The relationship between dietary compounds and colonic microbiota has garnered scientific interest due to its impact on host health (1, 2). From a health perspective, changes in colonic microbial function are more relevant than alterations in composition. Imbalances in microbial metabolite production may distinguish individuals with disease from healthy controls, despite individual variations in microbial taxonomy (3–6). Numerous metabolites play crucial roles in the bidirectional communication between the colonic microbiota and the host, such as neurotransmitters, polyamines, vitamins, bile acids, and organic acids (7). Among them, the short-chain fatty acids (SCFAs), acetate, propionate, and butyrate, offer numerous benefits to the host, including maintaining colonic barrier integrity, serving as an energy source for colonocytes, and exerting neuroprotective effects (8–11).

Most of our current understanding about diet-colonic microbiota interactions centers on the effects of complex carbohydrates on the adult microbiota. In contrast, the impact of other dietary compounds, such as fatty acids and polyphenols, remains underexplored, particularly in underrepresented populations such as infants and older adults (12). A better understanding of how dietary compounds affect colonic microbial function across diverse human populations is crucial. Clinical trials are standard approaches for evaluating this impact and assessing potential health outcomes from diet-microbiota interactions. However, they are time- and resource-consuming, and often challenged by technical and ethical concerns (13). Given these limitations, mathematical models show promise as complementary tools to reduce the cost and time of microbiota investigations (14, 15).

Various models have been proposed to investigate diet-colonic microbiota interactions, including kinetic-based (16–19), agent-based (20–22), and genome-scale metabolic models (GEMs) (23–26). GEMs use metabolic reconstructions, a mathematical representation of a microorganism’s metabolism, and flux balance analysis (FBA) to predict microbial metabolite production as fluxes (units of concentration per time) (27). Metagenome-scale community metabolic models (MGCMs) extend this concept to microbial communities (14, 15, 28). Among MGCMs, the Microbial Community model (MICOM) stands out for its user-friendly approach, extensive documentation, and pre-made workflows that range from data preparation to visualization (14, 29).

However, experimentally validating the predictions of MGCMs is challenging (30). A recent study compared measured SCFA fluxes from ex vivo fecal incubations with predicted fluxes obtained from MICOM under the influence of dietary fibers (31). An agreement between predictions and experimental measurements was reported for propionate and butyrate. Nevertheless, while this study partially validated in silico predictions for the influence of isolated dietary fibers, the model’s accuracy in predicting how whole foods shape colonic microbial function remains unexplored. Importantly, foods contain multiple dietary compounds that can interact during digestion and modulate their collective impact on colonic microbes (32, 33).

This study aimed to evaluate MICOM’s accuracy in predicting microbial SCFA production in real-life feeding scenarios for weaning infants. Predicted acetate, propionate, and butyrate fluxes were compared with fluxes measured experimentally from a published in vitro fecal fermentation study (34). The in silico simulations were designed to match the experimental setup as closely as possible, which examined how complementary foods, alone or combined with infant formula and/or other foods, affected major SCFA production by the colonic microbiota of weaning infants.

This study compared the predicted production of acetate, propionate, and butyrate in silico using MICOM (14) with experimental values obtained from a fecal fermentation study (34). Our focus on assessing the model’s accuracy in predicting the production of health-relevant metabolites, rather than microbial growth, was driven by the lack of available data on in vitro individual microbial growth rates. Additionally, from a health perspective, microbial function is more informative than composition, as colonic microbes are functionally redundant, and the colonic microbiota of healthy individuals maintains similar functionality despite taxonomic differences (3, 6). The workflows for in silico simulations and statistical comparison between in silico and in vitro outcomes followed published protocols (31, 37). Another strength was the assessment of the model’s accuracy in predicting SCFA production by the colonic microbiota of weaning infants under realistic infant feeding patterns, including whole foods, foods combined with other foods, infant formula, or both.

