Liquid biopsy enables non-invasive cancer screening and monitoring by analyzing tumor-derived biomarkers in the blood — but those signals are often sparse and noisy. Exai Bio has advanced AI-powered liquid biopsy using novel small-RNA biomarkers, and in recent work two generative AI models for cell-free RNA — Exai-1 and Orion — reported meaningful gains in signal denoising and early cancer detection. According to a joint Databricks and Exai Bio write-up, these advances were built on Databricks’ Lakehouse architecture and cloud AI infrastructure, which integrate large genomic datasets and provide managed ML tooling (MLflow, Workflows, scalable clusters) for training large multimodal models on thousands of patient samples.
Multimodal foundation models for liquid biopsy
Exai Bio’s research introduces large generative models tailored to liquid-biopsy data, integrating sequence information, molecular abundance, and metadata to learn representations of cancer-associated RNAs.
Exai-1 (a cfRNA foundation model) is a transformer-based generative model that combines RNA sequence embeddings with cell-free RNA (cfRNA) abundance profiles. It was pretrained on a large-scale dataset — described in the accompanying Nature Machine Intelligence paper as over 306 billion sequence tokens — to learn biologically meaningful latent structure of cfRNA expression. By drawing on both sequence (via embeddings from the RNA-FM language model) and expression data, Exai-1 aims to enhance signal fidelity, reduce technical noise, and improve detection by generating synthetic cfRNA profiles. In practice it can denoise sparse cfRNA measurements and augment datasets, and classifiers trained on Exai-1’s reconstructed profiles consistently outperformed those trained on raw data.
Orion (an oncRNA generative classifier) is a specialized variational autoencoder for circulating orphan non-coding RNAs (oncRNAs) — small RNAs secreted by tumors. Orion uses a twin-VAE architecture: one input is a count vector of cancer-associated oncRNAs, the other a vector of control RNAs (such as housekeeping RNAs), each fed to a separate encoder. Its training combines adversarial and classification objectives, including a triplet-margin loss that pulls together same-phenotype samples (cancer vs. control) and separates different ones, reducing batch effects and technical variation. On a cohort of 1,050 non-small cell lung cancer (NSCLC) patients and matched controls, Orion reported 94% sensitivity at 87% specificity for detecting NSCLC across all stages — roughly 30% better than standard methods on the pooled data, per the Nature Communications paper.
Together, the two models form a scalable AI framework for liquid biopsy: Exai-1 provides a general-purpose cfRNA “language model” that generates realistic RNA profiles and boosts downstream classifiers, while Orion specializes that approach for lung cancer screening.
The enabling infrastructure
These results are powered by Databricks’ integrated data and AI platform, with several capabilities highlighted:
- Unified Lakehouse (Delta) storage: sample, lab, and experiment metadata live in Delta tables, preventing silos and enabling real-time analysis. Exai’s blood-sample corpus (serum and plasma) and thousands of prior small-RNA-seq datasets are registered in Delta and can be filtered and combined quickly for training.
- Scalable compute: cloud-native clusters let researchers spin up GPU or high-memory instances without heavy DevOps, with auto-termination and cost dashboards keeping budgets in check.
- Workflows: multi-stage ETL and training pipelines run with defined dependencies and parallelism, without hand-written orchestration.
- MLflow for MLOps: managed experiment tracking and a model registry capture hyperparameters, datasets, metrics, and artifacts across dozens of models, supporting ablation studies and reproducibility.
- Reproducible environments: container services and Git-based repos with CI/CD lock software dependencies, so every team member runs models in the same environment.
In effect, Databricks provided a turnkey MLOps stack — Spark ingestion, MLflow tracking, Workflows orchestration, and elastic auto-scaling compute. Related platform patterns for governed, model-driven biomedical data appear in this case study on data governance.
Impact on detection and biomarker discovery
- Better early detection: by amplifying cfRNA cancer signal against background blood RNA, the models aim to catch cancer earlier. Exai-1’s denoising yields clearer signals even in small-volume samples, and Orion’s generative embeddings achieved high sensitivity for early-stage lung cancer — improvements that could support more reliable blood-based screening.
- New biomarker insights: learning from raw RNA data reduces the bias of targeted panels. Orion identified numerous novel oncRNAs from tissue and public data (TCGA), with significance then evaluated in blood.
- Generative data augmentation: Exai-1 can simulate realistic cfRNA profiles by sampling its decoder, improving classifier training where real samples are scarce — helpful for rare cancer signatures.
- Scalable collaboration: biologists, bioinformaticians, biostatisticians, ML scientists, and data engineers work in one environment, which the team credits with supporting a rapid publication cadence across multiple cancer types.
Limitations and what to watch
- This is a vendor-and-partner account; the headline accuracy figures (94% sensitivity at 87% specificity) come from specific research cohorts and require prospective, multi-site validation before clinical claims can be generalized.
- Reported sample counts vary between the infrastructure narrative and the peer-reviewed papers; the figures here follow the published Nature papers where they differ, and specific numbers should be confirmed against those primary sources.
- Synthetic data augmentation can improve benchmarks but risks encoding model artifacts; independent evaluation on held-out real samples is essential.
- Diagnostic performance in screening depends heavily on the tested population’s cancer prevalence; sensitivity and specificity alone do not determine real-world positive predictive value.
- None of this constitutes medical advice; liquid-biopsy tests of this kind remain subject to regulatory review.
Looking ahead
The Exai Bio–Databricks collaboration illustrates how state-of-the-art generative models and modern cloud architecture together can push cancer diagnostics forward: Exai-1 and Orion show that deep generative learning can extract useful signal from liquid biopsies, supported by a Lakehouse that unifies heterogeneous biomedical data and managed ML tools that make large, reproducible experiments practical.
Sources: Karimzadeh et al., “A multimodal cell-free RNA language model for liquid biopsy applications,” Nature Machine Intelligence (2025); Karimzadeh et al., “Deep generative AI models analyzing circulating orphan non-coding RNAs enable detection of early-stage lung cancer,” Nature Communications (2024); Databricks documentation and blog.
