About GENT3

From GENT to GENT3

GENT3 is the third generation of a lineage that started as GENT and became GENT2 — a platform that, over its lifetime, has drawn roughly 131,000 visits a year and been cited in more than 500 published studies. That track record is the foundation GENT3 builds on, not a clean-slate rewrite: the same MAS5-based normalization pipeline, the same Affymetrix U133A/U133Plus2 platforms, and the same tissue-matched normal-vs-tumor design carry forward unchanged, so results from GENT3 stay directly comparable to anything published using GENT2.

What changed is what accumulated since GENT2's data freeze. Public repositories like GEO kept adding studies on the same platforms — enough that this update adds 138 new datasets and roughly 5,900 additional samples, plus sex/gender metadata that GENT2 didn't have. GENT2 also had a known analytical gap: cancer is molecularly heterogeneous, so a gene can be dramatically overexpressed in only a subset of tumor samples rather than uniformly across a tissue — a pattern a simple tumor-vs-normal average tends to wash out. Prior outlier-detection methods (COPA, OS, ORT, MOST, LSOSS) could flag some of this heterogeneity but didn't give researchers an interpretable way to rank genes by both how strong and how recurrent the overexpression was. Closing that specific gap — and keeping pace with the literature researchers need to interpret results against — is what GENT3's two new modules, described below, were built to do.

What's New in GENT3

What GENT3 Is

GENT3 is a research tool for exploring gene expression differences between normal and tumor tissue across large public microarray datasets. Instead of a point-and-click query builder, GENT3 is driven through a chat interface: type a gene symbol, a structured command, or just ask in plain English or Korean, and an AI agent parses the request, runs the appropriate analysis, and helps interpret the result.

Data & Scale

GENT3 integrates over 73,000 samples across 72 human tissue types, drawn from GEO microarray studies on the Affymetrix HG-U133A and HG-U133 Plus 2.0 platforms and reprocessed with the same MAS5-based pipeline as GENT2. All of it is indexed with Apache Lucene (via PyLucene) for sub-second lookup across roughly 4.7 billion individual expression records, each sample labeled Normal or Tumor.

Analysis Capabilities

The two deep-dive analyses are built directly into GENT3. They surface naturally during conversation — as a suggested next step after an analysis, or when you ask about them directly — and always ask for confirmation before running, since each call queries external databases.

Example: From a Pattern to a Drug-Target Case

Take ERBB2 (HER2) in breast cancer, walked through in the order GENT3's modules are typically chained together in an actual chat session:

1. /gex:ERBB2 — the starting question. It compares normal and tumor tissue samples across every tissue on the GPL570 platform; in the breast row, averaged across 6,727 breast tumor samples against 475 normal samples, ERBB2 is significantly higher in tumors (log2FC ≈ 0.93, roughly 1.9-fold, p = 6.2×10⁻³³, FDR = 7.2×10⁻³²) — a solid but unremarkable "up in cancer" result by itself.
2. /h:ERBB2, tissue:breast — that average turns out to hide something. ERBB2's H-index in breast cancer is 9.625, among the highest of any gene in that tissue (top 0.4% of ~22,880 genes analyzed), because about 9.6% of breast tumor samples show at least a 9.7-fold increase over the normal-tissue mean. The overexpression isn't a mild shift spread evenly across tumors — it's concentrated, and dramatic, in one subgroup.
3. /meta:ERBB2, tissue:breast, prog_type:os — does that subgroup pattern matter clinically? Pooling Cox hazard ratios from 94 independent studies is inconclusive on its own (HR ≈ 1.00, 95% CI crossing 1, p = 0.41, I² = 37% heterogeneity) — a reminder that averaging across every breast tumor, HER2-high or not, can dilute a subgroup-specific survival effect just as easily as it diluted the plain expression average in step 1.
4. That ambiguity is exactly where the deep-dive modules earn their keep. The biomarker deep-dive weighs that same meta-analysis result against current literature and correctly reports it as inconclusive on its own — recommending the stratified validation a real biomarker claim would need, rather than overselling a weak number. The druggability deep-dive asks an independent question instead: is ERBB2, as a protein, a viable drug target regardless of this particular survival result? Combining AlphaFold structural confidence (mean pLDDT 74, high confidence), its protein kinase domain, and close to 11,000 PubMed articles on HER2-targeted therapy, it returns a druggability score of 0.52 (high confidence) — consistent with ERBB2's independently established status as a validated cancer drug target.

The point isn't that every step confirms the last one — it's that when one lens doesn't fully answer the question, GENT3 lets you ask it a different way instead of stopping at a single number.

Technology

GENT3 feels light to use — type a gene name, get an answer in seconds — but that simplicity sits on top of a genuine convergence of technologies built specifically for scale. PyLucene bridges Java's Apache Lucene search engine into the Python stack, and the roughly 4.7 billion expression records are split across multiple index shards where a single Lucene index would exceed its own 2³¹−1 document limit. That's what makes sub-second lookups possible across a dataset this size, and it's why the platform stays fast and light to use even as its analytical scope — six distinct analysis modules plus a locally hosted LLM — keeps growing.

Bulk Data Download

The raw gzip-TSV source data behind GENT3's analyses is available for direct bulk download, for anyone who wants to run their own analysis rather than go through the chat interface.

Both archives are plain tar (not further compressed — the contents are already .gz). Extract with tar xf <file>.tar. Column layout is documented in the chat interface's help text and the project README.