Designed to bind

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

The de novo design of proteins capable of binding small molecules to facilitate drug delivery is challenging as it requires the simultaneous optimization of protein sequence and structure, along with ligand conformation in the protein. Now, Fry et al. report an optimization algorithm termed neural iterative selection–expansion (NISE) that overcomes this issue. The strategy relies on two neural networks, one to design protein sequences based on an input backbone, and another to predict the structural complex of input protein and desired ligand. This approach enables the design and refining of protein sequences, simulation of protein–ligand structures and ligand conformations by comparing them to the previous input.

The team designed protein binders for exatecan, a clinically approved anticancer drug. After generating single-chain four-helix bundle protein scaffolds and computing ligand conformers, these structures were used as initial inputs for NISE. The algorithm produced a series of candidate designs, from which four were selected based on the extent of ligand burial and the ligand–protein interactions. Protein expression enabled the determination of dissociation constants, with three pairs presenting Kd values <10 μM; the highest-affinity design (Kd of 120 nM) was renamed the exatecan–protein interaction construct (EPIC). This construct was then improved using the sequence-design neural network, leading to a double-mutant protein, EPIC(Q51N/M97L), with a Kd of 1.2 nM. Comparison of crystal structures of exatecan bound to EPIC or the single-mutant EPIC(Q51N) showed that the higher-affinity variant provided a more buried environment for the ligand. The team also selected the mixed α/β NTF2 fold as starting input for NISE and designed an apixaban (an anticoagulant) binder, obtaining apixaban-binding protein exemplar (APEX) with a Kd of 80 pM.