AI Designed the Protein. Does It Stay Folded?

If an AI model gives you a convincing protein design, I’d want to know what happens once someone actually makes it.

Does it have the kind of structure we expected? Does that structure hold up when the temperature changes?

As I learn more about protein characterization, this is the part of AI protein design I keep coming back to. Those experimental questions help decide which designs deserve more time and development.

Start with the secondary structure

Proteins contain structural features such as alpha helices and beta sheets. These are part of what we mean by secondary structure.

Circular dichroism, or CD, measures the difference in how a sample absorbs left- and right-circularly polarized light. In the far-ultraviolet region, the resulting spectrum provides information about protein secondary structure. This methods paper explains how those spectra are interpreted.

For a protein designed around helices, I’d want to see whether its measured spectrum supports that expectation.

That is a useful check, but it does not establish the exact arrangement of atoms or prove that the protein has adopted the precise fold in the model.

How AI was used to build a tougher protein

An eLife study published on April 2, 2026 shows how AI can help build tougher protein scaffolds. The researchers started with a three-helix protein from human spectrin.

The tools had different jobs:

1. RFdiffusion proposed structures with an added fourth helix.

2. ProteinMPNN designed amino-acid sequences for the new structures.

3. ESMFold and AlphaFold2 screened whether those sequences were predicted to fold into the intended shapes.

Physics-based simulations then assessed resistance to heat and pulling forces, helping select three designs for laboratory testing.

The goal was a tougher protein building block for future biomaterials and molecular devices. A possible application is a protein-based material that needs to tolerate stretching or heat during use.

What the researchers measured with Chirascan

The researchers used an Applied Photophysics Chirascan CD system to measure secondary structure and its response to temperature.

The three tested designs showed the characteristic helical CD minima near 208 and 222 nanometres. During temperature-dependent measurements from 20°C to 100°C, all three retained substantial helical structure at the upper end of that range.

For me, this makes the instrument’s role clear: it gives the design team experimental evidence to compare with the model.

Retained helical structure is meaningful evidence of structural resilience under those conditions. It does not establish biological activity at that temperature.

In the current Chirascan range, the V100 and Q100 use a large-area avalanche photodiode (LAAPD) detector. Applied Photophysics reports that its higher light-detection efficiency improves CD sensitivity. That helps separate small spectral differences from measurement noise, particularly when sample or measurement time is limited.

Ask what happens after heating

For a design I wanted to advance, I’d also ask whether its structure recovers after cooling.

Temperature-dependent CD can follow unfolding and refolding. If heating causes a permanent change, I’d interpret that result differently from a protein that recovers its starting spectrum after cooling.

A complementary approach I’d consider is modulated scanning fluorimetry (MSF) on SUPR-DSF. It uses repeated heating and cooling while monitoring the protein’s natural fluorescence to assess recovery after each cycle.

Nicoya’s MSF application note describes a non-reversibility onset temperature, or Tnr, indicating when recovery begins to decline. For designs with a measurable fluorescence response, this makes MSF suited to comparing recovery across candidates or buffer conditions in a plate-based workflow.

That would be an additional experiment for AI-designed proteins; SUPR-DSF was not used in the eLife study. CD follows secondary structure, while MSF follows fluorescence-based recovery.

The practical question is what the protein needs to withstand. A candidate intended for a particular process needs evidence under conditions relevant to that process. A high-temperature result is one part of that decision.

Give the next design cycle useful data

“Garbage in, garbage out” applies to experimental data too.

Automated CD on the Chirascan Q100 offers a practical way to collect consistent spectra from more protein designs. Linking those spectra to sequences, measurement conditions, and repeat results, including designs that fail, would build datasets for training and independent testing of future models.

Experimental stability data already have a role in model development. A 2024 study fine-tuned a protein language model to predict folding stability in small protein domains using proteolysis-derived measurements. CD would contribute different information about secondary structure and its response to temperature.

For me, this is the bigger opportunity: use high-throughput CD to build reliable experimental datasets that could help make the next generation of protein models stronger. Design, measure, learn, and use that evidence to improve the next round of predictions.

---

Desmond Gilmour

Email: [e-mail]

Phone: [phone]

I work at Nicoya Lifesciences. The views expressed here are my own and do not necessarily reflect those of Nicoya. This is a personal publication, not an official company communication.