Targeting arbitrary regions of intrinsically disordered proteins

Abstract

A general, robust approach to design proteins that bind tightly and specifically to intrinsically disordered regions (IDRs) of proteins and flexible peptides with minimal cost would have wide applications in biological research, therapeutics, and diagnosis. However, the lack of defined structures and the high variability in sequence and conformational preferences has complicated such efforts. Herein, we have built layers of work to solve this problem with two main computational methods, i.e., bottom-up and top-down. As for the bottom-up approach, we built a landscape to first develop components of geometric matching and amino acid sidechain-oriented pocket recognition for regular peptides; we then generalized it to the broad non-regular landscapes combined with deep learning tools under a rule of optimal binding geometric constraints, achieving specific recognition toward arbitrary unstructured protein sequence space. As for the top-down approach, we leveraged the power of deep learning, trained neural networks to predict and co-fold a disordered target and a designed binding protein to it all together. Using these computational methods, we have designed binders to more than 50 broadly diverse unstructured targets, including highly polar targets. Experimental testing of dozens to hundreds of designs per target yielded binders with affinities better than 100 nM in most cases, and in the pM range straight out of the computer in five cases. Co-crystal structures of designed binder-target complexes as well as NMR structures with isotope labeled peptide targets were closely consistent with the design models. All-by-all in vitro binding crosstalk experiments for representative designs binding diverse targets show they are highly specific for the intended targets, with no crosstalk even for the closely related peptides. Designs were shown functional in a number of downstream assays indicating the therapeutic, diagnosis, intracellular tracking potentials. These methods were applied in the biologically relevant cancer target RAS to distinguish the four distinct isoforms in cells to the degree antibodies have never achieved. Our approach thus could provide a general solution to the intrinsically disordered protein and peptide recognition problem, while paving a road to site-specific recognition of post-translational modifications (PTMs) and enzymatic functional designs as well.