Drug Discovery and Development · global
Finding Molecular Glue Targets No Longer Relies on Luck Alone: Proteome-Wide Platform Maps the CRBN Degradation Landscape
The research team connected geometric deep learning with high-throughput experiments into a single screening pipeline, identifying 43 new candidate binding proteins and revealing how adjacent zinc fingers influence degradation; however, there is still some distance between “can bind” and a safe, selective drug.
Molecular glues can bring otherwise unrelated proteins close to the cell’s clearance machinery, prompting the degradation of disease-causing proteins. This drug strategy has demonstrated value in diseases including multiple myeloma, yet the search for new molecular glues has long had an element of chance: researchers often observe a drug effect first and only then trace which protein was degraded. A study published in *Nature Biotechnology* now presents a scalable platform that seeks to turn this work into a more systematic target-search process.
The study used cereblon (CRBN), the substrate receptor of an E3 ubiquitin ligase complex, and the drug pomalidomide as its test system. In an assay called GluePCA, two inactive enzyme fragments were fused separately to CRBN and candidate proteins. When a molecular glue brings the two into proximity, enzyme function is restored, allowing yeast growth and sequencing readouts to serve as binding signals. The team began by screening 8,363 individual zinc-finger structures and 5,293 adjacent zinc-finger pairs, identifying more than 210 zinc fingers capable of binding the CRBN–pomalidomide complex.
This map not only expanded the candidate list but also revised the understanding of recognition rules. Deep mutational scanning showed that the primary zinc finger directly contacting the molecular glue is not the only determinant; neighboring zinc fingers on either side can strengthen or weaken overall binding and can even convert mere contact into actual degradation. The newly identified ZNF852 and ZNF341 could undergo pomalidomide-promoted degradation in a cell-based reporter system. A single zinc finger from REST only bound, while degradation appeared only after an adjacent zinc finger was included, providing a concrete demonstration of this context effect.
To extend the search across the entire proteome, the team then used MaSIF-mimicry to analyze the geometric and chemical features of protein surfaces, looking for structures whose shapes mimic known CRBN substrate interfaces. From the predictions, the researchers selected 1,959 truncated candidate protein structures for experimental testing with GluePCA, ultimately recovering 6 known substrates and identifying 43 new candidates showing pomalidomide-dependent binding. Some candidates even lacked the canonical G-loop structure previously often considered necessary, suggesting that the explorable target space may be broader than existing rules indicate.
The study also used RNF39 to demonstrate how this binding map could connect with chemical development. The team screened compounds against an RNF39 domain and found molecules that significantly enhanced its interaction with CRBN, one of which induced degradation of RNF39’s RING finger domain in cells. Blocking the relevant ubiquitination pathway rescued the degradation phenotype. This is not yet a drug candidate, but it shows that a “potential binder” can serve as a starting point for subsequent molecular-glue optimization.
The core value of this work lies in connecting computational prediction, parallel binding assays, and analysis of structural rules, rather than merely producing a list of AI-generated candidates. The study database GSE298800 contains 42 samples covering single zinc fingers, paired zinc fingers, deep mutational scanning, and treatment conditions involving pomalidomide and another molecular glue, ALV2. Processed data are available for download, and the raw sequencing data have also been linked to the Sequence Read Archive, enabling external researchers to reproduce the analyses.
However, this map primarily describes “whether binding can occur,” which does not mean that the full-length protein will necessarily be cleared effectively and selectively in human cells. GluePCA uses a yeast overexpression system, while the proteome-wide screen used truncated domains. Intracellular expression levels, full-length protein conformations, tissue differences, unintended substrates, and toxicity could all alter the final outcome. For the platform to become a predictive drug-design tool, its transferability must still be demonstrated across more E3 ligases, different molecular-glue scaffolds, disease models, and animal studies.