Read across toxicity predictions with nano-lazar

Christoph Helma, Denis Gebele, Micha Rautenberg

in silico toxicology gmbh

Requirements for nanoparticle read-across

• Nanoparticle characterisation
• Toxicity measurements

eNanoMapper particle characterisation

• Nanoparticles imported: 464
• Nanoparticles with particle characterisation: 394
• Nanoparticles with toxicity data: 167
• Nanoparticles with toxicity data and particle characterisation: 160

eNanoMapper toxicity endpoints

• Toxicity endpoints: 41
• Toxicity endpoints with more than one measurement value: 22
• Toxicity endpoints with more than 10 measurements: 2

Selected data

Protein corona dataset Au particles (105 particles) Toxicity endpoint: Net cell association (A549 cell line)

Read across procedure

• Identify relevant properties (statistically significant correlation with toxicity: 14 from 30 properties)
• Calculate similarities (weighted cosine similarity with correlation coefficients as weights)
• Identify neighbors (particles with similarity > 0.95)
• Calculate prediction (weighted average from neighbors with similarities as weights)

Algorithms for feature selection, similarity calculation and predictions may change in the future.

Future development (I)

• Validation of predictions
• Applicability domain/reliability of predictions
• Accuracy improvements:
  • additional data
  • feature selection
  • similarity calculation
  • predictions (local regression models)

Future development (I)

• Usability improvements:
  • additional data (extension of applicability domain, additional endpoints and chemistries)
  • inclusion of ontologies
  • inclusion of protein corona characterisation?
  • particle characterisation without experimental data
    • descriptor calculation from core and coating chemistries
    • ontological descriptors

nano-lazar

Your comments, ideas, recommendations?