Overcoming roadblocks in computational roadmaps to the future for safe nanotechnology

Machine learning methods offer great promise for modelling and prediction of beneficial and adverse properties of nanomaterials. This is because they are fast, do not rely on underlying mechanistic models, use open source accessible technologies, and have proven effective in modelling patterns in complex data in many other domains [13]. Being data driven methods, they are critically dependent on the information used to generate them. Data quantity, quality, diversity, and range are all important for generating robust and predictive ML models with broad domains of applicability.

3.1.1. Data quantity

3.1.2. Data quality

3.1.3. Data diversity

The first step in addressing nano-safety issues is characterizing nanomaterials in relevant biological environments (describing the so-called biologically relevant entity) so that subsequent examinations can be linked to specific features of nanostructures. Nanoparticles can agglomerate, dissolve, have photocatalytic properties, and form persistent or dynamic complexes with biological molecules or ions. These properties can substantially affect the biological behaviour of nanomaterials [70–75]. While some progress has been made with regard to when, where, and how to characterize nanomaterials, this remains to be a challenge because the novel properties of nanomaterials that differ from those of the bulk materials may change significantly under different conditions and/or over time. These dynamic properties of nanomaterials are highly sensitive to the surrounding environment and the properties may interact (i.e. are interrelated), making the characterization process even more complicated (e.g. the zeta potential of nanoparticles varies greatly depending on whether they are pristine or in different types of biological fluids).

Ideally, the physicochemical characterization of nanomaterials should be performed in relevant biological matrices because pristine nanomaterials become modified upon introduction to biological fluids [76]. The surfaces of nanomaterials are immediately covered by biological components, resulting in the formation of bio-corona that modifies materials characteristics such as particle size, surface charge and state of aggregation [77]. The corona varies with environment and over time because more abundant macromolecules that bind first are gradually replaced by less abundant but more tightly bound macromolecules. More research is needed on how the physicochemical properties of nanomaterials (e.g. size, surface chemistry, dissolution, aggregation, etc) change as a function of biological conditions and whether these transformations make nanomaterials more or less toxic than their pristine form. This will provide more detailed information on biological effects of nanomaterials that can be used by regulators to more accurately assess the risks that they pose to the human or environmental health.

Fortunately, significant progress is being made in characterizing the nanoparticle corona as a function of nanomaterials physicochemical properties and the environment in which they are found. A number of studies in the past year have reported a range of spectroscopic, mass spectrometric, and chromatographic methods for characterizing nanoparticle coronas, some of which can be used to study the dynamic properties of biocoronas (figure 4) [78–87]. These data have enabled development of explanatory and potentially predictive models of nanoparticle corona composition [88–93] and dynamic behaviour that form the basis for rational design and development of nanomaterials that attract little or no corona [94].

Zoom In Zoom Out Reset image size

Solubility and biopersistence of nanomaterials in biological fluids are other key considerations that need to be addressed in nanosafety evaluation, since they directly influence internal exposure to a nanomaterial. The biological activity of soluble nanomaterials may be largely due to the dissolved material rather than the particle per se, e.g. ZnO. The rate of dissolution is affected by nanoparticle physical form (e.g. particle size, shape, coating, core doping) potentially providing methods for rational control of dissolution to minimize adverse biological effects [95]. Despite its critical importance, it is still unclear which nanomaterials characteristics affect solubility and particularly, biopersistence [96, 97]. More research is also required to understand how nanomaterials dissolution in biological fluids or persistence in the body influence biodistribution and biological impacts, and how this can be modified by formulation, coatings, or other methods.

Molecular descriptors, mathematical entities that encode relevant structural and physicochemical properties of nanomaterials, can be generated by theoretical methods or by standardized experiments. They are one of the most important elements in computational modelling studies in medicinal/environmental chemistry, toxicology, pharmacology, genomics and drug design [98–100]. Computed theoretical descriptors that capture important structural properties of nanomaterials provide diverse sources of chemical properties and a broad coverage of the vast chemical property space describing all possible nanomaterials. Experimentally derived descriptors (e.g. size shape, solubility, agglomeration, etc) are of very limited use in ML models of biological effects of nanomaterials, not only because of their time and resource intensiveness (even using high throughput methods), but also because they are not available for designed or otherwise hypothetical materials not yet synthesized [101].

A large number of theoretically well-founded and chemically interpretable descriptors can now be directly calculated from molecular structure using various commercial and open-source software packages [102]. However, almost all existing molecular descriptors and fingerprints currently used are not nano-specific, rather have been 'borrowed' from the pharmaceutical domain for organic small molecules. Some of these descriptors are not directly applicable to nanomaterials due to the polydispersity of nanoparticle physical properties, the size and shape of nanomaterials, incomplete characterization of their compositions, and their complex interactions with each other and with biological macromolecules and ions. In other words, the interfacial properties of nanomaterials are highly affected by external conditions, and hence, cannot be simply reflected by their composition or chemical structure.

The specificity of nanoscale properties and the lack of nano-specific molecular descriptors that can capture them constitutes a long-recognized but still only-partially addressed roadblock. Clearly there is a strong need for better descriptors for nanomaterials, providing an important challenge for researchers, but progress in this area has been disappointingly slow. However, interesting developments in deep learning have shown that imaging and related methods coupled with deep learning algorithms may provide a paradigm shift in the way nano-specific descriptors are generated for training predictive models in the future. For example, Russo et al [103] used convolutional neural networks (CNNs) to encode complex nanostructures into data structures suitable for machine learning modelling. Nanostructures were represented by virtual molecular projections, a multidimensional digitalization of nanostructures, and used as input data to train predictive CNN models of nanoparticle biological properties. The CNN neurons recognized distinct nanostructure features critical to activities and physicochemical properties (figure 5). Russo et al proposed that this 'end-to-end' deep learning approach effectively digitizes complex nanostructures for data-driven ML models and that it can be applied to rationally design nanoparticles with desired activities.

Zoom In Zoom Out Reset image size

Apart from their efficacy in ML models, another important property of nanomaterials descriptors (and descriptors more generally) is interpretability. That is, how easily can the most important descriptors controlling biological responses, identified by ML models, be used to rationally design improved materials and to understand the mechanisms by which they elicit their biological properties. Interpretable descriptors can be related easily to chemically recognizable features in nanomaterials. This is an important, ongoing issue for ML modelling of physicochemical and biological properties of molecules and materials generally [100]. Traditional molecular descriptors currently used in ML models are often arcane, and new descriptors are needed can be easily interpreted in terms of their biological (e.g. toxicity), physical (e.g. photocatalysis, solubility) and interaction (e.g. agglomeration, corona structure) properties. To this end, research by De et al and Sizochenko and coworkers has described new methods for encoding the properties of nanomaterials that are both efficient and interpretable [104, 105]. De et al used a set of 23 descriptors obtained from the periodic table or derived from them. These included the molecular weight of a metal oxide, number of oxygen atoms, metal electronegativity, metal oxidation state, number of valence electrons and atomic number. These were found to be as effective as descriptors generated by time consuming quantum chemical methods. Sizochenko et al used a 'liquid drop' model for nanoparticles in which a nanoparticle is represented as a spherical drop in which atoms or molecules are densely packed, and the density of cluster is equal to mass density (figure 6). It distinguishes the nanoparticles' surface molecules from the other molecules in volume. They also employed a simplex representation of molecular structure to encode the first level of organization of nanoparticles. Here, any molecule can be represented as a system of different simplexes (fragments of fixed composition and topology) that differentiate atoms in a simplex by their type and other properties such as electronegativity, lipophilicity, van der Waals interactions, etc.

Zoom In Zoom Out Reset image size

Text string representations of molecular entities, such as SMILES and InChI, are being used increasingly to generate descriptors for ML models, often in conjunction with deep learning algorithms that generate latent features from them. InChI keys are textual identifiers for chemical substances that provide a standard way to encode molecular information. The recent report of the first definition of NInChI (an InChI representation for diverse types of nanomaterials) suggests their potential for use as unique identifiers for database searches. As they encode essentially all relevant properties of complex nanomaterials in a hierarchical and compact text format, they should allow generation of new types of nanospecific descriptors that incorporate the complex structures, different nanoforms of the same material, diverse types of nanomaterials, and provenance information [28].

New 'surrogate' methods of generating information useful for modelling properties of nanomaterials are showing considerable promise. Spectroscopic, mass spectrometric, proteomic, and genomic 'fingerprints' of nanomaterials, and/or the effects they have on biological systems, can serve as useful descriptors when coupled with efficient sparse feature selection methods that identify a small pool of most relevant features. These have proven very useful in modelling the biological properties of other types of materials. For example, ion mass profiles generated by Time-of-Flight Secondary Ion Mass Spectrometry (ToF-SIMS) are very useful surface chemistry descriptors for ML models [106]. Given the importance of surface chemistry to nanomaterials, it is very likely such surface analysis fingerprints will prove equally useful for generating robust ML models of nanomaterials properties.

