Research Radar — 2026-10-10

Summary: InSTability assigns each spatial-transcriptomics spot a score for how much its cluster membership changes when only the random seed changes. The authors report results across seven methods and 18 benchmark samples, linking unstable regions to poorer reference-label agreement and inconsistent downstream differential-expression gene sets.

Why it matters: A single apparently clean tissue partition can conceal algorithm-dependent uncertainty. Localizing unstable spots makes spatial-domain and marker claims easier to audit before biological interpretation.

Why for Yiru: Useful as a sensitivity check before treating spatial domains or niche boundaries as evidence for cell–cell communication, cell-state transitions or regulatory programs. It complements rather than validates causal interpretation.

Summary: GRASS-MIL models cell-type neighborhoods across spatial platforms and expresses each sample-level prediction logit as an exact sum of signed neighborhood contributions. It distinguishes simulated tissues with identical global composition but different architecture. Reported AUROCs are 0.74 for ovarian treatment status, 0.57 for lung relapse and 0.62 for mortality.

Why it matters: Shared cell-type labels may enable studies across incompatible molecular panels, while direct prediction decomposition helps locate hypotheses in tissue. The lung results also expose the gap between interpretable patterns and strong outcome prediction.

Why for Yiru: Relevant to spatial niche analysis and cross-platform learning. Neighborhood contributions could guide follow-up perturbations, but are associations learned by a predictor rather than causal mechanisms.

Summary: Tangermeme provides composable sequence edits, prediction, attribution and design tools around trained genomic models. The article adds variable-length seqlet discovery and diagnostics for silent DeepLIFT/SHAP failures, and shows that BPNet and Beluga can use different motif patterns while nominally predicting MYC binding.

Why it matters: Interpretation code can fail even when a predictor looks accurate. Reusable operations and convergence checks make learned sequence rules easier to inspect and compare across models.

Why for Yiru: Useful for regulatory-genomics and AI-for-science workflows that move from predictive performance to candidate motifs, variant effects or experimental construct design.

Summary: PEtracer couples accumulating genetic marks to single-cell transcriptomes across 16 mouse embryos. The journal reports lineage trees for more than 1.4 million cells and roughly 75% division resolution, enabling comparisons of fate bias, restriction timing and progenitor contributions across replicates.

Why it matters: Recorded ancestry adds an experimental reference for testing developmental trajectories inferred from expression alone.

Why for Yiru: A strong dataset for single-cell lineage/state integration, uncertainty-aware trajectory benchmarking and predictive developmental modeling.