Nature Biotechnology
(2026) Cite this article
Molecular glue degraders (MGDs), such as pomalidomide, induce degradation of non-native substrates by the cullin-RING E3 ligase 4 (CRL4) through its substrate receptor cereblon (CRBN). Here, to explore CRBN programmability, we tested whether reported CRBN–MGD substrates are part of a network of latent CRBN interactors, proteins capable of MGD-induced CRBN binding without detectable degradation. Leveraging a highly parallel protein complementation assay (GluePCA) to measure MGD-induced interaction between CRBN and zinc fingers, we identified ~210 zinc fingers bound to CRBN–pomalidomide, where top binders are already reported as degraded by dedicated MGDs. To map latent CRBN–MGD interactions proteome-wide and define the accessible CRBN interaction space, we combined artificial intelligence-derived protein surface queries (MaSIF-mimicry) with GluePCA. This pipeline identified 6 known and 43 novel CRBN–pomalidomide binders, including orthogonally validated hits. We find that these binders provide privileged starting points for MGD development. We expect this binding-focused workflow to be applicable to other MGD–E3 ligase systems, potentially extending the scope of this emerging drug class.
Molecular glue degrader (MGD) drugs, such as lenalidomide (Revlimid), have advanced the treatment of multiple myeloma1, targeting zinc finger (ZF) transcription factors (TFs) Ikaros and Aiolos for degradation through CRL4CRBN E3 ubiquitin ligase2,3,4,5,6,7. Lenalidomide derivatives are founding members of the family of MGD drugs, which engage target proteins through their conserved glycine β-hairpin α-turn (also known as G-loop). The Ikaros and Aiolos Cys2–His2 (C2H2) ZFs belong to the largest class of human TFs, with more than 675 members, all highly related in structure8,9. In recent years, a steadily growing number of ZFs, including Helios10,11, WIZ12,13, ZBTB16 (ref. 14) and SALL4 (refs. 15,16), have been identified as being degraded by CRL4CRBN through bespoke MGD derivatives. In addition to ZFs, other protein classes and folds, such as GSPT1 (ref. 6), VAV1 (ref. 17) and kinases (CK1α and others)4,7, also serve as MGD-dependent substrates of CRL4CRBN. The majority of these contain a G-loop motif, present in more than 2,500 proteins in the human genome18. Recently, non-G-loop motifs have been reported, further extending the target space17,19,20. The wide potential reach of these MGDs is contrasted by their apparent specificity in cellular degradation21,22. Hence, a key question for the development of novel therapeutics is which proteins, or folds, are potential CRBN–MGD targets and how specificity can be achieved.
Historically, MGD discovery has been driven by phenotypic screening of large chemical libraries23. Recent approaches include degradation reporters20,24, mapping CRBN–MGD interactomes by mass spectrometry17,19,25 and computational approaches17,19. Thus far, there is no integrated, scalable workflow available to comprehensively map the accessible CRBN–MGD interaction space proteome-wide. Here we present a highly parallel pipeline to computationally and experimentally identify latent CRBN–MGD interactions, including substrates beyond the canonical G-loop motifs. We utilized a combined workflow of highly parallel protein complementation assay (GluePCA) to measure MGD-induced interactions and surface-based computational tools (MaSIF-mimicry) to identify, proteome-wide, which proteins are already measurably engaged with CRBN–MGD through generic, existing MGDs and thereby constitute bona fide starting points for the development of specific degraders.
We hypothesize that many proteins found to be degraded through MGD derivatives are already CRBN-engaged through generic MGDs, such as pomalidomide, yet not sufficiently to drive degradation. Such latent proximity inducers are prime candidates for degraders, as evident from a number of studies on CRBN as well as other E3 ligases correlating cellular degradation with biochemical binding properties, reporting that an approximately two- to threefold increase in target binding to CRBN–MGDs suffices to convert CRBN–MGD-bound proteins into robustly degraded ones24,26. Hence, knowing which proteins latently interact with CRBN–MGD E3 ligase complexes, even if not yet degraded, will inform CRBN target discovery.
