Machine-learning models that predict enantioselectivity are trained on reactions mined from the chemistry literature. Enantioselectivity is how strongly a chiral catalyst favors one mirror-image product over the other, usually reported as enantiomeric excess ($ee$). These models are trained mostly on fixed columns: substrate, ligand, metal, solvent, and sometimes temperature. Many of the "activity cliffs" and contradictions in these datasets may not be chemistry the model failed to learn. They may be conditions the dataset never recorded. Examples are whether the ligand and metal were pre-mixed before the substrate was added, a trace additive, or an ion-pairing co-catalyst. We propose to find these hidden conditions automatically. A small logic layer checks the extracted reaction claims for contradictions. Each contradiction it cannot explain sends an extractor back to the two source papers to find the attribute that separates them. That attribute is then added to the schema.
In short, the schema grows only where the data contradicts itself. This borrows the core move of conflict-driven clause learning in SAT solvers, where a conflict produces a new learned constraint, and applies it to information extraction: a conflict produces a new schema attribute. We test the approach on the roughly 12,600-reaction asymmetric olefin hydrogenation corpus. We ask two questions: whether the recovered conditions are real, and whether they improve few-shot $ee$ prediction beyond the current meta-learning methods.
Core Claim
Literature catalysis datasets contain many pairs of reports that look inconsistent. Two such reports share the same substrate, ligand and metal but give very different $ee$, or they use mirror-image ligands but report the same product handedness. We claim that a meaningful fraction of these pairs are explained by conditions stated in the experimental text but missing from the dataset's fixed columns. We further claim that a conflict-directed search finds these conditions more precisely and more cheaply than extracting every possible qualifier up front.
The motivating chemistry is well established. Reagent addition order decides stereocontrol in cinchona-mediated Reformatsky reactions (Enantioselective Reformatsky Reaction Using Cinchona Alkaloids as Chiral Ligands). A pre-formed cyclometalated iridium complex governs selective imine hydrogenation (Revisiting the Stereodetermining Step in Enantioselective Iridium-Catalyzed Imine Hydrogenation). Surfactants amplify $ee$ with dimeric cinchona catalysts (Highly Enantioselective Epoxidation of 2,4‐Diarylenones by Using Dimeric Cinchona Phase‐Transfer Catalysts). These variables rarely survive into fixed extraction schemas.
Relation to Prior Work
Two strands motivate the gap.
The first strand is fixed-schema literature mining in catalysis. Synergizing a knowledge graph and large language model for relay catalysis pathway recommendation extracts 27,760 reactions into an expert schema of 5 classes and 29 attributes. It reaches a mean extraction F1 of 0.9113, but it has no stereochemical attributes and cannot add new ones.
The second strand is data work on asymmetric hydrogenation. Data-Driven Insights into the Transition-Metal-Catalyzed Asymmetric Hydrogenation of Olefins documents publication bias and sparse olefin–ligand coverage in the asymmetric hydrogenation of olefins (AHO) dataset. It diagnoses the problem but does not repair records.
The prediction methods we build on and compete with are A meta-learning approach for selectivity prediction in asymmetric catalysis and Bayesian Meta‐Learning for Few‐Shot Reaction Outcome Prediction of Asymmetric Hydrogenation of Olefins. The first uses prototypical networks over fingerprints and graph embeddings. The second adds deep-kernel Gaussian processes with a parameter prior, called ADKF-prior. Both treat each reaction record as complete. We leave their models unchanged and change the records and features they see.
A separate line, Highly accurate prediction of reaction performance and enantioselectivity with a uniform machine learning protocol, shows that 3D quantum-chemistry descriptors capture activity cliffs. This constrains our claim. Some cliffs are genuinely structural, and our method must not "explain away" those with spurious text conditions.
The schema-induction method comes from Beyond Predefined Schemas: TRACE-KG. TRACE-KG induces entity and relation types by LLM clustering with deterministic validators, and it attaches condition qualifiers with provenance links to source text. It reports 84.8% relation compatible coverage in held-out-ontology tests on DBpedia-WebNLG. TRACE-KG extracts qualifiers wherever it can and never checks whether the resulting claims are mutually consistent. We reuse its extraction and provenance machinery, but drive it with contradictions.
