Integrated proteomic screening reveals design principles of CRBN molecular glue degraders

By integrating deep proteomic and ubiquitinomic profiling of a 960-compound library with interpretable machine learning, this study systematically expands the CRBN neosubstrate landscape to include over 124 new targets and establishes design principles for the rational development of next-generation molecular glue degraders.

The Gist

Imagine your body is a bustling city, and inside every cell, there's a massive recycling plant called the Proteasome. Its job is to break down old, broken, or unnecessary proteins (the "trash") to keep the city running smoothly.

Sometimes, the city needs to get rid of specific, dangerous proteins that are causing trouble—like a corrupt official or a broken machine. This is where Molecular Glue Degraders (MGDs) come in. Think of them as tiny, super-smart "glue" molecules. They don't just stick to the trash; they stick to a specific trash collector (an enzyme called CRBN) and then grab a piece of trash they've never seen before, forcing the collector to take it to the recycling plant.

This paper is like a massive, high-tech detective story where scientists tried to figure out exactly what kinds of trash these glues can pick up, and how to design better glues to catch specific targets.

Here is the breakdown of their adventure:

1. The Great "Trash Hunt" (The Screening)

Previously, scientists knew a few things these glues could catch (like a few specific bad proteins). But they wanted to know the whole list. Could they catch hundreds of different things?

• The Experiment: The team took a library of 960 different glue molecules (think of them as 960 different keys) and tested them on human cells.
• The Method: They used a super-powerful microscope (Mass Spectrometry) to take a "snapshot" of the cell's proteins before and after adding the glue. If a protein disappeared, it meant the glue had successfully tagged it for recycling.
• The Result: They found that these glues didn't just catch a few things; they caught over 230 different proteins. Even better, 124 of these were brand new discoveries that no one knew could be caught before!

2. The "Ghost" Problem (Filtering the Noise)

Here's a tricky part: Sometimes, when you break one thing, a whole bunch of other things fall apart because the system gets confused. It's like pulling a specific brick out of a wall; the whole wall might collapse, but you only wanted to remove that one brick.

• The Solution: The scientists used a special "mutant" cell line where the most common target (GSPT1) was unbreakable. By using this "unbreakable" cell, they could ignore the "wall collapse" effects and see exactly which proteins were being directly targeted by the glue. This helped them separate the real targets from the accidental victims.

3. The "Secret Handshake" (How the Glue Works)

How does the glue know which protein to grab?

• The Old Theory: Scientists used to think the glue only grabbed proteins that had a specific "handle" (called a G-loop) sticking out of them.
• The New Discovery: About half of the new proteins they found didn't have this handle. They were "G-loop-less."
• The Analogy: Imagine you thought a key only fit locks with a specific shape. But then you found out this key also fits locks that look completely different. The glue is much more versatile than we thought!
  • Example 1 (IRAK1): They found a glue that catches a protein involved in inflammation (IRAK1) even though it has no "handle." They figured out exactly which part of the protein the glue latches onto.
  • Example 2 (BCL6): They found a glue that catches a protein involved in blood cancer (BCL6). They proved it works by showing the glue physically grabs the protein in a test tube.

4. The "AI Detective" (Machine Learning)

With 960 different glues and 230 different targets, there was a mountain of data. How do you figure out the rules?

• The Tool: They used Artificial Intelligence (AI), specifically a type called "Interpretable Machine Learning."
• The Metaphor: Imagine you have a million different keys and a million different locks. You want to know: "What feature of the key makes it open this specific lock?"
  • The AI looked at the chemical "fingerprints" (the shape and atoms) of the glues.
  • It learned that certain shapes on the glue are like specialized tools. Some shapes are good for catching "Type A" trash, while other shapes are good for "Type B."
  • The Big Win: The AI didn't just guess; it found a specific chemical shape that acts as a "switch" to make the glue catch one target (WEE1) instead of another (CSNK1A1). This gives drug designers a blueprint: "If you want to catch Target X, add this specific shape to your glue."

