Revealing small non-coding RNA signatures in non-small cell lung cancer using machine learning

Abstract

Non-small cell lung cancer (NSCLC) is ranked as the second leading type of cancer among both genders in the United States. Despite a decline in death rates due to better smoking cessation and treatment improvement, NSCLC in non-smokers is still increasing. Early detection of cancer can significantly increase the overall survival of patients. Small non-coding RNA (sncRNA) such as tRNA derived fragments (tRDF) and piRNA have been identified in different types of tissue in next generation sequencing dataset. Increasing evidence has shown that both tRDF and piRNA expressed aberrantly in cancers, which is expected with novel biomarkers on the identification of cancer cases. Up till now, no comprehensive study has investigated these two types of sncRNA expression profile and their biological functions in NSCLC. With the development of different machine learning techniques, more methods can be applied in the classification of cancer and non-cancer cases based on the small RNA-seq data. In this work, this study utilized the multi-center small RNA-seq raw data from NSCLC combined machine learning methods with the aims of 1) identify promising signatures from tRDFs for NSCLC diagnosis, prognosis and its biological function; 2) identify promising signatures from piRNAs for NSCLC diagnosis and its potential in liquid biopsy. Besides the ML application for cancer detection, this study took advantage of the Transformer architecture to understand the contextual interaction between miRNA/piRNA and mRNA via target sites on specific genes, which is the last aim 3) prediction of miRNA/piRNA target gene and target site by integrative T5 model of miRNA/piRNA-mRNA binding and targeting experiment data. The study comprehensively analyzed multi-center data across thousands of NSCLC patient samples from tissue, plasma, and exosome datasets. For tRDFs, diagnostic and prognostic signatures were identified through expression profiles, which six diagnostic signatures achieved AUC values up to 0.90 in independent validations. For piRNAs, the study established and validated a novel piRNA-Based Tumor Probability Index (pi-TPI) by integrating plasma and exosomal piRNA signatures. The pi-TPI model incorporating five ribosome-derived piRNA signatures distinguished NSCLC patients from healthy individuals with AUCs exceeding 0.80, achieving 0.85 in plasma cohorts and 0.96 in non-cancer versus cancer subgroups. In parallel, a deep learning framework was developed named Contrastive miRNA-mRNA Sequence model (COMIMS), using transformer-based encoders and contrastive learning to predict sncRNA-mRNA interactions with high precision. The COMIMS framework achieved an AUC of 0.96 for 3’UTR and 0.99 for CDS miRNA-mRNA interaction prediction, confirming exceptional generalizability. When fine-tuned on piRNA-mRNA data from CLASH-seq experiments, COMIMS maintained high performance with AUC as 0.96 in predicting 5’UTR binding sites in lung cancer cell lines. The integrated study provides a comprehensive atlas of sncRNA-mRNA regulatory networks in NSCLC and introduces robust diagnostic and prognostic frameworks based on tRDFs, piRNAs, and miRNAs. The combination of experimental validation and deep learning-based prediction underscores the potential of sncRNA biomarkers for early, non-invasive detection and molecular stratification of lung cancer, paving the way toward precision diagnostics and therapeutic intervention.