1186P - SmartCS-LPLLM: Enhancing early cancer detection through ctDNA methylation analysis leveraging large language models

Background

Early cancer screening using circulating tumor DNA (ctDNA) faces challenges due to low abundance and a high signal-to-noise ratio. We aimed to develop a robust screening model that overcomes these limitations.

Methods

Low-pass whole-genome bisulfite sequencing (Low pass-WGBS) was utilized with the high-efficiency WATCHMaker (7K0101-096) library preparation kit for the optimization of cell-free DNA (cfDNA) sample processing, with sample loss minimized and molecular conversion efficiency enhanced. Thirteen cancer-specific differentially methylated regions (DMRs), including those related to lung and liver cancers, were targeted in the analysis. The SmartCS-LPLLM model, a single-molecule multimodal early cancer screening model based on large language models, was developed. Cancer signals were precisely identified by this model through the analysis of cfDNA features, including methylation scoring, sequence length, terminal motif characteristics, and sequence linguistic features.

Results

Reanalysis of public data from BMC Medicine (CRA001537) demonstrated the SmartCS-LPLLM model's significant improvement in differentiating hepatocellular carcinoma (HCC) from non-HCC samples, with an increased AUC value of 0.967. In a blind test of 12 cfDNA samples, the model accurately classified all 5 liver cancer samples. Notably, the model has been enhanced to accurately identify ctDNA at a concentration as low as 0.05%. Furthermore, during the model's construction, it was observed that the highest accuracy was achieved when the DMR region was 120M, with the single-molecule read-level model achieving a 85% accuracy rate in distinguishing tumor from healthy reads.

Conclusions

The SmartCS-LPLLM model, integrating biological features like methylation and copy number variations (CNVs), provides a precise clinical strategy for early cancer screening. Its performance in blind tests confirms its robustness and suitability for identifying low-abundance ctDNA samples, indicating significant clinical relevance.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.