BGPT Evidence layer for scientific AI — claims with supporting evidence, scope, and limitations.
Updated October 12, 2026
Every claim gets a one-sentence evidence-qualified summary. Quality dimensions (novelty, rigor, generality, usefulness, reproducibility, explanatory depth) are rated 0–100% with stated limitations — uncertainty is never hidden.
Available via MCP for Claude, Cursor, and other compatible AI tools.
Bacteroidota methyltransferase BF2170 methylates bezafibrate into a prodrug, raising gut-wall permeability ~20-fold in Caco-2 assays and increasing plasma drug levels in gnotobiotic mice — suggesting microbiome composition could silently alter drug exposure.
Novelty 90% • Quality 90% • Limits: mono-colonization mice, no AUC, in vitro-only substrate screens.One engineered conjugative plasmid carrying entry-exclusion genes blocked >99.9% of in vitro acquisition of IncM/L/C/A resistance plasmids and prevented gut establishment in co-housed mice — a population-level “vaccine” concept proven only in E. coli with small mouse groups.
Novelty 90% • Quality 70% • Limits: n=3 mice, no statistical testing, single strain, patent conflict declared.A multi-layer microfluidics-genetics study shows Caballeronia insecticola wraps its flagellum to traverse micrometer constrictions, with hook stiffness causally controlling wrapping and infection success — a quantitative rule that may extend to other polar-flagellated pathogens.
Novelty 90% • Quality 90% • Limits: primary evidence from one insect-symbiont species; rectangular channels.First structural evidence of full-envelope bridges carrying double-stranded DNA, enabling chromosomal exchange up to 1.1 Mbp (~27% of the genome) between clinical strains — far beyond known transformation or phage transfer, though the conduit mechanism remains unexplained.
Novelty 100% • Quality 70% • Limits: one strain pair tested functionally, machinery unidentified.IRSp53 drives Arc capsid assembly and extracellular-vesicle release from dendritic filopodia, and transferred Arc mRNA reduces AMPA receptors in neighboring neurons — a proposed new form of synaptic plasticity shown only in cultured rodent neurons so far.
Novelty 90% • Quality 80% • Limits: no in vivo functional validation; cultured neurons only.Only one high-quality MMR–autism study was available: a UK matched case-control study (1,294 cases, 4,469 controls) reporting an adjusted odds ratio of 0.86 (95% CI 0.68–1.09) — no association; inflating the evidence base with non-MMR studies would itself be a scientific error.
Honest single-study answer; no pooled estimate possible from supplied records.Neural cellular automata evolved in silico show aging emerges as loss of goal-directed morphogenesis after development rather than damage accumulation — conceptually bold but proof-of-concept only, with no empirical validation and no public code.
Novelty 80% • Quality 70% • Limits: minimal 16×16 model, single target pattern, industry COI disclosed.Enformer applied to 34 hominin and ape haplotypes predicted 37,592 lineage-specific regulatory elements, validating 7 of 9 in luciferase assays with R = 0.81, including a causal single-variant SRF-motif enhancer — though most predictions remain mechanistically unexplained.
Novelty 80% • Quality 70% • Limits: cis-only scope, small validation set with circularity risk.First functional genetic evidence (RNAi) that a nicotinic acetylcholine receptor subunit (chrnal-E) promotes wakefulness in a cnidarian, with cholinergic drugs bidirectionally altering pacemaker firing — but small samples, no rescue experiment, and a preprint-not-peer-reviewed status temper certainty.
Novelty 90% • Quality 80% • Limits: n=3 sleep-deprivation arm, species-identity confound, preprint.Single-nucleus ATAC-seq of haploid mouse testis nuclei yielded >25,000 genome-wide meiotic crossovers across 3,842 gametes and confirmed elevated crossovers in Fancm-knockout mice — a scalable alternative to pedigree-based recombination mapping with subtelomeric bias and batch-effect caveats.
Novelty 90% • Quality 90% • Limits: n=3 per genotype; lower per-cell genetic-distance recovery than prior methods.A leakage-controlled benchmark (CodonBench) shows reported codon-LM gains collapse or reverse under gene-held-out splits, and a memorization canary (AUC 0.739) beats every neural model — a transferable audit template for thin-signal machine-learning claims.
Novelty 90% • Quality 80% • Limits: frozen embeddings only; human synonymous-variant scope.Animal and cell studies consistently show inflammatory signaling and oxidative stress, and microplastics amplified liver cadmium accumulation in zebrafish by 46–184% — but the causal human link remains nearly empty, making dose–response translation the field's biggest blind spot.
Quality 80% (framework paper) • Limits: industry funding from plastics-sector bodies; purely theoretical.A CDC-authored minireview synthesizes evidence that influenza A, particularly H5N1, can reach mammalian gut tissue and that ingesting contaminated material (e.g., raw milk) can establish infection — though judged a low but theoretically possible human risk, with mechanisms of intestinal tropism unknown.
