Planning and Analyzing a Low-Biomass Microbiome Study: A Data Analysis Perspective

Abstract

As investigations of low-biomass microbial communities have become more common, so too has the recognition of major challenges affecting these analyses. These challenges have been shown to compromise biological conclusions and have contributed to several controversies. Here, we review some of the most common and influential challenges in low-biomass microbiome research. We highlight key approaches to alleviate these potential pitfalls, combining experimental planning strategies and data analysis methods.

Low-biomass microbiome studies show great potential alongside major controversies. This review surveys key methodological challenges and discusses experimental planning strategies and data analysis methods that can address them.

Advances in microbiome research have spurred a growing interest in identifying microbial DNA in tissues or ecosystems with low microbial biomass, including tumors [1, 2], lungs [3, 4], placenta [5], and blood [6], as well as in nonhuman ecosystems such as the deep biosphere [7] or glaciers [8]. While such endeavors yielded many success stories, they have also fueled several controversies and contradictory results [9–13], highlighting a need for a deeper understanding of the common pitfalls, approaches, and limitations of low-biomass microbiome studies [14–19]. For example, the human placenta was once claimed to harbor a microbiome [5], but additional research revealed that the results were driven by contamination [12, 13]. Another article analyzing the tumor microbiome [20] was recently retracted amid concerns regarding misclassification of human DNA and machine learning approaches [10, 21]. In this review, we survey common challenges encountered in low-biomass microbiome studies, offer insights into how they can be addressed, and highlight remaining methodological gaps. We focus on aspects of study design and data analysis, rather than experimental approaches, which have been reviewed elsewhere [16, 22–27]. We note that while some have classified “low-biomass” quantitatively (eg, <10 000 microbial cells/mL [28]), we advocate for considering biomass as a continuum, with certain challenges having a stronger effect the fewer microbes are present in the ecosystem.

KEY ANALYTICAL CHALLENGES OF LOW-BIOMASS MICROBIOME STUDIES

Low-biomass microbiome studies have been undertaken using a variety of approaches, including 16S ribosomal RNA gene amplicon sequencing, metagenomics, and metatranscriptomics [1–8]. Below, we discuss key challenges that have emerged in various studies. While issues of host depletion and abundance estimation are not applicable to 16S sequencing, it provides limited phylogenetic and functional resolution.

Host DNA Misclassification

Metagenomic or transcriptomics data originating from low-biomass human microbiome studies would typically consist mostly of sequences originating from the host (eg, in the tumor microbiome only approximately 0.01% of sequenced reads were estimated to be microbial [2, 29]). While sometimes referred to as “host contamination” [23, 30], we find that this term is somewhat inaccurate, as host DNA is expected to genuinely be present in the ecosystem. It also misspecifies the main issue, which is that unaccounted host DNA can be misidentified as microbial [10] (Figure 1A). In most cases, this will generate noise that will impede the ability to identify signals in the data. However, if the levels of undetected host DNA are confounded with a phenotype, host DNA misclassification might result in artifactual signals.

Figure 1.

Challenges of data analysis in low-biomass microbiome research. A–D, Illustrations of 4 challenges in microbiome studies that are particularly complicating low-biomass studies: host misclassification [10] (sometimes called “host contamination”), in which DNA from the host, such as a human, gets incorrectly classified as bacterial (A); contamination [17], the introduction of microbial DNA during sample collection and processing that can be misclassified as belonging to the low-biomass microbiome of interest (B); well-to-well leakage [14], in which the contents of a sample can transfer to other samples, usually spatially adjacent to it during processing (C); and processing bias [31], the compositional impact of the varying efficiency each experimental and analytic stage has for different microbes, such that the relative abundance of different microbes can be over- or understated, depending on the sample composition and the experimental and analytic methods used (D). E–H, A hypothetical experiment illustrating how failing to account for these technical challenges during study design and computational analyses can lead to incorrect conclusions. This is an analysis of 108 samples (54 cases and 54 controls), of which 106 have an identical microbial composition. By construction, there is no underlying association between the phenotype (case/control) and any microbe. However, if the distribution of samples across plates is confounded with the phenotype (E and G), then technical factors of contamination, leakage, and bias will manufacture a difference in the measured communities between cases (F) and controls (H). I, As a result, implementations of standard differential abundance methods without technical corrections would identify multiple significant differences between the microbiomes of the case and control samples, despite the fact that their true compositions are nearly identical.

