Abstract
The burgeoning literature on uncertainty analysis shows the need for accessible and transparent information about the limitations of knowledge associated with predictive models for environmental decision-making. Using qualitative analysis, we examine how experts involved in the development of genomic selection (GS) for Canadian public forestry conifer breeding assess and communicate uncertainty. GS is a bio-digital technology characterized by big data compilation, sophisticated statistical analysis, and high-throughput genome sequencing. While GS applications in forestry have the potential to increase yields, reduce errors, and improve the selection of resilient trees in the face of climate change, our data revealed barriers that impede more comprehensive discussions about uncertainty, including assumptions that uncertainty can (and should) be eliminated through the availability of more data, tacit commitments to the application of GS in commercial forestry operations, deterministic assumptions about linear gene-to-trait outcomes, and difficulties discussing uncertainty in collective settings. Uncertainty talk is uncomfortable as it can be perceived as a threat to applied research goals, but uncertainty talk is also a necessary, productive, and generative way to encourage transdisciplinary and inclusive discussions at early stages of predictive model deployment for environmental applications.
Introduction
In today's complex, turbulent, interconnected, globalised world, uncertainty must be embraced as perhaps more central than ever. Scoones and Stirling (2020)
The importance of assessing and communicating uncertainties in environmental and natural resource decision-making is widely acknowledged in academia and beyond. Numerous crises have been blamed on failures to acknowledge the complexity of socio-ecological systems and their associated uncertainties in the collective understanding of how such systems operate (Berkes, 2007; Power, 2007). Early and ongoing assessments of uncertainty have many advantages, such as promoting anticipatory preventative approaches to environmental challenges, offering antidotes to technological hype, and enhancing discussions about potential alternatives.
Signaling the importance of uncertainty, several frameworks, mostly developed from quantitative perspectives, are available in the environmental management, Earth system modeling, and geography literature (Brown, 2004; Pastor et al., 2020; Quigley et al., 2019; Spiegelhalter and Riesch, 2011). Calls for critically oriented, qualitative approaches to uncertainty analysis emerge from the recognition that quantitative approaches can all-too-easily imply that environmental problems can be solved through the acquisition of more data and the development of a more sophisticated analysis (Stirling, 2018). Increasing attention to qualitative, critical approaches to uncertainty in geography and cognate disciplines interrogate and situate uncertainties in relation to power, place, cultures, and governance (Senanayake and King, 2021; Scoones and Stirling, 2020; Wynne, 1992; Wynne, 2005).
In line with these efforts, this paper examines how experts involved in the creation of predictive genomic models make sense of uncertainty. Leveraging theoretical insights from science and technology studies (STS), we ask: How do experts involved in the development of genomic selection (GS) to improve conifer tree breeding for Canadian public forestry assess and communicate uncertainties? GS is a bio-digital technology that, along with the promise to increase accuracy, precision, and expediency, adds layers of complexity and uncertainty to environmental and natural resource decision-making. The term bio-digital signals technological developments in computational power, statistical analysis, and high-throughput genomic sequencing, in which the generation and interpretation of large amounts of data are central (van der Elst, 2019). Emerging tools in this domain, such as GS, have been met with increasing enthusiasm by governments, industry, and some scientists as a means of responding to environmental and economic challenges in forestry. Findings are derived from the authors’ involvement as social researchers in a large-scale Canadian applied genomics project, whose primary funding agency mandates the inclusion of ethical, environmental, economic, legal, and social dimensions of genomics. In previous work, we outlined the ways in which uncertainty assessments can not only foster, but also limit, public communication of scientific research findings (Blue et al., 2022; Blue and Davidson, 2021). Extending these insights, this paper explores how discussions of uncertainty take shape within interdisciplinary, applied research settings.
When exceedingly large, complex, and dynamic systems are modeled, as is the case with the application of GS for long-lived conifer trees in Canadian forestry, high levels of uncertainty are to be expected. Our findings indicate that those closest to and most knowledgeable of GS simultaneously acknowledged uncertainty, but also expressed considerable discomfort in discussing uncertainties that were not readily quantifiable. Notably, we found that in collective settings, engaging in discussions about uncertainty is generally difficult and, in some instances, emotionally fraught. Our findings support the conclusion that collective discussions about uncertainty in applied research settings are uncomfortable because such discussions reveal information that can be perceived as threatening to applied research goals (Rayner, 2012).
