Regional sequence and embedding divergence among five grass OsNAC25-like proteins

Imagine you are a detective trying to find a long-lost cousin in a massive, bustling family reunion. You have a photo of your own relative, but the family tree is so huge and the cousins so many that you need a clever way to spot the right person. In the world of biology, this "family reunion" is the vast collection of proteins that make up living things, and the "cousins" are proteins that share a common ancestor. Scientists use special tools called NAC transcription factors to help plants read their genetic instructions, kind of like a manager reading a to-do list to tell the cell what to do. These managers have a very specific uniform: a sturdy, unchanging badge on their chest (the N-terminal domain) that identifies them as part of the team, and a more chaotic, variable jacket on their back (the C-terminal region) that might change depending on the specific job they do.

Recently, a new kind of detective tool has emerged: protein language models. Think of these as super-smart AI that has read every book in the library of life. Instead of just looking at the letters in a protein's name, these AIs understand the "vibe" or "style" of the protein, turning it into a mathematical fingerprint called an embedding. If two proteins have similar fingerprints, they might be related. Another tool, structure prediction, tries to guess what the protein looks like in 3D, giving it a "confidence score" to say how sure it is about the shape of the badge versus the jacket. The big question scientists wanted to answer was simple: Do these high-tech AI fingerprints and shape guesses match up with what we already know? Do they clearly show that the sturdy badge is more similar across different grass species than the wild, variable jacket?

---

The Grass Family Reunion: A Tale of Badges, Jackets, and Outliers

In this study, a researcher named Neil Sumanth decided to test these high-tech tools on a small group of five grass proteins. He started with a well-known rice protein called OsNAC25 and found its "look-alikes" in four other grasses: foxtail millet, sorghum, finger millet, and teff. He didn't just look at the whole protein; he split each one in half at residue 180. The first half (residues 1–180) contained the famous, sturdy badge. The second half (residues after 180) was the variable jacket.

The Badge vs. The Jacket: What the Numbers Said
When Neil compared the actual letter-by-letter sequences (the "spelling" of the proteins), the results were exactly what you'd expect. The badges (residues 1–180) were very similar across the five grasses, with an average match of 0.480. The jackets (residues after 180) were much more different, with an average match of only 0.346. It's like finding that all five cousins have the exact same family crest on their chests, but their jackets are all different colors and patterns.

The structure prediction tool, ESMFold2, agreed with this. It gave the badges a higher "confidence score" (a measure of how sure it is about the shape) than the jackets. The average score for the badges was 0.865, while the jackets scored 0.811. For most of the grasses, the tool was much more confident about the badge's shape. However, for sorghum and teff, the difference was tiny (only 0.004 and 0.002 respectively), suggesting that for them, the jacket wasn't necessarily a mess; it was just as predictable as the badge.

The AI "Vibe Check": A Surprise Twist
Here is where the story gets interesting. When Neil used the ESM C language model to compare the "vibes" (embeddings) of these proteins, the results didn't follow the simple "badge vs. jacket" rule.

If the AI was working perfectly, it should have shown that the jackets were more different from each other than the badges were. In six out of ten comparisons, the jackets were more different. But the pattern wasn't clean. The biggest surprise came from the finger millet candidate.

In the AI's "vibe space," the finger millet protein was a total outlier. It was so different from the others that it stood far away in the mathematical map.

• When comparing the full proteins, the finger millet was 0.163 to 0.240 away from the others.
• When comparing just the badges (residues 1–180), it was still 0.157 to 0.248 away.
• Meanwhile, the other four grasses (rice, foxtail millet, sorghum, and teff) were practically hugging each other, with distances as small as 0.0096 to 0.0296.

It's as if you walked into a room where four cousins looked almost identical, but the fifth cousin (finger millet) was wearing a completely different outfit, had a different accent, and stood on the other side of the room, even though they were all supposed to be wearing the same family badge.

What This Means (and What It Doesn't)
The paper is very careful not to over-explain this outlier. The researcher suggests a few possibilities: maybe the finger millet protein belongs to a different branch of the NAC family tree entirely, or maybe the way the data was collected (the "search" used to find it) accidentally picked a different cousin. The study does not prove that the finger millet protein has a different job or function. It simply flags it as a "priority suspect" that needs more investigation.

The main takeaway is that while the old-school sequence comparison and the 3D shape confidence scores both agreed that the "badge" is more conserved than the "jacket," the new AI "vibe check" was messy. It didn't just see the badge vs. jacket difference; it saw that the finger millet candidate was a weird one.

The Bottom Line
This study didn't solve the mystery of the finger millet protein, but it did a great job of pointing a flashlight at it. The results suggest that if you want to know if these grass proteins are true biological cousins (orthologs) or just look-alikes, you can't just rely on a quick similarity search. You need to do a deeper, two-way check and build a proper family tree. For now, the finger millet sequence is the one that needs a second look, while the other four grasses seem to be on the same page. The tools are powerful, but they still need a human detective to make sense of the outliers.