Symbioses are widespread in nature, and beneficial digestive-tract symbioses have been shown to be critical for host health [1]. The benefits and contributions of the gut microbiota include enhancement of digestion efficiency, provision of nutrients and vitamins, and procurement of digestive enzymes [2]. Members of the microbiome can contribute to host health by detoxifying allelochemicals from plants, such as tannins, flavonoids, and alkaloids, along with creating colony resistance against possible pathogens [3]. When the host feeds on a nutrient-poor diet, the reliance on the physiological capabilities of the microbes is even greater. Mammals that feed on a cellulose-rich diet, such as ruminant cows, require a gut bacterial community to generate energy for the host due to the inability of mammals to produce cellulases [4, 5]. In contrast, some insects can produce cellulases and sometimes harbor protist symbionts that are crucial to the breakdown of the wood meal [4]. These insects rely on the bacterial symbionts to provide a source of energy in the form of short-chained fatty acids (SCFAs) and for nutrients that are present at low amounts or are absent in plant food sources, such as nitrogen, amino acids, sterols, and many B vitamins [6, 7].

Insect-feeding habits have been shown to partially dictate the microorganisms present in the gut. Cockroaches fed a low-protein and high-fiber diet showed decreases of Streptococci and Lactobacilli in their gut, coinciding with the reduction of acetate and lactate production [8]. The gut of the American cockroach, Periplaneta americana, is populated by a higher abundance of protozoa when fed a high-cellulose diet [9]. The house cricket, Acheta domesticus, shows reduced production of H₂, CO₂, and SCFAs when fed a high-protein diet compared to other diets [10]. A comprehensive study on higher termites showed that diet plays a role in shaping the gut microbiota. Bacteria with the ability to degrade cellulose were observed to be present in higher abundances in wood-feeding termites compared to termites fed diets without cellulose [11]. In contrast, termites that feed on humus and soil have a more alkaline gut environment, and bacteria that live in more alkaline environments were shown to be more abundant in these termites [12]. Reticulitermes flavipes, the eastern subterranean termite, is a wood-feeding lower termite that harbors protist, bacterial, and archaeal symbionts. The protists are thought to aid in the breakdown of cellulose and lignocellulose, while the bacteria and archaea use the breakdown products to produce nutrients for the symbiotic community and the host [13, 14].

Although microbiomes are being extensively studied, temporal studies are limited in their scope, and predictive in silico modeling of microbiome dynamics is lacking. Only a few studies have attempted to model a microbiome, learning about the dynamics between members of the community and environmental factors [15]. Reasons for the lack of microbiome models include the inherent complexity of most communities, the computational difficulty of modeling highly nonlinear relationships, and the need to account for the effect of many external influences, such as substrate, temperature, pH, and micronutrient concentrations. One of the few studies performed in this area involved using an artificial neural network (ANN) with Bayesian network inference to predict the relative abundance of a microbial taxon in the English Channel as a function of its environment [16]. While this method was successful at modeling how the environment shapes the microbiota, it did not answer the question of how to identify important taxa or environmental factors once the dynamics are learned. Similarly, previous studies observed which organisms in the rhizosphere microbiome are important for disease protection in plants, but there is no current method to determine which taxa or environmental factors influence the growth or decline of these organisms based on a learned model [17]. Combining the qualitative knowledge of the bacterial members of a microbiome with quantitative in silico modeling of microbiomes is key for identifying influential organisms in the microbiome, as well as for learning how members of a microbiome work synergistically or antagonistically.

The hindgut microbiome of the lower termite, R. flavipes, is suitable for testing predictive models because of the detailed understanding of the community members. Furthermore, R. flavipes is capable of feeding on different types of wood and can be easily maintained in the lab [18]. An important aspect of the hindgut microbiome is that a large proportion of the taxa is consistently present in individual termites [18]. These taxa are considered to be part of the “core” community. In this work, the composition of the R. flavipes hindgut microbiota was monitored by sequencing the V4 region of the 16S rRNA gene over time following dietary changes. An algorithm inspired by Larsen et al. [16] used a deep artificial neural network (ANN) to learn the dynamics of the microbiota resulting from external influences, such as changes in the diet, and from changes in the relative abundance of the taxa in the microbiome observed in the 16S rRNA gene sequencing data. The ANN was then trained on this data, and a sensitivity analysis was performed to determine the accuracy of the model. When used in conjunction with microbial community analyses, the ANN-learned dynamics allowed for an in-depth analysis of the microbiome to understand taxon-taxon and taxon-substrate interactions.

Our study revealed that altering the food source of R. flavipes affected the overall composition of the hindgut microbiota without affecting the members of the core taxa. The MDS analysis revealed that after 7 days on a new diet, the overall community differed significantly from the community at day 0. This change in the community structure was driven by the less abundant microorganisms that were separate from the core microbiota. In comparison, the control group that was continuously fed mulch was not significantly different from day 0 after 7 days (f = 2.03, p = 0.029). A previous study fed R. flavipes either a wood or a paper substrate, sampled the hindgut microbiota after 7 days, and reported the similarities and differences between the hindgut bacterial communities [19]. They observed that termites from the same colony that were fed different diets were more similar to each other than termites from different colonies that were fed the same diet after 7 days. However, data for later time points were not reported in the study, time points at which changes in the hindgut community were observed in the present study. Our results are similar to those reported from wood-eating cockroaches and higher termites. In wood-eating cockroaches that are closely related to termites, the members of the core microbiota were shown to be stable during dietary changes [24]. In higher termites that do not harbor protist symbionts, it was shown that the dominant members of the hindgut microbiota remained stable and that only less abundant members were affected [12]. It is interesting that these three studies consistently determined that the core was not affected by dietary changes. This highlights the critical contributions of the individual taxa to the overall function of the hindgut microbiota and animal host. The most abundant organisms in a symbiotic habitat have been shown to perform important functions within the environment, essentially securing their constant presence in the environment [25]. A study of the gut microbiota of 37 adults over 5 years showed that approximately 60–70% of the strains remained stable throughout the study. Furthermore, the most stable organisms were also the most abundant, which supports our findings [26]. The compositional stability of the core microbiota is consistent with the idea that the host and its microbes form one functional unit upon which selection acts, as has been proposed in the holobiont theory [27–29]. In our study, mostly non-core members exhibited altered abundances and contributed to shifts in community structure when the food source was altered suggesting that these variable microbes play a critical role in expanding the holobionts capacity to occupy new niches.

The major exception to the stability of the core microbiota occurred in the group that was starved. Starved termites have been shown to lose their symbiotic hindgut protists, which are associated with Treponema, belonging to the order Spirochaetales, and Endomicrobia extracellular and intracellular symbionts, respectively. Thus, the observed decrease in the relative abundances of Treponema and Endomicrobia was expected. While the Treponema exist as both protist symbionts and free-living bacteria, the Endomicrobia are strict endosymbionts of the protists. It was therefore not surprising to observe the greater drop in the abundance of Endomicrobia in starved termites compared to Treponema [30, 31]. It was interesting to note that the Endomicrobia were most abundant in the spruce-fed termites, which was also observed in pine-fed termites by Huang et al., suggesting that this substrate may create a hindgut environment that enriches for the protist and Endomicrobia populations [20].

The ANN was used to evaluate the effect of a particular taxon or diet on the other taxa, determine the connectivity of a particular taxon to other taxa, and predict the composition of the community at any given point in time. A temporal comparison of the bacterial OTUs from the day 0 gut communities to end-point communities for each diet showed that dietary changes influenced the composition of the hindgut microbiota in R. flavipes (Fig. 1). Predictions from the in silico model corroborated these results, as the number or strength of correlations between substrate and taxa were greater in each diet (except maple) compared to the number/strength of correlations for the original mulch (diet of day 0 samples) (Fig. 5). Oak was the only substrate with significant correlations to the gut community, suggesting a disturbance to the gut microbes or the need for different metabolic strategies to breakdown its components. When grouped by order, the in silico model showed that Spirochaetales, the order to which the genus Treponema belongs, was the most connected order, with 17 total important correlations (Fig. 7b), while no correlations for Endomicrobia were observed. Treponema are acetogenic spirochetes that are hypothesized to contribute to the majority of the acetate production in the termite hindgut, which solidifies its importance to the community as acetate is a vital short chain fatty acid (SCFA) to other microbial members and the termite host [32, 33]. Termite gut Treponema live associated and disassociated with protists, and Endomicrobia are strict endosymbionts of the protists. If the protists and their endosymbionts are required at a constant level for the degradation of the lignocellulose, irrespective of its source, it may be that their abundance is decoupled from auxiliary changes in the community structure.

