Letter to the editor: comment on “Polyethylene terephthalate microplastics exposure enhances the risk of ulcerative colitis: insights from multi-omics integration, machine learning, and molecular docking reveal intestinal toxicity mechanisms”

Dear Editor, We write to express our earnest concerns regarding the article “Polyethylene terephthalate microplastics exposure enhances the risk of ulcerative colitis: insights from multi-omics integration, machine learning, and molecular docking reveal intestinal toxicity mechanisms” (DOI: 10.1097/JS9.0000000000003476)1. The study’s integration of multi-omics and machine learning approaches to investigate the toxicity of polyethylene terephthalate (PET)-microplastics (PET-MPs) in ulcerative colitis (UC) is commendable for its conscientious effort. However, we would like to highlight several methodological and experimental design issues that may impact the interpretation of the findings. Limitations in target prediction for PET-MPs While some studies on PET-MPs have similarly employed chemical informatics databases (e.g. ChEMBL, Swiss Target Prediction) for target prediction, it is critical to note that these tools are designed for small molecules with defined receptor-binding domains2. PET-MPs, as polymeric aggregates, may exert toxic effects not only through specific ligand–receptor interactions but also, and perhaps more significantly, through physical interactions such as membrane disruption and oxidative stress. Therefore, the direct application of molecular docking to PET-MPs may overestimate the biological relevance of the predicted “binding affinities.” We encourage the authors to address this methodological limitation in the discussion. Inadequate experimental design in animal studies The study exposed mice to a combination of 3% dextran sulfate sodium (DSS) and PET-MPs. However, the absence of a control group receiving PET-MPs alone makes it impossible to determine whether the observed exacerbation of colitis is due to a direct toxic effect of PET-MPs or merely represents a synergistic interaction with DSS. Specifically, without a PET-MPs-only group, the claim that PET-MPs “enhance the risk of UC” lacks robust experimental support. Another core issue lies in the unjustified selection of the PET-MPs dosage. The authors administered 0.5 mg/mouse/day based on a previous study investigating polystyrene microplastics (PS-MPs) in a multi-organ toxicity model3, which is not directly applicable to the investigation of PET-MPs in UC. PET-MPs differ from PS-MPs in properties such as density, surface hydrophobicity, and intestinal retention time, all of which affect bioavailability and toxicity. Furthermore, no pre-experimentation (e.g. dose-escalation studies) or specific literature support is provided to justify the relevance of this dosage to UC exacerbation. Furthermore, the phenotypic assessment appears insufficient to robustly support the claim of exacerbated colitis. Although the authors mention “histopathological scoring” in the Methods section and present statistical results, the specific scoring criteria – such as the scale ranges and parameters for crypt loss, inflammatory cell infiltration, or epithelial damage – are not described. This lack of transparency regarding the scoring system makes it difficult to assess the rigor and reproducibility of the histological conclusions. Additionally, despite including histology, the study overlooks standard clinical endpoints for DSS-induced UC models, such as the Disease Activity Index4, which typically integrates weight loss, fecal characteristics, and hematochezia. Consequently, the presented results provide an incomplete picture of disease severity. It therefore remains unclear whether the observed effects in the DSS + PET-MPs group represent a true aggravation of UC pathology or are merely a reflection of non-specific systemic toxicity. Unvalidated machine learning model generalizability The integration of three machine learning algorithms (LASSO, SVM-RFE, and Random Forest) is methodologically sound. However, the exceptionally high area under the curve (AUC) values (0.98–0.99) reported for both the training and validation sets raise concerns about potential overfitting, particularly given the relatively small sample size typical of GEO datasets. Furthermore, reliance on a single external validation cohort (GSE92415) may be insufficient to robustly demonstrate the model’s generalizability across diverse populations. The inclusion of additional independent validation cohorts is widely considered an essential step to more rigorously assess the model’s performance and ensure its robustness. Conclusion We acknowledge the authors’ effort in addressing this timely topic. However, the methodological limitations highlighted above, particularly those concerning target prediction and the incomplete animal experimental design, warrant a cautious interpretation of the findings. To enhance the robustness of future research, we suggest that subsequent studies should focus on employing calibrated docking approaches, implementing more comprehensive animal experimental designs, and performing robust model validation. Addressing these points will be crucial for building a more solid foundation for understanding the mechanisms underlying PET-induced UC. This correspondence adheres to the TITAN Guidelines 2025 for artificial intelligence (AI) reporting standards5.