Nucleosome Interactome Analysis Using Generative Deep Learning Methods

In recent years, deep learning models have introduced a paradigm shift in structural bioinformatics, yielding significant advancements in protein structure analysis and design. Neural networks are now being used not only to predict protein structure geometry, but also to identify interfaces of interacting proteins and generate de novo molecules. A significant leap in experimental success of AI-modelled structures suggest deep understanding of sequence context and protein geometry by neural networks. However, it still not well understood to which extent generative deep learning models are able to consistently identify hallmark features of protein-protein interactions and are capable of revealing novel hotspot interfaces on the surface of proteins. To explore the capabilities and limitations of AI models in this domain, we conducted a study using nucleosomal complex as a benchmark for profiling protein-protein interactions with peptide binders. Nucleosomes serve as a foundation of genome compaction, epigenetic annotation, and facilitate interactions between chromatin proteins and DNA. highly conserved histone core of the nucleosome presents an ideal target for protein-protein interaction (PPI) profiling. Utilizing a dataset of experimentally resolved nucleosomes interacting with other proteins, we investigated the capacity of state-of-the-art deep learning networks to identify core features of interactions between histones and nucleosome-binding peptides. Our findings demonstrate the ability of generative AI to identify distinct hotspots on the histone core surface and generate binders capable of interacting with it. These results highlight the growing understanding of PPI mechanisms by these models, while also revealing their limitations. Furthermore, our study presents methodological insights that can guide future applications of these models.