Mining Conditional Degron Events to Deorphanize E3 Ubiquitin Ligases

Cells get rid of specific proteins by attaching a small tag called ubiquitin to them, which sends them to the cell's shredder, the proteasome. The machines that choose which proteins get tagged are called E3 ubiquitin ligases; humans have more than 600 of them, and for a large fraction we do not know which proteins they tag. Meanwhile, decades of biomedical papers describe exactly the events we need — a protein disappearing after a phosphorylation, a lysine residue being ubiquitinated, a degradation that stops when one ligase is removed — but in free text, one isolated observation per paper, invisible to any database. This proposal argues that the fastest route to matching orphan E3 ligases with their substrates is not a bigger screen but a linguistically disciplined reading of the literature: a relation-extraction system built specifically for the conditional, three-part grammar of ubiquitination events, coupled to a pairing model that ranks candidate ligases for experimentally mapped degradation signals.

The proposal sits deliberately at the interdisciplinary edge of NLP. The field's own history shows that its tools become foundational for other sciences precisely when they absorb those sciences' structure — Inoue et al. document NLP becoming core infrastructure for sociology and management research, and Amblard et al. argue for training that treats signal, syntax, and knowledge representation as one discipline. Ubiquitin biology is a case where the standard generic extraction schema ("protein A interacts with protein B") discards exactly the information that matters: the ligase, the substrate, the lysine or degron motif, and the trigger that makes the degradation conditional.

Core Claim

Existing biomedical text mining cannot deorphanize E3 ligases because it flattens ubiquitination into a generic binary interaction, and existing enzyme–substrate machine learning cannot either, because joint neural models trained on sequence pairs collapse into memorizing average reactivity — the "CPI Paradox" demonstrated by Machine learning modeling of family wide enzyme-substrate specificity screens across six dense enzyme–substrate screens. I claim that (a) a dedicated extraction schema for conditional proteolytic events, with nested entity recognition for multi-subunit ligase complexes, can recover far more usable (ligase, substrate, site, trigger) tuples than generic protein–protein interaction (PPI) extractors; and (b) these tuples, used as conditioning data and structural-accessibility filters for a pairing model, will rank the correct ligase for an orphan degradation event substantially better than either text co-occurrence alone or naive joint sequence models.

Research Program

The program has three linked components.

First, a schema and corpus. I will define an event schema capturing the four roles that the surveyed literature shows to be decisive: the E3 ligase (or a multi-protein complex like an SCF or cullin-RING assembly), the substrate, the ubiquitination site or degron motif, and the conditional trigger — phosphorylation, conformational opening, or metabolic state. The motivating cases come straight from the file: MITF degradation coupled to c-Kit dual phosphorylation with the ligase unknown (Hemesath et al.), EZH2 clearance after β-hydroxybutyrate with the ligase unknown (Yum et al.), and WAVE2 ubiquitination at K45 exposed only when the complex unfolds after T-cell activation, again with the ligase unnamed (Joseph et al.). These are precisely the tuples generic extractors lose. A key sub-problem is entity recognition: Zhou et al. showed that 44.42% of true relations on BioCreative VI Track 4 were lost because the upstream tagger missed gene mentions, and ligase nomenclature is worse — nested complex names, TRIM family members, subunit-level naming. The two-stage nested named-entity architecture of Gupta et al., which ranked first on SeeDev (F1 0.3738), will be adapted to recognize ligase complexes and modification-specific substrates.

Second, the extraction model. Rather than generic binary PPI extraction, the model will be trained on the new schema, with knowledge-graph priors in the spirit of Zhou et al.'s TransE embeddings over IntAct/BioGrid triples — but extended with event-type embeddings so that "ubiquitinates at K" is distinguishable from "stabilizes" or "binds." The output is a large, weighted graph of asserted ligase–substrate–site–trigger events, each with the sentence, trigger conditions, and citation attached. The extraction component will also harvest the negative and orthogonal evidence the file shows is abundant: deubiquitinases like ATXN3 that stabilize substrates by removing ubiquitin (Wei et al.) must not be conflated with ligases, and a generic extractor will conflate them constantly.

Third, the pairing model. For a substrate with a mapped but unassigned degron — the exact situation produced by Larsen et al.'s atlas of >63,000 peptides across 1,626 transcription factors, where the ligase is unknown for nearly all discovered degrons — the model ranks all ~600 candidate ligases. It combines three signals: text-mined prior probability from the event graph; protein language model representations with the substrate-binding-domain pooling strategy of the CPI-screening paper, pooling only ligase residues within a structural radius of the substrate-recognition domain rather than averaging over the whole sequence; and structural accessibility of the degron itself, since Larsen et al. found only 32% of degrons are surface-exposed, meaning buried degrons require ligases active in quality-control or conditional-exposure regimes rather than static recognition. The design lesson from the CPI Paradox is explicit: the ligase representation must be conditioned on the specific degron, never averaged. No reported result found on Lacuna for any existing E3-ligase–substrate pairing model, which is the open slot this component fills.

