The major resources, such as Gene Ontology (GO), that capture the annotation of biochemical and functional knowledge of proteins, compounds, reactions, and interactions have made important contributions for genomics (Ashburner et al., 2000; Carbon et al., 2021). GO has enabled a major step forward in genomics by constructing a controlled vocabulary that can be used in the annotation of any molecular process or reaction, in any organism. This is the foundation that supports a large number of tools to perform computation on functional data in the three kingdoms of life. However, as is frequent in life, a strength can pose at the same time a limitation. Figure 1 shows the most commonly annotated top-level terms of the GO categories that are utilized in the description of knowledge of Escherichia coli K-12, one of the best-characterized microbes.

At the first level below the root node, “biological process,” there are classes such as cellular processes, biological regulation, response to stimulus, localization, biological adhesion, and metabolic processes. A different classification, MultiFun, devoted to E. coli, has 10 major classes: Metabolism, Information Transfer, Regulation, Transport, Cell Processes, Cell Structure, Location, Extra-chromosomal origin, DNA Site, and Cryptic Gene (Riley, 1993). These major categories are the roots of a hierarchy of several subclasses, and individual genes are classified into one or multiple categories in a structure similar to that in Figure 1.

These numerous classes and their subclasses are the result of essentially a bottom-up approach of annotation of individual gene functions across different species within the GO and, in the case of MultiFun, the result of years of work specifically in the annotation of E. coli genes by Monica Riley in the earlier days of genome analysis (Riley, 1993; Serres and Riley, 2000). They have the common aim of describing all known functions of gene products. Both GO and MultiFun are examples of classifications used in the annotation of functions of genes as they continue to be uncovered.

While these classifications help in the annotation of gene properties, they are not adequate to fully explain the functioning of an organism. Ideally, the major groupings in a classification system would correspond to the major concepts needed to grasp functional microbiology, or the physiology of a multicellular organism. The prevailing view of a cell is that of an evolving and self-replicating chemical factory run by a program, coded in the DNA, and whose main (if not sole) agenda is to survive and proliferate as much as possible, hence ensuring the transmission of its genome. This involves adapting to a changing and often stressful, competitive environment. This general principle is not conveyed at the high-level description of the functional genomic classifications. Why is there such a large discrepancy between the major classes of functional classification systems and the main concepts required to understand the major principles of microbial physiology?

Function, as defined in GO (Thomas, 2017), is ‘‘a specific objective that the organism is genetically programmed to achieve.’’ Note that this definition encompasses what GO refers to as ‘‘molecular function’’ and ‘‘biological process.’’ The contributions of both annotations for a gene product identify the properties within a larger ‘‘system’’ it belongs to. However, these annotations and classes are too granular for general principles such as fitness, survival, and viability to emerge. These types of concepts, concerning the properties of entire organisms (phenotypes, in the wide meaning of the term), are described in other ontologies such as Ascomycete Phenotype Ontology.¹ There is no easy way to computationally navigate between these organism-level observations and the complement of reactions and signaling processes that mediate them.

The granularity of genes may be too low for understanding where a physiological capability achieved at the molecular level is located in the larger cellular context. The structure of the GO contains three different classifications, defined as: molecular function for the activities of individual genes, biological process for pathways and larger processes also defined as specific objectives that the cell or organism is genetically programmed to achieve, as already mentioned, and cellular component to indicate where the gene products are active².

A less-detailed level of description, to understand processes, could be that of biochemical pathways, or similarly, the level of activities of collections of genes within transcription units, operons, and even at the level of regulons, that is to say, genes subject to regulation by the same transcription factor (TF). An annotation effort at this level is illustrated by the molecular and functional descriptions of regulons or genetic sensory-response units (GENSOR Units) in RegulonDB (Gama-Castro et al., 2016). Take for instance the TrpR (tryptophan regulator) regulon of E. coli. In RegulonDB, the collective function of 9 genes is summarized as: ‘‘The presence of tryptophan inhibits the expression of genes needful for transport and synthesis of tryptophan from serine and conversion of erythrose-4P to chorismate, a precursor of tryptophan’’.³ This would be called tryptophan homeostasis in GO. This example shows a good direction of succinctly describing genetic capabilities of groups of genes. The Gene Ontology Consortium, with the development of the Causal Activity Model (GO-CAM), now provides a mechanism to express this type of information (Thomas et al., 2019).

