2. Context and historical background

MGB research weaves together several strands of earlier investigation from neuroscience, gastroenterology, and microbiology. The exploration of connections between the gut and brain has a particularly long and venerable research history. Early psychologists William James and Carl Lange are seen as forerunners of brain-gut-axis research (e.g., Eisenstein Reference Eisenstein2016), although they made limited claims about these connections. James merely insisted that “visceral stirrings” should be conceptualized as part of the emotion of fear (1884). Subsequent research continued to connect emotional responses to visceral signals. In the early twentieth century, for example, Walter Bradford Cannon observed that “the movements of the stomach immediately stopped” when “a female [cat] with kittens turned from her state of quiet contentment to one of apparent restlessness” (Cannon Reference Cannon1909, p. 484). He postulated that these changes depended on the sympathetic nerve supply (Cannon Reference Cannon1911).

More fine-grained studies followed. The administration of adrenaline, which is released by the host after activation of the sympathetic nervous system, was discovered to lower the number of pathogenic bacteria needed to establish a generalized infection or to kill the animal (Evans et al. Reference Evans, Miles and Niven1948; Renaud & Miget Reference Renaud and Miget1930). Although these effects were attributed to decreased recruitment of white blood cells (Evans et al. Reference Evans, Miles and Niven1948), it was realized much later that adrenaline also diminishes the bactericidal activity of these cells (Qualliotine et al. Reference Qualliotine, DeChatelet, McCall and Cooper1972). Other experiments revealed that adrenaline actually reduces host mortality after the injection of bacterial toxins (Chedid & Boyer Reference Chedid and Boyer1953; Hodoval et al. Reference Hodoval, Morris, Crawley and Beisel1968), which suggested that this hormone has different effects on living bacteria and bacterial fragments. Another molecule, acetylcholine, which is released by the parasympathetic nervous system of animals, was also shown by other research to be produced by a strain of the bacterium Lactobacillus plantarum (Stephenson & Rowatt Reference Stephenson and Rowatt1947).

By the 1980s, the term brain-gut axis had become a common label for investigations of these connections (e.g., Aziz & Thompson Reference Aziz and Thompson1998; Gastroenterology 1980). A variety of important findings emerged about gut microbes, their cell wall components, and nervous systems or behavioural states (Bluthé et al. Reference Bluthé, Dantzer and Kelley1992; Hart Reference Hart1988; Lyte Reference Lyte1993). Further extensive experimentation on adrenaline and noradrenaline showed that they stimulate the growth of some bacteria (Lyte & Ernst Reference Lyte and Ernst1992) and that some microorganisms are themselves able to produce these substances (Asano et al. Reference Asano, Hiramoto, Nishino, Aiba, Kimura, Yoshihara, Koga and Sudo2012; Tsavkelova et al. Reference Tsavkelova, Botvinko, Kudrin and Oleskin2000). This body of evidence that microbes can influence the gut-brain axis, and in turn be influenced by the brain-gut axis, forms an important basis for more recent developments in MGB research.

From the microbiological angle, intestinal microbes have long been studied for their effects on human health, from the perspectives of both individual pathogens and more systemic community effects (see Haenel Reference Haenel1961; Savage Reference Savage2001). Although work in the 1980s had begun to examine mechanistically how specific intestinal microorganisms might affect mammalian brain states (e.g., Brown et al. Reference Brown, Price, King and Husband1990; Jeppsson et al. Reference Jeppsson, James, Hummel, Brenner, West and Fischer1983), it is only in the last decade that brain-gut-axis research has been able to take advantage of methods that reveal the full diversity of microorganisms inhabiting the human gut. This expanded capacity for the molecular analysis of microbial communities in host organisms is what is now called microbiome research.

Microbiome research developed on the basis of tools that allow analysis of the DNA sequence of entire microbial communities (microbiota). The DNA is directly extracted from microbiota in their natural environments (Handelsman et al. Reference Handelsman, Rondon, Brady, Clardy and Goodman1998; see sect. 5 for more detail). “Microbiomes,” the molecular sequences of these communities, are analysed for compositional patterns and their associations with aspects of the environment. In the mid-2000s, microbiome researchers began to focus more closely on the human ecosystem: the human body and its complement of microorganisms, particularly gut microbes (Eckburg et al. Reference Eckburg, Bik, Bernstein, Purdom, Dethlefsen, Sargent, Gill, Nelson and Relman2005). As human microbiome research developed, key researchers began to use germ-free (GF) mice. These are mice that are born and live their lives without microorganisms until they are experimentally colonized; other GF organisms have been used historically for different purposes (Kirk Reference Kirk2012). Influential studies showed that giving GF mice microbiota transplants from obese hosts could bring about obesity (e.g., Turnbaugh et al. Reference Turnbaugh, Bäckhed, Fulton and Gordon2008). Although GF mice have many abnormalities (see sect. 7.3), they have become the gold experimental standard for causal claims in human microbiome research, which now includes gut-brain studies.

Despite all of these well-known MGB precursors, the current phase of microbiome-oriented gut-brain research often cites its starting point as 2004, when Sudo and colleagues (Reference Sudo, Chida, Aiba, Sonoda, Oyama, Yu, Kubo and Koga2004) used germ-free mice to reveal that “commensal microbes [are] affecting the neural network responsible for controlling stress responsiveness” (p. 271). Many of today's microbiota-gut-brain papers refer to the Sudo et al. paper as “seminal” (e.g., Mayer et al. Reference Mayer, Tillisch and Gupta2015, p. 926; Sampson & Mazmanian Reference Sampson and Mazmanian2015, p. 567) and as a “landmark” in the history of the emerging field of MGB research (e.g., Foster & McVey Neufeld Reference Foster and McVey Neufeld2013, p. 306). This 2004 paper emphasizes a simple potential treatment: probiotics. It also suggests that GF mice allow much of the complexity of microbiomes to be ignored: Mice either have microbiota or they do not. Both this paper and the earlier work have inspired attempts to merge multiple disciplinary perspectives, including those from psychiatry, pharmacology, psychology, neuroscience, immunology, microbiology, and gastroenterology. But in the process of drawing on so many approaches, key problems plaguing broader microbiome analyses were also included: the difficulty of identifying causal pathways and yet the tendency to suggest microbiota are bringing about specific host effects (see Hanage Reference Hanage2014).

