Soil PICRUSt2 results estimate the genetic functions that might be represented in a microbial community described by marker-gene sequencing. If you want to understand PICRUSt2 predicted metagenomes, KEGG orthologue or KO abundance tables, EC enzyme classifications, MetaCyc pathways, nearest-sequenced taxon index (NSTI) values or PICRUSt2 functional prediction limitations, begin with the essential question: were those genes directly sequenced, or were they inferred from reference genomes?
A PICRUSt2 prediction is not equivalent to shotgun metagenomics, a measured transcript, an enzyme assay or a soil process rate. It combines observed marker sequences with evolutionary placements and reference-derived gene-content predictions. That can reveal useful hypotheses about potential functions, particularly for exploratory comparison, but uncertainty depends on environmental reference coverage and the organisms being represented. In soil, many relevant lineages lack well-characterised reference genomes.
This guide walks through PICRUSt2 output files, NSTI interpretation, predicted gene families, inferred pathways, correction factors, compositional scaling, sample comparisons and practical biological limits. It offers original worked examples for readers explaining a soil microbiome report to students, researchers or a wider audience. Illustrative numbers are constructed to teach the calculations; they are not results of a laboratory experiment.
Quick answer: PICRUSt2 predicts rather than directly reads functions
PICRUSt2 is a computational workflow for predicting functional potential from marker-gene community profiles using phylogenetic information and genomic reference data. A typical input is a table of amplicon sequence variants (ASVs) with their counts across samples and representative nucleotide sequences. The tool places sequences into a reference phylogenetic context, predicts gene content and combines those predictions with sample feature abundances. It may also infer pathway profiles from predicted gene-family information.
The central reporting distinction is therefore ‘predicted functional potential’ rather than ‘measured function’. The output is model dependent, not a direct census of the genes recovered from all environmental genomes. If you see a list of pathways with numerical values, your first duty is to identify exactly how those values were generated. A correct calculation can still be biologically overinterpreted if the observation is mislabelled.
Why scientists use functional prediction at all
Many soil investigations begin with accessible bacterial or archaeal marker surveys. Those surveys identify variation in particular genetic markers but do not directly sequence every organism’s full genome. Shotgun metagenomics may be informative yet can require greater sequencing, laboratory and analytical resources. A prediction method can use existing marker data to formulate hypotheses about functional capacities that deserve more direct investigation.
A useful use case might be comparing the predicted distribution of gene families associated with carbon transformation across independently sampled soil management treatments. The output can prioritise candidates for qPCR, shotgun sequencing, transcript measurements, enzyme tests or incubation experiments. It should not be used alone to announce that carbon was fixed faster, greenhouse gases were reduced or a soil treatment is universally beneficial. Each of those claims needs evidence measuring the relevant outcome.
What you must know about the original marker survey
The prediction begins from a biological sample and an amplified sequence marker. Read how the soil was collected, which marker region and primers were used, how many biological replicates were collected and how DNA was extracted. The resulting ASV table has already been shaped by extraction efficiency, primer biases, PCR amplification, sequencing depth and denoising rules. Functional prediction inherits those limitations rather than cancelling them.
If a primer does not recover certain microbial groups, their predicted genes cannot appear reliably in the final profile. If one treatment was sampled at a different depth or extracted using a different laboratory protocol, the inferred difference may partly reflect the method. The soil ASV table guide explains why a read-count table is not a count of living microorganisms. That lesson remains true after predicted gene families are attached.
Begin with the actual input files and identities
A PICRUSt2 project needs representative sequence identifiers and abundance records that correspond correctly across its inputs. An ASV called Feature_17 in one file must identify the same nucleotide sequence in the other. If identifiers are silently changed between processing steps, a perfectly executed programme may associate abundances with the wrong sequences. That destroys the biological meaning while leaving outputs that still look orderly.
