eduKateSG · Why Science?
A functional-genomics value becomes interpretable when its Series, Sample, Platform and processing history stay connected
Trace a GEO study from its GSE Series through GSM Samples and GPL Platform, then separate submitter-supplied processed values from linked raw data and your own analysis.
Functional genomics can produce a table with thousands of rows and many samples, yet the hardest question often comes before statistics: what does each value represent? NCBI’s Gene Expression Omnibus, or GEO, organises submitted studies so readers can connect a Series, its Samples, the measurement Platform and related files.
Those connections matter. A column labelled “control” is meaningful only when its organism, tissue, treatment, time point, replicate role and processing history are known. This guide turns GEO into an evidence-reading journey for science learning, data literacy and reproducible reuse.
Work with a Series that has a manageable number of Samples. Create a study notebook containing an accession map, a sample design table, a file manifest and an analysis log. These four objects answer different questions: what records are connected, what was compared, what was downloaded and what you changed. Keeping them separate prevents a cleaned table from erasing its own history.
Begin in the browser, then use structured files only after the human-readable design makes sense. Automation can retrieve hundreds of fields quickly, but it cannot decide whether “day 0,” “baseline” and an empty time field mean the same thing. That decision needs documented biological reasoning and, sometimes, a limitation rather than a forced category.
For every count, ask for its denominator. Six files may describe three biological samples measured twice; twenty columns may include controls excluded from the main analysis. Sample size, file count and measurement count should therefore occupy different cells in the notebook. This simple discipline protects statistical interpretation.
The chapters use the word processed carefully. Processing is not a defect: alignment, normalisation and summarisation often make data interpretable. The scientific task is to identify which transformations the submitter performed, which values the repository holds and which transformations you add. Those layers allow a result to be reproduced or challenged.
This makes GEO especially useful for secondary science and introductory bioinformatics. Students practise fair comparisons, variables, units, graph reading and source evaluation using authentic records. A successful project is not the one with the most dramatic heatmap; it is the one whose samples, values and analytical choices can be traced from conclusion back to accession.
Choose the outcome before choosing the chart. If the question concerns a small gene set across two conditions, a transparent table and dot plot may communicate more than a clustered heatmap. The visual should preserve individual Samples wherever possible and label the processed scale. Decorative colour cannot repair uncertain groups or unknown units.
Keep exclusions in an explicit ledger. For every omitted GSM, record the reason, the decision date and whether the rule was defined before inspection. Excluding a failed measurement can be justified; removing an inconvenient value without a reproducible rule changes the evidence. A sensitivity check can show whether the conclusion depends on one decision.
Protect metadata during cleaning. Do not overwrite original treatment text, convert accession-like strings to numbers or allow spreadsheet software to reinterpret dates and gene symbols. Work on a derived copy, keep a read-only source file and test that the number of rows and Samples remains stable after every join.
When analysis code is used, include a small verification report before the results: file names, checksums where available, column-to-GSM mapping, sample counts, missing fields and transformation summary. These checks are less glamorous than a final graph, but they let another reader decide whether the data object still represents the archived study.
Ethical reuse also means respecting what the archive and consent context permit. Public availability does not erase the need to interpret sensitive human data carefully or to follow relevant access conditions. For a school project, prefer questions and outputs that teach method without attempting personal or clinical conclusions from group-level research records.
Plan a stopping rule as well. If essential group labels, units or processing details cannot be recovered, the correct outcome may be a documented data-quality report rather than a biological comparison. Explaining why an analysis would be unreliable is a valuable scientific result because it prevents uncertain inputs from acquiring false authority through computation.
At review time, ask a partner to select one plotted point and trace it backward through gene or feature mapping, processed value, GSM Sample, GSE design and GPL Platform. If any link breaks, repair the provenance chain before interpreting the pattern.
That trace is the practical definition of trustworthy reuse in this guide.
Chapter index
Chapters 1–9: from accessions to experimental design
- 1. Learn the three core accession types
- 2. Read the Series before the matrix
- 3. Treat each Sample as an evidence object
- 4. Decode the Platform
- 5. Build the condition table yourself
- 6. Separate kinds of replication
- 7. Preserve biological context and missingness
- 8. Distinguish processed and raw data
- 9. Read the processing description
Chapters 10–17: from processing and links to reproducible reuse
A GSE accession identifies a Series, a GSM accession identifies a Sample and a GPL accession identifies a Platform. These records link to one another, but they carry different parts of the study description. Copying them into one unlabeled list makes the design harder to reconstruct.