3. MGB research and its scope

In part because of its rich historical background, MGB studies draw on a considerable variety of methods and disciplinary approaches (see Supplementary Table 1). These methods are both experimental and descriptive. They focus on implementing microbiota-related interventions that can change specified brain and/or behavioural states. The targets of these interventions are usually disorders of various degrees, including depression (Jiang et al. Reference Jiang, Ling, Zhang, Mao, Ma, Yin, Wang, Tang, Tan, Shi, Li and Ruan2015; Park et al. Reference Park, Collins, Blennerhassett, Ghia, Verdu, Bercik and Collins2013), anxiety (Crumeyrolle-Arias et al. Reference Crumeyrolle-Arias, Jaglin, Bruneau, Vancassel, Cardona, Daugé, Naudon and Rabot2014; Neufeld et al. Reference Neufeld, Kang, Bienenstock and Foster2011a), autism (de Theije et al. Reference de Theije, Wopereis, Ramadan, van Eijndthoven, Lambert, Knol, Garssen, Kraneveld and Oozeer2014; Hsiao et al. Reference Hsiao, McBride, Hsien, Sharon, Hyde, McCue, Codelli, Chow, Reisman, Petrosino, Patterson and Mazmanian2013), schizophrenia (Severance et al. Reference Severance, Yolken and Eaton2016), posttraumatic stress disorder (Hemmings et al. Reference Hemmings, Malan-Müller, van den Heuvel, Demmitt, Stanislawski, Smith, Bohr, Stamper, Hyde, Morton, Marotz, Siebler, Braspenning, Van Criekinge, Hoisington, Brenner, Postolache, McQueen, Krauter, Knight, Seedat and Lowry2017), Parkinson's (Sampson et al. Reference Sampson, Debelius, Thron, Janssen, Shastri, Ilhan, Challis, Schretter, Rocha, Gradinaru, Chesselet, Keshavarzian, Shannon, Krajmalnik-Brown, Wittung-Stafshede, Knight and Mazmanian2016), and anorexia nervosa (Kleiman et al. Reference Kleiman, Watson, Bulik-Sullivan, Huh, Tarantino, Bulik and Carroll2015). But more general brain and behavioural states are also scrutinized, including fear (Bravo et al. Reference Bravo, Forsythe, Chew, Escaravage, Savignac, Dinan, Bienenstock and Cryan2011), stress (O'Mahony et al. Reference O'Mahony, Clarke, Dinan and Cryan2017), mood (Steenbergen et al. Reference Steenbergen, Sellaro, van Hemert, Bosch and Colzato2015), temperament (Christian et al. Reference Christian, Galley, Hade, Schoppe-Sullivan, Kamp Dush and Bailey2015), cognition (Magnusson et al. Reference Magnusson, Hauck, Jeffrey, Elias, Humphrey, Nath, Perrone and Bermudez2015), memory (Gareau et al. Reference Gareau, Wine, Rodrigues, Cho, Whary, Philpott, Macqueen and Sherman2011), and sociability (Desbonnet et al. Reference Desbonnet, Clarke, Shanahan, Dinan and Cryan2014).

When experimental effects are detected, mechanisms are often postulated to consolidate the links made between these brain and behavioural outcomes and the microbiota. Proposed intermediary mechanisms include the vagus nerve, inflammatory molecules, microbial metabolites and “neurotransmitters,” immune system mediators and responses, various “signalling” molecules and cells, the so-called leaky gut, and leaky blood-brain barriers (see Sampson & Mazmanian Reference Sampson and Mazmanian2015). None of these are uncontested as potential or adequate mechanisms. For example, the molecules often labelled “neurotransmitters” are not neurotransmitters for the microbes. Even if these molecules can cross the gut barrier and blood-brain or nerve barriers, they do not meet the criteria for neurotransmitters. These criteria require a neurotransmitter to be present in presynaptic elements, for it to be released in response to presynaptic depolarization and for there to be receptors on a postsynaptic cell (Purves et al. Reference Purves, Augustine, Fitzpatrick, Katz, LaMantia, McNamara and Williams2001). Another very problematic mechanism is the “leaky gut” and its highly disputed role in neurological disorders (e.g., Quigley Reference Quigley2016; Rao & Gershon Reference Rao and Gershon2016; see sect. 7).

An outline of some key studies in MGB research will help show the field's scope and trajectory of development. The now classic Sudo et al. (Reference Sudo, Chida, Aiba, Sonoda, Oyama, Yu, Kubo and Koga2004) paper serves as something of a template for much subsequent research. In that paper, Sudo et al. compare hypothalamo-pituitary-axis (HPA) responses to restraint stress in GF, specific pathogen-free (SPF), and conventional mice (i.e., unmanipulated microbiota). The study found that GF mice show higher post-stress corticosterone concentrations than SPF and conventional mice. In addition, higher corticosterone in GF mice was counteracted by the administration of probiotic bacteria (Bifidobacterium infantis). Because this occurred only to the 9-week-old mice and not the older ones (17 weeks), Sudo et al. (Reference Sudo, Chida, Aiba, Sonoda, Oyama, Yu, Kubo and Koga2004) postulated a crucial developmental stage for the HPA stress response that is determined by microbiota. These key findings of probiotic effects on physiology and behaviour, plus a developmental window of maximum effect, get taken up in numerous other MGB papers.

In 2009, O'Mahony and colleagues established that several consequences of maternal separation stress exist at adulthood: namely, visceral hypersensitivity, changes in gut microbiota, less exploration of novel environments, and more defecation. Those behaviours are often considered “anxiety-like” (see sect. 6 for further discussion). The relationship of such behaviour to microbes had already been explored in earlier work focused on single microbes (e.g., Lyte et al. Reference Lyte, Varcoe and Bailey1998). Following the new trend of focusing more broadly on the microbiota as a whole, Diaz Heijtz et al. (Reference Diaz Heijtz, Wang, Anuar, Qian, Björkholm, Samuelsson, Hibberd, Forssberg and Pettersson2011) and Neufeld et al. (Reference Neufeld, Kang, Bienenstock and Foster2011a) found that GF mice (i.e., no microbiota at all) display fewer anxiety-like behaviours than SPF mice in the light-dark box and elevated plus maze.

