PARADOX: pharmacological analysis of relation anomalies and drug contradictions extraction

The rapid expansion of biomedical literature has led to a growing number of contradictory pharmacological reports, where the same drug interactions are described with opposing outcomes across studies. Critically, many such reports are only apparent contradictions arising from mismatched experimental conditions such as dosage, strain, endpoint, or population, rather than logically incompatible claims under comparable settings. These inconsistencies complicate drug safety evaluation and hinder the reliability of automated biomedical knowledge systems. In this paper, we introduce PARADOX, a framework for the detection and analysis of pharmacological contradictions in scientific text. The system integrates BioBERT-based embeddings, a Siamese similarity model, and an iterative refinement using a snowball process to progressively extract and refine conflicting relational statements. Each candidate contradiction is semantically verified using GPT-4o-mini, producing a structured contradiction graph derived from validated outputs. Beyond detection, we introduce a three-stage decision-tree protocol to distinguish real contradictions from apparent ones and outline a conditional biological modulation map that summarizes how identical pharmacological inputs may yield divergent outcomes under different conditions. On more than 18,000 evaluated pairs, the model recovered 39 validated contradiction candidates with only three misses. Qualitative analysis shows both genuine logical conflicts and opposing statements better explained by contextual differences or missing metadata. The results highlight the potential of contradiction-aware mining to strengthen pharmacovigilance, enhance literature consistency checking, and support more robust biomedical reasoning.