Try this
Open a Series and find its linked Samples and Platforms. Build a small relationship map with the GSE at the top, GSM records below and GPL records to the side. Add any external archive links separately. The map shows why one accession cannot substitute for the others.
Use accession prefixes as a reading aid rather than a guarantee that every study has identical organisation. Record the human-readable title beside each identifier so a classmate can catch a transposed or mistyped accession.
Evidence checkpoint
Copy the GSE, GSM and GPL accessions into separate fields. Series, Sample and Platform records answer different questions.
The Series summary and overall design explain the submitter's intended experiment. Read them before downloading a matrix because column values alone rarely reveal the full comparison, treatment structure or sampling plan.
Try this
Rewrite the design as a question with named factors: which organism or model, which tissue or cell type, what intervention, which time points and what outcome was measured? If a factor is not stated, mark it unknown instead of guessing from a filename.
Then count the linked Samples and compare that total with the number expected from the design. A mismatch may reflect excluded controls, pooled material, technical repeats or an incomplete understanding. Investigate before running an analysis.
Evidence checkpoint
Read the Series summary and overall design before downloading a matrix. A table without its design cannot define the comparison.
CHAPTER 3 OF 17 · Identify GEO record types
3. Treat each Sample as an evidence object
A GSM record can contain source characteristics, treatment descriptions, extraction and labelling protocols, processing notes and links to data. It represents a submitted biological or technical sample, not merely a column in a spreadsheet.
Try this
Choose two Samples from different groups and compare their metadata field by field. Highlight the factors that should differ and any unexpected differences, such as collection time, sex, batch or protocol. These details can reshape a biological comparison.
Create one row per GSM in a design table. Preserve the accession even when you replace a long title with a shorter analysis label. The accession keeps every calculated result connected to the original record.
Evidence checkpoint
Treat each GSM Sample as a biological or technical object with source, treatment and protocol metadata, not merely a spreadsheet column.
The GPL Platform describes the measurement technology and feature identifiers used in an experiment. Depending on the technology, a row may represent a probe, transcript, genomic feature or another measurement target. It should not automatically be called a gene.
Try this
Inspect several platform annotations and note the primary feature ID, linked gene fields and annotation date where available. Ask whether multiple features map to one gene or one feature maps ambiguously. The answer affects aggregation and interpretation.
Platform annotations can age while gene nomenclature and reference sequences evolve. Save the mapping file or version you actually use. A later mapping may be useful, but it becomes a new analysis step that needs documentation.
Evidence checkpoint
Use the GPL Platform to understand feature identifiers and measurement technology. A probe, transcript and gene are not automatically interchangeable.
Do not infer groups from column order, colours or short sample names. Extract organism, tissue or cell type, treatment, dose, time point, genotype and other relevant factors from the Sample records into explicit columns.
Try this
Add a proposed group label only after the source fields are visible. Sort the table by accession and count each group. If one sample does not fit the planned comparison, inspect its metadata rather than forcing it into the nearest category.
This exercise is powerful for secondary science because it converts experimental design into observable data. Students see that a fair comparison depends on controlling variables and documenting exceptions, even when the experiment is reused from an online archive.
Evidence checkpoint
Build the condition table from Sample metadata and verify counts. Never infer groups only from column order or file names.
Biological replicates sample independent biological units, while technical replicates repeat measurement or processing. Repeated measurements across time or related individuals add other dependency structures. Treating them all as interchangeable inflates confidence.
Try this
Use Sample metadata and the publication to classify each relationship. Draw brackets around measurements that come from the same biological source. If the repository description is insufficient, report the ambiguity instead of declaring every column independent.
Count biological units at the level relevant to the claim. A study with six files may contain three biological samples measured twice. The file count is useful for download validation, but it is not automatically the experimental sample size.
Evidence checkpoint
Separate biological replicates, technical replicates and repeated measurements. They contribute different information to analysis.
