Interpretable Random Forest Model Based on Serum Biochemical Indicators for Discriminating Between Diquat-Exposed Rats and Controls

Acute diquat poisoning causes multi-organ injury and glycolipid metabolic disturbance. Liquid chromatography–tandem mass spectrometry (LC-MS/MS) serves as the well-established confirmatory method for diquat detection, but it may produce false-negative results at late poisoning stages. This study aimed to construct an interpretable machine learning model for discriminating diquat-exposed male SD rats from saline controls and screen serum candidate biomarkers. In total, 110 male Sprague–Dawley (SD) rats were divided into a diquat-exposed group (n = 50, 100 mg/kg gavage) and a control group (n = 60). Thirty-three independent test samples (15 diquat-poisoned rats and 18 control rats) were reserved for final model evaluation. Serum biochemistry was detected 48 h post exposure. After multistep feature selection performed exclusively on the training subset, five low-collinearity candidate biomarkers, including aspartate aminotransferase (AST), alkaline phosphatase (ALP), glucose (GLU), blood urea nitrogen (BUN), and triglyceride (TG), were retained. Tree-based ensemble models showed higher mean nested-CV AUC than logistic regression (LR) and support vector machine (SVM). Selected based on nested cross-validation metrics, the random forest (RF) model yielded a held-out test set AUC of 0.970, an accuracy of 0.879, a sensitivity of 0.733, a specificity of 1.000, and an F1-score of 0.846. Serum biochemistry showed no significant intergroup difference in AST, whereas ALT and ALP were decreased in diquat-exposed rats. These five markers reflect hepatorenal functional disturbance and glycolipid metabolic disturbance observed in this diquat-poisoned rat model; however, their toxicological specificity for diquat intoxication remains unconfirmed. Combined with a SHapley Additive exPlanations (SHAP) interpretation, this biochemistry-based model offers a novel toxicological analytical framework, yet its findings are only applicable to male SD rats and cannot be extrapolated to humans or other conditions. This study has limitations including the use of a single rat strain and one diquat dose; external validation on independent animal cohorts is required to confirm model generalizability.

Authors

Institutions

Publication Details

Journal
Toxics
Published
2026-11-16
DOI
https://doi.org/10.3390/toxics14111010
Primary Topic
Paraquat toxicity studies and treatments
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Interpretable Random Forest Model Based on Serum Biochemical Indicators for Discriminating Between Diquat-Exposed Rats and Controls

Min Zeng, Simon X. Yang, Qiao Liu, Jiali Ma et al.
1 links
Toxics
Paraquat toxicity studies and treatments
article

Interpretable Random Forest Model Based on Serum Biochemical Indicators for Discriminating Between Diquat-Exposed Rats and Controls

Min Zeng, Simon X. Yang, Qiao Liu, Jiali Ma, Xianchuan Wang
article en

Abstract

Acute diquat poisoning causes multi-organ injury and glycolipid metabolic disturbance. Liquid chromatography–tandem mass spectrometry (LC-MS/MS) serves as the well-established confirmatory method for diquat detection, but it may produce false-negative results at late poisoning stages. This study aimed to construct an interpretable machine learning model for discriminating diquat-exposed male SD rats from saline controls and screen serum candidate biomarkers. In total, 110 male Sprague–Dawley (SD) rats were divided into a diquat-exposed group (n = 50, 100 mg/kg gavage) and a control group (n = 60). Thirty-three independent test samples (15 diquat-poisoned rats and 18 control rats) were reserved for final model evaluation. Serum biochemistry was detected 48 h post exposure. After multistep feature selection performed exclusively on the training subset, five low-collinearity candidate biomarkers, including aspartate aminotransferase (AST), alkaline phosphatase (ALP), glucose (GLU), blood urea nitrogen (BUN), and triglyceride (TG), were retained. Tree-based ensemble models showed higher mean nested-CV AUC than logistic regression (LR) and support vector machine (SVM). Selected based on nested cross-validation metrics, the random forest (RF) model yielded a held-out test set AUC of 0.970, an accuracy of 0.879, a sensitivity of 0.733, a specificity of 1.000, and an F1-score of 0.846. Serum biochemistry showed no significant intergroup difference in AST, whereas ALT and ALP were decreased in diquat-exposed rats. These five markers reflect hepatorenal functional disturbance and glycolipid metabolic disturbance observed in this diquat-poisoned rat model; however, their toxicological specificity for diquat intoxication remains unconfirmed. Combined with a SHapley Additive exPlanations (SHAP) interpretation, this biochemistry-based model offers a novel toxicological analytical framework, yet its findings are only applicable to male SD rats and cannot be extrapolated to humans or other conditions. This study has limitations including the use of a single rat strain and one diquat dose; external validation on independent animal cohorts is required to confirm model generalizability.

ToxicsVol. 14(11)
Wenzhou Medical University (CN), Wenzhou Vocational College of Science and Technology (CN), Wenzhou Municipal Sci-Tech Bureau (CN), University of Guelph (CA)
Basic Research Project of Wenzhou City
Good health and well-being
Paraquat toxicity studies and treatments
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon