Computational design of conformation-biasing mutations to alter protein functions

Abstract

Most natural proteins alternate between distinct conformational states, each associated with specific functions. Intentional manipulation of conformational equilibria could lead to improved or altered protein properties. Here we develop Conformational Biasing (CB), a rapid and streamlined computational method that utilizes contrastive scoring by inverse folding models to predict variants biased towards desired conformational states. We validated CB across seven diverse deep mutational scanning datasets, successfully predicting variants of K-Ras, SARS-CoV-2 spike, β2 adrenergic receptor, and Src kinase with improved conformation-specific functions including enhanced effector binding or enzymatic activity. Furthermore, applying CB to lipoic acid ligase, a conformation-switching bacterial enzyme that has been used for the development of protein labeling technologies, revealed a previously unknown mechanism for conformational gating of sequence-specificity. Variants biased toward the "open" conformation were highly promiscuous, while "closed" conformation-biased variants were even more specific than wild-type, enhancing the utility of LplA for site-specific protein labeling with fluorophores in living cells. The speed, simplicity, and versatility of CB (available at: https://github.com/alicetinglab/ConformationalBiasing/) suggest that it may be broadly applicable for understanding and engineering protein conformational dynamics, with implications for basic research, biotechnology, and medicine.

Figure 1.. Conformational biasing (CB) approach for predicting mutations that shift conformational equilibria.

(A) Many proteins switch between alternative conformations. Specific conformations may exhibit unique binding, catalytic, and functional properties. (B) CB predicts mutations that shift conformational equilibria. (C) Computational workflow for CB. Mutant sequences for a protein of interest are paired with the protein’s structure in two or more conformational states. The structures can be experimentally-solved or computationally-predicted. The sequence-structure pairings are scored using an inverse folding model such as ProteinMPNN and plotted as shown. Conformation-biasing mutations maximize the difference in conformation-specific inverse folding scores.

Figure 2.. CB predicts protein variants with altered effector binding, receptor activation, and enzymatic activity.

(A) The G protein K-Ras switches between an inactive State 1 conformation (PDB: 8T71, bound to GDP) and an active State 2 conformation (PDB: 6XHB, bound to GTP-analogue). State 2 binds to the effectors listed. RMSD, root mean square deviation. (B) 3320 single mutants and 26761 designed K-Ras double mutants were scored using CB against State 1 and State 2 experimental structures in (A). Points are colored according to experimental binding data from Weng et al.(24), for RALGDS, a State 2-specific effector (left) or DARPin K27, a State 1-specific binder (right). (C) Violin plot showing distributions of experimentally-determined binding scores from Weng et al.(24) for top CB-predicted State 1-biased variants (blue) and State 2-biased variants (red). Mutations at the KRAS-effector binding interface that are likely to directly influence effector binding are excluded from analysis (Figure S1A). ****p < 0.0001. Additional K-Ras analysis in Supporting Text and Figure S1. (D) SARS-CoV-2 spike protein with ACE2 receptor binding domain (RBD) in “down” (PDB: 7XIX) versus ACE2-binding competent “up” (PDB: 7XO8) conformations. (E) Distributions of raw ACE2 binding scores (left) and expression-normalized ACE2 binding scores (right), for CB-predicted up-biased and down-biased S1 variants. **p<0.01, ***p < 0.001, ****p<0.0001. CB scatter plot in Figure S3A. (F) Same as (E) except groups are based on single conformation scoring using ProteinMPNN instead of CB. ****p<0.0001. (G) SARS-CoV-2 RBD variants were scored using CB against RBD up/down structures. Points are colored by experimentally-determined variant expression levels from Starr et al.(25) (H) Human β2 adrenergic receptor (β2AR) in inactive and active conformations (PDB: 3SN6 and 2RH1, respectively). The conformational change in transmembrane helix 6 (red) enables the active form to bind to G proteins and arrestin. Positions of five CB-predicted active-biased mutations that are corroborated by experimental data from Jones et al (30) (Z > 1)are labeled. (I) 5263 β2AR variants were scored using CB, and top active-biased, inactive-biased, and neutral variants are plotted by their experimentally-determined receptor basal activation scores (30). ****p < 0.0001. (J) Src kinase repositions its auto-inhibitory SH2/SH3 domains upon autophosphorylation, becoming active (PDB: 2SRC and 1Y57, respectively). AS, active site. (K) Distributions of Src kinase activities for inactive-biased, active-biased, and neutral variant populations. **p < 0.01, ****p < 0.0001. CB scatter plot for Src in Figure S3B.

Figure 3.. CB biases E. coli LplA towards open or closed adenylate ester-bound conformations.

(A) Two half-reactions catalyzed by LplA: adenylation of lipoic acid by ATP to generate lipoyl-AMP, and transfer of lipoyl onto lysine sidechain of an acceptor protein. LplA switches from a closed to open conformation between the first and second half reactions. Ad, adenosine. CTD, C-terminal domain of LplA. NTD, N-terminal domain. Structures from PBD: 1X2H, 3A7R, and 3A7A (left to right). Lipoic acid, lipoyl-AMP, and octyl-AMP shown as red spheres (left to right). (B) CB on LplA, using open (3A7R) and closed (1X2G) backbone structures. Closed-biased, neutral, and open-biased variants selected for SAXS (bold) and Trp fluorescence analysis are labeled. (C) Structures of BCN, lipoyl-AMP, and non-hydrolyzable analog lipoyl-AMS. (D) SEC-SAXS analysis of four biased variants (on W37V background, indicated by V superscript), in apo state or in complex with lipoyl-AMS. Fractional occupancy predicted using Oligomer (see Methods and Figure S5), from 2–5 independent measurements per variant. Errors, ±1 std. dev. (E) DENSS ab initio modeling of protein envelopes, fit with major LplA conformer detected under +lipoyl-AMS condition. A48N appears as a dimer at the high protein concentrations required for SAXS (model generated by AlphaFold3). (F) Fold-change in tryptophan (Trp) fluorescence for purified LplA variants on addition of lipoyl-AMS. Open and closed structures of LplA show positions of Trp sidechains in light blue. n=3, p < 0.01. (G) Correlation of Trp fluorescence data from (F) to fractional occupancy predicted by SAXS Oligomer analysis in (D). (H) Distribution of top open-biasing (red) and closed-biasing (blue) mutations in LplA structure. Figure S12 expands on the mechanisms by which these mutations may introduce conformational bias.

