AI Protein Switches: Designing Dynamic Multistate Molecular Machines

AI has enabled many innovations in protein science: new antivenoms, new nanocages, new ways to sequence proteins, and even more. But creating small, controllable movements remained out of reach; until today! Researchers have created a framework to create multistate proteins, that work just like natural ones! Let’s dive right in.

Programming Protein Switches

Proteins: Misunderstood Dynamic Machines

Proteins are dynamic machines. Many natural proteins switch between conformational states, in response to ligands, mutations, or simply to perform their function. This is the key to signaling, regulation and catalysis. These are subtle changes. A small shift a loop or helix is enough to modify the function of a protein.

And examples are common! Not just that; these proteins are often important therapeutics targets:

• Kinases: A small hinge movement between two lobes and the rearrangement of an activation loop toggles the kinase on and off. Kinases are target of many cancer therapies.

• G protein-coupled receptors (GPCRs): A vast, vast family of proteins. In these receptors, a slight rotation of transmembrane helices converts extracellular signals into an intracellular response. GPCRs are targeted by a whopping 30-40% of all drugs! So, yeah, kind of important.

• Calmodulin: A calcium-binding protein involved in everything from muscle contraction memory. The calcium ion binds to the protein and it causes a change of conformation, allowing calmodulin to wrap around its protein targets and turn on signaling pathways.

There is more, but this is probably enough. These small changes are fundamental for the function of proteins. And often, a single mutation or the binding of a ligand can change a structure from one stable conformation to another.

Why Can’t We Design That?

So, dynamics is central to protein function. At the same time, de novo protein design has focused on static structures. Why? Well, because designing specific and reversible conformation changes is really hard.

• Physics-based models can’t resolve the small energy differences between local conformers

• The powerful deep-learning models often lack interpretability, making it difficult to predict or control dynamic behavior.

This is why most existing switchable systems are based on domain-level movement. This is simpler to design, but it also doesn’t capture the fine dynamic behavior of real signaling proteins!

Designing Fine Tuned Protein Movements

This is where today’s paper comes in. The team here developed a framework to:

• Design dynamic proteins with switchable conformations

• Allow precise control over the switching, using ligands or mutations

• Merge deep learning-based sequence design with physics-based simulation, allowing mechanistic insights into the design!

Their design framework has 3 stages, let’s see them with an example.

Walking Through It: Designing Calcium Binders

Stage 1. Generating Alternative Protein Conformations

The researchers started with a calcium (Ca²⁺) binding protein called 1SMG (catchy name). This will be “state 1” of our dynamic protein.

The team computationally reshaped the loop III, helix C and Ca²⁺-binding site II, creating slight variations in the secondary structure. They then used Rosetta to design sequences optimal for single state stability and filtered them. This left them with around 1000 variants, similar to what happens in natural signaling proteins.

The next step is experimental: the team screened the 11 most promising candidates using yeast surface display, a proxy for expression and stability. From this, they selected #6306 (another catchy name) for its stability and for the presence of a highly distorted Ca²⁺ site, which makes it binding-incompetent, They also solved its structure, which matched AlphaFold2’s prediction almost perfectly, showing the power of AI-based protein tools!

With this, the authors had a validated state 2 for the protein, so they were ready to create a multistate dynamic protein.

Stage 2. Deep Learning-Guided Multistate Design

The next step was to design sequences that could adopt both state 1 and state 2, with fine control over which conformation they preferred.

The team used AlphaFold2 (AF2) to guide their filtering step. They first reversed mutations in #6306 (our state 2) back to the state 1 sequence once at a time, and used AF2 to see if the structure stayed stable. They found and marked as “tolerated” the mutations that didn’t disrupt state 2.

The authors designed new sequences for the region of interest, while keeping the rest constant using ProteinMPNN, a tool that creates protein sequences from protein structures. AF2 was used again to predict the conformational state for each design.

The result was impressive: the designed sequences were similar, but they differed at one key position: residue 89. Mutations in this site favored one state over the other:

• I89 (isoleucine): State

• S89 (serine): State 2.

• R89, N89 (arginine, asparagine): Mixed states.

Residue 89 is far from the calcium-binding site. A single mutation in this residue outside the active site can control conformational state, making it an allosteric site, a very important control mechanism for natural proteins!

Stage 3. Mechanistic Validation: NMR, Simulation and More

The researchers validated how these single mutations regulate switching using NMR and molecular dynamics (MD) simulations.

The NMR showed that there is a switch between I89 and S89 variants, and that the calcium-binding properties confirm the predicted state preferences. I89 binds calcium 10 times tighter than S89!

But the MD simulations were even more interesting. They observed transition between the two states on a low microsecond scale, similar to natural systems. In addition, with calcium bound, the protein stayed in state 1, showing orthosteric stabilization (when the stabilization comes from binding within the region of conformational change). The MD simulations also showed that the two states are stabilized by different hydrophobic interactions and hydrogen bonds, explaining the experimental findings!

To Conclude: Key Innovations and Applications

Great paper! The authors created a modular pipeline, which integrates combinatorial sampling, Rosetta design, deep learning, MD simulations and experimental validation! They rationally programmed conformational dynamics, using both distal mutations and binding-site stabilization.

I also enjoyed that, using NMR and MD simulations, they could figure out the mechanism of the switch. And it’s a (reasonably) efficient framework, not requiring too many experiments and relying on computational tools.

Well, this was super cool! And of course, it has many, many possible applications, in synthetic biology, drug discovery and protein engineering. I had to leave out many things, so go and read the paper here!

If you made it this far, thank you! What do you think of this article? Do you see some interesting applications? Reply and let me know!

More Room:

• Seeing G Quadruplexes: DNA and RNA can fold into G quadruplexes, and there is more and more evidences that they have biological functions. This minireview discusses recent advances in detecting G-quadruplex structures in DNA and RNA using optical probes that rely on fluorescence or phosphorescence lifetime, spectral ratios, or single-molecule detection, rather than intensity alone. It also highlights current limitations and future directions for G quadruplex research.

• Proteins, Patterns and DNA Origami: Antibodies are often used to imprint biological functions to DNA origami, but it’s a bit unclear how well they bind. This study maps antibody binding efficiency on DNA origami using atomic force microscopy, revealing that binding depends on site spacing and position. Different patterns show opposing trends due to electrostatic, steric, and cooperative effects. The findings help optimize DNA-based systems for diagnostics and therapeutics.

• Accelerating Enzyme’s Reactions: DNA nanostructures can also be used to accelerate enzymatic reaction. But how? This study shows that DNA scaffolds accelerate enzymatic reactions with hydrophilic substrates but slow them with hydrophobic ones. The effect is due to a dense water layer near the DNA surface, which alters local substrate concentration. This reveals how biomolecular interfaces can influence reaction rates.

• Know someone who’d love this?
  Pass it on! Sharing is the easiest way to support the newsletter and spark new ideas in your circle.

• Got a tip, paper, or topic you want me to cover?
  I’d love to hear from you! Just reply to this email or reach out on social.