Research Radar — 2026-10-11

Summary: CohortFM uses flow matching to reconstruct continuous molecular trajectories from sparse longitudinal patient samples, then groups them into cohort-level progression patterns. The authors test four infectious-disease settings and separately evaluate interpolation and forward prediction.

Why it matters: Separating molecular progression from calendar visits may make sparse patient trajectories easier to compare. Explicitly distinguishing interpolation from forward prediction avoids treating fitted paths as forecasting evidence.

Why for Yiru: Relevant to flow-based molecular-state models and sparse observation settings. The design suggests tests for progression alignment and held-out trajectory prediction, without establishing causal perturbation dynamics.

Summary: A point-process study of eight glioblastoma specimens links CosMx-derived tumor-proximity phenotypes to transcriptional states. Although transcript-level features outperform pathway scores, performance falls when an entire patient is held out, and harmonization does not fix transfer.

Why it matters: The reported 2.43 million cells still span only eight specimens. Patient-specific tissue architecture can limit transfer even when cell-level training sets appear enormous.

Why for Yiru: Directly relevant to transferring learned molecular relationships between dissociated and spatial data, and to separating statistical neighborhood structure from causal cell–cell signaling.

Summary: RNA fingerprinting maps new single-cell transcriptional responses onto genetic, chemical and cytokine perturbation dictionaries, assigning cells to one or more candidate perturbations. The Cell abstract reports genome-scale and combinatorial benchmarks plus applications to p53 stress responses, drug-dose mechanisms and influenza-associated B-cell states.

Why it matters: Perturbation atlases can support uncertainty-aware candidate explanations for new cell states. The earlier preprint makes assignment coverage and ambiguous credible sets explicit rather than forcing a unique label on every cell.

Why for Yiru: Highly relevant to genetic and chemical perturbation representations and combination inference. It also supplies a baseline for asking how far perturbation signatures transfer across cellular contexts.