Figure 4.. Conformationally-biased LplA variants show large differences in promiscuous protein labeling.

(A) Flow cytometry assay for measuring the promiscuous labeling activity of mCherry-tagged LplA variants. After conjugation to endogenous proteins, BCN is derivatized with fluorogenic methyltetrazine(MTz)-BODIPY. See Supplementary Text and Figure S7 for further discussion of this assay. (B) CB plot for LplA, with closed-biased, neutral, and open-biased mutants colored by their experimentally-measured promiscuous activities in (C). (C) Promiscuous activity histograms for closed, neutral, and open-biased LplA variants (all variants made on W37V background except WT). 2D BODIPY vs. mCherry flow cytometry plots for all variants shown in Figure S8. This experiment was performed 2 times with similar results (correlation between replicates shown in Figure S6D). (D) Mean promiscuous activity scores for each category of CB-designed LplA variant in (C). * p<0.05, **** p<0.0001. (E) Correlation between promiscuous activity from (C) and Trp fluorescence change measured in Figure 3F. Y139V (red cross) is an unstable variant that shows strong aggregation in HEK293T cells. (F) Ratio of specific to promiscuous activity for 12 purified LplA variants, based on in-cell promiscuous activity measurements in (C) and E2p specific labeling assay in Figure S6F–G. Errors, 1 std. dev. *p < 0.05. (G) Proposed mechanistic model for how conformational occupancy (K_conf) tunes substrate specificity in LplA. In lipoyl-AMP-bound LplA, the C-terminal domain (CTD) may act as a tethered competitive inhibitor, increasing K_M-Obs for protein substrates when K_conf (equilibrium constant between closed and open conformations) is low. Modeling of on-target and off-target labeling velocities shown in Figures S6J–K. (H) LplA variants can be used for either site-specific (top) or promiscuous (bottom) protein labeling in cells. POI, protein of interest. LAP is a 13-amino acid engineered LplA Acceptor Peptide(47). (I) Confocal microscopy of LAP fusion proteins tagged with coumarin or BODIPY fluorophores, catalyzed by W37V LplA (top) or closed-biased variant T57IV (bottom). For BODIPY labeling (left), HEK293T cells were treated with BCN for 10 minutes, then clicked with BODIPY-MTz, and imaged live. For coumarin (right), HeLa cells were labeled for 10 minutes with coumarin-AM2(22), then fixed and stained with anti-HA antibody to visualize actin-LAP. Scale bars, 10 um. (J) Ǫuantification of images in (I), analyzing colocalization between BODIPY and BFP (left) or coumarin and HA (right). 4 FOV per condition. ** p < 0.005. (K) In-gel fluorescence of cell lysates labeled as in (I). (L) The open-biased LplA variants A48N and N83V catalyze promiscuous labeling of endogenous cytosolic proteins in HEK293T cells. Cells were treated with BCN for 10 minutes before lysis, Click with mTz-BODIPY, and SDS-PAGE. No LAP fusion was expressed in these samples.

Figure 5.. Benchmarking CB.

CB bias scores for LplA plotted against experimental data from (A) SEC-SAXS, (B) Trp fluorescence, and (C) promiscuous labeling activity measurements. Variants are colored by CB-prediction: closed-biased (blue), neutral (grey), and open-biased (red). Variants marked with an X are likely to have a steric clash with BCN substrate in the active site (AS), explaining their low activities (Figure S6I). *r = spearman correlation calculated without steric clash variants, r = spearman correlation calculated for all variants. (D) Correlation between LplA promiscuous activities and CB scores, determined using ProteinMPNN, ESM-IF1, Frame2Seq, or ThermoMPNN. (E) CB runtime on a single RTX 4090 GPU using different inverse folding models, in minutes. Data was processed using consistent parameters. (F) K-Ras analysis performed using single structure scoring. *p<0.05, ****p<0.0001. Compare to Figure 2C. (G) Comparison of CB vs. single structure scoring using ProteinMPNN, for LplA, β2AR, Src, and Braf datasets. **p<0.01, ****p<0.0001. (H) Correlation between AFCluster-determined bias scores (difference in open vs. closed structural similarity score) and LplA promiscuous labeling activities. (I) BioEmu sampling (n=500) of 12 CB-predicted open-biased variants (red) and 12 CB-predicted closed-biased variants (blue). Samples were aligned to LplA open and closed structures to estimate fractional open/closed occupancy, which is plotted. Green points are combination mutants, in which a closed-biasing mutation is combined with an open-biasing mutation in the same protein. These were simulated (n=100) using BioEmu and plotted in the same manner. (J) Correlation between BioEmu predictions and LplA Trp fluorescence data for 24 variants. Red points are predicted by CB to be open-biased, and blue points are predicted to be closed-biased. For comparison, CB vs. Trp fluorescence correlation plot is in (B).