In silicoevolution of protein binders with deep learning models for structure prediction and sequence design

Abstract

There has been considerable progress in the development of computational methods for designing protein-protein interactions, but engineering high-affinity binders without extensive screening and maturation remains challenging. Here, we test a protein design pipeline that uses iterative rounds of deep learning (DL)-based structure prediction (AlphaFold2) and sequence optimization (ProteinMPNN) to design autoinhibitory domains (AiDs) for a PD-L1 antagonist. Inspired by recent advances in therapeutic design, we sought to create autoinhibited (or masked) forms of the antagonist that can be conditionally activated by proteases. Twenty-three de novo designed AiDs, varying in length and topology, were fused to the antagonist with a protease sensitive linker, and binding to PD-L1 was tested with and without protease treatment. Nine of the fusion proteins demonstrated conditional binding to PD-L1 and the top performing AiDs were selected for further characterization as single domain proteins. Without any experimental affinity maturation, four of the AiDs bind to the PD-L1 antagonist with equilibrium dissociation constants (KDs) below 150 nM, with the lowest KD equal to 0.9 nM. Our study demonstrates that DL-based protein modeling can be used to rapidly generate high affinity protein binders.

Computational design of highly active de novo enzymes

“…An integral part of Riff-Diff is its backbone refinement protocol, in which the diffused enzyme backbones are refined in an iterative fashion. Recent work showed that the quality of de novo backbone-side chain pairs improves when predicted structures are used for subsequent sequence design (45,46). Therefore, we implemented iterative refinement cycles into Riff-Diff during which the coordinates of idealized helical fragments of the artificial motifs were used as constraints to minimize the motif backbone RMSD.…”

Machine Learning-Guided Protein Engineering

“…One group of such methods uses the ML predictor as a black-box oracle to evaluate existing candidates. This evaluation is then used to approximate the “fitness” of sequences, which is in turn used to navigate the sequence landscape and generate a new set of candidates using tools such as evolutionary algorithms , or simulated annealing . However, approximating complex mutational landscapes using oracles representing estimated, simplifying distributions can harm the optimization process and prevent the optimal solution from being found .…”