Keywords

1 Increasing Role of Nanomaterials

The history of the “nanoworld” begun on December 29, 1959, being initiated by the classic talk given at the Annual Meeting of the American Physical Society by Richard P. Feyman [1]. “There’s plenty of room at the bottom” – he summarized his visionary ideas about libraries as small as a pin head and miniature machines able to penetrate human body via the blood vessel and act as microscopic surgeons. The “nano” prefix, in a chemical context, describes particles characterized by at least one diameter of 100 nm or less. When nanoparticles are intentionally synthesized to be used in consumer goods, they are called “nanomaterials” [2].

Nowadays, 50 years after Feyman’s lecture, nanotechnology has emerged at the forefront of science and technology developments and nanomaterials have found a wide range of applications in different aspects of human life. For example, nanoparticles of such inorganic compounds as TiO2 and ZnO oxides are used in cosmetics [3], sunscreens [3], solar-driven self-cleaning coatings [4], and textiles [5]. Nano-sized CuO has replaced noble metals in newer catalytic converters for the car industry [6]. Nanopowders of metals can be used as antibacterial substrates (e.g., the combination of the pure nanosilver ion with fiber to create antiodor socks) [7]. Finally, metal salts (i.e., CdSe quantum dots) have found many applications in electronics and biomedical imaging techniques [8, 9].

The discoveries of fullerene (C60) in 1985 by Kroto et al. [10] and carbon nanotubes in 1991 by Iijima [11] opened a new area of the tailored design of carbon-based nanomaterials. Carbon-based nanomaterials are currently used, among other applications, for synthesis of polymers characterized by enhanced solubility and processability [12] and for manufacturing of biosensors [13]. They also contribute to a broad range of environmental technologies including sorbents, high-flux membranes, depth filters, antimicrobial agents, and renewable energy supplies [14].

According to current analysis [15], about 500 different products containing nanomaterials were officially on the market in 2007. Most of them (247) have been manufactured in the USA, 123 in East Asia (China, Taiwan, Korea, Japan), 76 in Europe, and only 27 in other countries. It is interesting that the number (500) is two times higher than the number of nanoproducts in the previous year. Investments in nanotechnology industry grew from $13 billion in 2004 to $50 billion in 2006 and – if one can believe the forecast – will reach $2.6 trillion in 2014.

Without doubt, nothing is able to stop such a rapidly developing branch of technology and we should be prepared for (better or worse) living day by day in symbiosis with nanomaterials.

2 Their Incredible Physical and Chemical Properties

The astonishing physical and chemical properties of engineered nanoparticles are attributable to their small size. In the nanometer-scale, finite size effects such as surface area and size distribution can cause nanoparticles to have significantly different properties as compared to the bulk material [16]. For instance, by decreasing the size of gold samples one induces color changes from bright yellow through reddish–purple up to blue.

However, from the physico-chemical viewpoint, the novel properties of nanoparticles can also be determined by their chemical composition, surface structure, solubility, shape, ratio of particles in relation to agglomerates, and surface area to volume ratio. All these factors may give rise to unique electronic, magnetic, optical, and structural properties and, therefore, lead to opportunities for using nanomaterials in novel applications and devices [16].

New, characteristic properties of nanomaterials include greater hardness, rigidity, high thermal stability, higher yield strength, flexibility, ductility, and high refractive index. The band gap of nanometer-scale semiconductor structures decreases as the size of the nanostructure decreases, raising expectations for many possible optical and photonic applications [17].

With respect to the size of the grains, it has been suggested that nanomaterials would exhibit increased (typically 3–5 times) strength and hardness as compared to their microcrystalline counterparts. For example, the strength of nanocrystalline nickel is five orders of magnitude higher than that of the corresponding microcrystalline nickel [18]. Interestingly, the observed strength of crystalline nanomaterials is accompanied by a loss of ductility, which can result in a limitation of their utility [19]. However, some of the nanocrystalline materials have the ability to undergo considerable elongation and plastic deformation without failing (even up to 100–300%). Such machinability and superplasticity properties have been observed for ceramics (including monoliths and composites), metals (including aluminum, magnesium, iron, titanium), intermetallic elements (including iron, nickel, and titanium base), and laminates [20]. Although the atomic weight of carbon nanotubes is about one-sixth of the weight of steel, their Young’s modulus and tensile strength are, respectively, five and 100 times higher than those of steel [21]. In addition, nanoparticles, because of their very small sizes and surface/interface effects such as the fundamental change in coordination, symmetry, and confinement, they may exhibit high magnetic susceptibility. A variety of nanoparticles reveal anomalous magnetic properties such as superparamagnetism. This opens new areas of potential application for them, such as data storage and ferrofluid technology [22].

According to recent studies, nanoparticles may have also great potential in medical application, mostly due to their good biocompatibility that allows them to promote electron transfer between electrodes and biological molecules. For instance, the high biocompatibility of magnetite nanocrystals (Fe3O4) makes them potentially useful as the magnetic resonance imaging contrast agents [23]. One of the unique aspects of nanoparticles is their high wettability, termed by Fujishima [24] as superhydrophilicity. Depending upon the chemical composition, the surface can exhibit superhydrophilic characteristics. For example, titanium dioxide (TiO2), at sizes below a few nm, can decrease the water contact angle to 0±1° [24]. Nano-sized composites, due to the chemical composition and viscosity of the intercrystalline phase, may provide a significant increase in creep resistance. It has been demonstrated that alumina/silicon carbide composites are characterized by a minimum creep rate, three times lower than the corresponding monolith [25].

3 Nanomaterials can be Toxic

As mentioned in Section 14.1, different types of nanomaterials are increasingly being developed and used by industry. However, little is known about their toxicity, including possible mutagenic and/or carcinogenic effects [26]. Some recent contributions report evident toxicity and/or ecotoxicity of selected nanoparticles and highlight the potential risk related to the development of nanoengineering. Evidently, there is insufficient knowledge regarding the harmful interactions of nanoparticles with biological systems as well as with the environment.

3.1 Specific Properties Cause Specific Toxicity

It is well known that the most important parameters with respect to the induction of adverse effects by a xenobiotic compound are its dose, dimension, and durability. Conversely, it is well established that nano-sized particles, due to their unique physical and chemical properties discussed above, behave differently from their larger counterparts of the same chemical composition [26–31]. Because of the difference between nanoparticles and bulk chemicals, the risk characterization of bulk materials cannot be directly extrapolated to nanomaterials.