External Contamination

Among the most common issues plaguing low-biomass microbiome research is the unwanted introduction of DNA that originated from sources other than the environment being investigated. This external DNA, broadly described as “contamination,” can be introduced in various experimental stages, such as sample collection or DNA extraction, each with its own microbial composition [16–18, 22, 32, 33] (Figure 1B). Contamination is particularly relevant in low-biomass studies, because it will generally account for a greater proportion of the observed data. Similarly to host DNA misclassification, in most cases contamination will generate noise; however, if it is confounded with a phenotype it might result in artifactual signals.

Well-to-Well Leakage (“Cross-contamination”)

Another source of DNA that did not originate in the ecosystem itself are other samples that are processed concurrently, often nearby in space (eg, in adjacent wells on a 96-well plate; Figure 1C). Termed “cross-contamination,” “well-to-well leakage,” or the “splashome” [14, 33–35], this phenomenon can compromise the inferred composition of every sample. Importantly, we have recently demonstrated that well-to-well leakage into contamination controls violates the assumptions of most state-of-the-art computational decontamination methods [33].

Batch Effects and Processing Bias

“Batch effects” is a general term (and challenge) in biology, which describes the differences observed among samples from different laboratories or processing batches, and can be attributed to differences in protocols, personnel, reagent batches, or even ambient temperature [18, 36–39] (Figure 1D). In the microbiome field, these differences have also been attributed to variable efficiency of different experimental and analytic processing stages for different microbes, altogether termed “processing bias,” which may be increased by some experimental approaches used in low-biomass research [15, 31, 40–42]. Processing biases were shown to potentially distort inferred signals both when batches are confounded with a phenotype, but also when they are not, if the mean sample efficiency is associated with a phenotype [42]. Batch confounding and batch effects, which are exacerbated by processing bias, underlie how many issues, such as those discussed above, generate artifactual signals. Furthermore, the ability to distinguish between processing bias and contamination could complicate the interpretation of low-biomass research.

Identification and Classification of Microbes

Low-biomass microbial ecosystems are often understudied, and the microbes potentially residing in them might not be adequately represented in reference genomic datasets [26, 43, 44]. As a result, the accurate identification of microbes in these samples is often problematic. This is of particular concern when the identity of detected microbes serves to validate that the analytic pipeline yields sensical results [11, 45].

A Hypothetical Case Study

We demonstrate the risks inherent to these challenges with a hypothetical analysis of a simulated case:control dataset with 54 samples from cases and 54 from controls. For simplicity, we demonstrate a scenario in which 53 of the samples from each group have an identical distribution consisting of 2 taxa, along with 1 extra sample per group that contain monocultures of a third and fourth taxon. In this scenario, the cases and controls are processed separately. Consequently, the batch that contains the case samples (Figure 1E) is subjected to distinct contamination, well-to-well leakage, and processing bias, resulting in an observed dataset that is highly different from the true samples (Figure 1F). The control samples (Figure 1G) undergo slightly different contamination, well-to-well leakage, and processing biases, which would result in an observed dataset significantly different from that of the case samples (Figure 1H). As a result, and despite the fact that 98% of all samples are identical, an analysis of the 2 observed datasets would identify 6 taxa “associated” with case/control status (Figure 1I): 2 due to contamination, 2 due to well-to-well leakage, and 2 due to processing bias. While this example emphasizes the risk of the different challenges, we note that the reason underlying artifactual results is the confounding of the phenotype with the batch structure. In the absence of such confounding, the result of the aforementioned processes would be noise (Figure 2). In either case, these pitfalls are addressable through effective study design and analytics, which we outline below.