CHAPTER 7 OF 17 · Reconstruct the study design
7. Preserve biological context and missingness
Organism, anatomical source, cell type, disease state, treatment, time and batch can all influence a functional-genomics value. Record the factors relevant to the question and keep the submitter's terminology before normalising labels.
Try this
Mark missing information as missing. Do not convert an empty field into “untreated,” “healthy” or “zero hours.” Absence of metadata is a limitation on interpretation, not evidence for a baseline condition.
A compact data dictionary helps: list each analysis column, its source GEO field, allowed values and any recoding rule. This makes group definitions transparent and prevents a helpful-looking spreadsheet edit from silently changing the experiment.
Evidence checkpoint
Record organism, tissue or cell type, treatment, time point and batch when provided. Missing factors belong in the limitations column.
GEO commonly provides submitter-supplied processed values and may link raw high-throughput sequence data to another NCBI archive. These are related evidence objects with different formats, storage locations and analysis requirements.
Try this
For each file, label it raw, processed or metadata based on the record and documentation. Record where it is hosted and which accession connects it to the study. A counts table is processed even when it looks numerically simple.
Decide which level your project needs. Reusing a processed matrix can be appropriate for a learning exercise, while reprocessing raw reads answers a different reproducibility question and requires a documented computational pipeline.
Evidence checkpoint
Distinguish submitter-supplied processed values from raw data linked to a sequence archive or other repository.
Numbers become interpretable only after processing is understood. Background correction, alignment, quantification, normalisation, filtering and summarisation can all change what a value means and which observations remain.
Try this
Translate the submitter's processing paragraph into ordered steps. Attach each step to an input and output object. If software versions or parameters are missing, record the gap; do not invent a complete workflow.
Compare the archive description with the methods section of the linked publication. Differences may reflect summarisation rather than error. Cite the source that supports each step and keep your own later transformation in a separate section.
Evidence checkpoint
Read the processing description before comparing numbers. Background correction, normalisation and transformation change value meaning.
Expression tables may contain counts, intensities, ratios, logged values or normalised estimates. Magnitude alone cannot reliably identify the scale. A value near ten could mean something very different across technologies and processing methods.
Try this
Read the value description, Series matrix notes and publication. Check whether zeros, negative values or decimals are expected. Plot a small distribution only after the semantic meaning of the measurement is established.
If you log-transform, centre or standardise values, write the formula and base. Preserve the original column or file so the change can be reversed. Analysis-ready data should retain a path back to the submitted values.
Evidence checkpoint
Check whether values are logged, centred, counts or intensities. Never decide from magnitude alone.
Turning a platform feature into a gene symbol is an annotation step. One feature may map to multiple transcripts, and symbols can be renamed or reused across organisms. Store stable identifiers where available alongside human-readable labels.
Try this
Create a mapping table with source feature, destination identifier, database or file version, mapping date and rule for duplicates. Count unmapped and multiply mapped features rather than quietly dropping them.
This provenance matters when comparing a GEO study with resources such as Expression Atlas. Apparent disagreement may come from different mapping or processing choices rather than conflicting biology.
Evidence checkpoint
Preserve feature-to-gene mapping provenance. Platform annotations and gene nomenclature can change after submission.
For sequencing studies, GEO guidance connects processed submissions with raw-data records held in sequence archives. Keep the GSE and GSM accessions together with the relevant project, experiment or run identifiers that you actually inspect.
Try this
Build a join table rather than relying on browser history. One sample may link to several runs, and a run should not be mistaken for an independent biological replicate. Verify expected file counts and sizes before analysis.
State the boundary of inspection. If you used only the GEO processed matrix, do not imply that raw reads were reanalysed. If you downloaded raw files, document retrieval date, checksums where available and the pipeline applied.
Evidence checkpoint
When raw sequencing data are linked, keep GEO and sequence-archive accessions together and record which files you actually inspected.
Supplementary files can contain matrices, annotations, scripts or sample sheets. A file's presence does not guarantee that its format, column definitions or role are obvious. Treat each as an evidence object with a name and purpose.
Try this
Create a manifest containing filename, size, format, reported role and associated accession. Open a text preview and check headers before importing. Spreadsheet software can alter identifiers or dates, so use tools that preserve the original text when possible.
