Microbial source tracking helps researchers estimate which sampled environments best explain the composition of a microbial community found somewhere else. If you need to interpret FEAST source contributions, a SourceTracker-style soil microbiome mixture, source-versus-sink labels, stacked contribution bars or an ‘unknown source’ estimate, start by asking which environments were actually measured and treated as possible sources.
A 40% contribution assigned to irrigation water in a FEAST result is not automatically proof that 40% of living soil microbes physically arrived through that water. The output is a model-based attribution of the sampled microbial feature composition under a specified set of candidate sources. Shared environmental organisms, sampling gaps, temporal change, compositional data and laboratory artefacts may affect attribution. The estimate must be explained as conditional on the model and reference sources.
This guide shows how to read soil microbial source tracking from sampling design and ASV tables through FEAST input metadata, source and sink selection, estimated fractions, unknown contributions, uncertainty, validation and meaningful decisions. It includes hand-checked illustrative calculations and communication examples. No fictional percentages are represented as measurements of an actual field site, and no model output is treated as a direct tracing experiment without supporting evidence.
Quick answer: what is microbial source tracking?
Microbial source tracking compares the feature profile of a target community, called a sink, with profiles collected from candidate source environments. A soil sink might be a newly restored plot. Potential sources could include older neighbouring soils, irrigation water, applied organic material, windborne dust or plant-associated communities. A method such as FEAST estimates the mixture contributions that best explain the target profile under its statistical assumptions.
The output is conditional. The algorithm can only compare information it has been given, alongside its mechanism for handling unrepresented sources. A large assigned share suggests compositional similarity consistent with the model, not a unique historical proof of movement. A good report distinguishes the modelled origin hypothesis from independent physical evidence about transport, survival, reproduction and microbial activity.
Define the ecological story before choosing source labels
Suppose a researcher wants to understand whether a recently revegetated soil resembles the original site soil or soil introduced through restoration materials. Another study might investigate whether microbial communities in crop fields receive detectable signatures from irrigation water. These are different questions. The candidate sources, timing and appropriate controls must match the ecological process under investigation.
Write down which material is the sink and why each source could plausibly contribute microbes to it. Do not define a source solely because its sequence profile is statistically similar. Similarity can reflect shared habitat conditions, widespread taxa or common processing artefacts. A defensible story combines biology, geography, management history and sampling with the model-based result, while keeping each kind of evidence distinguishable.
What FEAST is designed to estimate
FEAST stands for Fast Expectation-mAximization microbial Source Tracking. The method uses microbial community count data and source–sink metadata to estimate contributions from labelled candidate sources and an unknown component. The official implementation describes a source-contribution matrix with a row for each sink and columns for candidate sources plus unknown, with each sink row summing to one under the model.
This structure makes the output intuitive to plot, but mathematical neatness should not be mistaken for physical certainty. The algorithm estimates mixing proportions according to a model of observed microbial features. It does not watch cells travel, count organisms entering a field or verify that all contributing organisms were alive and functionally active. Source tracking provides a hypothesis about community origins that should be evaluated against external evidence.
Decide what counts as a source sample
A source is an environment represented by measured samples and labelled as a candidate contributor. For soil restoration, source profiles might come from stored topsoil, undisturbed reference plots, compost-like amendment material or another documented input. Multiple independent samples are preferable when a source is heterogeneous, because one aliquot may not represent the full range of its community composition.
A named source category should not hide how material was collected. If a ‘compost’ source contains only one preparation lot sampled on one day, its profile may not represent every amendment used over a season. If irrigation water varies over time, a single grab sample may not describe the water entering the soil across the relevant period. The quality of attribution depends heavily on the representativeness of candidate sources.
Decide what counts as a sink sample
A sink is the target community whose profile the model attempts to explain using candidate sources. A soil core collected after amendment may be treated as a sink, but its history could include resident microorganisms, seed-bank organisms, colonisation from multiple directions and local growth. The term does not imply that the soil is biologically passive or that all its microbes are newly imported.
Keep each sink’s field metadata visible: collection date, depth, plant cover, recent irrigation, amendment schedule and site identity. If a sampling campaign pools multiple sink plots, the output reflects the pool rather than separate spatial variation. Statistical models cannot reconstruct whether a source contributed differently to individual plots if those plots were physically combined before sequencing.
Timing can make or break the attribution question
Microbes detected in a potential source after a sink was sampled cannot automatically explain earlier immigration to that sink. Soil communities change with weather, plant development and storage conditions, and source materials may change as they age. When a study investigates movement, sample timing should correspond to plausible transport windows and subsequent persistence.