These findings contribute to qualitative uncertainty analysis by identifying underlying assumptions that can serve as barriers to the acknowledgement of uncertainty in an under-studied area. Attention to uncertainty in the academic literature related to environmental applications of GS is relatively low in comparison with other sectors in which GS has been applied, such as medicine (Biesecker et al., 2019; Cheon et al., 2014; Howard and Iwarsson, 2018; Han et al., 2017; Moses et al., 2018; Pyeritz, 2017) and forensic science (Georgiou et al., 2020). We view this relative lack of attention to uncertainty as both a concern—given the interest and increasing uptake of applications of GS for environmental purposes—and an opportunity to both assess uncertainty and experiment methodologically with promoting “uncertainty talk.” By uncertainty talk, we mean efforts to open-up discussions about the knowledge gaps associated with predictive models at the early stages of development and application. Uncertainty talk provides opportunities to reflect critically about the perceived purpose of technological interventions for environmental problems. Since uncertainty talk extends beyond topics and ways of thinking with which conventionally trained scientists are accustomed, resistance is likely to emerge with such efforts, and ongoing capacity building is needed.
From managing risk to acknowledging uncertainties
Lessons from past controversies over genomic technologies, particularly debates surrounding genetically modified foods in the 1990s, highlight the importance of engaging in the early and ongoing consideration of a broad range of uncertainties during research and development phases (Scoones and Stirling, 2020; van der Sluijs, 2012; Wickson, 2007; Wynne, 1992). Failure to consider a breadth of uncertainties can lead to overconfidence in perceived benefits, an underestimation of perceived harms, a limited capacity to learn from surprises, a reluctance to engage with diverse values and perspectives, and a diminished capacity to collectively imagine alternative futures (Scoones and Stirling, 2020).
Risk assessment is the dominant institutional approach for assessing uncertainties associated with biotechnologies. Risk assessment is typically a technical, expert-driven process, in which the probabilities and magnitude of potential outcomes are quantified. Critics point to significant flaws, with the assumptions underpinning conventional risk-based analysis, including the erroneous belief that risk assessment is a value-neutral process, that probabilities and magnitudes of potential harm are knowable in advance of technology implementation, and that risks exist “out there” in the world, independent of social contexts (Porter, 2006; Stirling, 2018; Wickson, 2007). Risk assessments assume that outcomes are clear, and probabilities can be assigned unproblematically to potential outcomes. Although appropriate in some instances, risk-based approaches are not well-suited to situations of high complexity, for instance, when technological applications are new, and impacts are not well-understood, or in the context of highly dynamic and variable environmental and human systems. In the face of complexity, the limitations of risk assessment to illuminate uncertainty become clear: risk focuses on “a restricted agenda of defined uncertainties” in ways that narrow attention to and subsequently close contemplation about deeper dimensions of incertitude (Wynne, 1992: 114).
In light of these limitations, a growing number of scholars call for the need to supplement risk analysis with uncertainty assessments (see Scoones and Stirling, 2020 for a comprehensive overview). 1 In contrast to the focus on risk, uncertainty assessment expands awareness of knowledge limitations and socio-ecological variability with the view to open discussion about alternative possibilities and futures. As mentioned in the Introduction, numerous quantitative frameworks for uncertainty assessment are available in the academic literature, and many take the view that uncertainty, like risk, exists apart from the people and institutions conducting the assessment. In GS applications, this tendency is apparent in fields, such as medicine and forensic science, in which predictive modeling is used to identify the risk of disease, indicate legal culpability, and assist researchers and professionals in reporting and interpreting results (Biesecker et al., 2019; Han et al., 2017; Weck, 2018). Uncertainty typologies, such as these, tend to focus on dimensions of incertitude related to a lack of knowledge or data, highlighting the need for experts to inform end-users (e.g. research participants, patients, and decision-makers) about the uncertainties that inhere in genomic tests.
Missing in these typologies is the attention to the perspectival and contextual dimensions of knowledge production, including the perspectives of experts themselves. STS approaches to uncertainty analysis fill this gap. For instance, in an early influential analysis, Don MacKenzie used the metaphor of a “credibility trough” to capture the differences in perceptions of uncertainty. He concluded that those directly involved in the production of scientific knowledge, as well as those alienated from institutions that produce scientific knowledge, tend to be the most aware of the uncertainty and the most skeptical of claims of certitude (Mackenzie, 1990). Conversely, those positioned as consumers or end users of scientific knowledge tend to perceive less uncertainty and are more trusting and credulous (Mackenzie, 1990). Furthering MacKenzie's claim that different actors have different relations to uncertainty based on their relation to knowledge production, Wynne argued that the expression of uncertainty among technical experts is directly related to commitments to success: when the stakes are high for a successful application of knowledge, uncertainties are often discounted as irrelevant or ignored entirely (Wynne, 2005: for similar arguments, see Shepherd and Kay 2012). Saltelli et al. identified further limitations with the ways in which technical experts discuss and acknowledge the uncertainties associated with predictive models, including a deflation of attention to uncertainties often for purposes of expediency, ritual use of models to convey an impression of predictability and control, and reliance on unverified assumptions (Saltelli et al., 2020).