We know that only around one-fifth of regulons in E. coli K-12 encode genes that collectively belong to a single biological process; several regulons encode genes involved in multiple biological processes, whereas in 16% of the 189 known regulons, the gene products encode reactions that cannot be connected to build pathways (Ledezma-Tejeida et al., 2019).

When grouping genes into complex regulons, defined as those subject to regulation by multiple and exactly the same TFs, the functional homogeneity of their regulated genes increases considerably in terms of belonging to a dominant functional class [See Figure 3 in Ledezma-Tejeida et al. (2019)]. It is true that this has to be taken with a grain of salt, since currently described regulons are not yet comprehensive since new TF-gene interactions may still be discovered for many TFs. We need to wait for the full genome characterization of complex regulons to solidify this interesting observation. Certainly, finding biologically motivated classes of genes that together perform a dominant function would offer a highly desired intermediate granular level of functional description connecting gene functions with cellular capabilities. At present, we can only speculate whether this intermediate level will be first confirmed in E. coli K-12, present in other bacteria, and less likely, present also in eukaryotes. It will be important to keep capturing whole-genome TF-gene regulatory interactions. Next-generation functional analysis tools should be able to integrate the various levels of knowledge to allow more powerful querying.

One non-trivial issue is that—logically—we follow a conceptual frame of what a living system is and what is necessary for it to work according to the prevalent concepts and notions of the time. This rules out, therefore, far unknown requirements that may be disclosed in the future when more knowledge is acquired, and it raises the challenge of handling the sure unknown unknowns that still remain to be understood and cannot yet be assigned sound gene annotations. In fact, we always work with incomplete knowledge. There is no genome yet with all its genes functionally characterized. GO deals with this by annotating a gene directly to the root term (either “molecular function,” “biological process,” or “cellular component”) when any one aspect has not been characterized. This formally means that the gene enables some “molecular function,” participates in some “biological process,” and is active in some “cellular component.” Fully annotated organisms have at least this information for every gene.

Another confounding factor is semantic ambiguity, even in the scientific community. Unequivocal functionality/annotation of a gene can be assigned in many cases without ambiguity, especially when the meaning of a word or a concept is clearly defined in a dictionary or an ontology. But the information used to annotate genomes is extracted from research articles using natural language processing, and the meaning of a term used in an article may differ significantly from that of the ontological concept definition, leading to misinterpretation and inconsistencies in annotation.

A traditional approach to functionally annotate ORFs (open reading frames) on the basis of the function historically assigned a first biological description and then propagated it to related sequences. Without further evidence, this is a reasonable approach to a first “guess” of the function of an uncharacterized gene, but this can lead to confusion, since contradicting that initial misannotation can be difficult, if only because researchers are misled rather than just unguided in their hypothesis.

A revealing example is the function of the arch-famous Catabolite Regulation Protein, or CRP. The crp gene was first discovered and exhaustively studied in E. coli as a key component of the regulation of the equally arch-famous lac promoter in this species. It was later found that CRP also regulates a large number of other metabolic processes, in addition to utilization of lactose as a carbon source, both negatively and positively in response – inter alia – to fluctuations in intracellular cAMP levels. Yet, years later it was shown that virtually identical CRP orthologs found in other bacterial species [e.g., Pseudomonas (Milanesio et al., 2011)] have nothing whatsoever to do with metabolism but play a role in membrane-related functions. Finally, recent data suggest that in reality CRP in E. coli is not only a promoter-specific regulator but also an authentic genome-scaffolding protein (Heyde et al., 2021) which responds not just to cAMP but to other physiological effectors like cytidine (Lauritsen et al., 2021) and thus has an additional global role. This highlights the importance of maintaining ontologies and the databases relying on them up to date to represent as closely as possible the current state of knowledge.

