Large-scale exploration of protein space by automated NMR

This paper presents a scalable, automated pipeline that integrates protein design with high-throughput NMR spectroscopy to experimentally characterize the structure and dynamics of hundreds of de novo designed proteins, revealing conformational heterogeneity not captured by current computational models and establishing a foundation for data-driven statistical structural biology.

The Gist

Imagine you are an architect who has just invented a revolutionary computer program. This program can design thousands of unique, beautiful houses (proteins) in seconds. You can see the blueprints, the 3D models, and even predict how strong the walls will be.

But here's the problem: You've never actually built any of these houses. You don't know if the doors actually open, if the floors creak, or if the house sways in the wind. You only have the static blueprints.

This is exactly the situation scientists have been in with proteins (the tiny machines that make life work). Thanks to AI, we can now "design" millions of protein shapes on a computer. But we lack the tools to actually build them and test how they move and behave in the real world.

This paper introduces a solution: A high-speed factory that builds and tests these protein designs automatically.

Here is the story of how they did it, broken down into simple steps:

1. The Blueprint Factory (Designing the Proteins)

First, the team used two different AI "architects" (called RFdiffusion and Proteína) to generate 17,500 different protein blueprints.

• The Challenge: If you just pick the "best" looking blueprints, you mostly get simple, boring shapes (like straight tubes).
• The Fix: They used a smart sorting system to make sure they picked a diverse mix: some tube-like, some flat sheets, and some complex knots. They selected 384 unique designs to test, ensuring they covered a wide variety of shapes.

2. The Assembly Line (Making the Proteins)

Usually, building a protein takes a scientist weeks of manual work, like a craftsman hand-carving a statue. This is too slow to test 384 designs.

• The Innovation: They built a robotic assembly line.
  • The Parts: They ordered the DNA instructions (the blueprints) as synthetic code.
  • The Workers: They put these instructions into bacteria (E. coli), which act like tiny factories.
  • The Automation: A robot pipette moved liquids between 96 tiny wells at once. The bacteria grew the proteins in a special nutrient soup that made them glow with a radioactive tag (isotopes), which is necessary for the next step.
  • The Result: In just a few days, they purified hundreds of proteins. It was like going from hand-carving one statue a month to a factory churning out 300 statues a week.

3. The "X-Ray" Scanner (The NMR Machine)

Now they had the proteins, but how do you see if they work? They used a machine called NMR (Nuclear Magnetic Resonance).

• The Analogy: Think of NMR as a super-advanced MRI scanner for single molecules. While a regular MRI takes a picture of your whole body, NMR takes a "fingerprint" of a single protein, showing exactly where every atom is and how it's wiggling.
• The Speed: Usually, scanning one protein takes days. Here, they set the machine to scan a protein for just 45 minutes.
• The Outcome: They scanned all 379 successful proteins.
  • 62% (239 proteins) gave clear, beautiful fingerprints.
  • The rest were either too weak or clumped together (like a house that collapsed before the inspection).

4. The Discovery (What Did They Learn?)

This is where the magic happened. They compared the "real" protein fingerprints to the "computer" blueprints.

• The Good News: Most of the proteins built exactly what the computer designed. The AI architects were right about the shape!
• The Big Surprise: The computer blueprints were static (frozen in time), but the real proteins were alive.
  • The NMR scans showed that many proteins had parts that were wiggling, flopping, or changing shape.
  • The AI models, which are trained on pictures of frozen proteins, completely missed this movement. They couldn't predict that a specific loop of the protein would be floppy or that the protein would have a "second personality" (a second shape it occasionally switches to).

Why Does This Matter?

Think of it like this: For years, we've been trying to understand how a car works by looking at a 2D drawing of it. We know where the wheels go, but we don't know if the engine vibrates, if the suspension bounces, or if the doors stick in the rain.

This paper proves that we can now build and test hundreds of cars in a week.

• For Science: It opens the door to "Statistical Structural Biology." Instead of studying one protein for five years, we can study thousands to find patterns.
• For the Future: By feeding this massive amount of "movement data" back into AI, we can teach computers to not just design the shape of a protein, but to design how it moves and feels. This is crucial for designing new medicines that fit perfectly into the moving parts of our bodies.

In short: They built a robot factory that turns computer protein designs into real, testable molecules, revealing that nature is much more dynamic and "wiggly" than our current computer models ever imagined.

Technical Summary

1. Problem Statement

While recent advances in deep learning (e.g., AlphaFold, RFdiffusion) have revolutionized in silico protein structure prediction and design, there remains a critical bottleneck in experimental validation, specifically regarding protein dynamics and conformational heterogeneity.

• The Gap: Current computational models primarily map sequences to static structures. They struggle to predict energy landscapes, local flexibility, and multi-state behaviors.
• The Limitation: Nuclear Magnetic Resonance (NMR) spectroscopy is the gold standard for characterizing atomic-resolution dynamics but has historically been a low-throughput technique, typically limited to single proteins or small sets. This prevents the generation of the large-scale, standardized datasets required to train deep learning models on protein dynamics.
• The Goal: To establish a scalable experimental pipeline that bridges the gap between high-throughput protein design and atomic-resolution dynamic characterization.

