Table 1.

Main features of (Q)SAR models predicting cytotoxicity of metal oxide nanoparticles.

Source	Dataset	Endpoint of Cytotoxicity Measurement	n	R2 1	Software 2	Statistical Method	Descriptors
Escherichia coli
[25]	[25]	LD50	7	0.979	-	Multiple linear regression (MLR)	Metal cation charge
[20]	[20]	LD50	17	0.862	MATLAB	MLR	Enthalpy of formation of a gaseous cation
[26]	[20]	LD50	17	0.741–0.838	CORAL	Monte Carlo	SMILES-based optimal descriptor
[27]	[20]	LD50	17	0.933	Minitab 16	MLR	Energy gap, hardness, softness, electronegativity, and electrophilicity index
[28]	[20]	LD50	17	0.81–0.90	-	MLR	Electronegativity, charge of the metal cation corresponding to a given oxide
[29]	[20]	LD50	17	0.93	RandomForest package	Random forest (RF)	S1—unbonded two-atomic fragments [Me] … [Me], which were encoded based on Simplex representation of molecular structures (SiRMS)-derived descriptors [34,35], describing distance where potential reaches minimum at van der Waals interactions; rw—Wigner–Seitz radius; ρ—mass density; (CPP)—cation polarizing power; S2—SiRMS-derived electronegativity aligned descriptor of oxides molecules—in a sense of the acid-base property of oxides (this parameter increases with a number of oxygens in molecule); S3—tri-atomic fragments [Me]-[O]-[Me], which were encoded by SiRMS-derived descriptors, encoding electronegativity; and (SV)—proportion of surface molecules to molecules in volume
[30]	[20]	LD50	17	0.955		Ensemble learning	Oxygen percent, molar refractivity, and polar surface area
[31]	[20]	LD50	17	-	MATLAB	Read-across	Ionization enthalpy of the detached metal atoms
[18]	[20]	LD50	17	0.889–0.982	CORAL	MLR	SMILES-based optimal descriptor
[36]	[20]	LD50	16	0.91	-	MLR	Enthalpy of formation of a gaseous cation (ΔHMe+), charge of the metal cation (χox), and pEC50 of HaCaT
[33]	[20]	LD50	16	0.879	SYBYL X1.1 and SPSS statistics v.17	MLR	Enthalpy of formation of a gaseous cation (ΔHme+) and polarization force (Z/r)
[37]	[20]	LD50	16	0.79	CORAL	Monte Carlo	Quasi-SMILES
[38]	[20]	LD50	17	0.92	-	Counter propagation artificial neural network	Metal electronegativity by Pauling scale, number of metal atoms in oxide, number of oxygen atoms in oxide, and charge of metal cation
[39]	[20]	LD50	17	0.968	-	RF	Oxygen in weight percentage and enthalpy of formation of a gaseous cation
[40]	[20]	LD50	17	0.877 and 0.903	-	MLR and support vector machines (SVM)	HOMO energy, α-LUMO and β-LUMO energy, the average of α-LUMO and β-LUMO, the energy gap between the frontier molecular orbitals ∆E, and molar heat capacity
[8]	[20]	LD50	17	0.93	-	Partial least squares (PLS)	Charge of metal ion, metal ion charge-based SiRMS, number of oxygen atoms in brutto formula weighted by ionic potential, covalent index weighted by charge of metal ion, molecular weight of metal oxide weighed by size of nanoparticle, squared thickness of interfacial layer, van der Waals repulsion weighted by size of nanoparticle, and Wigner-Seitz radius weighted by size of nanoparticle
[32]	[32]	LD50	17	0.87	Self-written program	MLR	Electronegativity of metal and electronegativity of metal oxide
[41]	[41]	IC50	24	-	R	SVM	Conduction band energy and hydration enthalpy (ΔHhyd)
Human keratinocyte cell line (HaCaT)
[29]	[42]	LD50	18	0.96	RandomForest package	RF	S1, rw, ρ, (CI)—covalent index of the metal ion, S2, and (AP)—aggregation parameter
