Catching Ligand Teleportation with Verified Literature Grounding

A new family of "co-folding" models — deep networks that take a protein sequence and a small-molecule structure and output a predicted 3D complex — has made orthosteric docking (predicting where a drug-like molecule binds a protein's active site) nearly routine. But a recent evaluation on seventeen proteins with both an active-site ligand and a remote regulatory ligand found something alarming: the models place active-site ligands within 5 Å of their crystal positions, yet displace allosteric ligands — molecules that bind distant "volume knob" sites to tune protein behavior — by more than 10 Å, frequently teleporting them straight into the active-site pocket, because the training data in the Protein Data Bank is dominated by orthosteric complexes (Co-folding, the Future of Docking – Prediction of Allosteric and Orthosteric Ligands). This is a hallucination in the exact sense the LLM literature uses: a fluent, physically plausible output (Boltz-1 passes stereochemical PoseBusters checks about 90% of the time) that is factually wrong about the world.

The root cause is data fragmentation, not model architecture. Allosteric site knowledge exists — but it is scattered across decades of journal text, figures, and tables that no structured database captures, while the curated residue of it, ASD2023, required manual curation of 2,422 proteins and 100,320 modulators and still leaves most of its 66,589 predicted pockets experimentally unvalidated (ASD2023: towards the integrating landscapes of allosteric knowledgebase; Direction 2 Summary). This proposal argues the two problems solve each other: literature mining can supply the grounding that co-folding models lack, and hallucination-detection machinery developed for text LLMs — claim decomposition and verification against source documents — is exactly what makes such mining trustworthy enough to use.

Core Claim

Pocket-selection errors in co-folding outputs are detectable and rankable by grounding predicted ligand positions against literature-attested allosteric site evidence, where that evidence is itself extracted by LLMs and verified at the level of atomic claim-triplets rather than sentences. The claim has two halves that must be tested together: that verified extraction yields a grounding set clean enough to act as a spatial prior, and that this prior separates genuinely allosteric poses from teleported ones in generative ensembles where physical-validity filters and model confidence provably fail. Neither half exists today. The co-folding evaluation study documents the teleportation failure but explicitly leaves open any automated scoring or ranking function to filter it, and never touches unstructured literature (Co-folding, the Future of Docking – Prediction of Allosteric and Orthosteric Ligands). The knowledge-centric hallucination detector REFCHECKER decomposes generated text into (subject, predicate, object) triplets and checks each against reference text with three-way natural language inference — deciding entailment, contradiction, or neutrality — but has only ever been evaluated on general-knowledge and open QA contexts, never on biochemical literature or protein structures (Knowledge-Centric Hallucination Detection). And multimodal chemistry extraction exists in OpenChemIE, which links reaction diagrams, R-group tables, and condition text across a paper (OpenChemIE: An Information Extraction Toolkit For Chemistry Literature), but it extracts small-molecule reactions, not allosteric sites, residue positions, or protein–modulator relationships, and its own authors flag LLM-based relational synthesis as an open future direction.

Research Program

The first component builds the grounding substrate, provisionally called an allosteric triplet corpus. An LLM pipeline reads full-text allosteric papers (figures, tables, and body text, following OpenChemIE's multimodal architecture) and emits atomic triplets of the form (protein, has allosteric site near residues R), (modulator M, binds protein P at site S), (residue mutation X, abolishes allostery between S and the active site). Each triplet is then verified against the source passage by a natural language inference checker restricted to that passage — restricted deliberately, because REFCHECKER's authors found that model-based checkers cheat by leaning on their own pretrained weights rather than the reference context (Knowledge-Centric Hallucination Detection), a bias that would be deadliest precisely in allostery, where real mechanisms are counterintuitive: zinc binding at a flexible loop repositioning Asn244 to boost DNA editing in APOBEC3 (Marx et al.), or a single loop coordinating long-range coupling between ATP and substrate binding in protein kinase A (The αC-β4 loop controls the allosteric cooperativity between nucleotide and substrate in the catalytic subunit of protein kinase A). Triplets that pass also get cheap structural sanity checks: the residue must exist in the protein's sequence, and the attested site must map to a surface-accessible region of a predicted or experimental structure. Triplets attested by multiple independent passages, with conflicting attestations preserved rather than averaged, are marked highest-trust. This directly attacks the failure modes the hallucination literature documents in scientific summarization: stripped caveats turning conditional findings into facts (Evaluating large language models on medical evidence summarization) and entity or role swapping in multi-entity text (Léo Milecki).

