Loading Pyodide & ML libraries...
This may take 1–2 minutes on first run (internet required).
Initializing...
GEO ML Analyzer
Machine-learning classification of GEO series matrices and gene-expression tables — fully in your browser.
Classify a gene-expression matrix without uploading it
Load a GEO series matrix (GSE…_series_matrix.txt.gz) or a samples-by-features CSV/TSV. The tool detects the class-label column, ranks differential features with ANOVA, and compares logistic regression, SVM, random forest, gradient boosting and an MLP using stratified cross-validation with feature selection repeated inside every fold.
- 1. Load — drop a matrix or browse for a file.
- 2. Configure — choose the label column and CV settings.
- 3. Run & export — read the metrics and download a DOCX/PDF report.
All computation runs in your browser through Pyodide/WebAssembly. Nothing you upload is sent to a server; the Python packages are downloaded from a CDN on first use.
About the GEO ML Analyzer
The GEO ML Analyzer answers a practical question: given a gene-expression matrix with known sample groups, which classifier separates those groups, and which features drive the separation? It fits logistic regression, a support vector machine, random forest, gradient boosting and a multilayer perceptron, then reports accuracy, macro F1, precision, recall and out-of-fold AUC for each. Every score comes from your own file — there are no demo numbers or simulated results in the output.
Input is a GEO series matrix table (GSE…_series_matrix.txt, optionally gzipped) or a delimited CSV, TSV or semicolon-separated matrix. Compressed inputs are unpacked in the browser, including plain ZIP and TAR archives. GEO phenotype lines beginning with !Sample_ are converted into candidate label columns, so a series matrix usually needs no editing before analysis. A features-by-samples table is transposed automatically; a samples-by-features table is used as it is.
The pipeline avoids the most common source of inflated performance in small omics studies. In each of the 3 or 5 stratified folds, a one-way ANOVA F-test is recomputed on the training samples only, the top 50–500 features are selected, and z-score scaling is fitted on that training fold. The held-out fold is never used for selection or scaling, so the cross-validated metrics are not contaminated by selection leakage. Optional 20-shuffle label permutation supplies an empirical p-value for the best model.
What this tool does
GEO series matrix and archive parsing
Loads GSE…_series_matrix.txt(.gz) and plain CSV/TSV tables, including .bz2, .zip and .tar.gz inputs. It detects the encoding, finds the !series_matrix_table_begin marker, turns !Sample_ metadata into label candidates and transposes features-by-samples matrices automatically.
Five classifiers compared side by side
Logistic regression (max_iter 800), linear SVM, random forest (80 trees), gradient boosting and a multilayer perceptron are trained on the same folds. The comparison table shows accuracy, macro F1, precision, recall, AUC and the per-fold mean and standard deviation.
Leakage-free feature selection and scaling
ANOVA F-ranking and z-score normalisation are recalculated inside each training fold from training data only. Choose 3- or 5-fold stratified cross-validation and top-K values of 50, 100, 200 or 500 features.
Differential feature statistics
Every feature receives an ANOVA F statistic, p-value, Benjamini–Hochberg FDR q-value and an effect size (Cohen's d for two groups, maximum class-mean difference divided by SD for more). Class means are shown alongside the top 40 ranked features.
Permutation test
Optionally shuffles the labels 20 times and repeats the entire fold-wise pipeline for each shuffle. The resulting null distribution gives an empirical p-value for the observed accuracy, which is a stronger check than a single accuracy figure on a small cohort.
Publication-ready figures
Class distribution, PCA projections, a volcano or effect-size plot, a heatmap of the most significant features, per-class boxplots, ROC curves for the best model and for all models, a confusion matrix, a learning curve, sample-level QC and feature importance.
Reproducible report export
Download a DOCX or PDF containing the dataset fingerprint, methods, model and per-class tables, differential features, every figure, limitations and software versions. Also export metrics CSV, the full feature CSV, all figures as a ZIP, and the complete results as JSON.
Privacy by construction
Parsing, feature selection, model fitting, plotting and report generation all execute in the browser tab through Pyodide/WebAssembly. The dataset is written to a virtual file system in memory, never sent to an analysis server, and the tab's memory is released when you load another file.
Frequently asked questions
What is a GEO ML Analyzer used for?
