II. 3. Shotgun Metagenomics – The Fuller Picture, but Not the Whole Truth

Shotgun metagenomics reads the sample's entire DNA for a richer picture, but the interpretive uncertainties don't disappear — they merely shift.

Shotgun metagenomics does not sequence a specific gene, but the entire DNA content of the sample. This in principle allows the simultaneous identification of bacteria, viruses, fungi, archaea and the host organism's DNA [19], and even the mapping of functional genes – metabolic pathways, antibiotic resistance genes. The technology is genuinely impressive [3], [2].

Its limitations, however, are no less noteworthy:

• Decisive influence of the DNA extraction protocol: Researchers examining the same sample using 21 different methods obtained completely different microbial pictures. The extraction procedure itself determines the result – before a single sequencing step has taken place [19].
• Host DNA dominance: In stool samples, human cell DNA often overwhelms the microbial signal. Rare bacterial strains thus remain invisible, unless special "host DNA depletion" steps are applied – which can however introduce new biases [21].
• The decisive role of the bioinformatic pipeline: Processing the same raw sequencing data with different software (MetaPhlAn, Kraken, DIAMOND, etc.) produces completely different taxonomic and functional profiles. According to the CAMI (Critical Assessment of Metagenome Interpretation) international comparison study, pipeline choice influences the result more strongly than the biological differences between samples [22].
• Presence of functional genes ≠ function: Shotgun shows whether a given metabolic gene is present in the sample – but not whether that gene is active, expressed, or actually performs the given process in the gut. Inferring functional activity from gene presence is only possible indirectly.
• High cost, slow results: Full metagenomic analysis is considerably more expensive and time-consuming than 16S analysis, making it less suitable for clinical monitoring purposes.

The same thousand scattered LEGO sets – but now every single piece is photographed and measured. In principle, far more data is available. The problem: some photos have human fingerprints on them (host DNA), the quality depends on the camera settings (protocol), and running a different software package produces completely different identification results from the same photos (pipeline). The more detailed picture is indeed richer – but the interpretive uncertainties do not disappear; they merely shift.

Analytical algorithms: interpolation and hallucination

In shotgun metagenomics, algorithmic uncertainty is one level more complex [19], [22], since the analysis concerns not only compositional determination but also functional annotation. The software (e.g. MetaPhlAn, Kraken, HUMAnN) generates taxonomic and functional profiles from raw sequencing data – but both steps are characterised by interpolation and hallucination errors.

At the taxonomic level, interpolation means that the program matches an unknown genome fragment to the closest known genome and attributes it to that species [22]. If the database is incomplete – and significant portions of the gut microbiome still lack reference sequences – the program assigns the sequence to a taxon based on estimation rather than measurement. In functional annotation, by similar logic: if a gene fragment shows 60–70% similarity to a known metabolic gene, the software assigns its function accordingly – while at 30–40% divergence, the actual biochemical activity may be entirely different.

Shotgun-specific hallucination occurs particularly at the intersection of rare taxa and host DNA contamination: human DNA fragments are sometimes identified as bacterial sequences if the reference database matching algorithm fails to filter them adequately. In the CAMI comparison study, this resulted in different pipelines giving mutually contradictory species identifications in up to 50% of cases from the same sample.

What does this mean for the patient?

The appealing promise of shotgun metagenomics is completeness: it shows not only who is present, but in principle also what functional capabilities they have [19], [22]. This is undeniably a level deeper than 16S amplicon sequencing. In practice, however, the richer data brings the same interpretive difficulties – and in some respects intensifies them.

The biggest misconception about shotgun analysis is that the richness of the data inherently includes clinical clarity. It does not [19], [22]. The presence of a metabolic gene in the gut microbiome only means the gene is there – not that it is active, not that it is expressed in sufficient quantities, and not that the produced metabolite reaches tissues in clinically relevant amounts. Every step in the "gene → protein → function → clinical effect" chain requires its own measurement method.

Furthermore, the 200–400,000+ HUF price of shotgun analysis may create the expectation of more accurate or reliable guidance [19], [22] on the likely success of FMT or other interventions. This is not currently a well-founded expectation. The price reflects the complexity of the technology – not the quality of clinical decision support.

Where shotgun does have a genuine advantage over 16S: in specific, pre-defined clinical questions. For example, identifying antibiotic resistance genes in suspected hospital infection, or investigating viral and fungal communities in immunocompromised patients – these are indications where knowledge of the entire DNA content has real therapeutic implications. For FMT selection or general dysbiosis diagnosis, however, it does not currently possess demonstrated clinical utility beyond 16S [19], [22].

Shotgun Metagenomics — Four Methodological Pitfalls

Shotgun metagenomics provides a more comprehensive picture than 16S by sequencing total DNA, but it has its own well-documented limits:

1. Sample preparation strongly biases results. Costea et al. 2017 (Nature Biotechnology) compared the same stool samples processed by 21 different extraction protocols and obtained entirely different microbiome profiles — the extraction method alone determines the result [19]. Without an industrial standard, this makes cross-study comparison nearly impossible.

