Related Experiment Video
Updated: Aug 5, 2026
05:47
In Silico Modeling Method for Computational Aquatic Toxicology of Endocrine Disruptors: A Software-Based Approach Using QSAR Toolbox
Published on: August 28, 2019
Interpretable Machine Learning for Predicting Enrofloxacin Residues in Fish Using a Large Literature-Derived Database
Peilong Song1,2,3, Bo Rong1,2,3, Linhua Zhou1,2,3
1School of Marine Technology and Environment, Dalian Ocean University, Dalian 116023, China.
Foods (Basel, Switzerland)
|July 28, 2026
Summary
Enrofloxacin residues in fish pose food safety challenges. Machine learning models predict concentrations using factors like species and tissue, but cross-study generalization is limited, requiring standardized data.
Area of Science:
- Aquaculture
- Food Safety
- Pharmacokinetics
Background:
- Enrofloxacin use in aquaculture leads to fish tissue residues, complicating food safety and risk assessment.
- Predicting enrofloxacin concentrations is challenging due to variability from species, tissue, administration, and environmental factors.
Purpose of the Study:
- To develop a model for predicting matrix-specific enrofloxacin concentrations in fish using a literature-derived database.
- To identify key predictors influencing enrofloxacin residue levels in fish tissues.
Main Methods:
- Compiled a database of 1254 enrofloxacin residue records from 39 source groups.
- Trained eight machine learning models using eleven predictors and log1p-transformed concentrations.
- Evaluated models using random row-level and stratified source-grouped validation.
Main Results:
- The histogram gradient-boosting decision tree model achieved high performance (R²_log = 0.8889) on row-level validation.
- Stratified source-grouped validation showed significantly reduced performance, indicating poor cross-study generalization.
- ExtraTrees model demonstrated stable grouped validation performance, identifying species, tissue, dosing frequency, and sampling time as key predictors.
Conclusions:
- A transparent framework for residue pattern modeling and predictor prioritization was established.
- The study highlights the need for standardized residue depletion data for improved predictive accuracy.
- External validation is crucial for assessing the generalizability of residue prediction models.