In the same year, Bercik et al. (Reference Bercik, Denou, Collins, Jackson, Lu, Jury, Deng, Blennerhassett, Macri, McCoy, Verdu and Collins2011a) published findings of the effects of oral antibiotics on anxiety-like behaviour in the step-down and light preference tests. Comparisons were made after microbiota transplantations into SPF Balb/C mice (an inbred mouse strain widely used in immunology and considered to display a high level of anxiety-like behaviour, or “timidity”), National Institutes of Health (NIH) Swiss mice (an outbred strain that shows less anxiety-like behaviour, or greater “boldness”), or GF Balb/C mice. The study found that oral antibiotic treatment reduced anxiety-like behaviour and increased exploration of the behavioural devices used, and that this increased exploration did not involve autonomic nerves. In addition, Bercik et al. (Reference Bercik, Denou, Collins, Jackson, Lu, Jury, Deng, Blennerhassett, Macri, McCoy, Verdu and Collins2011a) reported that Balb/C recipient mice transplanted with NIH Swiss microbiota showed more exploration than their counterparts with only Balb/C microbiota. Conversely, NIH Swiss mice that received Balb/C microbiota transplantation displayed less exploration than those that were colonized with NIH Swiss microbiota. The success of these interventions suggested to many people in the field that the microbiota is a major causal agent in determining anxiety-like behaviour.

Making a narrower microbial intervention (i.e., just one microbe, not a community), Bravo et al. (Reference Bravo, Forsythe, Chew, Escaravage, Savignac, Dinan, Bienenstock and Cryan2011) used a probiotic bacterium (Lactobacillus) to manipulate anxiety-like and depression-related behaviours in mice. They examined depression-related behaviour with the forced swim test (measuring how long the animal was immobile) and anxiety-related behaviour by the number of entries on to the open arms of the elevated plus maze. They also measured the time spent freezing after fear conditioning with a mild electric shock. Probiotic administration reduced immobility during forced swim tests and increased the number of open arm entries in the elevated maze. Subdiaphragmatic vagotomy (severing the vagus nerve under the diaphragm) prevented these effects (however, see Bercik et al. [Reference Bercik, Denou, Collins, Jackson, Lu, Jury, Deng, Blennerhassett, Macri, McCoy, Verdu and Collins2011a], who found no role for the vagus nerve in modulating the effects of antibiotics on the behaviour of mice in the light-dark preference and step-down tests). Follow-up studies subsequently showed that the probiotic facilitates firing of vagal sensory fibres (Perez-Burgos et al. Reference Perez-Burgos, Wang, Mao, Mistry, McVey Neufeld, Bienenstock and Kunze2013). Findings such as these have given rise to the idea of “psychobiotics.” These are substances derived from microorganisms that can be used as treatments for improving mental health (Dinan et al. Reference Dinan, Stanton and Cryan2013). This notion has strong appeal inside and outside of the MGB field.

A study by Hsaio et al. (Reference Hsiao, McBride, Hsien, Sharon, Hyde, McCue, Codelli, Chow, Reisman, Petrosino, Patterson and Mazmanian2013) suggested how such interventions might work mechanistically. The authors used adult mice born from mothers that had been administered an immune stimulation (a viral mimic) during pregnancy. The pups were born with both a “leaky gut” and the behavioural features of autistic developmental disorders. The adult offspring displayed anxiety-like features in the open field, stereotypical behaviour, less social interaction, and fewer ultrasound vocalizations. Feeding Bacteroides fragilis to the impaired mice mitigated “obsessive” behaviours such as grooming and marble-burying. However, reduced sociability did not improve, which was attributed to developmental timing. B. fragilis was known from earlier immunological studies to improve immune defects (Mazmanian et al. Reference Mazmanian, Round and Kasper2008). Although Hsaio et al. did not isolate colonized B. fragilis in the mouse intestines, a metabolic mediator associated with this microorganism was restored to normal levels after probiotic treatment. Studies such as this, although still incomplete, hint at the potential mechanistic pathways that might underlie microbiota effects on brain and behaviour.

Many MGB studies, including those just discussed, are believed to be relevant to human psychiatric disorders. In addition, cognitive and behavioural processes that are not necessarily connected to any psychiatric disorder have also been linked to microbiota changes. Bravo et al. (Reference Bravo, Forsythe, Chew, Escaravage, Savignac, Dinan, Bienenstock and Cryan2011) showed that although no differences in the amount of behavioural freezing were observed immediately after mice received a foot shock, mice that were fed a probiotic showed more conditioned freezing the next day than probiotic-free mice. Diet also has effects. Non-obese antibiotic-pretreated mice were given microbiota transplants from animals fed a high-fat diet. The mice with the high-fat microbiota transplants displayed more conditioned freezing to a shock-signalling tone than did mice with transplants from animals on a control diet (Bruce-Keller et al. Reference Bruce-Keller, Salbaum, Luo, Blanchard, Taylor, Welsh and Berthoud2015). Gareau et al. (Reference Gareau, Wine, Rodrigues, Cho, Whary, Philpott, Macqueen and Sherman2011) observed that probiotics could reverse stress-induced deficits in novel object recognition. Antibiotic treatment of healthy mice from adolescence through adulthood was also found to impair novel object recognition in mice (Desbonnet et al. Reference Desbonnet, Clarke, Traplin, O'Sullivan, Crispie, Moloney, Cotter, Dinan and Cryan2015).

Whether about cognitive or emotional capacities, or aspects of psychiatric disorders, the potential implications of these and many other studies are striking. Many of the core findings and interpretations are echoed repeatedly in the general MGB literature, which is characterized by an abundance of reviews (see sect. 1 in the Supplementary Material). Some of this work then goes so far as to claim that microbes control the mind and that free will is thereby refuted (e.g., Lepage et al. Reference Lepage, Leclerc, Joossens, Mondot, Blottière, Raes, Ehrlich and Doré2013; Stilling et al. Reference Stilling, Dinan and Cryan2016; see sect. 7). Most of these reviews, as well as much primary research, proclaim that a conceptual and methodological revolution is underway in brain and behavioural research (e.g., Liu Reference Liu2017; Mayer et al. Reference Mayer, Knight, Mazmanian, Cryan and Tillisch2014). And yet much of the research is highly speculative regarding causation and mechanisms, some of it is contradictory, and many well-established methods are used in limited, mistaken, and even outdated ways, as we will show.

Although some scientific papers and popular essays have already pointed toward central problems for microbiome research (e.g., Eisen Reference Eisen2017; Shanahan & Quigley Reference Shanahan and Quigley2014) and warnings have been issued about MGB “hype” in particular (see comments in Smith Reference Smith2015; Zimmer Reference Zimmer2014), these discussions have not been based on detailed examinations of the core literature. Very commonly within the field, cautionary statements are embedded in strongly promotional overviews of MGB research (e.g., Mayer et al. Reference Mayer, Knight, Mazmanian, Cryan and Tillisch2014; Sherwin et al. Reference Sherwin, Dinan and Cryan2018). Our aim is to provide a more thorough critical and external analysis of the field for anyone who wants to understand human minds and behaviours, and their putative microbiome connections.

4. The 25 most cited MGB papers

In order to analyse the field more closely, we examined the most highly cited MGB papers in the last decade (Table 1). We chose this set of papers because of their importance to the established field, and particularly its experimental core. They have shaped the field and continue to structure it, as all their citations attest. Focusing on them allows us to probe deeply into influential methods and interpretations, which would be less effectively achieved in a comprehensive but relatively shallow overview of all existing literature. Although we recognize that this selection of papers will not include the most recent work in the field (some of which may be using improved techniques), our aim here is to capture the most recognized experimental work that has been the basis for the majority of reviews and subsequent studies, as well as media attention.

Table 1. The 25 most cited papers in MGB research^a

^a Papers were extracted using a combination of PubMed searches and Google Scholar citations. See the main text for detailed selection methods. Papers are ranked by the number of citations received.

To identify this central corpus of work, we carried out a PubMed search using the term “gut-brain microbiota” (date of access: May 25, 2017; updated July 11, 2018). We discarded all reviews, which formed a very high proportion of the literature (almost 50%; see Supplementary Material). This search found 325 articles. We then used Google Scholar citation counts for each article to rank all the papers with more than 150 citations (a total of 15). To supplement this core of highly cited papers, we also examined the references to open access articles within the original 325 articles. This strategy found another 9 highly cited articles. Finally, we conducted a third search using the looser term “brain microbiota.” This search found 867 articles. We inspected the most cited articles of this group, which revealed another 3 publications that had not appeared in our earlier “gut-brain microbiota” search. We slightly cropped this list to 25 papers, of which the lowest number of citations is just over 120 and the highest above 1300 (Table 1). We then analysed the text of these papers manually, with an initial focus on two categories of methodology: microbiome methods (sect. 5) and behavioural tests and statistics (sect. 6).

5. Microbiome methodology

Microbiome research relies on the rapid and extensive DNA profiling of bacterial and other microorganismal genomes in specified locations. This use of DNA sequencing tools to explore microbial biodiversity is often called metagenomics, meaning that it goes beyond the single-species genome analyses of genomics (Handelsman Reference Handelsman2004). It allows the investigation of microbial communities in a vast variety of environments, including those provided by animal hosts. These methods have liberated the study of microbial biodiversity from the constraints of pure culture. Pure culturing approaches require growing microorganisms in the laboratory, which is not feasible (yet) for many microorganismal groups.

In the simplest scenario for sequencing, the presence of species is evaluated with metagenomic methods, which can be performed in two ways. The first is tag (or amplicon) sequencing, usually of a particular stretch of a ribosomal gene. The second is shotgun sequencing, which captures all the genes in the environmental sample. Tag sequencing is still widely used despite being restricted to information about bacterial abundance and diversity. Shotgun sequencing provides more information about the total pool of genes present in the environment but requires more complicated bioinformatic analysis. In order to do more than catalogue taxa on the basis of genes, researchers also employ metatranscriptomic methods to find actively transcribed genes, and metabolomic analyses to quantify the output of bacterial metabolic pathways (see Knight et al. [Reference Knight, Vrbanac, Taylor, Aksenov, Callewaert, Debelius, Gonzalez, Kosciolek, McCall, McDonald, Melnik, Morton, Navas, Quinn, Sanders, Swafford, Thompson, Tripathi, Xu, Zaneveld, Zhu, Caporaso and Dorrestein2018] for an updated methodological primer). However, whether tag or shotgun methods are used, the bulk of microbiome research has yet to advance beyond gene catalogues, and this greatly limits what can be said about microbial effects on hosts and other environments. But as we will show, a surprising amount of the MGB research in our top-cited sample does not even achieve the cataloguing step.

The gut is home to the most studied but also the most complex human-associated microbiota. It contains hundreds if not thousands of different microbial species, of which bacteria are the main component and research focus. The relative abundance and diversity of bacteria can vary considerably from one individual human to another (Human Microbiome Project Consortium 2012). Difficulties in interpreting diverse and complex sequence data result in the main output of health-focused microbiome studies being simple correlations between the abundance of particular taxa and host-associated disease states. These association patterns do not allow cause and effect to be ascertained (de Vos & de Vos Reference de Vos and de Vos2012; Hanage Reference Hanage2014). Moreover, the great majority of investigation is done with faecal samples, which are unlikely to represent microbial activity in the gut itself, especially in the small intestine or in association with the mucosal surface (Gevers et al. Reference Gevers, Kugathasan, Denson, Vázquez-Baeza, Van Treuren, Ren, Schwager, Knights, Song, Yassour, Morgan, Kostic, Luo, González, McDonald, Haberman, Walters, Baker, Rosh, Stephens, Heyman, Markowitz, Baldassano, Griffiths, Sylvester, Mack, Kim, Crandall, Hyams, Huttenhower, Knight and Xavier2014; Momozawa et al. Reference Momozawa, Deffontaine, Louis and Medrano2011; Quigley Reference Quigley2017). Nevertheless, the sheer convenience of such samples continues to ensure their popularity.

