Closing the degrader design loop with pull-down oracles

Drugs that work by inhibition must jam a protein's active site, but many disease proteins have no such site. A different pharmacology avoids this: a PROTAC is a two-headed molecule with one head that glues the disease protein to a cellular destruction enzyme (an E3 ubiquitin ligase), so the cell's own waste-disposal machinery shreds the target. Machine learning tools now design the chemical linker that joins the two heads, and some of them are impressively fast — but not one of them has ever checked whether its molecules actually pull the target and the destruction enzyme together inside a cell. Every published degrader-design pipeline scores its molecules purely by computation. The proposal here is to close that loop: run a small, iterated design campaign in which the scores that guide molecule generation come from two experiments — a pull-down assay that measures whether the target physically co-precipitates with the E3 ligase, and a mass-spectrometry readout of which cellular protein complexes are collaterally destroyed, summarized at the pathway level.

Two observations make this worth doing now. First, the generative side is ready: CupOFLATTE assembles linkers directly in the 3D gap between two pinned binding heads and, on the BRD9–VHL benchmark, achieved 72% assembly validity where standard linker generators failed at 0% (CupOFLATTE) — yet its authors state outright that wet-lab synthesis and testing are the necessary next step, and its claimed "1.7x to 1.9x more active" molecules exist only as predictions from another model. Second, the measurement side has never been connected to design: co-immunoprecipitation (Co-IP) — an assay where an antibody against a chosen "bait" protein drags out of a cell lysate everything physically stuck to it — is the standard way to prove that an E3 ligase physically contacts its substrate (Curcumol study), and unbiased Co-IP followed by mass spectrometry plus pathway enrichment analysis has been used to map whole interactomes (Sun et al.) — but only to describe natural biology, never to steer a molecule generator.

Core Claim

The claim is that a molecule generator trained on a few dozen wet-lab measurements per round will find degrader leads that purely computational ranking misses, because the binding of the two heads and the geometry of the linker do not add up linearly to what actually happens in a lysate. Supporting evidence already exists in fragments: DeepPROTACs' ablations showed that removing any one of the three components — ligase pocket, warhead, or linker — badly hurt prediction, meaning degradation is an emergent property of the whole assembly (DeepPROTACs); and the lead-optimization literature documents that favorable computed binding energies routinely fail to translate into cellular activity (ScaffoldGVAE). More pointedly, the field's own predictors disagree about why molecules work: DegradeMaster builds explicit 3D geometry and reaches an AUROC of 0.8825 on the PROTAC-8K benchmark, while DegradeQuery ignores 3D coordinates entirely, reasons over target–ligand–E3 relations, and scores higher at 0.9065 (DegradeMaster; DegradeQuery). When two state-of-the-art oracles disagree on their internal logic, ranking candidates by either one and synthesizing the top hits is a bet, not a measurement. Our loop replaces that bet with data, and as a byproduct arbitrates the geometry-versus-relations dispute.

Research Program

We build a closed-loop campaign on one well-instrumented system, BRD9 recruited by the VHL ligase, chosen because the published numbers that define the frontier all live there: CupOFLATTE's 72% assembly validity and DegradeMaster's 88.33% BRD9 degradation-prediction accuracy (CupOFLATTE; DegradeMaster). The pipeline has three parts.

The generator reuses CupOFLATTE's coupled discrete flows, which grow the linker atom-by-atom in the space between the pinned BRD9 warhead and the VHL ligand, with validity gating that rejects broken valences and toxic substructures. We deliberately keep its human-specified attachment points as a controlled variable rather than trying to fix that limitation too — the question this proposal answers is whether experimental feedback matters, not whether every design choice can be automated at once.

The oracle is tiered, which is what makes a wet-lab loop affordable. Tier one, applied to every synthesized candidate, is a targeted Co-IP: pull down BRD9 from treated cells and blot for VHL, giving a quantitative readout of whether the molecule actually assembles the physical target–degrader–ligase complex in a cellular environment. Tier two, applied to the best performers, is total-target Western blotting plus degradation kinetics ($DC_{50}$ and $D_{\max}$, the concentration at which half the target is degraded and the maximum fraction degraded). Tier three, applied to finalists, is unbiased Co-IP-MS: pull down the target's complex and identify by mass spectrometry everything that co-precipitates, then run overrepresentation analysis — the standard statistical test of whether proteins in a recovered list belong to particular pathways more often than chance — against Gene Ontology and KEGG annotations, using tooling in the spirit of funkea's unified engines (funkea). Tier three yields the safety signal: which essential cellular complexes are being collaterally depleted alongside the target, the bystander-degradation problem that no current database records (PROTAC-DB logs binary potency only).