2. Methodology: The NMR-APP Pipeline

The authors developed a fully automated platform termed NMR-Automated Protein Production (NMR-APP). This pipeline integrates generative design, robotic liquid handling, and automated NMR acquisition.

A. Protein Design & Library Generation

• Generative Models: Two distinct backbone generation models were used to ensure structural diversity:
  • RFdiffusion: A well-validated model known for producing helical structures.
  • Proteína: A newer flow-matching model capable of generating more diverse folds, including $\beta$-sheets.
• Sequence Design: Sequences were designed using ProteinMPNN (temperature 0.1) for 17,500 backbones per model.
• Filtering & Selection:
  • Sequences were refolded in silico using Boltz-1.
  • Designs were filtered based on self-consistency (pTM > 0.7, C$\alpha$ RMSD < 2 Å).
  • Clustering: To avoid bias toward simple helical structures, designs were clustered using Foldseek. The final library of 384 proteins was sampled to ensure broad coverage of structural space (including $\alpha$-helices, $\beta$-sheets, and mixed folds).

B. High-Throughput Production

• Cloning: Automated via Golden Gate Assembly using an acoustic liquid handler (Echo 525).
• Expression: Proteins were expressed in E. coli (BL21(DE3)) in 96-deepwell plates using auto-induction with $^{15}$N-isotopic labeling.
• Purification: A two-step automated process:
  1. Chemical lysis followed by IMAC (Ni-NTA) purification in 96-well plates.
  2. Automated semi-preparative Size-Exclusion Chromatography (SEC) to isolate monomers.
• Yield: A median yield of 115 $\mu$g per sample was achieved. Only 1.3% of samples failed during production.

C. Automated NMR Screening

• Acquisition: A standardized 2D [$^{15}$N, $^1$H]-SOFAST HMQC experiment was recorded for 45 minutes per sample at 25°C on a 600 MHz cryoprobe.
• Automation: Once loaded, data acquisition and sample handling were fully automated, requiring no manual intervention.
• Quality Control: Automated scripts analyzed peak counts, intensity distribution, and signal-to-noise ratios to classify spectra quality.

3. Key Results

A. Throughput and Success Rates

• Scale: The pipeline successfully processed 384 de novo designed proteins.
• Success Rate:
  • 98.7% (379/384) of designs were successfully expressed and purified.
  • 62% (239/379) yielded high-quality NMR spectra suitable for detailed analysis.
  • 21% yielded low-quality spectra (detectable but noisy), and 16% yielded no interpretable data (likely due to aggregation or low expression).
• Efficiency: A single operator can produce and analyze hundreds of proteins per week. The throughput is limited by spectrometer time (approx. 32 samples/day), not personnel.
• Cost: The cost per sample is approximately $25 USD, dominated by synthetic DNA costs.

B. Structural Validation

• Fold Confirmation: For 9 selected proteins, full backbone assignment and structure determination were performed.
  • RMSD: The experimentally determined NMR structure of a representative protein showed an RMSD of 1.3 Å to the designed model.
  • Secondary Structure: Chemical shift analysis confirmed that designed proteins adopted their intended secondary structures (helices, sheets, coils) with high fidelity.

C. Dynamics and Multi-State Behavior

• Unexpected Dynamics: Despite being designed as static entities, many proteins exhibited local backbone dynamics (loops, flexible segments) detectable via NMR relaxation experiments ($T_1$, $T_2$, NOE).
• Multi-State Populations: Several designs showed minor conformational states (populations as low as 5%), often attributed to cis/trans proline isomerization.
• Computational Gap: There was no correlation ($R^2 < 0.12$) between computational metrics (e.g., pLDDT, RMSF from AlphaFold/Boltz) and experimental NMR relaxation parameters. This indicates current generative models, trained on static structures, fail to predict residue-specific dynamics in designed proteins.

D. Statistical Insights

• Spectral Clustering: Unsupervised clustering of HMQC spectra successfully separated proteins by fold class (e.g., $\alpha$-helical vs. mixed), suggesting machine learning can extract structural patterns from unassigned spectra.
• Dynamics Metrics: The Coefficient of Variation (CV) of peak intensities emerged as a robust, concentration-independent metric for detecting conformational heterogeneity.

4. Significance and Impact

• Paradigm Shift: The paper establishes a new regime of "Statistical Structural Biology." Instead of studying single proteins in isolation, researchers can now study ensembles of designed proteins to derive statistical laws governing sequence-structure-dynamics relationships.
• Data for AI: The pipeline generates the large-scale, standardized datasets necessary to train next-generation deep learning models that can predict protein dynamics and energy landscapes from sequence alone, independent of evolutionary information.
• Scalability: By decoupling protein production from manual labor, the authors demonstrate that NMR can be scaled to match the throughput of modern computational design, turning NMR from a bottleneck into a high-throughput discovery engine.
• Future Applications: The platform is extensible to larger proteins (via methyl labeling), natural proteins, and complexes, and can be used to study intrinsically disordered proteins (IDPs), allostery, and functional heterogeneity.

Conclusion

This work successfully demonstrates that combining generative AI design with automated robotics and NMR spectroscopy creates a powerful feedback loop. It validates that de novo designed proteins fold correctly but reveals that their dynamic behaviors are not yet predictable by current static models. The NMR-APP platform provides the essential experimental infrastructure to close this gap, enabling a data-driven understanding of protein motion and function.