The social construction of ignorance is another relevant dimension of uncertainty. Drawing on former US Secretary of Defense Donald Rumsfeld's discussion of ignorance as “unknown unknowns” (i.e. the things we “don't know we don't know” that can lead to unwelcome surprises), Rayner (2012) described a different type of ignorance: “known unknowns.” By “known unknowns,” Rayner meant types of knowledge strategically discounted, forgotten, or overlooked because of the potential threat the knowledge poses to institutional legitimacy. Known unknowns are “uncomfortable knowledge” that institutions manage in the following ways: deny the existence of the knowledge, denounce the knowledge as illegitimate, divert attention away from the knowledge by directing the focus elsewhere, or displace the knowledge by substituting it with a manageable surrogate. According to Rayner, the social construction of ignorance is inevitable, even necessary, because it allows institutions to function in the face of complexity. However, in some instances, uncomfortable knowledge must be acknowledged to avoid problematic outcomes. Addressing the challenge of uncomfortable knowledge requires a diverse range of values and perspectives in decision-making processes to ensure that all relevant information are accounted for.
Uncertainty talk: technical, environmental, and social dimensions
The purpose of uncertainty talk, as intended in this article, is to open discussions about uncertainties that inhere in predictive models in ways that account for perspectives and context. With uncertainty talk, we seek to capture the aspects of uncertainty that are hard or impossible to quantify and, as a result, often remain unaddressed in scientific settings. Existing approaches to uncertainty analysis with similar aims (Saltelli et al., 2008; van der Sluijs et al., 2005) require a rudimentary understanding of statistics and mathematics; since such knowledge is not widely shared, technically oriented uncertainty assessments can limit, rather than facilitate, communication and understanding. Similar to Stirling (2012), we use the term uncertainty in a general sense to mean incertitude (i.e. limits of knowledge in the face of variability and complexity). In relation to predictive models, we focus on three dimensions of uncertainty: technical, environmental, and social.
Technical uncertainty refers to uncertainties that arise in the process of turning inputs (e.g. data about genotypes and phenotypes) into outputs (e.g. GS models that rank individual trees for breeding potential). Technical uncertainty can refer to statistical uncertainties, such as gaps in knowledge associated with efforts to describe a population based on the samples drawn from that population, and model uncertainties, such as predictive statistical formulas trained on data sets. Statistical uncertainties are typically assessed and expressed quantitatively. Some technical uncertainties can be reduced through the application of more research (e.g. by gathering more data and refining model parameters). Other technical uncertainties are difficult to reduce. For instance, since model-based knowledge is conditional on assumptions, model uncertainty cannot be reduced by gathering more data. Models cannot be validated in the sense of being proven definitively true, but can be corroborated in the sense of meeting tests for internal consistency or demonstrating the capacity to explain or predict an event in a convincing and parsimonious fashion (Saltelli et al., 2008). Ironically, model uncertainty is seldom addressed by modelers themselves, although this situation is changing (Carslaw et al., 2018).
Environmental uncertainty refers to gaps in knowledge resulting from the variability in complex systems caused by actions at various stages within that system that can lead to unpredictable outcomes. Surprises in complex systems (“black swan events”), such as the 11 September 2001 attacks in New York City, the financial crash of 2008, and the 2021 Pacific Northwest heat dome, have confounded expert assumptions, predictions, and attempts at control. These events, among many others, have led to broader awareness of indeterminacy in disciplines such as ecology, earth systems, economics, and epidemiology (Taleb, 2007). Here, distinguishing between determinate and indeterminate systems is important. For determinate systems, the causal pathway between the initiating action and the eventual outcome can be anticipated with some degree of certainty. By contrast, indeterminate systems refer to instances in which causal pathways between action and outcome can only be known retrospectively, if at all. Although terms such as determinate and indeterminate imply a recognition of complexity, a critical difference lies with the potential for multiple, recursive pathways in indeterminate systems in which outcomes are dependent on myriad choices made at numerous points, where previous changes can alter future outcomes. Since each action can inform the direction of future actions, it is impossible to predict the outcomes of initial decisions in indeterminate systems (Wynne, 1992).
Social uncertainty refers to knowledge gaps and variability in problem framings, definitions, values, worldviews, and perspectives. One example is ambiguity arising from contending understandings, meanings, and values about the meanings of problems, solutions, and potential outcomes (Stirling, 2012). Ambiguity is inevitable in interdisciplinary and applied research settings that draw together different knowledge paradigms, methodologies, and assumptions. The social construction of ignorance is another example, including strategies to address uncomfortable knowledge (Rayner, 2012).
Uncertainties in GS applications for forestry
Background
The challenge for forestry, as with agriculture, is selecting and breeding organisms resilient to present and future environmental changes. This challenge is particularly pressing in long-lived conifer breeding as conventional approaches that rely on the selection of observable traits in mature “parent” trees are labor-intensive and slow and hence may prove inadequate in the face of rapidly developing climatic changes. The promise of GS is that, with the advent of next-generation sequencing technologies and bioinformatics, a higher genetic resolution can offer better insights for breeders than traditional genetic analysis methods. Over the past few decades, GS has been widely adopted in animal breeding programs and is increasingly being adopted for crop breeding programs (Budhlakoti et al., 2022).
GS is a type of marker-aided selection (MAS), where breeders select traits of interest based on the known genetic markers in an organism. Specifically, GS identifies dense genetic markers located across an organism's genome, the entire set of DNA (
Forest genomics research is taking place internationally, particularly in regions where forestry is a key sector, such as in Canada, the United States, Brazil, and Sweden (Isabel et al., 2020). Tree improvement, which includes the selection, breeding, and planting of trees for targeted traits, requires extensive investments of time, money, and expertise, often across diverse institutions that include multiple layers of government, universities, and industry. In Canada, federal and provincial governments, along with the forest industry, have made significant investments in tree improvement to address unprecedented challenges, such as climate change, and related extreme events involving drought, pests, and pathogens (Naidoo et al., 2019). One aim of Canada's tree improvement programs (i.e. research and development directed toward understanding and leveraging the genetic constitution of trees) is to select trees capable of withstanding harsh climates, including, in particular, the three provinces colloquially, but inaccurately, referred to as the Prairie provinces—Alberta, Saskatchewan, and Manitoba (these provinces actually have more forests than grassland prairies) (Carlisle, 1970). Some argue that genetic analysis is essential for the sustainable management of Canada's forests: “In order for Canada's tree species to survive predicted changes in climate it is crucial that tree breeders understand and select for genes associated with cold acclimation, drought tolerance and growth phenology in addition to pest resistance mechanisms” (Kumagai et al., 2010: 64). Yet, the unprecedented forest fires of 2023 serve as an example of the uncertainties and challenges faced by the forestry industry.
In the early 20th century, tree breeders did not have a well-developed understanding of the relationship between genetic structure and phenotypic traits (Barton et al., 2017). The advent of breeding formulas, such as the widely used breeder's equation first developed in the 1940s, distill complex technical information about phenotypes into parameters that breeders can manipulate to increase productivity across generations (Cobb et al., 2019). While early breeding selections in forestry were made based on observable (phenotypic) traits, the discovery of the double-helical architecture of DNA in the 1950s ushered in an era of molecular approaches to tree breeding. Ensuing decades witnessed increasingly sophisticated developments in genetic technologies that combined statistical insights with molecular biology techniques. For instance, beginning in the 1970s, researchers used the recombinant DNA (rDNA) technology, a process in which the DNA of one organism is introduced into another organism, to enhance the development of commercially desirable traits in trees, such as pest and frost resistance. Although supported by genetic scientists and governments, the use of the rDNA technology to enhance tree breeding was met with significant public disapproval and apprehension. Public concerns contributed to policy restrictions on the use of genetic modification for forestry on public lands in Canada that continue to this day.
Research into MAS began in the late 1980s and matured in the early 2000s, spurred by the development of the automated DNA sequencing technology and the discovery of single-nucleotide polymorphism (SNP) genetic markers (SNPs are genetic variations in alleles) (Nakaya and Isobe, 2012). MAS is often considered a more publicly acceptable form of biotechnology likely due to the lack of direct genetic manipulation. GS is essentially a scaled-up version of MAS that identifies multiple genetic markers across an organism's entire genome. Although the statistical methods and prediction models underpinning the application of GS for plant breeding were first introduced in the early 2000s (Meuwissen et al., 2001), the technological capacity to sequence and identify large numbers of markers across genomes was lacking at the time. This limitation was overcome by improvements in cost-effective, time-efficient, high-throughput sequencing technologies that enabled GS to move from concept to application in fields as diverse as medicine, agriculture, and forestry. The first empirical study on GS was applied on mice, with a continued increase in applications in structured breeding programs for dairy cattle, chickens, pigs, and a variety of agricultural crops (Nakaya and Isobe, 2012). In the early 2010s, research in the United States demonstrated that GS could be successfully used for commercially viable conifer trees (Neale, 2011).
Notably, at around the same time GS was developing in animal and plant breeding, genomic sequencing technologies were also being applied in medicine, as was the case with the high-profile Human Genome Project (HPG). Paradoxically, the results of the HGP cast doubt on assumptions that genes (i.e. DNA) could deterministically control for desired observable traits (Stevens and Richardson, 2015; Wynne 2005). One of the unexpected findings of the HGP was that observable trait expression was often dependent on many factors in addition to genetic material, including environmental and behavioral dynamics. In other words, greater knowledge about genome sequences did not necessarily translate into an increased predictive capacity to determine phenotypic traits. As a result, post-genomics— the era following the HGP—is often characterized as a move beyond the reductionist insights of molecular biology to include systems-biology, epigenetics, and broader “omics” approaches (e.g. proteomics and metabolomics). Post-genomics emphasizes “complexity, indeterminacy, and gene-environment interactions” (Stevens and Richardson, 2015: 3) to the effect that as molecular biologists Moore and Shenk put it, “contemporary biology has demonstrated beyond any doubt that traits are produced by interactions between genetic and nongenetic factors that occur in each moment of developmental time…(t)hat is to say, there are simply no such gene-only influences” (Moore and Shenk, 2017: n.p. emphasis in the original).
Methods
Empirically, this research draws on qualitative interviews and participant observation, including semi-structured interviews with researchers involved in the development of GS for forestry (38 in total) and personal observations while taking part in workshops (2) and team meetings (8). We organized two workshops, initially with the research team, and on a separate occasion with funders and end users, about how to communicate uncertainties associated with GS for forestry. Team meetings were held bi-annually throughout the tenure of the research grant. The participants in the workshops and team meetings were aware that we were present as observers, which may have influenced the responses. One-on-one expert interviews (collected between June 2017 and December 2018) revealed explicit and tacit assumptions about uncertainty, while observations from collective settings (e.g. workshops and team meetings) revealed how individuals grappled collectively with “uncertainty talk.” Interview participants were recruited through purposive sampling with the view to include a range of perspectives from people involved in the GS development for forestry using email invitations and personal introductions by colleagues. Disciplinary locations of interviewees included molecular genetics, chemistry, ecology, bioinformatics, and economics, ranging in rank from post-doctoral researchers to tenured professors. The majority of respondents (30/38) were male. Transcribed interviews and written observations from workshops and team meetings were coded collaboratively by all of the authors drawing on thematic analysis, a qualitative analytic process that examines the patterns of meaning by capturing key themes in a data set (Braun and Clarke, 2021). Thematic coding is an iterative process that involves familiarization, generation and refinement of themes, and writing up of results. To ensure anonymity, in-text quotations were distinguished using an anonymized numerical system that assigned an identifying number to quotes where letter R indicates respondent. All potentially identifying features, such as discipline, were not revealed in compliance with the formal ethics requirements.
Findings
As many of our respondents explained, the development of a GS model for forestry is a complex, time-consuming, and labor-intensive process, including the gathering and interpreting of genotypic and phenotypic data to establish a training population of trees. The development of a GS model requires interdisciplinary expertise from disciplines, such as molecular genetics, bioinformatics, chemistry, plant ecology, and economics. Genotyping involved the extraction of DNA from needle samples from select conifer trees, which was then sequenced into the genetic alphabet (A,T,G,C) and translated into a digital form to enable statistical analysis. Conifer trees—cone bearing and typically evergreen trees that comprise the majority of Canada's public forest land and provide a mainstay for the forestry industry—have large genomes: since it is cost prohibitive to sequence the entire conifer genome, a variety of complicated techniques with similarly complicated names are available to sequence genome segments, such as SNP chips, genotype by sequencing (GBS), and whole genotype sequencing (to list just a few). Each method offers distinct trade-offs in terms of the sequenced coverage of the genome versus cost effectiveness. Similarly, to supplement the bigger genomic picture from the fragmented sequences of DNA, several statistical methods are also available for imputation (i.e. filling in the gaps that necessarily result from sequencing genomes), including genomic linear unbiased prediction methods (GBLUP), genomic nonlinear Bayesian variable selection methods (Bayes A, Bayes B, BayesCπ, BayesDπ), and so forth. Notably, each sequencing technique and statistical method delivers different results (for elaboration, see Ala Noshahr et al., 2017; Habier et al., 2011). A myriad of other steps are involved in the development of a GS model, including the collection and documentation of phenotypic data to complete the establishment of the training population. Once refined, the goal of a GS predictive model is to enable the selection of individual trees with the highest probability of exhibiting desired traits to breed future generations of trees.