A powerful aspect of the neural network analysis was that after having trained the algorithm, it could be used to predict future values with relatively good robustness. This suggests that even for a community as complex as that found in the termite hindgut, and with a relatively sparse sampling frequency, the ANN was able to accurately predict the community composition. One limitation of this analysis is that due to complexity of the interactions, the ANN was performed on the level of taxonomic orders. It would be interesting to perform and evaluate the accuracy of this analysis on data from other time series, including humans. If future implementations are able to use lower taxonomic levels, the ANN could prove to be an important predictor of an unbalanced microbial community or dysbiotic state before it occurs. Understanding the effect of diet on a microbial population is valuable because it provides insight into the dynamics of the symbiotic niche. In a laboratory setting, it is necessary to consider the biological effect diet has on a host organism and its symbiotic bacteria. The ability to predict the taxonomic composition of a community is beneficial for forming hypotheses about an environment and can provide insight into the community dynamics within that environment.

Since the start of the high-throughput 16S sequencing revolution, scientists have reported on microbiomes, often focusing on the more abundant taxa or grouping the populations into phylotypes [34]. Many microbial community surveys draw focus to abundant organisms in order to generate conclusions on impact and importance of microbes. In order to view these bacterial populations in numerous samples, it has been standard protocol to show organisms at abundances greater than 1% [35–37], or group the low-abundant organisms into an “other” category [9, 38]. While this is a widely accepted method of determining correlations between healthy and diseased states and reporting microbial communities in general, researchers may be missing key organisms in the low-abundant population. The importance of low abundant organisms has been reported in other systems, such as Desulfosporinus in peatlands. Pester et al. reported finding Desulfosporinus at 0.006% of the 16S rRNA reads in peatland communities and determined that this small fraction of the community contributes significantly to the overall sulfate reduction [39]. Sogin et al. reports the importance of the “rare biosphere” in marine environments and explains the contribution of these low members to diversity and the gene pool [40]. Low-abundant microbes may contain a pool of genes, that under specific conditions become activated and carryout metabolic processes important to the overall community [41]. For example, the removal of rare microbes from freshwater samples resulted in the reduced ability of the microbes to neutralize toxins and pollutants [42], suggesting that the rare microbes can perform critical functions in an unfavorable environment. While some microbes are rare, they can contribute to the complete metabolic potential of the community if they are highly active, enhance or trigger the metabolic activity of more dominant members, or contain enzymes needed for complex metabolic processes that are not found in the dominant members [41, 43]. In this study, only four core taxa had greater than five significant correlations, while 13 non-core taxa had greater than five significant correlations (Fig. 7). Only five of the 15 taxa with significant correlations were abundant at greater than 1% of the community. This suggests that although the core and highly abundant taxa perform important functions for the community and are conserved throughout diet changes and time, the non-core and less abundant taxa may play important roles in shaping the gut community.

Gut microbiome dynamics is an important but challenging topic due to the taxonomic complexity of the community and wide range in abundance of its members. Using the complex microbiome of termites as a test case, we have developed a method using deep learning artificial neural networks for temporal taxon prediction based on environmental conditions and the original state of the microbiome. This method also allows the detection of connectivity of taxa with other taxa based on changes in the abundance of microbes during environmental changes. This deep learning approach revealed that low-abundant bacteria that often do not belong to the core community are drivers of the termite hindgut bacterial community composition.