Machine Learning Allowed Interpreting Toxicity of a Fe-Doped CuO NM Library Large Data SetAn Environmental In Vivo Case Study

The wide variation
of nanomaterial (NM) characters (size, shape,
and properties) and the related impacts on living organisms make it
virtually impossible to assess their safety; the need for modeling
has been urged for long. We here investigate the custom-designed 1–10%
Fe-doped CuO NM library. Effects were assessed using the soil ecotoxicology
model Enchytraeus crypticus (Oligochaeta)
in the standard 21 days plus its extension (49 days). Results showed
that 10�-CuO was the most toxic (21 days reproduction EC50 = 650
mg NM/kg soil) and Fe3O4 NM was the least toxic
(no effects up to 3200 mg NM/kg soil). All other NMs caused similar
effects to E. crypticus (21 days reproduction
EC50 ranging from 875 to 1923 mg NM/kg soil, with overlapping confidence
intervals). Aiming to identify the key NM characteristics responsible
for the toxicity, machine learning (ML) modeling was used to analyze
the large data set [9 NMs, 68 descriptors, 6 concentrations, 2 exposure
times (21 and 49 days), 2 endpoints (survival and reproduction)].
ML allowed us to separate experimental related parameters (e.g., zeta
potential) from particle-specific descriptors (e.g., force vectors)
for the best identification of important descriptors. We observed
that concentration-dependent descriptors (environmental parameters,
e.g., zeta potential) were the most important under standard test
duration (21 day) but not for longer exposure (closer representation
of real-world conditions). In the longer exposure (49 days), the particle-specific
descriptors were more important than the concentration-dependent parameters.
The longer-term exposure showed that the steepness of the concentration–response
decreased with an increased Fe content in the NMs. Longer-term exposure
should be a requirement in the hazard assessment of NMs in addition
to the standard in OECD guidelines for chemicals. The progress toward
ML analysis is desirable given its need for such large data sets and
significant power to link NM descriptors to effects in animals. This
is beyond the current univariate and concentration–response
modeling analysis.