2. Excess human (host) DNA in the sample. Particularly in tissue or skin samples, human DNA proportionally suppresses the microbial signal. Chemical host-DNA depletion steps are used to address this, but they can introduce new biases [476]. The issue is smaller for stool samples but not zero — especially in mucosa-sensitive states (e.g., active IBD).

3. Bioinformatics pipeline choice is decisive. The CAMI (Critical Assessment of Metagenome Interpretation) international benchmark systematically demonstrated that the same raw data processed by different software (MetaPhlAn, Kraken, MEGAHIT, mOTUs, etc.) yields different microbiome portraits. Pipeline choice influences the outcome more than the biological difference between samples [22].

4. General, method-dependent bias. McLaren et al. 2019 (eLife) comprehensively documented that microbiome measurements regularly contain systematic, method-dependent biases that prevent direct comparison of results across protocols without correction procedures [21]. Two shotgun analyses of the same patient in different labs may not reach the same conclusion.

The clinical message is the same as for 16S: shotgun metagenomics, too, is only suitable for longitudinal monitoring if the same sampling method, laboratory, and analysis algorithm are used throughout.

References

[2] Qin J, Li R, Raes J et al. A human gut microbial gene catalogue established by metagenomic sequencing. Nature. 2010. Link

Illumina-based metagenomic sequencing of faecal samples from 124 European individuals (576.7 Gb of sequence) yielded a catalogue of 3.3 million non-redundant microbial genes, approximately 150-fold larger than the human gene complement. Genes were largely shared across individuals, with over 99\% bacterial origin. The cohort harboured an estimated 1,000–1,150 prevalent bacterial species, each individual carrying at least 160 species. The study defines a minimal gut metagenome and a minimal gut bacterial genome based on functions present in all individuals and most bacteria. Findings establish a foundational reference for the genetic potential of the human gut microbiota.

[3] Turnbaugh PJ, Ley RE, Hamady M, Fraser-Liggett CM, Knight R, Gordon JI. The Human Microbiome Project. Nature. 2007. Link

Strategic framework outlining the Human Microbiome Project's approach to characterizing the microbial components of the human genetic and metabolic landscape. The initiative aims to establish how microbiota contribute to normal physiology and predisposition to disease. Serves as the foundational programmatic statement for large-scale population-level microbiome research.

[19] Costea PI, Zeller G, Sunagawa S et al. Towards Standards for Human Fecal Sample Processing in Metagenomic Studies. Nature Biotechnology. 2017. Link

21 representative DNA extraction protocols were tested on identical faecal samples and compared with library preparation and storage effects against biological within-individual variation. DNA extraction had the largest technical effect on metagenomic outcomes. Protocols were ranked by DNA quantity, quality, and biases in community diversity and Gram-positive/Gram-negative ratio. The authors recommend a standardised, transferable DNA extraction method validated using a mock community of known composition for human faecal metagenomic studies.

[21] McLaren MR, Willis AD, Callahan BJ. Consistent and Correctable Bias in Metagenomic Sequencing Experiments. eLife. 2019. Link

Marker-gene and metagenomic sequencing measurements are systematically biased toward detecting certain taxa over others, making taxon abundances generated by different protocols quantitatively incomparable and prone to spurious biological conclusions. The authors propose a mathematical model of experimental bias based on real-experiment properties and validate it with 16S rRNA and shotgun metagenomics data from defined bacterial communities. The model fits experimental data better than previous, more complex frameworks and offers a path to correcting bias.

[22] Sczyrba A, Hofmann P, Belmann P et al. Critical Assessment of Metagenome Interpretation -- A Benchmark of Metagenomics Software. Nature Methods. 2017. Link

The Critical Assessment of Metagenome Interpretation (CAMI) challenge benchmarked metagenomics software using highly complex realistic datasets from ~700 newly sequenced microorganisms and ~600 novel viruses and plasmids. Assembly and binning performed well for species represented by individual genomes but were substantially degraded by closely related strains. Taxonomic profiling and binning were proficient at high taxonomic ranks with a marked drop below family level. Parameter settings strongly affected performance, underscoring the importance of reproducibility. CAMI provides a roadmap for software selection.

[476] Marotz CA, Sanders JG, Zuniga C et al. Improving Saliva Shotgun Metagenomics by Chemical Host DNA Depletion. Microbiome. 2018. Link

To enable shotgun metagenomic sequencing of host-dominated oral samples, three commercial host-depletion kits, size filtration, and a novel osmotic-lysis + propidium-monoazide method (lyPMA) were compared in human saliva. lyPMA was the most efficient method, reducing host-aligned reads from 89,29 ± 0,03% in untreated samples to 8,53 ± 0,10%. Furthermore, lyPMA-treated samples showed the lowest taxonomic bias compared with untreated controls. The method is recommended for microbial-DNA enrichment from host-rich oral samples in metagenomic studies.