How does microbiome research feature in MGB studies? In general, most MGB papers are not microbiome-driven in the way many other health-related or environmental microbiome papers are. In fact, in MGB research, including our 25 most cited list, “microbiota” and “microbiome” are often used simply to indicate that microorganisms in the human body appear to be involved in producing observed effects. Despite many methodological advances in microbiome research, standard microbiome analyses are not carried out even in many of the most highly cited MGB papers.

There are four broad categories of “microbiota” methods in the 25 most cited MGB papers we analysed.

1. 1. Comparisons of behaviours in GF mice/rat microbiomes with conventionally colonized or SPF animals (e.g., Crumeyrolle-Arias et al. Reference Crumeyrolle-Arias, Jaglin, Bruneau, Vancassel, Cardona, Daugé, Naudon and Rabot2014; Gareau et al. Reference Gareau, Wine, Rodrigues, Cho, Whary, Philpott, Macqueen and Sherman2011; Sudo et al. Reference Sudo, Chida, Aiba, Sonoda, Oyama, Yu, Kubo and Koga2004). Sometimes a rescue experiment is performed in which a standard microbiota is transplanted into GF animals to investigate whether the phenotype can be reversed (Clarke et al. Reference Clarke, Grenham, Scully, Fitzgerald, Moloney, Shanahan, Dinan and Cryan2013; Diaz Heijtz et al. Reference Diaz Heijtz, Wang, Anuar, Qian, Björkholm, Samuelsson, Hibberd, Forssberg and Pettersson2011; Neufeld et al. Reference Neufeld, Kang, Bienenstock and Foster2011a; Reference Neufeld, Kang, Bienenstock and Foster2011b).

2. 2. Studies of normally colonized mice treated with antibiotics (Ait-Belgnaoui et al. Reference Ait-Belgnaoui, Durand, Cartier, Chaumaz, Eutamene, Ferrier, Houdeau, Fioramonti, Bueno and Theodorou2012; Bajaj et al. Reference Bajaj, Heuman, Sanyal, Hylemon, Sterling, Stravitz, Fuchs, Ridlon, Daita, Monteith, Noble, White, Fisher, Sikaroodi, Rangwala and Gillevet2013; Bercik et al. Reference Bercik, Denou, Collins, Jackson, Lu, Jury, Deng, Blennerhassett, Macri, McCoy, Verdu and Collins2011a; Desbonnet et al. Reference Desbonnet, Clarke, Traplin, O'Sullivan, Crispie, Moloney, Cotter, Dinan and Cryan2015). One study in our sample then re-colonized the animals with microbiota from obese and normal hosts (Bruce-Keller et al. Reference Bruce-Keller, Salbaum, Luo, Blanchard, Taylor, Welsh and Berthoud2015).

3. 3. Studies in which probiotics and placebos are given to human or other animal subjects (Supplementary Table 2).

4. 4. Standard microbiota studies that assess the experimental alteration of gut microbes (Supplementary Table 3). Some older methods are still used to describe the microbial community, such as denaturing gel electrophoresis (DGGE) or terminal restriction fragment length polymorphism (T-RFLP). But at least some MGB researchers are now turning to more contemporary methods such as quantitative polymerase chain reaction (qPCR), which is an amplification method that targets specific molecules and thus selected taxa, or shotgun DNA sequencing that encompasses the whole community.

For most of the interventions in the third category of “microbiota” methods (probiotics), Bifidobacterium sp. and Lactobacillus sp. are the probiotics of choice, with Lactobacillus helveticus being the most popular (Supplementary Table 2). These genera of organisms have long been traditional targets for claims about fermented milk products having digestive and physiological benefits (e.g., Metchnikoff Reference Metchnikoff1908). B. fragilis, the intervention microorganism in Hsiao et al.’s (Reference Hsiao, McBride, Hsien, Sharon, Hyde, McCue, Codelli, Chow, Reisman, Petrosino, Patterson and Mazmanian2013) study, is not found in fermented milk products but can be deployed according to the World Health Organization (WHO) definition of a probiotic: any live microorganism that is used to intervene in a human body to bring about health effects (Hill et al. Reference Hill, Guarner, Reid, Gibson, Merenstein, Pot, Morelli, Canani, Flint, Salminen, Calder and Sanders2014; however, see Shanahan & Quigley [Reference Shanahan and Quigley2014] for conceptual concerns). We will come back to probiotics and their implications in section 7.

An important observation to make here is that treatment with single or multiple probiotics is not strictly a “microbiota” or “microbiome” study. Normally, this term is reserved for studies in which microbiota samples are analysed bioinformatically after sequencing. In MGB probiotic research, however, researchers might not even profile changes in bacterial composition, and when they do, no differences may be observed (e.g., Tillisch et al. Reference Tillisch, Labus, Kilpatrick, Jiang, Stains, Ebrat, Guyonnet, Legrain-Raspaud, Trotin, Naliboff and Mayer2013). Surprisingly, even when microbiota are analysed for changes, very limited microbiome methodology is used (Supplementary Table 3). The methods that are employed are often not state-of-the-art. It is curious indeed to see much older qualitative methods, such as DGGE, being used for a publication in 2013 (Park et al. Reference Park, Collins, Blennerhassett, Ghia, Verdu, Bercik and Collins2013). Although a useful tool in the 1990s, community fingerprinting methods like DGGE and T-RFLP have long been superseded by more advanced quantitative sequencing methods. These newer methods allow closer analysis of the composition and potential function of microbial communities.

It is important to note, however, that microbiome research in general continues to have a “causality problem” despite improved sequence analysis tools (Hanage Reference Hanage2014). Many microbiome studies simply cannot isolate specific causes no matter how sophisticated their sequencing and bioinformatic tools; even the experimental work with microbiota transplants is not adequate to demonstrate whole-microbiome causality (O'Malley & Skillings Reference O'Malley and Skillings2018; see sect. 7). In this regard, MGB studies may have an advantage, in that they focus on single microorganisms (probiotics) or small groups of microbes that can be manipulated. However, a probiotic focus would not normally license claims about the whole microbiome, and even narrow probiotic causal claims are problematic (see sect. 7).