The crp gene example is not an isolated case: it illustrates a widespread phenomenon called exaptation, i.e., co-opting of a given biological object (or its gene thereof) for a function very different from what it originally evolved. How to translate such functional ramifications of the same gene into sound criteria for predicting authentic biological roles remains a serious challenge. A conserved mechanism (binding of a TF to a given target DNA in response to a distinct effector) may result in entirely different physiological outcomes. Annotations in genomic databases have traditionally merged molecular mechanisms of individual gene products and their physiological roles-in-context, thereby causing considerable confusion. GO-CAM alleviates this problem, by connecting gene function to activators, substrates, and larger processes.

The Gene Ontology Consortium has been using a phylogenetic-based approach to complement annotations directly derived from experimental data (Gaudet et al., 2011). Conserved orthologs are annotated with functions that are expected to be conserved in the target species, while taking care not to infer functions that can easily diverge following gene duplication. This increases the coverage of annotations, in particular for non-model organism species for which experimental data are more scarce.

In his reference book on bacterial physiology, Frederick Neidhardt suggests that the whole set of chemical reactions of bacterial cells can be categorized as: assembly, polymerization, biosynthetic, and fueling reactions (Neidhardt et al., 1990). This, or alternative higher-level descriptions, could help to identify the most conserved elements present in many organisms and facilitate the identification, or even the validation within new genomes, of plausible general principles of cell physiology.

Certainly, in addition to shared evolutionary origin, any substance has both physical and chemical properties that the cell has to deal with, and as a consequence, some aspects of their biochemistry are shared across many organisms. Processing of carbon sources and nitrogen sources happens with associated consequences in acidification or osmotic changes in E. coli and in many other bacteria. As a consequence, some aspects of the combinatorial nature of control may also be shared across many bacteria, such as co-occurrence of nitrogen with extrusion of acid, or the use of glucose and extrusion of acid stress; all of these are due to the common chemical ground of molecules and their properties. It is reasonable to assume that some combinations or integration of physiology will be shared across bacteria given some universal properties of such biochemical and chemical processes, even if the precise mechanisms vary from one species to the next. Most likely, shared combinations of physiology will be found in organisms that have commonalities in their environment.

The top nodes of the biological process GO, similar to the top 10 nodes of MultiFun, include categories that are devoid of physiological content, such as: metabolic process, cellular process, biological regulation, signaling, or transport. These ontology-driven concepts contribute to a way of ordering the activities of genes. An interesting link would be to map gene functions to a separate collection of categories, all of which entail physiological content, with a structure reflecting the organization of bacterial physiology. Bacteria are governed by the need to survive, grow, and reproduce. A first exercise would require an annotation effort that could capture the consensus perspective of a group of experts in the field. This conceptual effort would strongly benefit by defining a dialogue with genome-wide and phenotype characterization experiments with tools—similar to gene enrichment—that help both the interpretation of experiments and the improvement of the conceptual model. Briefly, we suggest that a mapping of the current GO categories with a new collection of categories—devoid of ontological requirements—that reflect the priorities of bacterial physiology generated by a group of experts in the field would be a step forward to connect gene categories with principles of microbial physiology.

It is important to note that the physiological needs of an organism under specific conditions select for given mechanisms available to fulfill the function—and not vice versa. The specific molecular mechanisms to achieve a given physiological function can vary considerably, even among not-so-distant evolutionarily related microorganisms.

