Explainable Prediction of Iron Carbide NP Cytotoxicity via Enalos Cloud

How NovaMechanics combined atom-level descriptors, automated machine learning, and SHAP explainability to build a regulatory-compliant cytotoxicity model for iron carbide nanoparticles — deployed as a free cloud web service.

Nanoscale Advances • 2026
View Publication

The Challenge

Assessing Biocompatibility of Emerging Magnetic Nanomaterials

Iron carbide nanoparticles (ICNPs) show outstanding potential for biomedical applications including MRI contrast, magnetic hyperthermia therapy, and controlled drug delivery. Yet their biocompatibility remains a critical concern, and traditional experimental screening is time-consuming and resource-intensive.

A key barrier is the lack of universally accepted descriptors for representing nanoparticle structural features computationally. Standard experimental characterisation captures size, coating, and surface chemistry, but misses the atom-level features that drive biological interactions.

186
Viability endpoints curated from literature
10
Human and murine cell lines tested
3
Iron carbide crystal phases modelled

Our Approach

A dual modelling strategy combining experimental features with atomistic enrichment

Curate toxicological data from literature

Collected cytotoxicity data for ICNPs across 10 immortalised cell lines, standardised to 24-hour exposure endpoints with colorimetric viability assays. Built a unified dataset of 186 data points covering diverse core compositions, coatings, and concentrations.

Generate atomistic descriptors with ASCOT

Used the ASCOT toolbox and a modified NanoConstruct to build spherical NP models from iron carbide crystal structures. Generated atom-level descriptors capturing surface and bulk atom features for three iron carbide phases: cementite, hexagonal, and Hägg carbide.

Build and compare two modelling strategies

Developed an evidence-based approach using only experimental features and an atomistic-enriched approach that incorporates computed structural descriptors. Automated ML on the Enalos Cloud Platform evaluated multiple algorithms to identify the best performer.

Validate and explain model predictions

Validated the resulting Random Forest model against OECD principles for QSTR development. Applied SHAP additive values and permutation importance to reveal which nanoparticle characteristics most strongly drive cytotoxicity predictions.

Deploy as cloud web service

Made the final validated model freely available through the Enalos CHIASMA Cloud Platform, and published all curated data through the NanoPharos database for community reuse.

Results at a Glance

RF
Best Model
Random Forest selected as the top-performing algorithm through AutoML
OECD
Regulatory Compliance
Model fully adheres to OECD principles for QSTR development
SHAP
Explainable AI
Shapley values reveal key input characteristics driving cell viability
Atomistic
Enriched Descriptors
Atom-level features from ASCOT improved predictive performance
Free
Web Service
Model deployed on Enalos CHIASMA Cloud for open community access
FAIR
Data Sharing
Curated dataset published through the NanoPharos database

Related Publication

Peer-Reviewed Paper

Atom-level descriptors and explainable prediction of iron carbide nanoparticles' cytotoxicity via the Enalos Cloud platform

Antoniou M., Varsou D.-D., Tsoumanis A., Melagraki G., Lynch I., Afantitis A. — Nanoscale Advances, 2026, 8(2):646–661 — DOI: 10.1039/d5na00549c