Microbial omics pitfalls, from sampling to the bioinformatics pipeline, can reduce or bias the information yielded (Gutleben et al., 2018; Kaster and Sobol, 2020). Typically, some trade-off must be made in the experimental design, for which options have been suggested (Franzosa et al., 2015). In metataxonomics, resolution is usually limited to the genus level, though it is the most commonly used omics input for SML (Table 1), wherein relative operational taxonomic unit (OTU) abundances form the feature set (Miao et al., 2020; Janßen et al., 2021; Kim and Oh, 2021). However, the use of OTUs is inherently limiting for retaining community information and can miss important taxonomic groups. Indeed, since the development of the more biologically meaningful amplicon sequence variants (ASVs; Callahan et al., 2017), the absence of ASVs in most metataxonomic studies is striking. As ASVs represent a more accurate basis for taxa assignment, it will be interesting to see how their application influences ML performances in future.

Metagenomics is highly sensitive for low-abundance taxa, but is rarely applied for SML and carries additional costs which may limit sampling and options for ML (Chen and Tyler, 2020). Importantly, metagenomic approaches do not always convey a clear advantage over the more cost-effective metataxonomic approach (Xu et al., 2014). The choice between metataxonomics and metagenomics is evidently not clear-cut and should be considered in light of the expected community under study, choice of sequencing platform, and research goals. Microbial omics inputs are most often derived from closed-reference databases, leading to inevitable loss of learnable biological information in environmental samples due to unclassified/misclassified data (Chen and Tyler, 2020). However, the development of ML and DL tools (Liang et al., 2020) for enhancing taxonomic classification in metagenomic data sets could prove helpful. Alternatively, the direct use of biological sequences (from microbial omics surveys) circumvents this issue (by forgoing categorical assignment), thereby permitting the inclusion of more comprehensive feature spaces, at the cost of reducing the immediate interpretability for the user. Informative abstractions of omics data, such as the use of K-mer distributions as a feature set, have shown success in both taxonomic (Fiannaca et al., 2018) subtyping (Solis-Reyes et al., 2018) and phenotypic (Aun et al., 2018) classification, and are applicable to environmental applications. Indeed, K-mer abstractions have shown predictive potential for classifying sample environment and host-phenotype (an environmental status) that excels over OTU features (Asgari et al., 2018). Environmental metatranscriptomics-led SML is currently limited. However, the approach has been shown to uncover the mixotrophic processes of protists in response to nutrient gradients in the Pacific Ocean (Lambert et al., 2021), thereby demonstrating that trophic modes can be readily predicted from metatranscriptomic data.

There is a broad selection of the SML tools to select from and each carries its own advantages and limitations (Goodswen et al., 2021). Not a single architecture performs best in all environmental application cases and users must make a trade-off in terms of interpretability, learning performance, computational costs, data requirements, and ease of implementation (Ghannam and Techtmann, 2021). At the outset, selecting a set of architectures can help to ensure the delivery of research goals. Random forest (RF) is a popular choice for microbial omics-driven SML for its learning capacity, straightforward implementation, and high degree of interpretability (Ghannam and Techtmann, 2021). For especially complex tasks, or where knowledge is limited, DL approaches (multi-layered architectures) have the highest performance, as they can self-learn (i.e., do not require user extraction of) the feature set (Christin et al., 2019). However, DL comes with elevated computational costs and low interpretability of the underlying model (“black box” effect) and requires large volumes of data (thousands of samples). Consequently, though very promising, DL approaches for environmental omics are currently limited.

Feature selection and engineering are crucial for generating meaningful SML-based ecological models. Reducing the feature space can help to limit overfitting, reduce computational costs, improve cross-study comparison, and improve generalized prediction performance across data sets (Ghannam and Techtmann, 2021). However, care is needed when reducing features for training as biologically meaningful features can be missed if feature selection is based on abundance. This is especially so when assessing anthropogenic perturbations of pollutants in the environment, wherein the rare microbiome (taxa representing <0.1% of the total community) comprise a significant reservoir of gene clusters that enable the utilization and degradation of xenobiotic organic compounds (Wang et al., 2017). Taking embedded approaches for feature selection (that can evaluate across the full feature space) (Wang et al., 2017) or a biologically driven feature selection method (such as taxonomically aware hierarchical feature engineering) (Oudah and Henschel, 2018) may help in optimizing feature selection in metataxonomics-driven ML applications. Feature selection methods designed for functional feature sets are still notably lacking in this space.

Conventional statistics require assumptions on the underlying data and care is needed, given the compositional nature of microbial omics data sets (Gloor et al., 2017). For example, conventional ecological models often assume monotonicity in relationships, which can hinder ecological explanations of community variance across study sites. By applying SML (allowing for non-monotonic feature capture), the ability to capture this variance can increase nine-fold (Fontaine et al., 2021). It is important to note that the goal of SML should not be to replace classical statistical modeling, but rather to complement it. Integrating these two approaches presents an promising opportunity to leverage their advantages for predictive environmental microbiology (Lopatkin and Collins, 2020) and monitoring. For multi-omics studies, feature selection and engineering becomes increasingly complex with the successive systems levels, and there is much to be done in this area. In such studies, functional data across systems levels will likely need to be empirically assessed prior to SML to identify the most informative biomarkers for learning (Xu et al., 2014).

