Keywords

2.1 Introduction

Risk assessment is a scientific exercise that aims at anticipating hazards. By convention, it entails a specification of this hazard, the collection of data about known occurrences and various calculations allowing to extrapolate the frequency, severity, probability of future occurrences for particular persons or organisations from a baseline of data.

Risk assessment has turned out to be constitutive of a type of regulation, known as risk regulation [1, 2], [3]. It took form progressively in the 1970s and 1980s thanks to the contributions of a series of more established disciplines such as actuarial sciences, geography and natural disaster research, physics, operational research, all of which converged towards the notion that the probability of hazardous events could be computed, and they may be prevented thanks to these calculations. Prediction has been one of the rallying calls for the scientists that united to give birth to the interdisciplinary movement of risk assessment [1, 4, 5].

Several decades later, the broad movement of digitalisation and the promises of artificial intelligence seem to be pushing the limits of risk assessment and herald an era of faster and more precise predictions. The conversion of more diverse and larger sets of information into storable, classifiable and analysable digital form, and the design and adoption of IT technologies allowing organisations to perform such tasks at a quick pace and minimal cost revive the ambitions of risk assessors to anticipate risks with precision and reliability. And indeed, most risk assessment practitioners have joined the call to accelerate and deepen the movement of digitalisation, embracing like many other sciences the age of big data [6]. The imaginary of continuous, non-human-mediated production of data to train and feed predictive machines [7], and quickly discover new cause-and-effect relationships in complex systems, has penetrated risk assessment and risk analysis [8].

Digitalisation, however, is a complex of several intertwined transformations. It entails a phenomenon of datafication (the generation of data about an increasing diversity of organisational activity), of computational innovation (the introduction of new computing technologies and infrastructures) and of theoretical modelling of the world (with the rise, notably, of a systems-vision). This chapter briefly reviews the history of chemical risk assessment methods developed by regulatory bodies and associated research groups, and the complex ways in which this would-be science has digitalised over the years. It does so to identify what is currently happening in this area and to better determine whether the ever-revamped technological promise of prediction is within closer reach than it was before. Clearly, digitalisation runs through the history of risk assessment. But the computing technologies available, the data generated and the theoretical visions thanks to which we can make sense of these data and turn them into meaningful predictions, are perhaps more aligned nowadays than they used to be, producing this sense of a fast and deep transformation of technologies for knowing what is safe.

2.2 Assessing and Computing Risks

From aircraft to chemicals plants, through food ingredients and chemicals, most of the technologies that are recognised as potentially hazardous are submitted to some form of risk assessment. Risk assessment is the informational element of risk regulation regimes [2]. It involves dedicated techniques and routine processes, through which the conditions of appearance of hazards may be determined (their frequency, severity, publics and places most affected…), and corresponding regulatory controls legitimately decided.

It can be applied ex ante to the development of the technology in question, informing the decisions to put it on the market or in use more generally. It may also take place alongside the use of the technology in question (it is then called monitoring, surveillance or vigilance). However, the epistemic and regulatory ideal behind risk assessment is that of the prediction of hazards before they occur (Shapiro and Glicksman 2003).

Risk assessment is a process that has always made massive use of science and particularly of modelling. Chemicals cannot or must not be tested directly in the human body, in the environment or in the industrial conditions in which they will be used, but are on the contrary tested in experimental, simplified conditions. The toxicity of the chemical is tested on animals (and toxicologists and chemists speak, tellingly, of the various “animal models” (i.e., species) that can be used to perform these experiments) or in vitro. In this sense, risk assessment is a model-based science, if we understand models as a simplified, scale-reduced analogue, or representation, of a system.

Modelling is a process that consists in formulating a series of equations to capture the functioning of a system, and informing the parameters of the equations with various measurements and data, in such a way that various states of the system may be simulated. Modelling allows extrapolating from the data points available (of which there may be a relative paucity), to other situations, scales and time periods. It is closely dependent on current knowledge of a system and on the capacity to imagine risks and accidents occurring within that system [9]. Risk assessment follows a systems perspective. It simplifies systems and the behaviour of agents in this system. In practice, it follows from the social construction of a “risk object” [10]: a technological element excerpted from this system, to which a number of potential hazardous effects may be attributed.