When evaluating all samples, the model demonstrated poor accuracy, with only acetate showing a weak correlation between predicted and measured outcomes. Interestingly, we observed increased accuracy when analyzing samples predominantly composed of plant-based foods. Acetate exhibited a moderate positive correlation, while propionate and butyrate showed a trend toward weak positive and weak negative correlations, respectively. These findings suggest that MICOM’s accuracy in predicting the function of the infant microbiota increases for media enriched in complex carbohydrates. This result is expected, considering that the influence of complex carbohydrates on colonic microbes is far better understood than that of other dietary compounds. Consequently, metabolic reconstructions of colonic microbes are often not curated for biochemical reactions involving amino acid and lipid utilization due to limited available data (51). Additionally, this result is likely driven by the functional capacity of the colonic microbiota of weaning infants. Observational studies have demonstrated that the weaning infant microbiota transitions from primarily degrading human milk oligosaccharides to metabolizing complex carbohydrates, while amino acid fermentation increases but remains less prominent compared to that of the adult colonic microbiota (52, 53).

Given that correlation analyses assess the strength of the relationship between two variables but do not indicate how these variables differ from each other, a Bland–Altman analysis was conducted to determine the limits of agreement (95% confidence interval) between predicted and measured SCFA production. Standard scores were used to account for differences in magnitude and unit between methods (54). Acetate had lower limits of agreement than propionate and butyrate, suggesting stronger concordance between predicted and measured z-scores, which aligns with the correlation results. Overall, most food and food-food combination samples fell within the 95% limits of agreement. However, some exceptions highlighted the model’s limited ability to accurately predict SCFA experimental production across the entire dataset.

Among food ingredients and food-food combinations, disagreements were observed between in vitro and in silico outcomes. For example, strawberries and the combination black beans-blackcurrants had positive z-scores for total SCFA production in vitro but negative z-scores in silico. On the other hand, some samples that drove the greatest total SCFA production in vitro, such as blackcurrants, raspberries, and the blackcurrants-strawberries combination, also resulted in high positive z-scores in silico. Sensitivity analyses assessing the impact of microbial composition on prediction accuracy indicated that community composition strongly influences simulation outcomes. However, the model consistently showed limited predictive performance across SCFAs and samples, with acetate showing better agreement in carbohydrate-rich samples. As a take-home message, our study demonstrates that emerging modeling approaches for diet-colonic microbiota interactions are imperfect and should not replace experimental methods. Instead, given their cost- and time-efficiency, and ability to leverage existing data, these models offer a valuable starting point for generating insights that can guide the design of in vitro and in vivo studies. Their further refinement and use as complementary tools represent a promising opportunity to advance diet-colonic microbiota research.

Future directions for the development of MGCMs include expanding the number of high-quality microbial metabolic reconstructions, which should be built using sequencing data that meet quality standards and subsequently curated with experimental data (55–57). Shotgun metagenomic sequencing data is preferable to 16S rRNA sequencing data, as it more accurately captures the metabolic potential of colonic microbes (58). Importantly, experimental data remain crucial for model curation and validation (59), with a notable need for more research into the behavior of colonic microorganisms in response to various dietary compounds. The accuracy of MGCMs could also benefit from incorporating omics data, such as metatranscriptomics, to better personalize the model’s conditions (60). Additionally, integrating dynamic FBA could improve the representation of changes in the colonic environment over time (24, 61). Finally, databases used for designing media in in silico simulations require further refinement to better account for the heterogeneity of dietary patterns across individuals.

Our results partially contrast with a previous study that reported agreement between MICOM’s prediction for propionate and butyrate and experimental outcomes from ex vivo fecal incubations with isolated dietary fibers using adult inoculum (31). The modest performance observed here is likely due to the different study designs. The fecal microbiota of weaning infants has distinct functionality compared to the adult microbiota, including a higher proportional production of acetate and lower of propionate and butyrate (62, 63). Furthermore, whole foods are complex matrices containing not only fiber but also other compounds, such as protein, fat, and phytochemicals, all of which impact the function of colonic microbes (64–66). Finally, while the referenced study used metagenomic data to build models at the species level and normalized SCFA predictions based on microbial biomass (31), our simulations were limited to the genus rank and did not incorporate biomass normalization due to using 16S rRNA sequencing data, thereby reducing metabolic specificity. This contrasting result highlights the need for further investigations of the model predictive accuracy across different study conditions and host populations, particularly concerning the production of health-relevant metabolites like butyrate (8, 9).

