`nano-lazar`: A framework for nanoparticle read across risk assessment

Christoph Helma, Micha Rautenberg, Denis Gebele

in silico toxicology gmbh, Basel, Switzerland

eNanoMapper

`nano-lazar` framework

Framework for reproducible read-across (k-nearest-neighbor) predictions

For a given query substance (nanoparticle)

Free choice of algorithms and parameters for

Reasonable default algorithms and parameters

Endpoint:: Net cell association (121 Gold and Silver particles)
Descriptors:: Calculated fingerprints (new development), measured physchem properties, protein interactions
Feature selection:: Correlation filter (measured properties)
Similarity:: Weighted cosine (measured properties), Tanimoto/Jaccard (fingerprints)
Local regression:: Weighted average, partial least squares (PLS), random forests (RF)

Five independent 10-fold crossvalidations
No fixed random seed for training/test set splits, to avoid overfitting and to demonstrate the variability of validation results due to random training/test splits.
Separate feature selection for each training dataset to avoid overfitting

Descriptors	Algorithm	RMSE	r²	% measurements within prediction interval
MP2D	WA	2.04 2.0 2.02 2.07 2.07	0.24 0.27 0.25 0.22 0.22	NA
MP2D	PLS	2.14 2.11 2.21 1.99 1.9	0.27 0.26 0.26 0.32 0.36	94 97 91 91 97
MP2D	RF	1.84 1.67 1.68 1.69 1.71	0.4 0.5 0.49 0.48 0.47	94 96 96 94 94
P-CHEM	WA	1.91 1.93 1.91 2.03 2.02	0.48 0.47 0.49 0.41 0.42	NA
P-CHEM	PLS	2.2 2.33 2.11 2.27 2.21	0.34 0.28 0.38 0.31 0.33	97 92 96 93 91
P-CHEM	RF	1.8 1.82 1.77 1.68 1.86	0.54 0.53 0.56 0.6 0.51	94 96 97 97 93
Proteomics	WA	1.94 1.63 1.7 1.61 1.76	0.49 0.64 0.6 0.64 0.57	NA
Proteomics	PLS	1.67 1.63 1.86 1.74 1.8	0.62 0.64 0.53 0.59 0.56	90 88 84 89 88
Proteomics	RF	1.66 1.69 1.81 1.68 1.6	0.62 0.61 0.57 0.6 0.65	89 89 89 87 89
P-CHEM Proteomics	WA	1.61 1.56 1.71 1.66 2.41	0.64 0.66 0.6 0.62 0.33	NA
P-CHEM Proteomics	PLS	1.74 1.67 1.78 1.71 2.18	0.6 0.62 0.59 0.61 0.43	91 90 86 85 86
P-CHEM Proteomics	RF	1.78 1.62 1.56 1.82 1.77	0.57 0.64 0.66 0.55 0.61	88 87 87 89 90

Gold and silver particles included!

Correlation of predicted vs. measured values for five independent crossvalidations with Proteomics descriptors and local random forest models

Best results: Protein interaction descriptors and random forest models
Most results show no statistically significant difference
Calculated fingerprints perform surprisingly well (RMSE comparable to best results) and do not require measured nanoparticle properties
Local random forest models with nanoparticle fingerprints can be useful for screening purposes without physical measurements

Manuscript submitted to Frontiers in Pharmacology