#1
Source: Nature Medicine | Authors: Zoller J, Kalz M, ... Kather JN, Friedrich MJ (DKFZ/Heidelberg/LMU Munich; multicenter) | Published: July 3, 2026 (online); PMID 42380678
Score 16/20PEER-REVIEWEDprospective-deploymentONC: primary
base 9 (Nature Medicine) + validation +3 (prospective 1-month silent-trial deployment on consecutive cases) + impact +2 (agent-assisted residents reached near-senior concordance; addresses uneven access to subspecialty tumor boards) + oncology +2 (primary) − hype 0
HemaGuide is a locally deployable, modular LLM agent that turns unstructured clinical documents into structured case representations, routes each case to a 'guideline', 'advanced', or 'molecular' mode, and grounds recommendations in disease-specific flowcharts plus a memory of >2,000 real-world tumor-board cases. In expert-blinded benchmarking on 45 high-complexity cases it improved concordance with tumor-board decisions across six foundation models. External validation on 555 independent cases from a second academic center gave 81.8% concordance across 47 entities, and a prospective 1-month silent trial on 64 consecutive, unselected cases achieved 82.8% concordance. Hallucinations occurred in 2 of 664 evaluated cases (0.3%), and the workflow ran under real-time conditions on commodity hardware (median 39 s for molecular classification). In a simulated practice study, agent-assisted residents reached near-senior concordance.
Why it matters: It is one of the more rigorously validated clinical-AI-agent papers to date — spanning internal benchmarking, external multi-site validation, AND a real prospective silent-trial arm — showing auditable, locally deployable decision support could help extend subspecialty tumor-board quality to centers that lack it.
Limitations: The silent trial was small (64 cases, single-center) and 'concordance with the tumor board' is not a patient-outcome endpoint; the simulated 'partially outperformed seniors' claim comes from a simulation, not real practice. Prospective outcome data and multi-site deployment are still needed.
Post angle: A heme tumor-board LLM agent with a real prospective silent-trial arm and a 0.3% hallucination rate — validation done right, but still concordance (not outcomes) and single-center prospectively.
#2
Source: Nature Medicine | Authors: Shen W, Moon I, ... Marbach D, Zitnik M (Harvard Medical School; Roche); pan-cancer | Published: July 3, 2026 (online); PMID 42399673
Score 15/20PEER-REVIEWEDexternal-retrospectiveONC: primary
base 9 (Nature Medicine) + validation +2 (external/multi-cohort retrospective, 16 clinical cohorts) + impact +1 (first-in-class generalizable pan-cancer ICI predictor with plausible stratification path) + oncology +2 (primary) + KOL +1 (surfaced by a curated handle) − hype 0 (abstract frames claims cautiously)
COMPASS is a pan-cancer foundation model that predicts immune-checkpoint-inhibitor (ICI) response from bulk tumor transcriptomes using a 'concept bottleneck' transformer, encoding expression through 44 biologically grounded immune concepts (immune cell states, tumor-microenvironment interactions, signaling pathways). Trained on 10,184 tumors across 33 cancer types, it outperformed 22 comparator methods across 16 clinical cohorts spanning seven cancers and six ICIs, improving accuracy by 8.5% and AUPRC by 15.7% on average, and generalized to cancer types and treatments not seen during fine-tuning. Patients COMPASS classified as responders had longer overall survival (HR 4.7, P<0.0001). Personalized response maps link expression to immune programs (e.g., TGFβ signaling, endothelial exclusion) associated with resistance.
Why it matters: Existing ICI biomarkers generalize poorly across tumor types and drugs; a single model that transfers across cancers and checkpoint inhibitors could inform indication selection, patient stratification, and trial design, while its concept bottleneck offers mechanistic, hypothesis-generating interpretability.
Limitations: Validation is entirely retrospective and computational — no prospective arm, no clinical deployment, and predictions come from bulk transcriptomes not routinely available at the point of care. The survival HR is from retrospective cohorts; this is a stratification/hypothesis tool, not a ready-to-use clinical test.
Post angle: A pan-cancer foundation model that predicts immunotherapy response across 33 cancers and 6 checkpoint inhibitors — impressive generalization, but retrospective only: frame as trial-design/stratification, not a bedside test.