Quantitative fingerprint information can also be extracted from microscopy images of nanomaterials, which then can be used to train predictive models. Although the idea of image-based nanodescriptors has been around for more than a decade [107], it still is not routinely used in molecular descriptor calculation tools, and remains at a more conceptual rather than practical level, with a few exceptions such as recently-released NanoXtract image analysis tool [108]. This uses electron microscopy images to generate 18 image nanodescriptors such as area, perimeter length, roughness, aspect ratio, and concavity/convexity (figure 7).

Zoom In Zoom Out Reset image size

Another recent and novel approach is the use of information entropy to characterize distributions of nanoparticles [109]. The development of more sophisticated image analysis approaches and the application of image-based nanodescriptors is expected to evolve in the near future with the adaptation of existing tools to nanospecific needs. The predictive power of ML models can be boosted by incorporation of experimental fingerprints from surrogate analyses. However, in common with other experiment-based descriptors, these approaches are not useful for predicting the biological properties of nanomaterials not yet synthesized.

Predictive models of biological effects of nanomaterials can also provide insight into how nanomaterials generate their biological effects, but only if the descriptors that constitute the model are physically or chemically interpretable. Therefore, special attention must be given to the interpretability as well as efficacy of nano-descriptors so that they can be used to develop models that are both predictive and explanatory.

Clearly, ML models are necessarily trained using limited datasets that cover a restricted range of physicochemical and biological properties. The relative dearth of large and diverse training datasets is one of the main limitations to the wide use of ML models of nanomaterials biological properties. Computational models are most useful for predicting the properties of nanomaterials that exist in, or near, their domains of applicability, the multidimensional space defined by the descriptors and biological responses of data used to train the models. When using these models for interpolation (e.g. to fill in gaps of studies using partial factorial design of experiments) or extrapolation close to the model domain, the predictions are generally reliable. The further predictions are made from the domain of applicability of the model, the less reliable the predicted properties will be. New modelling algorithms such as Nearest neighbour Gaussian Processes [110], and existing ML methods that exploit Bayesian statistics [111, 112], can provide estimates of the reliability of the predictions. As larger, more diverse data sets become available, and ML methods better able to estimate the prediction uncertainties are adopted, 'out of domain' issues should diminish.

The problem of predicting relevant in vivo responses to nanomaterials are likely to persist for some time. Generation of experimental in vivo data is particularly restricted by ethical, time, and cost issues. By using efficient design of experiment methods, the number of in vivo experiments can be minimized, and ML models could be used to predict in vivo responses to other nanomaterials within their domains of applicability. However, realistically, the limitations of experimental in vivo testing will remain or may become even tighter, so we are unlikely to see large increases in the amount of available in vivo data. However, ongoing development of surrogate bioassays, such as gene expression profiles (toxicogenomics) or new phenotypic assays, should identify assays that correlate better with in vivo responses than current in vitro assays do. Rapid advances in intrinsically high throughput 'omics' methods, transcriptomics, proteomics and metabolomics can generate large, rich datasets on the responses of cells or organisms to nanomaterials exposure [113, 114]. ML modelling methods could exploit these data in two main ways: using these expression profiles as descriptors to better predict in vivo responses; and using computed nanodescriptors to model the relationships between physicochemical properties of nanomaterials and important omics signatures or pathways identified through expression profiling. For example, Furxhi et al used physicochemical descriptors for nanomaterials and experimental exposure conditions, cell line, cell type and tissue to model biological outcomes from genome wide studies using Bayesian networks [115]. Dysregulated genes were identified using fold change on exposure to nanomaterials, and Gene Ontology analysis functionally annotated the selected genes with the relevant biological effects. The model predicted the impact on 9 different biological pathways with a binary accuracy of 80%–100% (50% is random chance). The Ahmad group has also been active in using ML to model toxicological properties of nanomaterials using omics data. They recently reviewed the potential of combining integrative omics with machine learning to profile nanomaterials and model their biological impact for safety and risk assessment [116–118]. Metabolomics now allows study of metabolic dysregulation in biological cells, tissues, and living objects caused by exposure to nanomaterials. Untargeted metabolomics (metabolic fingerprinting) can be valuable for risk assessment of nanomaterials. Genomic and metabolic fingerprints, used in conjunction with sparse feature selection and feature importance methods, can be used to train predictive and interpretable ML models of the possible adverse biological properties of nanomaterials.

Hybrid ML modelling methods, where experimental in vitro assay results and nanodescriptors are used to model limited in vivo data and allow it to be leveraged more widely and reliably, show promise [119, 120]. For example, Lee and co-workers reported that molecular descriptors for drugs combined with in vitro basal cytotoxicity was more successful in predicting human acute toxicity than descriptors or in vitro results alone [121]. Recent developments in high throughput surrogate methods for assessing nanomaterials toxicity have been comprehensively reviewed by Collins and co-workers [59].

