Joint variational autoencoders for multimodal imputation and embedding.

Area of Science:

• Computational Biology
• Genomics
• Neuroscience

Background:

• Single-cell multimodal datasets offer deep insights into cellular and molecular mechanisms.
• Data generation challenges and missing modalities limit current approaches.
• Existing machine learning methods often require fully matched data, lacking modality specificity.

Purpose of the Study:

• To develop an open-source machine learning model, Joint Variational Autoencoders for multimodal Imputation and Embedding (JAMIE), to address challenges in single-cell multimodal data analysis.
• To enable accurate imputation and embedding from partially matched multimodal datasets.
• To provide interpretable insights into cellular mechanisms using prioritized features.

Main Methods:

• Developed JAMIE, a model utilizing variational autoencoders to learn modality-specific latent embeddings.
• Aggregated embeddings from matched samples to identify joint cross-modal latent embeddings.
• Employed Shapley values for feature prioritization in imputation and interpretation.

Main Results:

• JAMIE effectively handles partially matched single-cell multimodal data.
• The model outperforms existing state-of-the-art methods in imputation accuracy.
• Identified prioritized multimodal features for imputation, offering novel mechanistic insights.

Conclusions:

• JAMIE provides a robust solution for analyzing incomplete single-cell multimodal data.
• The model facilitates a deeper understanding of cellular and molecular mechanisms.
• JAMIE enables high-resolution mechanistic insights from diverse biological data.