The second component spends this substrate on the structural side. For each target protein, we generate an ensemble of co-folding predictions (Boltz-1 is the practical choice: roughly 188 seconds per prediction in the evaluation study, so ensembles of dozens of samples are cheap) and score each sample by the agreement between where the ligand lands and where verified literature triplets say allosteric sites are. A sample is flagged as a teleportation when the ligand sits in the orthosteric pocket while verified triplets attest a spatially distinct allosteric site on the same protein. Two auxiliary signals supplement the literature prior: within-ensemble placement variance, adapting the internal-consistency idea that probing a model with variations of the same question reveals fabrication through erratic answers (Ayush Agrawal et al.), and a generic pocket prior from sequence-based predictors such as LaMPSite, which finds 3D binding pockets from sequence alone at a 66.02% success rate on COACH420 (Protein Language Model-Powered 3D Ligand Binding Site Prediction from Protein Sequence). The output is a re-ranked ensemble and a per-sample hallucination verdict, the automated scoring function the teleportation study explicitly lacked.

A third, smaller thread extends the grounding idea to a known input bottleneck: AlloReverse, which predicts allosteric pockets by tracing communication from the active site outward, requires an orthosteric ligand-bound structure as input and cannot process unliganded structures (AlloReverse: multiscale understanding among hierarchical allosteric regulations). Verified literature attestations of active-site residues and ligands can stand in for that structural input on targets where no bound structure exists, letting us ask whether textual grounding substitutes for the missing structure.

Baselines

The naive reference point is physical-validity filtering: apply PoseBusters to every ensemble sample, exactly as the evaluation study did. This is the "do nothing" filter because it is what any practitioner would run today, and the documented result shows why it fails here — Boltz-1 passes about 90% of PoseBusters checks while displacing allosteric ligands by more than 10 Å, so the filter has almost no power to separate teleported poses from correct ones (Co-folding, the Future of Docking – Prediction of Allosteric and Orthosteric Ligands). We will run it on the identical ensembles so the comparison is exact.

The strongest existing competitor is Boltz-1's own native sample confidence on the seventeen-target benchmark: the model's built-in mechanism for choosing among its own samples, run with the same ensemble size, seeds, and compute as our method. No reported result found on Lacuna for any method successfully placing allosteric ligands on this benchmark; the only documented allosteric-placement number is the failure itself (displacement greater than 10 Å with frequent teleporting into the orthosteric pocket), while orthosteric placement succeeds within 5 Å (Co-folding, the Future of Docking – Prediction of Allosteric and Orthosteric Ligands). Beating this baseline therefore means, concretely: our literature-grounded re-ranking places the allosteric ligand within 5 Å of its crystal position in the top-ranked sample for strictly more targets than native confidence does, and the median top-1 displacement across the seventeen targets drops below 10 Å. A second strong competitor supplies an alternative site prior: AlloReverse, which achieves 71.0% recall on labeled allosteric sites, recovers at least one allosteric site in 77.6% of its 58 test proteins, and reaches a ROC AUC of 0.758 (AlloReverse: multiscale understanding among hierarchical allosteric regulations). We will run it wherever its required ligand-bound input exists and swap its site predictions for our literature attestations in the same scoring rule, which isolates whether mined text adds anything beyond a purpose-built structural predictor. P2Rank, at 68.24% success on COACH420 versus LaMPSite's 66.02% and Fpocket's 35.09% (Protein Language Model-Powered 3D Ligand Binding Site Prediction from Protein Sequence), serves as a generic, allostery-agnostic pocket prior under the same protocol.

Two ablations isolate the proposal's key ingredients. Removing the literature grounding leaves an ensemble-consistency-only filter (flag samples whose ligand placement diverges from the ensemble consensus), which tests whether the expensive extraction pipeline matters beyond cheap self-consistency. Removing triplet verification leaves raw LLM extraction feeding the scorer, which tests whether verification is load-bearing or cosmetic; we measure this by the contamination rate of the grounding set — fabricated residues, swapped roles, and stripped caveats counted against a manually verified gold subset.

Why not just…?

Why not just fine-tune the co-folding models on allosteric data? This is the most serious alternative, and the evaluation study itself suggests it: models succeed on targets like glucokinase whose allosteric sites are well represented in the Protein Data Bank, indicating memorization is the mechanism and rebalanced training could fix it (Co-folding, the Future of Docking – Prediction of Allosteric and Orthosteric Ligands). But fine-tuning needs paired protein–ligand structural data, which is exactly what fragmentation starves: ASD2023's 66,589 predicted pockets are mostly computational predictions without experimental validation, not training-grade structures (ASD2023: towards the integrating landscapes of allosteric knowledgebase). Our approach is complementary rather than opposed — the verified triplet corpus is precisely the training and evaluation asset such fine-tuning would require — and we will settle the question empirically at small scale by comparing our inference-time grounding filter against a confidence-ranking baseline on the same ensembles, at a fraction of the cost of retraining. If a future rebalanced model eliminates teleportation outright, the corpus remains valuable and the filter becomes a cheap safety net for the long tail of poorly represented targets.