If the roadblocks detailed above can be removed or at least reduced, the nanotoxicity field will have available to it a relatively large number of substantial, diverse, and relevant data sets, and a suite of efficient nano-specific descriptors. There are a large number of efficient and well-validated ML algorithms available that can be trained by such data and descriptors to generate robust and useful models of important toxicological endpoints. Recent reviews have highlighted the role specific ML methods play in modelling nanotoxicity data [10, 122–126]. For example, Brown and co-workers wrote a review focusing on: using ML to analyse and extract new insights from large nanoscience data sets; applying ML to accelerate material discovery, particularly the use of active learning to guide experimental design; and the science behind memristive devices that can be used for hardware implementation of ML [127]. Ahmad et al have recently reviewed the application of deep learning ML methods to modelling of nanomaterials properties relevant to nanosafety (figure 8) [117].

Zoom In Zoom Out Reset image size

There is still a paucity of scientists with multidisciplinary domain knowledge required to apply new technological developments to (computational) nanotoxicology. Relatively few researchers are being rigorously trained in skills needed to develop robust models of possible adverse effects of nanomaterials and probe mechanistic aspects of nanomaterials biology interactions. Large multidisciplinary teams are the dominant paradigm in many areas of science, and nanosafety is no exception. The new H2020 initiatives referred to above, building on EU COST Actions before them, are helping build this critical mass—similar initiatives exist in the US [128] and Asia e.g. Asia Nanos Forum (www.asia-anf.org/working-groups/nano-safety-risk-management/).

With advances in the quantitative detection of small particles in complex samples, it has become clear that a wide range of processes and tools release nano-sized particles into the environment. The ability of some nanoparticles to pass through biological membranes opens up new possibilities in diagnostics and therapies but poses serious questions about human health risks associated with nanomaterial exposure. Experimental nanotoxicology, and its complementary computational equivalent, are making important contributions to understanding interactions of nanomaterials with biological systems, predicting possible adverse effects, developing safe-by-design methods for industry, and providing data and models to support regulatory decisions.

Computational nanotoxicology supports human health risk assessment of nanomaterials by integrating, unifying, and completing empirical evidence. However, the chemical and physical complexity of nanomaterials will continue to make modelling and prediction of relevant biological and mechanisms impacts a mathematical and computational challenge. Subtle differences in nanomaterials properties may dramatically affect in vitro and in vivo responses. When coupled with the dynamic behaviour of the biologically relevant form of specific nanomaterials in different media, considerably more effort is needed to model nanomaterials properties relative to a bulk material or small organic molecules with precisely defined structures and properties.

In summary, to overcome these challenges, future research should focus on:

• development of improved computable, nanospecific, interpretable descriptors that represent unique and biologically relevant features of nanomaterials,
• collection of post-transformational characterization data measured under conditions that mimic all relevant exposure scenarios,
• development of models that can robustly predict the composition of the corona as a function of biological medium and time,
• generation of large, information rich surrogate data sets and nanomaterials fingerprints based on imaging, surface analysis methods, proteomics, transcriptomics and related fields,
• adoption of relevant sparse feature selection algorithms and deep learning methods that can generate and identify relevant nanomaterials features from a large pool of possibilities,
• generation of larger sets of reliable in vitro and in vivo toxicity data for model development and calibration; further development of hybrid in vivo models based on molecular, physicochemical, and high throughput in vitro data,
• creation of robust, searchable and self-updating databases that facilitate sharing and comparison of data on nanotoxicology,
• training of researchers with both domain expertise and data analytics background, and establishment of tight networks of computational nanotoxicology researchers.

Predictions are difficult, especially about the future, and especially in light of the failure to achieve the milestones in the roadmap developed almost 10 years ago. With these caveats, a roadmap capable of addressing the above nanomaterials research challenges should be achievable within a decade. Impressive developments in ML methods and applications, especially deep learning, will overcome some of the issues identified in the review. On the one hand, these methods require a lot of data. On the other hand, they may simplify the development of nanospecific descriptors and may be amenable to using raw data in new way. Finally, publication of negative results from toxicological studies is critically important for the generation of cumulative knowledge in nanoscience, because no-adverse-effect is a positive result in nanosafety, and models need to understand what nanoparticle features produce adverse effects and which do not. In this way the dream of rational 'safe-by-design' nanomaterials will become increasingly realized. It is interesting to speculate how quickly these new, exciting potential solutions to computational nanotoxicology roadblocks will be adopted, and whether the next decade will indeed see a golden age of safe use of nanotechnologies that will enrich the lives of us all.