세미나

Protein analysis has witnessed a revolution through sophisticated AI methodologies. At the forefront are highly accurate structure prediction methods, alongside advanced language models. Structure prediction methods such as AlphaFold2 and ESMFold have generated an avalanche of publicly available protein structures. The AlphaFold database and ESMatlas contain over 214 and 620 million predicted structures, respectively, covering nearly every protein sequence in our largest protein reference databases. In this talk, I will discuss how this avalanche of structural data can be made searched and clustered and discuss its potential to improve genomic and proteomic annotation.
Additionally, I will cover how language models are changing protein predictions, exemplified by tackling the challenge of immunosuppressive tumors. Leveraging language models and homology search, we identified and validated effective kynureninases that degrade L-kynurenine, a key immunosuppressive agent in tumors. The most active of these enzymes significantly reduced tumor weight in a mouse model, demonstrating a new frontier in cancer therapy empowered by AI.