Check the sample names and metadata table before reading any predicted pathway plot. Record soil depth, sampling date, treatment, extraction batch and sequencing run. Confirm which sequences were retained after filtering and whether the prediction workflow excluded certain ASVs. An output should never be treated as a magical conversion from an unexplained bar chart into function. Every transformation must preserve the identity of its biological and computational units.
The reference phylogeny is a model of relationships
PICRUSt2 uses a phylogenetic reference to relate marker sequences to organisms with genomic data. The idea is that evolutionary relationships may carry useful information about inherited gene content. However, related organisms can have different functions, and some genomes contain genes acquired through horizontal transfer, gene loss or environmental adaptation. Sequence similarity is evidence for a prediction, not proof of the exact contents of an unsampled organism.
Soil communities include many lineages that are distant from thoroughly sequenced reference genomes. A placement may be technically possible while leaving substantial uncertainty about specific genes. The result depends on the reference tree, genomic trait database, evolutionary model and tool version. A careful explanation says ‘predicted from related reference genomes’ rather than ‘detected within the sampled soil organism’ when no such gene was directly sequenced.
What hidden-state prediction means
In this setting, a hidden state is an unobserved trait associated with a marker-sequence lineage, such as an estimated count of a gene family. The programme applies an evolutionary prediction method to the reference tree and measured trait states at reference tips. It then estimates the state for study features without direct genome observations. The exact model and assumptions vary with the selected algorithm and reference set.
This is a statistical inference, not a recovered sequence of the target gene. An ASV may receive a predicted value for several orthologue families even though the laboratory only amplified one marker segment. Readers should distinguish the marker observation, the reference genome measurements and the inferred gene-content vector. Keeping these three levels separate allows a report to explain both what the method contributes and what a follow-up experiment would need to validate.
Why 16S copy-number correction appears in the pipeline
Bacterial and archaeal genomes can carry different numbers of 16S ribosomal RNA gene copies. When an amplicon table is used to estimate organism-associated functional content, variation in marker copies may distort the relationship between reads and underlying genomic representation. PICRUSt-style workflows use predicted marker-copy information to help adjust abundance contributions before estimating community gene content.
Such correction is not guaranteed to recover absolute cell numbers. It depends on predicted copy number, marker recovery, primer behaviour and the relationship between amplified reads and original DNA. If the correction changes sample rankings, investigate which lineages are affected. In prose, call the values corrected model estimates, not observed microorganisms per gram of dry soil. A mathematical adjustment removes only the biases it was designed and able to address.
Nearest-sequenced taxon index is about reference distance
NSTI summarises the phylogenetic distance between an input feature and its nearest sequenced reference taxon under the relevant tree and definition. Lower values generally indicate closer reference relatives, while larger values suggest greater reference distance and, potentially, less reliable functional predictions. This is not a probability of accuracy, and it is not a percentage of sequence reads found to be correct.
There may be per-feature NSTI scores and a sample-weighted summary using feature abundances. The latter depends on the retained features and their counts. A sample with one poorly represented dominant lineage may show a different weighted NSTI from another containing many well-referenced lineages. Always state whether the plotted number represents one ASV, an unweighted average or a weighted sample summary.
A worked weighted-NSTI example
Imagine two input features in an illustrative soil sample. Feature A contributes 80% of the retained marker feature abundance and has NSTI 0.10; feature B contributes 20% and has NSTI 0.50. A simple abundance-weighted mean is 0.8 times 0.10 plus 0.2 times 0.50, equal to 0.18. This differs from an unweighted arithmetic average of 0.30. The weighting changes which feature influences the sample summary.
The calculation demonstrates why reading a single reported NSTI requires understanding its denominator. It does not imply that 18% of functional predictions are wrong. Nor does it establish an accuracy rate for a particular enzyme. Real validation depends on the genes, lineages and reference representation. A report can use NSTI to flag uncertainty but should never treat it as a calibrated universal error percentage.
Filtering distant sequences affects who remains
Some versions and settings of PICRUSt2 use NSTI-related thresholds to exclude features judged too distant from available genome references. If poorly represented sequences are filtered from one soil group more often than another, the prediction can become a profile of the better-represented fraction rather than the full marker survey. A colourful pathway chart may hide this exclusion unless the researcher reports it.