2.3 Layers of Transformation: A Historical Perspective on Digitalisation

The field of risk assessment broadly evolves towards an ideal of continuous modelling of large sets of data, to analyse and simulate processes at various scales of a system; a sort of integrated form of simulation, where one aims to describe and predict a greater number of aspects of a system, at a fine-grained level [11]. This broad ambition, however, concatenates several transformations that have affected risk assessment since it emerged four decades ago: the material capacity to generate data in greater quantities and great variety, thanks to the diffusion of sensors across the environment and living organisms, or datafication; the design and increasing use of new mathematical models to capture the complexity of systems and the occurrence of hazards within them; the rise of a complex-systems vision of things. Looking back at what has been developed in the field helps appreciate the path of technological development and epistemic change through which current applications have taken shape.

2.3.1 Mathematical Models: Technologies of Computing

The first computational tool used in this area was one that aimed to characterise the properties of molecules through systematic analysis of the relationships between their structure and their biological effects—so-called structure–activity relationships (SARs) [12]. A quantitative SAR is a statistical analysis of the biological activity of a group of two or more chemicals that have some structural similarity, as captured through a chosen descriptor of the chemical. The modelling of causal relations between chemical properties and biological impacts is rooted in fundamental chemistry. It rests on the conduct of multiple, strictly standardised experiments on molecules with the same kind of structure (cogeneric molecules). Once a sufficiently powerful set of data has been produced, a statistical analysis can be run, to capture the correlations between structural properties and the biological effects. The resulting correlations can then be used to formulate a mathematical equation—a model—to predict the effects of a molecule without physically testing it. The challenges that QSAR research is facing typically concern the generation of sufficiently large sets of comparable data across a whole class of chemicals (a highly intensive endeavour), and the availability of both training sets and alternative data sets to validate the models. Without such data, modellers end up producing an over-fitted or under-fitted model [13]. Connecting model development to larger sets of data made available by pharmaceutical companies is one of the key hopes here.

A second technique aimed at modelling dose–response relationships in biological organisms. The technique is known as physiologically based pharmacokinetics (PBPK). PBPK models consist in simplified descriptions of the physiological system exposed to a chemical substance. By modelling the organism and the biological mechanisms involved in the metabolism of the substance, one can compute the dose at which the substance will produce hazardous effects in the organism. Models represent relevant organs or tissues as compartments, linked by various flows (notably blood flows) in mathematical terms. The parameters are calibrated with data emerging from animal experiments or clinical observations. PBPK modelling really started in the 1970s, once sufficient data and computer tools became available to establish the doses at which anticancer medicines could be delivered to various organs. The application of PBPK to industrial chemicals started at the beginning of the 1980s, to define so-called reference doses for chemicals: the levels of concentration at which they can safely be considered to not cause harm. This could be done because of the accumulation of data about volatile chemicals (then under threat of regulatory restrictions): data about people’s inhalation of chemicals, data about biological metabolisation of these chemicals and data about the quantity of chemicals eliminated by the human body and exhaled. These data originated, notably, from the use of costly inhalation chambers. Once databases were elaborated, models started to be elaborated and calibrated in more reliable ways, for more chemicals, allowing to envisage the possibility to model together the chemical and the human body. In this field, the main challenge has always been the capacity to calibrate the model with realistic and varied biological data, to counterbalance the drive to make predictions based on more quickly produced, but less representative average values.

A third technique consists in developing what is called biologically based mechanistic models, to analyse the functioning of the human body and biological pathways inside those, as well as their interactions with substances. The resulting “biologically based dose response” (BBDR) models pursue the same kind of aim as PBPK—doing better than animal tests in terms of prediction of risk thresholds. Indeed, some of its champions are the same as for PBPK [14], and BBDR was also developed to counter or moderate regulatory drives on critical chemicals such as dioxin [1]. Instead of capturing biology through equations, as PBPK does, it banks on rapidly evolving knowledge of the cellular pathways through which chemical substances trigger potential toxicological issues. These theoretical models of biological organisms are supposed to guide the interpretation of empirical toxicological data. Much like PBPK, the reliability of this sort of modelling is limited by the data that are being used, and their capacity to represent “inter- and intraindividual heterogeneity” [15].