The learner turns these measurements into preferences. Rather than asking a model to regress exact potencies from a handful of noisy numbers, we reuse the preference-learning idea from POLO, which converted optimization trajectories into $O(T^2)$ pairwise comparisons and hit an 84% success rate with only 500 oracle calls in conventional lead optimization (POLO) — but here the preferences are laboratory outcomes, not cheap property oracles. Each round, synthesized candidates are ranked pairwise by a reward that combines the pull-down signal with a penalty on off-target pathway depletion:

$$r(c) \;=\; s_{\text{IP}}(c) \;-\; \lambda \, s_{\text{off}}(c)$$

where $s_{\text{IP}}(c)$ is the normalized co-precipitated-ligase signal for candidate $c$, $s_{\text{off}}(c)$ is the enrichment score of off-target functional modules in its Co-IP-MS profile, and $\lambda$ is a trade-off weight. The generator is re-fit to prefer high-reward molecules, the predictor is retrained on all accumulated labeled data, and the next batch is chosen by acquisition over the joint model — explicitly probing the regions where the 3D-geometric and relational predictors disagree, since disagreement marks where the models are uncertain and the data is informative.

A note on assay honesty: Co-IP has known failure modes — lysis can create artificial contacts between proteins that are separated in the living cell, and wash cycles can strip away the very transient contacts we care about. We therefore cross-check tier-one positives with proximity ligation assays, which fluoresce only when two proteins sit within about 40 nm inside intact cells, following the orthogonal-validation practice established in the Co-IP literature. A candidate that scores in Co-IP but not in situ is flagged as a possible lysis artifact and down-weighted rather than promoted.

Proposed Experiments

The computational prelude trains the two predictors on PROTAC-8K (8,636 target–ligand–E3 tuples, of which only 1,502 carry experimental degradation labels) and verifies we reproduce the published ordering before any synthesis: DegradeQuery at AUROC 0.9065 and accuracy 0.8500, DegradeMaster at AUROC 0.8825 (DegradeQuery; DegradeMaster). Ternary-structure scoring for the geometric predictor uses DeepTernary, which reaches a median DockQ of 0.65 on the 22-case experimental-structure benchmark in under than a second on a GPU and whose predicted buried surface area has been shown to correlate with degradation potency (DeepTernary); we will re-test that BSA–potency correlation against our measured $DC_{50}$ values, since it currently rests on literature data alone.

The loop itself runs three rounds of 24 synthesized BRD9–VHL candidates each, a budget chosen to be small enough to be realistic for an academic lab and large enough for pairwise preference learning. The primary endpoint is hit-rate efficiency: the fraction of synthesized compounds that are confirmed degraders ($DC_{50} \le 100$ nM and $D_{\max} \ge 80\%$, the same binary criterion DeepPROTACs used to label its 2,832-sample dataset) per round, compared across the loop and the baselines below. Secondary endpoints are the correlation between tier-one Co-IP signal and tier-two $DC_{50}$ — if pull-down strength predicts potency, the tiered oracle is validated as a cheap screen — and the pathway-level selectivity score from tier three. The round-three model is then evaluated prospectively: does its top-ranked unsynthesized candidate succeed when made, at a rate better than the frozen baselines would achieve? That prospective test, not retrospective metrics, is the decisive check.

The mechanistic sub-study uses the accumulated data to arbitrate the field's open contradiction: whether 3D geometry or relational context dominates degradation outcomes. Candidates where DeepTernary's geometric scores and DegradeQuery's relational scores disagree strongly are exactly the ones the acquisition function oversamples; the wet-lab outcomes of those candidates determine which predictor's internal logic better matches cellular reality.

Baselines

The do-nothing reference is random linker selection: draw linkers uniformly from the 806 linkers catalogued in PROTAC-DB (PROTAC-DB), attach them between the same fixed BRD9 warhead and VHL ligand, and put the survivors through the identical synthesis and assay cascade with the identical per-round budget. Its confirmed-degrader rate per round is the floor; any loop that cannot beat random sampling from a database of known-degrader parts is not learning anything.

