CellxPert introduces MCMC steering for multi-omic perturbation simulation

Per the arXiv submission, CellxPert (CellxPert) is a scalable multimodal foundation model that unifies single-cell and spatial multi-omics into a common representation space (arXiv:2605.00930). The paper states that CellxPert jointly encodes scRNA-seq, ATAC-seq, and CITE-seq measurements and directly incorporates MERFISH and imaging mass-cytometry as 2D or 3D spatial-visual layers (arXiv:2605.00930). The authors report four primary downstream capabilities: cell-type annotation across a 154-label ontology, LoRA-based efficient fine-tuning, genome-wide transcriptomic response prediction to in-silico perturbations (ISP), and multi-omic integration (arXiv:2605.00930). The submission was posted on 30 Apr 2026 and associated with the ICLR Machine Learning for Genomics Explorations Workshop 2026 (arXiv:2605.00930).

Per the paper, the central methodological novelty is an inference-time Metropolis-Hastings Markov-chain Monte Carlo (MCMC) sampler that uses the model's masked conditional distributions as the proposal kernel to transition to new transcriptomic states conditioned on perturbed genes (arXiv:2605.00930). The authors contrast this approach with prior token-based perturbation heuristics that delete or reorder tokenized gene expression ranks, arguing the MCMC trajectories avoid abrupt out-of-distribution artifacts and yield biologically interpretable state trajectories (arXiv:2605.00930). The model is presented as LoRA-compatible for parameter-efficient fine-tuning and evaluated on multiple single-cell benchmarks (arXiv:2605.00930).

Editorial analysis: Foundation models for single-cell genomics increasingly aim to unify modalities and support in-silico experiments. Observed patterns in similar research show that mechanics for generating perturbation counterfactuals matter: token-edit heuristics can produce artefactual outputs, while probabilistic, inference-time methods often produce smoother, more plausible trajectories at higher compute cost. For practitioners, this implies a tradeoff between inference complexity and biological plausibility when choosing perturbation simulators.

Editorial analysis: A model that can jointly represent transcriptomics, chromatin, surface proteomics, and spatial modalities is technically significant for multi-omic integration research. Demonstrated gains on PBMC68K, Replogle Perturb-seq, Systema, and BMMC benchmarks (arXiv:2605.00930) suggest the approach may be of interest to computational biologists working on perturbation prediction and single-cell annotation. However, claims about broader utility and generalization beyond the reported benchmarks require independent replication and task-specific evaluation.

Editorial analysis: Observers should watch for a full methods and code release, independent benchmarks reproducing the reported gains, runtime and compute cost comparisons for the MCMC steering, and adoption of LoRA checkpoints that enable community fine-tuning. The paper has an OpenReview entry linked to the submission, which may host reviewer discussion if the workshop posting proceeds (arXiv:2605.00930; OpenReview entry id available).