Baselines

The do-nothing reference point is PubMed co-occurrence: retrieve every sentence where a ligase name and a target name co-occur with "ubiquitinat" or "degrad" and rank pairs by co-occurrence count. It requires no modeling and is the natural first instinct; Quan et al. measured structured linguistic extraction improving F-score by 7.5 to 9.4 percentage points over exactly this baseline on the Genetic Association Database, so it is beatable but not trivially so. For the extraction component, the strongest existing competitors are the SVM with corpus weighting and rich parse-path features of Miwa et al., which reached an AUC of 89.3% on AIMed, and the TransE-augmented memory network of Zhou et al., which reached an F1 of 37.48% on BioCreative VI Track 4 against a 33.94% rule-based top baseline. Both will be retrained on the same new annotated corpus with the same compute and tuning budget, evaluated on the same held-out test set; because they emit binary relations, they will be evaluated both on binary recovery and on best-effort tuple recovery (filling schema slots by heuristics), so the comparison tests the schema, not just the model class. For the pairing component, the strongest baseline is the one the CPI study showed dominates: single-task Ridge regression on ESM-1b features, one model per ligase, which outperformed all joint neural CPI architectures on the "Big Six" screens. It will be run with the same training pairs, the same structural features, and the same hyperparameter search budget. Beating a baseline means, for extraction, higher tuple-level F1 with the conditional-trigger slot scored strictly; for pairing, higher recall@10 and recall@50 of the true ligase in a held-out setting, with a paired permutation test. Finally, an ablation isolates the proposal's key ingredient: the same pairing model with the text-mined event prior removed (structure and sequence only), and separately with the structural accessibility filter removed (text and sequence only). If the full combination does not beat both ablations, the interdisciplinary claim is falsified and I will say so.

Why not just…?

Why not just co-occurrence mine PubMed? This is the most serious threat, because co-occurrence is free and the corpus of ubiquitination papers is large. The evidence against it is direct: Quan et al. found structured parsing beat co-occurrence by 7.5–9.4 F-score points on GAD because incidental mentions (controls, distinct conditions in one abstract) generate false positives, and Zhou et al. note co-occurrence cannot capture directionality or interaction type. For deorphanization the failure is sharper: a sentence saying ATXN3 stabilizes ZEB1 by removing ubiquitin (Wei et al.) is a co-occurrence hit for the wrong direction entirely. That said, co-occurrence may win on raw recall for candidate generation, so Experiment 1 measures precision–recall curves for both and the pairing model is free to use co-occurrence as one feature; the claim is that the schema adds precision on conditional slots, not that co-occurrence is useless.

Why not just train a joint deep model on ligase–substrate sequence pairs? Because this exact design has been tested and it failed: across six dense enzyme–substrstrate screens (1,000–36,000 pairs each), joint CPI neural networks produced "blocky" predictions that ignored the paired substrate, and simple ESM-1b features with Ridge regression beat them (Machine learning modeling of family wide enzyme-substrate specificity screens). That paper's small-molecule substrates are arguably an easier conditioning problem than polypeptide degrons, so the paradox should bite at least as hard here. Experiment 2 therefore includes a naive joint sequence-pair neural model as a baseline expected to fail; if it instead succeeds on protein–protein pairs, that would itself be a finding worth reporting, and the proposal's pairing claim would need to be re-scoped around the text prior.

Why not just run bigger peptide tiling screens, as Larsen et al. did? Tiling screens are complementary, not substitutable. They screen linear 30-mer peptides, so they systematically miss phospho-degrons (the c-Kit–MITF case), conformational degrons (the WAVE2 case, where the signal is buried until the complex opens), and — decisively — they report only that a peptide is degraded, never which of the 600+ ligases did it (Larsen et al.). The proposal uses that atlas as input, not competition: it is the source of orphan degrons the pairing model ranks ligases for.

Why not just use the memory-network-plus-TransE extractor as-is? Zhou et al.'s system is the closest existing work and is included as a trained baseline. It treats all interactions as generic or mutation-associated triples, has no slot for ubiquitination sites, degron motifs, or triggers, and its upstream tagger lost 44.42% of relations to entity misses — lethal for ligase complex nomenclature. The comparison in Experiment 1 is designed to quantify exactly this gap rather than assert it.

Proposed Experiments

Experiment 1 (extraction): annotate a corpus of ~2,000 ubiquitin-system abstracts under the new schema, with a held-out test set and a time-split. Evaluate tuple-level F1 (all four slots correct) for the proposed model versus co-occurrence, the rule-based reference (which scored 33.40% F on AIMed in Quan et al.), the Miwa et al. SVM-CW system, and the Zhou et al. memory network. The time-split is the decisive check: does the system extract ligase assignments that were experimentally confirmed after the training cutoff? Each confirmed future pair is a free positive label.

Experiment 2 (pairing): assemble all literature-confirmed ligase–substrate pairs as ground truth (the file alone supplies dozens: RNF148–TSPAN15 Suo et al., Rnf40–Parkin Kou et al., FBXW7–RIG-I Liu et al., UBR5–TERT, TRIM15–Keap1/Nrf2, and more), and evaluate under leave-one-ligase-family-out and time-split protocols, reporting recall@10 and recall@50. Baselines: co-occurrence prior alone, structure-only, ESM-1b+Ridge single-task, naive joint CPI network, and the two ablations. No reported result found on Lacuna for E3-ligase–substrate pairing under any protocol, so every number here is new.

Experiment 3 (application and wet-lab closure): apply the full pipeline to the highest-value orphans documented in the file — the ligase behind c-Kit–triggered MITF clearance, the ligase behind ketone-induced EZH2 clearance, and the ligase that ubiquitinates WAVE2 at K45 after T-cell activation — plus a sample of buried and exposed degrons from the Larsen atlas. For the top-ranked predictions, validate by degrading or knocking down the predicted ligase (small-molecule degraders exist for ligases like cereblon, redirected to neo-substrates in Lin et al.), or by the massively parallel Perturb-seq screening format of Suo et al.'s E3-ligase screen, which already profiles ligase knockouts and could be re-scored for stabilization of the orphan substrate. The decisive check is unambiguous: does removing the predicted ligase stabilize the orphan substrate? Even one confirmed assignment, made from text and structure before any bench work, would validate the interdisciplinary bet that the ligase–substrate knowledge we need is already written down — just not yet machine-readable.