Scientific claims with supporting evidence, scope, and limitations — via web and MCP for Claude, Cursor, and compatible AI tools.
Updated September 19, 2026
Gut bacteria synthesize drug prodrugs in situ
Microbiome-aware pharmacokinetics with direct bioequivalence implications.
Bacteroidota enzymes methylated 15 of 170 carboxyl drugs and raised bezafibrate plasma exposure in gnotobiotic mice, but the work used mono-colonization models and has not been shown in complex human microbiomes.
An engineered conjugative plasmid blocked >99.9% of in vitro transfer and gut establishment of targeted AMR plasmids in mice, but mouse groups were small (n=3) and human-microbiome generality is untested.
Entry-exclusion shows self-limiting escape (resistance requires losing conjugation machinery), yet coverage across the clinical plasmidome and escape-mutant dynamics remain unmodeled.
Mechanistically plausible at 62% estimated likelihood, but the evidence supports neither human safety nor infection-control effectiveness, and plasmid loss may not equal loss of resistance.
A 1,500-fold conjugation-frequency gap between IncX1 and IncU plasmids suggests two spatial transfer regimes, which a single factorial biofilm design can statistically bracket.
Each link in the fiber-to-barrier mechanistic chain is individually supported, but no direct study tests UPFs themselves — the attribution rests on observational diet-pattern epidemiology.
Enformer identified 37,592 hominin–ape regulatory events with validated luciferase correlation (R = 0.81), yet ~35% show below-average divergence, suggesting acceleration-based scans under-sampled the causal repertoire.
Under gene-held-out splits, reported +2.3–14.3 pp codon-LM advantages collapsed to ~+1.6–2.2 pp or reversed, showing a general leakage-amplification problem in representation-learning evaluation.
Haploid sperm nuclei yielded 3,842 phased genomes and >25,000 called crossovers, confirming elevated rates in Fancm-deficient mice with per-cell recovery lower than prior methods.
~61% of the germline genome is discarded at the 12-cell stage at precise palindromic motifs, though the causal nuclease and elimination mechanism remain unidentified.
A. vaga repairs broken chromosomes progressively across generations via homolog-directed reconstruction (BIHER), though conclusions rest on one deeply sequenced clone and a diffusion model fit.
Heterogeneous motility distributions can reduce population energy expenditure while preserving exploration, but direct measurements of within-population heterogeneity remain a testable hypothesis, not an established phenotype.
Plausible but undemonstrated: supplied evidence shows motility is environmentally labile without directly measuring state distributions or their fitness consequences.
Localized resource use can suppress invaders while sparing residents, but supplied studies do not directly measure unused-niche preservation, leaving the design principle a testable inference.
Cell-type diversification proceeds through specialization, regulatory recombination, and physical circuit placement, but no single universal evolutionary principle is established by the evidence.
Local brain dynamics largely reverse between propofol anesthesia loss and recovery while global trajectories do not, but the rat study covers one anesthetic with possible neurovascular confounds.
RNAi knockdown of a nicotinic receptor subunit (chrnal-E) increased sleep in Cassiopea — first functional genetic sleep evidence in a pre-bilaterian — with small samples and no rescue experiment.
IRSp53 drives Arc capsid assembly and vesicle release that suppresses AMPARs in neighboring neurons in culture; intact-brain validation is absent.
Wixted's interference framework reframes forgetting as the adaptive price of continual encoding, though key claims rest partly on post-hoc reanalyses of archival data.
Resampling across ABCD, HCP, and UK Biobank confirms tiny effect sizes (median |r|≈0.01), and phenotype reliability — largely ignored — may govern required sample size more cheaply than scaling enrollment.
Durable AI memory records should carry type, provenance, confidence, and permitted use — persistence without provable origin is weak evidence, per the supplied design analysis.
Biologically plausible but unproven: evidence is mostly mechanistic and mouse work, with no randomized human trial showing survival benefit.
A bottom-up estimate puts tissue cost near ~41 kcal/day — tenfold below standard surplus prescriptions — but the model contains zero primary data and a key parameter derived from infants and animals.
Tirzepatide outperforms GLP-1 agonists on HbA1c (up to 2.4%) and weight (up to 22%), but the cited review is methodologically weak — corroborated by an independent 21-trial meta-analysis.
A convergent intrathecal response targets a specific EBV epitope with 99.36% specificity in validation, but it applies to roughly 8% of MS patients and causality is unresolved.
UK Biobank data show genetically predicted omega-3s raise AMI risk while measured levels appear protective, implying omega-3 levels may be a metabolic-health biomarker rather than a uniformly causal agent.
N=21,851 cross-sectional study links PM2.5 to sperm DNA damage, but age dwarfs the pollution effect (+14.4 pp DFI at ≥50) and smoking/BMI confounding is unmeasured.
A Navigation Guide review rates prenatal exposure as probably neurotoxic at moderate-quality evidence, with null results in underpowered studies systematically attributed to confounding — a one-sided interpretive asymmetry.
UV generates DNA damage while repair, apoptosis, and clonal competition determine which damaged cells persist — a coherent narrative review without systematic search methods or original data.
Durable 1918 H1N1 antibodies coexisted with zero SARS-CoV-2 cross-reactivity in 33 Sicilian centenarians, ruling out antibody cross-protection — but the cohort is essentially unreplicable.
CYP450-derived AFBO forms N7-guanine adducts that depurinate or persist as FAPy lesions, with adduct persistence rather than total count driving mutation — established for metabolism, less so for clonal selection.
otto-SR improved screening and extraction efficiency in a supervised workflow, but whether rare or context-dependent errors change final review conclusions was not demonstrated.
TxGemma-9B best separates actives globally while LlaSMol-Mistral concentrates more in the top 1% — model choice depends on the operational question, with no prospective compound validation yet.
Governance invariants (determinism, traceability, monotonic escalation) are testable software properties, but hazard detection remains proxy-based and partly circular.
Latent-diffusion imputation reached PCC 0.31 and clustering ARI up to 0.85 in COAD, with value concentrated in converting sparse assays into atlas-constrained state representations rather than reliable gene-specific inference.
A 4-amino-acid degron plus RNA switch enables real-time whole-plant RNA imaging, though only 2 of 12 switch designs worked and no independent replication exists.
The defensible visual story is a design hypothesis: endosomal release is siRNA-supported while tumor accumulation, receptor identity, and apoptosis remain formulation- and model-specific.
Across 4,940 passerine species, song complexity peaks in temperate open habitats and is not positively related to brain size after controlling for migration — refuting intelligence-signaling explanations at global scale.
Carollia bats phase-lock calls to low-noise phases of rhythmic noise with causal frontal-auditory-field evidence, though sample sizes were modest and a single species and masker frequency were tested.
Methane ebullition rises more than fourfold above 3,000 m elevation, matching a pressure-scaled degas model, but the evidence is correlational and biased toward the Tibetan Plateau.
No source reports survival fractions from an N-body grid — survival likely collapses at periastron rather than separation, requiring a new simulation to quantify occurrence corrections for ~half of stellar systems.
Hydrothermal networks produced autocatalytic peptides and vesicles, but the universal period-five theory rests on analogy rather than direct test, with no statistics and only two networks examined.
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.
Explore by Goal
Ask a Follow-Up
Trending