Multi-Omics Data Imputation

The fundamental mechanism involves learning a mapping function between a measured source modality (e.g., scRNA-seq) and an unmeasured target modality (e.g., surface protein abundance). Models are trained on paired multi-omics data from technologies like CITE-seq, where both modalities are simultaneously captured. Once trained, the model can impute the missing modality in new cohorts where only the source data exists, effectively translating between molecular languages.

TotalVI (total Variational Inference) is a seminal deep generative model for imputation. It uses a variational autoencoder (VAE) to learn a joint latent representation of RNA and protein data. Key features include:

• Models zero-inflated negative binomial distributions for RNA counts
• Accounts for batch effects and library size variation
• Provides uncertainty estimates for imputed values
• Enables downstream tasks like differential expression and clustering on the imputed data

scJoint leverages transfer learning to impute cross-modality data without requiring paired training examples. It uses a semi-supervised neural network trained on unpaired scRNA-seq and scATAC-seq data. The model learns to project both modalities into a shared latent space where cells of the same type cluster together, enabling label transfer and feature imputation across modalities through k-nearest neighbor queries in the latent space.

MultiVI extends the VAE framework to jointly model three omics layers simultaneously: transcriptomics, chromatin accessibility (ATAC-seq), and surface proteins. It learns a shared latent representation that captures the common biological signal across all modalities. This allows imputation of any missing modality from any combination of observed ones, creating a truly flexible multi-modal inference engine for single-cell data.

Imputation accuracy is assessed using multiple metrics on held-out paired data:

• Pearson correlation between imputed and true expression values
• Root Mean Square Error (RMSE) for continuous protein estimates
• AUROC for binary detection of protein presence/absence
• Preservation of cell-type clustering after imputation
• Downstream task performance, such as differential abundance testing, to ensure biological validity

A primary use case is integrating large-scale biobank cohorts where different omics were assayed in different subsets. For example, the UK Biobank has proteomics data for only a fraction of participants with genomics. Imputation models trained on the overlapping subset can predict proteomic profiles for the entire cohort, dramatically increasing statistical power for proteome-wide association studies (PWAS) and enabling multi-omics disease prediction models.

A generative deep learning architecture that learns a probabilistic latent representation of input data. In multi-omics imputation, VAEs are adapted to compress different modalities into a shared latent space, enabling the reconstruction of missing data types from observed ones. Key properties:

• Learns a smooth, continuous latent manifold
• Enables cross-modal generation without paired training data
• Models uncertainty in imputed values
• Architectures like scVAE and totalVI are foundational for single-cell multi-omics integration

A mathematical framework for finding the most efficient mapping between two probability distributions. Applied to multi-omics imputation, optimal transport aligns cells from different modalities by minimizing a cost function based on feature similarity:

• Maps cells between transcriptomic and proteomic spaces
• Handles unpaired datasets where different cells were measured for each modality
• Preserves local neighborhood structure during alignment
• Algorithms like Mowgli and SCOT leverage optimal transport for cross-modality prediction

A self-supervised learning paradigm that trains models to pull similar representations together and push dissimilar ones apart in embedding space. In multi-omics contexts:

• Aligns cells across modalities without paired labels
• Creates a shared embedding where transcriptomic and proteomic profiles of the same cell type cluster together
• Enables zero-shot imputation by finding nearest neighbors in the shared space
• Frameworks like CLIP-inspired models adapt this for biological modalities

A statistical framework for unsupervised integration of multiple omics data types. MOFA infers a low-dimensional set of latent factors that capture the principal sources of variation across modalities from the same set of samples:

• Handles missing data patterns natively
• Decomposes variation into factors shared across modalities and factors private to each
• Provides a principled foundation for understanding what information is transferable between omics layers
• Serves as a benchmark for evaluating imputation quality against known biological variation

Large, pre-trained neural networks like scGPT and Geneformer trained on massive corpora of single-cell transcriptomic data using self-supervised learning. These models:

• Generate general-purpose cellular representations
• Can be fine-tuned for cross-modality prediction tasks
• Learn transferable features that generalize across tissues and species
• Represent the state-of-the-art for imputing missing omics layers from transcriptomic observations alone
• Leverage attention mechanisms to capture complex gene-gene interaction patterns

Cellular Indexing of Transcriptomes and Epitopes by Sequencing — a multi-omics technology that simultaneously profiles the whole transcriptome and a panel of cell-surface proteins from the same single cell. This technology:

• Provides ground truth paired data for training imputation models
• Enables direct validation of transcriptome-to-proteome prediction accuracy
• Serves as the gold standard benchmark dataset for multi-omics imputation algorithms
• Demonstrates that RNA expression alone can predict surface protein abundance with high fidelity