A mixture of attention experts-embedded flow-based generative model to create synthetic cells in single-cell RNA-Seq...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNAseq) is crucial for understanding cellular heterogeneity in tissues, especially in cancer research.
• scRNAseq data presents challenges like high dimensionality, sparsity, and imbalanced cell-type distributions, necessitating advanced computational methods.
• Generative models, including Variational Autoencoders (VAE) and Generative Adversarial Networks (GANs), are increasingly used to address these challenges in scRNAseq data processing.

Purpose of the Study:

• To develop and evaluate novel Flow-based (FB) generative models for scRNAseq data analysis.
• To compare the performance of FB models against established generative techniques like VAE and GANs.
• To provide guidance for developing automated scRNAseq data analysis systems.

Main Methods:

• Developed a Masked Affine Autoregressive transform-embedded FB (MAF-FB) model and a Mixture of Experts (MOE) of attention mechanisms on top of MAF-FB, creating the MOE-FB model.
• Conducted large-scale comparative analyses using pancreatic tissue, Peripheral Blood Mononuclear Cells (PBMC), and Human Cell Atlas Bone Marrow datasets.
• Evaluated model performance using discrepancy metrics, automated cell-type classification, differential gene expression analysis, and cell-cell interaction inference.

Main Results:

• The proposed FB models, especially MOE-FB, consistently outperformed VAE, GAN, Gaussian Copula, and ACTIVA across all tested metrics.
• MOE-FB achieved high accuracy in cell-type classification (F1-score 0.90, precision 0.89, recall 0.92) on integrated pancreatic datasets.
• MOE-FB generated biologically relevant synthetic data, with inferred cell-cell interactions closely resembling original data (RMSE 0.65).

Conclusions:

• Flow-based generative models, particularly MOE-FB, offer a promising and effective approach for scRNAseq data analysis.
• MOE-FB demonstrates superior capability in handling scRNAseq data challenges, improving cell-type identification and synthetic data generation.
• These findings support the potential of FB models in advancing automated scRNAseq data analysis systems for biological discovery.