Molecular Targets Shape Reasoning Topology: Bidirectional Validation of Chemistry-Inspired Bond Classification for AI Reasoning Chains

Recent work classified AI reasoning chains using chemistry-inspired bond types, namely covalent (deductive), hydrogen (reflective), and van der Waals (exploratory), showing that structural stability of reasoning predicts training data quality (Direction A). We test the complementary direction: does the molecular target being reasoned about shape the topology of reasoning itself (Direction B)? Across 478 reasoning traces from five ADMET endpoints, we find that it does, with large effect sizes (Cohen's d = 0.94 and 1.75, both p < 10 −9, surviving Bonferroni correction). We collect chain-of-thought traces from Claude Opus 4.6 performing ADMET property predictions across CYP2D6, CYP3A4, CYP2C9, hERG, and P-gp, using live computational chemistry tools via the Model Context Protocol. A four-signal Semantic Bond Classifier, using sentence embeddings, discourse markers, positional distance, and an energy proxy, classifies reasoning bonds without requiring attention weights from open-source models. Of five pre-registered hypotheses, aggregate reasoning structure does not predict prediction accuracy (H1-H4, all p > 0.12). However, molecular targets strongly determine reasoning topology (H5): CYP2D6 reasoning (2D substructure-driven) shows significantly more covalent bonds than hERG reasoning (3D conformation-dependent), while hERG reasoning shows more van der Waals bonds than CYP2D6. Classifier ablation reveals semantic similarity as the dominant signal (∆ = 13.2 percentage points when removed). These results close a bidirectional loop: chemistry metaphors explain reasoning structure, and chemistry problems shape reasoning structure. The finding that molecular target type determines reasoning topology, with large, interpretable effect sizes, has implications for understanding how AI systems adapt their problem-solving approach to domain-specific demands.