The strongest existing competitor is the frozen in-silico pipeline: CupOFLATTE generation ranked by DegradeQuery, the current best predictor at AUROC 0.9065 on PROTAC-8K (CegradeQuery), with DegradeMaster's ranking as a second variant. This baseline gets the same compute, the same synthesis budget of 24 candidates per round, and the same assays; the only thing it cannot do is retrain on the outcomes. If the frozen pipeline's hit rate holds steady across rounds while the closed loop's rises, the value of experimental feedback is demonstrated directly. If the frozen pipeline matches the loop, the honest conclusion is that current predictors already saturate what this small campaign can teach — a finding worth publishing either way, and one that would redirect the field's effort from data collection to model architecture.

Two ablations isolate the proposal's ingredients. The first removes the enrichment reward by setting $\lambda = 0$, leaving a loop driven by the Co-IP pull-down signal alone; this tests whether pathway-level safety scoring changes which candidates survive, or merely decorates the selection after the fact. The second removes the pull-down oracle, running a loop driven only by end-point Western blots of target depletion; this tests whether physical-engagement measurements carry information that simple abundance measurements do not — the mechanistic distinction the Co-IP literature insists on (Curcumol study).

Why not just…?

Why not just trust the computational predictors and skip the wet lab? This is the serious objection, because CupOFLATTE's generated molecules are already predicted to be 1.7x to 1.9x more active than their starting points, and DegradeQuery outperforms everything without touching a pipette (CupOFLATTE; DegradeQuery). But those activity numbers are model outputs fed into other model outputs, a circularity CupOFLATTE's own authors acknowledge in calling wet-lab testing the necessary next step; and the lead-optimization literature documents computed binding energies that never became cellular activity (ScaffoldGVAE). The fair version of this objection is the frozen-baseline arm above, and if the frozen pipeline matches the loop we will report that. What the predictors cannot do even in the best case is measure off-target complex destruction, because no database contains it.

Why not just use Western blotting of the target protein? Measuring that the target's abundance drops tells you something happened, not what: the protein could be declining because transcription is repressed, translation is shut down, or the cell is dying. The curcumol study is the template case — Co-IP was required to show the compound disrupts the physical SKP2–p27 contact, which abundance measurements alone could not establish (Curcumol study), and the same mechanistic logic appears in work distinguishing direct physical complexes from indirect cascades (Chang Li et al.). For a degrader specifically, only pull-down of the ligase with the target proves the molecule works by induced proximity rather than by poisoning the cell. The Western-only ablation will quantify how much ranking power is lost without that signal.

Why not just optimize a conventional inhibitor of BRD9? BRD9's bromodomain is inhibitable, so this is a real threat to the testbed's relevance, and we concede the point narrowly: the scientific claim under test is about the loop, not about BRD9 as a drug target. The broader answer, established in the degrader literature, is that degradation is catalytic — one molecule destroys many target copies — and needs only a shallow surface-binding site, which keeps the modality relevant for targets without deep active sites (DeepPROTACs). Practically, BRD9–VHL is where the comparable published numbers live, so it is the only system where beating the frontier is a measurable claim.

Why not just do the enrichment analysis after picking the best compound? That is exactly what the curcumol and dendritic-cell studies did — profile pathways post hoc to explain a compound already chosen (Curcumol study; DNMT1 study) — and it cannot steer anything: a safety problem found after optimization is finished forces a restart. The $\lambda = 0$ ablation settles whether feeding enrichment forward actually changes selection; if it does not, we will report that post-hoc profiling suffices at this scale, which would itself narrow the field's open questions.

Decisive Checks

Three results would make the case. First, the closed loop's confirmed-degrader rate must rise across rounds while both random sampling and the frozen in-silico baseline stay flat, with the prospective round-four prediction succeeding at a rate the frozen baselines do not match. Second, the tier-one Co-IP signal must correlate with tier-two $DC_{50}$ across candidates, validating pull-down strength as a cheap screen that stands between computation and full degradation assays — and the DeepTernary buried-surface-area correlation must either replicate or fail against measured potencies, whichever the data says. Third, the disagreement-sampling between geometric and relational predictors must yield a clear verdict on which internal logic better predicts cellular outcomes, converting the field's quiet contradiction into a measured answer. Failure of any of these is reportable: the first failure would show current predictors saturate small-budget campaigns; the second would show pull-down is not a usable intermediate oracle; the third would show the dispute needs data at a scale this proposal does not reach. No outcome leaves the field where it started, which is designing degraders against scores no one has ever checked.