A standard interpretation in MGB research is to attribute differences in behaviour between GF and non-GF animals to the lack of microbiota in the former (ditto for antibiotic interventions, which deplete but do not fully remove the microbiota). Often the different treatments experienced by GF or antibiotic-treated mice are not remarked on. Few studies in our most cited sample provide controls that would enable singling out the effects of the microbiota itself (e.g., rescue of phenotype by re-infecting GF animals with a full community transplant or by reintroducing specific bacteria). Although GF models have yielded many interesting results, questions continue to be asked about how relevant they are to humans (Nguyen et al. Reference Nguyen, Vieira-Silva, Liston and Raes2015), because very few humans ever experience germ-free conditions. Although sometimes GF status is equated with environments that have high levels of hygiene and multiple antibiotic treatments (e.g., neonatal care facilities; see Clarke et al. Reference Clarke, Grenham, Scully, Fitzgerald, Moloney, Shanahan, Dinan and Cryan2013), for the majority of researchers these are not considered equivalent conditions at all.

Overall, there are very few studies in this highly cited group of papers that have an experimental approach genuinely able to demonstrate the impact of the microbiota itself on behaviour. Correlations are loosely interpreted as indications of potential mechanisms (however, see Bajaj et al. [Reference Bajaj, Heuman, Sanyal, Hylemon, Sterling, Stravitz, Fuchs, Ridlon, Daita, Monteith, Noble, White, Fisher, Sikaroodi, Rangwala and Gillevet2013] for a more sophisticated analysis of correlation networks of microbial metabolites). The conditions under which potential mechanisms might operate are not specified. For example, one study postulates “the existence of a gut–brain axis in alcohol dependence, in which the gut microbiota could alter the gut-barrier function and influence behavior in alcohol dependence” (Leclercq et al. Reference Leclercq, Matamoros, Cani, Neyrinck, Jamar, Stärkel, Windey, Tremaroli, Bäckhed, Verbeke, de Timary and Delzenne2014, p. E4491). Yet all that this particular piece of research demonstrates is a correlation between increased intestinal permeability and certain bacterial taxa. Less cited and newer studies may be making greater efforts to show microbiota causality of behaviour and brain function (see sect. 8), but, in general, invoking the whole microbiome, rather than specific members of it, will require methods that are carefully designed to deal with the complexities of thousands of interacting organisms and pathways.

One consequence of this complexity is that inter-individual variability between human microbiomes is so high that it is impossible – given most clinical sampling practices – to distinguish specific groups of patients or animals and to find the taxa most associated with different health states (e.g., Falony et al. Reference Falony, Joossens, Vieira-Silva, Wang, Darzi, Faust, Kurilshikov, Bonder, Valles-Colomer, Vandeputte, Tito, Chaffron, Rymenans, Verspecht, De Sutter, Lima-Mendez, D'Hoe, Jonckheere, Homola, Garcia, Tigchelaar, Eeckhaudt, Fu, Henckaerts, Zhernakova, Wijmenga and Raes2016). Frequently, when differences in bacterial composition are observed in the broader body of MGB literature, they are simple correlations from single studies rather than multiple comparative analyses. Considering that hundreds of taxa are involved in any gut community, it is not surprising that some correlations are found. The broader microbiome field (outside MGB) uses a range of statistical correction measures, and their implementation – although still imperfect – at least reduces gross false discovery rates (Knight et al. Reference Knight, Vrbanac, Taylor, Aksenov, Callewaert, Debelius, Gonzalez, Kosciolek, McCall, McDonald, Melnik, Morton, Navas, Quinn, Sanders, Swafford, Thompson, Tripathi, Xu, Zaneveld, Zhu, Caporaso and Dorrestein2018; Weiss et al. Reference Weiss, Van Treuren, Lozupone, Faust, Friedman, Deng, Xia, Xu, Ursell, Alm, Birmingham, Cram, Fuhrman, Raes, Sun, Zhou and Knight2016). For example, one of the reasons that standard parametric tests are not adapted to microbiome data is because of the issue of compositionality. Rapid changes to any single taxon in the microbiota are often measured as changes to all of the taxa, instead of reflecting true abundances. This property leads to extremely high false discovery rates. These ongoing issues add to the field's struggles to achieve causal explanations of phenomena such as disease, but their incidence in MGB research is exacerbated by weaknesses in the methods that are used in combination with microbiome analyses.

6. Neuroendocrine, behavioural, and statistical tests

Microbiome research in its standard sense (i.e., the sequencing and bioinformatic analysis of community genomes) might inform only a subset of MGB papers, and even when it is carried out, it is unlikely to be the methodological focus. Most of the methodology is in fact centred on rodent hormones and behaviour in different conditions. We divided the 25 most cited MGB papers into five categories according to their research focus relative to hormones and behaviour: (1) neuroendocrine “stress” axis; (2) emotion-mood: anxiety; (3) mood disorder: depression; (4) autism spectrum/developmental disorders; and (5) cognition (see Supplementary Tables 4a–4e). About half of the 25 top-cited papers are concerned with activation of the so-called neuroendocrine “stress” axis, which results in the production of stress-related glucocorticoid hormones (Supplementary Table 4a). All of these studies, save one, describe experimental work done in rodents. Sixteen of the top 25 papers explore anxiety, of which 13 studies were carried out on rodents (Supplementary Table 4b). A little less than a quarter (6) of the articles are related to depression, with the majority of that work being done in humans (Supplementary Table 4c). Only two studies present work on animal models of autism spectrum disorder (Supplementary Table 4d), and six address different forms of cognition (Supplementary Table 4e).

