A Deep-Learning Atlas of XPO1-Mediated Nuclear Export at Proteome Scale

This study leverages AlphaFold 3 to construct a comprehensive, deep-learning-based atlas of the human XPO1-mediated nuclear exportome, identifying over 3,000 high-confidence NESs—including noncanonical patterns and regulatory mechanisms—that significantly expand our understanding of nuclear transport and its dysregulation in disease.

The Gist

Imagine your cell as a bustling, high-security city. Inside this city, there is a massive, walled fortress called the Nucleus (the city hall), where the master blueprints (DNA) are kept safe. Outside the walls is the rest of the city (the Cytoplasm), where the construction crews (proteins) actually build things.

To keep the city running, blueprints and workers need to move back and forth through the city gates (the Nuclear Pore). But you can't just let anyone in or out; you need a specific security badge to pass through.

The Problem: The Broken Badge Scanner

For years, scientists knew about these "badges," called NESs (Nuclear Export Sequences). They looked like short strings of letters in a protein's code. However, the old way of finding them was like trying to find a specific person in a crowd just by looking at their shirt color.

• The old method: "If the shirt has three red stripes, they must be a VIP."
• The reality: Many people with three red stripes aren't VIPs, and some VIPs are wearing blue shirts with a hidden red stripe. The old scanners were full of false alarms and missed the real VIPs.

The Solution: A 3D Deep-Learning "X-Ray"

This paper introduces a new, super-smart system using AlphaFold 3, an AI that can predict the 3D shape of proteins with incredible accuracy.

Instead of just reading the "shirt color" (the sequence of letters), the researchers used AI to build a 3D hologram of over 4,000 human proteins trying to fit through the city gate. They simulated the interaction between the protein (the cargo), the gatekeeper (a protein called XPO1), and the security key (RanGTP).

The Analogy:
Imagine trying to fit a key into a lock.

• Old way: You look at the key's teeth on paper and guess if it fits.
• New way (This paper): You 3D-print the key and the lock, then actually try to turn it. You can see exactly how the metal bends, which teeth click into place, and if the key turns smoothly.

What They Discovered

By "turning the key" for thousands of proteins, they found some amazing things:

1. The "Secret Handshakes" (New Badge Types)
They found over 3,000 new badges that the old scanners missed. Some of these badges were shaped differently than anyone expected.

• Analogy: They found VIPs wearing "backwards" badges or badges that required a specific twist to unlock the door. The AI showed that the gate (XPO1) is much more flexible than we thought; it can accept keys that bend, twist, or even skip a step in the lock mechanism.

2. The "Shape-Shifting" Badges
Some proteins only show their badge when they change shape.

• Example: The protein ATG3 (involved in cleaning up cellular trash) hides its badge when it's alone. But when it meets the gatekeeper, it unfolds and reveals a hidden helix that fits perfectly.
• Analogy: It's like a spy who wears a trench coat. You can't see their ID badge until they take off the coat. The AI showed us exactly when and how they take off the coat.

3. The "Broken Badges" in Disease
They looked at proteins involved in diseases like leukemia.

• Example: In a type of leukemia, a mutation changes the shape of a protein called NPM. This mutation accidentally creates a new, strong badge that wasn't there before, causing the protein to get kicked out of the nucleus when it shouldn't be.
• Analogy: A typo in the blueprint accidentally printed a "VIP" badge on a janitor's uniform, causing the janitor to be let into the server room, causing chaos.

4. The "Double-Action" Switch
They found many proteins that have both an "Enter" badge (NLS) and an "Exit" badge (NES) right next to each other.

• Analogy: It's like having a "Push to Enter" and "Pull to Exit" button right next to each other on a door. The cell can flip a switch (like adding a phosphate tag) to block one button and activate the other, instantly deciding whether the protein stays inside or goes outside.

Why This Matters

This paper isn't just a list of new proteins; it's a new map of how our cells move things around.

