The Challenge
Scaling Nanosafety Assessment Beyond Traditional Experiments
Traditional methods for assessing nanoparticle safety are time-consuming, expensive, and rely heavily on animal testing. With over 5,300 commercial products containing nanoparticles and growing regulatory pressure under the European Green Deal and SSbD framework, faster computational alternatives are urgently needed.
Key challenges include dataset scarcity and class imbalance in nanotoxicology data, the lack of robust applicability domain methods, and the need for models that non-informatics experts can confidently use and interpret.
Our Approach
An end-to-end in silico workflow from data enrichment to cloud-deployed nanosafety models
Integrate experimental and computational data
Combined High Throughput Screening with High Content Imaging data from the NanoMILE project with atomistic descriptors calculated from molecular mechanics force fields, enriching the experimental feature space with structural information.
Address class imbalance with synthetic data
Applied SMOTE and ADASYN oversampling techniques to generate synthetic training samples for underrepresented classes, improving model performance on minority class predictions while maintaining dataset integrity.
Automate model selection and optimisation
Deployed an AutoML scheme on the Enalos Cloud Platform to systematically evaluate classification and regression algorithms, optimise hyperparameters, and select the best-performing models for each toxicity endpoint.
Develop a novel applicability domain
Created a three-step AD method combining bounding-box analysis, leverage approach, and a new local similarity assessment to classify each prediction as good, moderate, or poor reliability — enhancing user confidence.
Deploy as SafeNanoscope web service
Made the validated models available through the SafeNanoscope application on the Enalos Cloud Platform, with data shared through the nanoPharos database for community access and reuse.