A RAIR-ATC transcriptional axis and multimodal drug-response modelling reveal class-level vulnerabilities in thyroid cancer.

Introduction: Radioiodine-refractory papillary thyroid carcinoma (RAIR PTC) and anaplastic thyroid carcinoma (ATC) are clinically challenging thyroid cancer states, yet the molecular continuum between radioiodine refractoriness, anaplastic transformation and targeted drug sensitivity remains unclear. We hypothesised that RAIR and ATC share a common transcriptional axis that can be distilled into a transferable signature to prioritise targeted therapies in RAIR-like models. Methods: We integrated RNA-seq and microarray cohorts (TCGA-THCA, one RAIR vs. radioiodine-avid PTC cohort, and five ATC cohorts) to derive RAIR/ATC gene modules and annotate pathway activity (MSigDB Hallmark). A single-sample RAIR signature was applied to TCGA tumours and DepMap thyroid cell lines. We combined RAIR scores with PRISM drug-response AUC data (18 thyroid cell lines) to quantify differential sensitivity (ΔAUC) and trained an interpretable multimodal ridge model integrating cell and drug features to predict AUC. External transferability and pharmacology-oriented mechanisms were assessed in an independent thyroid RNA-seq cohort (GSE126698) using ssGSEA and mechanism-relevant pathways. Results: RAIR and ATC upregulated modules converged on an EMT-angiogenesis-inflammatory/interferon axis, with ATC representing a more extreme state. The RAIR signature stratified TCGA tumours with graded activation of these programs and was largely independent of BRAF and TERT promoter status, while preserving pathway associations when transferred to DepMap. In PRISM, VEGFR/KDR and RAF/BRAF inhibitor classes showed the most negative class-averaged ΔAUC values, indicating preferential activity in RAIR-high thyroid lines. The multimodal ridge model improved AUC prediction over cell-only or drug-only baselines and was concordant with class-level sensitivity patterns. In GSE126698, RAIR scores retained robust associations with hypoxia and interferon/inflammatory programs and aligned with drug-mechanism pathways including KRAS signaling up and PI3K-AKT-mTOR signaling. Initial Conclusion: A module-based RAIR signature captures a disease-focused component of a broader RAIR-ATC axis, transfers to cell-line models and can be embedded into an interpretable multimodal framework for drug-response prediction and targeted drug-class prioritisation, prioritising VEGFR/KDR and RAF/BRAF inhibitor classes as candidates for further translational evaluation in RAIR-like thyroid models.