For example, irrigation water collected months after a soil event may share organisms with the sink but cannot by itself establish what was present in the water during the event. Temporal matching and repeated sampling improve ecological relevance. If timing is uncertain, explain that the model identifies compatibility with the sampled candidate profile rather than documenting a particular historical source.
Read the feature matrix before reading percentages
FEAST takes a matrix of microbial feature counts across samples. Those features may be ASVs or other operationally defined taxa. Their identities, filtering and sequencing depth affect how each source resembles the sink. If candidate sources and sinks were processed using different marker regions or incompatible pipelines, a comparison may be technically invalid or misleading.
Use the same feature definitions across input samples and verify identifiers. Examine retained read totals, zeros, prevalence filtering and taxonomic aggregation. The soil ASV interpretation guide explains why sequencing counts are not exact counts of living microbial cells. Source tracking inherits the same measurement boundaries, even when it converts profiles into clean-looking fractions.
The role of source–sink metadata
The official FEAST input metadata uses fields such as environment label, SourceSink classification and an identifier indicating which source set is eligible for a given sink. A row labelled Source represents a candidate reference environment; a row labelled Sink is a target. The grouping identifier matters when different sinks are analysed against different sets of possible sources.
Mislabelled metadata can silently change the question being answered. If a sink is marked as a source, it may influence estimates for other targets rather than receive the intended attribution. If a source is assigned to the wrong group, a model might use environmental references that could not have contributed under the field design. Validate metadata against the sampling ledger before interpreting a single result.
Different sinks may have different eligible sources
A restoration study across several geographically separated plots may not have one universal pool of candidate sources. Some plots received amendment A, others amendment B; some used groundwater, others a surface-water supply. FEAST supports configurations in which candidate sources are grouped differently for distinct sinks. This is scientifically useful because a source that was never present at a site should not necessarily be treated as an equal candidate.
The drawback is that attribution values from sinks with different source sets are not automatically comparable. A named source could receive a different share partly because the menu of alternatives changed. Every plotted percentage should be interpreted alongside the exact candidate list used for that sink. A comparison of source contributions must keep the denominator and reference universe visible.
What an output contribution matrix means
The FEAST output is described as a matrix with sink samples in rows and source categories in columns, including an unknown component. Each row sums to one. If a row contains 0.45 for source A, 0.30 for source B and 0.25 unknown, the values sum to 1.00. They express modelled shares of the sink’s microbial feature mixture under the fitted method.
They do not necessarily mean 45% of total soil biomass arrived from A, 30% of all living organisms arrived from B and 25% came from a newly discovered third source. Such statements replace a compositional inference with an unwarranted physical mass balance. The row-sum rule is a property of the model’s mixture proportions, not evidence that all real-world microbial transport and growth is fully explained.
A worked example with two sources and unknown
Imagine a sink soil profile analysed against two source classes: irrigation water and amendment material. A hypothetical result estimates 0.35 for irrigation water, 0.50 for amendment and 0.15 for unknown. The arithmetic is 0.35 plus 0.50 plus 0.15 equals 1.00, or 35%, 50% and 15% on the output scale. That describes conditional model attribution of the observed community profile.
The conservative summary is: ‘The model attributed the largest share of the analysed sink profile to the sampled amendment reference, followed by irrigation water, with a remaining unknown component.’ Do not write ‘half the living bacteria came from amendment’ without a validated transfer study. The interpretation may change if the candidate-source catalogue, marker data or fitting assumptions change.
What the unknown component represents
The unknown fraction accounts for contribution not assigned to the specified source profiles under the method. It can reflect genuinely unsampled environments, organisms poorly represented by candidate references, variation within sources, temporal mismatch, technical noise or limitations in the model. An unknown fraction is not a taxonomic discovery, a named missing habitat or a direct percentage of uncharacterised species.
Suppose the unknown value is 0.60. It would be misleading to conclude that 60% of the soil microbiome came from an exotic, unidentified source. The correct statement is that the model did not attribute 60% of the fitted sink profile to the supplied candidates. To investigate the missing contribution, broaden source sampling and test model performance rather than inventing a dramatic origin story.
Unknown is not equivalent to ‘contamination’
Contamination can be one possible issue, but an unknown component is not inherently a laboratory contaminant. Soil communities contain resident organisms, rare taxa and complex ecological mixtures that are not fully represented by any modest set of environmental references. A source-tracking result can contain a large unknown component even when samples were collected and processed carefully.
