On the summit stage, BioMate turns a real myasthenia gravis dataset into a developability verdict in three plain-English turns. That is the surface. Underneath, those three turns fan out into roughly thirty discrete phases — format detection, pseudobulk aggregation, robust dispersion estimation, output inference, enrichment, and ADMET. This is the companion to the demo announcement: the engineering, phase by phase, with the real numbers from the run.

3
plain-English turns
~30
backend phases
4
raw data formats handled
36,601
genes, 20 samples
The Real Problem

The Hard Part Was Never the Statistics — It Was the Raw Data

Every public dataset arrives in the shape its depositor happened to use. Most bioinformatics tools quietly assume one canonical layout — usually 10x Genomics matrix bundles — and simply fail on anything else. The case for this demo, GSE227835 (single-cell RNA-seq of AChR+ myasthenia gravis vs. healthy controls), ships as dense per-sample expression tables — one gzipped genes×cells matrix per subject — not the matrix.mtx/barcodes.tsv/features.tsv triplet most pipelines expect. A tool that only reads 10x MEX would stop before computing a single statistic.

BioMate’s ingestion layer detects and normalizes four distinct raw-data shapes automatically:

Bulk series matrix
Pre-quantified expression from the GEO series matrix — rounded to integer counts.
10x MEX bundle
Sparse matrix.mtx.gz + barcodes + features per sample.
Per-sample tarball
A .tar.gz per subject that unpacks to MEX files.
Dense genes×cells TSV
One gzipped dense table per sample — the format that broke naive readers, now handled natively.

Whichever shape it finds, the ingestion workflow pseudobulk-aggregates the single-cell counts up to the sample level, harmonizes the metadata, auto-detects the condition and contrast, and emits three DESeq2-ready files — counts_matrix.csv, sample_metadata.csv, and contrasts.csv. The user never sees any of it. They asked one question.

The Full Pipeline

Six Stages, ~30 Phases, Three Turns

Here is the complete flow. The three conversational turns map onto Stages 2, 4, and 5; Stages 0, 1, and 3 run invisibly around them.

0 Input & Routing
  1. Input intake — accepts a GEO/SRA accession, a download link, a file upload, or a typed question.
  2. Accession & intent detection — recognizes GSE/GSM/SRA patterns and the analytical intent.
  3. Routing decision — single workflow, a suggested multi-step chain, or a disambiguation question.
  4. Parameter prefill — rule-based + LLM hybrid; auto-fills the accession and defaults.
↓  accession → ingestion
1 Ingestion & Format Normalization
  1. GEO fetch — retrieve the series matrix and supplementary files.
  2. Format detection — branch across the four shapes above (bulk / MEX / tarball / dense TSV).
  3. Download & extract — pull and unpack the raw archive for single-cell paths.
  4. Per-sample read — sparse readMM for MEX, dense table read for TSV.
  5. Pseudobulk aggregation — sum counts per gene per sample, with a minimum-cells filter.
  6. Gene alignment — intersect the gene set across all samples.
  7. Metadata harmonization — auto-detect the condition column (disease vs. control).
  8. Contrast auto-detection — choose reference and target groups.
  9. Emit DESeq2-ready CSVs — counts matrix, sample sheet, contrasts.
↓  3 CSVs become the next stage’s inputs (data-connected)
2 Differential Expression — DESeq2 Turn 1
  1. Stage inputs — move data from object storage onto the compute worker.
  2. Load & coerce — integer coercion, missing-value clamp.
  3. Gene filtering — drop low-count genes below the expression floor.
  4. Robust dispersion & test — size factors, then a parametric → local → gene-wise dispersion fallback so small or low-variance cohorts never dead-end, then the Wald test.
  5. Results & significant genes — ranked table plus the significant-gene list.
  6. Diagnostics — volcano and PCA plots.
  7. QC metrics & gate — evaluated against configured thresholds.
↓  every run emits into output inference
3 Output Inference & Findings
  1. Output resolution — locate result files across the publish directory and work directory.
  2. Findings generation — read the result tables and summarize the key numbers in plain English.
  3. Report assembly — a cited, reproducible methods-and-results record.
↓  DE gene list becomes the enrichment input (data-connected)
4 Pathway Enrichment — clusterProfiler Turn 2
  1. Gene-list input & ID mapping — map symbols to ENTREZ identifiers.
  2. Over-representation analysis — enrichGO across Gene Ontology terms.
  3. Simplify — collapse redundant terms.
  4. Visualization — barplot, dotplot, cnetplot, emapplot, goplot.
  5. Results & findings — enriched-term table, back through Stage 3.
↓  no data link — the drug is a mechanistic hypothesis, not a pipeline output
5 Drug Developability — ADMET Turn 3
  1. Molecule input — a candidate SMILES (here, iptacopan).
  2. ADMET prediction & gates — clearance, hERG, CNS permeability, hepatotoxicity, with QC thresholds — then Stage 3 findings.
Cross-cutting, every stage.

Live progress streaming (per phase and step), usage metering, durable run-state so long jobs survive worker restarts, and an automatic QC remediation loop — all running alongside the six stages above.

The Real Run

What Actually Came Out of GSE227835

Nothing below is illustrative. These are the numbers from the pipeline on the real deposit — the biology is the data’s, not the slide’s.

StageQuantityValue
IngestionSamples pseudobulked (MG + HC)20
IngestionGenes per sample36,601
IngestionCells per sample~5,800–8,600
Differential expressionGenes tested (non-NA)14,722
Differential expressionDE genes at padj < 0.05883
EnrichmentDE genes GO-mappable99 / 141
EnrichmentLead signatureTNF-response (GO:0071356)

The top differentially expressed genes — MCRS1, ANKRD49, NRTN, ETV7, TIGIT, BATF3, GBP1 — and the enriched cellular response to tumor necrosis factor term (padj ≈ 0.02) point to an inflammatory signature in this cohort. That is a genuine, reproducible readout of a public dataset — surfaced by asking a question in English.

Honesty About the Wiring

Two Real Data Connections, One Mechanistic Hypothesis

It matters to be precise about what is automated data flow and what is scientific reasoning:

  • Stage 1 → Stage 2 (data-connected). The pseudobulk workflow’s three CSVs are the exact inputs to differential expression. No hand-editing.
  • Stage 2 → Stage 4 (data-connected). The differential-expression gene list is the exact input to enrichment.
  • Stage 4 → Stage 5 (mechanistic hypothesis). The candidate drug is chosen on established biology — complement is a validated axis in AChR+ MG, and iptacopan is a clinically established complement-directed therapeutic. The ADMET assessment is a real computation; the selection of the molecule is a scientist’s hypothesis, not an output of the enrichment step.
Why draw the line so sharply?

Because a platform that blurs “the model computed this” with “a human reasoned this” is not one you can build a regulatory submission on. Every step in the pipeline is logged, cited, and reproducible — and the one step that is a hypothesis is labeled as a hypothesis.

The Takeaway

The Engineering Is the Product

Three plain-English turns are what the audience sees. The value is in the thirty phases they don’t: the format branch that reads a dataset other tools skip, the dispersion fallback that keeps a small cohort from dead-ending, the output inference that turns a results table into a sentence. That machinery is what turns “we have a public dataset somewhere” into “here is a cited, QC-gated, reproducible answer” — without a two-week wait in the bioinformatics queue.

See it live at the summit, or start from your own accession at www.biomate.ai.