Data leakage is a subtle but important aspect of ML, referring to the unintended use or influence of data (that should not be available at the time of prediction) during the training process. This often occurs when the features used for training hide within themselves the result of the prediction, resulting in an overestimation of performance of the model during validation (Chiavegatto Filho et al., 2021). Due to the subtleties with which this can occur, avoiding data leakage is challenging and should be evaluated on a case by case basis. Important aspects for consideration here have been discussed previously (Wirbel et al., 2021) and include (1) data filtering that is influenced by the target label and (2) the splitting of dependent data (e.g., replicates and time-series data points) across training and validation sets. The use of an externally generated test data set (handled separately from the training set) for additional validation checks can help (Oyetunde et al., 2019; Wirbel et al., 2021), though data leakage is seldom discussed in microbial omics papers that use SML. We urge future authors in this space to consider including at least a statement on leakage assessment in studies based on SML.

Anthropogenic impacts are motivating the development of cost-effective and scalable environmental bioassessment methodologies (Fruehe et al., 2021). Microbes have long been recognized as potential in situ biosensors for following human impacts (Su et al., 2011), allowing for highly accurate quantitative SML predictions of the perturbation. Indeed, metataxonomic data can be valuable for the prediction of a variety of environmental contaminants (Table 1), spanning from relatively inert plastics (Li et al., 2021) to petroleum hydrocarbons [which illicit strong responses with detectable influences even after the pollutant is degraded and undetectable by conventional measures (Smith et al., 2015)]. Hydrocarbonoclastic indicator species have also been identified as key biosensors in ML-based bioprospecting of hydrocarbon seepage from subsurface reservoirs and can improve the likelihood of success in drilling for new assets (de Dios Miranda et al., 2019; Chitu et al., 2022). The same approach is also being explored as the potential early-warning indicators of leakage from hydrocarbon transport lines (Shaheen et al., 2011). Indeed, the SML of microbial fingerprints has even demonstrated reasonable predictions (accuracies of 72–85%) of the future production of hydrocarbon reservoirs (using metataxonomic input) (Zijp et al., 2021) which can facilitate decision-making for enhanced asset management. These approaches thereby have real potential for reducing the carbon footprint and ecological impact of upstream oil and gas activities.

Microbes have proved valuable as ecological assessment indicators in multiple diverse environments (Astudillo-García et al., 2019; Glasl et al., 2019; Hermans et al., 2020; Chen et al., 2021). Moreover, improvements in sequencing technologies are facilitating the upscaling and deployment of omics-based ML for more ambitious environmental monitoring and mitigation applications (Wang et al., 2021). These indicators can reveal important relationships for land management, when conventional field measurements are unhelpful (Chang et al., 2017). Indeed, the SML of microbial 16S rRNA abundances can directly predict soil productivity in arable land and risks posed for agriculture (Yuan et al., 2020). USML is routinely applied via ordination techniques to establish the organization of microbiome data in relation to their environmental parameters. However, in instances where conventional ordinations fail to determine clear relationships, SML may still yield community subpopulations that can serve as predictors for environmental parameters and processes of interest. For example, the influence between temperature and key phosphate and glycogen-accumulating organisms involved in the enhanced biological phosphorous removal processes of a set of wastewater treatment plants (WWTPs) in South Korea was identified using an SML approach, resulting in findings with clear implications for WWTP design and operation (Oh and Kim, 2021). Additionally, the SML of metabarcoded environmental DNA (eDNA) can provide superior performance for environmental quality monitoring over conventional bioindicator values for marine aquaculture monitoring (Fruehe et al., 2021). Furthermore, RF learning of eDNA has been shown to outperform conventional taxonomy-based biotic indices assessments (Cordier et al., 2018). Biodiversity in microbial communities can also be a useful proxy to assess the environmental impact of anthropogenic perturbations through changes in biotic indices (Aylagas et al., 2017). In these ways, SML is a useful means to improve environmental monitoring programs.

The predictive power of ML for monitoring environmental status also enables sample origin to be established (Raza et al., 2021). Microbial metrics have proved to be exceptionally sensitive indicators of human impacts on freshwater environments (Liao et al., 2018). Indeed, via ML modeling, the partitioning of microbes along complex anthropogenic xenobiotic gradients from urban and agricultural runoffs is sufficient to identify the origin of water samples from the 30 most abundant taxa (Wang et al., 2021) and is able to resolve sample origin depth and local salinity in the Baltic Sea (Alneberg et al., 2020). Such origin tracing carries the potential to inform for public health by accurately predicting the origins of fecal contaminants in public waters (Chen et al., 2021; Raza et al., 2021) and the source of food-borne pathogen outbreaks (Wheeler, 2019). The ability to identify sample origin sources is likely to be of critical importance moving forward for tracing runoffs from agricultural and industrial entities to ensure compliance with environmentally mindful legislation. It will be interesting to see whether this sort of tracing application will lend itself to following waterbodies in other settings, or indeed, other mobile elements within the environment (forensic analysis of migratory animals under conservation management, for example). Given the perceived stability in the gut microbiome, it is possible that this approach could also be extended as a biological tagging approach for following animal populations at the center of conservation efforts.

