The Changing Concept of Actionability in Precision Medicine

Regulators: Consortia, Guidelines, Recommendations

The agents contributing most directly and visibly to the establishment and standardization of the actionability concept were groups issuing statements and recommendations. This is common in medicine, where protocols, standards, and guidelines are of central importance. Their presence became encompassing in Western biomedicine in the 1960s and 70s, tightly connected to the relation between scientific research and clinical practice (Eddy, 1990). Research protocols, aimed at securing verifiability and replicability of procedures beyond local contexts, served as a model for the standardization of care and, today, coordinate clinical action with scientific, legal, economic, and political requirements. As research shows, such protocols are of particular importance for medicine wherever technological tools are applied (M. Berg, 1998) and wherever management of test results is at stake (Weisz et al., 2007). These are exactly the situations in which the first standards for deciding on actionability emerged.

The earliest instances (around 2010) of guidelines giving definitions and descriptions of actionability were concerned, on the one hand, with the topic of ‘return of findings’—whether and when to report study results to participants—in research contexts (Fabsitz et al., 2010; National Cancer Institute, 2011), and, on the other hand, with ‘translat[ing] laboratory test results into actionable prescribing decisions for specific drugs’ in pharmacogenomics (Relling & Klein, 2011, p. 464). These very first documents were recommendations issued by public health bodies, for example, the National Institutes of Health (NIH) Office of Biorepositories and Biospecimen Research and the US National Heart, Lung, and Blood Institute. Shortly after, these and similar agencies started creating independent working groups, normally composed of experts from prestigious medical institutions and tasked specifically with elaborating recommendations on how to understand and assess actionability. Occasionally their personnel overlapped and they merged into new groups, depending strongly on funding periods in a highly project-oriented area. The most important guidelines for our matter (based on cross-references and authors’ prestige), are the ones issued by the Actionability and the Somatic Working Group of the NIH-funded ClinGen consortium, and by the ACMG. A central objective of such guidelines was establishing a list of variants that should always be reported.

There is a strong bias towards actors from the United States, even exceeding what the general predominance of US-based literature production would lead to expect. Owens (2022) hypothesizes that this might be due to the litigious US-American context, where professional inaction easily leads to legal consequences. Information on the actionability of genetic findings, according to this explanation, is crucial not only for patients, but also for their physicians with guidelines providing assurance in matters of accountability. Initiatives outside the US tend to draw on guidelines from US institutions as international standards, because they were earlier to be established. In addition, as US agencies are often major funders of international teams, their work is often published under these agencies’ banner.

Repositories: Databases, Knowledgebases, Digital Platforms

Not less important are databases, mostly providing information about associations between genetic entities, phenotypes, and drugs. These platforms are much more than spaces for storage of static information. Rather, they commonly serve as interactive digital tools for precision-medical data interpretation (Cambrosio et al., 2020). Some are structured around differences in populations (e.g., 1000 Genomes Project), disease types (e.g., DECIPHER for rare genetic disorders), or specific types of alterations (e.g., dbSNP for small variations). Stevens (2013) has argued that changes in the structure of repositories like GenBank decisively shaped the development of a molecular-biological episteme. For the case of actionability, meta-reviews Klicken oder tippen Sie hier, um Text einzugeben.have gone so far as to call determining the clinical actionability of variants the central aim of such databases (Gao et al., 2019; Li & Warner, 2020).

Easy accessibility is important to ensuring databases their central role. Research on the origins of biomedical repositories Klicken oder tippen Sie hier, um Text einzugeben.has shown how far-reaching the early regulatory decision to make public access and standardized formats compulsory turned out to be (García-Sancho, 2012; Strasser, 2019). While some of the databases analyzed offered upgrades to versions with additional functionality or datasets, and some sold licensed local installations integrated into a medical center’s IT infrastructure, most were publicly available. The US-bias is not as strong in this area as it is in that of regulators.

Drawing on Ainscough et al. (2016, p. 806), databases containing information about variant actionability (e.g., CIViC) can be imagined as the peak of a pyramid, building on databases containing functional annotations (e.g., DoCM), which build on databases containing clinical assertions (e.g., ClinVar), which build on databases containing observed variants (e.g., ICGC, COSMIC, TCGA), which build on discovery and pathogenicity classification. Climbing this pyramid, the numbers of recorded variants decrease, while complexity and effort at curation increases. The development of the field saw a continuous shift of focus from the beginning to the end of the precision medicine pipeline, that is, from discovery to screening (beginning of the 2010s), to diagnostic testing (around 2013), to therapeutic application (around 2016). Roughly speaking, the closer to the peak of the pyramid, the later a database was introduced. Along the way, some databases disappeared, some lost significance, but mainly new databases filled niches created by the overall evolution of precision medicine.

Actionability is part of this development. From 2012 onward, the term began to appear in the self-descriptions of major projects and databases. The MyCancerGenome project referred to ‘a new term, actionable mutation’ (Swanton, 2012, p. 668), while databases such as CanDL (2014), COSMIC (from version 69, 2014) and DGldb (from version 2.0 in 2016) incorporated it into their frameworks.

Tools: Assays, Algorithms, Decision Support Software

Among the most important technologies enabling precision medicine are new sequencing modalities, some of them capable of reading millions of DNA fragments in parallel, thus allowing high throughput, speed, and scalability. Reading and aligning these fragments, ‘calling variants’ (identifying differences from a reference genome), and assessing their potential phenotypical impact requires bundles of digital algorithms. While databases and guidelines often directly work towards establishing a professional reading of actionability, the various sequencing assays and software products providing the field with modalities for finding actionable variants in genomes tend to be more indirectly involved in this conceptual work. Since these technologies normally serve the purpose of variant classification, which is antecedent to variant interpretation, whenever their providers refer to actionability, they tend to cite external definitions.

Nevertheless, providers of assays and prediction tools defined their products as detecting clinically actionable variants, while discrepancies between facilities prompted early calls for standardization (e.g., Manolio et al., 2013). Many of the guidelines were answers to these calls.

Studies: ‘Maps of the Actionability Landscape’

A last category comprises contributions to what the field sometimes calls the actionability landscape (Suehnholz et al., 2024). In close connection with the capabilities of sequencing facilities and large-scale data repositories, research teams began designing (mostly retrospective) cohort studies to determine the prevalence of specific genetic alterations in the general population, as well as the average number of actionable mutations identified when analyzing an individual’s exome or genome. The results of such studies help, among other things, to assess the penetrance of gene-disease associations: Only for a very small number of genetic variants do all carriers display a pathological phenotype, so that the strength of connection between genotype and phenotype can be given as a percentage of affected people displaying the pathology. These insights then feed back into the assessment of the variant’s actionability: the higher the penetrance, the more likely it should be considered actionable.

The first trials in the NIH National Library of Medicine mentioning actionability in the context of molecular medicine were registered in 2011 ¹ and the numbers of associated publications peaked in a phase between 2014 and 2017. In them, the term actionable alteration is often explicitly defined (often in an annex or supplementary information section) in, or similar to, the following way: ‘An actionable alteration was defined as an alteration that was either the direct target or a pathway component that could be targeted by at least one approved or investigational drug’ (Schwaederle et al., 2015, p. 1489).

Assuming a broad understanding of ‘actionability landscape mapping’, I included a second type of document into this category: meta-studies, reviews, commentaries, and opinion pieces discussing the implications of the evolving maps, be it on a conceptual level or for the organization of precision medicine. Several reviews, starting from around 2016, point out considerable discrepancies in the numbers of (potentially) actionable alterations between the landscape studies, sparking discussions about the viability of the category itself: ‘the definition of “actionability” used to assess the clinical utility of molecular profiling varies widely across different studies and institutions. Reports published in the past 2 years … suggest that between 30% and 94% of patients harbour actionable mutations’ (Berger & Mardis, 2018, p. 359). Another focus of these more conceptually aimed pieces is a potentially problematic discrepancy between the emerging meaning of clinical actionability and what individual patients would practically regard as actionable.

Establishing Actionability as a Quality of Variants

The first and most noticeable development concerns what is commonly labeled as actionable. Actionability, in a technical sense, gradually shifted from being used as a label for test results in knowledge management into a quality possessed by genetic variations. Having recorded information on this aspect systematically (variable 4), it was relatively straightforward to compare whether documents ascribing actionability to the same entities also stemmed from the same clinical contexts (variable 5).

The first discussions on actionability as a technical term in precision medicine took place in the debate about ‘return of incidental findings’ at the end of the 2000s (see Owens, 2022, pp. 5–7). Since everyone’s genome harbors genetic alterations that are not phenotypically expressed, the

[p]henomenon of possible incidental genomic findings—the incidentalome—threatens to undermine the promise of molecular medicine. In particular, the application of comprehensive genotype and functional genomic measurements across the general population is likely to yield unexpected incidental findings for nearly everyone. (Kohane et al., 2006)

The first context where this problem practically manifested itself was molecular-biological research. Given that the earliest applications of genetic sequencing were scientific, trial participants were the first to face the question of whether they should be informed of dormant genetic risks. Accordingly, actionability first emerged in bioethics debates and discussions about clinical implications of biomedical studies. The earliest mention of actionability in my corpus described the results of a survey on public concerns regarding the establishment of a national genetic biobank, published in a bioethics journal (Murphy et al., 2008, p. 39). The first guideline mentioning actionability (Fabsitz et al., 2010), and the first scoring system applied to actionability concern the ‘return of findings’ in study contexts. This is an update of Bookman et al. (2006), and marks the introduction of actionability as a technical term. The ethical and legal aspects incentivized, among other things, the creation of the ‘Return of Results consortium’ (in 2012; later turned into the Clinical Sequencing Evidence-Generating Research program) by the National Human Genome Research Institute and the eMERGE network.

My analysis suggests that the issue of incidental (or ‘secondary’ ² ) findings passed into the clinic around 2010, in the context of screening and diagnostic tests. Here, findings are incidental/secondary in that they do not refer to those pathologies for which the diagnostic test or screening had originally been ordered. Instead of a bioethical problem, return-of-findings is treated as a selection problem in the face of information overload. In developing next-generation sequencing (NGS) into a clinical diagnostic test the idea of ‘binning’ the findings according to their actionability (J. S. Berg et al., 2011, 2016)—sorting them into internally consistent but revisable tiers—emerged.

Before terminological use converged towards actionable variants/alterations in the mid-2010s, expressions like actionable genotypes, clinically actionable regions of the genome (in research on copy number variations), actionable prescribing decisions, and actionable genomic events (in pharmacogenomics) were found. Specifications such as clinically or medically actionable were more frequent in the early years of precision medicine and tended to be shortened to actionable over time. Until about 2014 it was not uncommon to find actionability written in quotation marks, indicating that it was still considered new terminology. The consolidation of the terms variant and alteration—more neutral than mutation—then turned the marker of a difference (the variation) into a positive entity to which, in a subsequent step, actionability was ascribed as a quality. Actionability turned from the evaluation of an administrative process (whether to take medical action for this person) to a selection criterion in clinical practice.

On a broader scale, this shift from actionable findings to actionable genes/variants is tied to the general path of precision medicine from exclusive research and preventive genetic screening to diagnostic genomic profiling of patients with a disease, to ‘post-genomic’ evaluation of treatment options based on molecular markers (Richardson & Stevens, 2015). Around the turn of millennium, the label genetic (medicine) was still standardly associated with hereditary (disease) and visions of future molecular medicine focused on the idea of testing ‘pre-patients’ for non-symptomatic genetic risks. The subsequent development turned increasingly towards diagnostic facilities for the undoubtedly ill, and starting from the first half of the 2010s, the merely diagnostic tests were more often integrated with therapeutic indications—not least via the concept of actionability.

The data reflect such an expansion by indicating an increasing interest in actionable somatic alterations and non-hereditary disease (variable 5) around 2016. In that year, somatic alterations still made up only about 2% of the usage of ClinVar dataset and ‘a similar, though distinct, language is often applied for interpreting somatic variants. For example, germline variants may be categorized as pathogenic, while somatic variants are often categorized as diagnostic, prognostic, or predictive biomarkers’ (Ritter et al., 2016). The most important genetic databases in the first half of the 2010s were the Human Gene Mutation Database (HGMD) and Online Mendelian Inheritance in Man (OMIM) for which the term actionability played no role. This changed with the following generation of somatic variant databases and understandings of actionability shifted in at least two respects. First, with the shift away from hereditary disease, informing and screening relatives disappeared from the spectrum of possibly-indicated medical actions comprised in actionability. Second, the effect of the single variant became even more uncertain the more the interplay of various potentially pathogenic variants came into focus.

The actionable variant is so established today that its recency is easily forgotten. Moreover, history is often apocryphally reported. Two early precision medicine landmark publications engaged in retrospective terminological adoption (Ashley et al., 2010; Lawrence et al., 2014). In both cases, though the term was not used in these publications, later authors claimed that they referred to actionability. ³ As such, actionability and the epistemic object of the variant co-evolved into key elements of precision medicine, usable as inputs and outputs of testing devices and databases, as those moved increasingly into the clinic—conceptual and material infrastructures stabilizing each other.

Delineating Actionability Against Other Categories and Scores

Another development concerns the relationship between actionability and similar categories. The data I recorded (variable 2) indicates a particular importance of clinical utility in this regard and a general trend towards increasing independence of actionability from neighboring concepts. While the early years of the development see a wild proliferation of such scales, starting with the already mentioned ‘binning system’, from about 2018 on, consolidation has been particularly strong in oncology, where most US-American institutions follow the OncoKB actionability scale and most European ones the ESMO scale.

Focusing on the early phase, particularly rich resources were the presentations and discussion that took place at the NIH ‘Characterizing and Displaying Genetic Variants for Clinical Action Workshop’ in December 2011 (see National Institutes of Health [NIH], 2011). Participants included the most influential experts working on the concept at the time. The discussion repeatedly turned to whether the term actionability possesses a specific meaning at all and if it should be given up in favor of the already established clinical utility—a term used synonymously in many other early instances.

My analysis also suggests a strong correlation between its movement towards independence and the development of scoring and classification systems (variable 3), used as secondary, mostly ordinal, actionability scales. These draw on information from different parts of the precision medicine pipeline. Three among them stand out: First, in 2015, the ACMG consensus recommendations defined a since-then standardly used five-tier pathogenicity score (pathogenic, likely pathogenic, uncertain significance, likely benign, benign) of which it is stated that it was ‘already in use by a majority of laboratories’ (Richards et al., 2015, p. 406) and which can be traced back to a set of IARC Recommendations (Plon et al., 2008). In many instances the classification comes with recommended actions for each tier. Second, the concept of levels of evidence was adopted from the evidence-based medicine movement. The foundations for its connection with actionability were laid in oncology, with the elaboration of a ‘Tumor Marker Utility Grading System’ by the American Society of Clinical Oncology (ASCO), containing a six-tier utility scale and an additional scale of five levels of evidence (Hayes et al., 1996). Third, where precision medicine touched on pharmacogenomics, the idea of ‘the druggable genome’ (Hopkins & Groom, 2002) was sometimes adopted in the form of a druggability score, prominently in the DGldb database.

Thus, although binning systems are part of precision medicine’s toolbox from the early 2010s, a stabilization of actionability tier system began around 2018. Actionability was increasingly used as a binary category (a variant is either actionable or not) or at least as an ordinal scale, with secondary grading systems (i.e., pathogenicity, druggability, and level of certainty/evidence) additionally qualifying it. Actionability thus emancipated itself from earlier concepts like clinical utility through increasing reliance on scores, artifacts characteristic of an emerging algorithmic logic (Fourcade & Healy, 2024).

Concretizing Actions

The meaning of the term actionability depends on the context and the type of suggested action (variable 1 in my template). With the growing diversification of usages for gene variant interpretation, represented in my data by a gradual expansion from human genetics to actors from other subfields (variable 5), this was increasingly being acknowledged. An illustrative case is the following statement of the Evaluation in Genomic Applications in Practice and Prevention (EGAPP) Working Group:

For disease-free patients, actionability implies an effective intervention to delay or prevent clinical manifestations …. For patients with undetected disease, actionability includes screening …. For symptomatic (but clinically unrecognized) patients, actionability includes alterations in patient management …. Other actions include family management … and avoidance of circumstances for the patient …. Sufficient support for actionability is derived from a practice guideline, expert-derived guidance, or a systematic review. If no such guidance or review exists, our process defines the gene/condition pair as not actionable. (Goddard et al., 2013, p. 724)

Note the vagueness of the indicated interventions. It is mainly due to the relative emphasis on disease-free, asymptomatic, and ‘clinically unrecognized’ patients. Prevention tends to be more general than intervention.

With the transfer of genetic testing into the clinic came a concretization—an overall trend towards increasing specification of actions. This was closely linked to the rise in approvals of precision medical drugs. In pharmacogenomics and in the descriptions of gene panel usage, short and concrete definitions can already be found during the earliest stage of the conceptual development, mostly of the following type: ‘Actionable genomic alterations are those for which there is an FDA-approved treatment or that serve as eligibility criteria for later phase clinical trials’ (Lanman et al., 2015, p. 3). Consequently, the more targeted therapies were approved, and the more precision medical trials were conducted, the more concrete actionability can be found. Around the middle of the 2010s, the introduction of such drugs was anticipated, as is evident from the fact that terms like ‘actionable in the future’ and ‘near actionable’ temporarily appeared on the scene. This future becomes present with the skyrocketing number of targeted drug approvals in the year 2017 (Suehnholz et al., 2024).

The actionability landscape studies fulfil an intermediary role between genomics and pharmacogenomics here. On the one hand, they used gene panels and often adopted their actionability definition in methods sections. On the other hand, the frequency statistics derived from them not only aimed to quantify how many patients have certain genetic aberrations, but sometimes also the numbers of available drugs for such patients, for example as a ‘percentage of medications with actionable germline pharmacogenetics’ (Relling & Evans, 2015).

With the growing importance of an actionability notion tied to the availability of approved drugs, the overall spectrum of the concept had widened to span the area between druggability and clinical utility. In the early years, professionals argued that a broad definition of actionability should be accepted (Dienstmann et al., 2014; National Cancer Institute, 2011), one that would not be limited to druggability. During this first period of precision medicine, testing for the presence of a disease-related genetic variation was separated from the testing for resistance or response to drugs and susceptibility to side-effects: Just as screening is different from diagnostic testing, so diagnostic testing was technically and semantically distinguished from pharmacogenomic testing (McCarthy et al., 2013). Analogously, pharmacogenomic databases (like the Pharmacogenomics Knowledge Base, PharmGKB, the Drug-Gene Interaction Database, DGIdb, the Therapeutic Target Database, TTD or the Genomics of Drug Sensitivity in Cancer database, GDSC) existed separately from variant databases. As whole-exome and whole-genome sequencing gained traction, the positions were brought closer together, since they created the possibility of using a single test for discovering both possibly pathogenic variations and variations that influenced the patient’s drug reaction. A watershed for the reconciliation of diagnostic and therapy-oriented logics of genomic analysis was reached around the year 2017, with the introduction of several databases specializing in therapeutic actionability, prioritizing predictive over diagnostic biomarkers. This especially included the Cancer Biomarkers Database, providing predictive software for in silico gene interpretation, JAX-CKB, noting that the selection of actionable variants could only be an intermediate step in therapeutic variant interpretation, and OncoKB, placing actionability in the center of its setup. The shift in database development towards clinical application was also represented by the fact that OncoKB, PMKB, and CIViC, all databases from this new generation, included a system of actionability tiers. Their webpages did not spot the disclaimer standardly seen on those of older databases, stating that the content ‘should be considered as an educational and research tool only and not be used for medical purposes’. Being actionable meant that action was warranted, and over the course of precision medicine’s development the available actions become narrower in scope (focused on the administration of specific drugs) yet broader in number as more drugs were approved. This movement has made the meaning of actionability increasingly specific.

Standardizing Standards

Most of what the prolific sociological literature on standards in health care contexts has established (Timmermans & Berg, 2003) also applies to molecular medicine (Hedgecoe et al., 2023; Timmermans, 2015). Attempts to set clear rules about when to speak of actionability were elaborated early on in the development of precision medicine. However, I focus on the development of second-order standardization, that is, the standardization of standardizations over time. My identification of this process is mainly based on the analytic variable of cross-references (variable 7 in my template). This harmonization of standards is closely connected to the creation of a growing number of precision medical institutions. Standing out among them are Molecular Tumor Boards (MTBs)—interdisciplinary expert panels issuing treatment recommendations based on a molecular oncological rationale for cancer patients. Their internal guidelines often commit them to using one certain standard for assessing actionability.

More broadly, the harmonization of standards is also considered a desideratum in the field. This is instantiated by one type of literature, namely meta-studies collecting numbers of actionable variants reported from the same sample by different tools or institutions. Such comparative meta-studies exist for databases (Lazareva et al., 2024; Pallarz et al., 2019), assays (Perakis et al., 2020), software, population studies, and MTBs (Rieke et al., 2018; van der Velden et al., 2017). They usually find the numbers diverging considerably, which leads to calls for more uniform standards.

Subsequent attempts at harmonization of standards involved working groups, consortia, and managers of databases, most notably the OncoKB knowledgebase. For example, in 2018 OncoKB adapted its categorization system to align with the ESCAT scale of the European Society for Medical Oncology (Mateo et al., 2018) and in 2021 it announced that it had become ‘the first somatic human variant database to be recognized by the FDA’ (OncoKB, 2021).

The tendency of patient and clinician advocacy groups to claim a (re)turn to broader actionability concepts—‘from medically actionable to patient actionable genes’ (Gornick et al., 2019; Sebastian et al., 2022)—is a critical reaction to the growing technicality that results from (standardization of) standardization.

The Translational Problem

The m-dimension of the actionability concept—the fact that it presupposes a context which not only allows for, but also requires, action—highlights both the first conceptual development described above and the particular importance of actionability in medicine.

As noted, the earliest conceptualizations of actionability stem from the context of ‘return of findings’ to research participants (i.e., from the frontier between science and medicine). Thus, the first problem actionability helps solve is deciding whether a certain data point, a genetic finding, should be treated as medical information. Actionable means medically relevant; clinically actionable variants are not just any gene mutations, associated with arbitrary phenotypical features like hair loss or eye color, but potential indicators for serious illness; as such, they require action. That it is not only possible, but also necessary to act on findings, results from the sense of urgency characteristic of the clinical medical context. This additionally helps make sense of the fact that in the early consolidating years the label actionable often appears in combination with the qualification medically or clinically. At the beginning of the first development described above (from the actionability of incidental findings to targetable actionable variants), actionability is the reaction to an unforeseen secondary problem in medical research—faced ‘when research seems like care’, as the title of one article puts it (Miller et al., 2008). Here lies the key to answering my first research question about the particular relevance of actionability in medicine. Its original background is the correspondence between the abstract m-dimension of actionability and medicine’s general aversion to inaction.

But this leads to the question of why an additional category is necessary, given that the label of the (potentially) pathogenic alteration already existed. To see the additional value of actionability, we must consider that the decoding of the human genome has led to the insight that most diseases with a strong genomic component are not monogenetic. Mendelian diseases—where the exact pathogenic mutation responsible for the pathology can be pinpointed—are exceptions. In diseases with complex etiology, molecular alterations interact. If a variant is undoubtedly associated with a pathological phenotype, action must be taken—no further classification needed. But the more uncertain the effect of a variant, the more physicians need to know whether to take action or not. In this sense, the actionability concept attends to the clinical implications of the biological untenability of the one-gene-one-trait assumption (Tempini & Leonelli, 2021, p. 2).This helps clarify which problem is solved by ascribing actionability to alterations instead of findings. In the latter form, it is standardizable. Lists of alterations that should always be reported can be written to help transfer the finding from the laboratory into the clinic without subscribing to the obsolete one-gene-one-disease assumption.

The Information Problem

This solution, however, comes with far-reaching implications. Classifying a finding as actionable means assessing the variant relative to the person whose cells carry it and to the situation in which she finds herself. Meanwhile, identifying the variant as actionable means assessing it independently of its carrier. This conflicts, however, with the insight that molecular alterations interact and thus their phenotypical effects depend on context.

This contradiction is acceptable while the number of actionable alterations per patient is small and while the focus is mainly on hereditary monogenetic disease. It becomes problematic once the inevitable incompleteness of information conveyed by a single variant is combined with an excess of data. While in most cases a single genomic alteration is insufficient to explain and predict a pathology, technological developments in genomic sequencing (i.e., increasing number of variants included in DNA microarrays or increasing affordability of whole-exome sequencing) lead to the return of more than a single potentially pathogenic variant per patient. Using the model of the three dimensions combined in actionability, this can be described as an asymmetry characterized by an overhang of the k-dimension, which I term the information problem.

Practically, this problem takes two different forms in two different settings. While the translation problem is situated at the bottleneck between medicine-as-science and medicine-as-clinical-care, the information problem appears in the realm between genetic screening and diagnostic testing. In screening, the excess of the k-dimension is mostly over the m-dimension (more is known than must be acted upon). The analogous problem to the incidentalome introduced above is, thus, overdiagnosis: Not all potentially relevant alterations found for an individual require action. The ever-growing list of variants between the guideline versions I analyzed testify to the buildup of the problem. In diagnostic testing, on the other hand, it manifests as an excess of the k-dimension over the c-dimension (more is known than can be acted upon), due to the focus on potential causes for a given illness, instead of potential illness given potential causes in screening.

Information overload is thus a problem in both preventive screening for hereditary disease and the diagnosis of complex pathologies. In both settings, the solution is a filter that pre-selects a limited number of variants for referral to a decision-making stage after the research or test laboratory. Actionability is well-suited to providing this solution due to its multidimensionality. Selection criteria, the m- and c-dimensions, are built into it. To sharpen these selection criteria, the concept needed to be more clearly delineated against similar ones and the recommended actions needed to be concretized. The spread of documents assessing variant penetrance around 2013 provides an approximate timeframe for this.

In the early 2000s, literature on geneticization drew attention to potential implications of the information problem in genomic testing, shaping the public imagination of a geneticized future in which everyone was constantly at (genetic) risk, turning them into pre-symptomatically ill pre-patients. Actionability can be understood as an empirical solution to the concrete manifestations of these predictions in screening. When the binary distinction between illness and health loses its meaning because there is nothing left that can be ascribed the value healthy, secondary distinctions are needed. The distinction between actionable and non-actionable can be used as such a secondary distinction.

One consequence is a dynamization of the actionability concept. New insights from clinical studies and newly developed therapies are constantly changing the contents of all three actionability dimensions. Regularly published lists and static repositories are not suited to representing this type of category. Curated knowledge bases, on the other hand, are a technical enabling condition. The new generation of (mainly cancer-focused) knowledge bases, for example, are characterized by combining pharmacogenomic and diagnostic data formerly stored in separate databases.

Another consequence is a shift of focus from screening towards diagnostic testing. In practice, genomic testing, especially with the expansion from hereditary genetic diseases to those caused by somatic alterations, has greater importance for the undoubtedly ill than for pre-patients. Attributing this trend directly to the influence of the actionability concept would be an overstatement. A central factor, independent of actionability, is that patient care is institutionalized within organizations, whereas general prevention is not. Nevertheless, the medical bias against inaction, combined with the growing practicality of the actionability concept, gives precedence to a setting where practitioners are confronted with the question of what to do rather than whether to do something. Acting as a solution to the information problem, adaptions to the actionability concept inadvertently pushed genomic sequencing further into the clinic.

The Complexity Problem

While providing a solution to the information problem, sharpening the profile of actionability through the weighting of its three dimensions led to a new difficulty: Now it becomes necessary to determine how the single dimensions should be integrated and weighted. I term this the complexity problem because it results from the attempt to compress complexity into a single measure, while at the same time holding it accessible. Note that this is closely related to the problem resulting from contradictions between standardization and individualization. At first glance, the ascription of actionability as a context-insensitive quality of some genomic variants seems to contradict the promise of personalization, which is central to precision medicine.

To see how further adaptions to the actionability concept helped address these complications, it is helpful to take a closer look at the contrasting concept of clinical utility. This concept was already firmly established when actionability entered the field. Owing to standardization efforts from the time of the first genetic screening tests in the 1970s, an NIH Task Force on Genetic Testing had introduced the threefold distinction between ‘analytical validity’, ‘clinical validity’, and ‘utility (to those tested)’ (Holtzman & Watson, 1997). The first two of these concepts—referring, respectively, to the accuracy of a test at determining the genetic indicator and the accuracy of the indicator at determining the risk of disease—stemmed from test statistics. Utility, on the other hand, was described as representing the ‘benefits and risks of positive as well as negative test results’ (Holtzman & Watson, 1997). This is not an easily formalizable description. Clinical utility was always a vaguely defined concept, the risks and benefits it supposedly measures being case-specific and distributed across various domains. This reflects the vagueness of the early actionability concept, not yet clearly distinguished from utility.

Clinical utility attempts to integrate multiple dimensions into a single measure. Actionability does the same through the combination of the three dimensions. Their initial similarity stems from the fact that they both address the complexity problem. The advantage of the actionability approach over the utility approach in managing it is apparent when considering that utility is a gradual measure and whenever quantified it is usually calculated as a weighted sum of its component dimensions (‘risks and benefits’). Actionability, on the other hand, is inherently discrete, corresponding to the shift from actionable findings to actionable variants. While knowledge can vary in utility, a variant can be actionable or not. This binarity, however, does not preclude the introduction of levels of actionability; most actionability scores I surveyed come in tiers and scales. Importantly, however, these are secondary and ordinal measures. The actionability level is commonly presented as an additional measure, once the actionability of a variant is established. For example, in the four-level scheme proposed by the Clinical Pharmacogenetics Implementation Consortium of the Pharmacogenomics Research Network (CPIC), levels A and B are combined to mean ‘actionable’ and levels C and D to mean ‘not actionable’ (Relling et al., 2020, p. 172).

Each of actionability’s three dimensions can be continuously graded: via the certainty of what is known (k-dimension), the gravity of the pathology (m-dimension), and the effectiveness of the available interventions (c-dimension). But the resulting score is typically separated from the actionable/not-actionable distinction itself. While clinical utility was an attempt to address the complexity problem by integrating various dimensions of risks and benefits into a single measure, actionability did the same by decomposing the problem. This made actionability easily operable, its tiers being sorted, categorized, cross-referenced, and dynamically updated in knowledge bases. In short, molecular medicine responded to the complexity problem by adapting actionability in a way that distributes the complexity over secondary (and tertiary) scores while keeping the actionability category itself simple.

The Integration Problem

This arrangement, however, only shifted the contradictions of standardization-with-personalization and of complexity-reduction-with-complexity-conservation. Consider the frequent concerns found in my document analysis regarding discrepancies in actionability assessments between panels, decision support platforms, and landscape studies. For an outside observer, they are not overly surprising. After all, precision medicine presents itself as being particularly personalized, with treatment tailored to each patient’s individual genetic makeup. Variation and discrepancies would be expected in such a setting. Standardization and second-order standardization are often presented as a solution to these shortcomings. However, as we saw, while standards abound, the problem persists. Once the significance of the actionability label has been confined to instructions on where an expert should look, an institutionalized system of further steps is necessary. Secondary actionability scales must be established, curated, and interpreted through expert assessment. This brings me to a fourth and last problem. The documents I analyzed contain numerous claims that general and non-specialized practitioners lack the knowledge and training necessary to interpret actionability results from genomic tests. The reiterated problem-solving described so far, through which actionability was repeatedly adapted, produced a system with its own logic—one not easily compatible with non-specialized general health care. I call this the integration problem.