The biological activity of nanoparticles and their unique properties causing harmful effects are highly dependent on their size. Nanoparticles, because of their small size, may pass organ barriers such as skin, olfactory mucosa, and the blood–brain barrier [32–34], readily travel within the circulatory system of a host, and deposit in target organs. This is not possible with the same material in a larger form [35]. Indeed, reduction of the particle’s size to the nanoscale level results in a steady increase of the surface to volume ratio. As a consequence, a larger number of potentially active groups per mass unit is “available” on the surface and might interact with biological systems [35]. This is one possible explanation why nano-sized particles of a given compound are generally more toxic than the same compound in its larger form [36].

However, Oberdörster et al. [37] suggested that the particle size is not the only possible factor influencing toxicity of nanomaterials. The following features should be also considered:

  • size distribution,

  • agglomeration state,

  • shape,

  • porosity,

  • surface area,

  • chemical composition,

  • structure-dependent electronic configuration,

  • surface chemistry,

  • surface charge, and

  • crystal structure.

Natural and anthropogenic nanoparticles gain access into the human body through the main ports of entry including the lungs, the skin, or the gastrointestinal tract. The unique properties of nanoparticles allow them not only to pnetrate physiological barriers but also to travel throughout the body and interact with subcellular structures. Toxicological studies show that nanoparticles can be found in various cells such as mitochondria [38, 39], lipid vesicles [40], fibroblasts [41], nuclei [42], and macrophages [43].

3.2 Oxidative Stress

Depending on their localization inside the cell, nanoparticles can induce formation of reactive oxygen species (ROS), for instance, superoxide radicals, hydroxyl radicals reactive nitrogen [44], sulfur [45], and other species stressing the body in a similar manner to the effect of ROS [46]. This results in oxidative stress and inflammation, leading to the impacts on lung and cardiovascular health [16].

It is worth noting that normally, due to the presence of antioxidant molecules (i.e., vitamin C and glutathione), the body’s cells are able to defend themselves against ROS and free radicals damage. However, when a large dose of strongly electrophilic nanoparticles enter the body, the balance between reduced glutathione (GSH) and its oxidized form (GSSG) is destroyed [47] and the unscavenged oxidants cause cell injuries by attacking DNA, proteins, and membranes [48]. At the cellular level, oxidative stress is currently the best developed paradigm depicting the harmful effects of nano-sized particles [31, 49, 50].

3.3 Cytotoxicity and Genotoxicity

The mechanism of oxidative stress occurring at the molecular level is mainly responsible for observed cytotoxic and genotoxic effects induced by nanoparticles. Cytotoxicity of selected nanospecies has been confirmed by many researchers. For example, fullerene (C60) particles suspended in water are characterized by antibacterial activity against Escherichia coli and Bacillus subtilis [51] and by cytotoxicity to human cell lines [52]. Single multiwalled carbon nanotubes (CWCNTs and MWCNTs) are also toxic to human cells [41, 53]. Nano-sized silicon oxide (SiO2), anatase (TiO2), and zinc oxide (ZnO) can induce pulmonary inflammation in rodents and humans [54–56].

Epidemiological studies have shown that nanoparticles might be genotoxic to humans [57]. Irreversible DNA modifications resulting from the activity of ROS may lead to heritable mutations, involving a single gene, a block of genes, or even whole chromosomes. DNA damage may also disrupt various normal intracellular processes, such as DNA replication and modulate gene transcription, causing abnormal function or cell death [16, 44, 58]. Until now, more than 100 different oxidative DNA lesions have been found. The most investigated OH-related DNA lesions is 8-hydroxydeoxyguanosine (8-OHdG) [59], which may be induced by several particles such as asbestos, crystalline silica, coal fly ashes. Oxygen free radicals may overwhelm the antioxidant defense system by mediating formation of base adducts, such as 8-hydroxydeoxyguanosine, and therefore play a key role in initiation of carcinogenesis [60].

3.4 Neurotoxicity

Data on neurotoxic effects of engineered nanoparticles are very limited, but it has been reported that inhaled nanoparticles, depending on their size, may be distributed to organs and surrounding tissues, including the olfactory mucosa or bronchial epithelium and then can be translocated via the olfactory nerves to the central nervous system [61]. There is also some evidence that nano-sized particles can penetrate and pass along nerve axons and dendrites of neurons into the brain [33]. Recent studies confirm the translocation of nanoparticles from the respiratory tract into the central nervous system; for example, inhalation with 30 nm magnesium oxide in rats showed that manganese can be taken up into olfactory neurons and accumulated in the olfactory bulb [34].

The particles at the nanoscale may also gain access to the brain across the blood–brain barrier [2]. There is experimental evidence that oxidative stress also plays an important role in neurodegenerative diseases and brain pathology, for instance, Hallervorden-Spatz Syndrome, Pick’s disease, Alzheimer’s disease, or Parkinson’s disease [62].

3.5 Immunotoxicity

The effects of nanoparticles on the immune system are still unclear. Although the reticuloendothelial system (RES) is able to eliminate nanoparticles, several toxicological studies have suggested that nanoscale particles’ interaction with the defense activities of immune cells can change their antigenicity and stimulate and/or suppress immune responses. Direct experiments showed that dendritic cells and macrophages uptake of nanoparticle–protein complexes may change the formation of the antigen and initiate an autoimmune response [16]. Several studies have also reported that nanoparticles may induce damage to red blood cells (erythrocytes). Bosi et al. [63] have studied the hemolytic effect of different water-soluble C60 fullerenes. Preliminary results indicate that hemolytic activity depends on the number and position of the cationic surface groups. However, no clinically relevant toxicity has yet been demonstrated [64].

3.6 Ecotoxicity

Nano-sized particles such as volcanic ash, dust storms, or smoke from natural fires have always been present in the environment. However, the recent progress of industry has increased engineered nanoparticle pollution. The unique size-specific behavior and specific physical–chemical properties, in combination with toxicity to particular living organisms, may also result in harmful effects on the level of whole environmental ecosystems [65].