Why not just use rule-based extractors or standard named-entity recognition instead of LLMs? The precedent is discouraging wherever evidence is scattered across modalities: ReactionDataExtractor 2.0 reached only 8.8% accuracy against the Reaxys database on cross-modal extraction, where OpenChemIE's integrated pipeline reached 64.3% (OpenChemIE: An Information Extraction Toolkit For Chemistry Literature). Allosteric evidence is similarly split — site descriptions in text, residue tables, affinity measurements in figures — and regex for mutation patterns cannot assemble a (site, modulator, effect) relation from that scatter. The honest caveat cuts the other way too: OpenChemIE's molecule recognition module accounts for 64% of its pipeline failures, and unconstrained LLMs invent references and swap entities (Let Me Do It For You: Towards LLM Empowered Recommendation via Tool Learning; Can We Automate Scientific Reviewing?), which is why our pipeline bolts triplet verification on rather than trusting generation.

Why not just run geometry-based pocket finders on crystal structures? On COACH420, Fpocket achieves a 35.09% success rate and DeepSite 53.07%, both well below sequence-based and graph-based learning methods (Protein Language Model-Powered 3D Ligand Binding Site Prediction from Protein Sequence), and more fundamentally, static geometry cannot see cryptic pockets that are collapsed in unliganded structures and open only on conformational change — the defining property of many allosteric sites (Discovery of cryptic allosteric sites using reversed allosteric communication by a combined computational and experimental strategy). Literature attestations, by contrast, record pockets observed under exactly the fluctuating or liganded conditions where they exist.

Why not just check whole sentences instead of decomposing into triplets? On the KNOWHALBENCH benchmark of 2,100 responses and 11,000 annotated triplets, triplet-level checking beat sentence-level checking by an average of 10 points in Macro-F1, and REFCHECKER beat SelfCheckGPT and FacTool by 18.2 to 27.2 points (Knowledge-Centric Hallucination Detection). A single allosteric sentence routinely stacks a correct site, a correct modulator, and a wrong residue; sentence-level verdicts blur exactly the error our spatial prior would inherit.

Proposed Experiments

The first experiment evaluates extraction and verification. We hold out a set of ASD2023 entries and their source papers, build a manually verified gold triplet set, and measure precision and recall of the extracted (protein, site, modulator) relations. The decisive comparison is verified versus unverified extraction: the contamination rate of the unverified grounding set should be high enough (fabricated residues, swapped roles, stripped caveats, each counted by type) to explain any downstream scorer failure, and triplet-level verification should show a gap over sentence-level checking consistent with the roughly 10 Macro-F1 point advantage reported on general-domain benchmarks (Knowledge-Centric Hallucination Detection). A companion probe for checker bias runs the context-restricted NLI checker against an open-book checker on genuinely counterintuitive but true mechanisms — the APOBEC3 zinc-loop case and the kinase αC-β4 loop case above — and counts how often each falsely flags a true claim, directly measuring the internal-knowledge-bias risk in a domain where novelty is the point.

The second experiment is the decisive structural test on the seventeen-target benchmark of paired orthosteric and allosteric ligands (Co-folding, the Future of Docking – Prediction of Allosteric and Orthosteric Ligands). To avoid circularity — grounding the scorer in the very structural knowledge the models memorized — the literature corpus for each target is restricted to publications predating deposition of the reference crystal structure, so the grounding set cannot contain the answer key. For each target we generate a matched ensemble, score all samples under each method (PoseBusters only, native confidence, consistency-only, AlloReverse-prior, P2Rank-prior, full literature grounding), and report top-1 allosteric placement within 5 Å, median displacement, and teleportation-detection precision and recall against the ground truth of displacement greater than 10 Å. A subgroup analysis splits targets by how well represented their allosteric sites are in the Protein Data Bank, testing the hypothesis that literature grounding helps most exactly where memorization fails most. The third, smaller experiment applies literature attestations as a stand-in for AlloReverse's required ligand-bound input on unliganded targets, measuring site-recovery against its reported 77.6% protein-level recovery where comparable (AlloReverse: multiscale understanding among hierarchical allosteric regulations). If the grounding filter moves allosteric placement from documented failure to within-5 Å success on even a majority of the seventeen targets, the claim stands; if consistency-only performs as well, the honest conclusion is that the literature pipeline is unnecessary and the cheap signal suffices — a result worth knowing, and one the ablation is designed to force.