The rockefeller university

Event Details

Type

Other Seminars

Speaker(s)

Anton Feenstra, Ph.D., Associate Professor; Director of Bioinformatics, Vrije Universiteit, Amsterdam

Speaker bio(s)

Protein interactions are crucial for understanding biological functions and disease mechanisms, but predicting these remains a complex task in computational biology. Understanding the binding site whereby proteins interact can shed light on how protein interactions come aboutt, and we can predict interface positions from sequence using deep learning. We previously showed that traditional `hand-crafted' features each contribute positively. Here, we show that PIPENN-EMB, using an ensemble of Unet, Rnet and Rnn networks and exploring the using of ProtT5-XL embeddings offers substantial improvement over `hand-crafted' features, reaching state-of-the-art performance with an AUC-ROC of 0.815 on the ZK448 benchmark set, on par with structure-based methods. We also study the enrichment on the PPI interface of treatment resistance-related mutations in Micobacterium tuberculosis (MTB). We furthermore show that PIPENN-EMB yields stable AUROC also for proteins with less than 30% sequence identity to the training dataset. The ability to predict interface regions from sequence opens up the possiblity to do such analyses truly at proteome scale.

Open to

Campus and Guests