MetaCH

Circulating tumor DNA is one of the most useful biomarkers in oncology, but reading it correctly depends on knowing where each mutation came from. A large share of the somatic variants found in blood plasma originate not in a tumor but in clonal hematopoiesis (CH) — the age-related expansion of mutated blood cell clones. CH accounts for more than 75% of cfDNA variants in people without cancer and sometimes more than half of them in people with cancer, so mistaking a CH variant for a tumor variant can distort treatment selection, disease monitoring, and early detection alike. The standard remedy is to sequence a matched white blood cell fraction alongside the plasma, which roughly doubles the cost, is impractical at scale, and is simply unavailable for archived samples.

MetaCH — a metaclassifier for clonal hematopoiesis detection — is an open-source framework from AstraZeneca Oncology R&D that assigns each variant in a plasma-only sample a probability of being CH-derived rather than tumor-derived. Its distinguishing feature is that it does not reason over hand-engineered variant annotations alone. Every classification stage consumes gene and variant embeddings produced by METk, a companion representation model pretrained on pan-cancer sequencing from TCGA and cBioPortal, so the mutational context a variant sits in is carried into the decision as learned features. The work was posted as a preprint in December 2024 and published in npj Precision Oncology in 2025.

The name is worth pinning down: the repository and journal article use MetaCH, the preprint abstract used MetaCHIP, and an unrelated metagenomics tool for horizontal gene transfer detection is also called MetaCHIP.

#Key Features

• Plasma-only operation: Classification requires only the variant call, its allele frequency, and the cancer type — no matched white blood cell sequencing, which makes retrospective and archival cohorts tractable.
• Pretrained mutational embeddings: METk supplies 128-dimensional variant embeddings learned with a StarSpace-style contrastive objective and 8-dimensional gene embeddings learned by a continuous-bag-of-mutated-genes model over patient mutation profiles.
• Three-stage stacked design: A cfDNA-trained classifier and two sequence-based classifiers, one for CH-oncogenic and one for CH-non-oncogenic variants, each emit a score; a logistic regression meta-classifier fuses the three into a single CH likelihood.
• Fixed checkpoints for new cohorts: The repository ships the trained classifiers, so the framework is applied to an incoming cfDNA cohort as-is rather than refit per dataset.

#Technical Details

METk was trained on 85,648 pan-cancer patients spanning more than 59 tumor types from TCGA and cBioPortal, encoding each variant with 2- to 5-mers drawn from a 20-base-pair window around its position, and pairing the resulting embeddings with functional prediction scores derived from SnpEff and SnpSift. The cfDNA-based classifier was trained on a 124-patient non-small cell lung, breast, and prostate cohort sequenced with a 508-gene panel at over 60,000× depth, using 436 tumor-derived and 914 blood-derived variants. The two sequence-based classifiers were trained on much larger public sets: 57,210 tumor variants from MSK-IMPACT and 3,778 CH-oncogenic plus 2,967 CH-non-oncogenic blood-derived variants. Each classifier was selected and tuned with the FLAML AutoML library under stratified five-fold cross-validation.

Across four independent external validation cohorts with matched white blood cell ground truth, MetaCH reached an area under the precision-recall curve of 0.89 (auROC 0.84) on the Leal gastric cancer dataset, 0.88 (auROC 0.85) on the Chabon lung cohort, and 0.71 (auROC 0.84) on the Chin breast cohort, consistently ahead of prior machine-learning approaches. Performance was weakest on the Zhang cohort (auPR 0.22), where plasma was sequenced at 1000× but matched blood at only 400×, producing missed CH calls in the ground-truth labels themselves. Removing the canonical CH genes DNMT3A, TET2, and ASXL1 from external validation dropped the overall auROC from 0.79 to 0.73, showing that the model retains predictive power beyond the few genes that dominate CH.

#Applications

The immediate beneficiaries are translational oncology groups running plasma-based assays where matched blood sequencing is missing or unaffordable: retrospective analyses of banked trial samples, minimal residual disease monitoring, and early-detection studies where a CH false positive is expensive. Because the framework returns a per-variant probability rather than a hard call, it slots into an existing variant-calling pipeline as a filtering or annotation step.

#Impact

MetaCH demonstrates that representation learning over large public tumor genomics transfers usefully into a narrow clinical genomics decision — the embeddings, not the tabular features, carried most of the signal in permutation importance tests on external data. Its scope is deliberately narrow: a single binary judgment per variant, accuracy that varies with the sequencing depth and gene panel of the incoming cohort, and a demonstrated bias toward the gene panels present in training. The authors are explicit that connecting auPR to clinically meaningful false-positive rates remains unfinished work before the model is used to guide care.