If contamination is suspected, examine extraction blanks, laboratory control sequences, processing batches and known contaminant patterns. An unknown share alone provides no proof of contamination or procedural failure. Likewise, a low unknown share does not certify that a study is contamination-free: candidate reference profiles may share contaminants or technical signatures with the sink.
A low unknown fraction does not prove the source list is complete
When all analysed sink features can be assigned among the candidate-source patterns, the unknown fraction may be small. This can look reassuring, but it does not establish that the actual physical origin of every microorganism has been identified. Closely related or environmentally ubiquitous organisms can produce similar signatures across multiple habitats. A model may find a strong composition fit even if the true contributing environment was not sampled.
Try to validate attribution with known-mixture experiments and independent physical histories. A good explanatory analogy is matching an unidentified document to a set of known writing styles: a strong match can be informative while leaving open the possibility that an unrepresented writer had a similar style. The model expresses what is plausible within its reference set, not the entire world of possible origins.
NAs are not zero contributions
The FEAST documentation indicates that an NA in its contribution matrix may mean a source was not used in the analysis for that particular sink. This is different from a numerical zero, which may mean a fitted contribution of zero or a value rounded to zero under a used-source model. Conflating the two can make a stacked plot or group average mathematically misleading.
When summarising several sinks, do not replace all missing source entries with zeros without first checking why the source is absent. Compare sinks only over compatible eligible-source configurations or explicitly explain the difference. The integrity of a source-tracking figure depends on distinguishing ‘model found no contribution’ from ‘candidate was not part of this comparison.’
Read scientific notation correctly
Some outputs contain extremely small numerical values, such as 2.7e-14, meaning 2.7 multiplied by ten to the power of minus fourteen. These represent values close to zero on the contribution scale. Display software may round them to 0.00%, and the apparent precision of tiny digits can exceed the meaningful certainty of the model.
Avoid describing a contribution of 2.7e-14 as a confirmed biological transfer event. It is ordinarily more informative to report that the fitted contribution was negligible at the chosen reporting precision. Still, preserve the underlying exact numbers in analysis files when needed for reproducibility. Rounding is a communication choice, not proof that the ecological possibility has been ruled out.
Read percentages with their conditional denominator
Every sink’s estimated contributions add to a total of one under the model. If the output says 45% from source A, the denominator is the fitted sink mixture, not necessarily dry soil mass, DNA molecules, individual cells or time-integrated migration. The percentage can change simply because another candidate source is added, removed or recharacterised.
Consider a model fitted with only sources A and B, then refitted with A, B and C. A’s estimated share might fall even if the underlying sink sample is unchanged. The estimate is relative to the collection of candidate profiles and modelling assumptions. This is why scientific statements must mention the candidate-source set and resist treating percentages as immutable properties of the soil.
Similar sources are difficult to distinguish
Irrigation water and amendment material could share many bacterial taxa, especially if they interact before reaching a field. Two neighbouring soil habitats may also have similar community profiles. When sources overlap strongly in feature composition, a model may struggle to attribute their shared signatures uniquely. Small changes in source sampling can shift apparent allocations between them.
A responsible analysis investigates how separable the source profiles are before interpreting precise differences. If a paper claims that 37% came from A and 36% from B, the tiny gap may not be meaningful when the two reference communities are nearly indistinguishable. Report uncertainty and acknowledge ambiguity rather than ranking sources with unjustified precision.
Source diversity is not automatically source strength
A candidate source with many observed ASVs may overlap with the sink simply because it includes a broad set of features. That does not mean it transported more cells or exerted a larger ecological influence. Richness depends on sampling effort, extraction and sequencing; fitted mixture contribution depends on how those detected patterns match the sink under the model.
Do not assume that the most diverse reference contributes the most. Inspect feature overlaps, unique and shared signals, sample replication and relevant environmental history. The same alpha-diversity measures describe within-source richness and evenness, but they answer a different question from source attribution.
A source-like taxon does not prove transport
Suppose a bacterial ASV occurs in compost and in a treated field after compost application. That pattern is consistent with transfer but also compatible with a resident soil organism, an organism present from another route, independent selection or a shared background community. Detecting the same marker sequence does not establish the direction, timing or viability of transport.
A stronger experiment might sample before and after application, track candidate sources over time and include untreated controls. Genomic resolution may help distinguish close lineages, but even precise identity alone does not prove which event occurred. Explain source tracking as supporting an origin hypothesis, with independent field evidence strengthening or weakening it.
