A safety net, not a substitute. Every result must be confirmed by standard laboratory antimicrobial susceptibility testing and reviewed by a qualified healthcare or laboratory professional.
Two directions we want to take GATCHA next: grounding recommendations in the patient's own record, and making results easier to read and question.
Right now GATCHA scores an antibiotic panel from the genome alone. The next step is to let a clinician attach the patient's record so the report reflects that specific patient, not just the isolate.
Cross-reference the record against current or recent antibiotic courses so drugs already tried (and failed) are ruled out directly, instead of only inferred from resistance genes.
Flag or hide drug classes the patient is allergic to, or is otherwise contraindicated for, before they ever reach the ranked recommendation list.
Same principle as the rest of GATCHA: prefer an explicit, auditable exclusion over a silent one, and never guess past a known contraindication.
We plan to integrate an LLM as a reader on top of the analysis results, not as part of the prediction pipeline itself — the calibrated per-drug models stay the source of truth.
Turn the evidence receipt (genes, mutations, calibrated probabilities, no-call reasons) into a short summary a clinician can scan in seconds.
Let the user ask follow-up questions about a result — e.g. why a drug was no-called, or what evidence backs a given prediction — grounded strictly in that report's own data.
As with the rest of GATCHA, sequence data itself stays out of the LLM's context — only the already-computed report would be summarized or discussed.