Synthetic data only. Educational use only. Synthetic data, demonstration only. Not a medical device. Not for unsupervised clinical decision-making.

EGFR L858R in metastatic NSCLC

The starting input

A synthetic sample report in the Foundation Medicine F1CDx layout. NSCLC, lung, FFPE biopsy, panel size 324 genes. The sample is fully synthetic, watermarked Synthetic data — demonstration only, and built to match the F1CDx report structure. It is not a Foundation Medicine document and contains no real patient data. It lives in our repo at data/sample_reports/synthetic/f1cdx_format_nsclc_egfr-l858r.pdf. Try it live on the vendor sample showcase.

What the parser extracts

The deterministic Foundation Medicine parser pulls:

• One variant: EGFR exon 21 L858R, VAF 38%, transcript NM_005228.5, HGVS p.L858R / c.2573T>G
• Three biomarkers: TMB 4 mut/Mb, MSI-stable, no germline pathogenic findings
• Two FDA CDx flags: osimertinib, dacomitinib (NSCLC EGFR L858R indications)
• Trial-section pointer: TRIALS TO CONSIDER block at page 6

What the evidence join adds

With the variant locked, /v1/lookup joins the reference graph and returns:

• CIViC: variant ID 33, multiple curated evidence items, top assertion EGFR L858R is associated with sensitivity to EGFR inhibitors in NSCLC
• ClinVar: pathogenic, multiple submitters with criteria provided
• openFDA: osimertinib indication for first-line NSCLC with EGFR exon 21 L858R substitutions
• ClinicalTrials.gov: 47 actively recruiting NSCLC trials matching EGFR L858R
• AMP/ASCO/CAP tier: I-A

Live at /lookup.

What the CDS surface produces

Engine 2 routes the parsed report into deterministic templates, not LLMs. The recommendations card surfaces the FDA-approved first-line option, the second- and third-line options, the relevant CPIC pharmacogenomics flags (none here, since EGFR is the somatic driver, not a CYP), and the matching trial NCT IDs. The educational-use banner is sticky on the page.

Live at /recommendations.

Where the LLM is and isn't

On a clean F1CDx like this one, parsing is fully deterministic: the Foundation Medicine vendor parser reads the layout, with AWS Textract OCR as the fallback for image-only PDFs. A long-tail variant LLM fallback is on the roadmap and is not part of the production parse path today; if it ever runs, it is clearly flagged in the response. The final clinical surface is rendered from deterministic templates, never an LLM.

What this case does not show

Vendor edge cases. A Caris MI Profile reports the same biology in a different layout. A Tempus xT formats CDx flags differently. The point of UNMIRI is that the canonical schema is identical across all three. See the side-by-side compare view for that.

Related references