus NextGen Antibody. Because these algorithms have performed so well over the last 7 years since their introduction in 2013 for general protein structure prediction, many of us expected that they would perform well for antibodies once properly adapted and tuned.
In the fall of 2019 we completed this work, and our internal testing showed clear superiority, producing more accurate structures than any other method. The gold standard, though, would be a test by a third party, judged by their own quantitative criteria, across a relatively large number of antibodies.
Now, in July 2020, we’ve completed such a test over 26 antibodies with NextGen against the latest Schrodinger software and two other top-performing software packages. We were very pleased to find that NextGen outperformed all of the other methods in this rigorous test, and now we are publicizing these results for the first time in a scientific blog post, before making a more extensive manuscript available.
This is an important step forward for Cyrus, but more importantly it promises more accurate results in antibody efficacy and safety predictions, and ultimately a variety of better and more effective antibody drugs produced by Cyrus algorithms for a wide range of diseases. For example, better models from NextGen could enable faster development of an antibody drug, or make certain diseases susceptible to antibody drugs for the first time. Better models could also enable the invention of second-generation versions of existing drugs, such as the popular “TNF-alpha inhibitor” arthritis drugs, with fewer immunogenic side effects or less frequent injections.