2025-06-26 カリフォルニア大学サンフランシスコ校(UCSF)
<関連情報>
- https://www.ucsf.edu/news/2025/06/430256/how-ai-can-help-us-make-new-medicines-faster
- https://www.science.org/doi/10.1126/science.adr7094
ディープラーニングによる動的タンパク質の設計 Deep learning–guided design of dynamic proteins
Amy B. Guo, Deniz Akpinaroglu, Christina A. Stephens, Michael Grabe, […] , and Tanja Kortemme
Science Published:22 May 2025
DOI:https://doi.org/10.1126/science.adr7094
Editor’s summary
Many signaling proteins and enzymes respond to binding of a small molecule or ion by shifting dynamics to favor one structural conformation over another. This behavior is key to biological function, but engineering these properties in designed proteins is very challenging. Guo et al. developed a computational approach to designing such dynamic proteins that can sense and respond to binding of a calcium ion. Starting with a static protein that binds a calcium ion, the authors identified potential alternate conformations and used AlphaFold2 predictions to identify sequences that were compatible with both structures. Validation with molecular dynamics simulations and nuclear magnetic resonance experiments confirmed a dynamic, multistate design that could be shifted into a single conformation upon ion binding. —Michael A. Funk
Structured Abstract
INTRODUCTION
Deep learning has greatly advanced the design of static proteins with new-to-nature (de novo) structures. A clear next challenge is to design the types of tightly controlled, dynamic conformational changes that underpin natural protein functions de novo. For instance, many superfamilies of protein regulators undergo functionally critical changes to the relative orientation of secondary structural elements within a domain (such as helix rotation in kinases to form active sites, kinking of helices in G protein–coupled receptors to expose binding interfaces to downstream signaling partners) and couple diverse effectors to modulate functionally distinct state populations and thereby activity. Yet, these conformational changes have been inaccessible to de novo design, despite their importance.
RATIONALE
Pioneering work to design protein conformational switches has focussed on side-chain rearrangements or large-scale hinge-like domain motions where most of the atomic-level intra-domain interactions are preserved. However, no general method has been described to design the intricate controllable intradomain conformational change mechanisms prevalent in natural regulators de novo, thereby leaving many potential functionalities unexplored. In particular, it is challenging to accurately model the small energetic differences between states at this scale with traditional physics-based models, while “black box” deep learning–based models limit our biophysical understanding of the designed system. As a proof of principle, we reasoned that by combining the high performance and speed of advanced deep learning methods with the interpretability of molecular simulations, we could not only design controllable intradomain modes of motion de novo, but also understand the atomic-interaction networks underlying them.
RESULTS
We describe a general method to design dynamic proteins that uses deep learning to guide the search of sequence and structure space during multistate design. A key element of our approach is restricting the search space by first identifying the minimal set of residues required to define each user-specified conformational state through an in silico mutational scan. Specifically, we assess whether point mutations to increase sequence identity between states perturb computationally predicted structures. Focusing sampling at positions found to be important for determining state preference then allows us to generate sequences with high similarity but diverse state population distributions. This approach not only designs protein switches that exchange between our prespecified states on a functionally relevant timescale (i.e., the low-microsecond regime) but also a mechanism to tune the conformational equilibrium through minimal changes to sequence. The sequence context shared between designs helps to reveal how small sequence differences may alter the conformational ensemble on a biophysical level.
We validate our predictions by extensive experimental dynamics analyses as well as four structures solved by nuclear magnetic resonance spectroscopy, confirming atomic precision of our designs. We find that the amino acid identity at just one sequence position can substantially shift the distribution of states, ranging from highly skewed to roughly equal populations. We also show regulation of the state populations through Ca2+ concentration—a common secondary messenger—by incorporating a Ca2+ binding site within the designed dynamic region, which adopts a binding-competent conformation in only one of the two states. In turn, we demonstrate the allosteric effect on Ca2+ binding affinity by distal mutations that tune the conformational equilibrium. Finally, we find close agreement between our deep learning–based design predictions, physics-based simulations, and experimental data, allowing us to extract and reprogram state-specific distinct atomic interaction networks.
CONCLUSION
Our results validate a general approach to designing proteins with distinct conformational states that can be specified by the designer. As in natural regulation, the equilibrium between the designed conformational states can be controlled both by ligand binding in the active site and by allosteric perturbations (mutations) to distal sites. We show that new modes of intradomain motion, inspired by those key to natural signaling, can now be realized through de novo design. This work provides a foundation for designing programmable signaling systems de novo, facilitating more complex behavior such as de novo signal integration or concerted motions coupled to energy inputs. Moreover, this approach can be applied to engineer non-native motions into natural proteins to control their activities.
Design of dynamic proteins with controllable motions, mimicking mechanisms in natural signaling switches.
We describe a general method that searches sequence and structure space guided by deep learning to design proteins that transition between user-specified structural states (red, blue). A binding site in the dynamic region couples effector binding (Ca2+) to the conformational equilibrium (left), which can in turn also be controlled by allosteric mutations (right). Experimentally determined structures show close agreement with the designed states, and physics-based simulations reveal state-specific networks that explain the observed allostery and predict new mutations that tune the equilibrium.
Abstract
Deep learning has advanced the design of static protein structures, but the controlled conformational changes that are hallmarks of natural 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 intradomain geometries of proteins, similar to switch mechanisms prevalent in nature, with atomic-level precision. We solve four structures that validate the designed conformations, demonstrate modulation of the conformational landscape by orthosteric ligands and allosteric mutations, and show that physics-based simulations are in agreement with deep-learning predictions and experimental data. 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.