2.3.2 Datafication

All of the above techniques, as briefly mentioned, have been limited by the slow and costly generation of data through in vivo or in vitro tests, as well as by the quality of the hypotheses that guide their interpretation. In terms of toxicity data, the game-changer has come from the genomics (and the corresponding toxicogenomics) revolution, namely from tools that can generate massive sets of data points about genetic events from a single experiment, and at high speed. “Omic” techniques, such as microarrays, make it possible to represent all the events in a biological system associated with the presence of a chemical substance. Robots allow multiple assays to be run on dozens or hundreds of substances day after day, generating massive sets of data, to be modelled by biologists. This toxicogenomic effort emerged a little after 2000s, after the three others introduced above.

Under the impetus of the chief of the US National Toxicology Program, Chris Portier (a biostatistician who had, among other things, worked in the area of PBPK and BBDR), a draft strategy was elaborated in 2003 “to move toxicology to a predominantly predictive science focused upon a broad inclusion of target-specific, mechanism-based, biological observations”. The Environmental Protection Agency embarked on a similar effort a few years later. These institutions soon developed together a vast effort known as Tox21, to conduct hundreds of assays on thousands of substances thanks to high-throughput technologies. The central character in this program is a robot from the Swiss company Stäubli that autonomously manipulates plates containing dozens of mini-petri-dishes, to conduct multiple assays on multiple chemicals at several dosages. The result is an immense set of data, in which biological patterns can hope to be detected. This is done, notably, through open data challenges: the Tox21 institutions have called for teams of computational biologists around the world to search through their data to generate such models. This is where machine learning enters the picture: models are being constructed from the ground up, through supervised exploration of the mass of data to identify (or learn) patterns [16].

2.3.3 Computational Risk Assessment: The Integrated Vision

At about the same time as the Tox21 effort took off, a panel of top toxicologists and specialists of the field of toxicity testing, led by Melvin Andersen, rationalised computational toxicology.

The addition of high-throughput toxicogenomic to previous developments allowed to envision a future in which data would be available for many possibly toxicity pathways concerning multiple substances, to radically change how the toxicity of chemical substances would be tested: not as an isolated object with defined properties, but as elements of a biological system acting at low doses through diverse biological pathways. In other words, a knowledge system that would be representative of the reality of how biological systems function in the current chemicalised environment. The risk assessment of chemicals, thus, has evolved in the same manner as supporting disciplines such as biology, towards a more computational, systems-based style of analysis [17, 18].

The resulting “vision” was published by a branch of the US National Academies (the National Research Council) and heralded as the right guiding vision [19]. Interestingly, the vision seems to cap all previous efforts in the area of model-based, predictive toxicology: efforts in QSAR (to characterise properties of a substance), PBPK and BBDR (to formulate biomathematical models of the organism) and in high-throughput in vitro testing were now the building blocks of a knowledge system allowing to “evaluate relevant perturbations in key toxicity pathways” [19, p. 7], as opposed to simply measure the levels at which an object, taken in isolation, may prove harmful.

2.4 Discussion

The current development of artificial intelligence rests on a discourse about the all-powerful machine learning methods, and their unabridged capacity to learn from data, thanks to powerful computers. Risk modellers often resort to short-cutting claims such as the one that they can predict risks thanks to better maths and bioinformatics. In holding that discourse, modellers in the area of chemical risk assessment illustrate the fact that digitalisation colonises risk assessment of chemicals, just as it has colonised other areas of scientific practice.

Historians of science and technology have noted that the digital is a lingua franca in sciences; a form of generic technology that produces comparable epistemic effects across disciplines [20]. In the case that is documented here, one can see that computing technologies and theoretical, systems-based visions were both, in some ways, borrowed from neighbouring spaces. In the present case, one sees the application of deep learning late in the process, in the context of the Tox21 program. One also sees the importation of a robotic technology from industrial fields. To give one further example: PBPK modelling has developed and gained credibility thanks to the use of generic programming languages (e.g., Fortran), allowing more people to engage in this area, generate more models, creating an emulation/comparison of models, resulting in the improvement of the technique altogether.

