#5 · ML / AI methods

Explicit latent structure for single-cell perturbation response and multi-omic modeling

The dominant thread is learned latent spaces applied to predicting single-cell responses to perturbation and to building integrated cell representations. PRiMeFlow applies flow matching in gene expression space to genetic and small-molecule perturbations, validated on PerturBench, while SAVE pairs a conditional Transformer with flow matching and gene block attention for multi-condition generation. DeSCOPE, a lightweight conditional VAE, predicts genetic perturbation responses across transcriptomic and epigenomic data. scDisent disentangles expression and regulation latent variables with a sparse causal module for multi-omic integration, and ProtoCloud, a published self-explaining model, uses cell-type prototypes with uncertainty estimates. The thread is not universal: layer-wise evaluation of scFoundation and Tahoe-X1 shows optimal layers vary by task and cellular context. What is new is explicit structure imposed on latent space, including conditioning, disentanglement, causal sparsity, and prototypes.