In the pioneering report on the non-human toxicity of fullerene, Eva Oberdörster [66] observed that manufactured nanomaterials can have negative impacts on aquatic organisms. Water-soluble C60 fullerenes cause oxidative damage (lipid peroxydation in the brain) and depletion of glutathione in the gill of juvenile largemouth bass (Micropterus salmoides) at a concentration of 0.5 ppm. However, these results might be disputable, because the authors used the organic solvent tetrahydrofuran (THF) to disaggregate C60 fullerenes, THF is classified as a neurotoxin [67].

Subsequently, Lover and Klaper [68] observed the toxicological impact of nanoparticles of fullerenes (C60) and titanium dioxide (TiO2) to Daphnia magna: C60 and TiO2 caused mortality with a LC50 value of 5.5 ppm for TiO2 and a LC50 value of 460 ppb for the fullerene. In this case the authors also used THF for solubilization of hydrophobic C60, thus the results are also of lower credibility. Interestingly, in similar experiments by Andrievsky et al. [69] with “fullerene water solutions” (hydrated fullerenes, C60 · nH2O), no mortality was observed.

In a later study, Adams et al. [70] confirmed the acute toxicity of selected nano-sized metal oxides against D. magna. He observed that SiO2 particles were the least toxic and that toxicity increased from SiO2 to TiO2 to ZnO. A further study by the authors [71] showed that these three photosensitive nanoscale metal oxides in water suspensions have similar antibacterial activity to Gram-positive (B. subtilis) and Gram-negative (E. coli) bacteria (SiO2 < TiO2 < ZnO). All the metal oxides nanoparticles tested inhibited the growth of both Gram-positive and Gram-negative bacteria; however, B. subtilis was more sensitive than E. coli.

Similar results have been observed for a bath of ZnO, TiO2, and CuO against bacterium Vibrio fischeri and crustaceans D. magna and Thamnocephalus platyurus [72]. The antibacterial effects of nano-sized metal oxides to V. fischeri were similar to the rank of toxicity to D. magna and T. platyurus; they increased from TiO2 to CuO and ZnO. It is also very important to recognize that titanium dioxide was not toxic even at the 20 g/l level, which means that not all nanoparticles of metal oxides induce toxicity.

Smith et al. [73] investigated the ecotoxicological potential of single-walled carbon nanotubes (SWCNT) to rainbow trout (Oncorhynchus mykiss) showing that the exposure to dispersed SWCNT causes respiratory toxicity – an increase of the ventilation rate, gill pathologies, and mucus secretion. Additionally, the authors observed histological changes in the liver, brain pathology, and cellular pathologies, such as individual necrotic or apoptotic bodies, in rainbow trout exposed to 0.5 mg/l SWCNT.

Mouchet et al. [74] analyzed the acute toxicity and genotoxicity of double-walled carbon nanotubes (DWNTs) to amphibian larvae (Xenopus laevis). The authors did not observe any effects at concentrations between 10 and 500 mg/l. However, at the highest concentrations (500 mg/l) 85% of mortality was measured, while at the lowest concentrations (10 mg/l) reduced size and/or a cessation of growth of the larvae were observed.

Summarizing this section, there is strong evidence that chemicals, when synthesized at the nanoscale, can induce a wide range of specific toxic and ecotoxic effects. Moreover, even similar compounds from the same class can differ in toxicity. The available data on toxicity are still lacking; thus, more comprehensive and systematic studies in this area are necessary and very important.

4 “NANO-QSAR” – Advances and Challenges

As demonstrated in this book, quantitative structure–activity relationship (QSAR) methods can play an important role in both designing new products and predicting their risk to human health and the environment. However, taking into account the specific properties of nanomaterials and their still unknown modes of toxic action, this class of compounds seems to be much more problematic for QSAR modelers than the “classic” (small, drug-like) chemicals.

4.1 Description of Structure

Until now, more than 5000 different descriptors have been developed and used for the characterization of molecular structure (Chapter 3). In general, the descriptors can be classified according to their dimensionality. Constitutional descriptors, so-called “zero-dimensional,” are derived directly from the formula (e.g., the number of oxygen atoms). Descriptors of bulk properties, such as n-octanol/water partition coefficient or water solubility, are classified as “one-dimensional” descriptors. Topological descriptors based on the molecular graph theory are called “two-dimensional” descriptors and characterize connections between individual atoms in the molecule. “Three-dimensional” descriptors reflect properties derived from the three-dimensional structure of a molecule optimized at the appropriate level of quantum-mechanical theory. “Four-dimensional” descriptors are defined by molecular properties arising from interactions of the molecule with probes characterizing the surrounding space or by stereodynamic representation of a molecule, including flexibility of bonds, conformational behavior, etc. [75–79]. Only a little is known about applicability of those “traditional” descriptors for the characterization of nanostructures. Some authors [80–82] postulate that the existing descriptors are insufficient to express the specific physical and chemical properties of nanoparticles. Thus, novel and more appropriate types of the descriptors must be developed.

A group of nanoparticles is structurally diversified. In fact, this group has been defined arbitrarily in some way, taking into account size as the only criterion of the particles’ membership. Therefore, structures as various as nanotubes, fullerenes, crystals, and atom clusters as well as chemical species of such different properties as metals, non-metals, organic compounds, inorganic compounds, conductors, semi-conductors, and isolators were put together into one single group. Since nanoparticles are not structurally homogenous, a common mechanism of toxicity cannot be expected for all of them. In consequence, toxicity and other properties should be studied within the most appropriately chosen sub-classes of structural and physico-chemical similarity.

What is the best way to define the sub-classes? The answer might be given based on a stepwise procedure recommended by the OECD guidance document on the grouping of chemicals [83] (see also Chapter 7). Along with the guidelines, the following eight steps should be performed:

  1. 1.

    Development of the category hypothesis, definition, and identification of the category members. The category can be defined based on chemical similarity, physico-chemical properties, toxicological endpoint, and/or mechanism of action, as well as in terms of a metabolic pathway.

  2. 2.

    Gathering of data for each category members. All existing data should be collected for each member of the category.

  3. 3.

    Evaluation of available data for adequacy. The data should be carefully evaluated at this stage according to the commonly accepted protocols (i.e., according to the appropriate OECD guidance).

  4. 4.

    Construction of a matrix of data availability (category endpoints vs. members). The matrix is to indicate whether data are available or not.

  5. 5.

    Performing of a preliminary evaluation of the category and filling data gaps. The preliminary evaluation should indicate if (i) the category rationale is supported and (ii) the category is sufficiently robust for the assessment purpose (contains sufficient, relevant and reliable information).

  6. 6.

    Performing of additional testing (experiments). Based on the preliminary evaluation (especially evaluation of the robustness), additional experiments and group members for testing can be proposed.

  7. 7.

    Performing of a further assessment of the category. If new data from the additional testing are generated, the category should be revised according to the criteria from step 5.

  8. 8.

    Documenting of the finalized category. Finally, the category should be documented in the form of a suitable reporting format proposed by the guidance.

The currently proposed [82] working classification scheme for nanostructured particles includes nine categories:

  1. 1.

    spherical or compact particles;

  2. 2.

    high aspect ratio particles;

  3. 3.

    complex non-spherical particles;

  4. 4.

    compositionally heterogeneous particles – core surface variation;

  5. 5.

    compositionally heterogeneous particles – distributed variation;

  6. 6.

    homogeneous agglomerates;

  7. 7.

    heterogeneous agglomerates;

  8. 8.

    active particles;

  9. 9.

    multifunctional particles.

This classification has been adapted from the original work of Maynard and Aitken [84].

What types of structural properties should be described within the groups? As previously discussed in Section 14.3, the diameter of a nanoparticle is important, but it is not the only one possible factor influencing the mode of toxic action. The additional structural characteristics which must also be appropriately expressed are size distribution, agglomeration, shape, porosity, surface area, chemical composition, electronic configuration, surface chemistry, surface charge, and crystal structure. In contrast to the classic QSAR scheme, an entire characterization of a nanostructure may be impossible only when computational methods are employed. Novel descriptors reflecting not only molecular structure, but also supra-molecular pattern (size, shape of the nanoparticles, etc.) should be derived from both computational and experimental techniques.

The fastest and relatively easy step of characterizing the structure is the calculation of constitutional and topological descriptors. An interesting and very practical idea in this field is to replace a series of simple descriptors by one, so-called “technological attributes code” or “SMILES-like code” [85–88]. For instance, a nanoparticle of ceramic zirconium oxide, existing in bulk form and synthesized at a temperature of 800°C can be expressed by the code “Zr,O,O,CER,%E” [80]. Similar to the simplified molecular input line entry system (SMILES), the international chemical identifier (InChI) might also be used directly as a descriptor of chemical composition [89]. Another possibility is to apply descriptors derived from either molecular graph (MG) or the graphs of atomic orbitals (GAO) theory [90–92]. In the first case, vertexes in the graph represent atoms, while edges represent covalent bonds. In the second method, vertexes refer to particular atomic orbitals (1s, 2s, 2p, etc.), while edges connect the orbitals belonging to different atoms (Figure 14-1). Based on the molecular graphs, Faulon and coworkers [93–96] have developed the signature molecular descriptor approach for the characterization of fullerenes and nanotubes. The signature is a vector including extended valences of atoms derived from a set of subgraphs, following the five-step algorithm:

  1. 1.

    constructing of a subgraph containing all atoms and bonds that are at a distance no greater than the given signature height;

  2. 2.

    labeling the vertices in a canonical order;

  3. 3.

    constructing a tree spanning all the edges;

  4. 4.

    removing of all canonical labels that appear only one time;

  5. 5.

    writing the signature by reading the tree in a depth-first order.

Figure 14-1.
figure 14_1_156654_1_En

Molecular graph (MG) and graph of atomic orbitals (GAO) for SnO2 (vertex numbering and vertex degrees). [90–92, 132]

Full size image

The signature descriptor can be utilized not only for direct QSAR modeling, but also for calculating a range of topological indices (i.e., the Wiener index).

Without doubt, simplicity of calculation is the most significant advantage of the topological descriptors. However, in many cases these two-dimensional characteristics are insufficient to investigate more complex phenomena. In such a situation, a more sophisticated approach must be employed to describe the structure appropriately. As mentioned previously, quantum-mechanical calculations can deliver useful information on the three-dimensional features (see Chapter 2). Among others, they include: molecular geometry (bond lengths, valence, and torsion angles), electron distribution, ionization potential, electron affinity, surface reactivity, and band gap. When performing quantum-mechanical calculations, there are always two important assumptions to be introduced. First one is an appropriate molecular model; the second one is the appropriate level of the theory. Both assumptions are closely related: when the model (system) is too large, the calculations at the highest levels of the theory are impossible, because of large computational time and other technical resources to be required [97].

Small fullerenes and carbon nanotubes can be treated as whole systems and modelled directly with quantum-mechanical methods. Among the theory levels, the density functional theory (DFT) recently seems to have been accepted as the most appropriate and practical choice for such calculations. Indeed, DFT methods can serve as a good alternative for conventional ab initio calculations, when a step beyond the means field approximation is crucial and the information on the electron correlation significantly improves the results (e.g., Hartree–Fock – HF method in conjunction with Møller-Pleset the second-order correction – MP2). Unfortunately, even “small” fullerenes and carbon nanotubes (containing between 40 and 70 carbon atoms) are, in fact, large from quantum-mechanical point of view. Therefore, the “classic” ab initio calculations might be impractical because of the reasons mentioned in the previous paragraph, whereas DFT can be performed in reasonable time.

The functional commonly utilized for DFT is abbreviated with the B3LYP symbol. In B3LYP calculations (Eq. 14-1) the exchange-correlation energy EXC is expressed as a combination (a0, aX, and aC are the parameters) of four elements: (i) the exchange-correlation energy from the local spin density approximation \( (\rm{LSDA},\ {\rm E^{LSDA}_{xc}})\), (ii) the difference between the exchange energy from Hartree–Fock and LSDA , (iii) Becke’s exchange energy with gradient correction , and (iv) the correlation energy with Lee-Yang-Parr correction [98, 99]:

$${\rm{E}}_{{\rm{XC}}} = {\rm{E}}_{{\rm{XC}}}^{{\rm{LSDA}}} + {\rm{a}}_0 ({\rm{E}}_{\rm{X}}^{{\rm{HF}}} - {\rm{E}}_{\rm{X}}^{{\rm{LSDA}}} ) + {\rm{a}}_{\rm{X}} {\rm{E}}_{\rm{X}}^{{\rm{B88}}} + {\rm{a}}_{\rm{C}} {\rm{E}}_{\rm{C}}^{{\rm{LYP}}} $$
((14-1))

