Tc-43/NEK7_Molecular_Glue_Designs_9NFQ

NEK7 Molecular Glue Designs — CRBN/DDB1/NEK7 (PDB 9NFQ)

434 de novo molecular glue candidates designed into the CRBN–NEK7 neosubstrate interface of PDB 9NFQ (CRBN–DDB1–NEK7 with MRT-3486), generated by the Technetium TC-43.ai agentic AI engine — a physics-based generative platform with reinforcement-learning-guided optimisation. Target selection and structural analysis by Claude Code.

CRBN-recruiting molecular glues degrade a neosubstrate by creating a protein–protein interface that does not otherwise exist. NEK7 is a NIMA-related kinase and an established CRBN neosubstrate; 9NFQ is its only ternary complex structure.

The designs engage the tryptophan cage as an aromatic cage

CRBN's Trp380 / Trp386 / Trp400 triad is usually described as the pocket that accepts a glutarimide. Thalidomide and its analogues present a non-aromatic imide ring to it: in 7BQU, 8OIZ, 9FJX, 9NFQ and 9NFR the ring stacking on W380 (3.6–4.0 Å) and W386 (4.0–4.4 Å) is the glutarimide itself. Their only aromatic ring — the phthalimide benzo — points out of the pocket, so none of the clinical IMiDs makes an aromatic π–π stack with W386.

These designs do, and they use the cage in both geometries available to it:

tryptophan	face-to-face (parallel)	face-to-edge (T-shaped)	median centroid distance
W380	222 / 433	1	3.87 Å
W386	239 / 433	19	3.82 Å
W400	0	227 / 433	4.86 Å

W400 is engaged exclusively edge-on, W380 and W386 almost exclusively face-on — the three tryptophans form a cage that supports both stacking modes at once. 221 of 433 designs engage all three; 226 make face-to-face and face-to-edge contacts simultaneously.

Both interactions are present: hydrogen bonds and π-stacking

MRT-3486 itself makes only four hydrogen bonds in 9NFQ, and its two CRBN anchors are backbone, not side chain: His378 C=O at 2.88 Å and Trp380 N–H at 3.25 Å, plus Asn351 ND2 (3.05 Å) and NEK7 Arg35 NH1 (2.53 Å). The tryptophan cage supplies stacking; the 378–380 backbone supplies the hydrogen bonds.

The designs reproduce both halves:

	designs
≥1 hydrogen bond across the interface	433 / 433 (median 3, up to 7)
canonical pair recovered — His378 C=O and Trp380 N–H	323 / 433
hydrogen bond to Trp380 N–H	365 / 433
hydrogen bond to His378 C=O	331 / 433
hydrogen bond to NEK7 (Gly57, Glu37, Arg35, Arg50, Asp56)	167 / 433
π-stacking on all three cage tryptophans	221 / 433
canonical H-bond pair AND all three π-stacks	196 / 433

interaction_score is a 0–1 composite over six terms: canonical H-bond pair (0.20), total H-bond count (0.10), π-stacking across the Trp cage (0.20), cage occupancy (0.10), NEK7 contact recovery (0.20), CRBN contact recovery (0.20). Rows are sorted by it. The top 20 is unchanged when either the H-bond or the π-stacking weight is doubled (20/20 overlap), and 16/20 survives doubling the NEK7 term.

Not sorted by Vina. Over this set Vina correlates with the interaction score at Spearman +0.17, and the single best Vina value (−15.20 kcal/mol) ranks 377th of 434.

Synthesizable and drug-like

	value
synthetic accessibility (Ertl SA)	median 3.81, range 2.38–5.72	none above 6
Lipinski violations = 0	409 / 434
MW	423 (390–451)	8 above 500
cLogP	2.05 (−0.3 to 4.2)	16 above 5
TPSA	109 Ų
QED	0.56 (0.40–0.69)
Fsp3	0.35	4.6 rings, 5 rotatable bonds

Interface recovery

Scored against MRT-3486's own contacts in 9NFQ (12 CRBN residues, 9 NEK7 residues):

• 433 / 434 place their centroid within 5 Å of the native ligand; 198 within 2 Å
• 378 / 434 contact all five tri-Trp cage residues
• 426 / 434 bridge the CRBN cage and NEK7 — the defining requirement for a glue
• median recovery: 10/12 CRBN contacts, 5/9 NEK7 contacts; the best design recovers 9/9

complementarity_score is a 0–1 composite of cage occupancy (0.30), NEK7 contact recovery (0.35) and CRBN contact recovery (0.35). Rows are sorted by it, not by Vina — over this set Vina correlates with complementarity at Spearman +0.17, and the single best Vina score (−15.20 kcal/mol) ranks 377th of 434 on interface recovery.

Files

file	contents
designs.csv	434 rows × 31 columns — the index the dataset viewer reads
ligands.sdf	433 poses with bond orders, stereochemistry from 3D, and all annotations
receptor_9NFQ.pdb	DDB1 (A) + CRBN (B) + NEK7 (C) + Zn, identical for every run
structures/	434 per-design ligand poses; concatenate with the receptor for a complex

The receptor is byte-identical across all 434 runs and matches deposited 9NFQ exactly (CRBN Cα RMSD 0.000 Å without superposition), so it ships once rather than 434 times. CRBN numbering is 1:1 with UniProt Q96SW2. One design is absent from the SDF — bond-order assignment from its SMILES failed — but is present in designs.csv and structures/.

Honest limits

These are computational designs, not experimental hits. No synthesis, no binding or degradation data. Poses come from docking into a single rigid receptor conformation, so induced fit is not modelled. The cage motif occupying the pocket is most often a 2-pyridone (238) or lactam (54) rather than a glutarimide (21) — non-glutarimide CRBN binders do exist in the PDB, but this chemotype is unvalidated for CRBN and the stacking geometry reported above is a hypothesis to test, not an established binding mode. 110 of 434 designs carry a formal charge, 11 of them a dication, which is a permeability liability worth filtering on. Hydrogen bonds are assigned from heavy-atom donor–acceptor distance (2.5–3.5 Å) because the receptor carries no hydrogens, so donor–H–acceptor angles are not checked. Stacking geometry uses ring-centroid distance, interplanar angle and lateral offset.

Citation

Designs: Technetium TC-43.ai engine, TechnetiumTx. Analysis: Claude Code. Receptor: RCSB 9NFQ — Petzold et al., Science 389:eadt6736 (2025).