#13 · ML / AI methods

Deep generative and contrastive models integrate heterogeneous omics and text data

The common thread is learning shared representations from heterogeneous, partially observed biological measurements. PlantBiMoE combines bidirectional Mamba with sparse mixture-of-experts for plant genome language modeling, evaluated on a 31-dataset benchmark. mmContext contrastively embeds omics (AnnData .obsm) and text, enabling zero-shot cell type and disease prediction. CytoOne applies a Bayesian deep generative framework to CyTOF data spanning batch correction, differential analysis, and visualization. AUTOENCODIX standardizes autoencoder training and evaluation, and Scouter uses LLM gene embeddings to predict responses to unassayed perturbations. Explicit transfer is most evident in Prompt-CBP, which reports cross-species promoter prediction under leave-one-species-out evaluation. DSSMReg addresses regulon inference from scRNA-seq and does not share this transfer aim.