Answers are grounded in this paper and cite their sources.
Pesticide tank mixing is widely used to improve the efficiency and flexibility of crop protection. However, mixture performance depends not only on the molecular structures and physicochemical properties of the constituent pesticides but also on practical application contexts, including formulation types, target crops, target organisms, application methods, and technical requirements. These context-dependent factors make potentially incompatible or antagonistic combinations difficult to identify from molecular structures and physicochemical properties alone; in registration texts the corresponding constraints appear as documentary mixture-restriction evidence rather than as experimentally measured outcomes. Although pesticide registration documents contain authoritative information on practical use conditions, most of this information remains embedded in unstructured text and is rarely incorporated into computational analysis of registration-derived mixture restrictions. This study addresses binary prediction of registration-derived restriction associations for pesticide pairs. Positive labels are derived from mixture-restriction evidence recorded in pesticide registration texts, which primarily reflect antagonistic interactions. The model outputs a relative association score for registration-derived restriction evidence and is not intended to represent other risk types such as physical incompatibility, chemical instability, or phytotoxicity, nor does it provide experimentally verified incompatibility conclusions. This study developed FMPIP as a computational framework for predicting pesticide-pair associations with restriction evidence derived from pesticide registration information. The current study was designed as a retrospective computational evaluation rather than a prospective validation of physical, chemical, or biological mixture incompatibility. To address this limitation, we propose a structured knowledge-driven framework integrating registration-text structuring, multi-source knowledge graph construction, and graph-based registration-derived restriction association prediction. First, a previously validated single-stage entity–relation extraction method is adapted as the knowledge-generation component to transform pesticide registration descriptions into structured usage-context knowledge. Direct statements indicating mixture restrictions are excluded from predictive inputs and are used only to generate registration-evidence-derived supervision labels, thereby separating label-generating evidence from model features. Second, non-label registration-derived contexts are integrated with molecular information and physicochemical attributes to construct a multi-source pesticide knowledge graph. Third, FMPIP is developed to estimate relative restriction-association scores for registration-derived restriction evidence by combining graph-based random neighborhood sampling with attention-based aggregation. FMPIP provides a relative association score for registration-derived restriction evidence. Within the retrospective benchmark, the score ranks pesticide pairs by their association with registration-derived restriction evidence; however, ranking-oriented metrics and practical baselines are not evaluated in this study, so practical prioritization capability is not demonstrated. The model output should not be interpreted as a direct prediction of physical, chemical, or biological mixture compatibility under specific application conditions. The knowledge-generation module achieved an F1-score of 96.6% on the pesticide registration dataset. For the downstream restriction-association prediction task, FMPIP achieved an AUC of 90.4%, an AUPR of 84.8%, and an F1-score of 84.4% on PID, outperforming all graph-based baselines re-evaluated under the same deduplicated, pair-disjoint protocol (AUC 85.5-85.9%, AUPR 72.3-75.3%, F1-score 71.9-76.3%). Ablation analysis showed that, after deduplication, removing any single information source did not reduce performance, suggesting that the contribution of registration context observed in the earlier, non-deduplicated evaluation. An order-invariant evaluation, in which pair scores do not depend on the storage orientation of a pair and the classification threshold is selected on validation data, obtained an AUC of 90.8% (SD 0.5), an AUPR of 82.5% (SD 1.0), and an F1-score of 85.0% (SD 0.5) on PID, confirming that this comparison is not an artifact of pair storage. The resulting scores should be interpreted as retrospective recovery of held-out restriction labels within the benchmark rather than as evidence of experimentally verified incompatibility or of practical prioritization capability. Independent regulatory or experimental validation is required before these predictions can be interpreted as evidence supporting prospective practical mixture decisions. These findings illustrate how regulatory pesticide-use text can be converted into structured contextual knowledge for retrospective mixture-restriction analysis.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: