Beyond “Trapped Pets” and “Red Buttons”: Bioinformatics as an Experimental Discipline

Computers have been used since the 1950s in various biological contexts, including molecular biology. Key pioneering figures include people such as Robert Ledley, Joshua Lederberg, and Walter Goad, who attempted to develop simulations and data modeling techniques to answer biological questions (November 2012; Stevens 2013; Strasser 2017). However, as documented by Stevens (2013) and Strasser (2017), many molecular biologists were initially reluctant to use computers, deeming them pointless and thus hindering early attempts at introducing computational projects in biology. But since the data deluge started in the early 1980s, computational assets such as databases started to attract the interest of molecular biologists, to the extent that a new discipline called “bioinformatics”¹ has slowly emerged. While the data management dimension has played a central role, the development of early computational projects based on software and tools to model biological data has nonetheless continued in parallel. However, historical investigations have focused especially on the data management and sequence analysis aspects of bioinformatics, and to our knowledge, a detailed history of other dimensions of bioinformatics has yet to be written.

The situation is slightly better in philosophy of science. Philosophical works engaging with the epistemology of computationally-driven disciplines such as genomics have only very recently started to discuss bioinformatics as a discipline per se, without implicitly assuming that it is just a set of tools and solutions to store and sometimes analyze data. In particular, recent works by Leonelli (2016, 2019), Strasser (2017), and Stevens (2013), while describing in great detail the epistemic ramifications of data management practices in the biological context, have also attempted to go a step beyond the view of bioinformatics as merely data management and automated analysis, by emphasizing the experimental dimension² and the proper goals of this discipline (see  Appendix 1 for an overview of these positions). By building on these attempts, the goal of this article is to develop a full-blown account of bioinformatics as an endeavor with its own epistemic goals. In our work, bioinformatics is understood as a discipline that, by engaging in experimental activities with virtual experimental systems (whose origin is nonetheless material) through the development of new computational³ tools, generates new kinds of data that wet-lab biologists cannot create. Our account coherently integrates the aspects described by Stevens, Leonelli, and Strasser, with novel facets of bioinformatics practices in the molecular biological context. By doing so, we shed new light on bioinformatics as a unique epistemic and experimental culture.

The motivations for the present work are two.

One motivation is to fill a gap in the philosophical, historical and sociological literature on the nature of bioinformatics. This has been often seen in relation to other biological disciplines, with only recent works trying to characterize bioinformatics in its own right (as discussed in  Appendix 1). By building on these attempts, this paper should be seen as contributing to the philosophical, historical, and sociological understanding of a discipline that has been often considered ancilla biologiae, rather than a proper subfield of biology.

The second motivation comes from an unfavorable situation in which many bioinformaticians work, which, to be improved, requires a richer epistemic account of bioinformatics practices. In particular, we refer to what has been called “the trapped bioinformatician,”⁴ or “the pet bioinformatician”⁵ syndrome. There is a tendency in laboratory-based groups to consider bioinformaticians as valuable resources to manage, curate, and analyze the data that “wet-lab” biologists generate, but not so much as pursuers of genuine biological questions themselves. This implies that bioinformaticians will have difficulties in completing their own projects, which often require bench biologists to reciprocate the time that bioinformatics practitioners have spent in analyzing data belonging to wet-lab projects. This culture reflects a view of bioinformatics practice as mere “red button-pushers” that initiate automated analysis procedures.⁶ This situation, we claim, generates a divide. On the one hand, wet-lab biologists think about bioinformatics mostly in instrumental terms—data management and analysis—while bioinformaticians feel that they generate genuine biological knowledge themselves. On the other hand, wet-lab biologists who generate large amounts of data in high throughput experiments often lack the expertise to analyze such datasets, thus having to rely on bioinformaticians for important steps, decisions, and biological interpretation of these experiments, but the nature of such decisions and its impact is underestimated. We call this divide “epistemic alienation”: bioinformaticians generate genuine biological knowledge, but they are excluded from the intellectual category of “knowledge makers”; at the same time, wet-lab biologists cannot make sense of important parts of their own work as they need bioinformaticians’ inputs to interpret their own experiments. The divide, which implies a subordination of bioinformaticians to wet-lab biologists, has been well documented by the massive, decade-long STS study of bioinformatics culture by Andrew Bartlett, Bart Penders, Jamie Lewis, and others (Lewis and Bartlett 2013; Bartlett et al. 2016; Lewis et al. 2016; Bartlett et al. 2017). One highlight of their studies is that “many view bioinformatics as a ‘service’, rather than a scientific field in its own right … [this] renders the intellectual contribution of bioinformaticians invisible, hidden in the ‘black-box’” (Bartlett et al. 2017, p. 2). One consequence of this view is a shady distribution of credits among wet-lab scientists and bioinformaticians. The results of this study are supported by an impressive variety of empirical evidence,⁷ corroborated by more insights (Markowetz 2017; Grabowski and Rappsilber 2019; Way et al. 2021), though with recent slight improvements (Calder et al. 2021).

Epistemic alienation has a strong sociological and political component. As Bartlett et al. (2017) say, science is necessarily tied to “institutional and organizational arrangements” (p. 2) which shape power dynamics. From this point of view, there is not much that we can do in the present article. However, what we can do is to dismantle philosophical prejudices lying at the roots of epistemic alienation. Therefore, we argue against the prejudices that bioinformaticians do mostly data management and that their work can be increasingly automated, and most importantly that they cannot produce novel biological knowledge by working on purely computational projects. An epistemic account of bioinformatics practice can show that there is more to this discipline than just data management and automated data analysis, and that bioinformatics is indeed an experimental science, as much as molecular biology is. The emphasis on “experimental” is essential, given the old illustrious theme in molecular biology (Strasser 2017) that those who generate genuine biological knowledge are the ones doing the experimental work.

The structure of the article is as follows. In Section 2 we identify the philosophical assumptions behind epistemic alienation and the idea that bioinformatics should be subordinated to wet-lab biologists. We introduce the concept of “epistemic driver,” which designates those scientific actors leading a research project and co-opting other people’s labor to achieve their own epistemic goals. We explain how, in biology, being an epistemic driver is strictly connected to experimentation, understood as a particular kind of material intervention aimed at creating new data types or new data that are indications of biological phenomena. But, a popular view claims, bioinformaticians do not do that: they only manage data and initiate automated procedures. What impedes bioinformaticians from acting as epistemic drivers is therefore twofold: they do not do experiments, and they do not have material access to phenomena. In Section 3, we delineate in detail the case of RNA velocity as a paradigmatic example of bioinformatics experimentation, by showing how the model and data type of RNA velocity is discovered by various formal interventions on data that have been “converted” from “real-world data.” In Section 4 we develop our own account of bioinformatics as an experimental practice by showing in which sense cases like RNA velocity are instances of biological experimentation, and by distinguishing two types of “experiments” in bioinformatics that we call “soft” and “hard.” Finally, in Section 5, we address the concern about the missing materiality of bioinformatics experimentation. All in all, this will show that bioinformaticians can be epistemic drivers.

Our starting point is the notion of an “epistemic driver.” We define an epistemic driver in a scientific group as an individual who, in leading a research project, produces scientific knowledge and co-opts other individuals’ expertise to achieve his or her own epistemic goals. An epistemic driver controls the unfolding of a scientific project. This is akin to making “path-dependent” decisions that ended up framing the general discovery strategy of a scientific project. Concretely, this means deciding the experiments to perform, how results should be interpreted, and how the efforts of other individuals should be allocated to achieve an epistemic goal that he or she chooses. Furthermore, this is also going to influence the “story” or the “narrative” that will be written in scientific articles.⁸ When we argue that epistemic drivers are co-opting other people’s work, we are not saying that they force other individuals to work on their behalf. As we will see, in biological labs there are different projects, and hence different epistemic drivers, and by offering one’s own services for another project, reciprocity is expected (Knorr-Cetina 1999). A realistic picture is that, within each laboratory, there is an intricate network of projects and hence of epistemic drivers. It is possible to zoom-out and identify groups of individuals that, in principle, can be epistemic drivers, and groups of individuals that cannot. For example, in a biological laboratory PhD students and postdoctoral researchers usually lead their own projects, and hence have their own epistemic goals, while technicians do not. This means that PhD students or postdoctoral researchers can become, at least in principle, epistemic drivers, while technicians cannot because they only provide a service to epistemic drivers. In this context, bioinformaticians have struggled to be recognized as epistemic drivers.

The concept of an epistemic driver is useful to describe general situations in experimental research, regardless of the field of study. Highly collaborative research groups—including consortia—will have different projects where an individual (or, more rarely, a group of individuals) has a specific research question, studies the existing literature to identify relevant gaps, designs and executes experiments, interprets the data and compiles these interpretations in a communicable form such as visualizations or reports.

In order to understand how bioinformaticians may be denied the role of epistemic drivers, it is important to describe exactly in which sense traditional bench biologists can be defined as such. Two aspects must be emphasized at the onset.

First, the figure of the epistemic driver can be investigated from two perspectives. On the one hand, there is a socio-cultural point of view, emphasizing the power dynamics leading some specific professional, academic, and scientific figures to become drivers rather than others. A second angle concerns the characteristics of epistemic drivers as such, in particular in the context of molecular biology or, to use Morange’s expression (2008), macromolecular biology, which includes disciplines developed from the molecular vision, such as systems biology, the various -omics, etc. We are interested in the latter angle, even though there might be much to say about the former.

Second, to understand the epistemic reasons for being epistemic drivers in macromolecular biology, we have also to consider (a) the environment in which biologists work, and (b) the conditions of possibility for discovering how biological phenomena are constituted.

Let us start with (a), namely the laboratory. An important ethnographic study investigating the epistemic dimension of macromolecular biological labs is Knorr-Cetina’s classic Epistemic Cultures (1999), which we use as a starting point. According to Knorr-Cetina, macromolecular biology is a discipline characterized by “object-oriented processing,” which is the continuous manipulation and production of material objects, such as plasmids, cell lines, etc., that are generated and used following protocols. The laboratory has a two-tier structure that is characterized by material objects. The first provides and maintains the materials necessary in a laboratory, while members of the second use the working material for experimental work in ways that are dictated by their epistemic goals.

We need to zoom-in to the second layer in order to grasp (b), namely the conditions of possibility for discovering how biological phenomena are constituted, to which only certain practitioners (i.e., the epistemic drivers) have access. In this layer, there are “massive transformations [brought] to bear on objects” (p. 85). In her rich description of the nature of object-oriented processing, Knorr-Cetina emphasizes the importance of the experiences, bodies, and senses of biologists. In order to be a good biologist, one has to develop sophisticated experimental skills, which means being able to tinker with experimental systems in efficient ways. The lives of biologists are characterized by “daily interactions with material things … the need to establish close relationships with the materials” (1999, p. 86). Good biologists have to develop a deep personal knowledge of their own experimental systems (Rheinberger 1997). This is because protocols have to be adapted to the specificities of the materials biologists are working with, requiring an ability to ‘feel’ the experimental system (Keller 1983) in a way that protocols have to “be negotiated in practice with obdurate materials and living things” (Knorr-Cetina 1999, p. 88). In this context, a necessary condition for being an epistemic driver is having access to experimental systems and being able to manipulate them. In other words, the concreteness and materiality of experiments seem to play a central role.

Let us start with “experiment” and “experimental.” It is beyond the scope of this article to provide a precise account of “experiments”—the topic and the literature would require a separate book-length treatise. What we will do here is highlight a few aspects associated with experiments that are important in this context.

There is a general way of understanding the term “experimental” as designating a “broad range of research practices, including both experimentation intended to control and experimentation intended to analyze” (Strasser 2017, p. 14). Here we especially emphasize the aspect of “intervention/manipulation” in experimentation by focusing on aspects of laboratory science that “interfere with the course of that aspect of nature that is under study” (Hacking 1992, p. 33). Parker (2009) characterizes experiments as investigative activities involving intervention “on a system in order to see how properties of interest of the system change,” where an intervention is “an action intended to put a system into a particular state” (p. 487). Knorr-Cetina captures this specificity in biology, noticing that many experiments subject “specimens to procedural manipulations … experiments deploy and implement a technology of intervention” (1999, pp. 36–37). Rheinberger also emphasizes the intervention/manipulation dimension of experimentalists, by stressing that experimentalists (in his case, molecular biologists) are “tinkerers” rather than engineers (1997, p. 32). But just tinkering is not enough. Tinkering with biological systems is a necessary aspect of experimental activities, but one can tinker in non-experimental ways. Experimenting is (a) tinkering with a system’s parts in a controlled setting, (b) recording unforeseen consequences in order to understand the parts’ behavior, (c) with some biological questions in mind. Tinkering without a question or interest in mind is not experimenting, nor is it experimenting to tinker in cases with known consequences (like repairing a system we understand perfectly) or without controls. The mention of controls is particularly salient, because experimenting is not mindless tinkering; it needs “confidence-building” strategies (Franklin 1986; Parker 2008), namely ways of checking whether the experimental activity is at least internally valid—this requires practices of calibration, consistency of results with known intervention or even with theory, robustness, etc.

Let us now turn to the “concreteness” or “material” component. Experiments defined in this way are central because the way biological phenomena are produced and/or maintained cannot be directly observed, and biologists have to find creative (though reliable) ways to force experimental systems to “reveal” something about biological phenomena. One central way in which this is done is by manipulating experimental systems in order to generate either novel data types or simply new data that can constitute evidence for phenomena (Bogen and Woodward 1988; Leonelli 2016). On this account, data are “the marks that some section of the world [i.e., in this case, specific biological phenomena] makes when it moves through some recording field” (Lowrie 2017, p. 9). Biologists manipulate experimental systems in ways that will force the phenomenon to leave new types of traces (especially if they want to discover something new) or just specific traces that they know are indicative of a specific phenomenon.⁹ In the tradition of macromolecular biology as depicted by Knorr-Cetina, these experimental activities have an important material dimension: new data types or specific traces are created by materially manipulating experimental systems. To take a common example, in order to study the biological phenomenon we call “genome,” researchers have to literally shear genomic DNA molecules into fragments using enzymes (restriction endonucleases), insert these fragments into circular DNA molecules that can be amplified by bacteria (plasmids), and insert plasmids into specialized bacteria strains (transformation). These are subjected to several amplification processes, including one that emits a specific fluorescent signal for each of the A, C, T, G nucleotides, thus resulting in a sequence of light signals that are detected by a machine, and converted to sequences by an image recognition software. Genomes as biological phenomena are thus (as Rheinberger would say) brought to light by tinkering with biological systems in a way that certain new types of marks/signals/data indicating characteristics of genomes are created. Without material engagement, knowledge cannot be created.

To sum up, there are some epistemic requirements for being an epistemic driver in macromolecular biology. In particular, one has to be able to engage in experimental activities in the way defined above, which means being able to generate new data types (or at least data we know are indications of phenomena) by means of intervening (in the way defined above) materially on experimental systems. Take a fictional, though realistic example. Consider a laboratory where Alice, a postdoctoral wet-lab researcher, is carrying out a research project based on an idea that she has discussed with her supervisor. Alice develops a sense of ownership of the project: she studies, prioritizes experimental work, designs individual experiments, and interprets data either on her own or in a discussion with other colleagues, including her supervisor. Her material work and her choices shape the narrative of the project in that they represent a logical and biologically motivated ordering of steps, connected by deductive and inductive activity. She is the one who, beyond taking these steps, is tracing them and choosing a path forward with more or less support and guidance from her supervisor. Alice is, in brief, driving her project.

There is a sense in which bioinformaticians are not epistemic drivers, which is when they provide support for projects of wet-lab biologists. This includes tasks like aligning reads to an annotated genome, performing quality controls, performing hypothesis testing using statistical methods, etc. The computational biologist can merely act as support to help a macromolecular biologist reach their own epistemic goal, e.g., knowing the transcriptional response to the knock-out of a particular transcription factor. This is not something specific to bioinformaticians. Indeed, it is typical of wet-lab biologists as well, for instance when providing orthogonal validation, i.e., an attempt at confirming a result using different molecular techniques or perturbing a system in a different way.

But it is possible to deny, on epistemic grounds, the status of epistemic drivers to bioinformaticians qua bioinformaticians. This is reflected in the view that data management and automated data analysis are all that bioinformatics can possibly offer. More explicitly, this can be expressed by saying that bioinformaticians (1) do not do experiments, and (2) do not have access to the material world of biological phenomena. 1 and 2 are indeed strictly connected: in order to have access to biological phenomena, you need to have material access to them, and in order to do this, experiments are required. To put it differently, if the epistemic goals of macromolecular projects (i.e., inferring the configurations of biological phenomena by collecting novel data or generating new types of data) can only be achieved by having material access to those phenomena, and in order to have this you need to engage in direct experimental activities (in the way defined above), then a bioinformatician is cut out by definition (or at least, under the conception that bioinformatics is only about data management). Because of the lack of material interaction with experimental systems and the inability to do any tinkering (in the way defined in Section 2.1.1), bioinformaticians cannot generate especially new data types. In other words, bioinformaticians cannot even in principle discover how biological phenomena are constituted, and hence they cannot be epistemic drivers. This epistemic preconception is prior to obstacles related to the social structure of biology that make it hard for bioinformaticians to be epistemic drivers—before even discussing the latter, the epistemic matter has to be addressed.

These considerations are compatible with the evidence gathered by STS studies mentioned in the introduction (e.g., Lewis and Bartlett 2013; Lewis et al. 2016). The subordination of bioinformaticians to bench biologists can be summarized by saying that “bioinformaticians do not perform experiments … [T]heir practice involves the manipulation of the primary inscriptions produced by biologists, rather than the transformation of the natural world through inscriptions” (2013, p. 249). This suggests that materially and directly “transforming” the natural world through a system of experiments is seen as a necessary condition to be an epistemic driver in macromolecular biology in the first place. The lack of experiments challenges the possibility for computational biologists to be epistemic drivers, and to have their own epistemic goals like wet-lab biologists do.

In summary, there is a view according to which bioinformaticians cannot be, in principle, epistemic drivers, because of their inability to engage materially in experimental activities (as defined above) to construct new types of data that can become evidence for answering biological questions.

In the previous section, we have reconstructed the view that bioinformaticians cannot be epistemic drivers. This is based on the assumption that bioinformaticians do not engage in material experimental activities, and what they do is just data management and superficial analyses. The emphasis on the material and the experimental is motivated by the particular context of this article: macromolecular biology and the prominent role that experimenting has played in it.¹⁰ To counteract this view, we need to rebut both the charge against the lack of “experimental activities” and the charge against the lack of “material engagement.” In order to argue for these things, we will use a specific case study, namely RNA velocity (La Manno et al. 2018). This is an important case for a variety of reasons. First, the computational model of RNA velocity as revealing stable features of a genuine biological phenomenon stems from unique properties of gene expression rather than being the rote application of a model borrowed from other disciplines. Second, the single steps taken by the investigators who created RNA velocity amount to a form of intervention (as defined in Section 2) in an experimental fashion. Third, this experimental intervention happened in silico, but traces of materiality can indeed be found, showing that bioinformatics is not as detached from the material as wet-lab biologists seem to think. Fourth, the outputs of RNA velocity go beyond a simple analytical application of statistical models but actually bring forth a new kind of biological datum.

This is how we proceed. In Sections 3.1 and 3.2, we illustrate the main aspects of this case study. In Section 3.3, we introduce some preliminary considerations as to how RNA velocity is a case of bioinformatics experimentation. This is before elaborating a full account of bioinformatics experimentation (Section 4) that is also sensitive to the materiality concern (Section 5).

To discuss RNA velocity, we need to briefly look at the dynamics of gene expression. For every given transcript in the majority of cell types, the temporal sequence of events (transcription, splicing, modification, export, translation, degradation) is completed in a matter of hours, with transcription and splicing being the longest processes. The different rates at which each of these events happen depend on many biophysical parameters that are both influenced by a cell’s current state (e.g., in terms of pH, temperature, concentrations of ions, type and amount of proteins, etc.) and by locus- and transcript-specific features (e.g., sequence, length, subcellular localization, etc.). Greatly simplifying, the relative importance of each of these steps can be classified as follows: if a cell transcribes a gene G at a transcription rate that exceeds the degradation rate, then this gene is being up-regulated. If the gene G is produced at a rate that matches the degradation rate, it is at a steady state, as its amount does not change in time. Finally, if G is degraded at a faster pace than it is produced, it is being down-regulated.

Given this picture, some researchers (see Zeisel et al. 2011; Gray et al. 2014; Gaidatzis et al. 2015) had an intuition: if intronic RNA (i.e., the abundance of unspliced, immature RNA) and exonic RNA (i.e., the abundance of mature, already-spliced RNA) could be measured separately for every single transcript that is being made by the cell at a given time, then it is possible to infer how new a transcript is, using the ratio between spliced (“older”) and unspliced (“newer”) transcript as a proxy for their relative age, and its degradation dynamics. This intuition is based on the knowledge accumulated by macromolecular biology on these phenomena. If the relationship between unspliced and spliced transcripts holds and reveals a temporal trend, it should be possible to 1) model the relationship over time by observing its change across samples taken at different time points and therefore 2) predict the amount of a spliced transcript at a future time point. Given sequencing results at different time points, quantities for spliced and unspliced transcripts can be plugged in a model of gene expression that makes use of ordinary differential equations to describe the relationship between rates of transcription, splicing, and degradation.

RNA velocity is a computational model expressing the relationship over time between unspliced and spliced transcripts. The relation is expressed in such a way that the model can predict the amount of a spliced transcript at a given time. The phenomenon that RNA velocity models is, more precisely, the trajectory of the gene expression state of a cell.

It is important to be more precise on how RNA velocity is related to data, phenomena, and theory, and in which sense RNA velocity creates a new kind of datum. We can understand the relation between RNA velocity, data, and phenomena by considering how these fit into a widely known account, such as Bogen and Woodward’s famous view (1988). The phenomenon here is the set of dynamics governing gene expression in a cell, which is a process characterized by stable features that can be identified across different experimental contexts. As a process, gene expression is explained by a number of well-characterized mechanistic models (that, together, constitutes the theory of molecular biology, see Ratti 2020). The trajectory of gene expression is one aspect of the general phenomenon of gene expression. The way this trajectory is represented in the model of RNA velocity is influenced by those mechanistic models. By using RNA velocity, a new type of datum is created, namely data about the trajectory of gene expressions. The idea is that RNA velocity models data on transcripts. One might be tempted to think of this data as raw, but data on transcript is nonetheless a “data model,” at least in the sense of “corrected, rectified, regimented, and in many instances idealized version of the data” (Frigg and Hartmann 2020), that we gain from certain experimental procedures. By modeling these “data models” on transcripts, RNA velocity generates a new kind of data model that provides evidence for a specific aspect of the biological phenomenon of gene expression (that is, its trajectory).

Now that the general framing is clear, let us consider the nature of RNA velocity in depth.

In order to characterize more precisely the biological phenomenon captured by RNA velocity, bioinformaticians have created a new data type by experimenting computationally, rather than materially. Here we describe the steps of this experimental activity (La Manno et al. 2018).

Quantifications undergo total depth normalization, in which each read count for each gene in each cell is divided by the total amount of read counts in the cell. This yields comparable quantities u_i and s_i.

Assumption 2: La Manno and colleagues assume that the splicing rate β is the same for all transcripts, such that β = 1. While this is not exactly true, as splicing can be influenced by several factors, it is a necessary simplification to be able to use the instantaneous measurements (i.e., without temporal information) of u and s. This also means that all the other rates will be expressed as units of the splicing rate.

The goal of this model is to be able to model the “RNA velocity”, expressed as the first derivative of the amount of spliced transcript with respect to time, dS/dt. If the model holds, it becomes possible to extrapolate the amount of spliced transcript S at a (not too distant) time t even if the time is not observed. To be more precise, the model allows us to predict with reasonable levels of confidence the expression dynamics (spliced mRNA) of genes in the near future, thus indicating a direction of change for these genes. The “near future” is limited by the biophysical dynamics of transcription and splicing, i.e., these predictions hold for changes that happen in a few hours.

Assumption 3: at steady state there is no change in spliced transcript abundance; mathematically: ds/dt = 0.

Taken together, assumptions 2–3 result in γ = u/s and α = u. If these assumptions were compatible with the complexity and biological features of the phenomena of interest, extrapolating spliced transcript quantifications would be trivial. However, the steady state assumption (assumption 3) only holds for cells or tissues that do not undergo changes such as differentiation or response to a stimulus. RNA velocity becomes interesting only in a dynamic picture, such as a developmental process; in fact, it would allow researchers to extrapolate “future states” of gene expression based on the data currently available. Moreover, the production rate α is unknown and difficult to measure without specialized experiments. For these reasons, La Manno and colleagues drop assumption 3 and a constant value of α from assumption 1, and make another set of two alternative models, each with its own assumptions.

The dynamics of each specific transcript can be represented as the progression of the combination of spliced and unspliced quantities, i.e., different solutions to a system of differential equations. The geometrical representation of these solutions constitutes a phase portrait (Figure 1A). More precisely, for every acceptable pair of u_i and s_i values—that is, for every pair that can represent a solution to these equations—there is a point in space; connecting these points along their variation in time creates the phase portrait. This representation is useful to understand the relationship between the parameters, the quantities, and their progression in time. According to equation (11), equilibrium points (i.e., points where velocity is 0 are reached where γ is equal to u/s, or where both u and s are 0, or when α = u. Then, the fit of a regression line going through the diagonal of the phase portrait represents the steady state approximation for γ (Figure 1A); in other words, a simple linear regression coefficient will be accurate if and only if all samples are at the equilibrium points of the phase portrait. However, as discussed previously, samples/cells undergoing differentiation or responding dynamically to a stimulus will be populating many other parts of the phase portrait, making a linear fit on their s and u values severely biased.

Accordingly, the RNA velocity authors use an “extreme quantile fit”: rather than trying to calculate the coefficient using all points in the phase portrait, they only consider points that lie at the extreme of their distribution (Figure 1B), thus getting closer to the steady state assuming degradation rates do not change along the trajectory.

This procedure, however, only works well when the extreme quantiles are close to the steady state. There may be genes that are up-regulated late or down-regulated early, so that we do not observe their steady state in the time period sampled in the experiment; in this case, their extreme quantiles will lie in the middle of the phase portrait, meaning the fitted γ will be still biased (Figure 1C, D). For these cases, the authors developed yet another model termed “structural fit,” which accounts for the number of exons, length of introns, and number of internal priming sites that can be captured by the sequencing technology.

Building these models require an understanding of the unique properties of transcripts and of the sequencing procedure. Thus we can see how so far the development of RNA velocity as an approach requires a high degree of data analytics (data processing, transformation, modelling with different mathematical approaches) but also of software development (coding all these implementations in an efficient and usable way for other practitioners): La Manno et al. had to write code to perform data preprocessing, phase portrait estimation, model fitting, estimation of the velocity vectors and their visualization, in a programming interface that can be applied to commonly used data representations for single cell RNA-sequencing.

Indeed, it can be useful to pause at this point to reflect on how the calculations undertaken under different assumptions, resulting in two alternative models, and the different attempts at fitting linear models to derive degradation rates amount to experimental tinkering. While the intuition of modeling gene expression using simple differential equations is far from new, there are a few important novel aspects in this approach.

The first is the concept of RNA velocity itself, which finds an important application in the field of single cell transcriptomics—an important intuition by Sten Linnarsson and Peter Kharchenko. As it is widely known, the innovation of single-cell transcriptomics lies in its ability to capture a large number of individual cells within a tissue/organ, as opposed to the “bulk” sequencing of the transcriptome of a tissue/organ. Therefore, analyzing a population of cells that is undergoing a transition in an un-synchronized fashion, such as a developmental process, means collecting a snapshot comprising different phases of the process itself, within a reasonable time frame. Conversely, in a “bulk” setting where the transcriptome of each cell is mixed, there is only one such average phase per sample. It becomes evident that population-level temporal dynamics (often called “pseudo-temporal”) can be inferred in a single cell dataset by virtue of these cells being, individually, at different stages of their progression along a specific biological trajectory—which is described by RNA velocity.

The second important innovation by the authors of the original RNA velocity paper lies in the techniques for the visualization of velocity. The opportunity to derive cell-specific velocity vectors within a dataset at single cell resolution brings forth an additional layer of complexity to the canonical single cell data analysis outputs. Originally, the visualization of single cell data posed an important challenge: if each cell is embedded in a space according to its gene expression values, meaning every cell is represented by a point, the coordinates of this point will be determined by the numeric expression values of n genes: points will exist in an n-dimensional space. For this reason, dimensionality reduction techniques have been leveraged to reduce complexity while retaining meaningful relationships in a two- or three-dimensional representation: in other words, points (cells) that are close together in this visual space are supposed to be similar, while points that are far away are supposed to be different. Several techniques have been proposed, at different levels of granularity: t-stochastic neighbor embedding (t-SNE, van der Maaten and Hinton 2008), uniform manifold approximation projection (UMAP, McInnes et al. 2018), partition-based graph abstraction (PAGA, Wolf et al. 2019), similarity weighted nonnegative embedding (SWNE, Wu et al. 2018), diffusion pseudotime (Haghverdi et al. 2016), to name a few. And, as researchers routinely discover, visualization techniques may be biased, imprecise, or contain assumptions that are at odds with what we know about the biological systems they are meant to represent, giving way to new, improved visualizations that should be “more faithful” to the underlying biology. Visualization of high dimensionality data is an active field of experimentation in computational biology (and machine learning in general) and, as every experimental field, it proposes partial solutions with advantages and pitfalls (see Chari and Pachter 2023 for the case of UMAP). These representations play a central role in the analysis of single cell data, as they are not only ways of summarizing an analysis output, but they are also de facto data models that are used for discovery, inference, and validation of hypotheses—a point emphasized by Stevens in his ethnography of bioinformatics (2013). One of the most important outputs of the RNA velocity procedure can be considered an enhancement to these visualizations: a two-dimensional representation of the velocity vectors, pointing to the future state of single cells, within the “transcriptional space.” The authors of RNA velocity devised a technique to draw velocity arrows on top of two-dimensional visualizations that were previously created, either at the level of single cells, or as a “vector field” that shows a summary of local velocity at every point of the transcriptional space. Thus, by looking at a UMAP visualization of single cells and overlaying their velocity vector field, researchers can literally see whether a certain cell population is progressing towards another, thus inferring that a differentiation progress is taking place with a certain directionality and intensity.

What we think RNA velocity shows is that bioinformaticians engage in experimental activities, understood as investigative activities involving interventions that, by manipulating existing data, can even create new data types, exactly like traditional macromolecular biologists.

The estimation of RNA velocity vectors is non-trivial, and presents many challenges to the original authors of the method. They make use of several models, alternative ways to fit degradation coefficients, and simulations that test the extent to which their models hold given differences in gene expression levels, equation rates, and their temporal dynamics. There was no a priori guarantee that RNA velocity would represent a relevant biological phenomenon once single cell transcriptomic data was used and processed. Comparisons to real-world datasets with different levels of ground truth are included as a validation of their experimental procedure.

Taking these considerations together, it can be argued that 1) the quantification of transcriptional dynamics in single cell data does not require ad hoc experimental procedures, rather a repurposing of existing data; 2) the extrapolation of a cell’s future transcriptional state is not only a biophysically motivated ordering of cells along a trajectory, but also a measurement of an unobserved instantiation of such a trajectory; 3) extensive tinkering with different models and assumptions was required to arrive at a final, usable data model. But RNA velocity is being investigated also by other groups, in direct competition with the original picture. As often happens with the discovery and characterization of other new biological phenomena, the publication of the RNA velocity paper sparked many enthusiastic reactions, and the community quickly started building on top of the original models and results. In fact, several alternative versions of RNA velocity estimation were published, which made use of different assumptions, different models, different representations and had different software implementations. The corpus of experimental work on the field of RNA velocity is growing—estimating RNA velocity, as an experimental activity, has a life of its own.

A large and complex study (Gorin et al. 2022a) published a few years after the first RNA velocity paper aims at laying down more rigorous foundations for the method, which implies an inevitable critique of the original work and most of its derivatives. We will not go through the details of the study as it is a very exhaustive treatise of the biophysical foundations of the model, but we want to highlight what we perceive to be important contributions, both to the field per se and to our argument in particular: the RNA-velocity strategy generates a new type of biological datum that is evidence for a specific biological phenomenon worthy of studying in its own right. The Gorin et al. study highlights the great potential of RNA velocity approach(es), but at the same time sheds light on several issues, motivated by a computational experiment: the same dataset analyzed through two different RNA velocity implementations yields two very qualitatively different results (Soneson et al. 2021; Gorin et al. 2022a). The first issue is the definition of RNA velocity itself, which can be interpreted in seven different ways. The second issue concerns different processing pipelines which potentially render some of the assumptions invalid, in particular considering spliced and unspliced molecules two mutually exclusive species and thus over-simplifying the complexity of alternative splicing. Then, assumptions made by different implementations are also quite diverse. Additionally, they critique the visualization of RNA velocity itself, following previous work by the same authors in which they address the larger issue of whether a visualization through severe dimensionality reduction is properly representing a biological phenomenon or not. Gorin and Pachter (2023) go through a rigorous study of these assumptions performing other computational experiments (such as the application of RNA velocity estimation to a dataset with no differentiation or stimulus) and conclude that the current implementations of RNA velocity reduce the complexity of the quantities they are trying to model, are lacking in biophysically motivated foundations, require restrictive assumptions, make use of arbitrary parameters and as a consequence do not result in reliable estimations of a future cell state. Their critique of current implementations goes as far as questioning whether RNA velocity can be useful at all or whether something can be salvaged by asking the Biblical question: “is there no balm in Gilead?” In the last few years, the Pachter group has worked on more biophysically motivated models of transcriptional activity which show, by the application of different models and mathematical frameworks, how genes can be classified in different ways (Gorin et al. 2022b), and how a precise modelling of stochasticity in gene expression and its measurement is required to describe transcription mechanistically using single cell sequencing data (Gorin et al. 2023). Interestingly, in this article Gorin and colleagues explicitly mention tinkering with their virtual system by way of “manipulation of generating functions” (Gorin et al. 2023) in purely experimental ways.

To summarize, RNA velocity has been developed through computational experimentation, made available as an analysis tool, been experimented with and heavily refined (with some refutation of its assumptions and models) as a result of both additional experimentation and mathematical formalization, and has taken on a life of its own. We argue thus that through RNA velocity a new type of biological datum is constructed that does not stem from modifications in wet-lab experimental procedures, but rather from an elegant and complex in silico system of experiments. By creating a new type of biological datum that can provide evidence for a specific biological phenomenon (i.e., the trajectory of gene expression states), the computational work seems to achieve the same kind of result that material tinkering performed by wet-lab biologists can achieve.

In the previous section, we have reconstructed an example of bioinformatics practice where a new data type providing evidence for a specific biological phenomenon is created through various experimental activities done in silico. In this section, we describe these activities at a more general level, by constructing a comprehensive account of the dimensions of bioinformatics as an experimental discipline consisting of three dimensions (data management; analytics, development). This account builds on previous analyses and observations of bioinformatics, most notably (Stevens 2013; Leonelli 2016; Strasser 2017) and discussed in  Appendix 1. The facets of our account will be illustrated by referring back to the example of RNA velocity.

Our account of bioinformatics counts three dimensions.

First, there is data management. As mentioned earlier, this is a central aspect of bioinformatics practice, given the importance of databases. It includes those practices geared at creating, maintaining, interfacing with and creating connections among biological databases in virtual spaces. Data management thus consists of creating standard formats, an easily accessible and navigable infrastructure, secure storage and updated records; from an end user perspective, it is the management of laboratory archives with special regard for high throughput data, making sure that datasets are properly stored and shared together with their metadata, and ensuring the reproducibility of the raw data processing steps.¹¹ In the case of RNA velocity, the practices associated with the management of single cell transcriptomics (which have been built on the foundations of the transcriptomics data management ecosystem) are instrumental in creating reproducible analyses, such as the storage and distribution of spliced and unspliced count matrices.

A second dimension of bioinformatics is analytics, which is the application of more or less established statistical models and computational procedures to gain a first level of biological interpretation of a given experimental outcome. Analytics include processing raw data into quantities of features of interests (such as genes, proteins, chromatin regions, etc.); applying quality control procedures to distinguish signal from noise, rule out technical artifacts, and remove systematic biases; applying mathematical frameworks to identify patterns and score relevant differences between experimental conditions, together with a measure of their uncertainty; visualizing results in a clear and informative way; etc. These aspects have been also emphasized by Stevens (2013) in his analysis of the epistemic roles of data visualization, and more recently by Leonelli (2019) in her ethnography of the SureRoot project. What these—and other examples—show is that modern analytics consists of different steps that can be combined together in different and novel ways; far from being “automated,” analytics allows the creation of analytical pipelines with varying degrees of flexibility. In the case of RNA velocity, analytics is a critical aspect, as it is a specific combination of several processing and mathematical modeling steps that creates a new data type, and has become, after its development, a rather standardized step in several single cell transcriptomics analysis pipelines.

Finally, development consists in the invention and programming of new mathematical and statistical frameworks, or the optimization of previously available frameworks, with a specific type of biological question in mind. A bioinformatician involved in development aims at writing software that tackles extant challenges in the generation and interpretation of results, such as identifying and implementing the use of the correct statistical distribution for a certain type of datum; integrating different data modalities to enhance the discovery of biologically relevant phenomena; using “first level” analytical results to predict more complex behaviors of a biological system, or non-trivial ways in which this system can be modified, and so forth. Writing software not only entails the theoretical exercise of finding the most appropriate models, operations or representations for the data, but also the practical aspect: implementing the algorithms, optimizing them, creating a usable interface and its documentation. A few examples of these new frameworks and implementations will be discussed in Section 4.2. There is empirical evidence suggesting that development is experiencing an exponential growth: even just considering the short time frame between 2017 and the first quarter of 2024, and limiting ourselves to the field of single cell biology, the number of bioinformatics tools has quickly surpassed 1700 units (Figure 2A; Zappia et al. 2018). Similarly, when querying the number of R packages distributed from the Bioconductor project’s first release in 2002, we observe a similar trend (Figure 2B). In the case of RNA velocity, the implementation and optimization of the processing and modeling steps, together with the creation of a user-friendly software interface (and several other improvements and iterations from other computational biology groups) that we have described in the previous section is a classic example of development.

These three aspects (i.e., management; analytics; development) are fundamental integrated ingredients of bioinformatics as a discipline. As such, they often complement each other and coexist within many declinations of bioinformatics practice, as Strasser, Leonelli, and Stevens have noted, even though without using the terminology employed here.

For instance, management can imply development, as bioinformaticians who want to distribute their software to a large community should, at a minimum, provide thoroughly tested software that has as few problems as possible, make it easily accessible and easily findable through metadata, provide clear documentation and instructions on how to use it, and—if the software is released as open source, which is in most cases—store the source code in repositories that allow version control. In order to make this collective endeavor easier and standardized, bioinformatics developers started projects such as Bioconductor (Gentleman et al. 2004; Huber et al. 2015), which hosts more than 1250 packages for biological data analysis and is maintained by the community on a volunteer basis¹².

Analytics and data management are intertwined as well, especially regarding the reproducibility of data analysis. Whereas wet-lab experimental procedures are usually succinctly described in the methods section of a paper, leaving their complete description to protocols (commercially or academically published), the gold standard for analytics reporting is to provide other researchers with the exact steps, i.e., the code used for the analysis, together with the expected outputs and the necessary inputs. This, in turn, requires bioinformaticians to provide their colleagues with access to the same software they used, as some outputs could depend heavily on the software version that was used. Thus, analysts need to manage representations of their workflows—aptly named “notebooks”—and digital snapshots of the software they used—“images” or “containers”—to ensure that their results are reproducible by anyone with sufficient skills.

But the most interesting integration is the one between analytics and development. On the one hand, developers usually master some facets of the analysis toolkit, even just to be able to benchmark the results of their newest algorithm against gold standard applications, or to generate data representations upstream or downstream of their inventions. On the other hand, analysts can combine different tools crafting pipelines which, at some levels of complexity, can be considered akin to development. In fact, analysis tools are often developed with the Unix philosophy in mind: programs should do one thing, and do it well (McIlroy et al. 1978). This translates to a highly modular analysis workflow in which every step can be carried out by several alternative approaches and/or tools. An expert analyst combines these tools and, in many instances, refines their input writing code that can be in the same language as the one of the tools they use. In some cases, entire analysis workflows can be packaged as single one-stop solutions, showing how blurry the line between analytics and development can be.

We have mentioned throughout this article that bioinformatics should be considered an experimental science, as much as macromolecular biology. But how exactly? We claim that the integration of analytics and development plays a central role.

First, let us draw a parallel between the ‘wet-lab’ biology pipeline and the computational one (Figure 3). In the case of macromolecular biology, the starting point is a question about how biological phenomena are produced and/or maintained, as these are often opaque. A material experimental system (which is taken to materially embed the biological phenomenon a biologist wants to explain) is perturbed or tinkered with certain inputs (e.g., reagents, conditions). The results of this tinkering are output data that are taken to be evidence for certain claims or mechanistic models (Craver and Darden 2013) about the biological phenomenon under scrutiny. As noticed in Section 2, this is not mindless tinkering. In fact, conditions of calibration, robustness, internal/external consistency, and specific biological questions will constrain the tinkering. In the case of bioinformatics, an analogous dynamic is at play. Consider the case of RNA velocity. The material experimental system is converted into a virtual system—by digitizing some of its features and embedding them in an appropriate numerical representation—but it is tinkered with anyway. This is not necessarily a specific aspect of virtualization. In fact, any experiment consists in isolating specific, robust signals from a (biological) system and recording them through an apparatus. However, recording a high amount of observations in parallel and embedding them into a virtual system—what biologists call ‘high throughput’ technologies—allows to retain several “hidden” relationships between data points or features for which we do not necessarily know the data generating process. In this virtual system it is possible to do new work that is experimental—though not material—to identify new relationships, new signals, new secondary inscriptions through new techniques. These novel biological insights are particularly interesting because they retain a “semi-material” property given the material origin of the virtual system, and their reliability is almost always confirmed by comparing them to a material experiment. Bioinformaticians interact with such a system by modifying inputs and/or rules, and monitoring changes in outputs. The way bioinformaticians interact with such systems makes ample use of the ‘confidence-building’ strategies exemplifying good experimental activities in material systems as emphasized in Section 2, including calibration, robustness analysis, and checking for consistency. Following these strategies reflects the notion of experimenting that we have formulated previously: bioinformatics is not just tinkering; rather, it involves controlled settings (which allow monitoring), entertaining biological hypotheses, and recording unforeseen consequences and effects. For instance, bioinformaticians do not know a priori what kind of statistical distribution or differential equation best fits the description of a particular biological phenomenon; indeed, they always clarify certain assumptions that make their models tractable, and test a set of proposed models with respect to “ground truth” data. However, ground truth is not always, if at all, attainable in biology, forcing researchers to use reasonable approximations or resort to orthogonal validation. Thus, bioinformaticians deal with data models that have potentially surprising behaviors due not only to their non-deterministic nature, but also to unobservable variables. As an example, linked to the case of RNA velocity, consider the cellular composition of a tissue as inferred through single cell RNA sequencing and unsupervised clustering. Most currently utilized clustering algorithms do not require the user to specify how many clusters are expected, nor their size, or their degree of separation. A clustering algorithm that operates in ideal conditions should be able to partition the data in biologically relevant (and, possibly, experimentally separable through other means) units such as “cell types” or “cell states.” A surprising behavior thus would be the identification of a previously unobserved (and unexpected) intermediate state between two well characterized differentiated cell types, or the existence of a continuum bridging what were previously thought to be isolated, highly stable transcriptional profiles. More precisely, what is surprising here is the fact that a specific novel clustering algorithm may or may not reveal “new” biologically plausible aspects of the virtual system that other algorithms could not reveal before. Tinkering on the data with different clustering techniques, or inventing an entirely new clustering technique, is what we refer to as experimental in this setting. Several approaches can be used to estimate or infer the unobservable variables of interest, which can then be correlated with the biological nature (e.g., whether these variables overlap with a disease or mutant state compared to a healthy/wild type control). Changes in algorithms, models, parameters all amount to different experimental procedure on the same data model, resulting in different outcomes. The evolution of bioinformatics tools—through analytics and development—shows that indeed our understanding of the same data can be furthered by testing and refining procedures. For instance, in the context of RNA-sequencing and differential expression analysis several authors over the years have suggested different statistical distributions and models to deal with read count data (Marioni et al. 2008; Anders and Huber 2010; Robinson et al. 2010; Trapnell et al. 2010; Law et al. 2014): using t-tests, linear models, generalized linear models for Poisson-distributed data, Negative Binomial GLM, etc. Eventually, after several experiments and benchmarking studies (Robles et al. 2012; Soneson and Delorenzi 2013; Germain et al. 2016), the field appears to largely favor the use of linear models and/or negative binomial generalized linear models with variance shrinkage, although it has been argued that other methods are more precise and reliable in particular settings (Li et al. 2022). This type of tinkering is done by integrating and modifying analytical tools and developing new software or computational infrastructures that can host the right set of tools. These are all forms of experimental activities.

We can be even more precise and distinguish between soft and hard forms of experimentation. Soft experiments in bioinformatics are the attempts at using new (or old, but refined) approaches and techniques to gain deeper understanding of data: existing approaches are being repurposed or extended to deal with biological data beyond their original scope, with no a priori guarantees regarding their reliability, robustness, fidelity to the natural process or interpretability. It is tinkering in a controlled setting by recording unforeseen consequences. Bioinformaticians operating on data models by applying analytical steps that create new representations of the model and new data types is the hard experimental nature of bioinformatics.¹³ In the soft experimental framework, the novelty lies in the use of the tool, and not in the tool itself. Conversely, in the hard experimental framework, the novelty lies in the tool itself and in the new type of data generated that can constitute indications for claims about biological phenomena. Regardless of whether an approach is soft or hard, it should be considered experimental by virtue of the tinkering, possibility of unexpected results, internal cohesiveness of the digitalized experimental system, and biological focus.

