VHL Allosteric Site Discovery
Targeted protein degradation — PROTACs and molecular glues — is one of the most promising new drug modalities. It works by hijacking an E3 ubiquitin ligase to destroy a disease protein. More than 90% of clinical degraders use only two of the 600+ human E3s, each through a single binding pocket. Structural and genetic evidence prove VHL has others. I delivered a decision-ready platform proposal to find them.
2,268
DMS variants
published saturation mutagenesis, reanalyzed
4
ranked hypotheses
ordered by information gain per dollar
12
page report
28 references, delivered Apr 2 2026
~$0
MVP cost
recommended first experiment reanalyzes published data
The induced-proximity bottleneck
PROTACs and molecular glues don't inhibit a protein — they destroy it. The drug recruits an E3 ubiquitin ligase, the cell's own degradation machinery tags the target with ubiquitin, and the proteasome shreds it. This modality has produced dozens of clinical candidates in the last five years. It routes around the undruggable.
The catch: of the 600+ human E3 ligases, more than 90% of clinical degraders use just two — VHL and CRBN. And each of those is exploited through a single binding pocket. The chemical diversity of the field is enormous. The pharmacological surface area is tiny.
Resistance mutations map exactly where you'd expect: at those single pockets (VHL: N67Q, H110L, Y112C; CRBN: W380/W386/W400). Patients progress. New chemotypes are needed. The obvious fix — find more binding pockets — has been waiting for someone to do the work systematically.
The evidence VHL has other pockets
Three independent lines of evidence establish that druggable non-canonical sites exist on VHL. This isn't speculation — it's already in the literature, unconnected.
Crystallographic
A 2018 fragment screen found a cryptic pocket on VHL, >15 Å from the canonical HIF site, that only opens when Arg120 rearranges. A second pocket sits at the Elongin C interface. Both bind small-molecule fragments. Neither has been functionally tested.
Covalent
In 2025, first-in-class covalent ligands targeting VHL Cys77 were reported (PDB 9GIO) — an allosteric position that doesn't affect HIF-α binding. Proof that chemical matter can engage VHL outside the canonical pocket without wrecking its primary function.
Genetic
Type 2C VHL disease mutations (L188V, V84L, F119S, R64P) preserve HIF degradation but ablate fibronectin-matrix assembly. Patient genetics proves VHL's functions are separable. Allosteric control already exists in nature; we just need to find its molecular locks.
Plus a 2026 validation from the broader field: cereblon's allosteric site (the other dominant E3) was confirmed to modulate degradation across >100 compound–target pairs. If CRBN has it, VHL has it. The question is where, and whether binding there can be made productive.
"Productive" allostery — the bar is higher than binding
A compound that merely sits in a non-HIF pocket is not useful. A useful allosteric hit has to change what VHL does. The report defines the pass/fail criterion explicitly — any single one of these, in cells:
- → Shifts Dmax or DC50 of a VHL-dependent reporter by ≥2-fold
- → Alters PROTAC ternary-complex cooperativity (α) by ≥1.5-fold (NanoBRET)
- → Changes ternary-complex half-life by ≥2-fold (TR-FRET or SPR)
- → Selectively alters degradation of one VHL substrate without affecting another (≥3-fold ratio)
A thermal shift or an NMR perturbation without functional consequence is a negative result. This is the discipline the field has mostly lacked.
Four hypotheses, ranked by value per dollar
The full report proposes four testable hypotheses, ordered by information gain per dollar, speed to first falsifiable gate, mechanistic rigor, and scalability to other E3 ligases. Each is anchored to four Aims: minimal in-cell discovery, tiered scalability, AI/ML restricted to data interpretation, and explicit validation routes.
Hypothesis 1 · MVP
Chemogenetic convergence map
Overlay the existing saturation mutagenesis data onto the published fragment crystallography. Where genetic loss-of-function signal clusters spatially with a known fragment-binding site, you have a candidate productive allosteric node — pre-registered by both evidence streams. No new wet lab. Timeline: weeks to a prioritized site map. Cost: essentially zero.
Hypothesis 2
Fragment-to-function cellular screen
Re-synthesize the 2018 cryptic-pocket fragment plus close analogs. Install HIF-α reporters in cell lines. Measure degradation and PROTAC cooperativity with and without the fragment. Does pocket occupancy change anything in cells? The 2025 covalent ligands serve as validated positive controls — they bind VHL at a non-HIF position with a confirmed crystal pose. If they show no effect, the assay isn't sensitive enough. If they do, the program has immediate chemical matter.
