1174P - FLAMINGO: Accurate cancer detection from ultra-low-pass whole genome sequencing of cell-free DNA

Background

Cell-free DNA (cfDNA) obtained from minimally invasive liquid biopsies holds promise as a biomarker for early cancer detection, offering a molecular snapshot of the patient's condition. However, current methods lack the sensitivity required for early cancer detection. RenovaroCube ("The Cube"), an AI platform for cancer diagnostics, recognizes that no single model or approach can achieve the necessary sensitivity alone. Thus, The Cube integrates multi-omic data, employing a library of trained models for various omic layers. One such model, Flamingo, focuses on detecting cancer from cfDNA sequencing data using fragmentomics.

Methods

Flamingo utilizes mapping coordinates from as few as 200,000 cfDNA fragments per patient, extracted from Delfi cfDNA WGS experiments. Fragment lengths, sequence motifs, and Shannon entropy of the motifs are analysed, serving as input for Flamingo—a neural network trained to differentiate cancer from healthy samples. Performance evaluation via cross-validation on a training cohort (n=136 healthy, n=127 cancer) and validation on a separate cohort (n=55 healthy, n=48 cancer) demonstrated Flamingo's robustness.

Results

Using the Delfi dataset, Flamingo achieved a median AUROC of 0.952 and a sensitivity of 74.2% at 98% specificity during cross-validation. In the validation set, Flamingo maintained high specificity (98.2%) and sensitivity (75.0%), with an AUROC of 0.966, confirming its generalizability beyond the training dataset. Flamingo and the Delfi model identified non-overlapping cancer cases, supporting our belief that multiple models and modalities are required for comprehensive cancer detection.

Conclusions

In summary, Flamingo, developed from minute amounts of cfDNA WGS data, performs comparably to state-of-the-art cancer detection models. Integrated into The Cube platform, Flamingo enhances cancer detection across diverse omic layers, facilitating early intervention and improved patient outcomes.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

D. Vessies, T. Bucho, E. Post, K. Roohollahi, J. van den Berg, A. Foster, S. Patel, A. Vandor, D. Makarawung: Financial Interests, Personal, Full or part-time Employment: RenovaroCube; Financial Interests, Personal, Stocks/Shares: RenovaroCube. F. van Asch: Financial Interests, Personal, Member of Board of Directors: RenovaroCube; Financial Interests, Personal, Stocks or ownership: RenovaroCube; Financial Interests, Personal, Ownership Interest: RenovaroCube.