Quality 80% • Limits: single-day ferret snapshot, no per-author COI statements, indirect human evidence.Equilibrium modeling of nucleobase synthesis in planetesimals explains the meteoritic absence of cytosine (rapid deamination) and thymine (oxidation by peroxide) while guanine, adenine, and uracil persist — an elegant framework with 3–4 orders-of-magnitude overproduction and an untested peroxide-selectivity assumption.
Quality 80% • Limits: no code release, proprietary ChemApp, kinetic effects ignored.A formal framework maps multigenerational AI model populations onto population genetics — training on model output as drift, real data as immigration, weight averaging as blending inheritance, and merging as recombination — with pre-registered falsifiers and reproducible artifacts but simulation-scale evidence and single-author, AI-co-produced writing.
Novelty 80% • Quality 70% • Limits: 3 seeds, LLM evidence in only two model families.A review of 50 studies finds roughly four in five report a significant positive association between social media use and adolescent depressive symptoms, with anxiety evidence weaker and sometimes U-shaped — entirely observational, narrative-only, and lacking per-study effect estimates.
Novelty 50% • Quality 60% • Limits: no meta-analysis, unvalidated exposure measures, no registered protocol.Each intermediate link is supported — fiber loss suppresses butyrate producers, SCFA loss weakens gut barrier and immune homeostasis, enabling systemic inflammation — but the chain has never been tested with ultra-processed foods themselves, resting on diet-pattern epidemiology plus borrowed mechanism.
Mechanistically strong, attributionally untested; UPF-specific intervention trials absent.Bats phase-lock echolocation calls to low-noise dips of 4-Hz noise in lab and wild settings, with frontal auditory field inactivation disrupting the timing — an energy-free sensing strategy demonstrated in one species and one noise rate, with modest sample sizes and post hoc modeling.
Novelty 90% • Quality 70% • Limits: n=14 lab, 8 inactivation; no data/code links.In 21,851 US men, higher PM2.5 was associated with increased sperm DNA fragmentation (estimate 0.73, p<0.0001), but the age effect (+14.4 pp at ≥50) dwarfs the pollution effect, and cross-sectional design, ZIP-level exposure proxies, and missing smoking/BMI data limit causal inference.
Quality 70% • Limits: duplicated effect estimates reported, contradictory SES findings, proprietary data.Perceived isolation predicted cognitive decline and Alzheimer's risk over 65 months in 823 dementia-free adults, surviving adjustment for depression — yet autopsy data showed loneliness unrelated to neuropathology, implying effects on cognitive reserve rather than disease itself.
Quality 70% • Limits: narrative review, small extremes-of-distribution samples, causal language beyond data.In the asexual bdelloid rotifer Adineta vaga, chromosome breaks are progressively repaired across multiple generations via homolog-directed reconstruction (BIHER) — a genuinely new genome-maintenance concept, but demonstrated in one clone with holocentricity inferred rather than shown.
Novelty 90% • Quality 80% • Limits: single deep-sequenced clone, generations untracked under desiccation.LDDM unifies docking, covalent docking, fragment linking, and de novo design in one generative framework with wet-lab validation: 79.5% top-1 docking success, 3.5-fold cathepsin S potency improvement, and micromolar hits from as few as 57 synthesized compounds — though potencies remain hit-level and pose prediction is weakly validated.
Novelty 80% • Quality 80% • Limits: 3 crystal structures for pose validation, no blinded baselines.A Navigation Guide review of 25 studies rates prenatal pyrethroid exposure as having “sufficient evidence” of adverse neurodevelopment at moderate quality, with thyroid disruption the strongest mechanism — while noting that two cohorts reported beneficial metabolite associations, an asymmetry that inflates apparent consistency.
Quality 80% • Limits: spot-urine biomarkers, infeasible meta-analysis, judgment-dependent ratings.Rutter's 2005 review concludes true ASD incidence is likely 30–60 per 10,000 (vs 4 in 1966), mostly explained by better ascertainment and broadened diagnostic concepts, with no epidemiological support for MMR or thimerosal causation — using exposure-withdrawal natural experiments as falsification tests.
Quality 80% • Limits: registry/ecological data, narrative method, residual true rise unexplained mechanistically.Thirty-three Sicilian centenarians carried ~50-fold higher neutralizing titers against 1918 H1N1 than septuagenarians yet showed zero SARS-CoV-2 cross-reactivity — a valuable negative result ruling out antibody cross-protection, with only 8 infections observed and no T-cell data.
Novelty 60% • Quality 60% • Limits: N=33, essentially unreplicable cohort, raw data on request only.A C. elegans preprint links a common preservative to fat gain and shortened lifespan via SKN-1/Nrf2 — hypothesis-generating for a GRAS preservative but unreviewed, worm-only, with unexplained non-monotonic dosing and classic worm-assay confounds unaddressed.
Novelty 50% • Quality 50% • Limits: key datasets “not shown,” no lifespan statistics, no n reporting.Combining dietary energy reduction with activity is the strongest-supported approach: ≥150 min/week prevents weight gain, 250–300 min/week supports clinically significant loss, and ~10% weight loss substantially improves liver histology in fatty liver disease — prevention and repair doses differ roughly twofold.