5. Why This Matters (The Takeaway)

This paper is a game-changer for drug discovery.

• Before: Designing these drugs was like shooting in the dark. You'd guess a shape, hope it worked, and if it didn't, start over.
• Now: We have a massive map (a database called NeosubstratesDB) showing exactly what these glues can catch. We also have a "rulebook" (the AI model) that tells us how to tweak the glue to catch exactly what we want, without catching the wrong things.

In short: The scientists built a giant net, cast it into the ocean of human proteins, and pulled up a massive haul of new "trash" they can now clean up. They also figured out the secret code to make the net catch only the specific trash they want, paving the way for smarter, more effective medicines for diseases like cancer and Alzheimer's.

Technical Summary

1. Problem Statement

Molecular glue degraders (MGDs) are small molecules that reprogram E3 ubiquitin ligases (specifically Cereblon/CRBN) to induce the proteasomal degradation of non-native "neosubstrates." While CRBN-based MGDs (e.g., thalidomide derivatives) are clinically successful, the field faces several critical bottlenecks:

• Limited Neosubstrate Landscape: The full scope of proteins targetable by CRBN-MGDs is unknown. Current knowledge is biased toward a few well-characterized targets (e.g., GSPT1, CSNK1A1) and often relies on isolated domain screens that do not reflect full-length protein behavior.
• Lack of Predictive Design Principles: The structure-activity relationship (SAR) for MGDs is notoriously steep; minor chemical changes can drastically alter target selectivity. There is a lack of large-scale, public datasets linking chemical structures to proteome-wide degradation profiles to train predictive models.
• Differentiation of Direct vs. Indirect Effects: Distinguishing between direct neosubstrate degradation and secondary effects (e.g., translation inhibition caused by GSPT1 loss) or off-target toxicity remains challenging.
• G-loop Dependency: While many neosubstrates contain a characteristic "G-loop" degron, recent evidence suggests CRBN can target proteins lacking this motif, yet the mechanisms and scope of these "G-loop-independent" targets are poorly defined.

2. Methodology

The authors employed a high-throughput, target-agnostic screening strategy integrating deep proteomics, ubiquitinomics, and machine learning.

• Compound Library: A library of 960 CRBN-based MGDs (primarily IMiD derivatives) was screened.
• Cellular Model: HEK293 cells overexpressing CRBN were used to enhance sensitivity.
• Proteomic Screening (DIA-MS):
  • Cells were treated for 6 hours (duplicates per compound).
  • Data-Independent Acquisition Mass Spectrometry (DIA-MS) quantified >10,000 protein groups per plate.
  • Rescue Strategy: To mitigate secondary effects caused by GSPT1 degradation (which halts protein synthesis), a subset of potent degraders was re-screened in a line expressing a degradation-resistant GSPT1 mutant (GSPT1-G575N).
  • Dependency Validation: Hits were re-tested with the NEDD8-activating enzyme inhibitor MLN4924 (to block CRL activity) and in CRBN-knockout (CRBN-/-) cells to confirm CRL4CRBN dependency.
• Ubiquitinomics:
  • A 30-minute treatment window was used to capture early ubiquitination events before protein depletion occurs.
  • K-GG Remnant Profiling: Enrichment and quantification of di-Glycine (K-GG) peptides to map ubiquitination sites proteome-wide.
  • Integration of ubiquitination data with 6-hour proteomic depletion data to identify "productive" degradation events (ubiquitination + depletion).
• Mechanistic Validation:
  • Affinity Enrichment MS (AE-MS): Used to map interaction domains for specific targets (IRAK1, BCL6) using truncation mutants and point mutations (e.g., G-loop mutants).
  • Cellular Context: Validation in primary human PBMCs and Huh7 cells to assess physiological relevance.
• Machine Learning (iML):
  • XGBoost Models: Tree-based gradient boosting was used to predict protein abundance changes based on molecular fingerprints (binary vectors of substructural features).
  • Interpretability: Variable Importance in Projection (VIP) and SHAP (SHapley Additive exPlanations) analyses were used to identify specific structural motifs driving selectivity for targets like CSNK1A1, ZFP91, and WEE1.

