From Tissue Archives to Liquid Biopsy: Transfer Learning for MicroRNA-Based Lung Cancer Diagnosis

Serum microRNA-based liquid biopsy holds a great application prospect in noninvasive lung cancer diagnosis. However, building an efficient serum-based miRNA classifier remains a challenging issue due to the nearly infinite miRNA combinations and the scarcity of large-scale, clinically annotated serum samples. Here, we develop a transfer learning strategy with feature space alignment integrating molecular detection, enabling effective domain adaptation and knowledge transfer from large-scale tissue data to limited serum data sets, and apply it to serum-based lung cancer diagnosis. A 4-miRNA panel (miR-139-5p, miR-10a-5p, miR-148a-3p, miR-30d-5p) is identified through genetic algorithm-driven feature selection and unsupervised clustering analysis, demonstrating high accuracy (AUC > 0.98) in tissue-based classification. Their expression data are accurately quantified via reverse transcription quantitative PCR in 89 clinical serum samples. The transfer-learned model ultimately achieves a high classification accuracy of 91.5% and a sensitivity of 92.2% on the clinical serum test set. We envision that the approach offers a cost-effective solution for high-accuracy liquid biopsy with limited samples.