Unavailability of experimental 3D structural data on protein folding dynamics and necessity for a new generation of structure prediction methods in this context

This paper highlights the scarcity of experimental 3D structural data on protein folding intermediates, demonstrates the failure of existing native-structure prediction tools like AlphaFold2 in modeling these non-native states, and advocates for the development of new, dynamics-aware methods to advance the understanding of protein folding and related diseases.

The Gist

Imagine a protein as a long, tangled string of beads (amino acids). To do its job in your body—like digesting food or fighting a virus—this string needs to twist and fold into a very specific, complex 3D shape, like a origami crane. This process is called protein folding.

For a long time, scientists have been great at figuring out what the final origami crane looks like. They have massive libraries (databases) of these finished shapes. But there's a big problem: we don't know what the string looks like while it's being folded.

Think of it like this: If you only have a photo of the finished crane, you can guess the final shape. But if you want to understand how the paper folds, or why it sometimes crumples into a messy ball (which causes diseases like Alzheimer's), you need to see the string at every step of the folding process. These steps are called intermediates.

The Big Problem: We Can't "See" the Steps

The authors of this paper went on a treasure hunt to find photos of these folding steps. They looked for two types of folding:

1. Post-translational (The "Free Fall"): The whole string is made, then it tries to fold itself in a test tube.
2. Co-translational (The "Assembly Line"): The string is being built bead-by-bead by a machine (the ribosome) inside your cells, and it starts folding while it's still being made.

The Hunt Results:

• For the "Free Fall" type: They found only two studies with actual photos of the folding steps. It's like trying to learn how to bake a cake by looking at only two blurry snapshots of the batter rising.
• For the "Assembly Line" type: They found only four studies. Even worse, these studies only looked at very short strings (small proteins).

Why is this hard?
Taking a photo of a folding protein is incredibly difficult. It's like trying to take a high-definition photo of a hummingbird's wings mid-flap with a camera that has a slow shutter speed. The protein moves too fast, and the "camera" (scientific tools like X-ray or MRI) usually only captures the final, stable pose.

The "Magic Crystal Ball" That Failed

Because we lack photos of the folding steps, scientists hoped that AlphaFold2 (a famous AI that predicts protein shapes) could act as a crystal ball. They thought: "If AlphaFold2 is so good at guessing the final shape, maybe it can guess the intermediate shapes too!"

The authors tested this. They fed the AI the sequence of a protein during its folding process and asked, "What does it look like right now?"

The Result: The AI failed miserably.

• The Analogy: Imagine you show a master architect a half-built house and ask, "What does the house look like right now?" The AI, trained only on finished houses, keeps drawing the finished house, ignoring the scaffolding and the half-built walls. It tries to force the intermediate shape into a final shape.
• The Data: When the AI tried to predict the folding steps, it was often no better than random guessing. It couldn't capture the messy, dynamic reality of a protein in motion.

The "Proxy" Solution (The Best We Can Do for Now)

Since we can't take photos of the steps and the AI can't guess them, the authors looked at a clever workaround they developed in previous work.

Instead of trying to predict the dynamic steps, they took the final finished shape and simply chopped off the end of the string, then chopped off a bit more, and so on.

• The Analogy: Imagine you have a finished origami crane. To guess what it looked like halfway through, you just unfold the last few folds. You aren't seeing the real folding process (which might have involved the paper crumpling differently), but it's a "proxy" (a stand-in) that gives you a rough idea of the progression.
• The Finding: Surprisingly, these "proxy" steps were just as good (or just as bad) as the AI's guesses. This proves that current tools are fundamentally missing the "physics" of how proteins move and change over time.

Why Should You Care?

Understanding these folding steps isn't just an academic puzzle.

• Disease: Many diseases happen because proteins fold incorrectly (misfold) and clump together. If we knew the "steps" where things go wrong, we could design drugs to stop the bad folding before it happens.
• New Tools Needed: The paper concludes that we need a new generation of tools. We can't just use tools designed for static, finished shapes. We need "video cameras" (new experimental tech) and "motion-simulators" (new AI models) that understand the dynamics of folding.

The Bottom Line

We have a library of millions of finished protein shapes, but we are blind to the journey they take to get there. Current AI, no matter how smart, is like a photographer who only knows how to take pictures of the destination, not the journey. To cure diseases and understand life at a molecular level, we need to develop new ways to watch the movie of protein folding, not just look at the final frame.

Technical Summary

1. Problem Statement

Protein folding is a dynamic process where an amino acid sequence transitions through a series of 3D conformational states (intermediates) before reaching its native structure. Understanding these folding intermediates is critical for elucidating folding mechanisms and misfolding-related diseases. However, the field faces a fundamental data bottleneck:

• Scarcity of Experimental Data: While databases like the Protein Data Bank (PDB) contain hundreds of thousands of native structures, data on non-native folding intermediates (both post-translational and co-translational) is extremely sparse. Experimental techniques (NMR, cryo-EM, X-ray) struggle to capture the transient, unstable, and rapidly fluctuating nature of these intermediates.
• Limitations of Current Predictors: Existing state-of-the-art structure prediction tools, most notably AlphaFold2, are trained and optimized to predict stable native structures. It is unknown whether these models can generalize to predict the 3D structures of non-native intermediates, which often possess distinct topologies and dynamics.
• Gap in Co-translational Folding: While some computational methods exist for post-translational pathways, there is a near-total absence of methods specifically designed to predict co-translational intermediates (folding occurring as the protein is synthesized by the ribosome).