Figure 2.

Unconfounded experimental design can avoid artifactual signals. Similar to Figure 1  E–H, but in which case and control samples are uniformly scattered across and within each plate. A–D, This hypothetical example illustrates how a microbiome experimental design in which case and control sample processing is not confounded by factors such as sequencing plate and well locations within each plate avoids the artifactual signals potentially generated by the challenges of contamination, well-to-well leakage, and processing bias.

EXPERIMENTAL STUDY DESIGN

Optimal experimental study design is essential for low-biomass microbiome studies. We offer some important considerations.

Avoid Batch Confounding by Optimizing Study Design

A critical step to reducing the impact of low-biomass challenges on subsequent data analysis is to ensure that phenotypes and covariates of interest are not confounded with the batch structure at any experimental stage (eg, sample shipment batches or DNA extraction batches). As demonstrated in Figure 2, in which both batches include a similar ratio of case and control samples, an unconfounded design increases the likelihood that experimental biases will mask true signals rather than introduce artifactual ones into downstream analyses. A broad definition of analytic goals, even for exploratory analysis, combined with study-specific considerations and domain knowledge, will help in identifying covariates and outcomes to consider when designing a study. While randomization of samples is helpful, we advocate for a more active approach in generating unconfounded batches, such as the approach proposed by BalanceIT [46]. If batches cannot be de-confounded from a covariate, such as in the case of a clinical site with a different case:control ratio than other sites, we recommend that rather than analyzing data from all batches together, the generalizability of results be assessed explicitly across batches [40, 47]

Use Process Controls That Represent All Contamination Sources

Even when best efforts were made to reduce batch confounding, addressing the challenges of low-biomass microbiome studies is critical to reduce the impact of residual confounding and to improve detection of true signals. Chief among these challenges is contamination (Figure 1B). While best laboratory practices can reduce contamination, they cannot eliminate it. It has therefore become standard to collect process controls whose contents represent contamination introduced throughout the study [1, 18, 32, 35].

Which controls to collect is a topic of ongoing investigation. Some have recommended focusing on control samples that pass through the entire experiment and therefore “represent” all contaminants concurrently [48, 49]. We caution, however, that in large studies this requires careful planning that ensures that these control samples are present in each batch, or they might miss certain contamination sources (see Figure 3 for an example). We and others therefore advocate for additionally identifying contamination sources and profiling them separately using process-specific controls [16, 32, 33] (Figure 3). The types of controls collected should be tailored to each study; examples include surface or adjacent tissue samples [1, 13, 50], empty collection kits [18, 35], blank extraction controls [13, 20], no-template controls [1, 14], or library preparation controls [20, 33]. For each of these, attention should be given to factors that may cause a difference in the contamination profile, such as manufacturing batches for swabs [13].

Figure 3.

“Whole experiment” controls might miss certain contamination sources. An illustration of how separate contamination sources are captured by different sets of controls. Certain process controls, such as the blank swab collected early in the pipeline, will contain combinations of downstream contamination sources. Samples from separate process batches, such as DNA extraction batches, can be subjected to different contaminants. Note that the blank swab does not represent the contamination source of batch 1 (dark purple).

We note that we are not aware of a general consensus or data-driven guidance on the required number of controls per contamination source. We have found that 2 control samples are always preferable to 1, and that in specific cases more controls are helpful [33], particularly when a high amount of contamination is expected. We anticipate that the optimal number of controls, as well as the prioritization between different types, will vary between studies and ecosystems. While it is best to collect process control samples for every possible source of contamination, this might not always be feasible, in which case careful analytic strategies and alternative decontamination methods that do not use controls should be applied.

