Science, amplified.

Rare glioblastoma variants are usually treated as ordinary glioblastoma because prospective variant-specific cohorts are scarce. That default is practical, but it risks erasing biology that should matter for diagnosis and trial design. We reanalyzed public cBioPortal clinical sequencing data from the MSK 2019 glioma cohort, comparing 18 gliosarcoma samples with 539 conventional glioblastoma samples, and used TCGA PanCancer Atlas and CPTAC 2021 glioblastoma cohorts as external GBM baselines. Gene-level Fisher exact tests, burden comparisons, and a primary treatment-naive sensitivity analysis were used to distinguish recurrent alteration patterns from treatment and sampling artifacts. In the full MSK comparison, gliosarcoma was enriched for TP53 alteration (14/18 versus 183/539, q = 0.0067) and RB1 alteration (8/18 versus 78/539, q = 0.0258), and depleted for EGFR alteration (1/18 versus 219/539, q = 0.0258). PTEN alteration was directionally higher but did not survive FDR correction. In primary treatment-naive tumors, the same directions persisted, but gliosarcoma sample size fell to 11 and false-discovery-adjusted significance was lost. Tumor mutation burden and fraction genome altered were not significantly different. These data support gliosarcoma as a pathway-distinct glioblastoma variant, not merely a globally more unstable GBM subset. The finding is small-cohort evidence, but it is strong enough to argue that gliosarcoma should be stratified rather than silently pooled in molecular studies and clinical trials.

Vineet Reddy

Relapsed pediatric B-cell acute lymphoblastic leukemia remains clinically hard because recurrence is not only a return of leukemia burden, but a selected biological state. We reanalyzed 49 paired diagnosis-relapse marrow expression profiles from GEO series GSE28460 and challenged the resulting relapse program in the independent diagnostic cohort GSE7440. Across 21,755 annotated genes in GSE28460, 505 genes changed at relapse at FDR < 0.10 with an absolute mean paired log2 fold-change above 0.35. The relapse program was dominated by cell-cycle and DNA-replication biology, with cell division, sister chromatid cohesion, mitotic nuclear division, DNA replication, and G1/S transition as the strongest enriched processes. Curated signature analysis recovered increased cell-cycle scoring and decreased glucocorticoid-response scoring at relapse. In 59 GSE7440 diagnosis samples with outcome labels, the top relapse-up minus relapse-down program did not distinguish later relapse from complete continuous remission (AUC 0.48; Mann-Whitney p = 0.80). This negative external test sharpens the interpretation: the signal is a relapse-state program, not a baseline bulk prognostic signature.

Vineet Reddy

The AI infrastructure boom is young, but it is not too young to measure its first labor-market channel. This paper studies eight high-confidence U.S. counties with AI-specific or AI-boom hyperscaler data-center milestones in 2024 and links them to BLS Quarterly Census of Employment and Wages county-industry records through 2025 Q3. The main result is clear: the first measurable local labor signal from AI data centers is in power infrastructure work, not permanent data-center operation. In NAICS 237130, power and communication line construction, average weekly wages rise 18.6 percent relative to matched controls over event quarters 0-4, and all four contributing treated counties have positive wage effects. The result survives the main identification checks: the leave-one-county-out minimum is 13.8 percent, the estimate remains positive under treatment timing shifts of plus or minus two quarters, augmented matching on pre-boom power-line levels and wage slopes leaves a minimum effect of 11.9 percent, six placebo sectors have wage effects no larger than 2.1 percent, and a within-match randomization test gives a one-sided p=0.002. Payroll in the same industry rises 23.8 percent on average but is more dispersed. By contrast, the combined skilled-trades index is not positive, and directly AI-specific data-processing employment has only one disclosed treated-county observation. AI data centers are already visible in local labor data, but the visible channel is concentrated in grid-adjacent construction trades before permanent operating jobs appear in public employment records.

George Pickett

AI systems are increasingly used to synthesize biological literature, but useful synthesis should be judged by whether its claims can be operationalized and checked against experimental data. We evaluated a claim-level protocol for testing AI-generated biological synthesis against public cancer datasets. The system generated twelve structured claims spanning two validation domains: CRISPR gene dependency in DepMap 24Q2 and pharmacogenomic drug sensitivity in GDSC release 8.5. Each claim was mapped to an explicit biomarker or lineage cohort, an experimental response variable, and a direction-aware concordance statistic. Six of seven dependency claims were supported after multiple-testing correction, including WRN dependency in MSI-high models, PAX8 dependency in ovary/fallopian-tube models, SOX10/MITF dependency in skin melanoma models, and BRAF/KRAS oncogene dependency in hotspot-mutant contexts. Drug-response synthesis was less reliable: BRAF mutation strongly predicted RAF inhibitor sensitivity and KRAS mutation partially predicted MEK inhibitor sensitivity, but ERBB2 mutation alone did not recover ERBB2 inhibitor sensitivity and BRCA1/2 hotspot mutation alone did not recover PARP inhibitor sensitivity. One EGFR claim was not directly verifiable because the positive cohort was too small. These results show that AI-generated synthesis can recover strong, established preclinical relationships, but that plausible biomedical language often hides biomarker mismatch, sparse cohorts, and missing biological context. The appropriate role is exploratory claim generation followed by explicit public-data validation.

Matthew Pacheco