Omnii-1 for enhanced multi-cancer diagnostics

BACK BACK

We fine-tuned Omnii-1 to predict a proprietary cell-free DNA biomarker it had never seen before and used the resulting model to enhance the sensitivity of a cell-free DNA test for early-stage cancer.

Detecting advanced cancer from a blood sample is often straightforward. Late-stage tumors usually shed enough circulating tumor DNA (ctDNA) into the bloodstream to stand out against the background of normal cell-free DNA (cfDNA). But detecting cancer at its earliest stages, when diagnosis has the greatest potential to improve patient outcomes, remains remarkably difficult despite decades of research.

Early-stage tumors shed very little ctDNA, making their signal difficult to distinguish from the much larger background of normal cfDNA fragments. Recent approvals of blood-based colorectal cancer screening tests¹ and progress toward multi-cancer screening² show how far the field has come, but detecting early-stage cancers across multiple tissues remains difficult. A useful test must recover these faint signals while maintaining very high specificity: because cancer is rare among asymptomatic individuals, even a small false-positive rate can lead to unnecessary follow-ups, imaging, anxiety, and costs.

To investigate whether a genome language model could help recover these faint signals, we partnered with ClearNote Health. Its approach combines cell-free whole-genome sequencing (WGS) with measurements of 5-hydroxymethylcytosine (5hmC), an epigenetic DNA modification associated with gene activity.³ Guler et al. demonstrated the potential of cell-free DNA 5hmC signatures for detecting early-stage pancreatic ductal adenocarcinoma.⁴

Omnii-1, the world's most advanced multimodal genome language model, learns relationships between biological modalities during pretraining on omics data across the tree of life. We wanted to test whether that shared biological context could help it interpret a new type of measurement and extract information useful for cancer detection. Its training corpus included DNA methylation measurements from cell lines and tissue samples, but no 5hmC data. ClearNote’s measurements therefore offered a concrete test of whether knowledge learned from other modalities could transfer to an unseen biomarker.

We approached the problem in two steps. First, we post-trained Omnii-1 to predict patient-specific cell-free 5hmC profiles from germline genetic variants and cell-free WGS. Then we tested whether the resulting features improved cancer detection and tissue-of-origin prediction at high specificity. The largest detection gains occurred in patients with tumor fractions below ichorCNA’s detection threshold,⁵ the setting in which the signal is hardest to distinguish from the background.

Post-training Omnii-1 to reconstruct cell-free 5hmC profiles

We post-trained Omnii-1 using a self-supervised objective to predict cell-free 5hmC profiles from individual-specific genome sequences and matched cell-free WGS profiles. Germline genotypes at autosomal biallelic SNP sites were inferred and phased from low-pass cell-free WGS using GLIMPSE2.⁶ Quality-filtered genotypes were incorporated into the hg38 reference to generate individual-specific genome inputs.

Despite never seeing 5hmC data in pretraining, Omnii-1 readily adapted to predicting base-resolution profiles of this new modality. We evaluated the model in 32 held-out patients and found strong agreement between predicted and measured 5hmC coverage. This agreement extended across a wide range of 5hmC abundance, and predicted profiles at individual genes closely resembled the measured signal at single-base resolution (Figure 2).

We next asked whether Omnii-1’s predictions preserved associations between germline genetic variation and 5hmC coverage. We compared alternate-allele dosage effects on measured 5hmC coverage with those from Omnii-1 reconstructions and matched WGS in the 32 held-out patients. Across the 250 held-out loci with the strongest discovery associations, Omnii-1’s dosage effects closely tracked those measured in 5hmC coverage, while effects estimated from matched WGS showed little correlation, suggesting that Omnii-1 recovered genetic associations with 5hmC that were not apparent in WGS coverage alone (Figure 3).

Omnii-1 improves early-stage cancer detection and tissue-of-origin prediction

We next investigated whether Omnii-1 could improve early-stage cancer detection and tissue-of-origin classification. Using each patient’s germline genotypes and cell-free WGS profile, Omnii-1 generated predicted 5hmC coverage and features describing the genomic distributions of fragment start sites and centers. We then used Omnii-1-generated features in classifiers for cancer prediction and compared their performance with that of regularized logistic regression classifiers built from measured 5hmC and WGS features.