3. Key Contributions

• NeosubstratesDB: The creation of an interactive, publicly accessible database containing the screening results for 960 compounds and >230 neosubstrates.
• Expanded Neosubstrate Landscape: Identification of 124 high-confidence, novel CRBN neosubstrates (previously unreported), expanding the known landscape significantly.
• G-loop Independence: Demonstration that ~53% of the newly identified neosubstrates lack a predicted G-loop degron, challenging the dogma that G-loops are strictly required for CRBN engagement.
• Design Principles: Derivation of specific molecular fingerprints that dictate neosubstrate selectivity, moving beyond empirical screening to rational design.
• Workflow Standardization: Establishment of a robust, high-throughput proteomic workflow (combining DIA-MS, ubiquitinomics, and ML) that can be adapted for other E3 ligases.

4. Key Results

• Screening Scale and Classification:
  • Out of 960 compounds, 526 were classified as "hits" (downregulating ≤10 proteins), 107 as "medium-active" (10–100 proteins), and 57 as "highly active" (>100 proteins, often driven by GSPT1 depletion).
  • The GSPT1-G575N rescue screen successfully unmasked neosubstrates that were previously obscured by translation defects.
• Novel Neosubstrates:
  • Identified 124 novel CRBN-dependent neosubstrates.
  • Diversity: Targets span kinases (IRAK1, ATR, WEE1), transcription factors (BCL6, ZFP91), chromatin remodelers, and zinc-finger proteins.
  • Clinical Relevance: 15% of identified neosubstrates are classified as strong selective dependencies in cancer (DepMap), and many are linked to non-oncological diseases.
• Mechanistic Insights:
  • IRAK1 (G-loop independent): Identified as a novel target. AE-MS mapped the interaction to the kinase C-lobe (residues 381–524), distinct from the conserved N-lobe. The degrader showed high selectivity over IRAK2/3/4.
  • BCL6 (G-loop dependent): Validated as a target. AE-MS showed that while a single G-loop mutation (G608N) did not abolish binding, a triple mutant (G598N/G608N/G636N) completely disrupted CRBN interaction, suggesting cooperative binding of multiple zinc fingers.
  • Off-targets: Identified CRBN-independent degradation of ER proteins (WLS, HMGCR) likely due to compound-induced ER stress.
• Machine Learning Insights:
  • XGBoost models achieved high accuracy (Pearson r up to 0.82) for frequently degraded targets.
  • Selectivity Determinants: Identified specific fingerprints driving selectivity. For example, fp1698 (a direct C-C bond from the isoindolinone core) was identified as a key determinant for WEE1 selectivity over CSNK1A1, recapitulating known medicinal chemistry findings.
  • Context Dependence: The study revealed that degradation is not driven by single "magic" fingerprints but by the molecular context (combinations of features). A promiscuous fingerprint might drive degradation in one context but not another depending on co-occurring structural features.

5. Significance

• Therapeutic Expansion: By identifying 124 new targets, including disease-relevant kinases and transcription factors, this work opens new avenues for targeted protein degradation therapies in oncology and other diseases.
• Rational Design: The integration of interpretable machine learning provides a framework for de novo design of MGDs. Researchers can now predict which chemical modifications will shift selectivity toward a desired neosubstrate while avoiding common off-targets (like CSNK1A1 or WEE1).
• Paradigm Shift: The findings challenge the strict reliance on G-loop motifs, suggesting a broader, more complex landscape of CRBN engagement that includes surface mimicry and multi-domain interactions.
• Resource Availability: The release of NeosubstratesDB and the standardized screening workflow democratizes access to high-quality TPD data, accelerating the development of next-generation degraders for other E3 ligases.

In summary, this paper establishes a comprehensive, data-driven framework for discovering and designing CRBN molecular glue degraders, bridging the gap between large-scale proteomic screening and rational chemical biology.