Minimize Well-to-Well Leakage and Account for It in Experimental Design

Our current ability to address well-to-well leakage analytically is limited, making study design and experimental measures even more important. To minimize leakage, high- and low-biomass samples (eg, stool and tissue samples) should not be processed together, as it has been demonstrated that such scenarios risk high leakage into the low-biomass samples [14]. It was further shown that using robotic systems bears higher risk of well-to-well leakage [14]. Beyond these observations, we are currently not aware of accepted standards or guidelines for addressing this challenge. We therefore recommend that laboratories performing low-biomass microbiome studies profile and calibrate the extent of this phenomenon with their protocols, potentially using monocultures [14, 26].

Well-to-well leakage is especially destructive for contamination controls, as it contradicts the assumptions made by most decontamination methods [33]. We therefore recommend that efforts be made to limit the number of samples processed adjacent to contamination controls. For example, for processing done on a 96-well plate, we have observed that leaving blank wells around them reduces well-to-well leakage [33]. Additionally, as noted above, well-to-well leakage can confound results by exposing samples from similar groups to the same leakage pattern, which risks creating artifactual results (Figure 1F, 1H, and 1I). Thus, just as with batch design, it can often be helpful to randomize sample locations, and it is important to ensure that there are no covariates of interests associated with processing order or spatial arrangement of an experiment (Figure 2).

Record and Publish Detailed Experimental Metadata

Correcting for factors such as batch effects, contamination, and well-to-well leakage requires in-depth sample metadata. It is therefore critical to record how each sample traversed through the experimental pipeline, including factors such as collection site and time, as well as batches of shipment, aliquoting, DNA extraction, PCR (polymerase chain reaction), library preparation, and other study-specific stages. The order of processing and location on processing plates should also be recorded. This information is important both for the initial analysis of a study, as well as for future reanalyses and meta-analyses performed by other investigators, and should therefore be made publicly available alongside the data. We advocate for its explicit inclusion in checklists such as STORMS (Strengthening The Organization and Reporting of Microbiome Studies) [51]. There are many additional relevant factors, such as which researchers were processing which samples, or what were the production batches of different reagents. In general, recording and examining as much information describing the experimental pipelines as possible gives investigators the best opportunity for successful data analysis.

DATA ANALYSIS

Effective data processing and analysis of the resulting data are needed to ensure robust interpretation of low-biomass data (Figures 2 and  4). Importantly, in every analytic decision made in these studies lies a tradeoff between discovery of novel signals and reporting of results that are conservative and high confidence. Analyses with several levels of confidence may therefore be beneficial. For example, Poore et al. [20] analyzed 4 datasets with increasing levels of decontamination stringency and 2 abundance estimation pipelines. We discuss 5 recommendations for sequencing-based low-biomass microbiome data analysis.

Figure 4.

Cross-batch or cross-study generalization demonstrates robust data analysis. Illustration of an analysis unlikely to suffer from inflated results due to contamination, bias, misclassification, or well-to-well leakage. The training set (A) is contained within a separate batch from the test set (B) and is processed separately. Both batches contain random dispersion of case and control samples, along with process controls to facilitate decontamination. C, A standard metagenomics processing pipeline for the training batch, which includes host depletion, reference mapping, decontamination, and bias correction. Afterward, some analysis of the resulting training dataset should be implemented to evaluate the biological relevance of the sequenced microbes, for example, by devising machine learning (ML) models that are predictive of a phenotype. D, Processing of the test samples should follow the same 4 steps as the training set, upon which the resulting samples should be used to assess the generalization of the associations or predictive models described for the training set. None of the processing or analysis steps, of either training or test samples, should incorporate test set labels. While we illustrate an example train-test split across 2 plates, certain situations might require relevant analysis split across patient populations, covariates, or more.