Ask for the fraction of sequences and, when useful, the fraction of total read abundance retained or removed under the chosen settings. Distinguish the relative share of predicted functions among retained features from the proportion of the original community that the model could represent. An apparent treatment difference may be partly a difference in prediction coverage. Sensitivity analyses and transparent exclusions are essential when interpreting soil lineages with poor genomic references.
‘NSTI is low’ does not validate every function
A lineage can be phylogenetically close to a reference while differing in gene content. Genome evolution includes gene gain, loss, recombination and horizontal transfer. Even strains assigned similar names can have distinct metabolic capacities. Consequently, a favourable NSTI does not guarantee that a particular rare function was predicted correctly, especially when the trait evolves rapidly or is carried on mobile genetic material.
A scientifically careful conclusion might say, ‘Reference proximity supports using the predictions as exploratory evidence, but individual gene assignments require independent verification.’ It is especially important to avoid direct ecological process claims when the inferred pathway has few supporting genes or involves genes with uncertain annotations. The metric is a guide to one important uncertainty dimension, not a substitute for biological proof.
Gene-family tables contain predicted quantities
Functional output tables may contain rows representing predicted orthologue or enzyme families and columns representing samples. Depending on settings, values may reflect predicted gene copies or scaled community-level contributions rather than raw sequencing reads. A label such as KO refers to a functional orthologue category associated with a database convention; it is not a count of demonstrated enzymatic reactions in soil.
Read the documentation of the exact PICRUSt2 version and selected output format. Do not assume a table named ‘metagenome’ was produced by sequencing metagenomic DNA. Ask whether values were predicted by feature, summed across features, normalised by sample or transformed for a plot. This prevents one of the most common reporting errors: calling inferred gene families ‘detected shotgun genes’ when they were obtained from marker data.
KO tables are not lists of observed proteins
A KEGG orthologue identifies a category used to connect genes with known biological knowledge. A sequence-associated prediction about a KO does not automatically establish that a protein was made or had catalytic activity. Some functions are shared across multiple pathways; annotation choices and database definitions can change. A single KO can support several hypotheses but rarely determines an entire process by itself.
If a study claims that nitrogen transformation increased because one KO’s predicted share rose, check whether that KO is specific to the claimed step. Is its gene function distinctive? Are alternative pathways possible? Is the environmental setting compatible with the process? A thoughtful reader can appreciate the predicted association without converting a database label into measured soil chemistry.
Enzyme Commission identifiers have their own scope
EC numbers classify enzyme-catalysed reaction types, but their relationship to genes is not always one-to-one. Different sequences can encode enzymes with the same classified activity, and one protein may be multifunctional. A predicted EC-family profile should be interpreted as modelled potential linked to a reaction category rather than a direct measurement of an enzyme’s activity in the sample.
Use an enzyme assay when the goal is to measure catalytic conversion under specified laboratory conditions. The soil enzyme activity guide explains why assay rates depend on substrates, conditions, blanks and reporting units. Predicted EC abundance and measured catalytic potential are complementary evidence, not interchangeable numbers.
Pathway predictions add another inference layer
A pathway profile may be derived from predicted gene families and a pathway database or reconstruction algorithm. It can report estimated pathway abundance, coverage or other scores. These are not automatically equivalent. Abundance relates to an inferred representation on a particular calculation scale; coverage describes how much of the modelled pathway structure is supported under chosen rules. Neither is a direct proportion of a biochemical reaction completed.
Read how pathways were mapped and whether alternatives, gap filling or shared gene families were considered. A pathway can be strongly represented in gene predictions while being inactive under the sampled soil’s temperature, oxygen level or substrate supply. Pathway coverage of 0.8 is not a measurement that a process ran for 80% of a cycle. Preserve the model’s technical meaning in plain English.