Strong cohort/RCT support; individual levers are dose- and context-dependent.The gut-disruption claim is mechanistically plausible with animal and in vitro support, but particle size switches the harm mode (microparticles: immune/dysbiosis; nanoparticles: barrier damage/apoptosis) and human causal data remain absent.
Quality 50% (narrative review, full text unavailable) • Reproducibility 30%.HIIT alters circulating exercise factors (PGC-1α +63% vs +33% vs no HIIT, p=0.047; reduced IL-6; increased BDNF/irisin), but no supplied evidence isolates extracellular vesicles — any “exosomal marker of HIIT success” claim rests on indirect extrapolation until EV-isolation studies are done.
Honest negative answer; closest signals are non-EV biomarkers.Two small naturalistic studies (n=8 basketball, n=21 soccer) found winning fans' testosterone rose and losing fans' fell, reversing the normal daily decline in winners — foundational for vicarious-hormone theory but with tiny all-male samples and no control for alcohol or timing.
Novelty 80% • Quality 60% • Limits: no individual-level data published; selective attrition.A FANCI stop-gain variant appeared in 3 of 101 Polish top-elite athletes and none of 1,899 comparators, but Bonferroni-adjusted p=0.069 misses significance and no carrier physiology was measured — the authors themselves concede chance cannot be excluded.
Quality 40% • Generality 20% • Limits: n=3 carriers, no mechanism tested, winner's-curse risk.Dual agonists outperform GLP-1 RAs on HbA1c (up to 2.4%) and weight (up to 22%) with similar GI tolerability, corroborated by an independent 21-trial network meta-analysis — but the reviewed synthesis itself is methodologically substandard, with conflicting search descriptions and no PRISMA flow or risk-of-bias appraisal.
Novelty 40% • Quality 40% • Limits: no registered protocol, placeholder co-author, narrative-only synthesis.A rigorous meta-analysis of 23 longitudinal studies (181,006 participants) found a 29% increased CHD risk (RR 1.29, 95% CI 1.04–1.59) and 32% increased stroke risk (RR 1.32, 95% CI 1.04–1.68) — solid observational evidence, though funnel asymmetry suggests the true effect may sit nearer ~18–19% than the headline 30%.
Quality 80% • Limits: unresolvable heterogeneity, variable confounder adjustment, high-income settings only.A VAE decoder renders “preferred visual experiences” for 11 navigation-network regions during virtual driving, generating a testable hypothesis of TPJ vehicle selectivity (p=0.005) — but with n=6 including two authors, qualitative validation only, and interpreter confirmation-bias risk.
Novelty 70% • Quality 60% • Limits: preprint, code not yet available, ±10 scaling unexplained.Postpartum depression affects roughly 6.5–12.9% of mothers, is most strongly predicted by antenatal depression, and is effectively managed with screening, psychological therapies, and lactation-compatible SSRIs — though the headline antenatal-treatment prevention figure (0% vs 92%, n=78) is implausibly large and needs replication.
Quality 80% (NEJM review) • Limits: narrative synthesis, self-reported screening data.Oral globin digest boosted hypertrophy and performance in zebrafish, mice, and C2C12 cells, with stage-specific peptides identified — but no isonitrogenous protein control, small n (4–7), missing raw data, and corporate conflicts of interest limit confidence, and the peptide-relay mechanism is inferential.
Novelty 90% • Quality 60% • Limits: no protein-control arm, relay mechanism untested (PHT1/mTORC1).Kota and Kurumba tribes each split into two cryptic clusters, and a rare North Asian mtDNA lineage (C4a2c1) dates to ~7 ka — but with no autosomal Northeast Asian signal, tiny per-cluster samples (N<5), and no data availability, the migration claim rests on a single surviving maternal line.
Quality 60% • Reproducibility 40% • Limits: Wikipedia citations, overstated title, no code/data release.Hydrothermal C,H,O,N,S reaction networks produced autocatalytic peptides, PNA monomers, and vesicles — modest incremental chemistry — but the universal “period-five” claim rests on analogy rather than direct test, with only two networks examined and no statistical support.
Quality 30% • Generality 20% • Limits: category errors in the Sharkovsky argument; universal claim untested.Godfray et al.'s landmark Science review frames food security as a threefold challenge (demand, sustainability, equity) and coined “sustainable intensification” — its 70%-by-2050 target became a reference point later validated by modeling, though the review is narrative with no methods, data, or code.
Quality 80% • Usefulness 90% • Limits: non-systematic, one author held ag-biotech shares (disclosed).No published N-body grid reports binary-survival fractions for habitable-zone planets, so the answer must come from new simulation — physical reasoning suggests survival collapses at periastron, meaning grids parameterized by separation alone can miss deadly eccentric configurations.
Open quantitative gap; fully reproducible by design with open-source integrators.Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.
Explore by Goal
Ask a Follow-Up
Trending