ARTFEED — Contemporary Art Intelligence

CellxPert: Single-Cell Foundation Model for In-Silico Perturbation

ai-technology · 2026-05-06

CellxPert represents a versatile multimodal foundation model that integrates single-cell and spatial multi-omics into a unified representation space. It encodes measurements from transcriptomics (scRNA-seq), chromatin accessibility (ATAC-seq), and surface proteomics (CITE-seq), while also incorporating MERFISH and imaging mass cytometry data as spatial-visual layers in 2D or 3D. This model facilitates four primary downstream tasks: annotating cell types across 154 overlapping identities (the largest label space to date), enabling efficient fine-tuning through Low Rank Adaptation (LoRA), predicting genome-wide transcriptomic responses to in-silico perturbations (ISP), and allowing seamless integration of multi-omic data across various assays and platforms. This research is available on arXiv with ID 2605.00930.

Key facts

  • CellxPert is a multimodal foundation model for single-cell and spatial multi-omics.
  • It encodes scRNA-seq, ATAC-seq, CITE-seq, MERFISH, and imaging mass-cytometry data.
  • Supports cell-type annotation across 154 overlapping identities.
  • Uses Low Rank Adaptation (LoRA) for efficient fine-tuning.
  • Enables genome-wide transcriptomic response prediction to in-silico perturbations.
  • Facilitates multi-omic integration across various assays and platforms.
  • Published on arXiv with ID 2605.00930.
  • Addresses limitations of current single-cell foundation models.

Entities

Institutions

  • arXiv

Sources