Functional genes can be shared across predicted pathways
One orthologue family may participate in multiple biochemical processes or pathway models. Adding abundance values from different pathways can double-count or otherwise misrepresent underlying evidence. Some programmes use probabilistic or rule-based criteria, making a pathway value non-additive with others. Readers should never assume a column of pathway scores can be summed to reveal total microbial activity.
When a report displays the ten most enriched pathways, inspect whether values are independent, comparable and normalised on the same scale. Separate the biological hypothesis from the arithmetic convenience of a bar chart. If two pathways share multiple predicted genes, their apparent co-enrichment may arise partly from the same input evidence. This is not necessarily an error, but the overlap should be disclosed.
Different functional databases answer different questions
Predicted functions may be described using KEGG orthologues, EC classes, MetaCyc pathways or other reference schemes. Each imposes its own vocabulary and biological groupings. A category called ‘carbohydrate metabolism’ in a broad database is not automatically equivalent to a precisely defined carbon transformation pathway. The choice of reference controls which functions can appear in the output.
When comparing studies, identify the database and version. A change in annotations may rename or reorganise categories even if the starting ASV sequences are identical. Preserve raw feature identifiers, analysis settings and versioned output files for reproducibility. Reporting only a figure without reference details can make an apparent biological difference impossible to separate from an annotation change.
Directly measured shotgun results are different evidence
In shotgun metagenomics, reads are obtained from many DNA regions and may be assigned to genes or assembled into longer sequences. PICRUSt2 infers functions from marker sequences and reference genomes without directly sequencing those functional genes in the environmental sample. Even if both methods produce tables with similarly named KOs, their evidence types differ.
The soil shotgun gene-and-pathway guide explains the directly sampled DNA approach. A meaningful validation study compares predicted and sequenced functional profiles under compatible processing and data conditions. A broad correlation does not ensure every specific gene or differential-abundance result agrees. Report agreement at the level actually tested.
Why prediction can perform differently across habitats
Reference genome collections are uneven. Frequently cultured or sequenced lineages may be well represented, while rare soil organisms, plant-associated specialists and many deep environmental lineages have sparse references. Prediction quality can therefore differ even between two soil horizons from the same site. An algorithm that performs well on one community should not automatically be assumed equally accurate for another.
Ask whether the study has relevant reference genomes, whether many ASVs are distant from those references and whether critical taxa are excluded. Different soil habitats may contain different proportions of organisms for which predictions are poorly supported. A responsible comparison includes reference coverage as part of the biological interpretation rather than treating every row in the output as equally reliable.
The connection between genome size and apparent abundance
Different microbial genomes contain different numbers of genes and copies of genes. Summed predicted gene-family contributions can reflect genome-size differences, marker copy corrections and compositional changes. A higher predicted count of a general function need not imply that the same number of organisms became more functionally specialised. The mathematical quantity is a modelled contribution under a chosen aggregation scheme.
When interpreting apparent changes, distinguish per-genome predictions from community-level aggregates. If a new group of organisms with large genomes becomes more represented in marker data, many gene categories may change together. That broad effect can be useful to study, but it cautions against attributing each difference to a separate adaptive response. Normalisation choices must be written next to the biological claim.
An illustrative example with two ASVs
Suppose ASV-A receives a predicted value of three copies of a hypothetical gene family per genome-equivalent contribution and ASV-B receives one. A simplified sample A has model-adjusted contributions of 100 for ASV-A and 50 for ASV-B. Summing the products yields 100 times three plus 50 times one, or 350 modelled gene contributions. If sample B has 50 and 100, the analogous total is 250.
This toy arithmetic illustrates how features with different predicted genome content affect the aggregate. It deliberately omits correction and normalisation steps that real pipelines may apply. The result is not 350 genes per gram of soil, nor evidence that 350 enzymes were active. It also depends on the validity of each feature-level prediction. The lesson is how the model combines input contributions, not a shortcut to biological abundance.
