On the characterization of protein interaction interfaces with computational approaches and experimental validations

Abstract

Proteins play crucial roles in virtually all cellular processes, and their functions are of- ten realized through interactions with other proteins via different interfaces. Thus, for a comprehensive understanding of protein functions, it is important to investigate the molecular mechanisms of protein-protein interactions (PPIs) by characterizing their interaction interfaces. Unfortunately, due to most high-throughput assays detecting only PPIs and not their interfaces, mechanistic information is scarce for most PPI data. The current thesis presents a collection of studies exploring the characterization of different types of PPI interfaces in the human protein interactome. Domains and motifs are two types of conserved functional modules that enable PPIs by forming domain-domain interfaces (DDIs) and domain-motif interfaces (DMIs). Chapter 2 and Article I delved into the characterization of DMIs and DDIs in PPIs by leveraging information from existing databases. Chapter 2 centered on automating the detection and scoring of DMIs in PPIs through a computer program. Article I focused on evaluating the DDIs annotated in the 3did database by manually curating a reference dataset of DDIs and identifying useful features to further score them. The program developed in Chapter 2 and the high-confidence DDIs from 3did were applied to PPIs detected in the human protein interactome to characterize their interfaces. AlphaFold-Multimer (AF-MM) is an artificial intelligence (AI)-based tool for pre- dicting the structures of protein complexes. Article II and III investigated the use of AF-MM to predict novel PPI interfaces. As there is a lack of a comprehensive as- sessment of AF-MM and its metrics, Article II systematically benchmarked AF-MM’s ability to predict different types of interfaces and established criteria for discriminat- ing good from bad structural models. Testing AF-MM using sequences longer than minimal interacting regions revealed that they are detrimental to AF-MM’s prediction performance, prompting the development of a fragmentation-based approach to en- hance AF-MM’s sensitivity. The approach was applied to PPIs detected in the human protein interactome to predict their interfaces, and some predicted interfaces were ex- perimentally validated. Similarly, article III applied the fragmentation approach on a protein pair whose interaction is important for piRNA amplification. Subsequent experimental validation also confirmed the interaction interface predicted by AF-MM. This thesis provides various approaches to leverage existing knowledge on different PPI interface types to predict PPI interfaces in the human protein interactome, paving the way towards a mechanistically annotated human protein interactome.