Sometimes, when a system is too large from the quantum-mechanical point of view, the calculations are practically impossible. The situation is very common for larger crystalline nanoparticles (i.e., nanoparticles of metal oxides: TiO2, Al2O3, SnO2, ZnO, etc.) and, in such cases, a simplified model of the whole structure must first be appropriately selected. In general, there are two strategies for modeling of crystalline solids: (i) an application of the periodic boundary conditions (PBSs) and (ii) calculations based on the molecular clusters. In the first approach, calculations for a single unit cell are expanded in the three dimensions with respect to the translational symmetry by employing appropriate boundary conditions (i.e., the unit cell should be neutral and should have no dipole moment). In doing so, the model includes information on the long-range forces occurring in the crystal. However, the cell size should be large enough to also be able to model defects in the surface and to eliminate the spurious interactions between periodically repeated fragments of the lattice [100–102].

In the second approach, a small fragment or so-called “cluster,” is cut off from the crystal structure and then used as a simplified model for calculations. The only problem is how to choose the diameter of the cluster correctly? This must be performed by reaching a compromise between the number of atoms (and thus the required time of computations) and the expected accuracy (and hence level of the theory to be employed). It is worth mentioning that the molecular properties can be divided into two groups depending on how they change with increasing size of the cluster (going from molecular clusters to the bulk form). They are (i) scalable properties, varying smoothly until reaching the bulk limit and (ii) non-scalable properties, when the variation related to increasing size of the cluster is not monotonic. Although the cluster models usually avoid the long-range forces, they have found many applications in modeling of local phenomena and interactions on the crystal surface [103].

As previously mentioned, in addition to calculated properties, experimentally derived properties may also serve as descriptors for developing nano-QSARs (Table 14-1). The experimental descriptors seem to be especially useful for expressing size distribution, agglomeration state, shape, porosity, and irregularity of the surface area. Interestingly, the experimental results can be combined with numerical methods to define new descriptors. For example, images taken by scanning electron microscopy (SEM), transmission electron microscopy (TEM), or atomic force microscopy (AFM) (Figure 14-2) might be processed with use of novel chemometric techniques of image analysis. Namely, a series of images for different particles of a given nanostructure should first be taken. Then, the pictures must be numerically averaged and converted into a matrix containing numerical values that correspond to intensity of each pixel in the gray scale or color value in the RGB scale. New descriptors can be defined based on the matrix (i.e., a shape descriptor can be calculated as a sum of non-zero elements in the matrix; porosity – as a sum of relative differences between each pixel and its “neighbors,” etc.) [104].

Figure 14-2.
figure 14_2_156654_1_En

Nanopowder – SEM image of nano-sized SnO2

Full size image
Table 14-1. Experimental properties for possible use as descriptors in nano-QSAR studies [105]
Full size table

Without doubt, an appropriate characterization of the nanoparticles’ structure is currently one of the most challenging tasks in nano-QSAR. Although more than 5000 QSAR descriptors have been defined so far, they may be inadequate to express the supramolecular phenomena governing the unusual activity and properties of nanomaterials. As a result, much more effort in this area is required.

4.2 Nanostructure – Electronic Properties Relationships

An important step related to the numerical description of chemical structure and QSAR modeling involves establishing a qualitative relationship between the structure of a nanoparticle and its various electronic properties.

The B3LYP functional and the standard 6-31G(d) Polple’s style basis set were applied by Shukla and Leszczynski [106] to investigate the relationships between the shape, size, and electronic properties of small carbon fullerenes, nanodisks, nanocapsules, and nanobowls. They found out that the ionization potentials decrease, while the electron affinities increase in going from the C60 fullerenes to the closed nanodisks, capsules, and open bowl-shaped nanocarbon clusters. In similar studies performed for capped and uncapped carbon nanotubes at the B3LYP/6-31G(d) level of theory by Yumura et al. [107, 108], the authors demonstrated that the tube lengths, edge structures, and end caps play an important role in determining the band gap expressed as a difference between the energies of the highest occupied and lowest unoccupied molecular orbitals (HOMO–LUMO) and vibrational frequencies. Wang and Mezey [109] characterized electronic structures of open-ended and capped carbon nanoneedles (CNNs) at the same theory level (B3LYP/6-311G(d)) concluding that conductivity of the studied species is strictly correlated to their size. Only very long CNNs structures have band gaps sufficiently narrow to be semiconductors, while the band gaps of very short and thin structures are too large to conduct electrons. Similarly, Poater et al. [110, 111] observed that the Parr electrophilicity and electronic movement described by the chemical potential increase with increasing length of the carbon nanoneedles and very “short” structures (containing four layers and less) have a HOMO–LUMO gap too large to allow conductivity. Moreover, Simeon et al. [112], by performing B3LYP calculations, demonstrated that a replacement of the fullerene carbon atom with a heteroatom results in a significant change of electronic and catalytic properties of the fullerene molecule.

Similar studies have been performed for crystalline metal semi-conductors with the use of the cluster calculations. As mentioned in Section 14.4.1, some electronic properties are scalable. They change with the changing size of the cluster until the bulk limit is reached. Known examples of such properties are the HOMO–LUMO gap (band gap) and the adiabatic electron detachment energy. For instance, the band gap of ZnO nanoparticles decreases with increasing diameter of the particle up to the bulk value observed for about 4 nm [113]. Similarly, the bulk limits of the HOMO–LUMO gap and the detachment energy for titanium oxide anion clusters of increasing size (increasing n) were reached already for n=7 [114, 115].

In the classic formalization of QSARs, electronic properties (e.g., HOMO, LUMO, ionization potential) have been utilized as “ordinary” molecular descriptors. As discussed above, this approach should be revised for nanoparticles, for which the properties vary with size of a particle and this variation cannot be simply described by a linear function. It is not out of the question that similar phenomena might be observed also for other types of the “traditional” descriptors and further studies in this area are required and strongly justified.