What normalisation changes
Functional outputs may be transformed into relative profiles, scaled totals or other reporting units. Once a sample is rescaled to a fixed total, each function’s share depends on the other predicted functions included. An increase in one predicted category’s fraction can coincide with unchanged absolute gene abundance, or with changes elsewhere in the mixture. This compositional constraint deserves explicit attention.
Before saying a gene ‘increased’, identify whether the plotted number is unnormalised predicted content, a scaled contribution or a percentage of all included predictions. Write ‘relative share of predicted orthologue abundance’ when that is what the chart supports. Avoid the stronger ‘organisms produced more enzyme’ without measured RNA, protein or reaction rates. A denominator can be as important as the numerator.
A common error: treating percentages as cell counts
Imagine a functional plot in which a predicted metabolic category rises from 10% to 20%. A reader may say that twice as many microbes acquired the ability. That conclusion does not follow. The change describes a share of the categories under the prediction and normalisation model. It does not directly count cells, establish the number of genomes or measure gene expression.
A precise alternative is: ‘The category’s relative representation in the inferred functional profile increased from 10% to 20%, under the reported pipeline.’ The result may still be interesting. It suggests a candidate process worth testing. The repaired sentence becomes more useful because a future experiment can be designed to measure the missing evidence instead of assuming that evidence already exists.
Another common error: calling predicted pathways measured flux
A pathway with high predicted abundance can be inactive when necessary substrates, oxygen conditions or other environmental prerequisites are absent. Soil is spatially heterogeneous: aerobic and anaerobic microenvironments may coexist at small scales. A bulk sample may contain organisms with pathway genes but not show the related process at the moment or depth of sampling.
If a claim concerns carbon dioxide release or nutrient transformation, connect the prediction to independently measured rates. The soil respiration guide covers gas-flux evidence under specific conditions. Neither respiration nor a prediction alone identifies every contributing organism. Combining evidence is valuable only when each technique retains its own meaning.
Why statistically significant predicted differences still need validation
Suppose a statistical test reports an adjusted p value below 0.05 for a predicted gene category. That is evidence about the modelled numbers under the test, not a verification of the gene’s presence in the original soil. Errors in reference-based predictions may be systematic; testing those predictions precisely cannot eliminate underlying model bias. An extremely small p value does not prove a pathway operated.
Describe the finding as a group difference in predicted functional profiles. Then evaluate feature coverage, biological replication, study design, compositional effects and false-discovery control. For specific actionable claims, prioritise direct evidence. The soil differential-abundance guide explains log-fold changes and adjusted significance in sequencing-based comparisons.
Rare and abundant functions need different care
Low predicted values may be driven by a few sparsely detected features, placing the result at the edge of both sequencing sensitivity and reference prediction quality. Dominant categories are typically more stable numerically, but their biological meanings may be broad. A dramatic fold change from a very small modelled baseline can sound impressive without contributing much to a community’s overall gene potential.
Inspect feature prevalence and input read depth before ranking results by fold change. Consider the difference between a statistically robust broad function and a poorly supported rare predicted enzyme. Both can inform research, but neither should be reported with identical confidence. If a result rests on one rare ASV with a large NSTI, say so, and treat it as a hypothesis requiring specific follow-up.
Predictive confidence is not a single number
A sample-weighted NSTI captures reference proximity, but overall prediction reliability also depends on sequence placement quality, marker representativeness, genomic trait variability, the selected database, pathway mapping and sample preparation. Some of these sources of uncertainty may not be expressed in an output table. A reader should avoid searching for one threshold that magically separates reliable science from unreliable science.
Instead, construct an evidence checklist: which traits are predicted, how variable are they across close relatives, which reference genomes exist, how much of the sample is represented, and are direct validation data available? This method encourages useful caution without pretending all model outputs are equally uncertain. It is more honest than labelling an entire study ‘accurate’ because its average NSTI looks favourable.
Version control is scientifically important
PICRUSt2 evolves, and changes in reference databases, default settings or pathway resources can change output even for the same input feature table. A study should record software version, reference set, relevant parameters and functional classification sources. If two analyses disagree, software provenance belongs on the comparison checklist alongside biological sampling and experimental conditions.
