Deep learning guided design of dynamic proteins

Abstract

Deep learning has greatly advanced design of highly stable static protein structures, but the controlled conformational dynamics that are hallmarks of natural switch-like signaling proteins have remained inaccessible to de novo design. Here, we describe a general deep-learning-guided approach for de novo design of dynamic changes between intra-domain geometries of proteins, similar to switch mechanisms prevalent in nature, with atom-level precision. We solve 4 structures validating the designed conformations, show microsecond transitions between them, and demonstrate that the conformational landscape can be modulated by orthosteric ligands and allosteric mutations. Physics-based simulations are in remarkable agreement with deep-learning predictions and experimental data, reveal distinct state-dependent residue interaction networks, and predict mutations that tune the designed conformational landscape. Our approach demonstrates that new modes of motion can now be realized through de novo design and provides a framework for constructing biology-inspired, tunable and controllable protein signaling behavior de novo.

Fig. 1.. Generalizable approach for the deep learning guided design of dynamic proteins.

(A) Schematic of design goal to engineer dynamic proteins in a two-state equilibrium that can be controlled by orthosteric ligands and allosteric perturbations. (B) Main stages of the approach: (1) De novo generation of alternative states that differ in their geometry (light blue region) using systematic conformational sampling (Fig. S2), followed by in silico and experimental validation of single state designs. (2) Deep learning (DL) guided sequence/structure search to focus sampling during multi-state design (MSD) at key positions for determining state preference and their neighbors. (3) Combination of physics-based molecular dynamics (MD) simulations, mutual information analysis (MutInf), and deep learning models to determine state-specific residue interaction networks and identify mutations capable of modulating the conformational landscape, followed by experimental validation. (C-D) Application to generate two designable states with distinct conformations coupled to ligand binding (Ca²⁺). (C) (Top row, light blue) Binding-competent state 1 structure with Ca²⁺ binding site (inset) (PDB ID: 1SMG) shown in two orientations. (Bottom row, teal) De novo generated alternative (binding-incompetent) state 2 model shown in two orientations. To couple Ca²⁺ binding to the designed conformational change, the Ca²⁺ binding site is significantly reshaped in state 2 to disfavor Ca²⁺ binding (inset). (D) Overlay of the NMR structure of a state 2 single-state design (teal) with its AF2 prediction (grey) shows excellent agreement (Cα RMSD = 0.98Å, excluding loops).

Fig. 2.. Two-state equilibrium in fast exchange shifted by allosteric mutations.

(A) The protocol in Fig. 1B predicted a family of sequences differing only at position 89 (X), where the amino acid identity at position X determined whether the 5 AF2 predicted models were entirely in state 2 (left), mixed (middle), or entirely in state 1 (right). Depicted AF2 models (grey cartoons, with reshaped region colored by AF2 pLDDT) are for the underlined amino acid at position X (shown as sticks). Small polar residues favored state 2 by hydrogen-bonding with the backbone of loop III (bottom, left), while bulky and/or hydrophobic residues favored state 1 by pushing loop III away from the central helix (bottom, right). (B) ¹H,¹⁵N-HSQC spectra of S89, N89, and I89, with several well-resolved peaks (shaded ovals) showing chemical shift changes consistent with a two-state equilibrium in fast exchange between state 1 preferred (I89), state 2 preferred (S89) and intermediate (N89). Inset shows ¹Hn chemical shift changes between I89 and S89 colored on the AF2 model of I89, consistent with the designed conformational change in the reshaped region (the reshaped region is circled by dashed line in panels B-E). (C) Agreement between the NMR structure of S89 (red) and its AF2 prediction (grey) (Cα RMSD = 1.31Å excluding loops). Inset shows the hydrogen bond formed between S89 and loop III in the NMR structure consistent with AF2 predictions in (A). (D) NMR models for I89 were consistent with sampling both state 1 (blue, left) (Cα RMSD = 1.67Å excluding loops) and state 2 (pink, right) (Cα RMSD = 1.31Å excluding loops), with key residue ILE 69 buried in proposal #1 (state 1) and solvent-exposed in proposal #2 (state 2). (E) ¹⁵N near-resonance R_1ρ relaxation rates for design I89 plotted per residue (top) and visualized on the AF2 structure of design I89 (bottom) indicate low microsecond exchange in the regions predicted to undergo significant conformational changes (median predicted change in Cβ-Cβ distances between states shown on the x axis colored by magnitude). Residue numbering of all designs includes the N-terminal thrombin cleavage site scar (grey in (A)).

Fig. 3:. Modulation of the conformational landscape by ligand binding.

(A) Changes in I89 ¹Hn chemical shifts upon adding 10eq of Ca²⁺ visualized on the AF2 model, showing significant chemical shift perturbations distal to the Ca²⁺ binding sites, particularly in the reshaped region and in neighboring residues. (B) ¹H,¹⁵N-HSQC spectra of S89, N89, and I89 with 10eq of Ca²⁺ show ligand-dependent chemical shift changes for residues in the reshaped region consistent with Ca²⁺ shifting the population distribution toward state 1 (arrows). (C) The Ca²⁺ bound NMR structure of I89 is excellent agreement with its AF2 prediction (Cα RMSD = 1.34Å, excluding loops), and the observed binding site backbone is consistent with the known EF hand binding motif with modeled-in Ca²⁺ (right).

Fig. 4:. Physics-based simulations reveal molecular mechanisms underlying switch behavior.

(A) Ca²⁺-free 1μs molecular dynamics (MD) trajectory of I89 showing the expected transitions between a binding-competent state 1 and binding-incompetent state 2 conformation (cyan helix, top) with a timescale of exchange of approximately 0.5μs. The protein remains in state 1 in the presence of Ca²⁺ (bottom). C⍰ RMSD is measured for the reshaped helix compared to the starting coordinates and coordinates Ca²⁺ consistent with an EF hand (inset). (B) Mutual informational analysis (heatmap) revealing a correlated network of residues (light blue) connecting the Ca²⁺ binding loop (green) with the allosteric mutation site (dark blue), shown in two orientations (top panels, reshaped helix circled by a dashed line). Panels on the right show interaction details of a key residue, ILE 69 (red, highlighted by dashed box in heatmap), linking the two regions. ILE 69 becomes solvent exposed in state 2, enabling the formation of distinct contacts by TYR 43. (C) State-specific interactions for state 1 and state 2 (colored) observed during a 1μs Ca²⁺-free MD simulation for I89. (D-F) In silico and experimental validation of state-specific interaction networks using mutations. (D) Frame2Seq (F2S) predictions for Y64F and K68E; score refers to the negative log-likelihood difference between the mutated and original sequence, where the negative change in score predicts the mutations disfavor state 2. (E) AF2 predictions showing higher pLDDT (greater confidence) for state 1 for mutants on the right compared to the original sequence on the left. (F) ¹H,¹⁵N-HSQC NMR data showing chemical shift changes consistent with an increased state 1 population in the mutants (arrows denoting direction of shifts from state 2 (S89) towards a larger population of state 1 (I89)).