4.3 Nano-QSAR Models

Regarding the five OECD principles for the validation of a (Q)SAR as discussed in Chapters 12 and 13, an ideal QSAR model, applicable for regulatory purpose, should be associated with (i) a well-defined endpoint; (ii) an unambiguous algorithm; (iii) a defined domain of applicability; (iv) appropriate measures of goodness-of-fit, robustness, and predictivity; and (v) a mechanistic interpretation, if possible. Unfortunately, it is extremely difficult to fulfill all of these principles for (Q)SARs applicable to nanomaterials. There are two main difficulties related to the development of nano-QSARs. The first one is lack of sufficiently numerous and systematic experimental data, while the second one is very limited knowledge on mechanisms of toxic action.

As we mentioned many times, regarding their structure, the class of nanomaterials is not homogenous, combining a range of physico-chemical properties, as well as possible mechanisms of metabolism and toxicity. Thus, it is impossible to assume one common applicability domain for all nanomaterials. Each mode of toxicity and each class of nanomaterials should be studied separately. Analyzing the literature data (Section 14.3) it must be concluded that even if a class of structurally similar nanoparticles is tested with the same laboratory protocol, the number of tested compounds is often insufficient to perform comprehensive internal and external validation of a model and to calculate the appropriate measures of robustness and predictivity in QSAR. For instance, Limbach et al. [116] have proposed two rankings of cytotoxicity of seven oxide nanoparticles based on the in vitro study of human and rodent cells. The rankings were as follows: (i) Fe2O3≈asbestos > ZnO > CeO2≈ZrO2≈TiO2≈Ca3(PO4)2 and (ii) ZnO > asbestos≈ZrO2 > Ca3(PO4)2≈Fe2O3≈CeO2≈TiO2, respectively, for human (mesothelinoma) and rodent cells. In another paper by the same research group, the authors have found that for four metal nanoparticles – namely, TiO2, Fe2O3, Mn3O4, and Co3O4 – the chemical composition was the main factor determining the formation of reactive oxygen responsible for toxicity toward human lung epithelial cells [117]. Obviously, the results cannot be combined together and a data set containing five or six compounds is too small to build an appropriately validated QSAR model.

Do these restrictions and problems mean QSAR modelers are not able to provide useful and reliable information for nanoparticles? We do not believe this to be true. The amount of data will increase along with increasing number of nanotoxicological studies. However, no one can expect the accumulation in the next few years of such extensive data for nanomaterials, as it is now available for some environmental pollutants, pharmaceuticals, and “classical” industrial chemicals [118, 119]. Despite the limitations, there are some very promising results of preliminary nano-QSAR studies which are reviewed below.

Toropov et al. [81] have developed two models defining the relationships between basic physico-chemical properties (namely, water solubility, log S, and n-octanol/water partition coefficient, log P) of carbon nanotubes and their chiral vectors (as structural descriptors). The two-element chiral vector (n, m) contains information about the process of rolling up the graphite layer when a nanotube is formed. It had been previously known [120] that the elements of the chiral vector are related to conductivity. At this point, Toropov et al. confirmed, using the QSPR-based research, that the vector is also strictly related to other properties. The models developed were defined by the following two equations (Eqs. 14-2 and 14-3):

$$\begin{array}{l} {\rm{log}}\;S = - 5.10 - 3.51n - 3.59m \\ R^2 = 0.99,\;s = 0.053,\;F = 126 \\ \end{array}$$
((14-2))
$$\begin{array}{l} {\rm{log}}\;P = - 3.92 + 3.77n - 3.60m \\ R^2 = 0.99,\;s = 0.37,\;F = 2.93 \\ \end{array}$$
((14-3))

The study was based on experimental data being available for only 16 types of carbon nanotube. To perform an external validation, the authors divided the compounds into a training set (n=8) and a test set (n test=8). Statistics of the validation were \(R^2_{\rm test}=0.99\), s test=0.093, and F test=67.5 and R 2 test=0.99, s test=0.29, and F test=5.93, respectively, for the models for water solubility and n-octanol/water partition coefficient. Without doubt, these were the first such QSPR models developed for nanoparticles. However, the ratio of descriptors to compounds (the Topliss ratio) was low, thus the model might be unstable (see discussion in Chapter 12 for more detail).

Another contribution by Toropov and Leszczynski [80] presents a model predicting Young’s modulus (YM) for a set of inorganic nanostructures (Eq. 14-4).

$$\begin{array}{l} {YM} = - 3720.0(\pm 39.9) + 3950.0(\pm 39.2){DCW} \\ R^2 = 0.98,\;s = 18.3,\;F = 761, \\ R_{{\rm{test}}}^2 = 0.90,\;s_{{\rm{test}}} = 34.7,\;F_{{\rm{test}}} = 51 \\ \end{array}$$
((14-4))

The model was calibrated with a training set of 21 compounds and validated with eight compounds, thus the Topliss ratio in this case was satisfactory. The values of DCW descriptor were calculated from the Smiles-like code, according to the following equation (Eq. 14-5):

$${DCW} = \prod\limits_{k = 1}^N {{CW}(I_k )} $$
((14-5))

where I k is the component information on the nanostructure (e.g., Al, N, BULK, refer to Section 14.4.1), CW(I k ) is the correlation weight of the component I k , and N is the total number of these components in a given nanostructure. The values of CW(I k ) were calculated by the Monte Carlo method with the software developed by the authors. The model was correctly validated and the authors demonstrated the possibility of the prediction the Young’s modulus for external compounds with QSAR.

Martin et al. [121] have proposed two QSAR models predicting the solubility of buckminsterfullerene (C60), respectively, in n-heptane (log S heptane) and n-octanol (log S octanol) (Eqs. 14-6 and 14-7):

$$\begin{array}{l} {\rm{log}}\;S_{{\rm{heptane}}} = 3.49( \pm 3.46) + 76.98( \pm 8.11){RNCG} - 9.56( \pm 2.25) {^2ASIC}\\ \qquad\qquad\qquad - 1.18( \pm 0.45){E}_{{\rm ee}}^{{\rm min}} (CC) \\ n = 15,\;R^2 = 0.90,\;s^2 = 0.18,\;F = 34.8, \\ n_{{\rm{test}}} = 3,\;Q^2 = 0.84,\;R_{50}^2 = 0.82,\; s_{50}^2 = 0.35 \\ \end{array}$$
((14-6))
$$\begin{array}{l} {\rm{log}}\;S_{{\rm{octanol}}} = 10.5( \pm 1.30) - 8.40 \times 10^{ - 2} ( \pm 7.71 \times 10^{ - 3} )^1 {IC} - 1.57( \pm 0.16)\\ \qquad\qquad\qquad{E}_{{\rm ee}}^{{\rm min}} ({\rm{C}}{\rm{C}}) + 0.88( \pm 0.15){RPCS} \\ R^2 = 0.96,\;s^2 = 0.078,\;F = 97.3, \\ {\rm{Q}}^2 = 0.93,\;{\rm{R}}_{50}^2 = 0.96,\;{\rm{s}}_{50}^2 = 0.10 \\ \end{array}\\$$
((14-7))