Avoid Misclassification of Host Reads

A substantial concern is that sequenced reads originating in the DNA of the host might be misclassified as microbial [10]. A key step to address this is to filter host reads by mapping them against reference genomes. These references have seen impactful updates in recent years and are still evolving, and orders-of-magnitude differences in mapping rates are observed between different references [29]. Therefore, the use of updated references, such as T2T-CHM13 [52] and the Human Pangenome Reference Consortium [53], is necessary. Furthermore, a recent analysis showed that the number of reads removed using subsequent reference updates has not reached saturation, indicating that human reads are potentially still missed [29]. While these are likely in low quantities, we recommend applying additional quality assurance steps, such as investigating taxa whose abundance across samples is correlated with the abundance of detected host DNA.

A complementary way to reduce human reads misclassification is to ensure that microbial reference genome databases do not contain host sequences, for example, using postprocessing methods [29, 54]. We note that some methods, such as Kraken [55], require the presence of the human reference genome also within the reference database, lest it results in misclassification [10, 55, 56], but this is not necessary for other methods such as Bowtie2 [57]. While it has been suggested that reference databases should only include complete genomes because these are less likely to contain errors [10], we find that this is unnecessarily restrictive. Recent assembly efforts have demonstrated incredible diversity not captured in datasets of complete genomes [43, 58, 59], and it is likely that many microbial inhabitants of low-biomass microbial ecosystems do not have complete reference genomes. We recommend, instead, approaches such as the inclusion of a simulated sample composed entirely of host DNA that is then processed with the same analytic pipeline to quantify the rate of false-positive misclassification of host DNA as microbial [29].

Interpret Microbial Abundance Estimation With Caution

In low-biomass microbiome research, the identity of taxa measured in the ecosystem is often used as a measure for analytical rigor or the effectiveness of decontamination [9, 11, 45]. We caution against this approach: Such ecosystems are often understudied, and misidentification of taxonomy is likely [60]. We also caution against the opposite: It is advisable not to heavily rely on specific taxonomic classifications without deeper investigation and potentially validation, even if these show some biological plausibility.

Additional approaches may be necessary to alleviate issues in abundance estimation. Homology between genomes and ambiguously mapped reads [61] can cause spurious read mapping to a particular genome and a false-positive identification. Examination of coverage breadth has been proposed for this purpose and can be incorporated into abundance estimation [56, 62], under the premise that a certain proportion of the genome needs to be covered to consider a microbe as potentially present in the ecosystem [29]. Further development of bespoke methods for draft genome references and low read counts, characteristic of low-biomass microbiomes, is necessary. Additionally, using absolute abundance measures such as spike-in taxa can be used to establish a limit of detection for low-abundance taxa [13].

An alternative approach that avoids reference databases is de novo assembly [43, 58, 59]. However, it is typically challenging in low-biomass microbiome studies, as obtaining the necessary sequencing depth is often difficult and expensive. Bespoke development of reference-guided co-assembly pipelines could greatly advance this field [63]. Finally, complementary measurements, such as cultures or fluorescence in situ hybridization (FISH) [64], which are independent of some complications that affect sequencing data, offer potential for additional validation.

Decontaminate Analytically, but Consider Uncaptured Contamination

“Decontamination” methods in microbiome research aim to computationally remove unwanted microbial DNA introduced during sample collection and processing (Figure 1B). The use of process control is strongly recommended [16, 17], and most decontamination methods use these controls to identify contamination [33, 48, 65] (Figure 4A and 4B). For instance, decontam [65] classifies species as contaminants based on their prevalence in controls and samples, while SCRuB [33] models taxonomic compositions of contamination sources, and accounts for the possibility of well-to-well leakage into process controls (Table 1). When using SCRuB, we recommend to sequentially model each contamination source using process-specific controls. Additionally, since precise contamination communities are specific to individual processing groups/batches, every SCRuB decontamination layer should be performed per individual batch or grouping (Figure 3).

Table 1.

