New paper out: doi.org/10.1016/j.im...

Here we created a QSPR for the prediction of nanomaterial ion release (particle dissolution).

We also investigated what effect artifically increasing datasets has on machine learning models.

What works:
- publisher site link
- preprint server link
- DOI / PMID / URN / SSRN / RePec / ArXiv link

What doesn’t work:
- shortlinks like bitly (DOI.org’s own service works)
- uploading the pdf to bsky
- hosting your work on your blog and sharing it from there
- deleting the link then posting

In doi.org/10.1016/j.en... we also highlight the importance of model evaluation and going beyond the OECD principles.

Applicability domain results can at times over- or underestimate prediction reliability.

Finally, it also shows how larger dataset sizes do not improve nano-QSAR performance

In doi.org/10.1016/j.en... + doi.org/10.1021/acs.... we expanded the idea of traditional QSARs by including species characteristic or ecological information to make them applicable to multiple species instead of just one

Variables related to the exposure conditions were largely unimportant. This is probably due to the lack of variance in exposure conditions (the experiments were largely standardized, so things like pH, temperature etc. were almost always the same)

We found that all ML algorithms performed pretty well but the tree based method (random forest) was the best. Identifying variables of importance led to mixed results that depended per model. In general, the molecular descriptors and p-chem properties scored high in importance

Most nano-QSARs focus on in vitro endpoints (because more data is available) but to a lesser extent on in vivo endpoints. The aim here was to look how different ML algorithms would perform on in vivo endpoints and identify important variables for nanotoxicity prediction

In doi.org/10.1016/j.ch..., we collected Daphnia magna nanotoxicity data from literature based on standardized lab experiments and created classification nano-QSARs

Will we only find out which it is by opening the (black) box (model)?