AI interpretation

Interpretation is an optional layer after pipeline outputs exist. Every summary cites computed values from completed runs — the LLM does not invent numbers.

Prerequisites

One or more completed oncology pipeline runs (communication, immune profile, mutation landscape, or survival)
Ollama credentials configured in backend-analysis (see repo root ai.md)

Launching interpretation

Open the study Interpret page or dispatch via the analysis API:

curl -X POST http://localhost:8001/oncology/jobs/interpret \
  -H "Content-Type: application/json" \
  -d '{
    "study_id": "your-study-id",
    "parameters": {
      "source_run_ids": ["onco-run-abc12345", "onco-run-def67890"],
      "interpretation_type": "summary"
    }
  }'

The source_run_ids list must reference completed runs whose artifacts exist on disk.

Interpretation modes

Oncology interpretation templates cover the major pipeline outputs:

Context	When	Output
Cell-cell communication	Communication run complete	Signaling axis summary with cited interaction scores and pathway rankings
Immune phenotype	Immune profile run complete	Deconvolution, TIDE phenotype, exhaustion, and repertoire narrative
Mutation landscape	Mutation run complete	TMB, driver alterations, signature, and oncoprint interpretation
Survival	Survival run complete	Kaplan-Meier and Cox findings with cited hazard ratios and p-values
Multi-modal summary	Multiple source runs	Integrated narrative connecting TME, immune, mutation, and outcome findings

Guardrails

Every interpretation artifact includes:

Research / non-clinical disclaimer
Metric citation validation — numbers must match run artifacts
Cautious language — hypotheses about biology and therapy response, not clinical recommendations

The interpret pipeline rejects requests when source run artifacts are missing or incomplete.

What interpretation does not do

Diagnose cancer or recommend treatment
Replace expert review of pipeline inputs and QC
Compute new metrics — it explains metrics already in artifacts
Fabricate survival statistics, TMB values, or interaction p-values without supporting data

Menu