Viability and activity remain unmeasured
DNA can persist after cells die, and marker counts do not necessarily reflect actively growing populations. Source tracking usually compares detected microbial sequence profiles. It does not independently test whether transferred microorganisms survived, reproduced, produced enzymes or changed soil ecosystem functions. Those outcomes require other measurements and carefully matched sampling.
If the reader wants to assess whether a soil amendment changed respiration, nutrient transformation or disease suppression, a mixture model cannot answer the whole question. The soil respiration guide covers measured gas responses, while enzyme assays and chemical studies provide different function-related evidence. These methods can complement attribution but should not be replaced by it.
Absolute cell numbers are not recovered by source fractions
A fitted contribution fraction cannot be multiplied by an unverified assumption about total cells to produce a trusted number of organisms from each source. Amplicon amplification bias, genome marker-copy variation and DNA extraction differences may alter relative read profiles. Even with qPCR-based total gene-copy estimates, source-specific absolute attribution requires justified calibration and comparable marker definitions.
For a source-specific quantitative interpretation, design a validation with known mixtures or independently measured markers where possible. Explain which amount is being estimated, in what units and with what recovery assumptions. Calling a FEAST fraction ‘cells per gram’ without such a bridge is a dimensional error: a fraction is unitless; cells per gram requires an actual measured reference quantity.
Rarefaction and coverage settings affect inputs
The FEAST implementation includes an input parameter for sequencing coverage or rarefaction depth. Library-size differences influence whether rare features are represented and can affect comparisons. Choosing a low common depth may discard much sequencing information, while a high threshold can exclude low-depth libraries. The programme’s behaviour should be checked against the documented version and study design.
Report how much data were retained, which samples were excluded and whether source profiles were measured at compatible depths. A result that changes drastically when the analysis depth shifts may not support a precise ecological attribution. Sensitivity analysis can help show which broad source conclusions are stable and which rely on uncommon features near the detection boundary.
Technical batches can masquerade as biological sources
Suppose soil sinks and all amendment source samples are processed together in one sequencing run, while water sources are processed separately in another. Shared laboratory contaminants or batch-specific artefacts could increase apparent similarity between sinks and the amendment references. A model could then attribute a source-like signature that partly reflects processing rather than ecology.
Balance sources and sinks across laboratory batches where practical, track extraction and library-preparation identifiers and include appropriate negatives. When source category and batch are perfectly confounded, statistical adjustment cannot reliably separate them. The right summary notes the limitation and recommends more balanced sample collection rather than declaring a definitive source hierarchy.
Field replication supports useful confidence
One sink sample can be fitted successfully, but a single value says little about how stable source contributions are across independent soil plots. A plot-level comparison should inspect the distribution of estimated shares among independent biological samples. A source that dominates one unusual core but not its neighbours may represent local heterogeneity, sampling noise or a transient event.
Show individual sink contributions or suitable uncertainty summaries, rather than hiding everything in group averages. If the study has repeated observations from the same plot, maintain that dependency when estimating variation. A broad source hypothesis becomes more persuasive when it recurs across independent sites and survives sensible changes to the candidate reference set.
Known-mixture tests check a different claim
A known-mixture experiment can combine previously characterised source materials at declared proportions, process the resulting samples and compare the model’s fitted estimates with those known proportions. This tests performance under the chosen experimental conditions. It may reveal bias, source overlap, low sensitivity or a large unknown component even when physical composition is controlled.
Such validation is not universally transferable. A benchmark combining clean laboratory cultures may not predict accuracy for complex soil with high DNA-binding minerals, changing environments and unsampled lineages. State the validation materials, sequencing pipeline and mixture range. A model that fits training samples well is not automatically calibrated for every field environment.
Leave-one-source-out analysis reveals dependence
One way to test sensitivity is to remove a candidate source class and rerun attribution. If shares shift strongly among the remaining sources or into the unknown component, the original percentages were highly dependent on the source catalogue. This need not invalidate the approach; it shows which conclusions rely on particular reference information.
Describe the observed stability at a sensible level. If amendment material remains the dominant assigned source under several plausible configurations, that is stronger evidence of a robust compositional pattern than a single untested ranking. If it disappears as soon as a related source is added, the correct interpretation is that the available sources are difficult to distinguish. Robustness is not a guarantee of historical truth, but it is useful evidence about model sensitivity.
