'Building Molecular QR Codes to Store Human Cell Identity'

We had previously identified a class of cancer-emergent small non-coding RNAs, termed oncRNAs, which are absent in normal tissues but actively expressed and secreted by tumors. In a systematic effort, we annotated oncRNAs across 32 human cancers and demonstrated that their presence and absence define highly specific, digital expression barcodes reflective of cancer type and subtype. We had also shown that a subset of these oncRNAs is detectable in circulation, enabling minimally invasive disease monitoring. To harness the diagnostic potential of these molecules, we developed Orion, a deep generative variational model that learns robust, generalizable embeddings of cell-free oncRNA profiles. Orion achieved high sensitivity and specificity for early-stage lung cancer detection and effectively disentangled technical confounders from biological signal. To scale cfRNA-based diagnostics across diverse biofluids and low-resource settings, we introduced Exai-1, a multi-modal foundation model trained on over 13,000 plasma and serum samples. By combining RNA sequence and abundance data, Exai-1 captured biologically meaningful structure, denoised sparse measurements, and enabled synthetic data generation to improve performance in downstream classification tasks. Together, these efforts establish a foundation for representation learning in cfRNA and enable next-generation liquid biopsy applications.