Technical

Abstracting Molecular Markers and Genomic Reports: Why AI Needs a Human in the Loop

Molecular reports now steer cancer treatment. AI makes abstraction fast; a trained reviewer makes it right. Here is how the two work together.

Mugdha Raut

Team Lead - Quality Assurance

October 19, 2026

9 min read

A single molecular report can decide which therapy a cancer patient receives. Biomarkers such as EGFR, ALK, KRAS, BRAF, HER2, BRCA1/2, MSI status and tumour mutational burden now shape treatment choices across many cancer types.

Yet these results rarely arrive as data. They come as long PDFs from next-generation sequencing (NGS) panels, tissue tests such as FoundationOne⁴ and liquid biopsies such as Guardant360.^5,6 Each lab has its own layout, vocabulary and appendices, and a patient may collect several reports across their journey: one at diagnosis, more at progression.

Registries, care programmes, trial screening teams and EHRs all need those findings as clean, structured fields. The question is no longer whether to use AI for this. It is how to pair AI with human expertise so that abstraction is both fast and right.

In short

Manual abstraction of genomic reports is slow and still error-prone.
AI reads fast, but misses context that changes clinical meaning: significance tiers, specimen type, negatives, germline findings.
The answer is a loop: AI extracts, rules flag, trained reviewers confirm what matters, and every correction makes the next run better.

324genes

analysed by a single tissue panel, FoundationOne CDx⁴

16,812patients

in a 2025 study separating blood variants from the tumour and from clonal haematopoiesis⁸

4tiers

in the AMP/ASCO/CAP system for classifying cancer variants, from strong significance to benign¹

01 · The problem

Why manual abstraction alone falls short

Manual abstraction is slow, and slow work is not automatically accurate. Experienced abstractors are skilled, but the task works against them in four ways.

TurnaroundA comprehensive panel can cover hundreds of genes.⁴ Reading every report line by line takes time, and backlogs grow faster than teams can hire.

Transcription errorsOne wrong character in a protein change, or a misplaced decimal in a variant allele frequency, quietly changes the meaning of a result.

InconsistencyThe same finding is recorded as "EGFR exon 19 deletion" in one row and "EGFR p.E746_A750del" in the next. Each may be correct, but mixed levels of detail make the data hard to count, compare or analyse.

FatigueDense tables, appendices and repeated gene lists invite skipped lines, especially late in a long queue.

The result is data that arrives late and still needs checking. AI alone fixes the speed but not the judgement, which is why neither extreme works on its own:

‍

Manual only

Abstractors read and key in every report

SpeedSlow

ConsistencyVaries

Clinical judgementStrong

AI only

Extraction released without review

SpeedFast

ConsistencyHigh

Clinical judgementLimited

Recommended

AI + human review

AI extracts, experts confirm what matters

SpeedFast

ConsistencyHigh

Clinical judgementStrong

02 · Where AI needs help

Why AI alone is not enough

AI reads a long report in seconds, but it can be confidently wrong, and in genomics a small error can change a treatment decision. These are the places where meaning hides in context:

‍

NomenclatureOne alteration can appear as a protein change, a cDNA change, an exon-level description or a legacy name. All must resolve to one standard form.³

Clinical significanceReports separate actionable variants from variants of uncertain significance (VUS). Extracting a VUS as actionable is a serious error.^1,2

Specimen and allele frequencyTissue and liquid biopsy results are read differently. A variant found in blood may come from clonal haematopoiesis, age-related changes in blood cells, rather than from the tumour.^7,8

Complex alterationsA fusion is only meaningful with both partner genes, and an amplification with its copy number. Extracting the gene name alone loses the finding.

Genomic signaturesTMB, MSI and HRD are reported differently by each assay, and TMB values and cut-offs are not directly comparable across panels.^9,10

What was not found"Not detected", "not tested" and "insufficient sample" mean very different things and must be kept apart.

Germline versus somaticSome findings may be inherited rather than tumour-only. They carry implications for the patient's family and need separate handling.

Therapy and trial tablesReports list drugs and trials linked to each finding. These are options, not the patient's treatment.

Amended reportsWhen a lab issues a corrected report, it must replace the original in the record, not sit beside it as a second result.

Two examples of what this looks like

A negative that was really "not evaluable"

Tissue NGS

Source report

ALK fusion: Not detected
Comment: tumour content below
assay threshold

AI extracted

ALK: Negative

Reviewer recorded

ALK: Not evaluable
(low tumour content)

A true negative and an uninformative test look alike as a single word. Recorded as negative, the result could close the door on retesting a patient who may still carry an actionable fusion.

A blood variant that may not come from the tumour

Liquid biopsy

Source report

DNMT3A p.R882H  VAF 1.2%
Note: possible clonal
haematopoiesis

AI extracted

DNMT3A p.R882H
Origin: tumour

Reviewer recorded

DNMT3A p.R882H
Origin: possible CH,
not attributed to tumour

DNMT3A is one of the genes most commonly affected by clonal haematopoiesis. The report's own note is easy to lose when values are pulled from a table, and keeping it changes how the finding is used.

Examples are illustrative, based on common reporting patterns rather than specific patient records.

Does the reading

Pulls every gene, value and signature from long, varied reports in seconds, at volume.

RULES

Flag the doubts

Automatic checks catch impossible values and route low-confidence fields for review.

HUMAN

Makes the judgement

A trained reviewer confirms what carries clinical weight before the data is released.

Try it

One report, two outcomes

Here is what an AI pulled from an illustrative lung cancer NGS report. Everything looks plausible. Switch to the reviewed version to see what a trained reviewer changed before release.

03 · The workflow

How AI and human review work together

The strongest model gives AI the volume and gives people the decisions that carry clinical weight. In practice, that is a six-step loop.

Every record passes automatic checks; flagged and critical fields go to an expert, and each correction improves the next run.

1. Configure by cancer type

Decide which markers matter before extraction starts

The AI then extracts what is relevant to the programme, not every gene on the panel. For example:

‍

LungEGFRALKROS1KRASBRAFMETRETNTRKPD-L1

BreastHER2BRCA1/2PIK3CAESR1

2. Extract and normalise with AI

Volume handled in seconds

Genes, alterations, allele frequencies, significance tiers, signatures, specimen type, assay and report dates are pulled out and mapped to standard nomenclature.

3. Run automatic checks

Catch the impossible, flag the uncertain

Rules catch impossible or inconsistent values: a frequency outside 0 to 100%, a gene not on the stated panel, a missing unit. Low-confidence fields are flagged.

4. Route to expert review

Where the human in the loop sits

A reviewer trained in molecular reporting checks every flagged field and every critical field, plus a sample of the rest, against the source report.

5. Release structured output

With a trail of who checked what

Verified data moves to the registry, EHR or analytics platform, ready to use.

6. Feed corrections back

Every fix makes the next run better

Reviewer corrections improve rules, prompts and templates, so the same mistake becomes less likely next time.

The effect compounds. As accuracy rises, fewer fields are flagged, review gets quicker, and experts spend their time on the genuinely difficult reports.

04 · Reviewer checklist

What the human reviewer checks

A good reviewer does not re-abstract the whole report. They verify the fields where an error would matter most:

‍

Patient and report identityThe right patient, specimen, collection date and report version.

Each actionable alterationGene, exact change and nomenclature match the source text.

Significance tierActionable findings and VUS sit in the right category.

Allele frequency and copy numberValues, decimals and units are copied correctly.

SignaturesTMB, MSI and HRD values carry their units and the assay's own threshold.

Negatives and failures"Not detected", "not tested" and "insufficient sample" are recorded distinctly.

Germline flagsAny germline or incidental finding is marked for the right handling.

No leakage from therapy tablesListed drugs and trials have not been recorded as treatment given.

‍

Each correction is logged with its reason. That log is what turns review into improvement.

05 · Measuring success

Measuring quality and promptness together

Speed without accuracy creates rework, and accuracy without speed leaves patients and programmes waiting. Track both, side by side:

‍

Measure	Type	What it tells you
Field-level accuracy	QUALITY	How often each extracted field matches the source report
Critical-field error rate	QUALITY	Errors in the fields that can change care, tracked separately
Turnaround time	SPEED	Time from report received to verified data released
Clean-record share	SPEED	Proportion of records released with no corrections needed; the higher it is, the faster data moves
Correction rate by field	QUALITY	Which fields the AI gets wrong most, and where to improve next
Reviewer agreement	QUALITY	Whether two reviewers reach the same answer on the same report

‍

When accuracy rises and turnaround falls at the same time, the combination is working.

06 · Readiness check

How ready is your molecular abstraction workflow?

Self-check

Tick what is already true for your team. Nothing is stored or sent.

The markers to extract are defined for each cancer type you cover. Gene and variant names are normalised to one standard, such as HGVS. Significance tiers, including VUS, are captured as their own field. "Not detected", "not tested" and "failed" results are recorded distinctly. Flagged and critical fields are checked by a trained reviewer against the source. Reviewer corrections are logged and fed back to improve extraction.

Fast and right, not fast or right

Molecular data is only useful when it is both timely and trustworthy. AI removes the bottleneck of reading; human reviewers protect the meaning. Neither one alone delivers both.

At Jonda Health, this is how we already approach laboratory data: JondaX pairs AI extraction with human review so that results arrive structured, harmonised and checked.¹¹ Molecular and genomic reports raise the stakes, which makes the human in the loop more important, not less.

Abstracting molecular reports by hand, or trusting AI without review?

Talk to us about building a workflow that gives you both speed and accuracy.

Talk to us

Frequently asked questions

Can AI extract molecular markers from NGS reports on its own?

AI can extract genes, alterations and values from long reports in seconds. What it cannot reliably do on its own is judge context that changes clinical meaning: significance tiers, specimen type, results that were not evaluable, possible clonal haematopoiesis and germline findings. A trained reviewer checks those fields before the data is used.

What is a variant of uncertain significance (VUS)?

A VUS is a variant whose clinical impact is not yet known. In the AMP/ASCO/CAP system it sits in Tier III, separate from variants with strong or potential clinical significance. Recording a VUS as actionable can mislead treatment decisions, trial matching and analysis.

Why do liquid biopsy results need extra care?

A variant found in blood may come from clonal haematopoiesis, age-related changes in blood cells, rather than from the tumour. Reports often flag this, and the flag needs to travel with the finding into the structured data.

Does human review slow abstraction down?

Not when it is designed well. Reviewers focus on flagged and critical fields rather than re-reading every report, and each correction improves the AI, so fewer fields are flagged over time and review gets quicker.

References

Li MM, Datto M, Duncavage EJ, et al. Standards and Guidelines for the Interpretation and Reporting of Sequence Variants in Cancer: A Joint Consensus Recommendation of AMP, ASCO and CAP. Journal of Molecular Diagnostics, 2017. DOI.
Association for Molecular Pathology. AMP Assesses Clinical Implementation of Past Standards and Guidelines for Interpretation and Reporting of Sequence Variants in Cancer. 2022. AMP.
den Dunnen JT, Dalgleish R, Maglott DR, et al. HGVS Recommendations for the Description of Sequence Variants: 2016 Update. Human Mutation, 2016. DOI.
Foundation Medicine. FoundationOne CDx. Foundation Medicine.
Oncology Nursing Society. FDA Approves First Liquid Biopsy NGS Companion Diagnostic Test for NSCLC. 2020. ONS.
Guardant Health. FDA Approves New Guardant360 Liquid CDx. May 2026. Business Wire.
CAP Today. Up Close on Clonal Hematopoiesis in cfDNA Testing. CAP Today.
Caris Life Sciences. Press release on "Characterization of plasma cell-free DNA variants as of tumor- or clonal hematopoiesis-origin in 16,812 advanced cancer patients", Clinical Cancer Research. February 2025. BioSpace.
Merino DM, McShane LM, Fabrizio D, et al. Establishing Guidelines to Harmonize Tumor Mutational Burden (TMB): Phase I of the Friends of Cancer Research TMB Harmonization Project. Journal for ImmunoTherapy of Cancer, 2020. PubMed Central.
Friends of Cancer Research. Friends of Cancer Research Releases Results of Phase 2 of TMB Harmonization Project. Friends of Cancer Research.
Jonda Health. JondaX Pathology module. jonda.health.