Cross-Omics Imputation

The fundamental mechanism relies on shared latent biological variation. A model learns the correlational structure between, for example, the transcriptome and the proteome from complete training samples. When a new sample has only RNA-seq data, the model predicts the missing protein abundances by mapping the observed gene expression through the learned cross-modal covariance matrix. This is distinct from simple mean imputation, as it preserves the conditional dependencies specific to that sample's molecular profile.

State-of-the-art methods use Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). A multi-modal VAE learns a joint latent space where both omics types are encoded. To impute, the model encodes the available modality into this latent space and then decodes it through the missing modality's decoder. Conditional GANs train a generator to produce realistic missing omics data that a discriminator cannot distinguish from real measurements, ensuring the imputed values respect the complex, non-linear distributions of biological data.

Methods like Multi-Omics Factor Analysis (MOFA) and Non-Negative Matrix Factorization (NMF) decompose multi-omics data into a low-rank structure. They identify a set of latent factors that explain variance across all modalities. Imputation is performed by reconstructing the missing data matrix from the shared factor loadings and the sample-specific factor scores inferred from the observed modalities. This approach is highly interpretable, as the latent factors often correspond to biological pathways or technical artifacts.

A model is pre-trained on a large, complete reference dataset (e.g., The Cancer Genome Atlas) to learn a general mapping between omics layers. This pre-trained model is then fine-tuned on a smaller, target cohort where imputation is needed. The pre-training step captures universal biological relationships, while fine-tuning adapts these relationships to the specific disease context or experimental platform of the target study, significantly improving accuracy in data-scarce scenarios.

This strategy uses Similarity Network Fusion (SNF) or Graph Convolutional Networks (GCNs). A patient similarity network is built from the available omics data. Missing values for a patient are imputed by propagating information from their nearest neighbors in the network who have complete data. GCNs formalize this by learning to aggregate feature information across the graph's edges, allowing the model to impute a patient's missing proteomics based on the proteomic profiles of genomically similar patients.

A critical characteristic of robust imputation is reporting prediction confidence. Bayesian models and VAEs naturally output a posterior distribution for each imputed value, not just a point estimate. This allows downstream analyses to be weighted by imputation certainty. For example, a biomarker discovery model can penalize or exclude features with high imputation variance, preventing the propagation of hallucinated molecular associations into clinical decision-making pipelines.