Priyankar Dey ^1* Sandeep Choubey ²

• ¹ Thapar Institute of Engineering & Technology, Patiala, India
• ² Indian institute of mathematical sciences, Chennai, India

Our understanding of the role of gut microbes in human health and disease has come a long way since John M. Whipps and colleagues first defined the term 'microbiome'. Since early 2000s, with gradual lowering of the cost of commercial DNA sequencing, health science has been flooded with 16S rRNA data. Unfortunately, despite a plethora of preclinical and clinical publications on the gutliver axis, the majority of our understanding on the microbiota-liver reciprocal interaction basically remains limited to correlation analysis. The most exciting part of such metagenomic studies is the sheer amount of 'big data' generated which is rather easy to correlate with physiological variables; the bigger the metagenomic data and number of independent variables, the more chances to find 'significant' associations. Notably, our limited understanding on the gut-liver axis is derived from the microbiome, not the microbiota. Therefore, the conclusions obtained through microbiome-related correlation studies often do not reflect causation, are not representative of universal phenomenon, and there has been almost no true microbial markers of dysbiosis linked to chronic liver disease. Translational potentials of pre-clinical microbiota-liver associations in clinical disease prediction and treatment have not been much successful despite enormous numbers of interventional and observational studies. In fact, the relevance of utilizing rodents' gut microbial signatures in understanding human diseases has been criticized due to the massive difference in the gut microbes between both species, attributed to the gastrointestinal biogeography and genetic makeup (Nguyen et al., 2015). Especially in high-fat diet model of metabolic disease, there is lack of consistency between good and bad microbes. In fact, diseases have also been associated with the strains belonging to probiotics (e.g., Lactobacillus) and commensals (e.g., Akkermansia, Faecalibacterium) (Dey and Ray Chaudhuri, 2023). Added to this is the fact that confounding results are obtained from chemicalinduced metabolic disease models that are physiologically mostly irrelevant (Dey, 2020) and data from germ-free mice that possess innate defect in various physiological processes (Jans and Vereecke, 2024). Clinical progression of metabolic liver disease is distinct from pre-clinical models of chronic liver disease in terms of timeline, cellular phenotype, pathological complexity, the difference in immune responses, and metabolic machinery (Liu et al., 2013). Further, the innate and general differences between human and animal disease models such as biological differences (e.g., genetic makeup, organ anatomy, liver functional capacity), extent of disease complexity (e.g., model specificity, co-influence of comorbidities), ability to perform controlled experiments, predictive validity, etc. make it challenging to conclude on clinical gut microbial phenotypes based on pre-clinical data. Although there has been emerging reports of in vitro models of human distal intestine (Qi et al., 2023), these models are physiologically irrelevant given their host independent nature. Due to the advent of culturomics techniques and controlled clinical studies, pre-clinical gut microbial patterns that were initially considered associated with disease conditions, such as Firmicutesto-Bacteroides ratio, enrichment of energy-harvesting species, specific metabolic functions, seem to be falling apart. For instance, -proteobacteria due to presence of lipopolysaccharide (LPS) were previously thought to simply cause hepatic inflammation. However, studies have identified that LPS-TLR4-inflammation axis cannot be generalized due to huge difference in LPS structure dictating the extent of immune response (Picarello, 2022), and that majority of the luminal LPS-supplying Bacteroides, rather displays an immunosuppressive characteristics (d'Hennezel et al., 2017). Today, it is being strongly recognized that the good, bad, and ugly nature of the gut microbiota is conditionspecific. Factors such as the availability of preferred nutrients, pathoadaptive mutations, and potentials to evade the mucosal immune response has been recognized as critical factors that define a specific microbial species as commensal or pathobiont (Dey and Ray Chaudhuri, 2023;Dey, 2024). The bottom line is that there is no proper clinical definition of dysbiosis and eubiosis in terms of specific microbial features. However, the only one aspect almost universally accepted is that loss of gut microbial diversity is associated with chronic liver disease, and when we talk about the diversity, it's the community effects, not disease causation by a single species. A recent systematic review and meta-analysis of 54 clinical studies have indicated substantial inter-study heterogeneity in gut microbial taxonomic identification, in which the enrichment of inflammation-inducing genera were more closely associated with non-alcoholic fatty liver disease, but no genera were identified to provide long-term disease risk predictive value (Su et al., 2024). To date, there have been more than 100 cross-sectional studies to identify the predominant gut microbes associated with metabolic liver disease, yet no absolute core microbiome related to progressive liver disease has been identified. Although longitudinal studies are considered superior to cross-sectional studies, and that informed understanding on disease pathogenesis can only be obtained through the former one, there has been only a few longitudinal studies undertaken in order to link the gut microbiome with liver health. One study from Kyoto (Japan) evaluated the gut microbial alterations from pre-transplantation to two months post-surgery in 38 liver transplant patients (Kato et al., 2017). Data show initial decline and later increase of microbial diversity, along with overall increase of Bacteroides, Enterobacteriaceae, Streptococcaceae, and Bifidobacteriaceae, while a depletion in the abundance of Lactobacillaceae, Enterococcaceae, Clostridiaceae, Peptostreptococcaceae and Ruminococcaceae was noted. The authors acknowledged that variations in antibiotic regimes, food, synbiotics, and patient heterogeneity (e.g., various donors and underlying diseases) likely influence the study's findings. Confounding variables and data on relative microbial abundance were among the limitations. In line, a relatively recent study from National Institutes of Health, USA, investigated the gut-liver axis in hepatitis C patients, taking into account varied degrees of fibrosis severity (Ali et al., 2023). The investigation was done six months after HCV was undetectable (n=23) and before (n=29) attaining a durable virologic response. Data suggest that increased hepatic fibrosis was correlated with Anaerostipes hadrus, while Bacteroides vulgatus with portal inflammation in HCV. A prolonged virologic response suggest that Methanobrevibacter smithii may have a