Applicability-domain-constrained prediction of landfill-relevant odor descriptors using integrated chemical knowledge.

Abstract

Malodor emissions from municipal solid waste treatment facilities are complex environmental exposure mixtures, making rapid identification of odor-relevant compounds difficult using olfactometry or compound-by-compound GC-MS interpretation. Here, we developed a knowledge-guided molecular learning framework for interpretable and reliability-aware prediction of waste-treatment malodors. The framework integrates Morgan fingerprints, large language model-derived structure-odor rules, deterministic functional-group descriptors, and StructKG-derived hierarchical structural semantics. On a curated dataset of 3756 molecules, the fused representation with XGBoost achieved the best performance, with an AU-PRC of 0.437 and an AU-ROC of 0.869. To improve transferability to external chemical space, we introduced a multi-label applicability domain combining local similarity density with neighborhood label inconsistency, increasing in-domain AU-PRC to 0.545. External validation using compounds detected at the Shanghai Laogang waste-treatment site showed that malodor-related predictions increased from 66% outside the domain to 93% inside it. Model interpretation and theoretical odor concentration-weighted attribution identified reduced sulfur compounds, amines, volatile fatty acids, and reactive carbonyls as major odor-marker classes, supporting targeted monitoring and control of waste-treatment malodors.