Integrating glycosylation in de novo protein design with ReGlyco Binder Design Filter

This paper demonstrates that integrating explicit glycosylation filtering using the ReGlyco tools and GlycoShape database into de novo protein design pipelines significantly improves the efficiency of identifying successful binders against heavily glycosylated targets, as validated by screening designs for the Nipah virus glycoprotein and providing a workflow for human erythropoietin.

The Gist

The Big Picture: Designing Protein "Keys" for Sugar-Coated "Locks"

Imagine you are a master locksmith trying to design a new key (a protein binder) to open a very specific, high-security door (a virus protein).

For years, scientists have been using powerful AI to design these keys. They can generate thousands of designs in minutes. However, there's a catch: The door isn't just a plain metal surface; it's covered in thick, moving velvet curtains (sugars/glycans).

Most AI designs ignore these curtains. They design keys that look perfect on a blueprint but, when you try to use them in real life, they get stuck in the velvet. They hit the curtains instead of the lock. This wastes a lot of time and money because scientists have to build and test thousands of "keys" that are doomed to fail before they even leave the lab.

This paper introduces a new "pre-check" system. It's like a virtual simulator that puts the velvet curtains on the door before you try to make the key. It filters out the bad designs instantly, saving time and money.

---

The Problem: The "Velvet Curtain" Effect

The Science:
Proteins in our bodies (and viruses) are often covered in sugar molecules called glycans. These sugars are huge, flexible, and constantly moving. They act like a protective shield.

The Analogy:
Think of the virus protein (NiV-G) as a castle wall. The sugars are like a dense, foggy forest growing right up against the wall.

• Old Method: Architects (AI) design a bridge to the castle wall, but they forget about the forest. The bridge crashes into the trees.
• The Result: Scientists build the bridge, try to cross it, and it fails. They have to start over. This happens with about 20% of the designs in recent competitions.

The Solution: The "ReGlyco" Filter

The authors created a tool called ReGlyco (part of the GlycoShape database). Think of this as a 3D "Traffic Cop" for protein design.

1. The Simulation: Before a design is sent to the lab, ReGlyco takes the blueprint and adds the "forest" (the sugars) to the target protein.
2. The Check: It asks, "Does this new key crash into the trees?"
   • If Yes: It flags the design as a "Fail" immediately.
   • If No: It gives a "Pass."
3. The Flexibility: Sometimes, the trees (sugars) wiggle, or the branch holding the tree (the protein sidechain) can twist slightly. ReGlyco has a special mode called ReGlyco Rotamer that checks if a tiny twist can save a design that looks like it's crashing.

The Real-World Test: The Nipah Virus Challenge

To prove this works, the authors looked at a recent global competition where scientists tried to design keys to stop the Nipah virus (a dangerous virus with a high fatality rate).

• The Setup: 1,200 new "keys" were designed by AI.
• The Reality: Only about 10% actually worked when tested in the lab.
• The ReGlyco Test: The authors ran all 1,200 designs through their new filter.
  • The Result: The filter correctly identified 236 designs that would fail because they crashed into the sugar shields.
  • The Efficiency: It did this in just 3 hours on a standard computer.
  • The Savings: If scientists had used this filter before the competition, they wouldn't have wasted time and money building those 236 useless keys.

The Demo: A Playground for Everyone

The authors didn't just keep this tool for themselves. They built a free, interactive "playground" (a Google Colab notebook) where anyone can try it out.

• The Example: They used Erythropoietin (hEPO), a hormone used to treat anemia, which is also heavily covered in sugars.
• How it works: You upload a target, the tool adds the sugar "forest," and then it uses AI to design mini-keys that fit around the trees, not through them.
• The Outcome: It shows you which designs are "Pass," "Borderline" (might work with a slight twist), or "Fail."

Why This Matters

In the world of drug discovery, time is money.

• Before: Design a key $\rightarrow$ Build it $\rightarrow$ Test it $\rightarrow$ It fails because of sugar $\rightarrow$ Start over. (Expensive and slow).
• After: Design a key $\rightarrow$ Virtual Sugar Check $\rightarrow$ It fails? Discard it. It passes? Build it. (Fast and cheap).

The Bottom Line:
By explicitly adding the "sugar curtains" into the design process, scientists can stop wasting resources on designs that are physically impossible to work. It's a simple but powerful step that makes the future of drug design faster, cheaper, and more likely to succeed.

Technical Summary

1. Problem Statement

Current AI-driven de novo protein design pipelines (e.g., RFdiffusion, AlphaFold 3) are revolutionizing structural biology but suffer from a critical limitation: they largely treat proteins as unmodified polypeptide chains, neglecting post-translational modifications (PTMs), specifically glycosylation.