The iterative loop resembles Logical-SAGE, which feeds solver error messages back to a language model to repair logic programs. There, feedback fixes the translation; here, feedback expands the vocabulary being translated into.
For the logic layer we consider Qiana. Qiana is a first-order logic in which formulas can be quoted and asserted to hold "in context c", written $ist(c, p)$. Contradictions stay local to a context, and theories compile to the input format of the Vampire theorem prover. Qiana has so far been demonstrated only on hand-built narrative theories. We treat its necessity as an open question, not an assumption.
Research Program
The pipeline has three parts.
The first part turns each AHO record into a claim scoped to its source paper, using papers where the dataset's source publication can be retrieved. A claim is roughly "in paper p, substrate s with ligand L and metal M under recorded conditions gives $ee$ e with configuration R/S". The claim is linked back to the paper's experimental section.
The second part is a small, hand-written background theory of chemically uncontroversial rules. The mirror-image ligand should give the mirror-image product under otherwise identical conditions. Duplicate reports of identical conditions should agree within a tolerance. Where a paper states a dependency explicitly, such as "without pre-complexation the background reaction is racemic", that dependency is added as a scoped rule. A prover then enumerates minimal sets of claims that clash under these rules. Some clashes are direct, such as the same conditions with different $ee$. Others appear only through a rule, such as the enantiomer-symmetry rule.
The third part handles each conflict. The extractor receives the two, or few, source passages and is asked for the attribute and values that distinguish the contexts. Examples are addition order, additive identity, water content, pressure, counterion, or co-catalyst. The answer passes deterministic validators before it enters the schema. The proposed values must appear verbatim, or as normalized spans, in the cited text. The attribute must actually separate the clashing claims. It must also be re-extracted for every other record, so that it does not create new conflicts elsewhere.
An accepted attribute becomes a new column that is back-filled across the corpus. The loop stops when the remaining conflicts are either resolved or labeled "unexplained by text". The unexplained set is itself a product: a candidate list of genuine structural cliffs or reporting errors.
The logic layer comes in three variants that we compare directly. The first is a grouping query with no logic. The second is standard first-order logic in which every predicate carries a paper argument, known as reification, which avoids explosion by construction. The third is Qiana, whose added value would come from quantifying over quoted claims. An example is a rule of the form "any claim in a context lacking pre-complexation is weakened", and ordering reagent additions as temporal contexts. We commit in advance to reporting the simplest variant that matches the others.
Baselines
The naive reference point is the AHO dataset used as-is, with no conflict handling. Protonet and ADKF-prior are trained and evaluated on the original features and splits. This measures how much any intervention helps at all.
The strongest existing predictor is Protonet with cluster-based task construction from A meta-learning approach for selectivity prediction in asymmetric catalysis. With full training support it reports an area under the precision–recall curve (AUPRC) of 0.9300 ± 0.0011 on AHO. On its 245-reaction out-of-sample test set, with 64 support examples, it reports 0.9341 ± 0.0053. ADKF-prior from Bayesian Meta‐Learning for Few‐Shot Reaction Outcome Prediction is the strongest method on substrate-held-out and time-held-out splits. The file quotes no scalar for it: no reported result found on Lacuna for ADKF-prior's AUPRC on those splits.
We rerun both methods from their released configurations on identical splits, seeds and tuning budgets. Our variants change only the input features and the training records. Beating these baselines means a gain larger than three standard deviations over at least five seeds. For Protonet_cluster that means exceeding roughly 0.933 AUPRC. The gain must also hold on the substrate-held-out and time-held-out splits, where unrecorded conditions should matter most.
The key ablation is non-directed qualifier extraction. TRACE-KG-style extraction pulls every condition qualifier from every source paper, and the resulting columns are fed to the same predictors, with the same extraction model and token budget. If this matches conflict-directed induction on both downstream AUPRC and attribute precision, the conflict loop contributes nothing beyond cost savings.
A second ablation fixes the schema. The Cat-KG 29-attribute schema is extended by a chemist before seeing any conflicts, with stereochemistry-relevant fields such as addition order, additives and water. The same extractor fills it.
Why not just…?