The high volumes of omics data are enabling large-scale meta-analyses (Zeller et al., 2014) that can provide a global view of microbial roles within major environments (Ramirez et al., 2018; Wu et al., 2019; Yuan et al., 2020). However, several challenges arise in such studies owing to non-standardized sample collection, extraction methods, and primer choice (Ramirez et al., 2018). Additionally, technicalities of sequencing platforms, variable library sizes, and environmental confounders can reduce concordance across omics studies (though SML is alleviating this issue) (York, 2021). ML tools are well suited for uncovering patterns within these challenging data collections. For example, a meta-analysis of soil microbiomes with SML was able to reveal microbiological indicators for predicting propensity for Fusarium wilt (Yuan et al., 2020), an agriculturally important pest. Additionally, a meta-analysis of global soil (Ramirez et al., 2018) and WWTP (Wu et al., 2019) communities provided macroecological insights into the microbial biogeography communities and confirmed the importance of the rare microbiome members as bioindicators. There remains significant scope for standardizing the workflows in both omics and SML. Such standardizations are crucial to mitigating common pitfalls; these enhance reproducibility and promote meta-analyses and data mining. An important limiting factor here is that many data sets are unavailable, uploaded to repositories without raw data or lacking metadata descriptions. This issue has been raised before (Ramirez et al., 2018) and impedes otherwise valuable work. For instance, bioprospecting of biosynthetic gene clusters with SML-based omics data mining can yield proteins with biotechnological potential (Correia and Weimann, 2021) for bioremediation, biodegradable plastic production, and sustainable biofuels (Haque et al., 2020; Keasling et al., 2021). We therefore urge that omics data sets be uploaded in their raw form with metadata made available.

The collective effects of anthropogenic perturbations are driving the consequences of climate change (notably, losses of ecosystem function, services, biodiversity, and habitat) at unprecedented rates (Giuliani et al., 2017). The actions of microbial communities are implicitly tied to geochemical cycling, global water chemistries, nutrient availabilities, and soil/plant health (Gorbushina and Krumbein, 2000; Falkowski et al., 2008; Lian et al., 2008; Dong, 2010; Panke-Buisse et al., 2014). Microbes are thereby drivers of numerous ecosystem services on which the global population relies (Marco and Abram, 2019). Understanding microbe–ecosystem interactions and functions is therefore central to their utilization in ecological models and biotechnologies for intervening on climate change. The generation of high-resolution spatiotemporal dynamics data and incorporation of different omics data sets can provide important insights into the molecular mechanisms behind climate changes responses and improve the accuracy of forecasting models (Herold et al., 2020; Layton and Bradbury, 2021). Together with their ubiquitous nature, the core roles of microbial communities afford us with a broad framework for potential microbiological tools with which the fundamental impacts of global climate change can be understood, monitored, predicted, and conceivably, mitigated. The short generation times of microbial community members and their predictable changes following changing environmental parameters (Larsen et al., 2012) open the possibility for their use as early-warning indicators of climate change-led impacts on macroecological networks (Shah et al., 2022) before further biodiversity loss is observable on the macroscale. Conversely, microbial contributions to climate change via carbon cycle-climate feedback and N₂O production (Bardgett et al., 2008) are an ideal candidate for predictive SML modeling and intervention. Indeed, predictive models from microbial omics data have also shown utility across a range of climate change-linked phenomena, including browning (Fontaine et al., 2021), eutrophication (Glasl et al., 2019), harmful algal blooms (Hennon and Dyhrman, 2020), and arability of soils (Chang et al., 2017; Hennon and Dyhrman, 2020; Yuan et al., 2020). omics in soil-plant, subsurface, and aquatic microbiomes is also central to making inroads in the development of carbon capture and sequestration (CCS) biotechnologies (Schweitzer et al., 2021). It will be interesting to see whether such developments benefit from SML-based modeling, which could prove useful for establishing taxa and metabolisms that predict stability and sequestration rates in CCS systems. Therefore, SML modeling can facilitate the establishment and optimization of carbon fluxes in microbial communities (particularly for the poorly characterized deep subsurface microbiome) and may also help to bridge bioenergy production to CCS, which is considered essential for many climate change mitigation plans (Hanssen et al., 2020). At present, the ability of microbes to inform on, and forecast, climate change impacts via ecological monitoring programs is perhaps the most immediately applicable area for the SML of microbial omics in climate change research. In this way, microbes can assist decision-makers for sustainable policies and intervention measures to ensure food security and maintain ecosystem services before further ecological detriment occurs (Cordier et al., 2021; Shah et al., 2022). The potential future applications in this space, however, are vast and may be key for realizing goals in global-scale climate management and engineering against climate change.