Protein engineering by highly parallel screening of computationally designed variants

Abstract

Current combinatorial selection strategies for protein engineering have been successful at generating binders against a range of targets; however, the combinatorial nature of the libraries and their vast undersampling of sequence space inherently limit these methods due to the difficulty in finely controlling protein properties of the engineered region. Meanwhile, great advances in computational protein design that can address these issues have largely been underutilized. We describe an integrated approach that computationally designs thousands of individual protein binders for high-throughput synthesis and selection to engineer high-affinity binders. We show that a computationally designed library enriches for tight-binding variants by many orders of magnitude as compared to conventional randomization strategies. We thus demonstrate the feasibility of our approach in a proof-of-concept study and successfully obtain low-nanomolar binders using in vitro and in vivo selection systems.

Keywords: Computational protein design; protein engineering; structural biology.

Fig. 1. Structure of the ubiquitin-USP21 complex.

Ubiquitin and USP21 are shown in gray and blue, respectively, with the engineered ubiquitin region shown in pink.

Fig. 2. Schematic of the parallel protein engineering strategy.

Computational protein design is performed on the human ubiquitin interface (positions 54 to 71) binding USP21 [Protein Data Bank (PDB) ID: 3I3T]. Fixed backbone design is performed on protein ensembles, whereas flexible backbone design is performed directly on the initial crystal structure. Unique sequences (2000), 18 amino acids (AA) in length, were extracted for each of the three design strategies and reverse-translated (NT) for synthesis on a DNA microarray. Libraries were constructed by amplifying the DNA microarray product and were subsequently screened by either phage display or Y2H. Deep sequencing of the final screening products recovered the designed ubiquitin variants that tightly bound USP21.

Fig. 3. Sequence logos of ubiquitin variants that tightly bind USP21.

(A to C) Sequence logos of computationally designed ubiquitin variants derived from MD, CONCOORD, and Backrub ensembles for (A) 2000 of the best ranked designed variants, (B) variants selected by phage display, and (C) variants selected by Y2H. (D and E) Sequence logos of ubiquitin variants selected for USP21 binding (D) from a biased naïve library screened by phage display and (E) predicted by random forest models from simulated naïve libraries following the NNK nucleotide randomization scheme. The simulated libraries were biased toward the ubiquitin wild-type nucleic acid composition 70% of the time. The x and y axes correspond to the designed ubiquitin positions and the information content in bits, respectively, as determined by WebLogo (39). Amino acids colored black, green, and blue correspond to hydrophobic, neutral, and hydrophilic residues, respectively.

Fig. 4. PCA of the designed ubiquitin variants.

(A and B) The first two principal components (PCs), capturing 35% of the variance, are formed from a set of 37 unbound ubiquitin and 137 bound ubiquitin crystal structures, which include the wild-type ubiquitin (WT Ub) and a tight-binding ubiquitin variant (Ubv21.4) bound to USP21. (A) Ubiquitin backbone models from the top 2000 scoring variants from each of the MD, CONCOORD, and Backrub ensembles were projected into the first two PCs. (B) MD, CONCOORD, and Backrub structures associated with the 215 ubiquitin variants found to tightly bind USP21 by phage display are projected into the first two PCs. (C and D) Sequences of 215 designed variants identified by phage display and 26 variants from a biased naïve library found to tightly bind USP21 in addition to the wild-type ubiquitin sequence were used for PCA over (C) 7 sequence positions (62, 63, 64, 66, 68, 70, and 71) engineered by the biased naïve library (25% of the variance) or (D) all 18 designed positions (21% of the variance).

Fig. 5. Ubiquitin variant validation.

(A) Dose-response curve of USP21 activity inhibition by wild-type ubiquitin and the Ubv10 ubiquitin design. IC₅₀ concentrations were evaluated as the ubiquitin concentration inhibiting USP21 activity by 50%. The Ubv10 dose-response curve has an IC₅₀ value of 4.2 nM. (B) A strong correlation exists between the IC₅₀ concentrations and the log-transformed deep sequencing counts.

Fig. 6. Microenvironment of the ubiquitin-USP21 interface.

Molecular models for four ubiquitin variants were superimposed on wild-type ubiquitin (PDB ID: 3I3T). (A to D) Designed residues (bold) for (A) Ubv1 (cyan), (B) Ubv2 (magenta), (C) Ubv4 (yellow), and (D) Ubv10 (green) are shown relative to wild-type ubiquitin (pink) as colored sticks. USP21 contact residues are shown as gray sticks. Wild-type ubiquitin Glu⁶⁴ was substituted for Phe, Asn, His, and Asn in Ubv1, Ubv2, Ubv4, and Ubv10, respectively, which removes the unfavorable charge interaction with USP21 Asp⁴³⁸. Ala⁶⁵ substitution for all variants removes the unfavorable charge interaction with Glu³⁹⁵. (E) Thr⁶⁶ was maintained in all variants making a hydrogen bond with Glu³⁹⁵, whereas Phe⁶⁸ facilitates hydrophobic interactions with USP21 residues Leu³⁷⁸, Phe⁴²³, and Leu⁴⁵⁰.