# Machine learning prioritization of antibiotic residues in aquatic foods reveals exposure-driven genotoxic risk mediated by BCL2

## TL;DR

This study uses machine learning to identify antibiotic residues in aquatic foods that pose genotoxic risks, showing how exposure levels drive these risks and how BCL2 can mitigate them.

## Contribution

A novel ML framework that integrates surveillance data and mechanistic validation to prioritize genotoxic risks from antibiotic mixtures in food.

## Key findings

- Prioritized mixtures of antibiotics reduced cell viability and induced DNA damage in hepatocytes.
- BCL2 overexpression significantly reduced DNA damage, confirming its protective role.
- ML revealed that exposure levels, not just toxicity, are key drivers of risk.

## Abstract

Antibiotic residues in aquatic foods pose genotoxic risks. Traditional monitoring focuses on individual Maximum Residue Limit (MRL) compliance, often overlooking cumulative multi-residue risks. We developed an interpretable machine learning (ML) framework integrating surveillance data (3719 samples, 17 antibiotics) with mechanistic validation to prioritize risks in Guangzhou (2021−2023). Exposure metrics and in silico hazard predictions were analyzed via clustering and ranking. Enrofloxacin, sulfamethoxazole, and 3-amino-2-oxazolidinone were prioritized, with risk drivers deconstructed via Explainable AI (SHAP). In L-02 hepatocytes, prioritized mixtures reduced viability (70.21 ± 7.49%), increased apoptosis (7.12 ± 2.75%), and induced DNA damage (tail DNA% 5.25 ± 1.03%) (all p < 0.05). BCL2 overexpression significantly attenuated this damage (tail DNA% 2.90 ± 2.65%, p < 0.05), confirming its role as a key functional mediator. This surveillance-to-mechanism workflow provides a data-driven paradigm for identifying mechanistic biomarkers and prioritizing food safety interventions, surpassing traditional compliance-based monitoring.

•Integrated exposure-toxicity metrics via ML to prioritize residues in aquatic foods.•SHAP revealed exposure outweighs intrinsic toxicity as primary risk driver.•Validated antibiotic mixtures induce DNA damage via BCL2-mediated apoptosis.•ML uncovered “risk-shifting” from restricted ENR to emergent SMZ violations.•A replicable paradigm translating big data into actionable policies.

Integrated exposure-toxicity metrics via ML to prioritize residues in aquatic foods.

SHAP revealed exposure outweighs intrinsic toxicity as primary risk driver.

Validated antibiotic mixtures induce DNA damage via BCL2-mediated apoptosis.

ML uncovered “risk-shifting” from restricted ENR to emergent SMZ violations.

A replicable paradigm translating big data into actionable policies.