Cancer stratification is moving beyond the tissue sample

The signal is leaving the tissue

Four preprints posted this week point the same way: the information that sorts cancer patients into treatment groups is increasingly found outside the tumor sample itself. In a preprint on glioblastoma, researchers report that panitumumab treatment significantly reduces surface EGFRvIII on extracellular vesicles derived from U87-EGFRvIII cells, and that treatment induces more significant proteomic changes in those vesicles than in the parental cells. The vesicles, the authors say, are potential circulating biomarkers for monitoring treatment efficacy and resistance. The claim is modest, potential rather than proven, but the direction is clear: the blood may carry the readout. m:…7407

A second preprint builds the same case from single cells. Integration of 244 single-cell RNA sequencing samples across 132 patients provides a robust atlas for cutaneous T cell lymphoma, according to the authors, and robust linear mixed-effects modeling of pseudobulk data identifies recurrent gene expression alterations in the disease. The molecular subtypes identified are predictive of clinical outcomes in independent datasets. That last point matters: the subtypes held up outside the cohort that produced them, which is the difference between a pattern and a tool. m:…7075

A third preprint adds splicing to the picture. Distinct splicing event patterns exist between short (under five years) and long (five years or more) survival cohorts in metastatic small intestine neuroendocrine tumors, the authors report, and CPSF2 expression is positively correlated with worse survival outcomes and increased vascular and perineural invasion. Alterations in core spliceosome components and splicing factors are associated with tumor plasticity and treatment resistance. In other words, how a tumor splices its RNA tracks with how aggressively it behaves. m:…8397

The fourth preprint closes the loop back to the microscope slide. CellBridge, a model built by multi-modal pretraining on 16 million cells with paired H&E and spatial profiles, outperforms leading pathology foundation models in macro F1 by up to 16 percentage points, according to its authors. Cell-ontology-aware conformal prediction allows for hierarchical label assignment based on morphological evidence sufficiency. The practical reading: molecular information can be transferred into morphological representations, so an ordinary pathology image can carry more than it used to. m:…3121

Why stratification is the prize

None of the four preprints claims a new treatment. What they offer is a better sort, and that is arguably the harder problem in oncology. A drug that works for a third of patients looks like a failure until the third is identified in advance. Vesicles that change more than their parent cells, subtypes that predict outcomes in independent data, a splicing factor tied to invasion, and slides that carry molecular signal all serve the same decision: who gets what, and how soon do you know whether it is working. This suggests the near-term value in cancer diagnostics sits in measurement and classification rather than in another compound. m:…7407 m:…7075

There is also a practical convergence here. Circulating vesicles can be sampled repeatedly, single-cell atlases give the reference map, splicing markers give a molecular readout tied to survival, and slide-based models work on material pathology labs already produce. Each approach lowers a different barrier: repeat sampling, reference data, prognostic signal, and existing workflow. A team building a stratification product does not need all four, but the fact that all four advanced in the same week suggests the field is competing on the same question from different directions. m:…7407 m:…3121

What to watch

All four results are preprints, not peer-reviewed findings, and the strongest claims carry the usual caveats. The vesicle biomarkers are described as potential, not validated. The lymphoma subtypes predicted outcomes in independent datasets, which is encouraging, but independent retrospective data is not a prospective trial. The splicing associations are correlations with survival and invasion, not proof that intervening on CPSF2 changes either. And CellBridge's margin over leading models comes from its own authors' benchmarks. The honest read is that the measurement layer for cancer stratification is improving fast, while the clinical proof that these measurements change outcomes is still ahead. m:…7407 m:…7075 m:…8397 m:…3121