TP-Agent: An Auditable LLM Agent for Predicting Pollutant Transformation Products in Oxidative Water Treatment

Predicting the transformation products (TPs) of organic micropollutants is important for characterizing their chemical fate during oxidative water treatment and prioritizing candidate structures for high-resolution mass spectrometry (HRMS) analysis. Limited labeled data and diverse reaction pathways make this task challenging. TP-Agent combines a large language model (LLM), training-only analog retrieval, and executable atom/bond edits to rank TP candidates with inspectable structural records. On 231 standardized-parent-disjoint reactions, TP-Agent achieved 78.8% strict Top-5 accuracy, compared with 65.8% for the bare LLM, 76.2% for standalone retrieval-augmented generation (RAG), and 56.3 ± 1.5% for TP-Transformer across three training seeds. Parent-clustered bootstrap analysis supported the gain over the bare LLM; the 2.6-percentage-point difference from RAG was not statistically significant. Retrieval supplied most of the accuracy gain, while graph execution added candidate-linked structural records. Among 1095 candidates, 92.7% had exact edit-replay or deterministic-projection records, measuring structural linkage rather than mechanistic validity. Where available, candidate-linked LLM rationales provide chemical hypotheses for inspection. Estimated parallel latency and API-equivalent cost were 3.5 and 7.4 times those of RAG, respectively. TP-Agent thus combines TP candidate ranking with traceable structural operations and inspectable chemical hypotheses for HRMS follow-up.