To identify CRBN–MGD-bound proteins on a large scale requires overcoming the limitations of traditional biochemical approaches, allowing us to profile thousands of protein interactions. We chose a highly parallel protein complementation assay27, in which a methotrexate-resistant murine dihydrofolate reductase (DHFR) is split into two fragments (‘DH’ and ‘FR’), which are fused to bait and prey proteins. The binding between the proteins reconstitutes DHFR activity, allowing yeast growth to serve as a proxy for interaction (Fig. 1a and Methods)28. We set out to test whether DH–bait and FR–prey complementation can also be induced by MGDs (GluePCA29). We first chose CRL4CRBN and the ZF family of TFs, in which the FR fragment was fused N-terminally to the ZF, and the DH fragment to CRBN (Fig. 1a). As a positive control, we tested an Ikaros construct (ZF2, residues 141–174), the minimal domain required for CRBN binding in the presence of pomalidomide24, which showed sustained growth, indicating binding between Ikaros ZF2 and CRBN in the presence of pomalidomide. By contrast, no binding was observed in the absence of pomalidomide, confirming that GluePCA is a suitable assay for the CRBN system (Fig. 1b and Extended Data Fig. 1a), albeit higher compound concentration is required in yeast than in mammalian cells. To quantify and compare growth between single-clone GluePCAs, we computed the area under each curve (AUC)30.
a, Schematic of the GluePCA and the vector used. Both FR and DH are N-terminally fused to the CRBN and ZF construct with a terminator in the middle. THF, tetrahydrofolate; DHF, dihydrofolate. b, Single-clone GluePCA in yeast, constitutively expressing DDB1ΔB, with FR–Ikaros ZF constructs and DH–CRBN, grown for 2,888 min in 100 µl selective media (SC-Ura/Ade/Met) containing pomalidomide and methotrexate (MTX). OD600 was measured every 15 min in a plate reader and normalized with a blank (biological replicates = 3; each point indicates the mean of the experimental replicates; error bars indicate 1 s.d.). Ki, inhibitory constant; ND, not determined. c,e, log2(FC) of sequencing read counts (pomalidomide or DMSO) and corresponding −log10(adj. P value) values for the single-ZF (c) and dual-ZF (e) libraries. Differential abundance was assessed using linear modeling and empirical Bayes-moderated two-sided t-tests implemented in limma. P values were adjusted for multiple comparisons using the Benjamini–Hochberg procedure (biological replicates = 3). Colored points are binders degraded with pomalidomide. Red circles indicate binders degraded only by bespoke MGDs (left to right: ZBTB11, Helios, FIZ1). d,f,g,h, 293T cells expressing the indicated ZF–GFP–IRES–mCherry (d and g: single-ZF constructs; f and h: dual-ZF constructs) construct were treated for 2 h, 4 h and/or 6 h with DMSO or pomalidomide at the indicated dose. Cells were analyzed by flow cytometry to quantify the GFP over mCherry ratio, which was normalized to DMSO for each treatment (biological replicates = 3; bar heights indicate the mean of experimental replicates, and error bars indicate 1 s.d.).
Next, we assessed the dynamic range of the assay with different lengths of Ikaros constructs, spanning ZF1 to ZF3 (Extended Data Fig. 1b), and Ikaros mutants. Ikaros ZF2–ZF3 (residues 141–195) and ZF1–ZF2–ZF3 (residues 117–195) showed fast growth analogous to stronger binding, compared with ZF2 alone. An IKZF1-Q146H mutant shows weaker growth by two- to threefold, while the G151N mutation showed no growth above background (Fig. 1b). These growth rates mirror previous in vitro biophysical half-maximal inhibitory concentration (IC50) measurements for CRBN complexes (IC50 in probe displacement: ZF1–ZF2–ZF3: ~460 nM; ZF2–ZF3: ~470 nM; ZF2: ~73,540 nM)24. GluePCA thus provides a rapid and sensitive readout of MGD-induced CRBN binding, matching the dynamic range of biophysical assays while eliminating the need for protein purification.
