The evidence layer for scientific AI — claims with supporting evidence, scope, and limitations.
High-LET radiation methylation findings are robust within their limited model scope, but the lead author's Genegoggle cofounder/shareholder role demands skepticism toward broad generalization.
The strongest causal evidence for RNA-mediated inheritance may overestimate effect sizes because injection experiments use supraphysiological RNA doses — a bias the field has largely overlooked.
Loss of Ncoa1 dismantles the basal Ucp1 program in female mouse subcutaneous fat via an NCOA1–GATA3 complex, though chromatin accessibility data suggest the true gate may be epigenetic and upstream.
Flagellar hook stiffness ratio (A_hook/A ≈ 0.02 vs 0.14) causally distinguishes wrap-capable from non-wrapping species, a parameter that could predict confined-niche colonization in polar-flagellated pathogens.
First structural demonstration of full-envelope conduits carrying chromosomal DNA up to 1.1 Mbp between clinical strains, but the mechanism of conduit formation and DNA traversal remains unknown.
A congenic cured/re-infected design shows the mycovirus AfuPmV-1M enhances lung virulence and in vivo ribavirin improves mouse survival, though evidence is limited to one isolate and murine models.
Synthesis of epidemiological and animal evidence that H5N1 can reach mammalian gut tissue via contaminated raw milk, with human consumption risk judged low but theoretically possible.
AlphaFold2 achieved median 2.2 Å accuracy across 97 CASP14 targets and enabled molecular replacement for 89 of 97, with strength concentrated in single-chain domains rather than multi-chain systems.
A plausible but untested mechanism — user vulnerability × engagement × sycophancy/hallucination — rests only on case reports, with the true attributable fraction of harm genuinely unknown.
A population-genetic framework maps training on model output to drift and model merging to recombination, with pre-registered falsifiers but simulation-only evidence for key mechanisms.
AI-designed proteorhodopsins absorbing at ~410 nm boosted C. necator growth, but proton pumping was never directly measured — only inferred from retinal bleaching.
First functional genetic evidence (RNAi) that a nicotinic acetylcholine receptor subunit promotes wakefulness in a cnidarian, though the preprint has small samples and an unresolved species-identity confound.
A VAE decoder renders intuitive scene hypotheses for 11 navigation regions during virtual driving, but validation is largely qualitative with two author-participants among n=6.
Loneliness predicted cognitive decline and Alzheimer's risk in 823 older adults independent of depression, yet autopsy data showing no neuropathology link suggest effects may act on cognitive reserve rather than disease.
Mind-blanking trials co-vary with slower reactions and reduced P3b/EEG connectivity in 62 participants, but a decisive probe-timing sensitivity test on deposited data remains outstanding.
Only one high-quality MMR–ASD study exists (case-control OR 0.86, 95% CI 0.68–1.09); no pooled estimate is possible from available inputs, and inflating the evidence base would itself be a scientific error.
Rutter's 2005 review concluded true ASD incidence is likely 30–60 per 10,000 with most of the rise from improved ascertainment and broadened diagnostic criteria, and no epidemiological support for MMR causation.
Roughly four in five of 50 studies report a significant positive association with depressive symptoms, but the synthesis is entirely observational with unvalidated exposure measures.
Meta-analysis of 23 longitudinal studies (181,006 participants) finds ~29–32% increased risk, but funnel asymmetry and heterogeneity imply the true effect may be closer to ~18–19%.
A NEJM synthesis covering screening algorithms and lactation-compatible treatment, though its strongest prevention claim (antenatal treatment) rests on one small observational study with implausibly large effect sizes.
The apparent contradiction resolves as a time-scale artifact: acute caffeine transiently raises blood pressure while habitual intake coincides with lower chronic mortality risk in large cohorts.
Animal mechanisms of inflammatory and oxidative harm are well supported, but the causal human dose–response link remains nearly empty — the field's biggest blindspot.
Particle size switches the harm mode — microparticles drive immune/dysbiosis responses while nanoparticles drive cell death — so predictions averaging across particle classes are likely miscalibrated.
Equilibrium thermodynamics elegantly explains why cytosine and thymine are absent from meteorites while G, A, and U persist, though the model overproduces bases by 3–4 orders of magnitude.
First ancient DNA from the region shows ~1300 years of genetic continuity with Afanasievo-related ancestry arriving via already-admixed Tibetan plateau groups rather than direct Steppe contact.
Thirty years of consistent rodent wound-healing data collide with a translational wall — fewer than 30 humans ever studied, zero RCTs, and entirely unregulated grey-market supply.
~150 min/week of activity prevents weight gain while 250–300 min/week plus resistance training is needed for meaningful loss and organ-level repair — minimum and optimal doses differ roughly twofold.
A narrative review of helminth, maggot, and leech therapy relies almost entirely on preclinical data, with the most defensible path — recombinant parasite-derived immunomodulators — still lacking any human trials.
Cross-species evidence (zebrafish, mouse, cell) for stage-specific myogenic peptide activity, but no isonitrogenous protein control, corporate conflicts of interest, and an untested proposed mechanism.
An unreviewed preprint implicates the SKN-1/Nrf2 pathway in a common preservative's obesogenic effect at doses 250-fold below the FDA limit, but non-monotonic dose effects and missing data are classic worm-assay red flags.
A Neural Cellular Automata model argues aging emerges from lost morphogenetic goals rather than damage itself — conceptually bold, but proof-of-concept only with no empirical validation.
Low GPD1L expression uncouples the glycerol-3-phosphate shuttle in human cells and CRISPRa activation shortens the segmentation clock by ~40–50 minutes, though evidence is gain-of-function only.
Crack-opening pods with wax diffusion timers held seeds in warmer, drier microclimates and accelerated emergence in trials — but only at one irrigated field site with the causal mechanism not experimentally isolated.
A stop-gain variant appeared in 3 of 101 top-elite athletes and zero of 1,899 comparators, but with corrected p=0.069 and no physiological data, chance cannot be excluded.
BGPT is available through MCP for Claude, Cursor, and other compatible AI tools — bringing evidence-qualified scientific claims directly into your workflow.
Updated: October 24, 2026Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.
Explore by Goal
Ask a Follow-Up
Trending