https://www.science.org/doi/10.1126/science.adr7094

Deep learning–guided design of dynamic proteins

Key figures

• Figure 1: Explains the full design workflow, from de novo generation of alternative states through deep learning–guided multistate design to mutation-based tuning of the conformational equilibrium.
  
• Figure 2: Shows that a family of nearly identical sequences differing at one allosteric position can shift population between two designed states and exchange on a low-microsecond timescale.
  
• Figure 4: Reveals state-specific interaction networks from MD and uses them to prospectively predict additional mutations that bias the switch toward one state.
  

1) Thesis (one sentence)

To address the gap of de novo designing controllable intradomain conformational switching, in an engineered EF-hand calcium-binding protein system, deep learning–guided multistate design causes tunable two-state dynamic proteins by restricting design to state-determining residues and coupling orthosteric and allosteric perturbations to distinct atomic interaction networks, supported by AF2 prediction, NMR structures and dynamics, calcium-binding assays, and molecular dynamics evidence.

2) Evidence card (three bullets only)

• Strongest result: (Fig. 2, Fig. 3, Fig. 4) The authors designed sequences that differ only at residue 89 yet shift the equilibrium between two intended conformations, validated by NMR structures for S89 and Ca2+-bound I89, by mixed-state behavior for apo I89, by an overall ~10-fold difference in Ca2+ affinity between I89 and S89, and by MD-supported exchange on the low-microsecond timescale.
  
• Method enabler: (Fig. 1; research type + tools) The key enabling method was a three-stage pipeline combining LUCS de novo state generation, Rosetta single-state design and filtering, AF2-guided in silico mutational scanning to shrink the multistate design space, position-tied ProteinMPNN sequence generation, and downstream validation by yeast display, NMR, Ca2+ titration, MD, mutual information analysis, Markov state modeling, and Frame2seq scoring.
  
• Critical limitation: (Fig. 2D, Discussion) The designed conformational landscape and exchange timescale were not directly specified during design, and even the best apo I89 NMR analysis did not converge to a single reshaped-region structure, indicating residual ensemble complexity beyond a fully deterministic two-state output.
  

Optional

Quote bank (2–4 short excerpts)

• Quote 1: “the amino acid identity at just one sequence position can substantially shift the distribution of states” (Research Article Summary, page 1)
  
• Quote 2: “new modes of motion can now be realized through de novo design” (Research Article Summary, page 1)
  
• Quote 3: “the final set of multistate designable residues included these positions and their neighbors” (Multistate design of dynamic proteins, page 3)
  

Key comparisons (1–3 lines)

• Compared to: Earlier de novo switches based mainly on hinge-like rigid-body motions, domain rearrangements, fold switches, or side-chain dynamics.
  
• Win: Achieves local intradomain secondary-structure reorientation with experimentally validated, tunable two-state equilibria controlled by both ligand binding and distal mutations.
  
• Tradeoff: Requires substantial iterative computational filtering and still leaves transition-state barriers and full conformational landscapes only indirectly characterized rather than explicitly designed.
  

Methods I might copy (protocol hooks)

• Construct design / Models: Starting state was troponin C variant 1SMG with E41A; the reshaped region was residues 49 to 72 in 1SMG numbering; LUCS generated alternative conformations and Rosetta designed ~20 single-state sequences per backbone; AF2 mutational scanning reduced multistate designable positions from 37 to 25; ProteinMPNN generated 10^4 multistate sequences.
  
• Conditions / Instruments: Yeast display used EBY100 with c-Myc and His-tag readouts; bacterial expression used BL21(DE3), induction at OD600 0.6 to 0.8 with 300 μM IPTG at 37°C overnight; NMR structures used ~400 μM uniformly labeled protein in 100 mM KCl pH 6.7 with 5% D2O, measured mainly at 298.1 K on Bruker 800 and 600 MHz instruments; temperature-series HSQC was 5° to 35°C in 5°C increments; R1ρ used spin-lock fields of 1.0, 1.25, 2.0, 2.5, and 3.0 kHz; CD used ~2 μM protein in 10 mM KCl pH 6.7 with melting at 208 nm from 25° to 95°C at 1°C/min.
  
• Readout / Analysis: Structural validation used ARTINA, TALOS, XPLOR-NIH, and PDB validation; Ca2+ affinity was measured by HSQC titration from 0.025 to 51.2 mM CaCl2 and fit with a two-site apparent K_d model; MD used GROMACS 2022.5 with a99SB-disp, triplicate 2-μs runs plus additional sampling, followed by mutual information analysis and five-state Markov state modeling.
  

Open questions / Theoretical implications (2–5 bullets)

• How general is the observation that a single distal residue can strongly bias a dynamic equilibrium once the major state-determining positions are correctly set?
  
• How often will a designed dynamic system require explicit negative design against off-target folded states or partially ordered intermediates that AF2 and NMR reveal only after the fact?