Do not assume historical defaults apply unchanged to every release. Some documentation is version specific, and methods borrowed from older tutorials may not match the current software. The PICRUSt2 project documentation provides the current entry point for installation, output interpretation and limitations. When a reader reproduces an older analysis, matching software and reference versions can matter as much as matching sample names.
Quality-control checks before interpreting a pathway figure
Verify that every input ASV has a valid representative sequence, that sample names match across files, and that the predicted output covers the expected inputs. Inspect failures of sequence placement and any NSTI-related exclusions. Review the per-sample contribution of excluded features, distribution of reference distances and warnings about functional predictions. A missing output file should not be silently treated as a legitimate zero.
Then check normalisation, pathway coverage and the exact functional database used. If one treatment group loses many features while another retains almost all, a comparison may mix biological and model-coverage differences. A good report includes this possibility before presenting apparently precise treatment conclusions. Quality control is part of interpretation, not merely a troubleshooting appendix.
Replicates and design remain essential
Technical sequencing libraries cannot substitute for independent soil plots. If a treatment is applied to one field and the control comes from one other field, sequencing many aliquots does not independently replicate the treatment. Soil depth, plant roots, location and season can confound a comparison. Even perfect functional prediction cannot determine which external factor caused an observed difference when the design does not separate those factors.
Use biological replicates and a design-appropriate analysis. Show group variability, not just means. Distinguish a correlation between predicted gene content and treatment label from evidence of a treatment effect. If predicted profiles change across seasons, repeated sampling and temporal dependence must be considered. This basic discipline protects interpretation more than an extra decimal place on any prediction score.
Case study: a misleading claim about nitrogen
Imagine a report saying ‘PICRUSt2 proves that treated soil removes nitrogen twice as fast.’ The supporting figure shows a rise from 6% to 12% in the relative representation of selected predicted nitrogen-related gene families. There are no measurements of nitrogen pools, gaseous products, gene expression or transformation rates. The claim moves from inferred genetic potential to realised ecological flux without evidence.
A repair would read: ‘The selected nitrogen-related categories represented a larger share of the predicted functional profile under treatment, suggesting a hypothesis for direct nitrogen-process measurements.’ This wording keeps the interesting pattern visible while identifying what is missing. Follow-up tests could include targeted genes, transcripts, substrate incubations and appropriate chemical measurements. The reader now knows how to evaluate the claim rather than just admire a headline.
What a careful comparison with shotgun data looks like
If matched samples have both PICRUSt2 prediction and shotgun metagenomic profiles, compare the same functional categories on compatible scales and document any differences in taxonomy, mapping and normalisation. A useful analysis may assess correlations, ranking agreement, false positives or task-specific predictive performance. Broad overall agreement does not imply high accuracy for every rare function.
In a written summary, specify whether validation was performed on the same soil samples or borrowed from a different habitat. An evaluation carried out in one biological system cannot guarantee performance everywhere. If the project has not undertaken direct validation, say predictions are exploratory. This is especially appropriate for soil, where many genomic lineages and ecosystem conditions may be underrepresented by references.
How to explain PICRUSt2 to a school learner
A helpful analogy is a library catalogue. You recognise a book’s series and author from its cover and use knowledge of related books to predict its likely chapters. The prediction may be informed and useful, but it is not the same as opening the book and reading every page. Here, the marker sequence supplies the familiar clue, the reference genomes supply examples, and the functional table is an informed reconstruction.
The analogy also reveals the weakness: related books can contain different chapters, and some authors have no catalogue record. A student should be able to say which information was observed and which was inferred. This skill extends beyond microbiology into scientific modelling, historical reconstruction, language translation and interpreting automated classifications. Precision starts with naming the kind of evidence.
