Missing Modality Imputation

The foundational task of computationally converting one data modality into another using encoder-decoder architectures. In genomics, this involves predicting an entire omics layer—such as chromatin accessibility profiles from DNA sequence alone or proteomic abundance from transcriptomic data—without ever measuring the target modality experimentally. The encoder compresses the source modality into a latent representation, and the decoder generates the target modality from that compressed signal.

A generative probabilistic framework that learns a joint posterior distribution from multiple input omics layers. MVAEs are central to missing modality imputation because they model a shared latent space from which any individual modality can be reconstructed. Key properties include:

• Product-of-experts inference: Combines evidence from available modalities to infer the latent state
• Conditional generation: Samples the missing modality from the learned conditional distribution given observed modalities
• Uncertainty quantification: Provides variance estimates for imputed values

A regularization technique where entire data modalities are randomly zeroed out during training to force the model to learn robust representations that handle missing clinical assays. By stochastically removing modalities like DNA methylation or copy number variation in each training batch, the model learns to perform inference from arbitrary subsets of available data. This directly prepares the architecture for deployment scenarios where specific omics layers are unavailable due to cost or sample limitations.

A shared, lower-dimensional mathematical representation where embeddings from distinct biological modalities are aligned. For missing modality imputation, the joint latent space serves as the information bottleneck through which cross-modal translation occurs. When RNA-seq and ATAC-seq embeddings occupy proximal positions for the same biological state, a model can traverse from the RNA embedding to generate the corresponding ATAC profile. The quality of imputation depends directly on the alignment precision within this space.

A self-supervised training paradigm that pulls paired omics profiles together in the latent space while pushing unpaired profiles apart. Applied to missing modality imputation, contrastive objectives ensure that the latent representation of a cell's transcriptome is maximally similar to the latent representation of the same cell's proteome. This alignment enables accurate cross-modal generation: given only transcriptomic data, the model can retrieve the corresponding proteomic latent code and decode it into protein abundance estimates.

A transformer component that allows one biological sequence modality to selectively query contextual information from another modality during the imputation process. For example, when predicting protein binding tracks from DNA sequence, cross-attention enables the DNA encoder to attend to relevant epigenomic features from a partially available modality. This dynamic weighting mechanism ensures that imputation is context-dependent, leveraging whatever complementary data is present rather than relying solely on static learned priors.