Bulk transcriptomes no longer need a model per cohort: PRTS reads a labelled cohort as context and edges TabPFN v3.5 on mean rank correlation
Synopsis
PRTS is a 220-million-parameter model that reads a labelled cohort together with new samples and returns binary, multiclass, continuous or censored-survival predictions in one forward pass, with no per-cohort training. It reached the highest mean rank correlation on 175 held-out tasks and 44 TCGA tasks, but the lead is small and comes mainly from prediction within a cohort; across cohorts and at unseen sites it is level with the strongest tabular foundation model, TabPFN v3.5, while running more than twice as fast.
Interpretation
Predicting clinical endpoints from bulk transcriptomes no longer requires fitting a separate model for each cohort: one set of weights treats a labelled cohort as context and returns a predictive distribution for new samples directly. The prior route gave every cohort and endpoint its own model, fitted from thousands of genes and about a hundred samples, and such models often failed when moved to another cohort, platform or laboratory; transcriptomic foundation models learn gene representations, but a new endpoint still needs its own labelled samples. On 175 held-out tasks the mean rank correlation was 0.376, 0.013 above the pre-specified strongest comparator TabPFN v3.5 (95% CI 0.004 to 0.022); on 44 TCGA tasks it was 0.385, 0.019 above (95% CI 0.011 to 0.028).
The model can read different cohorts because genes enter only as random codes, so it never sees gene names, study identifiers or label names. Random codes let the same weights accept whatever gene set a cohort measures on microarrays or by RNA-seq, and attention along genes through inducing tokens keeps cost linear in the number of genes. Pretraining used 6.04 million tasks generated from 581,503 public RNA-seq profiles in three passes of increasing width, the last over more than 10,000 features; at prediction time four forward passes are averaged, each with fresh gene codes.
The lead is not uniform: it holds within a cohort and disappears across cohorts and at unseen sites. On 113 within-cohort tasks it was 0.018 above TabPFN v3.5 (95% CI 0.005 to 0.029); across cohorts with no target labels it reached 0.341, a difference of +0.008 (95% CI −0.007 to 0.022). On TCGA tasks whose query patients came from sites absent from the context, the margin over TabPFN v3.5 fell to +0.007 (95% CI −0.007 to 0.021), and PRTS's own mean changed by −0.023 (95% CI −0.043 to −0.008).
Perspective
The demonstrated scope is retrospective public cohorts: 175 tasks from 174 held-out RealBulk cohorts and 44 tasks from 13 TCGA projects never used in development, with endpoints from clinical records, survival and DNA and none derived from RNA. The authors propose next steps of ablating the prior, gene codes and distributional heads, adding larger and more diverse corpora and other molecular layers, and running prospective clinical studies.
The authors state that no design ablation was run, so the lead cannot be apportioned between the prior, the gene codes and the distributional heads; when no development cohort carried the target concept, PRTS was level with TabPFN v3.5 and GeneICL. Public cohorts can share patients without declaring it, and a linkage audit tied 34 held-out cohorts to development cohorts, though the main comparisons held on the 137 tasks that touch none of them. TCGA random folds mix new patients with known sites, and no method received a cohort or site index; the authors also note that newer tabular foundation models are likely to close the remaining gap and that the cohorts are retrospective.
