⚛️ biophysics

Structural divergence in protein evolution of a photoreceptor family undetected by AlphaFold is observed by high sensitivity FT-IR spectroscopy

This study reveals that high-sensitivity FT-IR spectroscopy detects significant secondary structural divergence in photoactive yellow protein family members and their resurrected ancestors that AlphaFold failed to predict, highlighting a critical limitation of the AI model in capturing evolutionary structural changes despite high sequence similarity.

Original authors: Dohmen, R. L., Hoogerwerf, G., Xie, A., Hoff, W. D.

Published 2026-08-06

📖 4 min read☕ Coffee break read

CC0 1.0

Original authors: Dohmen, R. L., Hoogerwerf, G., Xie, A., Hoff, W. D.

Original paper dedicated to the public domain under CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the human body as a bustling city, and inside every cell, there are millions of tiny machines called proteins. These machines do everything from carrying oxygen to helping you see the world. But how do these machines change over millions of years to do new jobs? This is the story of evolution, but instead of looking at fossils, scientists look at the blueprints of these proteins. Sometimes, two proteins look almost identical in their blueprint (their DNA sequence), yet they act completely differently. It's like finding two cars with the same engine code, but one is a race car and the other is a slow-moving tractor. Scientists have been trying to figure out how these tiny differences in the blueprint lead to such huge differences in how the machine works. For a long time, they relied on taking pictures of these proteins under powerful microscopes, but that's hard to do for every single version. Recently, a super-smart computer program called AlphaFold started predicting what these proteins look like just by reading their blueprints, promising to solve this mystery instantly. But does the computer see everything?

This paper tells the story of a team of scientists who decided to test if this super-computer is truly all-seeing. They focused on a specific family of light-sensing proteins called "Photoactive Yellow Proteins" (PYPs), which act like tiny solar panels for bacteria, helping them sense blue light. In one species of bacteria, there are two versions of this protein, PYP1 and PYP2. They are 60% identical—like cousins who share a lot of family traits—but they behave very differently. PYP1 is a quick thinker; it senses light and resets its "alarm" in about half a second. PYP2, however, is a slow thinker; it takes a full minute to reset. That's a 100-fold difference in speed! The scientists wanted to know: as these two proteins evolved from a common ancestor, did their physical shapes change to cause this speed difference?

To find out, the researchers used a time-traveling trick called "ancestral sequence reconstruction." They didn't have a real time machine, but they used math to guess the DNA of the ancient ancestor that PYP1 and PYP2 came from, and even the intermediate steps in between. They then built these ancient proteins in a lab to see how they actually behaved. First, they asked the computer, AlphaFold, to predict the 3D shapes of these resurrected ancestors. The computer said, "No problem!" It predicted that all these proteins, from the ancient ancestor to the modern PYP1 and PYP2, looked indistinguishable from one another. When the scientists measured the tiny differences in the predicted atomic positions, the numbers were so small (less than 1.1 Ångströms) that they fell within the range of normal experimental noise. Based on these predictions, the computer concluded that the proteins kept the exact same shape while just tweaking their speed.

But the scientists weren't convinced. They decided to check the computer's work with a different tool: a high-tech scanner called FTIR spectroscopy. Think of this like listening to the protein's internal vibrations to see how its parts are arranged, rather than just looking at a static picture. When they scanned the proteins, they found something the computer missed. While the computer predicted the shapes were indistinguishable, the FTIR scanner showed clear differences in the "secondary structure"—the way the protein's backbone folds into spirals and sheets. Specifically, the proteins that behaved like the slow PYP2 had lost a significant amount of their spiral structure compared to the fast PYP1. The computer had completely overlooked these substantial structural shifts that were actually happening during evolution.

The paper concludes that while AlphaFold is an amazing tool, it isn't perfect. It can miss real, physical changes in protein structure that happen during evolution, especially when those changes are linked to how fast or slow a protein works. The scientists showed that by combining the "time travel" of ancestral reconstruction with the "listening" power of spectroscopy, they could uncover the true story of how these proteins diverged. The lesson here is that even the smartest computers might need a little help from real-world experiments to see the full picture of how life's tiny machines evolve.

Technical Summary: Structural Divergence in Protein Evolution Undetected by AlphaFold

Problem Statement
Functional and structural divergence within protein families is a fundamental mechanism of molecular evolution. While AlphaFold has revolutionized the prediction of atomic-resolution protein structures, its ability to capture subtle structural changes associated with functional divergence remains untested. The Photoactive Yellow Protein (PYP) family, specifically the two homologs PYP1 and PYP2 from Halorhodospira halophila, presents a unique case study. Despite sharing 60% sequence identity, these homologs exhibit a 100-fold difference in the lifetime ( $\tau_{pB}$ ) of their pB signaling intermediate and distinct peak absorbance wavelengths ( $\lambda_{max}$ ). Previous studies have identified residues affecting spectral tuning ( $\lambda_{max}$ ), but the structural mechanisms driving the dramatic kinetic divergence ( $\tau_{pB}$ ) remain poorly understood. Furthermore, it is unknown whether AlphaFold can detect the structural rearrangements required to produce such functional differences.