[31]	[42]	LD50	18	-	MATLAB	Read-across	Mulliken’s electronegativity
[42]	[42]	LD50	18	0.93	-	MLR	Enthalpy of formation of metal oxide, Mulliken’s electronegativity
[18]	[42]	LD50	18	0.961–0.999	CORAL	MLR	SMILES-based optimal descriptor
[36]	[42]	LD50	16	0.88	-	MLR	Enthalpy of formation of metal oxide (ΔHf) nano-cluster, electronic chemical potential of the cluster, and pEC50 of E. coli
[37]	[42]	LD50	16	0.79	CORAL	Monte Carlo	Quasi-SMILES
[39]	[42]	LD50	18	0.918	-	RF	10-based logarithm of solubility measured in mol/L (LogS), topological polar surface area (TPSA), Mulliken’s electronegativity
[8]	[42]	LD50	18	0.83	-	PLS	Atom charge-based SiRMS descriptor, charge of the atom weighted by the bond ionicity, charge of metal ion weighted by ionicity of bond, squared ionic potential, ion change-based SiRMS descriptor, number of oxygen atoms in brutto formula per interfacial layer, mass density weighted by ionicity of bond, Wigner-Seitz radius weighted by ionicity of bond, and ionicity of bond based SiRMS
[43]	[42,44,45]	Cell viability (%)	21	-	CORAL	Hierarchical cluster analysis (HCA) and min–max normalization	Quasi-SMILES
Transformed bronchial epithelial cells (BEAS-2B)
[46]	[46]	% of membrane-damaged cells	9	-	Weka	RF	Atomization energy of the metal oxide, period of the nanoparticle metal, nanoparticle primary size, and nanoparticle volume fraction
[6]	[6]	Cell viability (%)	24	-	-	Regression tree	Metal solubility and energy of conduction
[47]	[6]	Cell viability (%)	24	-	RandomForest package	RF	Mass density, covalent index, cation polarizing power, Wigner–Seitz radius, surface area-to-volume ratio, aggregation parameter, and tri-atomic descriptor of atomic charges
[48]	[48]	LD50	24	-	RapidMiner	SVM	Conduction band energy and ionic index of metal cation
[49]	[50]	% of membrane-damaged cells	24	0.68	CORAL	Monte Carlo	SMILES-based optimal descriptor, dose, and exposure time
[43]	[6,51,52]	Cell viability (%)	21	0.713–0.733	CORAL	HCA and min-max normalization	Quasi-SMILES
Murine myeloid cells (RAW 264.7)
[6]	[6]	Cell viability (%)	24	-	-	Regression tree	Metal solubility and energy of conduction
[47]	[6]	Cell viability (%)	24	-	RandomForest package	RF	Mass density, molecular weight, aligned electronegativity, covalent index, surface area, surface area-to-volume ratio, two-atomic descriptor of van der Waals interactions, tetra-atomic descriptor of atomic charges, and size in DMEM
[48]	[48]	LD50	24	-	RapidMiner	SVM	Conduction band energy and ionic index of metal cation
[53]	[53]	Lactate dehydrogenase (LDH) release	25	-	R	PLS	Metal cation charge, hydration rate, radius of the metallic cation, and Pauling electronegativity
Rat L2 lung epithelial cells and rat lung alveolar macrophages
[54]	[54]	Membrane damage (units L−1)	42	-	-	Multivariate linear regression and linear discriminant analysis (LDA)	Size, concentration, size in phosphate buffered saline, size in water, and zeta potential
[55]	[54]	Membrane damage (units L−1)	42	-	-	MLR and simple classification	Size, concentration, size in phosphate buffered saline, and size in water

¹ Missing R² value means that an SAR model was built instead of QSAR. ² If software record is missing, then it was not mentioned in the original paper.