← Latest papers
💻 bioinformatics

Elucidating enzyme-substrate specificity through co-folding foundation model

The paper introduces Boltz2ESI, an end-to-end framework that leverages a biomolecular foundation model to predict enzyme-substrate interactions through native co-folding, thereby capturing active-site plasticity and outperforming existing methods to accelerate biocatalyst discovery and pathway elucidation.

Original authors: Cheng, X., Seo, S., Huh, C., Chen, J., Jiang, S., Guo, P., Weng, J.-K., Kim, W. Y., Jin, W.

Published 2026-08-02
📖 7 min read🧠 Deep dive

Original authors: Cheng, X., Seo, S., Huh, C., Chen, J., Jiang, S., Guo, P., Weng, J.-K., Kim, W. Y., Jin, W.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 cell as a bustling, high-tech factory where tiny machines called enzymes are constantly at work. These enzymes are the master chemists of life; they grab specific raw materials (called substrates) and snap them together or break them apart to keep the organism alive. The magic of this factory lies in a perfect lock-and-key relationship: an enzyme's active site is a uniquely shaped pocket that only fits one specific chemical key. If the key is even slightly wrong, the machine won't start, and the factory grinds to a halt. For decades, scientists have tried to map out which keys fit which locks, but it's been like trying to solve a giant jigsaw puzzle in the dark. Traditional methods were slow, and computer programs that tried to predict these matches often treated the enzyme like a rigid, unchanging statue. But in reality, enzymes are more like flexible dancers; they wiggle, twist, and change shape the moment a substrate steps onto the dance floor to make the chemistry work.

This paper introduces a new, super-smart computer program called Boltz2ESI that finally learns to watch these molecular dances. Instead of guessing based on rigid shapes or just reading the enzyme's genetic code, this new tool uses a "foundation model"—a type of AI that has already learned the deep rules of how proteins fold and move. By simulating the enzyme and the substrate folding together at the same time, Boltz2ESI can see how the enzyme's pocket reshapes itself to hug the substrate perfectly. The researchers found that this approach is much better at predicting which chemicals an enzyme will actually work with, even for enzymes that look very similar to each other but have different jobs. This isn't just a theoretical win; the team showed it could successfully find the right enzymes for a complex plant-based drug pathway, proving that understanding the "dance" of molecules is the key to unlocking nature's chemical secrets.

The Problem: The Rigid Statue vs. The Dancing Partner

For a long time, scientists trying to predict enzyme behavior had two main tools, and both had a fatal flaw. The first tool looked only at the enzyme's genetic sequence (its DNA instructions). It was like trying to guess what a person looks like just by reading their name; it gave a general idea but missed all the tiny details that matter. The second tool looked at the enzyme's 3D shape, but it treated the enzyme like a stone statue. It would try to jam a chemical key into a frozen pocket. The problem is, enzymes aren't statues. When a real substrate arrives, the enzyme often snaps its arms shut, twists its body, or shifts its shape to grab the key tightly. This is called "induced fit." If your computer model treats the enzyme as a rigid block, it misses the most important part of the interaction: the movement.

Furthermore, many of these enzymes are "orphans." Scientists know they exist in nature, but they have no idea what they do or what shape their active pockets look like. Traditional methods needed a pre-drawn map of the pocket to work, which meant they were useless for these mysterious enzymes.

The Solution: The Co-Folding Dance Floor

Enter Boltz2ESI. The authors built this framework to solve the "rigid statue" problem by using a technique called native co-folding.

Think of it this way: instead of trying to force a key into a locked door, Boltz2ESI simulates the door and the key walking into a room together and figuring out how to fit. It uses a powerful AI model called Boltz-2, which has already learned the physics of how proteins and chemicals behave.

Here is how the process works, step-by-step:

  1. The First Look: The AI takes the enzyme's genetic sequence and the chemical formula of the substrate (written as a SMILES string) and runs a simulation to guess where the active site might be.
  2. The Dance: It then zooms in on that spot and simulates the enzyme and the substrate folding together without any pre-set rules. This allows the enzyme to naturally shift its shape, just like a real biological system would, to accommodate the substrate.
  3. The Context Check: Once the "dance" is done, the AI doesn't just look at the shape. It also checks the enzyme's evolutionary history (using a tool called ESM3) to see how similar enzymes have behaved over millions of years. It also analyzes the chemical details of the substrate (using Uni-Mol2 and Morgan fingerprints) to understand its specific features.
  4. The Verdict: Finally, it combines all this information—the shape of the dance, the family history, and the chemical details—to give a score. This score tells us how likely it is that this enzyme and substrate are a perfect match.

What They Found: Better Than the Rest

The team tested Boltz2ESI against the best existing methods using a massive dataset called ESIBank, which contains over 323,000 enzyme-substrate pairs.

  • Beating the Rigid Models: When they compared Boltz2ESI to older methods that treated enzymes as rigid statues (like EZSpecificity), Boltz2ESI won hands down. It achieved a score of 0.9156 (on a scale where 1.0 is perfect) in standard tests, while the rigid methods struggled around 0.89.
  • The "Unknown" Test: The real challenge was testing on enzymes and chemicals the AI had never seen before. Even in this tough scenario, Boltz2ESI maintained a strong score of 0.7664, significantly outperforming the rigid models.
  • The "Why" Matters: The researchers discovered something fascinating. If they just used the base AI model to predict how tightly a chemical sticks to an enzyme (binding affinity), it wasn't enough. Many chemicals stick tightly but don't actually get "processed" by the enzyme. Boltz2ESI succeeded because it learned to distinguish between a chemical that just sticks and one that actually works. It learned the difference between a hug and a handshake.

The Proof: Citrate Synthase and the Plant Pathway

To prove their point, the authors looked at a classic enzyme called citrate synthase. This enzyme is famous for a dramatic move: when its substrate arrives, a whole section of the enzyme rotates about 19 degrees to close a door around the chemical.

  • The Rigid Failure: When they tried to use old "rigid docking" methods, the computer couldn't see this rotation. The substrate ended up in the wrong spot, and the model gave a low confidence score.
  • The Co-Folding Success: Boltz2ESI, however, simulated the rotation naturally. It predicted the closed shape perfectly and gave a high confidence score of 0.9770, proving it could capture the "dance" that the rigid models missed.

They also tested the tool on a real-world mystery: the withanolide pathway in the Ashwagandha plant. This pathway produces medicinal compounds, but scientists were still figuring out which specific enzymes (P450s) were responsible for each step.

  • Boltz2ESI was given a list of 107 candidate enzymes and asked to find the right ones.
  • It successfully ranked the correct enzymes much higher than previous methods. For example, it found the enzyme for step 1 (CYP87G1) at rank 8, while the previous best method had it at rank 13.
  • It also correctly grouped together "paralogs" (very similar enzymes that do the same job in different parts of the plant), showing it understands the underlying biology, not just the surface-level data.

Why This Changes Things

The authors suggest that this approach is a major step forward for "de-orphaning" metabolic pathways—figuring out what unknown enzymes do in nature. By moving away from rigid, static models and embracing the dynamic, flexible nature of enzymes, Boltz2ESI provides a more accurate map of the chemical world.

While the tool is currently computationally heavy (it takes time to run these complex simulations), the results suggest that the future of enzyme discovery lies in models that can watch the dance, not just look at the frozen pose. This could accelerate the discovery of new medicines, better industrial enzymes, and a deeper understanding of how life's chemistry actually works.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →