RNA1-DA: A domain-adaptive RNA foundation model for forward and reverse translation

Introduction. Preclinical cancer models and patient tumors differ in cellular composition, molecular profiles, and environmental contexts. To enable translation between clinical and preclinical systems, we developed a new foundation model, RNA1-DA, with domain adaptation between tumor, cell line, organoid, and xenograft samples. We demonstrate that RNA1-DA enables key translational research tasks, including molecular subtype transfer, preclinical model selection, and drug response prediction. Methods. We previously described RNA1 - a transformer-based RNA expression foundation model trained on 182,383 bulk RNA-seq cancer samples with self-supervised and multi-task training. Here we develop RNA1-DA, which extends RNA1 to enable the joint integration of clinical and preclinical samples using (a) a layer to deconvolve cancer cell expression from tumor samples, and (b) a domain adaptation layer using an adversarial autoencoder to integrate RNA1 embeddings across sample types. We developed a systematic evaluation framework to assess clinical-preclinical alignment in RNA1-DA embeddings by measuring the preservation and transfer of disease identity, molecular subtypes, cancer driver gene biology, and drug response across systems. Results. Using RNA1-DA, we integrated 30,810 tumor tissue, 94,973 cell line, 714 organoid, 1290 cell line-derived xenograft, and 2526 patient-derived xenograft samples. RNA1-DA aligned key biological structure across systems, including disease identity, molecular subtypes, and driver gene alterations. Preclinical samples achieved 62-88% accurate disease classification based on proximity to clinical tumors, outperforming comparator methods. Thirteen TCGA and 61 RNA1-derived clinical molecular subtypings were systematically transferred from clinical to preclinical samples and showed concordance with canonical markers and genetic dependencies; for example, assigned breast cancer subtypes matched canonical cell line subtype annotations (p=1.8e-8) and known genetic dependencies (e.g., ESR1, ERBB2, CDK4). Translational model selection was further supported by a novel transcriptomic-genomic neighbor overlap metric, which demonstrated significant correspondence (p<0.05) between RNA1-DA embedding neighbors and driver gene alteration neighbors across all 15 cancers evaluated. In addition, RNA1-DA enabled improved cell line drug response prediction through multi-task fine-tuning on CTRP screens, achieving substantially higher performance than baseline methods (median Spearman correlation 0.60 vs 0.35). Together, these results demonstrate the utility of RNA1-DA in supporting key translational research tasks within a unified framework, including molecular subtype transfer, preclinical model selection, and drug response prediction.