There are several examples in the history of bioinformatics that can be described using our framework of hard and soft experimentation. Take for instance the creation of differential expression tools such as DESeq2 (Love et al. 2014). This is a soft experimental activity: the use of generalized linear models and Bayesian variance shrinkage greatly predates RNA sequencing, but its application and successful implementation required tinkering and innovation. The resulting data, i.e., fold changes (effect sizes) and corresponding statistical significance values, do not belong to novel data types as they are basically the same type of result one would construct from other tests in which group means are compared, with or without computational tools (e.g., in the case of qPCR). This is why we can consider the invention of these tools as a case of soft experimentation.

The case of Gene Set Enrichment Analysis (Subramanian et al. 2005), instead, constitutes a hard experiment (as it is the case of RNA velocity): ranked gene lists were tinkered with in ways that created a new data type, the Enrichment Score, a numeric value whose sign and magnitude is indicative of the activity of a pathway in the comparison of global gene expression programs across conditions. The Enrichment Score is also constructed by comparing observed data to an empirical null distribution built by random permutation of rankings, a common approach in statistical testing that is akin to constructing a virtual negative control.

Enumerating which bioinformatics tools constitute hard or soft experimental approaches is beyond the scope of this article, but we supply a table with a small subset of examples spanning the domain of transcriptomics and other high-dimensional genomics data analysis tasks.

By conjuring the possibility of hard and soft experiments in bioinformatics, we want to argue that bioinformaticians can indeed be epistemic drivers, because they generate new data or even new types of data that can potentially constitute new biological knowledge, and they do this by experimenting in an analogous way to how macromolecular biologists do.

However, remember that the argument against bioinformaticians being epistemic drivers was not only about experiments; materiality was also involved. Maybe bioinformaticians do experiments; but given that they do not materially manipulate and generate data (or, to use Lewis and Bartlett’s conceptual apparatus, they do not generate primary inscriptions), then they cannot be epistemic drivers. By relying again on the case study of RNA velocity, our response to the materiality concern is twofold. First, we show that the importance assigned to materiality is misleading. Second, even if materiality was indeed that important, there is still space for something that we call semi-materiality, which basically applies to what biologists do, at all levels, be they wet-lab biologists or in silico biologists.

Let us start by showing how misleading the idea of materiality can be. Intuitively, materiality is deemed important because it provides more direct access to biological phenomena. Given that bioinformatics lack this direct access, then they do not have the same grasp of biological phenomena that wet-lab biologists have. However, it is just not the case that it is in virtue of materiality that we have a more or less mediated access. In fact, the case of RNA velocity shows that, while we need a preliminary “material origin,” virtualizing the data (or making them “semi-material,” as explained below) is what provides us a better access to the phenomenon itself—just with traditional material access, the phenomenon captured by RNA velocity is inaccessible. Moreover, consider the variety of experimental systems that biologists use: in vivo, in vitro, animal models, etc.; these are often only proxies for various biological phenomena, and hence one may say that the access to phenomena is nonetheless mediated, and materiality plays no substantial role in making these systems more inferentially reliable.

But let’s say that materiality is indeed important (even if the importance is vague). How should we address this? Our response is that we are not advocating for an exclusively in-silico knowledge generation process: the origin of the data is always material, unlike in some cases of computer simulations. This aspect is not appreciated enough. We can draw a parallel to so-called virtual experiments, namely nonmaterial experiments on semi-material objects. Morgan (2003) describes computational experimental activities to investigate the strength of bones. Given the challenges of assessing strength in material settings, one line of investigation was to convert a real cow hipbone into a computerized image—cutting bones into thin slices, taking specific pictures of them, re-assemble these in high-quality 3-d computerized images, and then intervening on them by means of various models. Morgan emphasizes how this process “retains a high degree of verisimilitude of structure for each particular bone sample” (p. 223). By conserving important structural features of bones in the process of recording and converting, those computerized images have a semi-material status: intervening on the 3-d images is de facto an experimental activity, where mathematical models are used as experimental instruments. Virtual experimental systems are akin to semi-material objects: biological features are recorded, then converted, but nonetheless conserved. In the case of RNA velocity, we have seen that certain physical aspects of data sets are conserved in the virtual experimental systems, such as the differences in spliced and un-spliced transcript abundances in single cells.

The material origin of data has other consequences too. Given that aspects of materiality are conserved, there is the risk of importing factors that may be confounding—to paraphrase Morgan, bioinformaticians have to consider “all conditions and factors that are likely to interfere with the process of interest” (Morgan 2003, p. 219), exactly as traditional experimenters. In the case of RNA velocity, these confounding factors are the stochastic aspects of transcript quantification given by the material process of transcript capture and reverse transcription of a very low input; the presence of dying/stressed cells whose gene expression does not represent a physiologically relevant cellular state; the lack of complete knowledge of the structure of transcripts and their spliced forms. Moreover, the material origin comes with the possibility of discovering something hidden in the data that, for various reasons, could not be separated materially. This is noteworthy in contemporary biology: in cases like high-throughput recordings, an impressive number of observations are recorded, and these observations have hidden relations for which data generation processes are unknown. Virtualizing such high-throughput experimental systems means recording (by means of conversion) these “hidden relations” on a virtual experimental system, and transforming signals to create new data or new data types, as we have shown. Tinkering with multi-dimensional data within a controlled virtual experimental system means being able to separate signals that, in the normal material laboratory setup, would be simply impossible to distinguish. As much as in normal laboratory conditions results are produced by intervening on a (material) system, here results (e.g., new data or new data types) are produced by intervening on the (virtual/semi-material) experimental systems, unlike typical cases of modeling where one derives results just by means of mathematics (Morgan 2003). And it is in virtue of the fact that something is conserved in the transition from material to virtual that inferences based on computational experimental activities can be, at least in principle, reliable. Of course, this does not mean that reliability is established purely in silico—in fact, orthogonal and functional validation from wet-lab biologists is still required (but this is true of any paradigm, even in the wet one).

But one can push the materiality argument further, in somewhat unreasonable ways. One way to do this is by appealing to the intuitive distinction between primary and secondary inscriptions (Lewis and Bartlett 2013). The idea is that bioinformatics data might have the semi-material dimension we have argued for, but they are still “secondary inscriptions”—by not being able to generate “primary inscriptions,” bioinformaticians cannot in principle meet the epistemic desiderata for being drivers. This is a strong claim, likely to end any discussion. However, the distinction between primary and secondary inscriptions (Lewis and Bartlett 2013) is misleading. For instance, in sequencing a sample, which data are considered primary inscriptions? Is the sample itself? But the sample per se does not constitute a datum, and in order to become data (e.g., a sequencing read) it is manipulated by various technicians, including computational technicians (Stevens 2013). Therefore, the primary inscription is a sample manipulated to become sequencing data. This shows that primary inscriptions are both material and in silico at the same time, and that they are co-produced both by traditional biologists and bioinformaticians. This means that there is not really any substance to the “primary vs secondary” distinction: all data is likely to be semi-material. But if this is the case, then in principle there is no difference between the data used by bioinformaticians, and data used by wet-lab biologists, at least from this perspective. In conclusion, this shows that even the second concern (the materiality concern) has not any robust substance, and hence there is in principle no reason why bioinformaticians cannot be epistemic drivers.

The fundamental role played by computational biology in most life sciences projects has grown at a quick pace, so much that international consortia such as the Human Cell Atlas (Regev et al. 2017) require an effort in coordinating, developing, testing and communicating computational methods for data storage, analysis and visualization that is far beyond the—still impressive—work required to generate all the single cell atlases. In increasingly more cases computational biologists use the generation of an atlas as a good testing ground for a new computational method (e.g., Stephenson et al. 2021), or they drive highly complex analysis efforts to establish best practices in the field with no additional data generation required (e.g., Luecken et al. 2022). These computational scientists do control the narrative of their projects and are fully equipped by their environment to be epistemic drivers: this is, we claim, bioinformatics functioning as a proper discipline, rather than as just support for wet-lab biologists. This article is only a first step towards a comprehensive characterization—both philosophical, historical, and institutional—of bioinformatics.