A clear-cut boundary between the biology of mechanisms and physiology in our perspective is the distinction between activator and repressors vs. inducible and repressible systems. Terms like activator and repressor in the description of gene regulators are required when describing mechanisms. Furthermore, genes subject to activation can be induced or repressed just as genes subject to repression, in the sense of being negatively regulated, can be induced or repressed. For instance, the classic LacI repressor is a player of an inducible system since, in the absence of glucose, when lactose is present and is incorporated into the cell, allolactose will bind to LacI, provoking its unbinding from the operator sites and inducing transcription of the lac operon. In our opinion the terms “inducible” and “repressible systems” as a consequence of the appearance of a signal belong to physiology, whereas the type of regulator characterizes their mechanisms.

Cellular capabilities described as physiological activities are more likely to remain valid definitions as new research continues to expand and new mechanisms are discovered. An example of this challenge can be found in the discussion and update of major concepts of transcriptional regulation in bacteria (Mejía-Almonte et al., 2020). Whether the two levels (physiological, mechanistic) can be automatically distinguished through computational means and retrieved with the same ease that users enjoy with existing annotation platforms remains to be seen (Mejía-Almonte and Collado-Vides, 2019). Probably advanced, dedicated artificial intelligence approaches will have to be developed and adopted for boiling down the plethora of available data to predict useful and significant functionalities.

We envision that ontologies and databases could/would provide data to test general hypotheses, such as “physiology is more conserved than the specific mechanisms to achieve a given function or biological program.” For this to happen biologically major terms that distinguish molecular biology and mechanistic descriptions from physiology need to be more clearly defined and incorporated into formal frameworks.

Biology is a very rich discipline with knowledge that goes from genetics to behavior, from structure to evolution and dynamics. It is noteworthy that this special journal issue addresses ontology and physiology, whereas most microbial genomic databases are still struggling with mechanisms, processes, and genetics (Kanehisa et al., 2017; Karp et al., 2019; Santos-Zavaleta et al., 2019; Ng et al., 2020; Keseler et al., 2021). Genomic databases gather what we can call genotypic properties, where activities of biomolecules are described but, so far, in the case of microbial organisms, without the precise description of the growing conditions under which a subset of genotypic properties are being actively used.

Gene ontology is a highly used resource to compare pairs of experimental and control whole-genome transcriptional profiles to identify the enriched functions within the genes that show changed expression. The expanding genomic high-throughput technologies are accelerating the identification of the elements of transcriptional regulatory networks (transcription start sites, transcription factor binding sites, terminators, transcription units, and expression profiles) at the genomic level. How can we best combine these resources to characterize the dynamics of inducing and repressing complete biological processes and link them to their mechanisms? Combining genomes, functional gene annotations, and the deciphering of regulatory networks should improve our capability of mapping specific mechanisms with cellular collections of physiological capabilities.

The link genotype-phenotype has been most often assigned on the basis of heterogeneous experiments with data expressed in formats that typically lack semantic let alone quantitative standards amenable to computation. One ongoing trend to tackle this issue is massive testing of mutant collections grown under a large number of standardized culture conditions (nutrients, stresses, etc.). The so-called phenotypic microarrays (Bochner et al., 2001), along with available high-throughput analyses of transcriptomics, proteomics, and metabolomics of the corresponding strains are increasingly establishing the field of phenomics (Acin-Albiac et al., 2020). Phenomics attempts to assign functionalities to given genes on the basis of a data landscape much wider than the typical reductionist approach and linear logic of traditional genetics. Extensive analyses of this type are already available for E. coli (Nichols et al., 2011; Otsuka et al., 2015) and are quickly expanding to other species of interest. It comes as no surprise that phenomics has become a fertile target of machine learning approaches (Lürig et al., 2021), applicable to both wet datasets and text sources (Mao et al., 2016) in published literature. These new perspectives (which may also involve considerable investments in technological platforms) have the potential to change the way automated annotations and GOs will be established (Gkoutos et al., 2012) in the not-so-distant future.