Every VHL PROTAC uses the same warhead. We generated 279 alternatives that don't.

If you have worked on targeted protein degradation, you know the VHL story. Essentially every VHL-recruiting PROTAC in the literature is built on one chemical idea: a (4R)-hydroxyproline that mimics hydroxylated HIF-1α Pro564, packaged as VH032 or a close analogue. It is a genuinely good binder. It is also a monoculture.

That warhead is expensive to carry. It contributes a large polar surface area to a molecule that is already deep in "beyond Rule of 5" territory, and PROTAC permeability is limited by exactly that — polar surface area and hydrogen bond donors. Assembled VHL degraders in the PDB land around TPSA ≈ 209 Ų. You are paying for the warhead twice: once in molecular weight, once in permeability.

So we asked a simple question: does a VHL binder have to be a proline?

279 designs that occupy the VH032 site without the proline

We released VHL_NonProline_Binder_Designs_9GIO — 279 generatively designed molecules docked into the VHL substrate-recognition pocket of the VHL–ElonginC–ElonginB complex, using PDB 9GIO as the receptor. 9GIO is a convenient starting point because it contains VH032 itself, so the reference sits right there in the same structure.

278 of 279 contain no 4-hydroxypyrrolidine, and none contain a prolinamide. They are 252 distinct Murcko scaffolds across 274 unique molecules. Superposed onto 9GIO they overlap the VH032 site to within 0.13 Å.

The property picture is what we hoped for:

228 of 279 fall below VH032's TPSA — a median reduction of 22 Ų, one fewer hydrogen bond donor, two fewer rotatable bonds. Because the warhead's polarity is carried straight into the assembled degrader, that reduction propagates into the full molecule on the axis where VHL PROTACs are most constrained.

The part we didn't expect

The interesting result was not the property table. It was how these molecules hold the pocket.

We ran the same interaction analysis on VH032 in 9GIO and on all 279 designs:

VH032 binds as a substrate mimic, and its affinity rests on three sidechain hydrogen bonds. Sidechain donors rotate and cost entropy on binding; one of the three is geometrically marginal even in the crystal.

The designs converged instead on a backbone hydrogen bond — fixed by the fold, not by a rotatable sidechain — plus aromatic stacking against a pocket lined with Tyr98, Tyr112, Trp88, Trp117, Phe76 and Phe91.

And here is the bit worth sitting with: VH032 makes no π-stacking at all. In a pocket that aromatic, the substrate-mimetic warhead leaves that binding energy entirely on the table. Only 4 of the 279 designs hydrogen-bond Ser111 — and that turns out to be the point rather than a defect. Once you have a conserved backbone anchor and good stacking, reproducing the hydroxyproline contact is not required, and giving it up is exactly what frees the scaffold from the proline core and its polar-surface cost.

What this is, and what it isn't

These are computational designs, not measured binders. No affinity, no ternary complex, no degradation data. The docking scores rank poses; they are not affinities. A different binding mode is also an untested binding mode — the His110-plus-stacking anchor is mechanistically sound and more geometrically reliable than three rotatable sidechain contacts, but whether it delivers VH032-like affinity is an open experimental question. The lower TPSA is partly traded for lipophilicity (median cLogP 3.4 vs 2.2), and everything was docked into a single rigid receptor.

We are releasing them as hypotheses, with the receptor, all 279 complexes, an SDF with verified 3D poses, and per-design interaction counts so every claim above can be checked against the rows rather than taken on trust.

More design sets

This dataset is part of the Technetium GA-II GenerativeAI Design Sets collection — 18 open design sets spanning novel and under-drugged targets, including molecular glues, PPI blockers, GPCR ligands and E3 binders. Molecules are generated by the Technetium TC-43.ai engine; structural analysis and curation are done with Claude Code. All CC-BY-4.0.

If you work on degraders and want to pressure-test the non-proline hypothesis, the data is there. We would genuinely like to know if it holds up.