Exai-1

Liquid biopsy — reading disease state from molecules circulating in blood — has mostly been built on cell-free DNA. Cell-free RNA carries a complementary and arguably richer signal, including the orphan non-coding RNAs that tumors actively secrete, but it is punishing data to model: extremely high dimensional, extremely sparse, and available in cohorts of thousands rather than millions of samples. Exai-1 is a foundation model built specifically for that regime, learning a latent representation of cfRNA profiles that can then be reused across diagnostic tasks instead of training a classifier from scratch for each indication.

Exai-1 was developed at Exai Bio with Hani Goodarzi's lab at UCSF and the Arc Institute, and a collaborator at UC Davis. Its central design choice is multimodality: rather than treating each cfRNA species as an anonymous feature index, the model initializes each feature's embedding from the sequence-derived representation produced by RNA-FM, then scales that embedding by the measured abundance of the species in a given sample. Sequence identity and expression level therefore enter the model as a single unified vector, letting attention operate over relationships between RNAs that share sequence or structural character.

The work was posted as a preprint in March 2025 and published in Nature Machine Intelligence in December 2025. The model name appears only in the paper body, not in the title, which is why it is easy to miss in citation lists.

#Key Features

• Sequence-informed feature embeddings: The embedding layer is initialized from the top 32 principal components of RNA-FM's representation space, which together explain over 91% of the variance, so chemically and structurally related RNAs start out near each other.
• Variational transformer: A transformer encoder compresses the scaled embeddings into a latent Gaussian distribution that a transformer decoder samples from to reconstruct abundances; ablating either self-attention or the variational layer degrades both reconstruction and downstream classification.
• Task tokens for multi-task context: Dedicated tokens for cancer detection, tissue of origin, assay version, and biofluid type let the model absorb clinical and technical context alongside expression, and removing them consistently hurts downstream performance.
• Generative data augmentation: Reconstructed synthetic cfRNA profiles can double a small training set, raising cancer-detection AUC by an average of 0.032 in the low-sample regime.
• Cross-biofluid transfer: Classifiers trained in Exai-1's latent space on plasma retain performance on serum, where classifiers trained on raw abundance values collapse.

#Technical Details

Exai-1 was pretrained with masked language modeling over 306 billion tokens from 8,339 blood samples, with 2,079 used for validation and 2,596 held out for reporting, drawn from a 13,014-sample cohort split between 6,792 serum and 6,222 plasma specimens. The feature set was distilled from 9,491,734 annotated cfRNAs down to 7,349 species — 704 tRNAs, 610 snoRNAs, 761 yRNAs, 716 miRNAs, and 4,558 representative orphan non-coding RNAs — by filtering on prevalence and coefficient of variation, then selecting oncRNAs evenly across RNA-FM embedding space. Because self-attention operates on the compressed hidden space rather than raw features, the model has only 3.6 million trainable parameters. Reconstructing 25% randomly masked features in held-out samples, Exai-1 reached an R² of 0.89 (95% CI 0.88-0.89) against a dataset-average baseline at 0.57 (95% CI 0.55-0.59). On paired plasma and serum samples excluded from training, an XGBoost cancer classifier built on raw abundance scored AUROC 0.74 on plasma but 0.56 on serum, while the same setup on Exai-1 latents held comparable performance across both biofluids.

#Applications

Exai-1 is aimed at cfRNA-based diagnostics: early cancer detection, tissue-of- origin inference, minimal residual disease monitoring, and treatment response tracking from small blood volumes. Its most immediately practical contributions are denoising sparse profiles and generating synthetic samples to augment small cohorts, which matters most for rare cancers where assembling hundreds of cases is infeasible — demonstrated on an ovarian cohort of 48 samples, where augmentation improved AUC by 0.063. The biofluid-disentangling behavior also addresses a real operational problem, letting classifiers developed on one sample type and assay version be applied to archives collected under another.

#Impact

Exai-1 extends foundation-model methodology into liquid biopsy, a domain previously excluded from it by data scarcity, and does so with an architecture small enough that overfitting is a manageable rather than dominant concern. Its sequence-plus-abundance formulation is a reusable idea for any assay where features are biological entities with their own sequence identity. Caveats are substantive: the training cohort is proprietary to a single company, most of the underlying clinical data is released only on request under a data transfer agreement, and downstream evaluation is retrospective, so clinical utility remains to be established prospectively. The reproduction code is released on GitHub under GPL-3.0 and the pretrained parameters and cfRNA expression matrices are deposited on Zenodo, though under a non-commercial, no-derivatives license that limits reuse.