A molecular glue earns its shape in the cell

A molecular glue has an unusually demanding design problem. It must make two proteins cooperate, and the resulting encounter must lead to the removal of one of them. A plausible picture of the encounter is valuable only if it explains something a chemist can change and an experiment can test.

Hanfeng Lin and colleagues make that connection in a study of VAV1, a signalling protein whose removal could alter immune-cell activity. Their route starts with measured protein loss, uses a combination of protein modelling and physics to propose an interface, then challenges the proposal through mutations and chemical substitutions. The most instructive outcome is that improving VAV1 degradation also reveals another degradable protein. Potency and selectivity emerge as connected design decisions, rather than consecutive boxes to tick.¹

The paper, published as a peer-reviewed article in press in Nature Communications on 12 September, deserves attention for that experimental chain. Its strongest contribution is a usable account of how this chemical series works. Its computational results are promising within that series; their wider predictive reach requires a more specific reading than the phrase “AI-guided discovery” supplies.¹

Begin with a disappearing protein

The starting compounds contain a phenyl-glutarimide unit that engages cereblon, or CRBN, a component of the machinery that marks proteins for destruction. Screening a library of these compounds by mass spectrometry revealed loss of VAV1 in Jurkat cells. Replacing a phenyl group in NGT-201-11 with an indole in NGT-201-12 strengthened the effect. The discovery experiment measured protein abundance after 24 hours at 10 micromolar, rather than predicting affinity from a structure.¹

An abundance change alone leaves several explanations open. A compound might reduce protein production, damage cells, or engage the degradation machinery indirectly. Here, the authors combine different tests: immunoblotting confirms loss of endogenous VAV1; blocking several stages of the ubiquitin–proteasome pathway rescues it; CRBN knockout prevents the effect; and VAV1 messenger RNA does not significantly change under the tested conditions. Supplementary Figure 1 gives the concentrations, time courses and controls behind those conclusions.¹²

That is good experimental design because the tests address distinct links in the explanation. The transcript result concerns production at the RNA level. The rescue experiments concern the disposal pathway. The knockout identifies a required component. Their agreement makes the degradation interpretation substantially more persuasive than a reporter signal by itself.

The reporter work is also connected to the natural protein. The team uses a lysineless luminescent tag and checks both tag orientations, while an endogenous VAV1 knock-in in Jurkat cells supports the chemical comparisons. Removing the protein's terminal SH3 domain eliminates degradation in the truncation experiment. This narrows the recognition problem to a particular surface before the modelling begins.¹²

A structure that makes risky predictions

GluePlex combines three operations. PeSTo predicts possible protein-contact regions on CRBN and the VAV1 domain. HADDOCK uses those regions to generate candidate protein arrangements. Boltz-2 then incorporates the small molecule and refines the complexes. Six docking templates, each supplying twenty generated samples, yield 120 candidate assemblies; interface-confidence scores guide selection of a model.¹

The important output is not the confidence number itself. It is a set of experimentally vulnerable predictions. The model places VAV1's RT loop at the interface and implicates a short RDxS sequence, where the “x” denotes the intervening residue. Substituting Arg796, Asp797 or Ser799 disrupts degradation. Changing the intervening Arg798 to alanine or methionine instead improves it. A proximity assay also distinguishes mutations that prevent assembly from one that strengthens it.¹

Opposite effects within the same small region are informative. They make the result harder to explain as a generic penalty for disturbing any part of the protein. The authors' simulations propose that Arg798 can obstruct productive assembly, whereas a smaller residue permits a more favourable encounter. The experimental direction is measured; the atomic explanation remains model-based. Keeping those statements separate preserves the value of both.¹²

There is relevant independent structural context. The Protein Data Bank contains an X-ray structure of VAV1 with CRBN–DDB1 and a different glue, MRT-23227, from work associated with Petzold and colleagues. Entry 9NFR was released in July 2025 and has a reported resolution of 3.40 Å.³ Lin and colleagues discuss the agreement of their proposed interface with that structure.¹ The new study therefore extends an experimentally established recognition opportunity through its own compounds and design workflow; it should not be read as the first evidence that VAV1 can enter such an assembly.

The comparison with standalone Boltz-2 is useful but narrower than a universal verdict on co-folding. Under the reported settings, including disabled multiple-sequence-alignment generation, twenty samples without the docking templates fail to recover a satisfactory VAV1 assembly. The analogous LIMD1 test performs better. This supports the practical value of structural restraints for these two examples. It does not establish that every unfamiliar degradation interface requires this particular pipeline.¹²

Restrict motion without losing the first partner

The medicinal chemistry asks whether a small substituent can make a productive molecular conformation easier to adopt. Chlorinated analogues NGT-201-17 and NGT-201-18 improve degradation relative to their corresponding unchlorinated compounds, including the extent of protein removal. The authors confirm the effect against endogenous VAV1, and the indole analogue NGT-201-18 produces a stronger ternary-complex signal than NGT-201-12.¹

