Pre-clinical models rarely line up cleanly with patients. We introduce RNA1-DA, a domain-adaptation layer that aligns tumors, cell lines, PDX and organoids in a shared expression embedding — enabling drug-response prediction to transfer across the translational gap.
Systematic differences between sample types — tumor purity, microenvironment, culture artifacts — dominate naive expression embeddings, so models trained on one domain fail on another. RNA1-DA learns a representation in which biological signal, not sample provenance, drives the geometry.
A shared embedding for pre-clinical and clinical samples
RNA1-DA builds on the RNA1 foundation model, adding a domain-adaptation objective that penalizes sample-type separability while preserving subtype structure. The result aligns five sample types into one space where nearest neighbors share biology rather than origin.
Forward and reverse translation
The alignment supports translation in both directions: forward, predicting patient response from pre-clinical evidence; and reverse, selecting the pre-clinical models that best mirror a given patient population.
- Forward. Cell-line drug-response signatures predict patient outcomes in held-out clinical cohorts.
- Reverse. Given a patient subtype, rank PDX and organoid models by biological similarity.
The translational gap has always been a data-alignment problem as much as a biology problem. RNA1-DA treats it as one.— Data4Cure Research
Reproduce it on the platform
The full workflow — data harmonization, RNA1-DA fine-tuning, and response transfer — runs end-to-end in the Biomedical Intelligence® Cloud, grounded in the CURIE Knowledge Graph™.
RNA1-DA was introduced in two late-breaking posters at the 2026 AACR Annual Meeting.