• For Cancer: Many cancer cells rely on moving bad proteins out of the nucleus to grow. This map helps us see exactly which "doors" they are using, helping doctors design better drugs (like Selinexor) to jam those specific doors.
• For Future Science: It proves that we can't just read the genetic code to understand how cells work; we have to understand the 3D shape and how it moves.

In a nutshell: The researchers used a super-smart AI to build a 3D simulation of the cell's security system. They found thousands of hidden keys, discovered that the locks are more flexible than we thought, and figured out how diseases can break the system. It's like upgrading from a 2D paper map to a full, interactive 3D GPS for the cell's traffic control.

Technical Summary

1. Problem Statement

Nuclear export of proteins is primarily mediated by Exportin 1 (XPO1/CRM1), which recognizes short, leucine-rich Nuclear Export Sequences (NESs). Despite its critical role in cellular homeostasis and its dysregulation in diseases like cancer and neurodegeneration, a comprehensive map of XPO1 cargos remains elusive.

• Limitations of Current Methods: Traditional sequence-based prediction tools rely on linear hydrophobic spacing motifs (e.g., $\Phi$-X$_{2-3}$-$\Phi$-X$_{2-3}$-$\Phi$-X$_{2-3}$-$\Phi$). These methods suffer from high false-positive rates because they ignore the structural requirement that NESs must adopt a specific amphipathic $\alpha$-helical conformation to dock into the XPO1 binding groove.
• Contextual Blindness: Many NESs are conditionally formed (disorder-to-order transitions) or sterically occluded in full-length proteins, rendering them invisible to peptide-based assays or sequence scans.
• Structural Bias: Existing structural knowledge is limited to a small set of canonical complexes, failing to capture non-canonical binding geometries, reverse orientations, or the impact of post-translational modifications (PTMs) and cofactors on export competence.

2. Methodology

The authors developed a structure-resolved, deep-learning-based workflow to model XPO1-mediated export across the human proteome.

• Dataset Construction: A non-redundant set of ~4,000 human proteins was curated from BioGRID interactomes, the Human Protein Atlas (nucleo-cytoplasmic shuttling proteins), and disease-associated gene sets (e.g., cancer biomarkers, kinases, autophagy factors).
• AlphaFold 3 (AF3) Modeling:
  • Full-length cargo proteins were modeled in ternary complexes with XPO1, Ran-GTP, and RanBP1.
  • Modeling was performed in template-free mode for the cargo to avoid bias, using the experimentally determined XPO1-RanGTP-RanBP1 structure (PDB: 6CIT) as a scaffold for the receptor complex.
  • Five independent models were generated per cargo to assess reproducibility.
• Structure-Aware Scoring Framework:
  • Pocket Mapping: Hydrophobic pockets (P0–P4) within the XPO1 H11-H12 groove were defined based on the 6CIT structure.
  • Scoring Matrix: A composite "Model_Rank" score (0–21) was calculated based on:
    • Base Score (0–16): Pocket occupancy (P0–P4), helix length, pLDDT (local confidence), contact density with the groove, and interactions with gatekeeper residues (e.g., K579).
    • Reproducibility Score (0–5): Consistency of the NES window and pocket occupancy across the five AF3 models.
• Classification System:
  • Developed an Orientation-Aware Classifier that determines helix directionality (Forward vs. Reverse) and maps hydrophobic anchors to pockets.
  • Introduced new classes for non-canonical geometries:
    • "Like" Classes ($XX\_L$): Canonical spacing but with non-hydrophobic residues (e.g., Thr, Ala) at distal anchor positions (P3/P4).
    • Class 5: Internal pocket skips (e.g., P0-P1-P2-x-P4) where an amphipathic helix bypasses one pocket but re-engages downstream.
• Experimental Validation:
  • Selected candidates (e.g., CEP43, UBP12, ATG3, ANKZF1, MAP2) were fused to an SV40-NLS-GFP reporter.
  • Subcellular localization was assessed in HEK293T cells with and without Leptomycin B (LMB), a specific XPO1 inhibitor, to confirm XPO1-dependence.

