Large language model produces high accurate diagnosis of cancer from end-motif profiles of cell-free DNA

Jilei Liu ¹, Hongru Shen ¹, Kexin Chen ²

⁰Key Laboratory of Cancer Prevention and Therapy, Tianjin Cancer Institute, Tianjin's Clinical Research Center for Cancer, Tianjin Medical University Cancer Institute and Hospital, National Clinical Research Center for Cancer, Tianjin Medical University, Tianjin, 300060, China.

Briefings in Bioinformatics +

|

September 2, 2024

View abstract on PubMed

Summary

We developed iLLMAC, an AI model using instruction-tuned large language models (LLMs), to detect cancer from cell-free DNA (cfDNA) end-motif profiles. This novel approach shows high accuracy in diagnosing various cancers, including hepatocellular carcinoma.

Area Of Science

Artificial Intelligence in Oncology
Genomics and Bioinformatics
Liquid Biopsy for Cancer Detection

Background

Instruction-tuned large language models (LLMs) show promise in aligning with human intentions.
Accurate and early cancer detection remains a critical challenge in oncology.
Cell-free deoxyribonucleic acid (cfDNA) end-motif profiling offers a potential non-invasive biomarker for cancer.

Purpose Of The Study

To develop and evaluate an LLM-based model for cancer detection using cfDNA end-motif profiles.
To assess the performance of the model, named iLLMAC (LLM-based instruction-tuned LLM for Assessment of Cancer), across different datasets and cancer types.
To investigate the impact of the number of cfDNA end-motifs on diagnostic accuracy.

Main Methods

Developed iLLMAC, an instruction-tuned LLM, trained on plasma cfDNA sequencing data from 1135 cancer patients and 1106 controls.
Utilized cfDNA end-motif profiles as input features for cancer detection.
Evaluated model performance using area under the receiver operating curve (AUROC) on training and external validation datasets.

Main Results

iLLMAC achieved an AUROC of 0.866 for general cancer diagnosis and 0.924 for hepatocellular carcinoma (HCC) detection using 16 end-motifs.
Performance improved with more motifs, reaching AUROCs of 0.886 for cancer diagnosis and 0.956 for HCC detection with 64 end-motifs.
On an external test set, iLLMAC achieved AUROCs of 0.912 for cancer diagnosis and 0.938 for HCC detection, outperforming benchmark methods.

Conclusions

LLM-based instruction-tuning is effective for developing accurate cfDNA-based cancer detection models.
iLLMAC demonstrates significant potential for non-invasive cancer diagnosis, particularly for HCC.
The study highlights the utility of cfDNA end-motif profiles and advanced AI for advancing cancer diagnostics.

Keywords:

cell-free DNA early cancer diagnosis large language models