Encouraged by these results, we expanded GluePCA to pooled high-throughput screens. We screened a library of 8,363 individual C2H2 ZFs across 1,407 uniquely expressed genes and quantified binding propensities using next-generation sequencing (Extended Data Fig. 1b). This led to the identification of numerous ZF targets previously reported to be degraded by pomalidomide, including Ikaros, Aiolos, RNF166, WIZ, ZFP91, ZNF276, ZNF653, ZNF692 and ZNF827 (refs. 3,5,13,24,31) (Fig. 1c). Comparing our data with published single-ZF degradation screens, we identified 9 of 11 previously reported ZFs24. Among the top binders, we also identified SALL4, a validated pomalidomide target15,16, which eluded previous single-ZF degradation screens24. ZNF445 (ref. 20) was recently identified in a dual-ZF degradation screen; notably, we also observed degradation of its single ZF in mammalian cells (Fig. 1d, Extended Data Fig. 1c and Supplementary Fig. 1). Of the top 16 hits, the remaining two ZFs (PATZ1 and ZNF687)31,32 have been reported to be degraded in cells as full-length (FL) proteins. We also observed Helios, for which biochemical recruitment with pomalidomide had been previously observed6,24; yet, degradation required compound modification10,11. To further experimentally validate GluePCA hits, we posited that true hits would show dose-dependency. Using a log2(fold change (FC)) threshold > 2, defined by deviation from the diagonal (log2(FC) 15 µM = log2(FC) 5 µM), indicating a compound-dependent effect, and FDR-adjusted P value < 0.01 at 15 µM, all 201 hits showed dose-dependent binding (Extended Data Fig. 1d). These findings experimentally confirm previous predictions that more than 100 ZFs are compatible with CRBN–pomalidomide binding, with top binders being subjected to degradation24.
The majority of ZFs in proteins in the human proteome are found as multidomain tandem repeats of single-ZF domains33 (Extended Data Fig. 1b). While the ZF contacting the CRBN–MGD interface is sufficient for binding and degradation, adjacent ZFs can strengthen the interaction and trigger robust degradation20,24. The contribution of these flanking ZFs remains unclear. We refer to ZFs directly engaging the CRBN–MGD interface as main ZFs (mZFs), and preceding and subsequent ZFs as accessory ZFs (aZFs).
To dissect how multiple ZFs engage CRBN–MGD complexes, we designed a library containing all tandem C2H2 ZFs in the human proteome with a 6- to 7-residue spacing (Extended Data Fig. 1b), yielding 5,293 dual ZFs across 713 unique genes. We then experimentally defined true binders, by repeating GluePCA at a lower pomalidomide concentration. A total of 285 dual-ZF binder constructs with log2(FC) > 2 showed dose-dependent binding (Extended Data Fig. 1e). Overall, the top dual-ZF hits aligned with the single-ZF screen, including ZNF692, ZNF827, Ikaros, Aiolos, ZNF276, SALL4, PATZ1, ZFP91 and ZNF653, yet the rank order changed (Fig. 1e and Extended Data Fig. 1f). We also identified previously unreported pomalidomide-recruited ZFs: ZNF852 and ZNF341, a TF regulating STAT3 (ref. 34). In cellular ZF degradation reporters, both showed robust degradation with pomalidomide (Fig. 1f, Extended Data Fig. 1c and Supplementary Fig. 1). Furthermore, we observed that REST, a transcriptional repressor implicated in multiple cancers35,36,37,38, bound CRBN–pomalidomide as a single ZF. While REST ZF2 alone was not degraded, the dual ZF2–ZF3 construct was degraded in the presence of pomalidomide (Fig. 1g,h and Supplementary Fig. 1), underscoring the importance of aZFs in driving degradation.