Compare the manifest with the Series description and publication. If a promised object is unavailable or a file includes fewer samples than expected, record the discrepancy before attempting to repair it.
Evidence checkpoint
Treat supplementary files as evidence objects with names, formats and roles. A downloadable file is not automatically analysis-ready.
GEO’s SOFT format represents database entities and attributes in a structured text form. It can support reproducible parsing of Series, Sample and Platform metadata, especially when manual copying would be error-prone.
Try this
Parse a small record and compare selected fields with the human-readable page. Validate accession, title, characteristics and relationships before scaling the script to a large study. Structured does not mean interpretation-free.
Save the parser version and retain the raw export. If a field occurs more than once or uses inconsistent labels, make the handling rule explicit. Automation should expose messy metadata rather than hide it.
Evidence checkpoint
Use SOFT or other structured exports when they help preserve metadata, but validate parsing against the human-readable record.
An archive record contains submitter statements, a processed dataset supports observations and your analysis produces recomputed results. These layers may appear in one report, but they should remain visibly distinct.
Try this
Use headings such as “Repository description,” “Observed in downloaded values” and “Our analysis.” Place methods and code beside the third category. This prevents a submitted conclusion from being presented as if you independently reproduced it.
The same separation helps when reading Bgee, which integrates selected expression evidence under its own rules. Every resource adds interpretation that deserves a label.
Evidence checkpoint
Write observations, recomputed results and submitter statements in separate sections so provenance remains visible.
Cite the GSE Series for study context, relevant GSM Samples for sample-level claims, the GPL Platform for feature semantics and any linked raw archive record for raw files used. Add the primary publication when relying on experimental interpretation.
Try this
Place citations near the specific statement rather than listing one accession at the end of a long section. A reader should be able to tell whether a number came from the archive, a paper or your calculation.
Include access date and analysis date because records and software environments can change. Citation here is functional documentation: it allows another person to retrieve the same objects and evaluate the claim.
Evidence checkpoint
Cite the GEO accession, linked publication and any external raw-data record that directly supports the conclusion.
CHAPTER 17 OF 17 · Report reuse responsibly
17. Connect functional genomics to future learning
GEO joins biology, experimental design, statistics and computing. It can spark interest in genomics, ecology, health science or data engineering, but one reused dataset cannot promise a grade, qualification or career.
Try this
Use the Science Learning Hub for concept practice and the Education Hub for broader subject planning. Choose a next skill based on the evidence gap: fair comparisons, coding, graph interpretation or molecular biology.
The BioStudies guide offers a complementary view of study-level packages. Together, these resources show that data reuse begins with organisation and provenance long before an exciting result appears.
Evidence checkpoint
Connect data organisation, statistics and biological design to further science learning without promising grades or career outcomes.
Contents · Previous chapter · Continue at the Science Learning Hub
A final GEO study-reading routine
Before analysing a GEO dataset, reconstruct a compact design table with one row per GSM Sample and columns for the factors that define the comparison. Keep the GSE Series and GPL Platform accessions above the table. Then record which values are submitter processed, which raw files are linked elsewhere, and which transformations you perform yourself.
Finish with an audit count: number of samples expected, number found, number assigned to each group, number excluded and reason. This simple arithmetic can reveal a mistaken download or hidden sample before a sophisticated statistical method makes the error look convincing.
| Reading object | What it can support | Question to keep visible |
|---|---|---|
| GSE Series | Study summary, overall design and linked samples | What comparison did the submitter intend? |
| GSM Sample | Source, treatment, protocols and sample-level values or links | Which biological object does this record represent? |
| GPL Platform | Measurement technology and feature identifiers | What does each row or feature measure? |
| External raw-data link | Files held in another archive | Did I inspect the linked raw record or only GEO processed data? |
Sources and next reading
Substantive resource claims in this article were checked against the primary pages below. A database record is a structured route into evidence; it does not by itself prove a new biological conclusion. Recheck the live record, evidence labels and release context before using it in assessed work or research.
Primary sources
- NCBI GEO overview — accessed 24 October 2026.
- NCBI GEO submission and archive scope — accessed 24 October 2026.
- NCBI GEO high-throughput sequence data guidance — accessed 24 October 2026.
- NCBI GEO SOFT format and record entities — accessed 24 October 2026.