As this brief description suggests, multiple types of uncertainty are present in the development of a GS model, and even more types of uncertainty can be expected when a GS model is deployed in conifer forestry settings. For the most part, our interviews and group discussions revealed a broad acknowledgement of the uncertainties associated with the use of GS models for forestry. The following quote is indicative, in this case referring to the significant time frames involved in growing long-lived trees in Canadian forestry operations: If you’re in agriculture, you know, you plant a crop, it's ready in a year. Here, we have such long timeframes for everything we’re talking about. So uncertainty is almost a given. It's just a fact of life. R18
Paradoxically, in tandem with the broad-scale recognition of uncertainty as a “fact of life” in forestry, we also encountered confident declarations about the power of genomic science to provide trustworthy, reliable predictions by virtue of large genetic and phenotypic datasets paired with sophisticated statistical analysis. Although the level of confidence varied somewhat by person, most respondents described GS as an uncontroversial tool that would advance, improve, and enhance tree breeding operations by improving precision, accuracy, and efficiency. The advantage of GS over conventional approaches to breeding is that GS can provide breeders with information about trees that would otherwise not be available at the early stages of breeding. Excerpts from one interview, in particular, are illustrative of this confidence in the predictive capacity of GS. As this individual proclaimed, “breeders don’t need to wait for 30 years to know the wood density, I will tell them that when the seedling is seven days old” (R14). The assumption that GS can make prediction possible at the earlier stages of selection is rooted in a presumed deterministic relationship between genes and traits: Every single physiological function is actually determined by genes. I have never seen a function in any organism that's haphazard. Everything follows the blueprint. (R14) Predictive models have a level of accuracy. We never said the accuracy is 100 percent. We saw we have 60 percent, 70 percent, 73 percent, 80 percent… If I hear a scientist doing what we’re doing and saying I am hundred percent sure that I did that, I’ll say this is baloney. Because I know the technique. I know the shortcomings. I know the trade-offs that we make, I know the human errors that we do when we are measuring. (R14) For me it's all numbers. It doesn’t matter if it is wood density, it doesn’t matter if it's a tree resistant to drought or resistant to insects. It's all numbers. We crunch the numbers and develop a predictive model and the predictive model tells me which is best. (R14)
We often heard that another important factor in reducing GS uncertainty is the higher molecular resolution afforded by powerful genomic technologies. A common claim was that resolution and predictive accuracy improve with more data and a more sophisticated statistical analysis. The following quote is illustrative of this perspective: “when the resolution increases you see things better … when the resolution increases, we would expect our knowledge and our decisions will be much better” (R2). When asked to elaborate, this individual explained: Over the years, genetics has gone through a quantum leap, specifically with the introduction of genomics. The level of resolution became much higher and the accessibility to the entire genome became very evident. Therefore, when we trace our attributes from one generation to the other, we have better tools and better methods to be able to, with great certainty, determine the causal effects of those attributes. (R2)
Notably, while most respondents purported that scientists had the capacity to reduce uncertainty by generating more data, a repeated concern was that the unique features of forestry confound the GS predictive capacity due to high levels of variability associated with the scalar mismatch between laboratory research and the conditions that ensue when trees are actually planted on the land: When we are dealing with particular things relating to the environment, we are often looking at scales and time lengths that are far beyond what we normally do in a lab. Much of what we do in a lab has scales of hours or days or addressing a couple cubic meters of something. But if we are looking at the environment, we’re looking at square kilometers, hundreds, thousands of square kilometers. We’re looking at time periods of years, decades and centuries and when it comes to projecting and modelling, we are very poor at that as scientists. And we make and bet a lot on some cases the models that we have; some of which may be wrong or based on minimal data, some of which may not be reproducible in other jurisdictions. (R3)
In addition to the difficulties of predicting inconstant environmental conditions, such as water supply, long-lived trees also introduce social uncertainties that defy prediction. As one respondent pointed out, 30 years ago, who could have foreseen the marked effects the advent of the internet would ultimately have on the print industry? Looking forward, another respondent reflected: We’re measuring the traits of the trees now, in a given environment, but they’re usually harvested thirty years from now. And so, yes, the genetics now would be predictive of that trait, but that trait can evolve or change over time in response to the environment. And so, it might not be the trait that we want in thirty years or it could be if everything was status quo. (R5) We have no crystal ball that's going to tell us what those demands will be in, in even twenty years, let alone in 80 years. And it may be that the demands are going to be to leave those trees alone for their environmental and recreational attributes. (R18)
We also observed a marked difference between how uncertainty was discussed in interviews and in collective settings. While many respondents were frank about uncertainty in interviews, in collective settings in which we sought to encourage uncertainty talk, we encountered notable reticence bordering on resistance. Sometimes, we encountered emotional responses, such as frustration, suggestive of suspicion as to our motives, as if the conversation had veered away from constructive academic discourse to perceived attacks on GS, particularly, and genomic science more broadly. Some people openly acknowledged that they perceived uncertainty as something to avoid, which may help to explain discomfort in discussions of uncertainty and perceptual attempts to minimize it. For example, following an uncertainty workshop, a researcher remarked that a key revelation that emerged for them was that uncertainty is not necessarily a “dirty word.” Moments, such as these, which are not captured in formal interview data, provide a fuller picture of the perspectival landscape in which discussions of uncertainty unfold, and assessments of uncertainty take shape.