A leave-one-location-out test is even more demanding
Suppose all sources and sinks from the same field share geographic signatures. Randomly dividing individual samples into training and test subsets may exaggerate model portability because neighbouring soil material appears in both. A more stringent validation might hold out a whole field, region or time point, depending on the intended use. It then asks whether the model can attribute new sinks beyond the closely related original sampling context.
The evaluation design should match the operational question. If the goal is to trace water-associated signatures within one irrigation system, one kind of holdout may be appropriate. If the goal is a global soil-origin classification, a far broader reference and validation structure is necessary. Avoid claiming one laboratory benchmark answers both questions.
What a stacked contribution bar actually shows
A typical plot assigns each sink sample a bar whose segments represent fitted source fractions. Colours distinguish candidate sources and often an unknown category. The segment widths or heights sum to 100% by construction. This makes the graphic easy to read but can hide uncertainty and similarities among the sources themselves.
Inspect whether bars represent individual sinks or group averages, and whether sources differ among sinks. If the chart uses percentages rounded to integers, the visible segments may sum to 99% or 101%; that is a display issue, not an ecological anomaly. A caption should state the model, marker, candidate-source set and whether uncertainty intervals or replicates are presented elsewhere.
Why pie charts may imply more certainty than the data support
A pie chart with 48% amendment, 32% water and 20% unknown can look like a precise inventory of where all microbes came from. The visual reinforces a whole divided into definitive origins, even when contributions are conditional estimates from overlapping reference profiles. The chart should not be allowed to overrule the method’s limitations.
A helpful caption might say: ‘Estimated contributions to the observed sink feature composition under the selected FEAST sources, including a modelled unknown component.’ That short description signals the correct level of inference. If source profiles are highly similar or under-sampled, make that limitation visible near the figure rather than hiding it in distant methodological text.
Distinguish source tracking from environmental gradients
If samples from a sink are more similar to wetland sources than dryland sources, the method might assign a larger wetland-related contribution. Yet common wetland-associated organisms may also arise independently through environmental selection within the sink. The observation could reflect a compositional environmental gradient rather than historical immigration. Source tracking alone may be unable to distinguish those explanations.
A beta-diversity analysis can show overall sample similarity, but it likewise does not prove movement. The two methods have different analytical outputs and related biological limits. Use physical histories and temporal sampling to test whether inferred source contributions correspond to plausible transport.
Source tracking and community networks have different jobs
A co-occurrence network explores statistical relationships among microbial features across samples. Source tracking compares a sink with environmental reference profiles. A network edge does not establish that one organism transported another, while a source attribution does not show that organisms directly interact. These are separate types of inference.
For readers encountering both figures in a paper, identify which unit each graph connects. An edge might connect taxa; a source-tracking bar connects a sink profile to candidate environment labels. The soil microbial network guide explains why correlation and causation must remain separate. The same caution applies to hypothesised transport.
A restoration example with uncertain sources
Imagine an illustrative restored meadow sampled after topsoil was brought from a donor area. Candidate source samples include donor topsoil, surrounding native meadow soil and irrigation water. FEAST attributes much of the sampled sink profile to donor topsoil, but the unknown fraction remains substantial. The field history makes donor soil a plausible contributor, while the unknown fraction shows that references do not fully explain the observed composition.
The disciplined conclusion is that the restored community shares a strong modelled component with the donor reference under the sampled marker and source catalogue. It is not proof that all assigned organisms colonised successfully or that restoration improved soil function. Complementary monitoring can test persistence, plant establishment and process measures independently.
Reading changes through time
Suppose modelled donor-topsoil contribution declines across three sampling dates while the local-source estimate increases. That pattern could be compatible with community succession or local colonisation, but it could also reflect changing reference profiles, sequencing effects, growth of resident organisms or altered environmental selection. A time-series result needs consistent sample processing and source definitions.
Report whether the same source references were applied across dates and whether the candidate environments were resampled. A moving fraction is a relative composition result, not a directly measured migration rate. If the study wishes to estimate turnover or transport flux, more targeted modelling and physical information will be necessary. Clear terminology prevents temporal charts from claiming processes they did not observe.
False precision and rounding choices
Model outputs can contain many decimal places because computers store numerical estimates at high precision. Ecological sampling and reference uncertainty rarely justify treating every digit as a biological measurement. A source contribution of 0.372846 should not become an assertion that 37.2846% of viable organisms came from that source. Report a reasonable level of numerical precision alongside the confidence supported by validation.
A useful rule is to let uncertainty and study design, rather than software defaults, determine significant digits in the text. If performance is not calibrated, use approximate shares or ranked patterns and state the limitations. Retain exact output values in the underlying data so others can reproduce the calculations. Readability and scientific integrity can coexist.