Context and scientific vocabulary
Terms such as ‘predicted’, ‘inferred’, ‘abundance’, ‘coverage’, ‘functional’, ‘significant’ and ‘validated’ can carry different meanings in everyday and technical English. ‘Functional potential’ does not guarantee observed activity. ‘Coverage’ may refer to represented features or inferred pathway components rather than a sample’s physical area. ‘Significant’ describes a statistical result under assumptions, not automatically an important environmental outcome.
Use the cross-linguistic transfer guide to examine how apparently familiar words can change meaning across contexts, and translation and meaning across languages to practise preserving evidence relationships. A good science glossary names the original variable, unit and inference limit rather than offering an isolated synonym.
A short checklist for a parent, learner or researcher
Ask what marker was sequenced and whether it adequately represents the biological question. Identify the source and version of the reference genomes. Inspect the proportion of ASVs and reads retained, the NSTI distribution and the treatment of distant taxa. Confirm the predicted family or pathway’s definition and numerical scale. Find independent soil samples, variation, statistical assumptions and any validating experimental evidence.
Then compose two sentences: one reporting the predicted observation and one describing what has not been directly measured. If neither sentence can be written without guessing, request the methods or source table. This habit is useful for reading papers, marking student explanations and preparing evidence-based environmental decisions. It replaces impressive but vague claims with observations another person can test.
Frequently asked questions
Does PICRUSt2 read functional genes directly from soil DNA? No. It predicts functional content from marker-gene community data and genomic references. Is NSTI a model accuracy percentage? No. It is a phylogenetic reference-distance measure under the relevant model. Does a lower NSTI always guarantee that every function is correctly inferred? No. Gene gain, loss and reference limitations still matter.
Can PICRUSt2 use any marker without extra considerations? No; reference resources and workflow compatibility must match the marker and analysis. Does a higher KO value prove higher enzyme activity? No. Can a pathway have high inferred coverage but low measured flux? Yes. Do two studies with different pipeline versions necessarily produce comparable numbers? Not without checking methods. Can PICRUSt2 help design a follow-up experiment? Yes, by prioritising functional hypotheses for independent tests.
Independent practice with explained answers
Suppose sample A has 90% of its marker reads represented by features with moderate reference proximity, while sample B has only 50% represented because many divergent features were excluded. A predicted pathway is twice as prominent in A. What is the first caution? The modelled populations differ in reference coverage, so the apparent pathway difference may partly reflect which organisms contributed to the prediction. Compare retained abundance and sensitivity before drawing biological conclusions.
Another question: a PICRUSt2 table shows a 1.5-fold predicted increase in an enzyme-related category but a laboratory assay records no clear change in catalytic activity. Is that impossible? No. Predicted genetic potential and measured enzyme behaviour differ. A third question: a feature’s NSTI is 0.4. Can you translate that as 40% incorrect gene predictions? No. The units and calibration do not permit that claim.
Further sources and learning routes
Read the PICRUSt2 repository and official documentation, its documentation and tutorials, and the QIIME 2 feature-table documentation for the observed marker inputs. The authors’ documentation explains source placement, predicted traits, reference coverage and version-dependent output conventions. Methodological resources are starting points for correct interpretation, not substitutes for checking the actual study’s settings.
For adjacent evidence, compare the soil amplicon ASV guide, soil shotgun functional-gene tables, soil differential-abundance results and soil enzyme activity. Each resource answers a different measurement question, so choose the one that matches the claim you want to evaluate.
Closing lesson: inferred potential deserves precise language
PICRUSt2 can make a marker-gene survey more informative by attaching testable hypotheses about genome content and pathway potential. Its value is greatest when readers understand what is directly sequenced, what is inferred from evolutionary and genomic references, and what remains to be demonstrated experimentally. Predicted families and pathways are not imaginary, but they are not direct measurements of soil function.
When you encounter a PICRUSt2 chart, name the marker, reference set, NSTI context, output scale and key exclusions before announcing a biological conclusion. Then ask what independent result would verify the proposed function in that soil. That disciplined translation transforms an appealing functional plot into a clear explanation grounded in evidence, uncertainty and the next sensible experiment.