The symbols \(R_{50}^2\) and \(s_{50}^2\) refer to leave-50%-out cross-validation. The authors applied CODESSA descriptors, namely, RNCG – relative negative charge (Zefirov’s PC); 2 ASIC – average structural information content of the second order; \(E_{\rm ee}^{\rm min}\) (CC) – minimum exchange energy for a C–C bond; 1 IC – first-order information content; and RPCS – relative positive charged surface area. Interestingly, the models were calibrated on 15 compounds including 14 polycyclic aromatic hydrocarbons (PAHs) containing between two and six aromatic rings and the fullerene. Although values of solubility predicted for the fullerene seem to be reasonable, the authors did not validate the applicability domain of the models. Indeed, the structural difference between 14 hydrocarbons and the fullerene is probably too large to make reliable predictions for C60 (the polycyclic hydrocarbons are planar, but the fullerene is spherical). In addition, the experimental values of log S for 14 PAHs ranged from –3.80 to 0.22 in heptane and from −3.03 to −0.02 in octanol, while the experimental values for the fullerene were −4.09 and 4.18 in heptane and octanol, respectively.

An interesting area of nano-QSAR applications is estimating solubility of a given nanoparticle in a set of various solvents. In that case, the main purpose of molecular descriptors is to correctly characterize the variation in interactions between the particle and the molecules of different solvents [122]. In fact, it means that the descriptors are related to the structure of solvents rather than to the nanoparticle structure.

Murray et al. [123] have developed a model characterizing the solubility of C60 in 22 organic solvents by employing three following descriptors: two quantities, and υ reflecting variability and degree of balance of electrostatic potential on the solvent surface and the surface area, SA (Eq. 14-8).

$${\rm{log}}(S \times 10^4 ) = - 29.0\left[ {\sigma _{{\rm{tot}}}^2 /(SA)^{3/2} } \right] + 1.28\left( {\upsilon \sigma _{{\rm{tot}}}^2 } \right)^{1/2} + 1.53 \times 10^{ - 9} (SA)^4 - 2.72$$
((14-8))

Although the model is well fitted (R=0.95, s=0.48), nothing is known about its predictive ability, because the model has not been validated.

A set of linear models built separately for individual structural domains, namely alkanes (n=6), alkyl halides (n=32), alcohols (n=6), cycloalkanes (n=6), alkylbenzenes (n=16), and aryl halides (n=9), was published by Sivaraman et al. [124]. The models were based on connectivity indices, numbers of atoms, polarizability, and variables indicating the substitution pattern as molecular descriptors for the solvents. The values of R 2 for particular models ranged between 0.93 (alkyl halides) and 0.99 (cycloalkanes) with the corresponding values of s from 0.22 (alkyl halides) to 0.04 (cycloalkanes). The authors concluded that it was impossible to obtain a unified model that included all solvents. However, when the first three classes of solvents (i.e., alkanes, alkyl halides, and alcohols) were combined together into one model, the results of an external validation performed were satisfactory.

As well as linear approaches, non-linear models have been constructed. For instance, Kiss et al. [125] applied an artificial neural network utilizing molar volume, polarizability parameter, LUMO, saturated surface, and average polarizability as structural descriptors of solvents. They observed that for most of the solvents studied (n=126) solubility decreases with increasing molar volume and increases with polarizability and the saturated surface areas of the solvents. The reported value of s in that case was 0.45 of log units. The values of R 2 and F were 0.84 and 633, respectively.

In another study [126] the authors proposed modeling with both multiple linear regression with heuristic selection of variables (HM-MLR) and a least-squares support vector machine (SVM). Then they compared both models with each other. Both models were developed with CODESSA descriptors [127]. Interestingly, the results were very similar (the model using SVM had slightly better characteristics). The values of R 2 for the linear and non-linear model were, respectively, 0.89 and 0.90, while the values of F were 968 and 1095. The reported root mean square errors were 0.126 for the linear model (HM-MLR) and 0.116 for the model employing SVM. When analyzing all the results it might be concluded that the main factor responsible for differences in the model error is related to the type of the descriptors rather than to the mathematical method of modeling.

Recently, Toropov et al. [89] developed an externally validated one-variable model for C60 solubility using additive optimal descriptors calculated from the International Chemical Identifier (InChI) code (Eq. 14-9):

$$\begin{array}{l} {\rm{log}}\;S = - 7.98(\pm 0.14) + 0.325(\pm 0.0010)\;{DCW}({\rm{InChI}}) \\ n = 92,\;R^2 = 0.94,\;Q^2 = 0.94,\;s = 0.25,\;F = 1540, \\ n_{{\rm{test}}} = 30,\;R_{{\rm{test}}}^2 = 0.94,\;s_{{\rm{test}}} = 0.35,\;F_{{\rm{test}}} = 437 \\ \end{array}$$
((14-9))

The descriptor DCW(InChI) is defined as the sum of the correlation weights CW(I k ) for individual IChI attributes I k characterizing the solvent molecules. The example of the DCW(InChI) calculation is presented in Table 14-2. The values of CW(I k ) were optimized by the Monte Carlo method.