To conclude this piece and introduce future works, we formulate in the remaining space one open question that we did not address in depth for the sake of brevity, but still deserve a mention in our conclusions, and further elaboration in its own merit.

We have motivated the need for a complete account of bioinformatics practice by mentioning the problem of epistemic alienation. We have co-opted the term alienation directly from Karl Marx’s posthumous Economic and Philosophic Manuscripts of 1844, where alienation (Entfremdung) is defined in terms of the estranged relationship between (1) the laborer and the act of production, (2) the laborer and the product itself, and (3) the laborer and their very own essence (Gattungswesen). Often, there are power dynamics in the biological community preventing bioinformaticians from having access to important decisions regarding experimental design, hypotheses to be tested, and raw data generation procedures, (1). The bioinformatician thus receives data that they are supposed to analyze and convert into biological knowledge, with little room for original interpretation and contribution to the narrative of the study, or freedom to suggest additional experiments (2). Thus, the bioinformatician is systematically denied the status of epistemic driver, although they still consider themselves (and expect to be considered) scientists (3). Our intuition is that the same Marxist lens allows us to look at the opposite face of the coin as well: a wet-lab scientist who produces high throughput data is often unable to follow it up through its analysis, leaving important choices in the hands of bioinformaticians (1), who are still required to translate the wet-lab’s scientist experiment into biological knowledge (2). Without this intermediation, the wet-lab scientist cannot carry out their project and control its narrative, a fundamental aspect of epistemic drivers (3). The open question regards the standing of our theory in the real world: does recognizing the experimental nature of bioinformatics provide a cogent and natural justification for bioinformaticians to become epistemic drivers? Are some bioinformaticians more experimental than others, and does this correlate with their ability to become epistemic drivers? And, if a transformation could be brought upon the field by shifting norms and practices, would the epistemic alienation experienced by both wet-lab and computational scientists be greatly reduced, if not entirely dissolved, creating more collaborative and harmonious research environments?

In this appendix, we review the STS and HPS literature on bioinformatics that shows that there is more to bioinformatics than just data management, and that bioinformatics indeed has an experimental dimension. However, we also point out where these attempts are too timid, in particular when they struggle to connect the experimental dimension of bioinformatics to its own distinctive epistemic goals.

A first comprehensive perspective on the relation between bioinformatics, experiments, and experimental biology comes from the work of Bruno Strasser. In Collecting Experiments: Making Big Data Biology, Strasser (2017) analyzes the relationship between data collections and experimental life sciences from the early twentieth century to the big data era. Despite big data biology being usually described as implying unresolvable dichotomies (e.g., data-driven vs hypothesis-driven), according to Strasser it should be seen as a hybridization of two ways of knowing: the comparative and the experimental. The hybridization is so transformative that a new discipline was born precisely to make sense of this integration; and that discipline is bioinformatics. In Chapters 5 and 6 of his book, Strasser reconstructs the origins and early development of bioinformatics, where databases are central, given their role in handling the data deluge in biology.

If bioinformatics must be understood as a hybridization between the comparative and the experimental, and the comparative is exemplified by databases, then where is the experimental ingredient? One important role of databases that Strasser emphasizes is providing “unique shortcut for experimental investigations” (p. 238), which is done through data manipulation and analysis. Accordingly, we can interpret the experimental ingredient in two ways. First, it lies in the aid that databases provide to experimenters. Second, it consists of the operations of tinkering with data and computational infrastructures that are necessary to build databases in the first place.

While this is certainly a precise characterization of bioinformatics that is grounded in historical evidence, the account has some limitations. In particular, it does not thoroughly consider whether bioinformatics practices characterized in such a way generate a distinctive kind of biological knowledge. Strasser’s account leaves the reader with the doubt that bioinformatics is simply a discipline developed to help wet lab biologists navigate the data deluge, rather than an endeavor with its own epistemic goals. Moreover, the tinkering he describes does not necessarily correspond to “experimental” tinkering, as it is a tinkering often described being motivated by data curation concerns, rather than with specific biological questions in mind.

Another comprehensive attempt to make sense of the role of computational tools—and hence potentially bioinformatics—in contemporary biological research comes from Sabina Leonelli. Her work is focused on providing a comprehensive picture of what she calls “data-centric biology” (2016), where the data deluge and the efforts to handle it are central. As in Strasser’s work, databases in Leonelli’s work play a pivotal role, given that “community databases have … come to play a crucial role in defining what counts as knowledge of organisms in the postgenomic era” (2016, p. 146). But in her work, computational work in biology is more than just databases, and in fact goes hand in hand with the experimental, in at least two ways.

First, there is the relation between data management and experimental knowledge. According to Leonelli, in order to interpret data available on databases and to assess their evidential value, biologists require “some degree of familiarity with the target system that data are taken to document, and the experimental conditions under which that system is studied … they need to have some experience in the experimental manipulation of actual organisms” (2016, p. 93). This motivates a careful handling of the information about procedures, protocols (and, in general, what she calls “embodied knowledge”) and, on the side of data curators, an attempt to formalize such information into metadata (2016, 4.1). This means that the very understanding of experimental knowledge is filtered through computational work.

A second aspect has emerged more systematically in her recent work¹⁴, especially Leonelli (2019). In this article, Leonelli describes some insights of a recent empirical investigation she performed within the phenotyping SureRoot project, where she looked into the challenges of processing and analyzing imaging data. She identifies seven stages of data production and data processing, where different experts (e.g., technicians, data managers, computer scientists, biologists) handle different phases. What is important here is what computer scientists do. Leonelli’s description of their work has an interesting “experimental” flavor. In coding, computer scientists engage in plenty of trial-and-error by tinkering with existing codes or programs in order to build appropriate software for data analysis. Interestingly, at this stage computer scientists do not focus on biological questions, but rather on “which properties of the images at hand would be more easily and reliably amenable to analysis through existing computational tools” (2019, p. 12). In the phase of image analysis, computer scientists literally play around with different designs and properties of charts to visualize results; they are “data games” which are “cheaper than running experiments, with more trial and error allowed” (2019, p. 13). These operations suggest that computer scientists are experimenting with different tools and formats in order to facilitate how biologists will interpret data. It is important to emphasize that computer scientists do not have in mind biological questions per se, but then a negotiation with biologists during image analysis makes these processes iterative, exactly with biological aims in mind. All these operations that Leonelli documents reveal that there is a distinctive kind of experimental, tinkering-based work that is central in computer scientists’ routine in biological contexts. However, to our knowledge these insights are not generalized to bioinformatics qua discipline.

A third author who engages significantly with bioinformatics is Hallam Stevens (2013). Stevens’ main thesis is that computers have not been adapted to the biological context in order to be useful; rather, biological problems have changed in order to accommodate the use of computers. He says that “[b]ioinformatics is not just using computers to solve old biological problems; it makes a new way of thinking about and doing biology in which large volumes of data play the central role” (p. 14). In other words, the availability of big data sets has pushed biologists to find new ways to handle them, turned their attention to computers, but at the same time changed the focus of biological problems because computers can only answer some questions and not others. This goes in the direction of establishing bioinformatics as a discipline with its own goals, and the ‘experimental’ nature of bioinformatics seems to play a key role in establishing the goals of such a discipline. But despite promising suggestions on how the experimental nature of bioinformatics can lead to a distinct type of bioinformatics knowledge (by discussing, among the others, Dayoff’s saga and Genbank), not much is said about the nature of this knowledge. He suggests that it is a kind of statistical knowledge, but this would not make it distinctive from other biological disciplines that use statistics. Nonetheless, Stevens’ insights are essential to develop our account of bioinformatics.