A major limitation of our study is the difference in outcomes obtained from distinct methods, as fluxes predicted in silico are not equivalent to the concentration of metabolites measured in vitro. To address this limitation, we followed a published protocol that used standard scores to compare results from different methodologies and validate MICOM’s predictions of propionate and butyrate production (31). However, quantifying SCFA fluxes using static in vitro fermentation is a limited representation, as metabolites accumulate and substrates are depleted over time (67). Over long fermentation periods, these conditions diverge from the dynamic conditions and steady state assumption inherent in FBA-based models, which assumes no accumulation of substrates in the system (27).

Additionally, the accuracy of models based on FBA strongly depends on the quality of the metabolic reconstructions used in the simulations. A recent systematic evaluation of FBA-based tools, including MICOM, reported low prediction accuracy of microbial growth rates compared to experimental data when using the AGORA metabolic reconstructions (68). The authors highlighted that semi-curated metabolic reconstructions were not sufficiently accurate for predicting the behavior of microorganisms (68). Our simulations were performed using the second version of AGORA, which was generated via a semi-automated pipeline and manually refined (36). However, it is important to recognize that these reconstructions may contain gaps or inaccurately assigned biochemical reactions (51).

Another limitation was the use of pooled fecal data, justified by the absence of individual-level data in the in vitro fermentation study. Although pooling fecal samples is a common practice in such studies, it inevitably alters the original microbial community structure, reducing inter-individual variability and functional resolution. These changes are likely to influence microbial functional dynamics in ways that are not captured by the taxonomic profile of the pooled sample (69, 70). Using amplicon 16S sequencing data limited model construction to the genus level, potentially masking metabolic differences between microbial species (31, 71). Additionally, this approach reduced the number of taxa included in the simulations and hindered the normalization of SCFA predictions by bacterial biomass. Notably, bacterial species that are present in low abundance yet biologically meaningful (keystone species) may be underrepresented when using 16S rRNA sequencing data (72, 73). Moreover, relying on microbial relative abundances as a proxy for absolute quantities, which can be biased by interdependencies between taxa and may not accurately reflect the true composition of the microbial community (74), can influence the model predictions. To overcome this limitation, simulations should be based on absolute microbial biomass, estimated via metagenomics or sequencing combined with qPCR (75, 76).

Furthermore, the in silico media were designed using the Virtual Metabolic Human database (38), which lacks information on diverse cooking methods and food ingredients. Notably, indigenous foods available in New Zealand and used in the experimental work (34), such as kūmara (sweet potato variety), were not included in the database and had to be substituted with the most similar available food. Finally, while the in vitro study used a static protocol to mimic the digestion and absorption of dietary compounds (34), MICOM, like other diet-microbiota models, currently does not account for digestion. Instead, it represents absorption by diluting the dietary compounds identified to be absorbed by the human intestine using a scalar (14). Not accounting for digestion is a potential major source of variation between in silico and in vitro outcomes, as digestion influences food structure and, consequently, the nutrient profile accessible for microbial fermentation (77). However, this limitation is intrinsic to the modeling framework, as computationally simulating food digestion is challenging. Although promising tools are emerging (78), they have not yet been integrated into MGCMs.

This study evaluated the accuracy of the metagenome-scale community metabolic model MICOM in predicting SCFA production by the fecal microbiota of weaning infants under realistic complementary infant feeding patterns, using static in vitro fecal fermentation data as a comparator. A weak positive correlation was observed between predicted and measured acetate production. The agreement between predicted and measured SCFA fluxes improved when analyzing samples predominantly composed of plant-based foods. These findings suggest that the model more accurately replicates experimental results when simulating media rich in complex carbohydrates. Despite disagreements between experimental and simulated outcomes for specific SCFAs, the model identified samples with the highest total SCFA production in vitro. This exemplifies the model’s limitations as a replacement for traditional experimental methods but supports its potential as a complementary tool. Further model development is essential to improve its accuracy, particularly for media-rich in fat and protein. Refined versions of the model would contribute to advancing research on the relationship between diet and microbiota.