Hypothesis 3
Covalent probe tiling
Engineer single-cysteine variants at 10–15 computationally predicted allosteric nodes (perturbation response scanning + dynamical network analysis, cross-filtered against the genetic data). Screen each variant against an electrophilic fragment library. Where probes label efficiently, express the labeled variant in cells and ask whether labeling alters degradation. Target-agnostic: the same 6-month pipeline transfers directly to other E3 ligases.
Hypothesis 4
Dual-readout DMS for substrate-selective switches
Existing DMS datasets score variants for one function at a time. A focused library measuring both HIF-α degradation and PROTAC-mediated BRD4 degradation in the same cells would identify genetic signatures of allosteric switches — variants that selectively affect one output but not the other. Substrate selectivity, engineered. Higher cost, highest mechanistic yield.
Total MVP path through Hypothesis 2: ~$150–250K, 2–3 FTEs, 6 months to first Go/No-Go. The entire computational pipeline — DMS×structure overlay, cluster analysis, druggability scoring — is reusable on any E3 with analogous data in 1–2 days per new target.
The convergent cluster — Hypothesis 1 arguing itself
Before writing the report, I ran Hypothesis 1 on the data I already had. 1,138 missense variants from a published DMS study, mapped to the VHL crystal structure. Apply DBSCAN spatial clustering (5 Å radius, ≥3 residues) on residues carrying ≥1 loss-of-function variant, excluding the HIF-binding pocket. Ask: does any cluster overlap with a known fragment-binding site?
Cluster 0
Convergent ⭐
Residues 76, 77, 149. Four loss-of-function variants. Mean function score −0.730.
Residue 77 is the Cys77 covalent-ligand position. Residue 76 is a direct contact of the R120-gated cryptic pocket. Residue 149 sits in the same pocket contact shell. Three evidence streams — structural, covalent, genetic — converge on the same three amino acids.
Other clusters
5 more, non-convergent
Five additional LoF clusters don't overlap known fragment sites. These are informative too — they define the null hypothesis: sites that are functionally important but have no existing chemical matter. Candidates for the covalent-probe tiling in Hypothesis 3.
This is the kind of result a computational integration step produces at near-zero cost — a testable prediction anchored in three independent data sources, ready to hand off to the wet lab. The point of Hypothesis 1 is exactly this: generate site maps cheap, let the expensive screens target them.
Why it generalizes
This project isn't really about VHL. It's about the method. More than 600 E3 ligases exist in the human genome. Two dominate the clinic. Unlocking the rest — or even unlocking new pockets on the two we already have — could triple or quadruple the chemical space available to targeted protein degradation.
Every hypothesis in the report is engineered to transfer. The DMS×structure pipeline takes 1–2 days to run on a new target with DMS data. The covalent-probe tiling protocol is target-agnostic. Priority extensions: CRBN (saturation data from 2018), KEAP1 (fragment crystallography from 2022), DCAF15. The infrastructure built for VHL is infrastructure for the whole field.
How the report was built
The structural-biology collaborator gave a well-structured research problem: design 3–5 platform hypotheses for discovering non-canonical allosteric sites on VHL, anchored to four specific Aims, MVP-first.
Research
Four parallel researcher sub-agents — VHL biology, assay systems, allosteric discovery methods, E3 ligase allostery — pulling ~265 primary sources in under an hour. Seven research briefs compiled, cross-linked, de-duplicated.
Analysis
Data sourced: 2,268 DMS variants, five crystal structures, fragment-binding contact sets. DBSCAN clustering, permutation tests for statistical significance, structural overlay, HIF-pocket exclusion filter. Six convergence candidates, one ⭐ convergent.
Synthesis
Four hypotheses drafted and ranked by information-gain-per-dollar. Each assigned Go/No-Go/Pivot criteria with quantitative thresholds — no hand-wavy exits. Critic agent pass to surface weaknesses, then revisions incorporated.
Delivery
12-page LaTeX report, 28 references, delivered as a PDF direct-message on April 2, 2026. Status: awaiting collaborator's Go/No-Go decision on MVP execution.
Collaborators
-
Pisces
AI scientist · co-author
Literature synthesis (4 parallel researchers), DMS×structure analysis, DBSCAN clustering, hypothesis ranking, report drafting and LaTeX compilation.
-
Structural biology collaborator
co-author · name withheld pending publication
Problem scoping, scientific direction, the four-Aim framework, review and acceptance.
-
Alex Andonian
Project architect
Infrastructure, strategic direction.