MIT AI peptides show 50% success

MIT researchers said this week that an artificial intelligence method they call FragFold designed short protein fragments that could alter biomolecular condensates in about half of the cases they tested. A May 8 bioRxiv preprint from Andrew Savinov, Jibin Sadasivan and colleagues at the Massachusetts Institute of Technology said the team screened 2,235 fragments computationally and then tested 18 candidates in experiments. Nine of those 18 designs succeeded, according to the preprint. The authors said the work covered proteins tied to stress response, viral infection, neurodegeneration and cancer-related signaling. ### What exactly did the MIT team build? FragFold is an AI-driven method for finding protein fragments that bind to larger proteins and change how they behave, according to the May 8 preprint. The paper said the approach was used to discover fragments that either inhibited or enhanced formation of biomolecular condensates — membrane-less clusters of proteins and nucleic acids that organize cellular activity. (biorxiv.org) A February 20, 2025 MIT News article described FragFold as an extension of earlier work from Gene-Wei Li’s lab that used AlphaFold-based predictions to identify inhibitory protein fragments in E. coli. That earlier version was reported in the Proceedings of the National Academy of Sciences, MIT said, and the new condensate work applies the same general strategy to a different biological problem. (biorxiv.org) ### Which proteins were tested, and where did the 50% figure come from? The MIT preprint named four condensate-forming proteins: G3BP1, the SARS-CoV-2 nucleocapsid protein, TDP-43 and focal adhesion kinase, or FAK. G3BP1 is linked to stress granules, TDP-43 is widely studied in ALS and related neurodegenerative disease, SARS-CoV-2 nucleocapsid is a viral structural protein, and FAK has been implicated in cancer signaling, according to the paper and background material from MIT and other scientific sources. (news.mit.edu) The 50% figure came from 9 successful designs out of 18 candidates selected for laboratory follow-up after the computational screen, the preprint said. The authors wrote that they tested only three to five candidates per target protein and that each of the four proteins had a success rate of at least 40%. ### What did the experiments show in cells and test systems? (biorxiv.org) The preprint said the designed fragments changed condensate behavior across the four targets and that predicted binding modes matched whether a fragment enhanced or inhibited condensate formation. In the FAK experiments, the authors wrote that one inhibitory fragment identified a domain interaction required for phase separation and that mutational analysis supported that mechanism. (biorxiv.org) The same FAK fragment also suppressed FAK condensate formation in living mammalian cells, according to the paper. That makes the cell-based result narrower than a therapeutic claim: the preprint reports control of condensate formation in vitro and in cells, but it does not report animal efficacy or human testing. ### Why are cancer, ALS and SARS-CoV-2 part of this story? (biorxiv.org) The disease links come from the proteins the team chose to test. TDP-43 is a central protein in ALS research, SARS-CoV-2 nucleocapsid is part of the coronavirus replication machinery, and FAK is a cancer-related signaling protein whose condensate behavior has been studied in tumor biology, the sources show. (biorxiv.org) MIT’s authors wrote that dysregulated condensates contribute to disease states including neurodegeneration, cancer and viral infection. That framing appears in the introduction to the preprint, where the team presents the method as a way to map and control the interactions that drive condensate formation. ### How much of this is established, and how much is still preliminary? (biorxiv.org) May 8, 2026 is the date on the bioRxiv preprint, which means the work has been disclosed publicly but not yet certified by peer review. The authors listed are Savinov, Sadasivan, Kyle J. White, Jack D. Rubien, Gene-Wei Li and Lindsay B. Case, with Savinov and Case marked as corresponding authors. (biorxiv.org) The next concrete milestone is a peer-reviewed publication or additional experimental follow-up from the MIT group. For now, the public record is the May 8 bioRxiv manuscript, which reports the 2,235-fragment screen, the 18 experimental tests and the 9 successful condensate-control designs. (biorxiv.org)