Traditional biomedicine tends to proceed in two steps: first, conditioned by a set of symptoms, signs and markers, a diagnosis is made in the form of a selection of a disease category from a spectrum of different options. Then, conditioned by the diagnosis plus a set of patient characteristics, a therapy decision is made in the form of a selection of treatment options from a spectrum of drugs and interventions. Diagnosis and treatment are two steps connecting three types of entities: (1) symptoms, signs, and markers are mapped onto (2) diseases/conditions, which are mapped onto (3) drugs/treatment devices. These components do not change in the precision medical actionability paradigm. On the contrary, genetic variants fit easily into the category of (bio)markers and the tripartite structure coincides perfectly with the three actionability dimensions. The ClinGen Actionability Working Group, for instance, selects the three entities ‘gene’, ‘phenotype’ and ‘intervention’ for actionability scoring (Hunter et al., 2016) and in an interview given at the launch of the ‘COSMIC Actionability’ database, its curator described ‘the unit within the database’ as ‘the “drug mutation disease triple”’ (Jupe, 2021). What changed radically in the actionability paradigm was the form and sequence of relating the components of the arrangement upon each other.

In the classical arrangement, classificatory diagnosis is at the heart of medical practice. It is supported by various sources of clinical evidence, including test results. In the context of ‘return of findings’ in which actionability was introduced into precision medicine, this is still the underlying logic. However, this logic changed when diagnostic and pharmacogenomic testing were combined into one single practice. Actionability was the juncture between them. Treatability formed part of the actionability diagnosis, skipping the classification stage. In classical biomedicine, the Archimedean point is the disease classification, in precision medicine it is the variant. Armstrong (2019) has accordingly described this as a transition ‘from classification to prediction’. This does not mean that disease classification is completely excluded from precision medicine. However, the hierarchical relationship between it and the variant is reversed. This is most evident in precision oncology, where the classification of the tumor type, which formerly guided treatment decisions, is now just one of the criteria determining the actionability tier of a variant for a specific patient (Laßmann & Hummel, 2021). Classification did not disappear from diagnosis, but was degraded to the rank of one data point among others—an exact reversal of the roles traditionally taken by tumor type and genetic test results.

The solution to the problem of integrating this logic with the work of existing health care institutions is (infra-)structural. While fashionable, well-funded, and considered cutting-edge, precision medicine still takes up a very confined place inside the medical system as a whole, and it is often isolated in specialized centers. Sometimes, genetic tests do enter the clinic as merely another tool (Kaufman et al., 2024), but often the shift ‘from test to prediction’ is a rather isolated phenomenon for specialized precision medical institutes. Variant interpretation is done by specialists and forwarded to treating physicians in the form of treatment recommendations. Precision medicine has its own tests, its own experts, and its own drugs. Kuiper et al. (2025) have recently interpreted this relative isolation of expert variant interpretation as the result of successful boundary-work by geneticists afraid of losing their domain of specialization. Approaching the phenomenon from the perspective of the local shift from diagnosis to prediction offers an alternative reading. Rather than being implemented into routine practice as one modality among others, precision medicine has carved out a niche, especially in oncology, through reiterated problem solving.

Institutionalizations like that of consortia issuing guidelines and cancer centers running Molecular Tumor Boards (MTB) function as boundaries for this niche. Within them, the traditional relationship between the variant and the disease type can be reversed. But this is only possible because the niche is still part of an environment where the more traditional rules predominate. MTB recommendations, for example, may be issued almost exclusively based on considerations of molecular genetic actionability (Hofmann & Esposito, 2025). But afterwards, they are fed back into the system of the medical clinic with classical utility considerations. MTBs only make recommendations. Making actionability a central category helped shape precision medicine into a highly innovative practice. But, at least for the moment, it can only be innovative against the background of common practice—a temporary solution to the integration problem lies in boundary-work.

Thus, the evolution of actionability shows how a conceptual category serves as a flexible device for continually reorganizing, rather than resolving, typical tensions in biomedicine (between research and care, data abundance and interpretability, standardization and personalization, innovation and coordination with existing structures)—tensions that result from technological and regulatory developments as well as from the conceptual reorganizations themselves.