beneficial effect on indicators of the severity of liver disease. Although this longitudinal investigation made it possible to compare HCV-infected individuals with a supposedly improved viral clearance state, but they were unable to report gut microbial patterns in healthy controls due to unavailability of samples. Another recent study from Stanford examines the gut microbial composition at various body site (including gut) and correlated with host multi-omics, immunological, and clinical indicators (including hepatic) (n=86) over a 6-year period in order to comprehend the dynamic interaction between the human microbiomes and host throughout health and illness (Zhou et al., 2024). Despite, identifying microbial compositional and diversity patterns associated with host physiological parameters and metabolites, no temporal associations were derived between the gut microbiota and the measured liver-specific parameters (e.g., transaminases). Emerging studies although claims causative effects of oral microbiota in the pathogenesis of chronic metabolic disease, including liver disease (Gupta and Dey, 2023), there has been no longitudinal study undertaken in this line till date. Thus, understanding the true nature of gut microbial dynamics under the course of hepatic disease pathogenesis and remission remains largely unknown. Beyond availability of longitudinal gut microbial data, a fundamental difficulty in analyzing large-scale microbiome big-data lie in their high dimensionality (Advani and Ganguli, 2016). In classically designed experiments, a small number of carefully selected variables (V) are measured to test a specific hypothesis, with a large number of measurements (M) for each variable. Thus, the measurement density is very large (M/V → ∞). Such datasets are referred to as low-dimensional, and much of classical statistics operates within this framework. In contrast to this classical scenario, recent technological capacity for high-throughput sequencing has led to a different statistical regime. It is commonplace to simultaneously measure many variables (V) such as the abundance of hundreds of taxa at the individual level. However, due to constraints on time or resources, often it is possible only to make a limited number of simultaneous measurements. Thus, while both M and V are large, the measurement density (M/V) is much smaller than in conventional experiments. Such datasets are referred to as high dimensional, i.e., they consist of a small number of points in a high-dimensional space. Microbiome datasets consist of the composition of thousands of microbial taxa in an individual gut. Hence, understanding the role of gut microbiome in liver health requires us to employ machine learning (ML) approaches to dissect such a high-dimensional dataset. Utilizing these approaches enhances our understanding of the complex host-microbe reciprocal, helping in tracking disease progression over time or monitoring treatment responses, which is valuable for personalized medicine.In recent years, ML approaches have been widely used to shed light on how gut microbiome impact on various liver diseases. These studies have identified microbiome biomarkers associated with metabolic liver diseases (Ruuskanen et al., 2021;Zhang et al., 2021;Liu et al., 2022;Park et al., 2024). However, not only that majority of these studies are mostly cross-sectional in nature, these ML approaches suffer from a number of limitations. First, it is extremely difficult to detect patterns in microbiome data sets that can provide useful biological insight with translational value. High dimensional data analysis techniques such as PCA, ICA, or t-SNE, are useful at reducing dimensionality and detecting patterns. However, since the resulting axes represent linear combinations of a large number of features (e.g., taxa abundance), interpreting the analysis or making experimental predictions are often difficult (Donoho and Tanner, 2009) (Furchtgott et al., 2017). Identifying patterns becomes even more challenging as the proportion of relevant features decreases, at higher taxonomic levels (Donoho and Tanner, 2009). In fact, decades of research related to the gut microbiome have shown that specific combinations of a few microbes can be associated with a disease. This indicate that the composition of a small subset of microbes may be most relevant for making accurate computational inferences. Therefore, there is a need to develop novel methods to detect sparse patterns in high dimensional datasets.The second challenge is to extract meaningful models from high dimensional datasets that can predict interventional outcomes and guide translational implications of research findings. Building complex models describing microbial networks involve hundreds of parameters. Unfortunately, in most cases the parameters remain completely unknown. As Von Neumann once said: "with four parameters I can fit an elephant, and with five I can make him wiggle his trunk." The challenges involving microbial networks are clearly exacerbated since even the simplest of models would require hundreds of parameters. Hence, the available data are always limited and can never constrain the space of models completely. Given the under-constrained nature of the problem, there exist an infinite number of combinations of the parameters that can successfully replicate the data. Hence, making predictions based on such underconstrained gut microbial networks has proven to be challenging. Moreover, most of the studies trained their models on specific datasets such as Western or Chinese data sets that may not generalize to different populations, limiting their overall applicability. Often these studies employed a small patient population size. Some of these studies suffered from the absence of population characteristics such as the lack of lifestyle information (e.g., diet, socioeconomic status, tobacco, alcoholism, etc.). Also, in some case, lack of patient clinical, metagenomic and metabolomic profiles, hampered garnering a comprehensive understanding of the various causal aspects of disease. The complexity of the gut microbial dynamics and its function in metabolic liver disease remains largely unknown. The progressive character and course of liver disorders are not well captured by pre-clinical and cross-sectional investigations. We suggest using cutting-edge ML methods in conjunction with longitudinal research to address these issues. Hence, it is imperative to first develop methods to detect sparse relevant features that can then be used to find patterns in the data and train predictive models. By addressing these limitations and building new computational approaches, we can fully harness the potential of ML approaches to deepen our understanding of the gut microbiome's role in liver disease and other areas. By doing this, we can pinpoint certain gut microbial patterns associated with the advancement of liver disease, resulting in the development of more potent preventative and therapeutic approaches. For gut-liver axis research to reach its full potential, machine learning must be included.