A new episodic learning capability lets Synthetic Cognition process DNA as sequences rather than static features. Benchmarked across 57 genomic classification tasks against DNABERT-2, NT-v2, and HyenaDNA, it won outright in 52% of them — using a fraction of a percent of the pretraining data those foundation models require.

With just 10 participants from the BCN02 HIV cure trial, Synthetic Cognition jointly analyzed transcriptomics, methylomics, and gut microbiome data, surfacing 320 genes and 1,295 methylation sites that separate viral controllers from non-controllers — with perfect classification on held-out patients.

On a public multi-omics breast cancer benchmark, Synthetic Cognition reached 0.934 accuracy classifying tumor subtypes — ahead of the previous best-performing graph neural network — while also showing exactly which molecular features drove each classification.

Extended to single-cell data, Synthetic Cognition clustered 206 cells using RNA expression and chromatin accessibility together, matching or beating dedicated joint-dimensionality-reduction methods built specifically for this task.

In a public hospital-run competition where only 2% of cases were positive, a Synthetic Cognition ensemble reached 2nd place (AUC 0.592) predicting sepsis risk in infants — without generating any synthetic samples for the minority class.