3. Key Contributions

• Proteome-Scale Atlas: Generated a high-confidence atlas of >3,000 NESs across ~4,000 human proteins, significantly expanding the known XPO1 exportome.
• Structural Grammar Expansion: Defined new structural classes of NESs that were previously undetectable by sequence rules, including:
  • Reverse Orientation (1a-R): NESs binding in the C-to-N direction.
  • Class 5 (Internal Skip): Motifs with internal gaps in pocket occupancy.
  • "Like" Variants: Motifs tolerating polar residues (Thr/Ala) at anchor positions, suggesting regulatory potential via PTMs.
• Mechanistic Insights:
  • Demonstrated that Zinc coordination (in ANKZF1) and phosphorylation (in Cyclin B1) act as structural switches that expose or occlude NES helices.
  • Identified juxtaposed NES-NLS motifs (within ±50 residues) in ~19% of cargos, suggesting a mechanism for mutually exclusive import/export regulation (e.g., in ANKZF1).
• Resource Release: Created a searchable web portal (nuclear.exportome.org) providing 2D groove plots, domain maps, and structural models for the entire dataset.

4. Key Results

• Benchmarking: AF3 accurately reproduced known NES structures (e.g., PDPK1, APC) with backbone RMSDs < 1.0 Å and correctly identified PTM-induced export loss (Cyclin B1 phosphorylation at Ser147).
• Novel Discoveries:
  • ATG3 (Autophagy): Identified a high-confidence NES in the linker region between catalytic subdomains. AF3 showed the NES is occluded in the apo-state but exposed upon XPO1 binding. Validation confirmed XPO1-dependent export.
  • RQC Pathway: Discovered a cluster of NESs in Ribosome Quality Control components (e.g., ZNF598, NEMF, ANKZF1), suggesting a nuclear interface for translation-coupled quality control.
  • ANKZF1: Revealed a zinc-dependent NES. In the absence of Zn, the NES is disordered; with Zn, it forms a stable helix. Mutating the NLS adjacent to the NES blocked nuclear entry, confirming a coordinated shuttling mechanism.
  • MAP2: Validated a Class 5 NES (internal skip) that drives XPO1-dependent export, proving that non-canonical spacing can still be functional.
• Disease Relevance: The atlas resolved ambiguous NESs in disease-associated proteins (e.g., CEP43 in ciliogenesis, UBP12 in deubiquitination) and identified how frameshift mutations in NPM1 (AML) generate de novo NESs.
• Scoring Correlation: High-scoring NESs (Score > 15) correlated strongly with robust cytoplasmic export in reporter assays, while lower scores (10–14) often indicated weaker or conditional export.

5. Significance

• Paradigm Shift: Moves NES discovery from a sequence-motif search to a structure-resolved, pocket-level framework. This resolves the "degeneracy" problem where many sequences look like NESs but fail to bind structurally.
• Regulatory Mechanisms: Highlights that export competence is dynamically regulated by protein folding, cofactor binding (Zn), and PTMs, rather than being a static sequence feature.
• Therapeutic Implications: Provides a mechanistic basis for understanding how mutations alter nucleocytoplasmic trafficking in cancer and neurodegeneration. It offers a framework for predicting the efficacy of XPO1 inhibitors (e.g., Selinexor) based on the structural integrity of specific cargo NESs.
• Future Directions: The identification of "extinct" or "dormant" NESs suggests evolutionary plasticity in nuclear transport, and the atlas provides a target list for systematic mutagenesis to explore the "structural grammar" of nuclear export.

In summary, this study leverages AlphaFold 3 to construct the first comprehensive, structure-based map of the human exportome, revealing that XPO1 recognizes a much broader and more regulatable set of structural geometries than previously appreciated.