Most of the studies we examined do not explicitly justify their methodologies. They seldom address potentially confounding effects (e.g., maternal separation and water avoidance stress) that may complicate interpretation and limit the generalizability of findings. The adequacy of particular behavioural tests and measures is rarely discussed and seems to be taken for granted (admittedly because many other studies have done so). For example, following Sudo et al.’s initial Reference Sudo, Chida, Aiba, Sonoda, Oyama, Yu, Kubo and Koga2004 work, about half of the papers in our top-cited sample measure corticosterone in relation to gut microbiota in rodents. Although most of this subset of papers examines corticosterone in the context of stress – a framework laid down by formative research published 60 years ago (Eik-Nes & Samuels Reference Eik-Nes and Samuels1958; Gold et al. Reference Gold, Singleton, Macfarlane and Moore1958; Persky et al. Reference Persky, Hamburg, Basowitz, Grinker, Sabshin, Korchin, Herz, Board and Heath1958) – it is worth recalling that non-stressful events, such as meal consumption, also increase the circulating concentration of this glucocorticoid (Toda et al. Reference Toda, Morimoto, Nagasawa and Kitamura2004; Wang et al. Reference Wang, Dourmashkin, Yun and Leibowitz1999). Adrenalin can equally be considered a stress hormone (Mormède et al. Reference Mormède, Andanson, Aupérin, Beerda, Guémené, Malmkvist, Manteca, Manteuffel, Prunet, van Reenen, Richard and Veissier2007). In other words, there can be confounding factors at play in any observation of stress responses.

The appropriateness of animal models for human disease is seldom argued for and yet is of crucial importance for the implications of these studies. Not only do mice and humans have different gut structure and neuroanatomy, different microbiota, and different evolved behaviours (see Arrieta et al. Reference Arrieta, Walter and Finlay2016; Nguyen et al. Reference Nguyen, Vieira-Silva, Liston and Raes2015), but there are also acute problems of “translation” into clinical practice when it comes to claims about stress, anxiety, and depression. Behaviours that may be normal for mice (e.g., fearfulness and timidity) are not normal or desirable for humans, and vice versa. Moreover, no self-report-based evaluations can be made on rodents to gain better insight into the organism's experience. Although terminology about findings related to disorders is generally appropriate in the 25 papers we examined most closely (e.g., “anxiety-like” and “depression-like”), we nevertheless found several instances of terms for multidimensional human disorders (e.g., “anxiety” and “depression”) being applied to the unidimensional rodent results (see Supplementary Tables 5a, 5b).

Translational issues arise in any research that extrapolates from rodent models to human function (Zeiss & Johnson Reference Zeiss and Johnson2017) but are particularly pertinent to neuropsychiatric disorders (Homberg Reference Homberg2013). In rodent behavioural studies, interpretations of results obtained in the open field, elevated plus maze, light-dark box, and forced swim tests have frequently been criticized. Indeed, some critical reviews recommend finding new animal paradigms to investigate anxiety (Belzung & Griebel Reference Belzung and Griebel2001). Some authors go so far as to say that “evidence in support of the validity of the plus-maze, the light/dark box and the open-field as anxiety tests is poor and methodologically questionable” (Ennaceur Reference Ennaceur2014, p. 55). Other authors consider increased immobility in the forced swim test an adaptive passive coping strategy rather than a measure of the behavioural despair that is indicative of human depression-like behaviour (Commons et al. Reference Commons, Cholanians, Babb and Ehlinger2017; Molendijk & de Kloet Reference Molendijk and de Kloet2015).

When articles from our 25 most cited papers do take notice of translational issues, they may not take them seriously. For example, Hsiao et al. (Reference Hsiao, McBride, Hsien, Sharon, Hyde, McCue, Codelli, Chow, Reisman, Petrosino, Patterson and Mazmanian2013, p. 1456) quote Bourin et al. (Reference Bourin, Petit-Demoulière, Dhonnchadha and Hascöet2007) as saying that “mapping an animal's movement in an open arena” allows researchers “to measure … anxiety.” Crucially, however, Bourin and colleagues are arguing that it is important to specify whether the open field test is used under dimly lit conditions to measure mere locomotor activity, or whether implementing it in bright light is testing innate rodent anxiety of open spaces during the day. Bourin et al. (contra Hsiao et al.’s interpretation) go on to urge caution about interpreting findings as having implications for anxiety disorders (Bourin et al. Reference Bourin, Petit-Demoulière, Dhonnchadha and Hascöet2007). In the broader MGB field (i.e., beyond the top-cited papers), there are some examples of researchers supplementing or changing their reliance on the open field and elevated maze plus tests (e.g., Bassi et al. Reference Bassi, Kanashiro, Santin, de Souza, Nobre and Coimbra2012; Goehler et al. Reference Goehler, Park, Opitz, Lyte and Gaykema2008), in order to avoid the confounding of anxiety-like behaviour with simple alterations in locomotor activity patterns (Swiergiel & Dunn Reference Swiergiel and Dunn2007). Most commonly, however, if mentioning these issues, MGB researchers merely note them and then very pragmatically continue with animal model manipulations and interpretations.

To conclude our methodological analysis, there are reasons to think that the statistical analyses carried out by some MGB studies in our most cited sample are not appropriate (see Supplementary Table 6). In particular, one-way analyses of variance (ANOVAs) or Student's t-tests are frequently employed when the experimental design includes more than one independent variable. In such cases (e.g., Ait-Belgnaoui et al. Reference Ait-Belgnaoui, Durand, Cartier, Chaumaz, Eutamene, Ferrier, Houdeau, Fioramonti, Bueno and Theodorou2012; Reference Ait-Belgnaoui, Colom, Braniste, Ramalho, Marrot, Cartier, Houdeau, Theodorou and Tompkins2014; Ohland et al. Reference Ohland, Kish, Bell, Thiesen, Hotte, Pankiv and Madsen2013), two- or three-way ANOVAs are required. In many biological situations, the effect of one factor on an outcome of interest often depends on other factors. Hence, when two or more independent variables or factors (such as microbiota status and stress) are studied, it is important to address both the effects of those factors independently and their interaction with the dependent variable being measured (e.g., behaviour in a specific test). Several of the 25 most cited papers did not do this (Supplementary Table 6). Finally, in a few of the MGB papers we analysed, statistically negative results (P > 0.10) are presented as if they are positive findings. For example, non-significant findings after intervention strategies on the microbiota are still used to argue for potential microbiome effects (see Bailey et al. Reference Bailey, Dowd, Galley, Hufnagle, Allen and Lyte2011; Bravo et al. Reference Bravo, Forsythe, Chew, Escaravage, Savignac, Dinan, Bienenstock and Cryan2011; Tillisch et al. Reference Tillisch, Labus, Kilpatrick, Jiang, Stains, Ebrat, Guyonnet, Legrain-Raspaud, Trotin, Naliboff and Mayer2013). It would be much more straightforward to say “no effect is found” without assuming other methods or future experiments on larger cohorts will find the desired outcomes.