This all too short historical overview has tried to specify, in contrast with this discourse, what are the area-specific conditions of a digital transformation. I have emphasised, first, that risk assessment has been, from the very start, a computational practice: a kind of science that asserted its scientificity through the development of gradually more complex modes of calculation of risks, moving towards the mathematisation and modelling of more and more aspects of the functioning of biological systems. There is certainly a degree of novelty in the current introduction of a variety of machine learning methods, but risk assessment has always used some means of computation, and the artificial intelligence methods that are being experimented now have a certain degree of continuity with previously used methods.

Second, it appears that the application of sophisticated means of computation and machine learning algorithms may not fulfil the promise of prediction, if it is not matched by equivalent investments in datafication. Computational modelling, indeed, does not mean doing without data, and without the various means available—including experimental ones—to generate, collect, curate and classify them. What one learns from the history above is that the generation of data is a necessary condition for moving towards more digital risk assessment. In fact, as can be gathered from the brief descriptions above, the various families of modelling techniques have been restricted by the same problem: the availability, diversity and representativeness of the data that are being modelled.

A simple conclusion to draw from this is that artificial intelligence will represent an innovation and a new leap in modelling capacities, in so far as it is matched by the parallel deployment of larger infrastructures of data allowing to document the various elements of these complex systems, rather than an isolated risk object and its effects. Failing the full datafication of the systems that scientists want to model, prediction will stay focused on these particular objects, as they have always been. As can be gathered from the brief description above, various risk objects are construed by computational systems over time. QSAR looks at the properties of molecules and models classes of chemicals. PBPK looks at the dose of chemicals in the human body and models physiological systems. In Tox21, it is the biological pathway that is the object of knowledge. These are heterogeneous objects, and the systems that are in place to know these objects are distinct, and not necessarily compatible. They may be, quite simply, the incarnation of different ways of modelling or predicting [21].

In the case of Tox21, even though a holistic vision has emerged, eventually, there is no assurance that these knowledge systems can be further integrated, or that the current development of artificial intelligence will bring coherence to past developments. It is so because there is ontological politics involved: a search, which may be contentious, for a realistic definition of what the problem is. A risk can be defined in reductive ways, assigned to an object that is deemed easier to regulate and control (i.e., the molecule). Or a risk can be defined in a more diffused, systemic manner and lead to the exploration of chains of causation between objects forming a complex system. The more one evolves towards modelling complex systems, the more complex it becomes to intervene in and regulate these systems, since modelling will reveal complex chains of causation and an intertwinement of causes. In the present historical case, this is illustrated by the fact the ontology of the “dose”, “threshold” and of the risky object—the chemical substance to which a risk can be attributed—loses ground. This raises the issue of how decision criteria are forged in the space of knowledge systems that are designed to turn out complex correlations, rather than to isolate linear causation chains between an agent and an effect.

Overall, then, the ideal of digitalisation and the epistemic ambition to predict what is happening in systems may be capped by the establishment of data systems. A gap remains between the imaginary of digitalised prediction and the actual breadth of data systems. The various levels at which digitalisation unfolds—datafication, computational turn, theoretical visions of what may be modelled—reinforce one another, but they are not necessarily accorded in practice. For instance, with digitalisation and the big data revolution comes the “end-of-theory” claim: the notion that data-driven sciences will be fully empirical, learning from the bottom-up, by the mere, intensive exploration of data, to recognise patterns in complex systems, without the assistance of a priori theory about how these models are constituted and work.

Again, the history outlined above shows that this is unlikely to happen, as there is no pure and atheoretical exploration of data: data are generated by infrastructures that enact a certain theoretical vision of the world, namely in this case, a vision in terms of chemical substances and the measurement of doses of chemicals in bodies and environments. Generating data according to different theories is a process that will take ample time. Epistemologically, it would be legitimate to think that this is simply not possible: data and metrics necessarily enact a certain theoretical vision of what needs to be counted, of all that is happening out there in the world.