A new VAE tries to stop single-cell genomics research from squeezing three incompatible goals into one representation.
The model, called scTrilemma, is a variational autoencoder for single-cell RNA sequencing data. Instead of forcing every signal through one embedding, it splits the work three ways: gene-level expression is gated directly, the cell's context is routed through the decoder, and the model's prior is conditioned on unlabeled pseudo-bulk data, all under a single reconstruction objective with no labeled targets needed. The team tested it in a zero-shot setup across successive releases of the CZ CELLxGENE Census, a large public single-cell dataset. It held up on biological-state, differential-expression, and pathway structure across multiple disease settings, and the code is public on GitHub.
Single-cell analysis is notoriously label-free. There's no fixed ground truth for what counts as signal versus noise in a given cell, so one representation has to satisfy three demands at once: preserve identity, resist nuisance context, and keep raw expression fidelity intact. The authors call that tension the representation trilemma. Most approaches pick two of the three and eat the cost on the remaining one. scTrilemma's architecture avoids that trade by giving each demand its own output path instead of cramming all three into one bottleneck.
The authors' own latent interventions found a limit to the fix: once nuisance context is removed, identity and fidelity still pull against each other, so the trilemma shrinks to a dilemma rather than vanishing.