Grounding the genomic report in non-small-cell lung cancer.

Roughly half of lung adenocarcinomas carry a targetable driver. EGFR exon-19 deletions and L858R, KRAS G12C, ALK and ROS1 fusions, BRAF V600E, MET exon-14 skipping, and RET and NTRK fusions. Each maps to an approved or trial therapy, and each is reported against a specific transcript and evidence tier.

Real-world data shows how concentrated this is. In a recent comprehensive-genomic- profiling cohort of advanced NSCLC, a clinically actionable alteration was found in about 45% of patients — KRAS G12C (~18%) and EGFR (~14%) leading — each classified against an ESCAT actionability tier.^[1]

Precision matters more than breadth. A report that says EGFR L858R has to name the right transcript (NM_005228.5), the right coding change (c.2573T>G), the right protein consequence (p.Leu858Arg), and an actionability tier a tumour board can trust.

In that same real-world cohort, only about a third of patients with an actionable alteration went on to receive matched targeted therapy.^[1] And across studies, biomarker-matched targeted therapy is associated with roughly double the real-world survival of unmatched treatment.^[2] A biomarker that is detected but mis-coded, buried or never reaches the board.

A general-purpose language model will answer any of these questions fluently. That is exactly the problem. Ask it twice and you may get two different HGVS strings, a plausible-but-wrong COSMIC identifier, or a confident therapy claim with no traceable citation. In a molecular report, a fluent guess is worse than no answer — it looks authoritative and it is not reproducible.

EGFR V843I (c.2527G>A, p.Val843Ile), a rare germline EGFR variant. Its HGVS is confirmed by VariantValidator; the normalization is unambiguous. The actionability question is where it turns dangerous because the published evidence on V843I is thin and conflicting. Several reports describe EGFR-TKI resistance, a few describe response to a third-generation TKI, with no accepted, consensus actionability tier.^[3]

An ungrounded model tends to fill that void with something fluent and wrong. The grounded AI does the opposite: it returns the confirmed HGVS and validated codes, then flags the variant for human review because it cannot ground a single actionability tier. The report is held for a human rather than guessed.

A grounded report is only useful if downstream systems can read it. Clarus emits HL7 FHIR conforming to mCODE and the HL7 Genomics Reporting IG, and therapies and trials linked to CIViC and ClinicalTrials.gov. An independent assessment by the HL7 CodeX / GenomeX group found that about 87% of the semantic content of real-world somatic reports is already representable in this standard^[4], so the report drops into a LIS and the board's view as structured, validated data.

We're starting with non-small-cell lung cancer (NSCLC). It has the most actionable driver mutations of any solid tumour — and the highest stakes for getting the match right. More tumour types are next.