Table 14-2. Illustration of the DCW calculation using pentane as an example (InChI: 1/C5H12/c1-3-5-4-2/h3-5H2,1-2H3). The value of DCW(InChI)=6.9256652 [89]
Full size table

All of the above models refer to physico-chemical properties as the endpoints, thus they are also termed quantitative structure–property relationships (QSPRs). Currently, there are only a small number of QSARs related directly to nanomaterials’ activity. In 2007 Tsakovska [128] proposed the application of QSAR methodology to predict protein–nanoparticle interactions. In 2008 Durdagi et al. published two papers [129, 130] presenting QSAR-based design of novel inhibitors of human immunodeficiency virus type 1 aspartic protease (HIV-1 PR). In the first work [130] the authors developed a three-dimensional QSAR model with comparative molecular similarity indices analysis (CoMSIA) method for 49 derivatives of fullerene C60. The values of R 2 and Q 2 for the training set (n=43) were 0.99 and 0.74, respectively. The absolute values of residuals in the validation set (n=6) ranged from 0.25 to 0.99 logarithmic units of EC 50 (μM). The second model [129] were characterized by lower values of the statistics (n=17, R 2=0.99 and Q 2=0.56). However, in that case the predictions for an external set of compounds (n test=3) were possible with an acceptable level of error. In addition, the authors proposed nine novel structures indicating possible inhibitor activity based on the model obtained. They concluded that steric effects play the most important role in the inhibition mechanism as well as electrostatic and H-donor/acceptor properties. However, the last two types of interactions are of lower importance. Similarly, SMILES-based optimal descriptors have been successfully applied for modeling HIV-1 PR fullerene-based inhibitors [131]. The model reported by Toropov et al. [131] was described by the following equation and parameters:

$$\begin{array}{l} pEC50 = - 31.6 + 0.125\;{DCW} \\ {n} = 8\;{\rm{R}}^2 = 0.90\;{Q}^2 = 0.85\;{s} = 0.35\;{F} = 58\;({\rm{subtraining}}\;{\rm{set}}) \\ {n} = 7\;{\rm{R}}^2 = 0.52\;{Rm}^2 = 0.13\;{s} = 1.27\;{F} = 5\;({\rm{calibration}}\;{\rm{set}}) \\ {n} = 5\;{\rm{R}}^2 = 0.99\;{Rm}^2 = 0.96\;{s} = 0.18\;{F} = 367\;({\rm{test}}\;{\rm{set}}) \\ \end{array}$$
((14-10))

Rasulev et al. [132] developed a QSAR model for the cytotoxicity to the bacterium E. coli of nano-sized metal oxides. They successfully predicted the toxicity of seven compounds (namely, SnO2, CuO, La2O3, Al2O3, Bi2O3, SiO2, and V2O3) from the model trained on the other seven oxides (ZnO, TiO2, Fe2O3, Y2O3, ZrO2, In2O3, and Sb2O3). The model employing the SMILES-based descriptor DCW is given by Eq. (14-11):

$$\begin{array}{l} - pLD50 = 1.32(\pm 0.031) + 0.27(\pm 0.0080)\;{DCW} \\ n = 7,\;R^2 = 0.99,\;s = 0.053,\;F = 539; \\ n_{{\rm{test}}} = 7,\;R_{{\rm{test}}}^2 = 0.82,\;s_{{\rm{test}}} = 0.241,\;F = 23 \\ \end{array}$$
((14-11))

The DCW descriptor in this case is defined as the following (Eq. 14-12):

$${DCW} = \sum\limits_{i = 1}^N {CW}(SA_k ) $$
((14-12))

where the SA k is a SMILES attribute, i.e., one symbol (e.g., “O,” “=,” “V”) or two symbols (e.g., “Al,” “Bi,” “Cu”) in the SMILES notation. Numbers of double bonds have been used as global SMILES attributes. They are denoted as “=001” and “=002.” “=001” is the indicator of one double bond and “=002” is the indicator of two double bonds.

Although we strongly believe in the usefulness and appropriateness of QSAR methodology for nanomaterial studies, the number of available models related to activity and toxicity is still very limited. When analyzing the situation, it seems that the main limitation is insufficient amount of existing experimental data. In many cases, lack of data precludes an appropriate implementation of statistical methods, including necessary external validation of the model. The problem of the paucity of data will be solved only when a strict collaboration between the experimentalists and QSAR modelers is established. The role of the modelers in such studies should not be restricted only to rationalization of the data after completing the experimental part, but also they must be involved in the planning of the experimentation. Since the experiments on nanomaterials are usually expensive, a kind of compromise between the highest possible number of compounds for testing and the lowest number of compounds necessary for developing a reliable QSAR model should be reached. Regarding the limited amount of data and high costs of the experiments, the idea of applying novel read-across techniques enabling preliminary estimation of data (Chapter 7) [82, 133] is very promising. However, no one has yet tried to implement this technique to nanomaterials.

5 Summary

Without doubt, a large and increasing aspect of the near future of chemistry and technology will be related to the development of nanomaterials. On one hand, due to their extraordinary properties, nanomaterials are becoming a chance for medicine and industry. But, on the other hand, the same properties might result in new pathways and mechanisms of toxic action. In effect, the work with nanomaterials is challenging for both “types” of chemists: those who are searching for and synthesizing new chemicals and those who are working on risk assessment and protection of humans from the effects of these chemicals.

When analyzing the current status of nano-QSAR, the four noteworthy suggestions for further work can be made:

  1. 1.

    There is a strong need to supplement the existing set of molecular descriptors by novel “nanodescriptors” that can represent size-dependent properties of nanomaterials.

  2. 2.

    A stronger than usual collaboration between the experimentalists and nano-QSAR modelers seems to be crucial. On one hand, it is necessary to produce data of higher usefulness for QSAR modelers (more compounds, more systematic experimental studies within groups of structural similarity, etc.). On the other hand, a proper characterization of the nanomaterials structure is not possible only at the theoretical (computational) level. In such situation, experiment-based structural descriptors for nano-QSAR might be required.

  3. 3.

    It is possible that the current criteria of the models’quality (the five OECD rules) will have to be re-evaluated and adapted to nanomaterials. This is due to the specific properties of chemicals occurring at the “nano” level (i.e., electronic properties change with changing size) and the very limited number of data (problems with the “classic” method of validation which is biased to small, low molecular weight molecules).

  4. 4.

    Greater effort is required in the areas of grouping nanomaterials and nano-read-across. This technique might be useful especially at the initial stage of nano-QSAR studies, when the experimental data are scarce.

In summary, the development of reliable nano-QSAR is a serious challenge that offers an exciting new direction for QSAR modelers. This task will have to be completed before the massive production of nanomaterials in order to prevent potentially hazardous molecules from being released into the environment. In the long term, prevention is always more efficient and cheaper than clean-up.