Uncertainty talk as uncomfortable knowledge
These findings offer key insights into the importance of position and context for uncertainty assessment. While conventional accounts of uncertainty often assume that scientists, as producers of knowledge, are the most familiar with the uncertainties related to their work, as suggested by MacKenzie's “certainty trough”, our findings paint a more nuanced picture by showing that scientists in applied research settings simultaneously acknowledged, but also overlooked, uncertainties depending on the context of interaction. This dynamic was best reflected in the discrepancy between the private acknowledgement of uncertainty in one-on-one interviews and its absence in collective settings. In interviews, at least some respondents were quite reflective regarding uncertainty, particularly with respect to technical uncertainties. Yet, we observed an overall lack of attention in collective settings, including team meetings and workshops, to discussions of uncertainty that extended beyond quantitative framings or even defensive postures when we encouraged uncertainty talk. Not only in collective settings, but also throughout interviews, an emphasis on the technical dimensions of uncertainty translated into calls for more scientific data and research, but neither for further scrutiny of social uncertainties, such as the assumptions that inform the development and application of genomic models, nor for the need for broader deliberations about the limitations of predictive knowledge in the face of environmental change. As a consequence, some areas of uncertainty, particularly social and environmental uncertainties, were not addressed at all in collective discussions.
Overall, the dynamics we encountered by engaging uncertainty talk with GS researchers resemble those described by Wynne, who identified tendencies among genomic scientists to simultaneously acknowledge “complexity as the limits of predictability” alongside a bracketing of genomic complexity and associated uncertainties (Wynne, 2005: 69; see also Saltelli et al., 2020). As Wynne elaborated, “predictive modelling requires a commitment to deterministic explanations,” which in turn reflects an “instrumentalist epistemology of modern science,” to support the commercial applications of applied science (Ibid, 2005: 76). For example, deterministic explanations were apparent in references to DNA as a blueprint, a metaphor that embodies an assumption that genetic material, on its own, can control and shape life (Stelmach and Nerlich, 2015). Social and biological researchers have written at length about the limitations of metaphors, such as blueprints, codes, alphabets, or books, to describe DNA. These metaphors reflect and maintain overly simplified assumptions about DNA's role in an organism's development (e.g. see Avise, 2001; Kay, 2000; Keller, 2002). Genomic metaphors evolve—or at least should evolve—in line with scientific insights. For instance, alternative metaphors that reflect the already known complexity and dynamism of DNA compare genomes with dynamic ecosystems or communities, where epigenetic (i.e. molecular factors external to DNA) and environmental factors also play a role in an organism's development (Avise, 2001). The use of metaphors, such as “DNA as blueprint,” suggests the tenacity of deterministic assumptions in applied research settings, assumptions which most likely inform the development and interpretation of GS models.
In addition, the narrowing of attention to quantifiable uncertainty was a dominant theme throughout the project, with direct implications for perceived confidence in GS biotechnologies and openness (or lack thereof) to discussions about the potential limitations of predictive knowledge. Numbers flatten, simplify, and reduce knowledge into useful information for decision-making (Ezrahi, 2004). Putting numbers on estimates of uncertainty has political resonance, particularly in domains that are vulnerable to criticism from outsiders, as Theodore Porter and others have argued (Ezrahi, 2004; Porter, 2006). The quantification of uncertainty lends legitimacy and authority to predictive science and enables scientific actors to maintain positions of authority and prestige, which insulates assumptions from scrutiny and contestation from outsiders. Scientists can be resistant to discussing uncertainty publicly or even with other researchers outside their disciplinary domains perhaps in part because doing so can be seen as tantamount to moving outside the confines of confidence conveyed by a position of authority. Admitting uncertainties, particularly incertitudes that are not readily quantifiable, challenges the fallacy of control central to modern science and environmental management. This is what Porter called the modern compact of science in society (Porter, 2006; see also: Arora, 2019; Stirling, 2007). The legitimacy of the modern compact requires faith in the impersonal and objective nature of numbers and in the capacity of the humans authorized to interpret these numbers to manipulate, engineer, and control natural systems.
The concept of uncomfortable knowledge and the broader social construction of ignorance shed light on why the dynamics of uncertainty talk may have unfolded as they did. Uncertainty talk in applied research settings is uncomfortable because it can be perceived as a threat to the legitimacy of applied research findings and to the legitimacy of applied research projects as a whole. Genomic scientists, along with funding agencies and industry partners that support GS research, have a vested interest in maintaining impressions of control in the management of uncertainty to facilitate the translation of research findings into industry applications. In applied settings, scientists are positioned as both producers of knowledge and beneficiaries of the uptake of that knowledge in ways that disincentivize open discussions of uncertainty. GS researchers and our own research team members may well have recognized that open acknowledgement of uncertainty could threaten the very auspices of the research project itself, the primary purpose of which was to identify ways to apply GS to forestry, and, by extension, the funding committed to such projects. Additionally, scientists conventionally trained to assess uncertainty quantitatively may resist expressions of uncertainty in forms for which they are unfamiliar, and that the lack of familiarity can feel threatening. The degree to which genomic scientists have been subject to critical scrutiny in the public sphere, as was the case with the use of rDNA in forestry, has also likely generated feelings of threat.
