Publication
Proteins are foundational biomolecules of life playing a crucial role in a myriad of biological processes. Their function often requires interplay with other biomolecules, including proteins themselves. Protein-protein interactions (PPIs) are essential for maintaining cell homeostasis, but are also involved in the progression of several diseases, being pathogenic, neuro-degenerative or cancer related. Therefore, PPI engineering has always been at the basis of several protein-based therapeutics and other biotechnology tools. However, most PPI engineering strategies so far relied on extensive experimental optimization or computational tools that depend on prior knowledge. Indeed, challenges remain for protein targets where no structural or experimental data are available, or for interfaces that involve non-protein components such as small molecules.
To explore and address these limitations, this work aims to leverage machine-learning and physics-based methods for the design of de novo protein interactions with therapeutic potential, that will ultimately be characterized and validated with established laboratory techniques.
The first part of this thesis showcases the translational capabilities of PPI designs. For this purpose, we rationally designed switchable protein-based therapeutics by integrating a previously established chemically-disruptable heterodimer (CDH). To optimize this OFF-switch system, we employed in silico methods based on physics-driven predictions, followed by rigorous in vitro validations to enhance its switchability in solution. This resulted in the development of a protein therapeutic exhibiting significantly improved drug-based controllability in mice models.
Nevertheless, most antibody and protein therapeutics discovered using experimental methods are agnostic to where and how these proteins engage their respective target. Despite recent advances, predicting an amino acid sequence that binds to a specific interface remains a major challenge for the field. To address this, a geometric deep learning framework, called MaSIF, was developed in our group to predict PPI interfaces and their corresponding binding partners based solely on the vectorized geometric and chemical features of the protein surface, also known as "fingerprints". In this work, we improved MaSIF by leveraging a database of small binding motifs to design novel protein binders for four therapeutically relevant targets. All protein binders were validated experimentally and reached native-like affinities after pure in silico generation.
Finally, we generalized our framework to design drug-bound protein complexes via the formation of neosurfaces that arise upon small molecule binding. The versatility of our approach allowed us to computationally design and experimentally validate binders against three small molecule-protein complexes. All designs exhibited drug-dependent binding with native-like affinities and were functionalized as ON-switch systems for diffe