Application of Bayesian networks in determining nanoparticle-induced cellular outcomes using transcriptomics

Figures & data

Figure 1. A systems toxicology paradigm. The pathway aims to extrapolate all available toxicological information to humans using a combination of in silico methods and toxicokinetics modeling. Molecular level analysis provides toxicological information preceding cellular and physiological manifestation of effects. In vitro methods investigate cellular effects of direct exposure to a substance. In vivo studies provide an examination of a variety of tissues in controlled environments. Human studies require biomonitoring, sampling and tissues/fluids analysis. Population level effects is the culmination of this extrapolation with supporting evidence typically requiring longitudinal observations and large-scale data gathering.

Figure 2. Workflow of BN construction and evaluation.

Table 1. Processed Dataset (PD) inputs and outputs extracted from literature review and their states (categories)/numerical ranges and abbreviations used in this study.

Nodes	Abbreviations	States (categories)
Inputs
NPs Physicoch-emical Properties
Nanoparticle	NP	10 categories: Ag, Au, Polymeric NPs, CuO, ZnO, TiO2, SiO2, Fe2O3, Polystyrene NPs, CoFe2O4 NPs
Core size (nm)	Core size	4 discretized ranges: 0 to 10, 10 to 50, 50 to 100, >100
Shape	Shape	5 categories: Spherical, Ellipsoid, Unknown, Nanotube, Clusters
Surface coatings	Coatings	21 Categories: None, Silica, Citrate, PAH, DMSA, Madelic acid, Lipid bilayer, Hybrid lipid, Methoxyl, Diethyl glyc, Gluconic acid, Citric acid, folic acid, chromium, starch, surfactant, polymeric, colloidal, polyethylenimine, mercaptopropyl trimethoxysilane, carboxyl
Specific Surface Area (m2/g)	Surf. area	9 discretized ranges: 0 to 18, 18 to 36, 36 to 54, 54 to 72, 72 to 90, 90 to 126, 126 to 198, 198 to 216, >216
Zeta Potential (mV)	Zeta	5 discretized ranges: -30 to -15, -15 to 0, 0 to 15, 15 to 30, >30
Exposure Conditions
Exposure dose (μg/ml)	Dose	8 discretized ranges: 0 to 3, 3 to 8, 8 to 12, 12 to 30, 30 to 50, 50 to 75, 75 to 100, >100
Exposure duration (h)	Time	5 discretized ranges: 0 to 6, 6 to 12, 12 to 24, 24 to 48, >48
In vitro model characteris-tics
Cell line	C. line	22 categories: SH-SY5Y, 293T, A549, CACO-2, HDF, PC3, THP-1, HEPG2, VSMC, HACAT, HMDM, JURKAT-T, MDDC, MCF-7, IMR-90, U251, HELA, HMEC 184, EAHY926, SAE, RKO, SK MEL-28
Cell type	C. type	2 categories: Cancer or Normal
Tissue	Tissue	12 categories: Kidney, Brain, Lung, Intestinal, Skin, Prostate, Liver, Colon, Cardiovascular, Breast, Cervix, Blood
Microarray	Array	4 categories: Genechip_Affymetrix, WHG_Agilent, Beadchip_Illumina, Superarray_Bioscience
Outputs
Cellular Effects
Cell cycle and Proliferation responses	C. cycle	Binary mode (Triggered, No effect)
Cell death and Apoptosis responses	C. death
DNA damage and Repair responses	DNA dam.
Cell adhesion and Extracellular matrix responses	C. adhe
Inflammation and Immune responses	Inflam.
Unfolded protein responses (UPR) and Endoplasmatic  reticulum (ER) stress	Stress
Metal Ion Responses	Metal ion
Angiogenenesis responses	Angiogen.
Cytoskeleton organization responses	Cytoskel.

Figure 3. Results of 10-fold cross-validation Correlation Feature Selection algorithm applied in processed dataset. (A) Inter-relationship of input variables, (B) relationship between the input and output variables.

Figure 4. Graphical structure of the BN-PD model. Center: Structure based on correlation between the inputs. Perimeter: BN-PD structures based on output prediction ability of inputs.

Figure 5. Accuracy of the classifiers. The validation of the models is carried out for each cellular effect separately. (A) Internal validation using a 10-fold cross-validation process with the training 60% datasets. (B) External validation using the validation 40% datasets. (C) Reliability validation using the reliability dataset. BN-PD: BN constructed from pre-processed dataset; BN-K2: BN automated constructed using K2-algorithm; BN-RD: BN constructed from raw dataset.

Figure 6. Sensitivity of the classifiers. The validation of the models is carried out for each cellular effect separately. (A) Internal validation using a 10-fold cross-validation process with the training 60% datasets. (B) External validation using the validation 40% datasets. (C) Reliability validation using the reliability dataset. BN-PD: BN constructed from pre-processed dataset; BN-K2: BN automated constructed using K2-algorithm; BN-RD: BN constructed from raw dataset.

Figure 7. Matthews correlation coefficient of the classifiers. The validation of the models is carried out for each cellular effect separately. (A) Internal validation using a 10-fold cross-validation process with the training 60% datasets. (B) External validation using the validation 40% datasets. (C) Reliability validation using the reliability dataset. BN-PD: BN constructed from pre-processed dataset; BN-K2: BN automated constructed using K2-algorithm; BN-RD: BN constructed from raw dataset.

Table 2. Normalized information gain for each model outcome (bold) by input attributes. Only the top five most influential attributes are shown.

Angiogenesis responses		Metal Ion responses		Inflammation and Immune responses
Input attribute	Information Gain	Input attribute	Information Gain	Input attribute	Information Gain
Exposure dose	1.00	Exposure dose	1.00	Exposure dose	1.00
Nanoparticle	0.61	Nanoparticle	0.82	Cell line	0.25
Microarray	0.37	Zeta potential	0.12	Coating	0.01
Surface area	0.27	Cell line	0.03	Shape	0.00
Exposure duration	0.21	Tissue	0.02	Tissue	0.00
Unfolded protein responses and Endoplasmatic reticulum		Cell death and Apoptosis responses		Cell adhesion and Extracellular matrix responses
Input attribute	Information Gain	Input attribute	Information Gain	Input attribute	Information Gain
Nanoparticle	1.00	Exposure dose	1.00	Exposure dose	1.00
Cell line	0.73	Cell line	0.24	Nanoparticle	0.45
Zeta potential	0.44	Tissue	0.20	Microarray	0.14
Exposure dose	0.42	Coating	0.05	Tissue	0.13
Coating	0.08	Nanoparticle	0.03	Coating	0.01
Cell cycle and Proliferation responses		Cytoskeleton organization responses		DNA damage and Repair responses
Input attribute	Information Gain	Input attribute	Information Gain	Input attribute	Information Gain
Exposure dose	1.00	Exposure dose	1.00	Exposure dose	1.00
Cell line	0.32	Coating	0.59	Cell line	0.59
Nanoparticle	0.28	Nanoparticle	0.42	Nanoparticle	0.43
Microarray	0.20	Microarray	0.23	Exposure duration	0.31
Coating	0.15	Exposure duration	0.14	Cell type	0.25

Data availability statement

The data that support the findings of this study are available at (Furxhi 2019).