Classifiers using Omnii-1 features improved early-stage cancer detection sensitivity at 98% specificity compared with classifiers using 5hmC features alone or 5hmC and WGS features combined (Figure 4). The largest gains occurred in patients with tumor fractions below the detection threshold of ichorCNA⁵, suggesting that Omnii-1 features are capturing ctDNA signals missed by other methods (Figure 5).

ClearNote’s multi-cancer early detection dataset includes 851 cancer patients across eight tissues and cancer stages I–IV. We evaluated tissue-of-origin classification using all cancer patients, regardless of whether their samples were classified as cancer. We evaluated each tissue separately as a binary classification problem: did a cancer originate in that tissue or in any of the other seven? Adding Omnii-1 features improved tissue-of-origin discrimination, measured by averaging the eight ROC curves with each tissue contributing equally (Figure 4).

Finally, we evaluated cancer detection sensitivity across patient age and sex. Classifiers with Omnii-1 features improved sensitivity across all groups, with smaller differences between groups than those observed with classifiers using 5hmC alone or 5hmC combined with WGS features. This held for both early-stage and all-stage cancer detection (Figure 6).

The potential impact of Omnii-1

Even small sensitivity gains could matter at screening scale. Applying the historical stage I–II cancer incidence rate from Hubbell et al.⁷ to the approximately 110 million US adults aged 50-79⁸ gives roughly 733,000 annual early-stage diagnoses. Using this incidence-based proxy, a hypothetical one-percentage-point sensitivity gain would correspond to approximately 7,300 additional detections at full participation, assuming the sensitivity gain generalizes across cancer types.

Hubbell’s screening scenarios imply approximately 0.23 cancer deaths averted within five years per cancer intercepted. In this illustrative example, applying that modeled ratio to the full-participation example suggests roughly 1,700 deaths could be averted within five years, per percentage point increase in sensitivity.

Omnii-1 and the future of diagnostics

Omnii-1, adapted to a biomarker absent from its pretraining, provides features that can be used to improve early-stage cancer detection and tissue-of-origin classification. The gains in patients with low tumor fractions suggest that a broadly trained biological model can help extract useful information from faint molecular signals. Prospective studies are needed to confirm these detection gains in screening populations and establish whether they translate into earlier treatment and fewer cancer deaths.⁹

Partner with us

Partners can bring patient cohorts, molecular assay data, and clinical validation capabilities. We can work together to train and validate Omnii-1-powered diagnostic models, with a particular focus on improving early-stage detection and extracting more information from existing assays.

We're hiring → join us

Citation

bibtex

@misc{omnii_diagnostics_2026,
  title        = {Omnii-1 for enhanced multi-cancer diagnostics},
  author       = {Siranosian, Benjamin and Poli, Michael},
  year         = {2026},
  month        = {Oct},
  url          = {https://www.radicalnumerics.ai/blog/omnii-diagnostics},
  organization = {Radical Numerics}
}

References

U.S. Food and Drug Administration. Premarket approval records for Guardant Health’s Shield (P230009; July 26, 2024) and Freenome’s SimpleScreen CRC (P250029; July 24, 2026).
GRAIL. FDA Advisory Committee Votes in Favor of Approval of GRAIL’s Galleri Multi-Cancer Early Detection Test. September 23, 2026.
He, B. et al. Tissue-specific 5-hydroxymethylcytosine landscape of the human genome. Nature Communications 12, 4249 (2021).
Guler, G. D. et al. Detection of early stage pancreatic cancer using 5-hydroxymethylcytosine signatures in circulating cell free DNA. Nature Communications 11, 5270 (2020).
Adalsteinsson, V. A. et al. Scalable whole-exome sequencing of cell-free DNA reveals high concordance with metastatic tumors. Nature Communications 8, 1324 (2017).
Rubinacci, S. et al. Imputation of low-coverage sequencing data from 150,119 UK Biobank genomes. Nature Genetics 55, 1088–1090 (2023).
Hubbell, E. et al. Modeled Reductions in Late-stage Cancer with a Multi-Cancer Early Detection Test. Cancer Epidemiology, Biomarkers & Prevention 30, 460–468 (2021).
U.S. Census Bureau. Age and Sex, 2024 American Community Survey 1-Year Estimates, Table S0101.
Mandrik, O. et al. Modeling the Impact of Multicancer Early Detection Tests: A Review of Natural History of Disease Models. Medical Decision Making 45, 1013–1024 (2025).