Single-cell and deep learning identify hypoxia-responsive lncRNAs predicting outcomes in colorectal cancer

Emerging evidence highlights hypoxia-responsive long non-coding RNAs (lncRNAs) as potential modulators in tumor biology. In this study, we explored the significance of a hypoxia-responsive lncRNA molecular signature (HRLPMS) and the therapeutic implications of hypoxia-responsive lncRNAs in colorectal cancer (CRC). To assess the significance of HRLPMS, we integrated bulk transcriptomic and proteomic data, single-cell RNA-seq (scRNA-seq), spatial transcriptomics (ST) data, therapy-specific clinical cohorts, and our in-house data. We further evaluated the TME characteristics, somatic variations, drug sensitivity, and applied multiple machine learning (ML) and deep learning (DL) algorithms to validate the prognostic power of HRLPMS. Pan-cancer analysis revealed that HRLPMS functions as a risk factor across most cancer types. In CRC, HRLPMS was associated with chromosomal instability, adverse pathological characteristics, and poor survival outcomes, as confirmed by Cox, ML, and DL models. This signature was notably enriched in immune and stromal cell populations, such as fibroblasts. Distinct patterns of somatic variation were observed between the high- and low-HRLPMS groups. Cell-state analysis indicated that low-HRLPMS cells, characterized by immune and inflammatory features, predominated during early-to-middle pseudotime, whereas high-HRLPMS cells emerged later, exhibiting angiogenesis and extracellular matrix (ECM) remodeling characteristics. Further analysis demonstrated that APP-CD74 interactions may mediate immunosuppression and tumor progression. Furthermore, high-HRLPMS patients showed evidence of benefit from fluorouracil plus bevacizumab and a trend toward improved response to preoperative chemoradiotherapy. We found that HRLPMS represents a promising prognostic tool for CRC, with the potential to refine therapeutic strategies and enhance patient outcomes through tailored treatment approaches.