Overview of Public Resources Commonly Used to Address Challenges in Low-Biomass Microbiome Research

Category	Subcategory	Methods	Description
Host depletiona	Alignmenta	BWA [66] Bowtie2 [57]	Used to filter sequenced reads that align to the genome of the host.
	Database depletiona	Conterminator [54] Exhaustive [29]	Removes elements similar to host genomes from microbial databases.
Host referencea	Human genomea	T2T-CHM13 [52]	Human genome reference.
	Human pangenomea	HPRC [53]	Genetically diverse collection of reference human genomes.
	Human transcriptsa	GENCODE [67]	Collection of human transcript references for RNA-sequencing data.
Microbial abundance estimationa,b	Alignmenta	BWA [66] Bowtie2 [57]	Used to compare sequenced reads to microbial genomes by aligning the entire length of each read.
	K-mer–based approachesa	Kraken [55] KrakenUniq [56]	Used to compare sequenced reads to microbial genomes by comparing k-mers within each read.
	Marker genesa	MetaPhlAn [68] mOTUs [69]	Assess microbial composition by comparing sequenced reads against microbial marker genes. Would typically be sensitive to the low read count in low-biomass samples.
	Coveragea	KrakenUniq [56] ICRA [62]	Account for the fraction of reference taxa genomes that is covered by sequenced reads.
	16S analysisb	DADA2 [70] Deblur [71]	Identify and quantify microbial abundance using 16S ribosomal RNA amplicon sequencing data.
Decontaminationa,b	Contaminant classificationa,b	decontama,b [65] Squeegeea [72]	Use information from collected samples to infer if a taxon is a contaminant and should be removed from downstream analyses.
	Control-free decontaminationa,b	decontama,b [65] Squeegeea [72]	Methods that identify and remove contamination without the use of process controls.
	Composition modelinga,b	microDecona,b [48] SCRuBa,b [33]	Processes collected samples to infer and remove a contamination source from a dataset.
Well-to-well leakagea,b	Handle leakage into controlsa,b	SCRuBa,b [33]	Remove potential well-to-well leakage from controls to avoid compromising the decontamination process.
	Strain-level comparisona	inStraina [34, 73]	Compare which pairs of samples share similar strains.
Bias/batch correctiona,b	Batch correctiona,b	ComBat [74] ConQuR [75] Voom [76] SNM [77] MMUPHin [78] PLSDA-batch [79]	Transformations that strengthen microbial composition similarities across multiple batches, often while preserving differences within specified covariates.
	Bias correctiona,b	radEmu [41] DEBIAS-M [40]	Methods that explicitly model and correct for processing bias.

Abbreviations: BWA, Burrows-Wheeler Aligner; ConQuR, Conditional Quantile Regression; DADA2, Divisive Amplicon Denoising Algorithm 2; DEBIAS-M, domain adaptation with phenotype estimation and batch integration across studies of the microbiome; HPRC, Human Pangenome Reference Consortium; ICRA, iterative coverage-based read assignment; MetaPhlAn, metagenomic phylogenetic analysis; MMUPHin, meta-analysis methods with a uniform pipeline for heterogeneity in microbiome studies; mOTU, marker gene–based operational taxonomic unit; PLSDA, partial least squares discriminant analysis; SCRuB, source-tracking for contamination removal in microbiomes; SNM, supervised normalization of microarrays; T2T-CHM13, Telomere-to-telomere assembly of the CHM13 cell line.

^aDenotes relevance to metagenomic and metatranscriptomic sequencing.

^bDenotes relevance to amplicon sequencing.

Importantly, it is likely that not all contamination sources are adequately captured in process controls [13, 20], necessitating complementary approaches for identifying contamination that do not require such controls. For example, decontam's frequency test identifies an inverse correlation between taxa abundance and sample concentration as evidence for contamination [65]. Squeegee compares samples from different ecosystems that have a procedural similarity, such as similar DNA extraction kits, and considers taxa shared among them as contaminants [72]. Using lists of “known contaminants” or “kitomes” is common, particularly in low-biomass microbiome studies. However, we advocate for more experiment-specific data-driven approaches, as these lists often contain known commensals and it is possible for similar species to both inhabit an environment of interest and be introduced as contamination.