2. Methodology

The authors employed a multi-pronged approach combining literature curation, computational evaluation, and comparative analysis:

• Systematic Literature Search: The team conducted a comprehensive search for experimentally determined 3D structures of folding intermediates. They categorized findings into:
  • Post-translational intermediates: Folding of the full sequence in solution.
  • Co-translational intermediates: Folding of nascent chains as they emerge from the ribosome.
• Data Curation and Cleaning: They identified specific studies providing 3D data in the PDB. For co-translational data, they analyzed 11 distinct intermediates (17 conformations) from four studies. They applied strict quality control, discarding conformations with significant discrepancies between deposited and modeled sequences (e.g., >50% data loss).
• Computational Evaluation (AlphaFold2):
  • The authors tested whether AlphaFold2 (via ColabFold) could predict the 3D structures of the identified co-translational intermediates.
  • They input the specific amino acid sequences of the intermediates (excluding the ribosome) into AlphaFold2.
  • Metric: Structural similarity was evaluated using the Template Modeling score (TM-score). A TM-score > 0.5 indicates the same overall fold, while < 0.5 suggests different folds.
• Comparison with "Proxy" Intermediates:
  • The authors compared AlphaFold2 predictions against "proxy" intermediates. These are synthetic intermediates derived by truncating the native structure of a protein to simulate the growing chain length during co-translation.
  • This comparison aimed to determine if AlphaFold2 offers any advantage over simple structural truncation of the native state.
• Review of Existing Computational Methods: The paper surveyed existing methods for post-translational pathway prediction (e.g., Pathfinder, FoldPAthreader) and recent simulation-based approaches for co-translational folding (e.g., Monte Carlo, Molecular Dynamics).

3. Key Contributions

1. Centralized Resource: The authors created a centralized repository (available on GitHub) organizing all available experimental 3D structural data for folding intermediates, highlighting the extreme scarcity of such data.
2. Critical Assessment of AlphaFold2: This study provides the first rigorous evaluation of AlphaFold2's ability to predict co-translational intermediates.
3. Benchmarking "Proxy" vs. AI: The paper introduces a novel comparison between AI-predicted intermediates and "proxy" intermediates (truncated native structures) to establish baselines for folding dynamics prediction.
4. Identification of Emerging Methods: The authors identify and discuss pioneering methods (published around the time of their submission) that explicitly model co-translational folding using Monte Carlo and Molecular Dynamics simulations, noting that these are the first attempts to address this specific gap.

4. Key Results

A. Data Scarcity

• Post-translational: Only two studies were found providing experimental 3D structures of intermediates (covering 2 proteins, 2 intermediates each).
• Co-translational: Only four studies were found (covering 4 proteins, 2–4 intermediates each).
• Conclusion: The field lacks the large-scale, high-resolution experimental data required to train or validate new predictive models.

B. AlphaFold2 Performance on Intermediates

• Native Structures: AlphaFold2 performed excellently on native conformations (TM-scores 0.80–0.96).
• Non-Native Intermediates: AlphaFold2 failed to accurately predict non-native co-translational intermediates.
  • For 9 out of 11 non-native conformations, TM-scores were < 0.5 (ranging 0.14–0.47), indicating the predicted structures had different folds than the experimental intermediates.
  • Even for the remaining 2, scores were only marginally above 0.5.
• Implication: AlphaFold2 is biased toward predicting the stable native state and cannot capture the dynamic, non-native conformations of folding pathways.

C. Comparison: AlphaFold2 vs. "Proxy" Intermediates

• When comparing experimental intermediates to "proxy" intermediates (truncated native structures), the structural similarity (TM-score) was similarly low (0.21–0.53).
• Finding: AlphaFold2 predictions did not significantly outperform the simple "proxy" baseline. In some cases, the proxy was better; in others, AlphaFold2 was better, but neither was consistently accurate.
• This suggests that current AI models do not inherently learn the physics of folding dynamics better than a naive truncation of the native state.

D. Existing Computational Approaches

• Post-translational: Methods like Pathfinder and FoldPAthreader show promise but rely on coarse-grained data or qualitative validation rather than direct 3D structural comparison.
• Co-translational: Recent studies (Wang et al., 2025; Tao & Xiao, 2025) using Monte Carlo and Molecular Dynamics simulations that explicitly model the ribosome tunnel show potential, though they lack direct validation against experimental 3D intermediate structures due to data scarcity.

5. Significance and Future Directions

• Paradigm Shift Needed: The study concludes that the current generation of structure predictors (like AlphaFold2) is insufficient for studying protein folding dynamics. A new generation of methods is required that explicitly incorporates biophysical factors such as:
  • The vectorial nature of translation.
  • The geometry of the ribosome exit tunnel.
  • Translation rates and codon usage.
  • Chaperone interactions.
• Data-Driven Innovation: The authors draw a parallel to the early days of the PDB and CASP competitions, arguing that progress in native structure prediction was catalyzed by the gradual accumulation of data. Similarly, the field of folding dynamics needs a coordinated effort to:
  1. Organize existing sparse data.
  2. Develop new experimental technologies (e.g., 19F NMR, time-resolved cryo-EM) to capture transient intermediates.
  3. Create specialized AI models trained on these dynamics rather than just static native states.
• Evaluation Frameworks: Future methods must be evaluated not just on native structure accuracy but on their ability to reproduce folding pathways and intermediate structures, potentially using dynamic Protein Structure Networks (PSNs) as a comparison metric.

In summary, this paper serves as a critical wake-up call for the computational biology community: while we can predict static protein structures with high accuracy, we are currently blind to the dynamic journey of folding, necessitating a new era of data collection and algorithmic development.