Following Fisher, it is standard in the life sciences to consider P < 0.05 as statistically significant and, conversely, that P > 0.05 indicates a non-significant difference (Habibzadeh Reference Habibzadeh2013). In this context, it is not possible to talk about “marginally significant” or “partially significant” (Habibzadeh Reference Habibzadeh2013), or as noted previously, “potentially significant.” At best, a statistical trend can be inferred when 0.10 < P < 0.05, provided there is sufficient statistical power. But if anything, studies in the life sciences tend to be underpowered, which has led several authors to make a plea for the use of more stringent cut-offs for P values and to consider only P < 0.01 as statistically significant (e.g., Colquhoun Reference Colquhoun2014; Vidgen & Yasseri Reference Vidgen and Yasseri2016). MGB research has yet to reflect on this advice.

These behavioural and statistical testing problems are by no means exclusive to MGB research. In fact, they are common throughout rodent-based behavioural neuroscience (Button et al. Reference Button, Ioannidis, Mokrysz, Nosek, Flint, Robinson and Munafò2013). But in MGB research, these weaknesses are compounded by the fact that it is misleading in some of the papers even to refer to microbiomes because no such analysis is done. Even when it is, superseded methods are providing very low-quality analyses. It is difficult of course to do everything well in interdisciplinary research, but, in some instances, it seems as if MGB papers are simply invoking the term “microbiome” without appreciating the minimal methodological commitments with which the term is normally accompanied.

7. Strong claims and interpretations

Although many of our 25 most cited papers use fairly basic reasoning, with limited mechanistic detail, they do not by and large indulge in the overinterpretation and overstatement to the extent we found in some of the broader MGB research literature. However, both our smaller sample of top-cited papers and the larger body of literature we examined divulge many examples of papers in which strong claims – such as “conclusively demonstrate” and “conclusive proof” (e.g., Ait-Belgnaoui et al. Reference Ait-Belgnaoui, Colom, Braniste, Ramalho, Marrot, Cartier, Houdeau, Theodorou and Tompkins2014; Bravo et al. Reference Bravo, Forsythe, Chew, Escaravage, Savignac, Dinan, Bienenstock and Cryan2011) – are offset by more conservative elaborations, sometimes in the very same paper (e.g., Christian et al. Reference Christian, Galley, Hade, Schoppe-Sullivan, Kamp Dush and Bailey2015; Foster & McVey Neufeld Reference Foster and McVey Neufeld2013). We are tempted to diagnose this as a case of “double-dipping,” when cautionary statements are belied by much more dramatic claims. We believe this strategy influences the public uptake of MGB research. In the following sub-sections, we discuss a selection of the overblown conclusions or speculations that help inflame the field, from the most abstract to the highly practical. We do this in order to show how misinterpretation may arise and propagate, especially in the review papers that are so dominant in MGB literature (between 40% and 50%; see sect. 1 in the Supplementary Material).

8. Summarizing our findings

To its credit, MGB research is driven by hypothesis testing, but it mostly proposes and confirms loose conjectures about microbial involvement in brain and behavioural states. Microbiome research (outside MGB) is very technology driven and often fishes around after analysis for some sort of hypothesis that might reasonably be based on the data. Neither extreme of this continuum of practice is desirable for the maturation of microbiome research. In fact, we could see in MGB research the potential to integrate and balance these two ways of doing science. Very importantly, this merger would bring more microbiome depth to MGB research, which our analysis shows is missing and misunderstood.

We also showed how MGB research has many other compounding methodological and interpretive issues. But might all the issues we have identified just be signs of a young field? Will it not get better all of its own accord, given enough time? We agree it is important not to inhibit new approaches as they develop. But a strong foundation seems important for future development, rather than ongoing reproduction of a rough-and-ready approach. We have taken a critical approach to this emerging field, partly because we see the same claims repeated over and over again. They achieve a wider reach with every iteration. Using evolutionary, ecological, microbiological, neurological, immunological, biochemical, genetic, molecular, and developmental perspectives to bolster a narrow band of results both overreaches and also displays limited acquaintance with some of the well-established knowledge in these fields. These limitations matter not only for the future of a field, but also for the status of scientific activity in these challenging times. As we suggest and others have argued (e.g., Hanage Reference Hanage2014; Perez-Muñoz et al. Reference Perez-Muñoz, Arrieta, Ramer-Tait and Walter2017), overblown claims damage the credibility of the field and cause harm to the general social reception of science.

A topic worthy of further social scientific investigation is why microbiome research in general is so popular with the public, and whether public perspectives on microbiome research are changing how people think about health, including mental health. We speculate that reasons for the public uptake of microbiome research findings, including MGB, are to do with its perceived “naturalness” and the “holism” of the science, as well as the strong potential for microbiome-related therapies to be self-administered and even “DIY” rather than imposed by technical experts. There are many good aspects to any such trends. But MGB research should be aware of these tendencies and their possible relationships with anti-scientific claims (e.g., anti-vaccination; anti-psychotropic medication). It could be well worth working with relevant public health and media experts on how to communicate this exciting body of work responsibly.

9. Conclusions and future directions

Despite the critical picture we have painted, we see MGB research as a field full of promise, with important implications for understanding the relationship between the brain and the rest of the body. Existing MGB findings point to an ongoing need for more connected research that is able to investigate the complex interactions occurring in multipathway systems. Expecting cure-alls to emerge from these early days in which the puzzle pieces have barely been recognized, let alone joined up, seems contradictory to the spirit we assume to be motivating MGB inquiry. Our findings indicate the tension between a field-wide recognition of complex networks of causes and effects versus expectations of a simple all-efficacious treatment. As we noted in the introduction, our critical overview of MGB research is from outside the field itself and does not presume it can provide the detailed advice necessary to lead the field forward. This has to come from within the field. Nevertheless, we can use our findings of the current state of the field to propose some general pointers about how the field might develop and what it should avoid in that development.