Speaker
Description
Nanomaterials are increasingly being used in areas including environmental remediation, biomedical engineering, and advanced manufacturing. However, this growth presents a continuous challenge: traditional in vitro assays used to evaluate nanoparticle toxicity are time-consuming and costly, and scale poorly as new nanomaterials continue to emerge. This study investigates whether machine learning approaches may provide a rapid and scalable alternative by predicting toxicity outcomes based on physicochemical properties, without relying on direct biological experiments.
Here, we use a publicly available nanoparticle dataset containing core size, hydrodynamic size, surface charge, surface area, exposure dose, and exposure time. Previous work on this dataset often stopped at reporting accuracy, with little discussion of why a model makes a given prediction or how such results would generalise beyond the research setting. In this work, four classification algorithms are compared: Random Forest, Support Vector Machine, K-Nearest Neighbors, and Extreme Gradient Boosting. SHAP (SHapley Additive exPlanations) is then applied to the best-performing model to identify which physicochemical properties most influence its predictions.
The aim is a screening tool that is not only accurate but interpretable, one that researchers could use in early-stage safety assessment without having to justify a black-box result.