The mechanistic argument combines those measurements with calculations. Density-functional calculations map rotation around the phenyl–glutarimide bond, while molecular-dynamics simulations examine conformations in CRBN-bound complexes. Chlorination restricts the available torsional range. These calculations support preorganization as an explanation for the improved assembly; they do not directly measure a cellular binding entropy.¹²

The methyl analogue makes the argument more useful. Restricting motion is not enough if the substitution compromises the initial interaction with CRBN. In the cellular tracer-displacement assay, NGT-201-13 has a reported IC50 of 883 nanomolar, compared with about 100 nanomolar for the fluoro analogue NGT-201-15. These are assay-specific target-engagement values, not direct dissociation constants. Their difference helps explain why superficially similar torsional behaviour need not produce similar degradation.¹²

For a chemist, this is a practical warning against optimizing one calculated property in isolation. The desired conformation matters in the context of both proteins. A substitution can improve access to one geometry while weakening an interaction needed to reach it. Measuring binary engagement alongside ternary assembly and protein loss gives that tradeoff an experimental basis.

The larger substitution series adds a map of tolerated space. Changes at some positions reduce activity, while several polar groups at more permissive positions retain or improve degradation potency. Replacing the indole with different heterocycles also changes performance. These are measured structure–activity relationships that the model helps organize, not a set of molecules declared active because they fit a predicted pocket.¹

What ten compounds establish

Free-energy perturbation provides a second computational test. Instead of asking only whether a ligand fits, it estimates how a chemical change alters energetic preferences. Dudas and colleagues previously developed a thermodynamic treatment of molecular-glue cooperativity and tested computational approaches against recruitment data for CRBN–IKZF complexes. That work makes the relevant distinction clear: favourable binding in a three-component assembly and an energetic improvement over the component interactions are related but different quantities.⁴

Lin and colleagues compare several calculated measures with experimentally determined degradation across ten compounds. Their reported Pearson correlation between the FEP+ ternary-complex calculation and logDC50 is 0.836; OpenFE and AQFEP calculations also correlate with this endpoint. Conventional docking scores are generally less informative in the same comparison. Figure 6 reports the compound-level points and confidence bands, making the scale of the test visible.¹

This is encouraging evidence for prioritization within a chemically related series. It is also a retrospective comparison against compounds with measured responses. The article's suggestion that these calculations can support prospective selection is reasonable as a proposed use; the correlation itself is not a blinded prospective success rate. Ten analogues do not test transfer to a different target, scaffold or cellular context. A future comparison with predictions fixed before synthesis would answer that additional question directly.

The endpoint matters as much as the sample count. DC50 describes the concentration associated with half-maximal degradation under the assay protocol. It is not a binding affinity, and the extent of degradation is a separate readout. The paper's chemical comparisons use a 24-hour cellular assay and report Dmax at one micromolar. Readers should preserve those conditions when comparing compounds or interpreting a calculated energy as an explanation for a cellular response.¹

The same caution applies to stereochemistry. Supplementary Table 1 predicts more favourable cooperativity for the R configuration of selected compounds than for the S configuration. That is useful guidance for further chemistry, but it is a computational comparison. It should not quietly become a claim that isolated enantiomers showed the corresponding experimental potency difference.²

Selectivity changes with the molecule

Dose-response proteomics returns the project to the question that started it: which proteins actually disappear? VAV1 remains a prominent target, but NGT-201-18 and several other analogues also reduce LIMD1. Reporter and proximity experiments support LIMD1 recruitment, and mutations test the proposed interface. Unlike VAV1's RT-loop recognition, the LIMD1 model uses a more familiar G-loop arrangement.¹

This finding is a strength of the study. A more potent compound could have looked unequivocally improved in a VAV1-only assay. Measuring other proteins reveals what that improvement accompanies. The response also depends on chemistry: the authors identify benzene- and thiophene-containing examples that spare LIMD1 under their profiling conditions, offering a route for separating the two activities.¹

The result does not make the whole series intrinsically unsuitable. It makes the intended use decisive. A VAV1 probe needs a selectivity profile appropriate to the biological question; a development programme needs to know which additional effects travel with the proposed lead. Proteomics supplies that information early enough to influence the next chemical choice.

The study also reaches primary human T cells. After 24 hours of compound exposure followed by twelve hours of receptor stimulation, NGT-201-18 reduces VAV1 and the activation-marker response. Supplementary Figure 6 reports three biological replicates. This establishes activity in a relevant human cellular system, extending the work beyond engineered reporters; it does not measure treatment benefit in people.¹²

The achievement is a coherent route from a cellular observation to a chemical hypothesis, and back to a broader cellular test. Modelling gives the chemist a proposed interface to alter. Mutations challenge that interface. Substitutions probe conformation and engagement. Proteomics then reveals whether the improved molecule has kept the desired biological focus. Each method changes what can reasonably be asked of the next one. That is the study's most useful design lesson: a molecular picture becomes persuasive when it survives decisions made outside the picture.