We speculate that power dynamics may also play a role in suppressing discussions of uncertainties that may have threatened the viability of the research project. Power dynamics played out tacitly not only between junior and senior members and between men and women, but also between disciplines, particularly those involved in the development and interpretation of predictive models (e.g. molecular biologists and bioinformaticians) and those located in holistic research disciplines, such as ecology and critical social science, where predictive capacity and reductionist assumptions are routinely questioned. Team members who endorsed a more cautionary approach to uncertainty may have felt disinclined to air such cautions in collective settings due to their positionality within the team. Addressing these power dynamics is clearly a key challenge to ensuring that diverse ways of knowing inform uncertainty assessments, including knowledge paradigms that challenge the modern compact of scientific objectivity.
Conclusions: Opening up “uncertainty talk” for predictive models
This paper advanced uncertainty talk for two reasons: to examine how experts involved in the creation of GS for forestry discussed uncertainty; and, to open up collective discussions about uncertainty associated with predictive models. We defined uncertainty along three registers: technical, environmental, and social. Uncertainty talk, as we proposed in this paper, aligns with a broader cultural shift, in which the reductionism of science and idealization of precision are increasingly challenged in the face of a broader recognition of complexity and uncertainties. The use of predictive bio-digital models, such as GS, both enables and resists this cultural shift. As our interviews revealed, most scientists recognize that models cannot provide definitive conclusions, and that scientific research is rife with uncertainties, particularly in applied interdisciplinary settings. Yet, our findings also highlighted the challenges of engaging collectively in uncertainty talk, particularly if the acknowledgement of non-quantifiable uncertainties is perceived as a threat. In the case of GS for forestry, uncertainty talk challenges the belief that dynamic biological, ecological, and social processes are fundamentally predictable and controllable through quantification. In the face of perceived threats to scientific and research legitimacy, experts are likely to downplay uncertainties and complexities that challenge assumptions about scientific prediction and control.
While we do not deny the usefulness of bio-digital predictive models, such as GS, for forestry, there is reason for concern if these technologies, anchored in as-yet-unrealized promises, are developed and embraced in the context of limited acknowledgement of uncertainties. Once implemented, GS will likely end up being far messier and complicated than was initially anticipated, leading to the potential for enthusiastically embraced anticipated outcomes to remain unrealized and for unanticipated outcomes to surface in their stead. Advanced high-throughput sequencing technologies, combined with conceptually robust insights from molecular biology and statistical analysis, will never eliminate uncertainties due to the inherent complexities in biological, technological, ecological, and social systems. Crucially, increased precision of a predictive genetic tool does nothing if underpinning assumptions are limited and if environmental and social variabilities are large. Since tree breeding and climate change operate over lengthy temporal scales, uncertainties are inescapably present, and the likelihood of surprise reinforces the need for broad, diverse, and ongoing discussions to facilitate social learning.
We conclude by underscoring the importance of qualitative assessments of uncertainty for predictive models and for supportive institutional arrangements to foster continued dialogue and learning. As with previous genetic technologies, such as rDNA and MAS, expectations in the initial stages of research and development are often deflated as the field matures and as difficulties and unanticipated outcomes inevitably arise. If implemented in forestry settings, GS will likely face a similar outcome. Uncertainty should be appraised and communicated as a routine part of applied science and interdisciplinary practice. Facilitating uncertainty talk requires challenging the presumption that uncertainty is problematic and something to be avoided or eradicated. Uncertainty talk provides an antidote to the idealization of precision and control: the generation of ever more data and advances in sophisticated technologies and statistical analysis cannot serve as a substitute for the challenges of negotiating differing values and perceptions about uncertainty in the face of technological, environmental, and social complexity.
Highlights
Qualitative uncertainty assessment necessary for bio-digital technologies, such as genomic selection;
Bio-digital technologies characterized by big data compilation, sophisticated statistical analysis, and high-throughput genome sequencing;
Expert interviews revealed barriers that impede more comprehensive discussions about uncertainty;
Barriers include belief that more data eliminates uncertainty, commitments to application, and deterministic assumptions about DNA.
Footnotes
Acknowledgements
The authors are grateful for the contributions from their research participants. They thank the two anonymous reviewers and the editor for the feedback on the previous versions of the manuscript. They also acknowledge support from the research team, and funders including Genome Alberta, Genome Canada, the University of Calgary and the University of Alberta, which made this research possible.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by Genome Alberta.