#28 · Tools & infrastructure

Benchmarks and diagnostics for inferring cell-level structure from single-cell…

The dominant thread is computational inference of cell-level and sample-level biology from single-cell and spatial transcriptomes, with increasing emphasis on stating when such inference fails. The GRN study isolates seven pathologies, including dropout, confounders and feedback, to test when causal versus correlational methods hold, while SCMBench benchmarks 23 domain-specific and foundation models for single-cell multi-omics integration. Spatially aware inference appears in DEGAS, which transfers patient-level disease attributes to individual cells with spatial smoothing to flag high-risk cells. Trajectory- and sample-aware approaches include MUSTARD for multi-sample scRNA-seq and PyFgsea's rolling-window enrichment along trajectories. Interpretability is addressed by scMarkerGene's Contribution Score matrix. Other items, including geneSync for RNA-seq symbol harmonization and nested-WGCNA on bulk data, do not share this single-cell focus. What is new is controlled stress-testing of inference under known pathologies, combined with spatially and sample-constrained cell-level assignment.