It tests whether sample groups in a gene-expression dataset can be separated by a classifier, and which genes or probes contribute most to that separation. Typical uses are checking whether a disease and control signature is real, comparing candidate classifiers on a small cohort, teaching cross-validation and feature selection, and producing a reproducible exploratory report before committing to a larger validation study. It reports cross-validated estimates on the dataset you supply; it does not claim performance on unseen cohorts.
Which file formats does the tool accept?
The primary input is a GEO series matrix table, usually named GSE12345_series_matrix.txt.gz. Plain .txt, .csv and .tsv tables with comma, tab or semicolon delimiters also work. Gzipped and bzip2-compressed files are read directly, and ZIP or TAR archives are unpacked to find the first table inside; a series-matrix member is preferred when several files are present. Excel workbooks, PDFs and raw FASTQ-style TAR archives are rejected with a message explaining what to download instead.
How does the tool find the class labels?
For a GEO series matrix it reads the !Sample_ metadata lines and turns suitable ones into ordinary columns, mapping values back to samples using geo_accession where possible. In a plain table it looks at object columns and low-cardinality columns, favouring names such as class, group, disease, label, type, condition or outcome. You can override the automatic choice from the dropdown or by typing an exact column name, and a hint reports the number of classes, their sizes and any problems before you run the analysis.
Why is feature selection repeated inside each cross-validation fold?
Selecting the top genes once on the whole dataset and then cross-validating leaks information from the test folds into the model, which makes accuracy look better than it is. This tool recomputes the ANOVA F statistic on the training portion of each fold, selects the top-K features there, fits z-score scaling on those same training samples, and only then evaluates the held-out fold. The reported numbers are pooled out-of-fold predictions, with fold-wise mean and standard deviation shown separately.
What do macro F1, AUC and the permutation p-value mean?
Macro F1 is the unweighted mean of the per-class harmonic mean of precision and recall, so a small class counts as much as a large one — useful when groups are unbalanced. AUC summarises the ranking quality of the model's out-of-fold scores across thresholds; for more than two classes it is a one-vs-rest macro average. The optional permutation test shuffles labels 20 times, reruns the full pipeline and compares the observed accuracy with that null distribution, giving an empirical p-value of (1 + number of null results at least as good) / (shuffles + 1).
Is my gene-expression data uploaded anywhere?
No. The file is read into browser memory and handed to Python running under Pyodide/WebAssembly inside your tab. Parsing, model fitting, plotting and report assembly all happen on your device, and nothing is posted to an analysis endpoint. The only network requests are for the runtime and libraries from cdn.jsdelivr.net on first load and the optional PDF library when you request a PDF. This makes the tool suitable for unpublished or embargoed expression data, subject to your institution's own policies.
How large a dataset can it handle?
The tool enforces a 500 MB input ceiling, but the practical limit is your device memory because the matrix, models and figures share the browser sandbox. A few thousand samples by a few tens of thousands of features is usually workable on a modern laptop; very large cohorts may exhaust memory or take a long time. If you hit a memory error, filter to the probes and samples you need, close other tabs, or run the same scikit-learn pipeline in a local Python environment.
Which model should I trust?
The tool marks the model with the highest macro F1 as best, but that ranking is a result of this dataset and these folds, not a universal verdict. Compare the cross-validated F1 and AUC together with the fold-wise standard deviation: a model with slightly lower mean F1 but much smaller fold-to-fold variation may be more dependable. With small classes, inspect the per-class table and confusion matrix rather than accuracy alone, and use the permutation p-value to judge whether the overall signal is stronger than label noise.
Can I use the results in a paper or thesis?
The reports are designed for reproducibility: they include the input file's SHA-256 fingerprint, class sizes, preprocessing choices, fold count, feature count, library versions and complete tables. That makes them useful as exploratory evidence, teaching material and a methods appendix. Before publishing a biological claim, validate the signature on an independent cohort and check the analysis against your field's reporting standards; feature FDR values here are discovery statistics computed on the same data used for ranking, not independent validation.
Does the tool work offline or need an account?
There is no account and no sign-in. An internet connection is required the first time you open the tool to download Pyodide, pandas, NumPy, SciPy, scikit-learn and matplotlib from a CDN, and when you first request a PDF; after those assets are cached, the analysis itself runs locally. If your network blocks the CDN, the runtime cannot start and the page will show a clear error rather than a silent failure.
Related tools
All processing happens locally in your browser. Nothing you upload is transmitted or stored, as described in the privacy policy.