Mojtaba Bahrami

The paper benchmarks gene expression reconstruction from single-cell latent representations, finding autoencoders perform best at low dimensionality and foundation models depend on decoder architecture.

Cellpin imputes and denoises spatial transcriptomes using a variational autoencoder trained on single-cell RNA-seq, outperforming existing methods without requiring cross-modality alignment.

Pertpy provides a modular framework for analyzing large-scale perturbation single-cell experiments, offering tools for metadata annotation and perturbation distance calculations.

scConcept improves single-cell representations by using contrastive learning instead of reconstruction, outperforming existing methods in multiple tasks.

This study benchmarks self-supervised learning techniques in single-cell genomics, finding masked autoencoders outperform contrastive methods in cell type prediction and data integration.

Pretrained on 57 million dissociated and 53 million spatial human and mouse cells, a 49.3M-parameter transformer predicts niche, region, neighborhood composition and density, and carries spatial labels onto scRNA-seq.

The paper evaluates distance metrics for single-cell RNA-seq data and recommends mean squared error for comparing populations and evaluating gene expression models.

This paper introduces scGAN, a method to reduce batch effects in scRNA-seq data by using a generative adversarial network, and shows it improves clustering and gene identification in psychiatric disorders.