Some have argued that observing different microbial abundances between similar samples that are processed with different kits is an indication of contamination [13, 45]. However, we find that this argument does not sufficiently distinguish between experimental contamination (Figure 1B) and processing bias or batch effects (Figure 1D). These have been shown to introduce changes even in high-biomass samples [15, 31, 38, 39]. Therefore, we believe that variations in abundances of particular microbes across sequencing kits alone should not be considered evidence for contamination (batch-specific differences in microbial presence, however, offer stronger evidence for contamination). In general, as we hope this discussion demonstrates, this is an area of ongoing research with substantial need for development of new, “control-free” decontamination methods.

Be Wary of Information Leakage in Batch Correction

Given how common batch effects are, many tools have been developed to address them [74–79] (Table 1). These are generally designed to reduce the differences between microbial compositions across different batches, while sometimes maintaining differences in specified covariates. Other methods, such as DEBIAS-M [40] or radEmu [41], were developed specifically to capture and correct for multiplicative processing bias proposed for microbiome studies [15, 31].

Batch- and bias-correction methods can be important for identifying generalizable microbiome patterns across batches and studies. However, the use of covariates and outcome variables by these methods can inflate the strength of observed associations. Particularly, in predictive modeling, it can compromise separation between training and test sets. As outlined in Figure 4, no processing steps should incorporate any test-set labels. Recently, application of batch correction methods have sparked controversy, as they were shown to introduce phenotype-associated values to sparse features [10]. While we demonstrated that this is often a benign result of using pseudocounts [80], caution should still be used when interpreting features manipulated by most batch correction methods. As with all data processing steps, it is recommended to perform exploratory data analysis to understand how different host depletion, reference mapping, decontamination, and batch/bias-correction pipelines have impacted analyses.

Use Generalization Benchmarks

The strongest indication for adequate handling of the various challenges of low-biomass microbiome studies is the detection of a robust association, in a nonconfounded setting, between a microbial feature and some covariate, such as a patient's phenotype. To ensure generalization, we recommend doing this in a predictive setting, via a train-test split that employs cross-validation within the training set for model selection and hyperparameter tuning. While optimal experimental design can alleviate concerns of batch confounding, we still recommend, when possible, to evaluate generalization across studies or batches such that residual batch confounding would make the benchmark harder rather than easier (Figure 4). Through such a benchmark, demonstrating that a model generalizes effectively to an independently processed test dataset is an effective justification for the legitimacy of the selected preprocessing steps, and the biological relevance of the detected associations.

DISCUSSION

There is a great opportunity for furthering our understanding of both human and nonhuman ecosystems through the analysis of low-biomass microbiomes. We discussed key challenges that have emerged in this field, which can introduce the possibility of both overstated and understated conclusions. We reviewed experimental design and computational analysis approaches that can mitigate these challenges, and highlighted areas in need of further methodological development.

While we described computational resources and tools that can be used for low-biomass microbiome research (Table 1), there is much need for further research and development of new methods. In particular, we note that recent controversies surrounding the potential misclassification of host DNA [10, 29] highlight a necessity for new methods. Additionally, there continues to be a lack of methods that can quantify and remove well-to-well leakage between samples, due to the computational challenge of confidently inferring the true source of the “leak” in most datasets. Finally, there is a need for new approaches that detect contamination in a data-driven fashion yet without the use of process controls, as these are not always available or sufficient. As this is a rapidly evolving field, we expect that the set of optimal tools will expand in the coming years.

Notes

Acknowledgments. We thank members of the Korem group for useful discussions.

Financial support . This work was supported by the Program for Mathematical Genomics at Columbia University and the National Institutes of Health (grant numbers R01HD106017 to T. K. and T15LM007079 to G. I. A.).