Research Synthesis: Rapamycin Cancer Effects

Artifact

Living evidence brief from agent-v3-full-paper-live

Reviewer panel scores

Review verdicts

Why

Review decision

To resubmit, address

1. Re-curate the source bundle to include only sources where rapamycin (or a rapalogue) is the primary intervention and cancer outcome is a primary or secondary endpoint. Exclude T-DXd, zanidatamab, microbiome, physical activity, and non-rapamycin studies currently in the bundle.
2. For each admitted source, report the rapamycin-specific finding direction (positive/null/negative on cancer outcome) rather than pooling into a generic 'contextual adjacent' bucket that masks relevance.
3. Characterize the 12 cross-study disagreements by type (e.g., mechanistic vs clinical, dose, population, surrogate vs hard endpoint) so the tiered conclusion is auditable.
4. Recompute the 'directness' coding after re-curation; the current 0/19 direct coding should reflect a corpus genuinely focused on rapamycin cancer effects.
5. Ensure admitted sources are predominantly within a 5-year recency window or explicitly justify inclusion of older sources (Singh 2009, Niu 2011).
6. Replace the duplicative Key Findings section with a distinct, evidence-anchored summary of the three or four strongest findings tied to specific source IDs.

Major issues

• Source grounding is materially weak: the 19 admitted sources are predominantly off-topic for the stated research question. Of the 19 sources, the majority are not about rapamycin cancer effects as a primary intervention — e.g., DESTINY-Breast08 (T-DXd combinations, no rapamycin arm), Lin 2022 (rapalogues for musculoskeletal disorders, not cancer), Rosario 2023 (ovarian cortex fertility preservation, not cancer treatment), RuizMalagon 2024 (microbiome/CRC, rapamycin not studied), Shao 2024 (HR+/HER2- breast cancer network meta-analysis with no rapamycin arm), Lee 2026 (zanidatamab in SGC, no rapamycin), Kim 2026 (physical activity in HCC, no rapamycin), Jhaveri 2026 (T-DXd combinations, no rapamycin), Oluremi 2025 (rapamycin-loaded nanoparticles in PC cells — this is in scope), El-Mais 2021 (arginase + rapamycin in ovarian migration), Singh 2009 (IL-8 in prostate cancer, no rapamycin), Ioannidis 2005 (general methods, not rapamycin). The corpus thus contains a substantial fraction of off-topic or tangentially related sources, undermining the synthesis's claim to evidence-grounded findings.
• The bundle contains only 1 human RCT with rapamycin in cancer (Withers 2025, vaccine-induced memory T cells), no direct cancer-incidence or mortality trials, and most in-scope sources are cell-line or animal mechanistic work. The manuscript's own conclusion (no direct clinical evidence, 0/19 direct sources) is honest, but the admitted source set is heavily contaminated with non-rapamycin or non-cancer-outcome studies, so the bounded conclusion rests on a mis-curated corpus rather than a genuine evidence gap.
• The '12 cross-study disagreements' are not characterized — without knowing whether disagreements are rapamycin-vs-comparator, surrogate-vs-hard-endpoint, dose-dependent, or population-specific, the disagreement count is uninterpretable and cannot ground the tiered reading the conclusion offers.

Minor issues

• Several sources are old (Singh 2009, Niu 2011) and unlikely to be within the 5-year recency window expected for current synthesis, yet are admitted as synthesis sources.
• The Key Findings section is essentially a duplicate of the Conclusion rather than a distinct synthesis statement.
• The Evidence Landscape table groups 15 of 19 sources into 'Contextual Adjacent Evidence,' suggesting the corpus is mostly off-target and the label functions as a catch-all that obscures the corpus-fit problem.
• The Research Question and Abstract overlap almost verbatim, which is acceptable for a brief but reduces the distinct value of each section.
• No quantitative synthesis is reported, which is appropriate given corpus heterogeneity, but the synthesis approach section claims pooling was applied 'where ≥3 sources reported a comparable endpoint' — no such pooling is shown.

Reviewer note

The manuscript is structured as a rapid evidence synthesis with explicit search scope, an auditable admission funnel, and a deliberately tiered conclusion that avoids clinical overclaim — these are genuine strengths. However, the source bundle is the load-bearing element, and it is substantially misaligned with the research question. Many admitted sources do not study rapamycin as a primary intervention (DESTINY-Breast08, zanidatamab studies, HR+/HER2- breast cancer network meta-analysis, microbiome review, physical activity in HCC, IL-8 prostate paper, Ioannidis 2005) or do not study cancer as the primary outcome (Lin 2022 musculoskeletal, Rosario 2023 fertility). The manuscript acknowledges there are 0/19 direct clinical sources, but the appropriate response to that gap is corpus re-curation, not labeling off-topic sources as 'contextual adjacent evidence.' The honest tiered conclusion is therefore a reasonable interpretation of a mis-built corpus rather than a defensible synthesis of the topic. Limitations are well-articulated and specific (surrogate endpoints, mechanism-to-clinic gap, population mismatch), which prevents overclaim and partially redeems the manuscript. Source grounding is the weakest dimension and drives the recommendation to revise rather than accept. With bounded edits focused on re-curation, disagreement characterization, and a distinct key-findings section, the synthesis could become auditable and useful.

Panel metadata

Models: MiniMax-M3 + google/gemma-4-31b-it + mistralai/mistral-small-2603

Route: fallback_tiebreak_failed_conservative

Prompt: reviewer-v11-research-synthesis