Why not just have chemists add the missing fields to a fixed schema? This is the most serious alternative. If the hidden conditions are a short, predictable list, it wins. We test it directly. A chemist pre-registers an extended schema before the loop runs, and we then measure what fraction of accepted conflict-resolving attributes fall outside it. If that fraction is small, say under a fifth, we will report that dynamic induction adds little for AHO. We would then reframe the contribution as the conflict audit itself. The Cat-KG paper's own restriction to 29 attributes suggests expert schemas omit stereochemical variables, but that is not evidence for this corpus specifically.
Why not just ask a large language model to read both papers and explain the discrepancy? The loop does exactly this at its core. What it adds is deciding which pairs to read, which requires detecting indirect conflicts through rules like enantiomer symmetry. It also adds validators that reject explanations whose values are not in the text. Logical-SAGE and the Cat-KG authors both report that unconstrained generation hallucinates in reasoning and chemistry settings. We measure it here: the same model, prompted per pair without validators, is scored on the hallucination rate of proposed attributes against expert annotation.
Why not just add a paper-ID argument to ordinary first-order logic, or skip logic and group records? Grouping finds direct conflicts. Reified first-order logic also avoids explosion. Either may suffice, and Qiana is retained only if it catches conflicts or expresses addition-order rules the others cannot. The three variants run on the same claims. We report conflicts found, conflicts resolved, and prover time. The formalism enters the final system only if it changes these numbers.
Why not just use 3D DFT descriptors, as the MALD protocol does, to capture activity cliffs? MALD addresses structural cliffs, meaning quantum-chemistry descriptors that must be computed for every molecule. We target cliffs caused by unrecorded conditions, and the two are complementary. The "unexplained by text" residue is where MALD-style descriptors should help. We check whether those residual pairs differ more in descriptor space than the resolved pairs do.
Proposed Experiments
The first experiment is an audit. We count candidate conflicts in AHO, direct and rule-mediated. Chemists annotate a stratified sample of 200 conflicts, labeling each as explained by stated conditions, a structural difference, or a likely reporting error. We then score the loop's separating attributes for precision and provenance-validated faithfulness.
The second experiment tests prediction. Protonet_cluster and ADKF-prior are trained on three inputs: original records, records with induced attributes as extra features, and records whose conflict tasks are rebuilt so that support and query sets share induced contexts. Evaluation covers random, substrate-held-out and time-held-out splits and the 245-reaction external set, using the fairness protocol above.
The third experiment measures cost: extraction tokens and prover time for the conflict-directed loop versus exhaustive qualifier extraction.
Three results would each settle a question. If induced attributes explain a substantial share of annotated conflicts and lift held-out AUPRC beyond seed noise, the hidden-condition hypothesis holds. If exhaustive extraction matches that lift, the value lies in qualifiers rather than in conflicts. If the pre-registered expert schema captures most accepted attributes, dynamic induction is unnecessary for this domain. We will report whichever outcome occurs.
References
- Enantioselective Reformatsky Reaction Using Cinchona Alkaloids as Chiral Ligands
- Revisiting the Stereodetermining Step in Enantioselective Iridium-Catalyzed Imine Hydrogenation
- Highly Enantioselective Epoxidation of 2,4‐Diarylenones by Using Dimeric Cinchona Phase‐Transfer Catalysts: Enhancement of Enantioselectivity by Surfactants
- Synergizing a knowledge graph and large language model for relay catalysis pathway recommendation
- Data-Driven Insights into the Transition-Metal-Catalyzed Asymmetric Hydrogenation of Olefins
- A meta-learning approach for selectivity prediction in asymmetric catalysis
- Bayesian Meta‐Learning for Few‐Shot Reaction Outcome Prediction of Asymmetric Hydrogenation of Olefins
- Highly accurate prediction of reaction performance and enantioselectivity with a uniform machine learning protocol
- Beyond Predefined Schemas: TRACE-KG for Context-Enriched Knowledge Graph Generation
- Logical-SAGE: A Logical Socratic Architecture for Guided Evolution in Neuro-Symbolic Reasoning
- Qiana: A First-Order Formalism to Quantify over Contexts and Formulas with Temporality