Vocabulary for careful translation
‘Source’ means a reference environment designated as a candidate origin; it does not always mean a physically proven contributor. ‘Sink’ is the target community being explained. ‘Contribution’ is a fitted profile proportion on a model scale. ‘Unknown’ represents the part not attributed to supplied candidates. ‘Similarity’ is not necessarily migration, and ‘detection’ is not necessarily activity.
These meanings can drift in ordinary language or translation across languages. Use the cross-linguistic transfer guide to examine how familiar words change technical meaning, and the translation-and-context guide to preserve the relationship between observations and claims. A good glossary defines the model’s units, not just its nouns.
A practical quality-control reading order
First identify the ecological transport hypothesis and sink definition. Confirm candidate sources are biologically plausible and were sampled at meaningful times. Examine the common feature table, marker region, depth and controls. Check metadata for correct Source and Sink assignments and source-set identifiers. Read fitted contributions, unknown values, missing entries and validation or sensitivity analyses.
Only after these checks should you interpret the stacked bars. Write down which claims concern compositional attribution and which require independent field evidence. If an input cannot be verified, mark the interpretation as conditional rather than filling the gap with an attractive story. A careful source-tracking summary should always be understandable from its source catalogue and sampling design.
Common mistakes and their repairs
Mistake one is treating model fractions as percentages of living microbial cells. Repair it by naming the fitted feature-composition scale. Mistake two is treating the unknown fraction as a newly discovered source. Repair it by explaining incomplete attribution. Mistake three is using source samples collected long after the sink event without temporal caution. Repair it through matched sampling and transparent limitations.
Mistake four is treating NA as a numerical zero when a candidate was never included. Mistake five is ignoring laboratory batch effects. Mistake six is comparing sinks with different candidate-source lists as if their denominators were identical. Mistake seven is mistaking a confident-looking bar chart for independently validated historical transport. Each repair restores a missing piece of the model’s definition or experimental design.
Frequently asked questions
Does FEAST directly watch microorganisms migrate? No. It models source contributions from community profiles. What does an unknown fraction mean? The share not attributed to supplied candidates under the model. Do contribution rows sum to one? The official output describes per-sink contributions summing to one, including unknown. Is NA the same as zero? No; NA can indicate a source not used for that sink.
Can a model assign a source even if no cells physically came from it? Similar community signatures and model limitations can produce misleading attribution, so independent evidence matters. Does 50% source contribution equal 50% of soil biomass? No. Can source selection change percentages? Yes. Are known-mixture validations useful? Yes, but performance depends on how well the validation resembles the field problem.
Independent practice and explained answers
A FEAST sink row reports source A equals 0.42, source B equals 0.33 and unknown equals 0.25. Do these add correctly? Yes: 0.42 plus 0.33 plus 0.25 equals one. Can you infer that 42% of the sink’s living cells came from A? No, because the fractions are modelled contributions to the analysed feature composition. The direct biological transport interpretation requires separate evidence.
Second, source C has NA for this sink. Should you plot it as a 0% estimated source without explanation? No; it may not have been eligible or used. Third, all source samples were extracted in one batch and all sinks in another. What concern should you raise? A processing effect could influence profile differences and attribution. A defensible analysis should check controls, batch balance and model sensitivity.
Methodological sources and useful learning routes
The FEAST project’s official repository describes inputs, source–sink metadata, fitted output matrices, unknown contributions and example usage. The original Nature Methods article on FEAST introduces the source-tracking framework. These sources explain what the method estimates; interpreting any particular soil result still requires the sampling ledger, input data and validation.
On eduKateSG, continue through soil ASV tables, soil microbial beta diversity and soil co-occurrence networks. Each provides a different lens on community profiles. Source tracking should be used when the actual question concerns plausible origins, not simply because the data can produce another colourful figure.
Closing lesson: estimated origin is not observed movement
Soil microbial source tracking can clarify which sampled environments most closely account for a sink community under a defined mixture model. It is especially useful for framing hypotheses about colonisation, amendment influence or environmental exchange. Its limits arise from incomplete reference sources, shared organisms, temporal change, compositional measurements and technical processing.
An accurate translation names the candidate sources, sink, marker, fitted units and unknown component, then explains what independent evidence would establish actual transport and ecological effects. A trustworthy report does not erase uncertainty merely because its percentages add neatly to 100%. It uses the model to ask better questions, preserve evidence and design more informative field studies.