Methodology
The authors employed a multi-faceted approach combining ancestral sequence reconstruction (ASR), computational modeling, and high-sensitivity spectroscopy:

Ancestral Sequence Reconstruction (ASR): Using a curated multiple sequence alignment of 984 extant PYP homologs, the authors reconstructed the Last Common Ancestor (LCA) of PYP1 and PYP2, as well as three intermediate ancestral nodes (Node 1A, Node 1B, and Node 2A) along the evolutionary paths to the extant proteins.
Protein Resurrection and Characterization: The genes for the LCA and intermediate ancestors were synthesized, expressed in E. coli, and purified. The functional properties of these resurrected proteins were characterized using UV/Vis absorbance spectroscopy to determine $\lambda_{max}$ and the lifetime of the pB intermediate ( $\tau_{pB}$ ).
Computational Structure Prediction: AlphaFold2 was used to predict the 3D structures of the resurrected ancestral proteins and PYP2. These predictions were compared against the high-resolution crystal structure of PYP1 (PDB: 1NWZ) using Root Mean Square Deviation (RMSD) calculations.
FTIR Spectroscopy: To experimentally validate the structural predictions, the authors optimized second-derivative Fourier Transform Infrared (FTIR) spectroscopy. This technique was used to probe:
- Active Site Hydrogen Bonding: Specifically the C=O stretching mode of Glu46, which correlates with $\lambda_{max}$ .
- Secondary Structure: Analysis of the Amide I region to detect changes in $\alpha$ -helical and $\beta$ -sheet content.

Key Results

Functional Divergence: The resurrected LCA exhibited a $\lambda_{max}$ of 446 nm and a $\tau_{pB}$ of 0.08 seconds, phenotypically similar to PYP1 (447 nm, 0.5 s). The evolutionary path to PYP2 involved a shift in $\lambda_{max}$ to 442 nm and a dramatic increase in $\tau_{pB}$ to ~52 seconds. The intermediate Node 2A already displayed PYP2-like kinetics ( $\tau_{pB}$ = 46 s), indicating that the kinetic divergence occurred early in the lineage leading to PYP2.
AlphaFold Predictions: AlphaFold2 predicted structures for all ancestral nodes and PYP2 with high confidence. Pairwise RMSD calculations between these predicted structures and the PYP1 crystal structure yielded values ranging from 0.1 to 1.1 Å, centered near 0.55 Å. These values are well within the range of variation seen among different crystal forms of the same protein (PYP1), leading to the conclusion that AlphaFold predicts PYP2 and its ancestors to be structurally indistinguishable from PYP1.
FTIR Spectroscopy Findings:
- Active Site: The FTIR spectra confirmed the relationship between Glu46 hydrogen bonding strength and $\lambda_{max}$ . PYP1, Node 1A, and Node 1B (447 nm) showed a Glu46 C=O stretch at 1726 cm $^{-1}$ . PYP2 and Node 2A (442 nm) showed a shift to 1723 cm $^{-1}$ , indicating a stronger hydrogen bond. Interestingly, the LCA (446 nm) also showed the 1723 cm $^{-1}$ shift, suggesting compensatory structural adjustments not captured by simple sequence analysis.
- Secondary Structure: The most significant finding was in the Amide I region. While PYP1, Node 1A, and Node 1B displayed characteristic peaks for $\alpha$ -helices (1644 cm $^{-1}$ ) and $\beta$ -sheets (1634 cm $^{-1}$ , 1688 cm $^{-1}$ ), Node 2A and PYP2 exhibited a near-complete loss of the $\alpha$ -helical peak at 1644 cm $^{-1}$ . The authors attribute the reduced amplitude in the second-derivative spectra to increased structural heterogeneity and significant adjustments in secondary structure.

Key Contributions and Significance
The primary contribution of this work is the demonstration that substantial structural divergence in a protein family can occur without being detected by state-of-the-art structure prediction tools like AlphaFold2.

Limitation of AlphaFold: The study reveals a critical limitation: AlphaFold2 successfully predicted the global fold of PYP2 and its ancestors as identical to PYP1, despite experimental evidence (FTIR) showing distinct secondary structure changes and a 100-fold difference in functional kinetics.
Validation of ASR and Spectroscopy: The paper validates the utility of combining ancestral sequence reconstruction with high-sensitivity FTIR spectroscopy to uncover evolutionary mechanisms. This approach successfully recapitulated the evolutionary trajectory, identifying that the shift to PYP2's slow kinetics and blue-shifted spectrum involved significant secondary structural rearrangements.
Mechanistic Insight: The results suggest that the functional divergence between PYP1 and PYP2 is driven by structural adjustments that alter secondary structure content and active site hydrogen bonding, changes that are not captured by the current generation of structure prediction algorithms.

Conclusion
The authors conclude that while AlphaFold is a powerful tool, it may fail to capture the subtle, functionally critical structural adjustments that drive protein evolution. The study highlights the necessity of experimental validation, particularly using spectroscopic methods capable of detecting secondary structure changes, to fully understand the structural basis of functional divergence in protein families.

More like this