Documentation

How GATCHA turns a genome into an antibiotic-response report, what each stage of the pipeline contributes, and how to move around by keyboard.

Model inputs

FASTA: an assembled, quality-checked genome (.fasta / .fa / .fna, up to 50 MB).

Species: the declared organism. Anything outside the validated set (E. coli, K. pneumoniae, S. aureus) is scored as no-call.

Antibiotic panel: the drugs to score. Unsupported drugs are hidden, never guessed.

Uploaded sequence data is processed as data only. It is never forwarded to any language model or third party.

Model outputs

Prediction: one of likely_to_work, likely_to_fail, no_call.

Calibrated confidence: a resistance probability after Platt or isotonic calibration on held-out genetic groups.

Evidence: the AMR features detected (gene, mutation, plasmid) with their database source, and whether each is a known determinant or only a statistical association.

Familiarity + OOD: how close this genome sits to the training distribution.

Decision receipt (JSON): a full audit trail of rules fired, model versions, thresholds, and no-call reasons.

The serving pipeline
01
Target gate

Screens species and QC before any model runs. Fails fast on the wrong organism, a poor assembly, or contamination.

Output · pass / fail with a reason

02
AMRFinderPlus features

The gold-standard annotator turns the assembly into known resistance genes, point mutations, and plasmids, collapsed into a hierarchical feature record.

Output · feature record (genes, mutations, drug classes)

03
Calibrated per-drug model

One interpretable model per antibiotic (logistic regression or XGBoost, chosen by grouped cross-validation), then Platt or isotonic calibration so the confidence means what it says.

Output · raw + calibrated resistance probability

04
No-call & OOD safety

A drug-target gate plus three triggers (low confidence, conflicting evidence, out of distribution) downgrade weak or risky cases to an honest no-call instead of a guess.

Output · final call, no_call_reasons[], lab_confirmation_required

Deep-learning encoders (ESM-2, DNABERT-2, a multi-task net) were benchmarked offline on K. pneumoniae for the record, but the live path stays on the fast, auditable calibrated baseline.

Keyboard shortcuts
  • Show / hide this shortcuts panelCtrl+?
  • Close dialogs / cancelEsc
  • Back to Dashboard (from any subpage)h
  • Go to Dashboardgthenh
  • Go to Documentationgthend
  • Go to About Usgthena

g is a leader key: tap g then a letter within about a second. Shortcuts are ignored while typing in inputs.