Having established GluePCA as a highly parallel binding assay to identify CRBN–pomalidomide binders across thousands of ZFs, we confirm the existence of a continuum of ~200 dose-dependent CRBN binders. The top hits either were previously reported as degraded by pomalidomide or showed robust degradation only with optimized compounds (that is, FIZ1 with compound 1 (ref. 39), ZBTB11 with WJ-01-306 (ref. 40), Helios with NVP-DKY709 (ref. 11) and ALV2 (ref. 10) (Fig. 1c)). The findings for REST further illustrate that the ZF degron that governs binding to CRBN–MGD complexes is more complex than a single ZF and that additional ZF interactions facilitate degradation.
To dissect how multiple ZFs engage the CRBN–MGD complex, we generated a deep mutational scanning (DMS) library of 89,918 single-point mutations from 124 dual-ZF binders. We subjected this library to GluePCA. For most constructs, one ZF was dominant, and mutating this ZF showed the most deleterious effect on binding (Fig. 2a and Extended Data Figs. 2a,b and 3a,b). This dominant ZF typically also scored better than its neighboring partner ZF in single-ZF GluePCA (Extended Data Fig. 3c) and was hence assigned mZF. DMS analysis revealed that while some ZFs were largely unaffected by N- or C-terminal aZF mutations, others showed enhanced or reduced binding (Fig. 2a and Extended Data Fig. 3a). The data suggested that individual mutations can affect both ZF fold integrity and CRBN interaction. Averaging the aZF mutation profiles did not reveal a conserved, easily identifiable binding patch (Extended Data Fig. 3b).
a, Example heatmaps of the deep mutational screen for ZNF516 and ZBTB16 (DMS) of the DMS library in GluePCA. b, Cryo-EM structure of SALL4ZF123 bound to DDB1ΔB~CRBN and pomalidomide (POM). CTD, carboxy-terminal domain; HBD, helical bundle domain; BPA and BPC, β-propeller A and C, respectively. c, Close-up of SALL4 ZF1, ZF2 and the helix fitted into our cryo-EM density (gray). d, Single-clone GluePCA in yeast with SALL4 ZF2Helix and ZF2 construct. ZF2 contains residues PQVKA (residues 433–436) preceding the helix, which were shown to benefit binding (biological replicates = 3; each point indicates the mean of the experimental replicates; error bars indicate the s.d.).
To identify potential binding sites for the aZF on CRBN, we reconstituted two dual ZFs: SALL4 aZF1–mZF2 and Ikaros mZF2–aZF3. Cryo-electron microscopy (EM) structures of these ZFs bound to DDB1ΔB, CRBN and either SALL4ZF1–ZF2–ZF3 or IkarosZF2–ZF3 in the presence of pomalidomide had an overall resolution of 3.5 Å (Extended Data Figs. 4a–e and 5a–e and Supplementary Table 1). The SALL4 structure had aZF1–mZF2, in a position consistent with the previously reported X-ray structure41,42,43 (Fig. 2b). We observed additional density located C-terminal of mZF2, resembling a sole α-helix absent from the crystallographic construct, which was not previously observed (yet predicted in AlphaFold2 (AF2)44) (Fig. 2c). Single-clone GluePCA assays showed that the presence of the helix strengthened binding ~1.4-fold (Fig. 2d). The IkarosZF2–ZF3 map also revealed density beside mZF2, which we assigned to aZF3 (Extended Data Fig. 3d). While we see the SALL4 N-terminal aZF1 interacting with CRBN amino-terminal domain (NTD) residues F102, I152, H353 and F391, the Ikaros C-terminal aZF3 interacts with CRBN NTD residues Q86 and H102 (Extended Data Fig. 3e,f). On the basis of the SALL4 and Ikaros structure, we conclude that the aZF N- or C-terminal of mZF can comprise up to three consecutive ZFs (aZF–mZF–aZF). In addition, our DMS screen suggests that the aZFs (N- or C-terminal) can either facilitate or weaken binding, as would be expected for neomorphic CRBN–MGD interactions45 (Extended Data Fig. 3a). While the mZF binding mode is relatively conserved, the aZF contributes through multiple and apparently non-conserved binding modes in line with a neomorphic, evolutionary non-conserved, binding mode.