• The Consequence: Glycans are heterogeneous, species-specific, and occupy significant surface volume. Ignoring them leads to a high rate of false positives in binder design, where predicted binders clash sterically with glycans on the target protein.
• The Cost: This results in inefficient experimental workflows, where thousands of computationally generated designs yield meager success rates in the lab, wasting time and resources.
• Data Scarcity: Existing AI models struggle with glycosylation because 3D structural data for glycans in the Protein Data Bank (PDB) is fragmented and insufficient for training robust predictive models.

2. Methodology

The authors propose a computational filter that explicitly integrates glycosylation into the design pipeline using the GlycoShape database and ReGlyco tools.

• Tools Used:

  • GlycoShape: A structural database containing ~920 unique 3D glycan structures derived from molecular dynamics (MD) simulations (approx. 2 ms sampling).
  • ReGlyco: A tool that links glycan structures to protein sites by minimizing a loss function.
  • ReGlyco Rotamer: A new tool developed in this work that allows the Asparagine (Asn) sidechain of the aglycon to sample different rotamers (using Dunbrack's library) to resolve steric clashes that are purely due to sidechain orientation.
  • Boltz-2 & AlphaFold 3 (AF3): Used for initial structure prediction of binder-target complexes.
  • RFdiffusion (RFD3): Used for generating de novo binder designs in the demonstration workflow.

• Workflow:

  1. Target Reconstruction: Reconstruct the target protein with explicit glycans based on glycoproteomics data (using ReGlyco Ensemble for dynamic representation).
  2. Filtering: Screen designed binders against the glycosylated target.
     • Mode 1 (Rigid): ReGlyco checks for clashes with glycans attached in a fixed conformation.
     • Mode 2 (Flexible): ReGlyco Rotamer allows the Asn sidechain to rotate to clear minor clashes.
  3. Triage: Designs flagged as "clashing" are discarded or refined before experimental testing.

3. Key Contributions

• Novel Filtering Pipeline: Introduction of the ReGlyco Binder Design Filter, a computationally efficient method to screen de novo designs against glycosylated targets.
• ReGlyco Rotamer Tool: Development of a specific module to distinguish between "true" steric clashes (impossible to bind) and "false" clashes caused by rigid sidechain modeling.
• Open-Source Resources:
  • A Google Colab notebook demonstrating the workflow for designing mini-binders against human erythropoietin (hEPO).
  • Public availability of the GlycoShape database and the filtered results of a major protein design competition.

4. Results

Case Study 1: Nipah Virus (NiV-G) Competition

The authors re-analyzed 1,201 designs from an Adaptyv Bio competition targeting the heavily glycosylated NiV glycoprotein (NiV-G).

• Experimental Baseline: Of the 1,201 designs expressed, only 116 (9.6%) were confirmed binders.
• ReGlyco Filtering (Rigid):
  • Flagged 251 designs as non-binders due to steric clashes.
  • 236 of these were confirmed non-binders (True Positives for filtering).
  • 15 were confirmed binders (False Positives for filtering; likely due to rigid modeling).
• ReGlyco Rotamer Refinement:
  • Reduced flagged designs to 138.
  • 133 were confirmed non-binders.
  • Only 5 confirmed binders were flagged.
  • Analysis of the 5: Visual inspection and re-prediction with AlphaFold 3 (without glycans) revealed that in 4 cases, the clash was irresolvable (incorrect pose). In 1 case ("Soft-Panda-Snow"), AF3 found an alternative non-clashing pose, suggesting that combining multiple prediction tools improves accuracy.
• Efficiency: The entire screening of 1,201 complexes took ~3 hours on a dual-core CPU.

Case Study 2: Human Erythropoietin (hEPO) Demo

• Demonstrated the workflow on hEPO (a smaller, heavily glycosylated hormone).
• Generated 20 mini-binders using RFdiffusion.
• Filtering Results:
  • 14 designs: Passed (no clashes).
  • 1 design: Borderline (cleared by Rotamer flexibility).
  • 5 designs: Failed (irresolvable clashes).
• This validated the workflow's ability to triage designs for smaller biologics within free cloud computing limits.

5. Significance

• Cost Reduction: By filtering out ~20% of designs that are guaranteed to fail due to glycan clashes before experimental expression, the method significantly reduces laboratory costs and time.
• Improved Success Rates: The integration of glycosylation moves the field from "protein-centric" to "glyco-protein-centric" design, addressing a major source of failure in current AI design pipelines.
• Accessibility: The tools and workflows are open-access, allowing researchers to integrate glycosylation constraints into their own de novo design projects without needing massive computational resources (unlike running full MD or AF3 on every candidate).
• Future Outlook: The study highlights that while AI tools are improving, explicit integration of PTM data (via databases like GlycoShape) remains essential for high-fidelity structural biology and drug discovery.