Omnii Diagnostics

We fine-tuned Omnii 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. A screening test must detect these faint signals while maintaining very high specificity. Cancer is rare among asymptomatic individuals, so even a small false positive rate can lead to unnecessary followups, imaging, mental anguish, and cost to patients.

Recent FDA approvals of blood-based colorectal cancer screening tests from Guardant Health and Freenome demonstrate the progress of cfDNA diagnostics.1 GRAIL's Galleri test, which uses cfDNA methylation to screen for multiple cancer types, also recently received a favorable FDA advisory committee vote, although full approval remains pending.2 Despite these advances, reliably detecting early-stage cancers across multiple tissues remains a largely unsolved problem.

We partnered with ClearNote Health to investigate whether Omnii could improve this tradeoff. By adapting Omnii to predict a previously unseen cell-free DNA biomarker, 5-hydroxymethylcytosine (5hmC), we improved both early-stage cancer detection and tissue of origin classification at high specificity.

ClearNote Health’s approach is particularly compelling because it combines cell-free whole-genome sequencing (WGS) with cell-free 5hmC profiling. 5hmC is an epigenetic DNA modification that's more correlated with actively expressed genes than standard DNA methylation — it tells you about what a cell is doing, rather than just its identity3. DNA fragments with 5hmC marks can also be selectively captured from a sample, avoiding bisulfite conversion and reducing the sequencing depth that's required.

ClearNote’s Avantect test for pancreatic cancer is commercially available in the United States for individuals at elevated risk of pancreatic cancer. They also have a test for multi-cancer early detection in otherwise healthy individuals under development.

The combination of a rich 5hmC biomarker, thousands of patient samples, and a clear clinical objective gave us an opportunity to test whether a genome language model could generalize to a previously unseen modality and extract cancer-associated signals missed by traditional approaches.

Omnii is the world's most advanced multimodal genome model trained on omics data across the tree of life. Its training corpus included DNA methylation levels measured in cell lines and tissue samples, but no 5hmC data.

First, we post-trained Omnii to predict patient-specific cell-free 5hmC profiles from germline genetic variants and cell-free WGS. Omnii was readily adaptable to this task, even though the model had never seen this modality during pretraining. We found impressive generalization, including the ability to predict the effects of germline genetic variants on 5hmC abundance in gene bodies.

We then used Omnii-generated features in a classifier for cancer prediction. In the multi-cancer early detection setting, Omnii features improved both cancer detection sensitivity and tissue-of-origin classification at high specificity. Omnii delivered the largest sensitivity gains in patients with tumor fractions below the detection limits of conventional methods, while reducing the classifier's dependence on patient age and sex.

We post-trained Omnii 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. Quality-filtered genotypes were incorporated into the hg38 reference to generate individual-specific genome inputs.

Despite never seeing 5hmC data in pretraining, Omnii readily adapted to predicting base-resolution profiles of this new modality. We evaluated the model in 32 held-out patients, finding 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’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 reconstructions and matched WGS in the 32 held-out patients. Across the 250 held-out QTLs with the strongest discovery associations, Omnii’s dosage effects closely tracked those measured in 5hmC coverage, while effects estimated from matched WGS showed little correlation, suggesting that Omnii recovered genetic associations with 5hmC that were not apparent in WGS coverage alone (Figure 3).

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

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

ClearNote’s multi-cancer early detection dataset includes patients with tumors from eight different tissues. Using the same feature sets, we trained classifiers to predict tissue of origin across cancer patients of all stages. Adding Omnii features also improved top-1 tissue-of-origin classification performance at 98% specificity (Figure 4).

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

Even small sensitivity gains could matter at screening scale. Applying the historical stage I–II cancer incidence rate from Hubbell et. al4 to the approximately 110 million US adults aged 50-795 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. 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. This is an illustrative extrapolation, not a prediction for Omnii. The benefit of additional detections may differ from the average benefit of screening, and translating earlier diagnosis into mortality reduction remains uncertain.6

Omnii, 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 screening studies are needed to establish whether these improvements translate into clinical practice.

We're hiring → join us